A curated, weekly updated collection of LLM-driven AI agents for biological and biomedical research, tagged by the five-dimension taxonomy of Qi et al., Briefings in Bioinformatics 2026. Updated 2026-10-23. † cited in the survey. Jump to 2026 · 2025 · 2024 · 2023.
Please cite: Qi C, Wang W, Jiang S, Liu Q, Song X, Fang H, Wei Z. Artificial Intelligence agents for biological research: a survey. Briefings in Bioinformatics 27(1), bbag075 (2026).
A reinforcement learning policy first generates candidate proton therapy plans, then LLM planning, chat and evaluation agents let physicians refine them conversationally with evidence from the reference plans. The planning agent justifies each recommended adjustment and the evaluation agent checks clinical metrics and constraint violations.
The paper evaluates different AI architectures for extracting clinical code status from discharge summaries, finding that single-agent systems perform comparably or better than more complex multiagent and reflective agent designs in capable models. The study highlights the importance of matching architectural complexity to model capability.
A constrained GPT-5.2 natural-language interface routes prompts to an adapted video segmentation foundation model and an axis-regression module to segment the left ventricle and estimate ejection fraction. With 20 labeled site-specific studies it reached a Dice of 0.940 and median ejection fraction error of 4.33 points on an independent Mayo Clinic cohort.
MedMCP is an agentic framework designed to make advanced medical image analysis tools accessible to non-technical users through natural language interaction. It demonstrates the potential for routine adoption of these methods by clinicians and researchers, enabling reproducible analyses with human oversight.
CS1 is a 4-billion-parameter model that answers many typed questions about a patient record in one pass with calibrated probabilities, deployed inside two agentic harnesses alongside a generalist model. On 963,408 held-out patients it beat larger open decision models on 13 of 14 tasks and generalized across 21 external datasets.
The paper evaluates the impact of multi-agent collaboration on reducing hallucination in clinical tasks, finding that while collaboration improves task performance, it does not consistently reduce hallucination. The effectiveness varies by model rather than topology or number of agents.
LeakBench is a framework designed to audit clinical modeling agents for prediction-time leakage and other critical issues, using synthetic tasks and real-world datasets. It demonstrates that even advanced models can fail in specific conditions without proper constraints or oversight, highlighting the need for robust auditing mechanisms.
EGIP is an interaction protocol designed to enhance the performance of clinical agents by enforcing structured evidence acquisition and decision-making processes. The system improves diagnostic accuracy across various datasets and language models, with clinician reviews indicating a significant increase in evidence-supported correct responses.
MedPAO-Assist structures radiology reports from dictation on a laptop CPU using an extract-ground-generate pipeline with SNOMED-CT grounding and a 0.6B-parameter model. It covers four imaging modalities with about 3.3 GB peak memory and roughly 31 seconds end-to-end latency.
FusionRare is a unified framework that integrates multiple AI paradigms to diagnose rare diseases, achieving superior diagnostic performance compared to existing methods. It demonstrates significant improvements in clinical decision-making by enhancing physician confidence and reducing diagnosis time.
The inClinico Agent is a machine-learning system that predicts the success of clinical trials based on registration-time features, providing a continuous Predictive Trial Index (PTI) for efficacy and safety/toxicity endpoints. The paper demonstrates high predictive accuracy with an AUROC of 0.828 for efficacy predictions and 0.809 for safety/toxicity predictions.
The paper evaluates the reliability of large language models in healthcare decision-making under uncertain conditions, using a synthetic benchmark with controlled perturbations. It highlights that while models perform well on clean data, their ability to escalate decisions when evidence is ambiguous or conflicting is significantly weaker.
The paper investigates the vulnerability of medical multi-agent systems to prompt injection attacks, where an attacker manipulates LLMs into incorrect decisions. It proposes a training-free defense mechanism that separates patient findings from clinical judgments and verifies each judgment against the findings alone, effectively reducing attack success rates while maintaining system accuracy.
TIR-EC is a reinforcement learning framework that integrates tools into large language models to improve their performance in emergency care prediction tasks. It demonstrates superior macro F1 scores compared to conventional ML methods and larger LLMs, while also showing more efficient use of clinical evidence.
TRACER is a multimodal system that integrates video and clinical records to provide minute-level action reports for patient care in hospitals. It demonstrates the ability to accurately detect patient actions and answer clinician questions, showing potential for supporting clinical decision-making without training on protected health data.
Evaluates retrieval-augmented agent configurations on 2,400 real MIMIC-CDM patient pathways across information gathering, diagnosis, treatment and guideline retrieval, with eight LLMs from 1B to 70B. Retrieval raised mean diagnostic accuracy from 55.6 to 64.3 percent, and autonomous query rewriting mattered most for retrieval fidelity.
An agent guided by an indexed clinical skill library synthesized longitudinal records of 811 myeloma patients and was compared against single-pass RAG, iterative RAG and full-context input. It reached 80.1% concordance with expert answers, above the roughly 76% ceiling of the alternatives, with larger gains on complex tasks.
HEM-Med is a hierarchical agentic memory framework designed for multi-center medication recommendation, addressing the challenge of transferring knowledge across hospitals while ensuring safety constraints. The system demonstrates improved performance in terms of both accuracy and safety compared to existing methods on real-world data.
PRISMATIC has specialty agents build and revise drug regimens for multimorbid patients on a blackboard, with agent influence weights adapted from past case outcomes and short-term discrepancy feedback. It targets prescribing errors that come from fragmented decisions across specialties.
This study fixes a four-evaluator verification stack and refinement policy for clinical note generation and swaps only the generator between a 4B on-device model and a frontier baseline. The small model reached 0.984 claim-level factual entailment against 0.992, with its remaining weakness in section arrangement turned into an explicit escalation decision.
TrialPD-Bench is an expert-annotated benchmark of 150 clinical trial protocol PDFs across four task categories for coding agents that select tools, run code and revise their approach. Across nine models in five harnesses, agents usually found the right evidence but failed at transcribing it into accurate structured output.
Oncology treatment planning is cast as an interactive reinforcement learning problem under partial observability, with process-aware rewards built from decision sequences in real colon cancer records. The agent must decide what information to gather before committing to a treatment, a setting where several trajectories can be defensible.
ProCARE is a framework that automates real-world studies by converting data profiles into Evidence Contracts and compiling study objectives into an intermediate representation, ensuring semantic consistency and traceability. It outperforms existing large language model agents in code executability and conclusion traceability on benchmark tasks, with validation through automated checks and expert review.
KGR-MAS is a multi-agent framework that uses knowledge graphs to enhance the reasoning and decision-making capabilities of large language models in medical settings. It demonstrates improved accuracy on clinical diagnostic tasks compared to larger, less specialized models.
LongCAF is a longitudinal context-constructive agentic framework designed to support hypertension self-management by integrating patient-specific context and interaction coordination. The system was evaluated in a real-world deployment over six months, demonstrating improved context completeness and high precision in safety escalation.
ActiveMedAgent is a framework that uses a cost-aware sequential logic to guide the acquisition of medical evidence, aiming to minimize costs while maximizing diagnostic accuracy. The system demonstrates its effectiveness by outperforming full-modality baselines in several benchmarks, highlighting its ability to identify when less information can lead to correct diagnoses.
MedZERO is a self-evolving framework designed for medical reasoning that combines an Examiner and a Reasoner to generate and solve complex medical questions, respectively. The system demonstrates significant improvements in accuracy on public medical reasoning benchmarks compared to existing methods.
The system uses a multi-agent architecture to answer patient questions about lung cancer screening CT results, demonstrating high accuracy in tool invocation and self-regulation. It was evaluated by physicians for clinical quality and patient understandability.
The Agentic AI Maternal Triage Framework uses predictive models and a retrieval-augmented generation-powered LLM to provide real-time risk stratification for maternal care, reducing alert fatigue in clinicians. The framework demonstrated high accuracy in simulation-based evaluations but requires further clinical validation before deployment.
PsyCIDRA pairs an interviewer agent that keeps working notes, loads expert skills and consults ICD-11 with a reasoning agent that reports hypotheses plus supporting, conflicting and missing evidence. It agreed with psychologists on whether to offer a hypothesis in 79.6% of blinded cases against 65.4% for direct prompting.
MedPrune treats medical visual question answering as a communication graph of department specialist agents and learns to prune both irrelevant agents and low-value connections. The sparsification is meant to improve reasoning while cutting the token cost of redundant agent chatter.
MedicalHarness evaluates the performance of large language models on medical tasks using different agent harnesses, revealing that the choice of harness significantly impacts outcomes. The study also identifies specific mechanisms within these harnesses that contribute to model performance.
The LifeLong Digital Twin system models a person's lifelong health trajectory by integrating various types of data into daily health states and lifelong health contexts. It demonstrates significant improvements in predicting disease trajectories and outcomes compared to existing methods, highlighting its potential for personalized healthcare.
ARISE is an autonomous planning framework designed to improve the interpretability and diagnostic performance of medical image analysis in inflammatory bowel disease (IBD). It demonstrates consistent improvement over baseline vision-language models by providing a transparent, step-by-step workflow that enhances clinical decision support.
The paper describes a human-AI collaborative workflow for mapping bilingual clinical text to SNOMED CT concepts, using an agent system with translation, abbreviation expansion, and vector-based retrieval modules. The study shows that the AI-assisted approach improves mapping accuracy and reduces time by about half compared to human-only or AI-only methods.
Sentinel is a confidence layer that reads a clinical text-to-SQL agent's reasoning, code and database outputs to refuse unanswerable questions, abort failing trajectories and withhold untrustworthy answers. On EHRSQL it raised the reliability score from +0.08 to +0.24 and delivered-answer accuracy from 54% to 69% as coverage fell.
An LLM controller maps clinical intent to 26 schema-constrained function calls across four radiotherapy phases while keeping patient and workflow state. It chose the intended function in 98.8 percent of 7,800 single-function executions and completed about 96.5 percent of cross-stage and real-patient workflows, with longitudinal state removal cutting completion to 84 percent.
The Orchestrator system uses a multi-agent architecture to assist in the diagnosis of secondary headaches by breaking down clinical reasoning into specialized agents, each focusing on specific red flag domains. The study demonstrated that this approach enhances diagnostic accuracy and transparency compared to single large language model methods.
PMxAgent is an open-source platform that automates pharmacometric analyses by integrating various tools into a Docker-based API server, enabling AI agents to perform tasks like non-compartmental analysis and exposure-response modeling. The system was benchmarked against other AI agents for accuracy and reproducibility in NCA workflows, demonstrating superior performance and deterministic results.
StrokeAgent is a multimodal agentic system designed to support decision-making in acute stroke care by integrating clinical records, laboratory results, ECG waveforms, and CT images. The system demonstrated high concordance with expert decisions and was rated highly for clinical utility and reasoning quality compared to large language models and neurologists.
The system uses a large language model to orchestrate an end-to-end workflow for PET/CT image interpretation and report generation, demonstrating high sensitivity in primary tumor detection but lower specificity in nodal involvement and distant metastasis assessment.
CXRAgent has a central director that orchestrates chest X-ray analysis tools, validates their outputs against visual evidence, plans the next diagnostic step and reasons over collected findings. It is built to adapt to new diagnostic tasks and to weigh how reliable each tool is.
The paper describes a multimodal AI system that assists audiologists in interpreting and reporting pure-tone audiometry and tympanometry results. The study evaluates the accuracy of the AI's performance in different modules, showing improvements in agreement rates for hearing loss types and acoustic reflexes when guided by specific skills or prompts.
EHR-RobustGym is a scalable and interactive environment designed to evaluate and train clinical agents on noisy electronic health records, aiming to improve their robustness in real-world scenarios. The system reveals significant performance drops when agents are exposed to noise, indicating the need for more robust training methods.
The paper describes an automated ICD coding system that uses a large language model and a multi-agent framework to assign diagnostic and procedural codes accurately. The system was evaluated using real-world medical records, demonstrating high precision and recall rates for both diagnosis and procedure codes.
STAM is a memory framework for clinical agents that tracks patient state transitions while preserving historical context. It demonstrated improved performance in downstream tasks and diagnostics compared to baseline methods across four longitudinal clinical benchmarks.
KlinikeBench evaluates language models in clinical settings beyond diagnostic accuracy by simulating full patient interactions, including history gathering and examination requests. The benchmark reveals a significant gap between diagnosis accuracy and interactive performance, with professional clinicians rating model-generated dialogues higher than real conversations.
PrivMeSA is a privacy-aware multi-agent system designed to manage medical consultations while minimizing patient data exposure. It uses reinforcement learning to balance task accuracy with privacy risks, demonstrating significant improvements in both clinical decision-making and privacy protection compared to direct remote consultation methods.
The paper presents a multi-agent framework for brain tumor instance segmentation and clinical report generation using YOLO architectures optimized with ASIO. The system demonstrates improved diagnostic precision and rapid inference latency, outperforming baseline methods in terms of mean Average Precision.
An LLM marks each finding present, absent or unknown while deterministic code computes score bounds and asks only the questions that could change the risk category, across six clinical calculators. On 1,200 synthetic emergency cases it matched asking everything with half the questions, while treating missing as normal under-triaged 8.5 percent of patients and an end-to-end agent asked irrelevant questions.
CLEAR-Med splits natural-language analysis of structured clinical data into an invocation agent that writes and runs SQL and a separate cross-provider validation agent that accepts, requests one repair, or abstains. On a 21-site neonatal encephalopathy table of 532 infants, the invocation agent answered 66.4% of benchmark queries correctly versus 12.0% ungrounded.
ECG-Scroll streams long ambulatory ECG recordings to an agent chunk by chunk so it must decide what to measure, store evidence as it passes and flag paroxysmal events without seeing the future signal. Because the signal is kept, answers are checked against objective ground truth with rule-based rewards.
Compares a single LLM call, five personas in one context and five blinded specialist agents merged by a moderator on curated case collections and 364 real emergency notes. The moderated team improved top-3 recall on curated cases but lowered it on real emergency notes, with blinded physicians finding no difference in error danger.
MACBT is a multi-agent system designed to support cognitive behavioral therapy (CBT) for depression, aiding clinicians in session preparation and documentation. The system includes a longitudinal memory module that tracks patient progress over time, demonstrating improved professionalism and clinical authenticity compared to existing systems.
SkinAgent AI is a multimodal framework that uses visual evidence and product information to provide skincare support, ensuring safety and privacy. The system demonstrated high accuracy in skin condition routing and skin-type estimation but showed limitations in tool selection and task completion.
DrGait is a training-free framework that uses Vision-Language Models to analyze gait videos for clinical applications, ensuring interpretability and accuracy by decoupling semantic reasoning from geometric perception through a structured workflow. It demonstrates the ability to generate transparent and audit-ready clinical reports while maintaining competitive diagnostic accuracy.
Eli5a v2.0 is an autonomous AI agent designed to generate patient information sheets for pediatric oral pathologies and rare head and neck tumors. The study found that Eli5a excelled in understandability, reading grade level, and actionability compared to other LLMs, highlighting its potential in creating accessible educational materials for complex medical conditions.
TrialAtlas is a multi-agent system designed to assist in clinical trial design and optimization by synthesizing literature, analyzing regulatory precedents, and providing integrated reasoning. It demonstrated superior performance compared to baselines in predicting the success of clinical trials and detecting deficiencies, as validated through both automated benchmarks and expert evaluations.
A two-level framework uses LLM planner, evaluator and supervisor agents to coordinate reinforcement-learning agents that adjust doses in lung IMRT treatment planning. On 62 retrospective lung cancer cases it improved plan constraint compliance and efficiency relative to knowledge-based planning.
SAFE-AI is a tool that uses an ontology-driven model to minimize hallucinations in large language models (LLMs) for detecting patient safety events, specifically medication errors, in medical records. The system demonstrated high accuracy and reliability in identifying epinephrine overdoses and delays in administration compared to human experts, showcasing its potential as a trustworthy clinical decision support tool.
The Genetic Severity Classifier is an AI agent that uses a ReAct-RAG architecture to classify the severity of genetic diseases based on ACMG and ACOG guidelines. It achieved high accuracy in classifying disease severity at both phenotype and gene levels, demonstrating its potential for standardized panel design in genomic screening.
MAESSCR is a medical education simulator in which several LLM agents play the patient, physical exam and diagnostic testing roles in text-based clinical encounters. In transcripts from 175 second-year students, clinician-educators found it followed scripted case details in most encounters but sometimes leaked diagnosis-changing information.
CTD-Agent is a retrieval-augmented multi-agent framework designed for integrated diabetes care, combining Western medicine, traditional Chinese medicine, and lifestyle management. In a single-case expert-rated pilot study, CTD-Agent received favorable ratings compared to general-purpose LLMs across several clinically important dimensions.
SpineAgent is a multi-agent system designed to generate spine MRI reports by integrating information from multiple imaging sequences. It demonstrates superior performance in spinal condition prediction and pathology localization, achieving state-of-the-art results through automated metrics and expert evaluation.
GeneResolver is a hybrid multi-agent pipeline that classifies rare disease cases based on their etiology and then prioritizes genes accordingly. It demonstrated improved accuracy in gene prioritization compared to existing tools, particularly when combining phenotype and genomic variant data.
The paper describes the development of an on-premise clinical AI agent that can perform diagnostic workflows autonomously, achieving high accuracy in disease diagnosis tasks. The system also includes a reliability framework to estimate uncertainty at decision time, which helps identify cases suitable for autonomous handling with high diagnostic accuracy.
Language model agents in primary care, specialty and subspecialty tiers consult 41 cohort-specific transformers trained on EHRs and make referral decisions through shared case memory without central control. On 3,000 test cases it reached 48.8 percent top-1 accuracy versus 19.2 percent for a single transformer and 12.7 percent for a single agent.
EMR is a self-evolving medical multi-agent system that uses experience mining and reuse to improve clinical decision-making. It demonstrates superior performance on medical reasoning benchmarks compared to existing systems, showing its ability to generalize across different specialties and LLM backbones.
ClinAgent is a conversational system that uses an agentic Retrieval-Augmented Generation (RAG) model to enable clinicians and researchers to query clinical trial information using natural language. The system integrates various tools such as ClinicalTrials.gov search, PubMed module, and a Python-based analyzer for efficient multi-turn interactions, demonstrating improved accessibility and synthesis of clinical trial data compared to manual methods.
A role-locked multi-agent LLM simulation of nurses, registrars and nurse managers was used to test how LLMs escalate deteriorating ward patients with and without ISBAR handover. Structured handover reduced hallucinations and shortened dialogues but did not reduce safety-critical omissions, with automated audits checked against ICU nurses.
Replays MedAgentBench tasks with all inputs fixed and compares the orders an agent actually files rather than its score, using six reliability metrics over 1,000 runs. Nearly all ordering groups produced different order sets across identical runs, often while the benchmark recorded the same verdict.
Uses a simulator of full emergency-department shifts to show agents diagnose well but miss timely critical actions, then adds a harness that rewrites its operating manual between shifts, a clinical skills library and three subagents splitting the patient queue. On held-out batches it improved critical-action correctness by 22 percent while keeping diagnostic accuracy.
MedRoundsQA is a benchmark for evaluating AI agents in multi-turn medical consultations, converting board-exam cases into structured dialogues. It demonstrates that moving from single-turn to multi-turn interactions significantly degrades diagnostic accuracy and highlights equity risks associated with patient personas.
AAD-Agent is an AI agent designed to assist in the rapid diagnosis of acute aortic dissection before imaging confirmation, using large language models for enhanced diagnostic accuracy. The system demonstrated consistent performance across different LLMs and settings, showing promise as a low-cost triage tool but with limitations on confirmatory diagnosis.
This platform combines a simulated patient agent, a tutor agent giving Socratic prompts without revealing the diagnosis and a turn-level evaluator agent for clinical interview training. In a randomized study with 100 medical students, the multi-agent condition improved final OSCE-based scores though not diagnostic accuracy.
The paper introduces a multi-agent AI system for automatic ICD code generation in medical claims, which includes agents for input extraction, coding with enhanced GPT-2 XL model, and report compilation. The system was evaluated on real-world data achieving 95% accuracy, demonstrating its potential to streamline clinical decision support and reduce human error.
The study evaluates an LLM agent's performance in clinical data analysis across five stages, using a dataset of patients with neovascular age-related macular degeneration. The agent reliably generates research questions and statistical analysis plans but requires expert oversight for code execution accuracy.
The paper introduces an agentic workflow that converts unstructured clinical free text into interpretable features for predictive tasks, using large language models (LLMs) and local extractors. The framework improves prediction accuracy while reducing computational costs compared to direct LLM processing of full corpora.
MedCP is an agentic AI framework that integrates large language models with electronic health records and a biomedical knowledge graph, enabling interactive human supervision for multi-step reasoning. It demonstrated improved performance in clinical research tasks and factual accuracy questions, while also facilitating the rapid replication of observational studies.
The Debate-Mixture-of-Agents (DMoA) framework uses a structured multi-agent approach to improve iterative diagnostic reasoning in complex clinical settings. It demonstrated significant improvements in diagnosis accuracy and safety rate compared to single-model approaches, highlighting the potential of multi-agent systems for clinical decision support.
VERDICT is an AI system designed for clinical trial matching that uses SMT solvers to ensure consistent application of decision policies and generation of accountable rationales. It outperforms other systems in terms of decision accuracy, policy consistency, and the quality of generated rationales as preferred by clinicians.
The system uses a multi-agent architecture to support clinical decision-making by retrieving patient similarity data from an electronic health record database and using large language models for analysis. It demonstrates the ability to generate temporally-aware clinical analyses through a case study, showcasing its potential in real-time clinical reasoning.
FairMedAgent measures how often a clinical LLM agent changes its action on identical repeated inputs, establishing an instability floor against which counterfactual fairness flip rates must be read. Across six models the floor ranged from 2.5% to 23.7%, and majority voting over five draws removed about 39% of it.
The paper introduces a multi-agent peer-reviewed reasoning method for medical question answering, where large language models independently generate and evaluate each other's chain-of-thought reasoning to produce final answers. This approach outperforms single-model and majority voting methods in terms of accuracy on benchmark datasets.
PlanningCopilot is an agentic system that integrates large language models and a validated Eclipse Scripting API module to autonomously generate treatment plans for lung radiotherapy. The study demonstrates its ability to produce clinically acceptable plans, meeting dosimetric requirements and improving optimization efficiency through self-learning.
The SAGE system uses a locally deployed large language model to autonomously optimize radiotherapy plans for prostate cancer patients, demonstrating its ability to generate clinically acceptable treatment plans. The study found that incorporating intra-case memory significantly improved sparing of the penile bulb while maintaining target coverage.
The Planning Copilot is an LLM-driven framework that refines radiotherapy treatment plans generated by knowledge-based planning models, improving their clinical goal achievement and portability across different institutions. It demonstrates significant improvements in plan quality metrics without the need for model-specific tuning or retraining.
ClinTraceBench provides 385 MIMIC-IV-derived verified dialogues with event-level provenance and nine task types to test whether compressed patient histories preserve longitudinal signal. Comparing eight history strategies, including two agentic memory systems, across four models it found compression loses relational facts that the agents fail to recover.
A specific LLM screener flags concerning growth trajectories in pediatric records, then an orchestrator runs a panel of agents, corrected by a trainer-injected learner agent, to rank differential diagnoses. In walk-forward simulation the screener reached 98 percent specificity a year before diagnosis, and expert review confirmed suspicion in a third of flagged controls.
The paper evaluates various AI methods for recommending CPT and HCPCS codes from clinical notes, focusing on recall within a fixed number of recommendations. It finds that zero-shot GPT-5.6-Sol achieves the highest recall among tested models.
This work defines an episode-level evaluation protocol for healthcare agents that scores state continuity, evidence traceability and escalation decisions rather than one-shot answers. It is instantiated as four task templates covering documentation updates, evidence retrieval, patient messaging and triage handoff, with cost-sensitive treatment of escalation errors.
DIASENTINEL is a multi-agent system designed for diabetes risk screening that integrates guideline-based recommendations with real-time patient data, ensuring reliable and auditable clinical decision support. The system demonstrates the capability to generate detailed reports grounded in ADA guidelines while maintaining patient privacy.
MASGR is a multi-agent system designed to improve medical referral decisions by constructing structured graphs that integrate evidence from various patient data sources. It demonstrates superior performance in handling complex cases compared to existing models, highlighting its capability to balance chronic disease management with emergency interventions.
GPAgentBench-2K is a benchmark designed to evaluate large language model agents in complex clinical decision-making scenarios, using real-world GP encounter data. The system reveals significant performance degradation and safety issues among leading models when dealing with more realistic action spaces.
FedEHR-Agents is a federated learning framework that enables autonomous clinical agents to collaboratively model electronic health records while preserving patient data privacy. It demonstrates improved performance in diverse clinical prediction tasks compared to local and federated baselines, highlighting the potential of federated autonomous clinical intelligence.
CaRDI is a planning framework for interactive medical diagnosis that selects questions based on clinical risk rather than uncertainty reduction, aiming to minimize diagnostic errors and severe omissions. Experiments demonstrate that CaRDI improves accuracy while requiring fewer questions compared to existing methods.
The GraphRAG framework integrates large language models with a pharmacotherapy knowledge graph to enhance the safety verification of clinical decision support systems. The system was evaluated on benchmark cases and showed improved accuracy in answering clinical safety questions compared to traditional retrieval augmentation approaches.
The paper describes an interoperable web platform that integrates clinical data using FHIR and LLMs to support medical data analysis and decision-making. The system demonstrates improved semantic interoperability and contextual reasoning, evaluated through both quantitative metrics and qualitative assessments.
MARLA is an autonomous agent designed to facilitate medical AI research by translating high-level clinical goals into executable research loops, coordinating specialized code agents, and refining strategies based on feedback. It demonstrates the capability to autonomously conduct complex medical AI tasks with reduced reliance on AI expertise.
MTDiag is a physician-validated multi-turn diagnostic dialogue dataset built from DDXPlus, MIMIC-IV and case reports, normalized to UMLS and ICD-10, with a patient utterance generator. It adds knowledge-grounded metrics for judging models as interactive diagnostic agents rather than static answerers.
EviDx builds interactive diagnostic environments from raw clinical cases and runs role-specialized agents with evidence tools and an observer harness that decides when enough evidence has been gathered to stop. A three-level evaluation covers execution robustness, reasoning dynamics and diagnostic outcome.
MediSkill-Evo is an interactive clinical agent that self-evolves its process knowledge to improve evidence-grounded decision-making in medical scenarios. It demonstrates significant improvements in diagnosis accuracy and treatment intent coverage compared to existing agents, while also reducing critical failures under stress conditions.
TrialCode Agent is a hybrid-large language model system designed to translate free-text clinical trial criteria into computable code sets, supporting various medical terminologies. The study demonstrated that while retrieval and broad family expansion can improve performance in specific cases, deterministic verification significantly enhances the validity and readiness of generated code sets for database use.
KMGen is an automated framework that extracts Kaplan-Meier curves and generates synthetic individual patient data, including adverse event trajectories. It demonstrates high accuracy in curve extraction and effective generation of realistic patient data across various clinical trial conditions.
DeepSAGE represents a first cognitive behavioral therapy session as eleven stages, with a controller judging stage completion and a reinforcement learning policy choosing therapeutic intentions that steer LLM responses. Against six alternatives it gave higher simulated client engagement and the best balance of stage-goal completion and efficiency, with expert review of the transcripts.
Combines medical knowledge retrieval and similar-patient reasoning across specialised analyst agents feeding a smaller integrator agent, varying roles while holding retrieval fixed. For in-hospital mortality it matched closed-model prompting on F1 while flagging more true high-risk patients, with smaller gains for readmission.
BioMed-Agent-RL orchestrates clinical grounding, lesion segmentation and reasoning expert models through a policy trained with preference optimization and group relative policy optimization. It is aimed at multimodal clinical cases where static agent pipelines mishandle noise and modality mismatch.
FRAC-MAS is a multi-agent system designed for automated and explainable bone fracture diagnosis, integrating deep vision models with conformal prediction to ensure statistical reliability. The system demonstrates improved clinical interpretability and safety by triaging cases into high-confidence and uncertain categories, outperforming single-agent systems in patient preference studies.
HiMA-MDD is a hierarchical multi-agent system designed to interpret multimodal clinical interviews for depression detection, breaking down the task into three layers of agents that handle evidence collection, symptom scoring, and profile verification. The paper shows that HiMA-MDD outperforms existing methods on the E-DAIC dataset.
MSM-Mem is a memory framework for medical AI agents that enables them to retain and utilize clinical experiences over time, improving their decision-making abilities. Evaluations show consistent performance improvements as the agent accumulates more experience.
ATHENA is a framework that uses neural architecture search to optimize Transformer-based models for clinical prediction from EHRs, enabling efficient and effective model selection across different hospitals. It demonstrates superior performance compared to baseline methods in various clinical tasks, showcasing its capability to transfer knowledge and maintain computational efficiency.
DDx-Finder is an open-source framework that uses large language models and the Model Context Protocol to extract clinical state information from electronic medical records and literature, aiding in differential diagnosis. The system demonstrates its potential for providing accessible and systematic diagnostic support by addressing operational challenges such as resource constraints and privacy concerns.
LabBridge is a large language model-based framework that standardizes laboratory test terminologies to the LOINC standard, enhancing semantic interoperability across different languages and institutions. The system demonstrated high accuracy in mapping laboratory tests from both Chinese and English datasets, outperforming baseline methods.
AGVF generates medical-necessity appeals through five agents for policy formalization, evidence retrieval, gap analysis, adversarial critique and gated synthesis, with a citation gate blocking unsupported claims. The authors prove the refinement loop terminates and validate the implementation on 1,000 synthetic appeal cases.
This framework sits between a patient and an unchanged downstream model, using a knowledge graph to build plausible hypotheses for an underspecified health question, find the missing patient details and ask targeted follow-ups. It was tested with five language models on two controlled ambiguity benchmarks including symptom-based diagnosis retrieval.
Compares single-agent and multidisciplinary-team LLM settings, with and without cohort summary statistics in context, for predicting neurological complications after acute type A aortic dissection surgery. Validated on 763 retrospective and 120 prospective patients against machine learning models such as random forest.
ED-Triage-Agent is a multi-agent framework designed to assist in emergency department triage by providing provisional acuity assessments based on symptom data alone and refining these assessments once vital signs are available. The system demonstrated high accuracy and sensitivity in prioritizing patient care, with performance comparable to existing benchmarks.
A locally deployed pipeline of rule-based and LLM agents restructures CT reports into anatomical sections and flags findings-impression mismatches, gender-anatomy conflicts and undocumented critical results. It flagged 14.1 percent of 638 reports, and two radiologists rated QA good or excellent in 84 percent of reviewed reports with no fabricated content.
ICD-Deepresearch is a clinical decision support system that forecasts future diagnosis codes for patients using a combination of predictive models and medical search tools. It demonstrates improved precision and recall compared to standalone systems, with higher usefulness ratings from physicians for the evidence it retrieves.
The PHA is a specialized metabolic health agent that generates accurate and trustworthy clinical reports by grounding its generation in four layers: HPP cohort data, domain-expert tools, declarative skills, and automated evaluations. It demonstrates significant improvements in structural completeness, numerical accuracy, and compliance with clinical language standards compared to general-purpose models.
DeLiriuMAgents combines a machine learning risk model, LLM specialist agents for emergency medicine, neurology and psychiatry, and retrieved evidence to predict delirium in emergency critically ill patients. It was validated on MIMIC-IV, a two-hospital Peking University cohort and eICU, with accuracy between 0.67 and 0.75, and clinicians reviewed its reasoning.
MARC v1 is a multi-agent framework for clinical AI reasoning that replaces monolithic large language model prompting with specialized agents for extraction, reasoning, answer generation, and evaluation. The system demonstrates the ability to coordinate these tasks deterministically, making it interpretable and accessible even to non-programmers.
GatorOnco is an agentic large language model designed for colorectal cancer treatment planning, which integrates time-sensitive clinical guidelines through a retrieval-augmented generation approach. In a randomized evaluation by oncologists, GatorOnco demonstrated expert-level performance in treatment planning and received higher ratings for readability and completeness compared to human experts.
AMIE was extended to real-time video consultations as a Gemini-based multi-agent system combining low-latency dialogue, clinical reasoning and audio-visual perception. In a randomized OSCE with 30 primary care physicians, 15 patient actors and 100 scenarios, evaluators rated the video configuration against physicians and the text-only version.
ATLAS is a framework that uses a coupled graph-policy distillation approach to personalize medication safety for older adults with multiple conditions. It structures guideline evidence into a medication-safety graph and updates this based on patient-specific information, achieving superior performance in decision-making benchmarks compared to existing systems.
The paper introduces an agentic AI-driven immersive simulation for training in HDR brachytherapy, using VR and mobile computing to create a risk-free learning environment. The system integrates a knowledge-aware assistant that provides real-time guidance based on clinical guidelines, validated through prototype deployment with promising results.
The paper evaluates DeepSeek-R1, a conversational large language model deployed in outpatient vertigo clinics to assist with vestibular diagnosis. The system showed high diagnostic concordance rates (around 80%) compared to specialist diagnoses, particularly for benign paroxysmal positional vertigo and Ménière disease.
CliniCARE-Bench has agents audit 750 MIMIC-IV-derived cases from 25 clinician-validated scenarios through a logged tool environment, returning yes, no or one of two kinds of indeterminate verdict. It scores evidence and policy grounding, process adherence, calibrated abstention and efficiency against independently adjudicated references.
ResidencyRL trains a clinical agent through simulated multi-turn encounters of up to 60 turns and 8 tool calls against rewards for diagnosis, management, communication, documentation and safety. On held-out evaluations it raised diagnostic accuracy under adversarial simulated patients from 81.0% to 88.0% and cut missed red flags by 31%.
The Nimblemind Multi-Agent System (nMAS) is an automated pipeline for heart-failure feature engineering from electronic health records, using a multi-agent architecture to integrate fragmented data and clinical guidelines. It demonstrated improved predictive performance in phenotyping heart failure subtypes and achieved high scores on rubric assessments of evidence support and methodological soundness.
QuanTiMedAI is a quantum-enhanced time-series model guided by an agentic AI for predicting mortality in cardiac arrest patients. It uses a large language model to discover clinically relevant features and a compact quantum recurrent network for temporal prediction, demonstrating superior performance with fewer parameters compared to existing methods.
ECHO is a locally-deployable conversational health assistant designed for chronic care management, integrating an agentic chatbot with clinical tools and safety layers. It demonstrates high accuracy in tool execution and safety classification, as well as effective multimodal speech assessment.
The paper introduces an LLM-driven agentic framework that can reconstruct, optimize, and customize deep-learning models for radiological image analysis on a consumer-grade PC. The system demonstrated improved performance in mammography analysis across multiple datasets and outperformed radiologists in a reader study.
AgentEYE is an autonomous multimodal AI agent designed to integrate ocular imaging data and medical evidence for ophthalmic diagnosis. It demonstrated higher diagnostic correctness, completeness, and safety compared to large language model baselines in both internal benchmarks and blinded evaluations by ophthalmologists.
DR. INFO is an agentic AI clinical assistant that retrieves from a curated clinical knowledge base, with an additional MTS-specific retrieval configuration. The study shows that DR. INFO significantly reduces undertriage rates compared to general-purpose language models like GPT-5.1 when used for emergency triage.
DRAgent is a multi-agent system designed for diabetic retinopathy treatment planning, integrating ophthalmology and endocrinology expertise. The system demonstrates improved performance over standalone large language models in generating individualized patient plans, as validated through both automated metrics and human evaluations.
CANOE is a multi-agent neuro-symbolic framework designed for care plan coordination in healthcare, using an arena-based argumentation system to resolve conflicts and generate acceptable intervention plans. The evaluation demonstrates that CANOE provides clinically correct and safe recommendations while also offering transparent explanations through its argumentative structure.
DoctorAgents builds and optimizes machine learning pipelines for small temporal clinical datasets using generation, validation and refinement agents that pass natural-language feedback as textual gradients. Across several clinical tasks it beat established AutoML baselines while yielding more interpretable task-specific representations.
This study tested whether committees of Gemini agents deliberating on clinical cases can be gamed by cues a clinician would ignore, across seven cohorts of text, imaging and ICU tabular data. Individual agents resisted the cues, but when two peers asserted the same wrong answer the agent under test adopted it 38% of the time, and only one of three oversight agents detected this reliably.
The Agentic OODA Framework is a multi-agent system designed for clinical decision support that breaks down the decision-making process into modular, stateful agents. It was evaluated in stroke care and showed significantly higher reliability and consistency compared to general LLM models.
The paper evaluates multi-agent LLM architectures for maintaining guideline compliance in clinical decision support systems under knowledge base distribution shifts. It shows that a stateful graph architecture with an independent safety floor mechanism can maintain high compliance even when the retrieval quality degrades, though it introduces additional latency.
The system translates natural-language clinical research requests into executable R workflows, ensuring privacy by retaining patient data locally. It was validated in a study on metabolic dysfunction-associated steatotic liver disease, demonstrating its capability to execute complex statistical analyses without manual coding.
MedSR-Copilot is a multi-agent system designed to automate medical systematic reviews under the PRISMA workflow, covering literature retrieval, screening, data extraction, and evidence synthesis. It demonstrates significant improvements in efficiency and accuracy when used as a copilot in human-AI collaboration studies.
The paper presents a large language model agent that assists in radiotherapy treatment planning for head-and-neck cancer patients, demonstrating its ability to generate plans comparable to those made by certified dosimetrists. The study highlights the potential of using reference-free LLMs to streamline clinical workflows and reduce variability in treatment planning.
An on-premises system layers data and knowledge, retrieval and tool-augmented reasoning, multi-agent collaboration and clinical application with safety governance for digestive malignancy management. Used before multidisciplinary discussions for 25 older patients, its conflict-detection module fired in five cases involving cardiovascular risk or antithrombotic therapy.
CORA is a retrieval-augmented large language model designed to provide source-linked clinical support for physicians, improving their decision-making accuracy. The study revealed that while CORA enhances physician accuracy and adoption of correct advice when supported by citations, it also reduces resistance to incorrect advice when such advice appears citation-supported.
A clinical knowledge graph routes therapeutic decisions for an LLM across eight psychotherapy modalities, tested with guideline vignettes and progressively disclosing simulated patient agents. It routed 116 of 117 vignettes appropriately, detected 96.7 percent of adversarial crisis probes and 98.5 percent of external suicidal-ideation dialogues, with fully auditable decisions.
The paper presents a multiagent large language model framework for recommending emergency treatments in acute ischemic stroke, which integrates clinical data and imaging information to provide accurate treatment suggestions. The system was validated using real-world cases and showed improved accuracy and safety compared to single-agent models, with potential to enhance decision-making among less-experienced physicians.
MA-RAG is a multi-agent framework designed to generate accurate and temporally consistent summaries of longitudinal Parkinson's disease assessments. It improves factual correctness and reduces hallucinations compared to baseline methods, as validated by clinical experts.
The authors built a Model Context Protocol server that lets LLM clients query the CIViC cancer variant knowledgebase through its API in natural language. In a benchmark it improved retrieval of clinical significance and ran much faster than GPT-5 Agent Mode browsing the web.
The paper describes the development and evaluation of a multi-agent system using large language models for advance care planning training in Chinese healthcare. The system, which includes an assistant, vignette, and evaluator agent, was evaluated through both automated metrics and human assessments, showing significant improvements over baseline models.
Swin-Qwen3 is a vision-language framework designed to generate accurate radiology reports from chest X-rays, incorporating a three-stage process for drafting, verification, and refinement. The system demonstrates modest improvements in clinical accuracy and factual consistency compared to simpler models but still requires further validation before clinical deployment.
ClinLens is a benchmark designed to evaluate large language model agents on longitudinal clinical data science tasks, focusing on patient-centric workflows. The system assesses the ability of these agents to maintain context and integrate multimodal evidence over time, revealing that while some agents perform well in execution success rates, they struggle with strict pass rates due to issues like incorrect patient episode grounding.
GuideSkill is a system that compiles clinical practice guidelines into executable functions to assist in differential diagnosis, improving diagnostic accuracy. It demonstrates significant performance gains over existing methods and receives positive expert evaluations for its clinically sound and reliable rules.
PatientAgentBench is a benchmark framework that evaluates AI agents designed to interact with patients in healthcare settings, assessing their ability to converse, reason about health records, and act on behalf of the patient. The system revealed significant clinical gaps among various models, particularly in triage quality, clinical safety, and workflow accuracy, highlighting the need for more comprehensive evaluation standards beyond static benchmarks.
Per-signal agents read ECG and PPG windows using measured features, assemble a patient rhythm profile and reason over retrieved guidelines, with a critic agent checking each conclusion against its cited guideline. It ranked first on diagnosis, significance and management tasks under both cardiologist and LLM-judge scoring.
GraphRareBench is a benchmark of 2,365 ontology-derived rare disease cases with fixed candidate pools, graph-defined hard confounders and source-linked evidence records. Tool-using agents reached MRRs around 0.72-0.75, close to supervised rankers, but differed greatly in how much of the supporting evidence they actually examined.
Pathology-CoT is an AI framework that captures and learns from the interactive diagnostic behavior of expert pathologists to navigate whole-slide images, propose regions of interest, and provide explainable diagnoses. The system demonstrated superior performance in detecting gastrointestinal lymph node metastasis compared to existing vision-language models.
The STACER Agentic System uses AI agents to simulate psychiatric consultations and provide structured feedback, demonstrating technical feasibility in supporting psychiatry residents' consultation skills. The system showed high behavioral fidelity and medical faithfulness but had limitations in facilitating rapport building.
PULSE is a knowledge-enhanced, domain-aware large language model agent designed to improve atrial fibrillation patient self-management. It integrates multimodal inputs and clinical knowledge bases to enhance the accuracy, integrity, utility, safety, empathy, and clarity of medical outputs compared to leading base LLMs.
A hierarchical multi-agent system combines a local guideline retrieval module with a cloud reasoning model to produce management recommendations for diabetic kidney disease. Twelve physicians rated its outputs on 267 multi-center cases as more accurate and safer than unaugmented LLMs, especially for renal-stage medication contraindications.
CardioAgent is a clinical decision support system that integrates an LLM with deterministic tools and guideline retrieval to enhance cardiovascular diagnosis accuracy. In evaluations, CardioAgent demonstrated improved accuracy over generic models by incorporating structured workflows and traceability features.
This proof-of-concept compared rule-based, single-agent and multi-agent LLM configurations, with a safety agent for abstention, for pre-screening patients for NSCLC trials. On 90 held-out synthetic patient-trial pairs, the study focused on which setups reduced false inclusion of ineligible patients.
The SHARE system uses a human-AI co-learning framework to accurately label disease activity from electronic health records, improving the quality of gold-standard labels and identifying ambiguous cases for further review. It demonstrates scalability in labeling large datasets while maintaining accuracy through efficient resource allocation.
The TCGA-Agent is an agentic system that uses a large language model to query and analyze The Cancer Genome Atlas clinical data, providing natural language responses with statistical analysis. It outperforms other methods in accuracy and completeness, especially for fields not covered by the curated CDR.
The Multi-Stage Iterative Clinical Reasoning Agent framework uses a multi-agent architecture to improve clinical reasoning in complex nephrology cases by iteratively refining its responses. The system outperformed both baseline large language models and static retrieval-augmented generation methods, demonstrating the importance of dynamic refinement for longitudinal medical reasoning.
RadFabric is an agentic AI system that integrates multiple CXR analytics models and VLMs to provide explainable diagnoses. It demonstrated superior performance in detecting various lesion types, especially uncommon ones, compared to existing state-of-the-art models.
The paper describes a compound AI platform for autonomous radiotherapy treatment planning that integrates multi-agent large language models and dose prediction tools. The system was evaluated on 60 retrospective cases, demonstrating its ability to achieve dosimetric criteria comparable to clinical reference plans.
PAUSE-Agents is a clinician-in-the-loop multi-agent AI system designed to draft structured handoff briefs for ICU-to-ward transfers, using role-specialized agents and safety checks. The system was evaluated by physicians who reviewed 100 drafts, showing high accuracy with minimal errors and omissions.
The paper introduces two autonomous agents, Iteration and Evolution, which optimize AI-enhanced electrocardiography models for structural heart disease by proposing and evaluating code changes. These agents improved the model's performance in external evaluations without requiring new data or human intervention.
JADE-Plus is a multimodal, agentic retrieval-augmented generation framework designed to assist in the diagnosis of jawbone lesions. It integrates vision-language models and an agentic verification loop for improved diagnostic accuracy and stability, outperforming other systems in clinical decision support tasks.
The paper evaluates agentic large language model systems for breast cancer treatment recommendations using real clinical cases and rubrics. The best system, Claude Opus 4.8 with D&C+SA pipeline, showed mixed performance improvements depending on the tool use and agent autonomy, highlighting persistent clinically relevant errors despite generating some clinically relevant recommendations.
The paper evaluates large language models' ability to engage in clinical reasoning by developing an agentic evaluation framework for hematologic oncology, finding that the best-performing model achieved only 68% accuracy. It highlights a failure of information-seeking under uncertainty as the primary limitation rather than medical knowledge deficits.
RadGuide AI is a nuclear medicine agent that integrates QA items, literature, and clinical trial data to provide traceable decision support for radiopharmaceutical development. The system demonstrated high accuracy in technical evaluations compared to other models, showing promise for clinical use after further validation.
PathFound is an agentic multimodal pathology model that makes an initial diagnosis, actively seeks more slide evidence and then refines its decision, using RL-trained reasoning models. Adding this evidence-seeking strategy improved diagnostic accuracy across several multimodal models.
SAGEAgent decides sequentially which diagnostic modalities to acquire for a cancer patient, weighing predictive gain against invasiveness, using tools that verbalize predictions plus episodic and semantic memory. It treats multimodal survival prediction as an acquisition decision problem rather than assuming all data are present.
LongMedBench is a benchmark for evaluating medical agents in long-horizon clinical decision-making using real-world EHR data. It assesses how well these agents can aggregate and use patient information over time, revealing that while they perform well with explicit timestamps, implicit temporal reasoning remains challenging.
MedicalAgentsBench is a benchmark designed to evaluate complex medical reasoning tasks using both internalized and externalized agent-based approaches. The study demonstrates that combining these methods can significantly improve performance in clinical decision support, with the best results achieved by integrating multi-agent systems onto an advanced single-agent model.
Octopus is a neuro-symbolic framework that uses multi-agent swarms to autonomously generate and test therapeutic hypotheses in colorectal cancer. It successfully identified Insulin-like Growth Factor 2 (IGF2) as a predictive biomarker for 5-Fluorouracil sensitivity, with the discovery validated through rigorous statistical analysis and independent mouse model experiments.
Agents build and then apply a severity rubric for type 2 diabetes from electronic records, covering domains such as glycaemic staging, beta-cell function and social determinants. On a synthetic cohort its open-weight version matched the proprietary pipeline, agreed moderately with algorithmic severity indices and separated mortality risk, differing from deterministic execution of the same rubric.
An LLM agent repeatedly writes EEG feature-extraction code, runs it, scores a tabular classifier and uses the diagnostics to refine the next feature module for interictal spike detection. On the VEPISET dataset the generated features reached an AUROC of 0.935, and artifact-aware search improved balanced accuracy.
ClawBench is a framework for evaluating large language models in the context of ACMG/AMP variant interpretation and calling, focusing on validity, safety, provenance, and reproducibility rather than just accuracy. It demonstrates that different types of errors arise from distinct architectural layers within the genomic workflow.
Sofiya is an LLM voice agent that phones patients before cardiac catheterization to give instructions, collect clinical data, answer questions and hand off to nurses. Across 1,606 calls to 1,431 patients it completed about 87% of calls in routine nurse-led use.
Introduces a benchmark of 2,268 clinical cases requiring single, multiple or nested medical calculators, and an agent framework that selects and chains calculator tools with structured validation of parameters. The framework scored best across all three settings compared with open, closed and medical LLMs.
KGRD is a diagnostic support framework for pediatric rare genetic disorders that uses knowledge-driven and data-driven inference over patient-level genomic and phenotypic data. It demonstrated improved diagnostic accuracy in a validation benchmark, increasing the mean Bond score of top-ranked diagnoses and raising correct identification efficiency.
TRUST is a dialogue system powered by large language models designed to conduct structured diagnostic interviews for PTSD, replicating clinician behavior. The study demonstrates that TRUST performs comparably to human clinicians in terms of clinical quality.
Doctor Xiao Yi is a specialized AI agent designed to provide post-transplant medical and psychological support during off-hours, addressing the limitations of general large language models. The study demonstrated that Doctor Xiao Yi effectively bridges the gap in care for transplant recipients by remaining active and responsive during non-working hours.
NOVA is a set of locally run LLM agents for knowledge questions, user interface actions and delegation that acts as a hands-free interface for a low-cost surgical navigation system. Agents were tested on two semisynthetic benchmarks and integrated into a working prototype.
ReclAIm is a multiagent framework designed to monitor and correct performance decline in medical imaging AI models through automated evaluation and fine-tuning. The system successfully detected performance discrepancies and restored model accuracy, demonstrating its potential for continuous improvement of AI systems in clinical settings.
RESCUE is a multi-agent system that uses machine learning to screen electronic health records for patients who may benefit from rare disease diagnostic testing. It demonstrated strong performance in identifying appropriate candidates, with high precision and likelihood of subsequent genomic assessments.
MARCH is a multi-agent system designed to generate 3D radiology reports for CT scans, emulating the hierarchical structure of radiology departments. It showed significant improvement in clinical fidelity and linguistic accuracy compared to existing models on the RadGenome-ChestCT dataset.
The Delirium Screening AI Agent uses a knowledge graph derived from clinical guidelines to autonomously select appropriate delirium assessment tools and generate guideline-aligned care recommendations. The agent demonstrated high accuracy in tool selection and conclusion generation, outperforming baseline models, with human experts validating its outputs for coherence and relevance.
Trains LLMs that actively search for evidence using clinician-designed, automatically refined rubrics as step-level process rewards in reinforcement learning. Across seven medical benchmarks it beat agentic-search baselines by 9 percent on average and raised evidence-based medicine rubric scores.
HemaGuide is a locally deployable LLM agent that structures clinical documents, routes cases to guideline, advanced or molecular decision modes, and grounds advice in guideline flowcharts and over 2,000 past tumor board cases. In blinded evaluation on 45 complex hematologic cases it improved concordance with tumor board decisions across six LLMs.
DDIAgents is a multi-agent framework designed to predict drug-drug interactions by dynamically routing relevant knowledge sources and orchestrating specialized expert agents. The system demonstrates superior performance on realistic DDI prediction benchmarks compared to existing methods, showcasing its ability to organize heterogeneous scientific data for adaptive reasoning.
HealthAgentBench is a benchmark suite for evaluating AI agents in complex, long-horizon healthcare tasks. It assesses the ability of agents to navigate real-world clinical workflows and reveals that even advanced models like Codex GPT-5.5 struggle with certain task categories.
PHO-Agents is a multi-agent system that uses large language models to predict health outcomes from electronic health records, improving prediction accuracy and interpretability. The system outperformed single-agent baselines in predicting mortality for acute kidney injury patients, onset of AKI in chronic kidney disease patients, and immune-related adverse events in cancer patients receiving immune checkpoint inhibitors.
PhenoAgent is a large language model framework designed to extract clinical phenotypes from free-text electronic health records by decomposing queries into structured formats and using self-correction loops. The system demonstrated high accuracy in phenotype retrieval, achieving macro-F1 scores comparable to physician evaluations and outperforming traditional baseline methods.
ATHENA-R1 is a treatment reasoning agent trained by reinforcement learning over 212 biomedical tools that at each step finds missing information, runs tools and folds in the evidence. Trained without human traces using self-generated trajectories, it beat language models across five benchmarks covering 3,168 drug reasoning tasks and 456 patient cases.
Virtual Speech Therapist classifies stuttering from speech samples and then uses LLM agents that generate, critique and refine individualized therapy plans before clinician review. Expert speech therapists judged the resulting plans to be consistent and clinically appropriate.
Five agents handle data analysis, objective weight adjustment, plan optimization, plan evaluation and reporting, with fine-tuning on expert weight adjustments, retrieval of similar cases and reinforcement learning from expert preferences. Dosimetric metrics of its plans did not differ significantly from expert clinical plans while planning was much faster.
The Agent-Based Conversational Dialogue method simulates multi-turn therapy sessions between an LLM clinician and an LLM-simulated aphasic patient. Comparing Claude, GPT and Gemini under different prompting and reasoning settings, it found distinct readability profiles, with few-shot prompting and advanced reasoning producing more accessible language.
MedGuards is a multi-agent system designed to detect and correct errors in medical text, enhancing patient safety by ensuring the accuracy of clinical notes. The system demonstrates significant improvements over existing methods across various metrics and datasets, highlighting its potential for safer deployment in real-world healthcare settings.
Combines triage routing, a database-backed patient memory and constraint-based retrieval to answer post-discharge follow-up questions grounded in prescriptions while checking cross-department drug conflicts. On 485 real follow-up cases, blinded clinical experts rated it above physician baselines for authority and safety.
The paper describes an agentic auto-discovery framework that uses AI to identify diastolic dysfunction phenotypes from surface ECGs, demonstrating its effectiveness in classifying LVDD severity with high accuracy. The system combines transfer learning and autonomous architecture optimization, achieving robust external validation across large datasets.
DeepBD is a grounded agentic workflow designed to prioritize genetic variants and assist in the diagnosis of birth defects. It integrates various types of evidence, including rule-based knowledge and biological context, to improve variant prioritization accuracy compared to existing methods.
An LLM agent queried a clinical database of notes, encounters, procedures and labs to complete the structured 30-day readmission review rubric used by physicians. Blinded physicians rated its reviews of 20 readmissions against those of two physicians each, and it was then applied to 100 more cases.
EHR-Complex is a benchmark designed to evaluate the performance of clinical agents in handling complex electronic health record (EHR) data through interactive database reasoning. The benchmark highlights significant challenges and limitations in current models' ability to accurately perform longitudinal multi-table aggregation and compositional reasoning, indicating a need for more robust clinical decision support systems.
This controlled study tested whether knowledge graph grounding, including an agentic natural-language-to-Cypher loop over PrimeKG, improves clinical question answering. Grounding did not help on MedQA across model sizes but did raise accuracy on counterfactual and novel-fact benchmarks, locating its benefit outside training knowledge.
VISTA Architect is a database-oriented AI system that integrates large language models with electronic health records to create a persistent, provenance-linked knowledge graph. It demonstrated high accuracy in reconstructing patient histories for multidisciplinary tumor boards and reduced preparation time without sacrificing precision.
A data agent converts structured clinical records to text and routes them to five fine-tuned domain expert agents, whose outputs a coordinator agent aggregates into staging and phenotype calls. On 1,066 patients from two neurology services it outperformed single multimodal LLMs and earlier medical multi-agent systems.
A LangChain workflow chains perception, retrieval, clinical reasoning and an audit agent that verifies the reasoning before predicting PHQ-8 depression severity from interview transcripts. On DAIC-WOZ it lowered mean absolute error from 5.35 to 5.02 relative to a single-agent baseline.
MedRLM is a framework that uses specialized agents to recursively inspect and synthesize patient information for clinical decision support, incorporating multimodal data like EHRs, medical images, and sensor signals. It demonstrates the potential of recursive reasoning in handling long-context clinical scenarios by integrating evidence from various sources into a Clinical Evidence Graph Memory.
MIRA is an autonomous agent in a sandboxed EHR that gathers histories, orders and interprets tests, generates differential diagnoses and writes treatment plans. In simulations on real patient cases it outperformed physicians in diagnostic accuracy.
AMIE was extended with an agentic system for multivisit disease management that grounds its reasoning in clinical guidelines and drug formularies using long-context retrieval. In a blinded virtual OSCE with 100 scenarios, it was non-inferior to 21 primary care physicians on management reasoning.
MedAgent is an evidence-based medicine agent trained with supervised fine-tuning, reward modeling and GRPO to run a workflow of question parsing, PICO extraction, evidence retrieval and grounded answering. It is designed to cite traceable evidence and respect evidence hierarchies in clinical answers.
TxBench-PP is a benchmark designed to evaluate AI agents' ability to make accurate decisions in small-molecule preclinical pharmacology by analyzing real-world assay data. The study found that no system could reliably recover preclinical pharmacology decisions, with the best performing configuration achieving only 59.3% accuracy.
EmoPatient is an emotion-directed patient simulator designed to train clinicians in palliative care communication by generating realistic emotional responses. The system demonstrates improved emotional realism and robustness across different conversational scenarios, enhancing the training experience for healthcare professionals.
PhysAssistBench is a benchmark designed to evaluate the interactive assistance capabilities of large language models in clinical settings, focusing on their ability to coordinate knowledge, communication, and system interaction. The evaluation shows that current LLMs are unreliable for such tasks due to their inability to reliably integrate these diverse skills.
RubricsTree is an evaluation framework for personal health agents that uses a hierarchical taxonomy of clinically-verifiable rubrics to assess the performance of AI systems in healthcare. It demonstrates superior alignment with expert evaluations and can significantly improve model performance when used as structured instructions or training rewards.
WEQA is a query-adaptive agent framework that uses an LLM controller to synthesize execution plans and route queries to specialized wearable analytical tools, demonstrating significant improvements in accuracy over existing methods through expert review and user studies.
The Agentic AI-based Framework uses a multi-agent system to mitigate premature diagnostic handoff and silent hallucinations in healthcare applications, ensuring thorough patient data collection before making diagnoses. The framework demonstrates improved diagnostic precision through deterministic orchestration constraints and epistemic uncertainty quantification mechanisms.
The paper specifies six ReAct tool-use agents with Reflexion self-correction and confidence-gated human review for monitoring patients across presurgical, intraoperative and postsurgical cardiopulmonary care. It replaces a rule-based multi-agent digital twin with LLM-mediated coordination and formally defines agent tool registries and escalation logic.
The paper evaluates the use of large language models in pharmacogenomics, finding that direct reasoning by these models is unreliable and unsafe. It proposes a trustworthy agentic genomics architecture where validated decision logic is encoded as versioned skills executed as code, ensuring deterministic and auditable clinical mappings across diverse populations.
AutoMedic is a multi-agent simulation framework designed to evaluate large language models as clinical conversational agents by transforming static QA datasets into virtual patient profiles for realistic dialogue simulations. The system demonstrates its effectiveness in providing a comprehensive evaluation of LLMs' performance in clinical settings, validated by human experts.
The Autonomous AI Agent integrates medical knowledge graphs with dynamic clinical data to enhance decision-making in emergency departments. It demonstrates superior performance across various tasks, including triage and drug interaction detection, compared to existing methods.
Five agents sharing one LLM backbone draft, check and audit answers to clinical questions against current regulatory data on withdrawn drugs. On 103 adversarial questions the agentic setup cut the rate of recommending banned drugs by about 53 percent across model families.
ClinAgent is an AI-assisted system designed to enhance clinical trial data processing and statistical programming, addressing the bottleneck in pharmaceutical research. It demonstrates its capability by successfully validating nine skills on a Phase 2 cardiovascular study dataset, though it highlights the need for expert review in complex domains due to varying accuracy levels.
AIMAH is an AI-driven virtual hospital platform for medical education that simulates patient interactions and provides structured assessment feedback. The system was validated through a study with medical trainees, demonstrating high realism, educational usefulness, and learner engagement.
Hospital agents keep patient records and retrieved cases local and send compressed latent KV blocks to a host agent that makes the rare-disease diagnosis. On a self-built cross-hospital benchmark it improved diagnosis while leaking less reconstructable clinical text than raw latent exchange.
ArogyaSutra is a multi-agent framework designed for multimodal medical reasoning in Indic languages, addressing the limitations of existing English-centric models. The system improves accuracy in multilingual medical reasoning and decision support by integrating tool grounding with dual-memory mechanisms.
GovernAgent is an LLM-driven framework designed to address data quality issues in clinical records by employing hierarchical governance mechanisms and constrained action planning. It improves the accuracy, efficiency, and reliability of clinical data governance while minimizing hallucinations.
The PATHFinder Agent is a conversational AI system designed to gather patient information and create personalized prenatal care plans aligned with ACOG's PATH guidelines. The study shows that GPT-5.2 performs well in generating tailored healthcare plans, though it still has gaps in antenatal testing recommendations.
The paper introduces an agent-based pipeline that converts raw, noisy audio recordings of doctor-patient conversations into structured transcriptions suitable for training large language models. The system demonstrated high accuracy in noise removal and dialogue structuring tasks compared to manual processing and other methods, leading to improved performance when used to fine-tune a medical LLM.
XMedFusion is a modular AI framework that uses multiple agents to interpret medical images and generate radiology reports, improving the reliability of automated diagnostic systems. The system demonstrates significant improvements in various evaluation metrics over existing vision-language models, showcasing its effectiveness in enhancing robustness and transparency in intelligent medical imaging.
Uses doctor, patient and supervisor agents to synthesize consultation dialogues under four diagnostic reasoning styles, then trains paradigm-specific LoRA branches merged into a shared model for an interactive doctor agent. On a new dynamic Chinese consultation benchmark it outperformed proprietary and medical baselines on diagnosis and process metrics.
The paper introduces gSage, a workflow-orchestrated agent designed for precision oncology clinical decision support. It demonstrates that restricting the agent to a curated set of tools and knowledge sources enhances its reliability and safety compared to broader access approaches.
The Multimodal Anonymizer is a locally deployable framework designed to deidentify identifiers across various formats in clinical data. It demonstrated high sensitivity and critical clinical preservation in automated and manual evaluations, supporting its feasibility for patient-level data anonymization.
The paper evaluates an agentic large language model system for assessing preoperative frailty in cancer patients, showing it outperforms a single LLM system in both binary classification and score generation tasks. The findings suggest that an agentic workflow can enhance the accuracy of automated frailty assessments from EHR data.
The paper evaluates a large language model agent for chart review in lung cancer guideline concordance, showing that the search-enabled agent achieves higher accuracy and reduces errors compared to sequential note review. The study suggests that tool-augmented LLM agents can effectively support scalable information extraction from longitudinal clinical documentation.
SymptomAI is a conversational AI agent designed for patient symptom assessment and differential diagnosis, deployed in a study involving 13,917 participants. The system demonstrated significant accuracy in diagnosing illnesses like influenza compared to clinician-provided diagnoses.
PSEBench is a benchmark designed to evaluate large language models in the context of patient safety event triage, focusing on their ability to interpret and apply jurisdiction-specific policies. The system demonstrates the capability trends of various LLMs and highlights areas for improvement in using AI for clinical decision support.
MedSP1000 is an interactive benchmark for evaluating clinical agents using standardized patient cases, which assesses their performance in dynamic decision-making scenarios. The system reveals that current large language models, even those specialized for medical use, are not yet reliable enough to be integrated into actual clinical practice.
Assigns department-specific doctor agents to each patient, retrieves complementary EHR evidence for them and updates only unresolved points across deliberation rounds before fusing the consensus with structured features. It improved mortality prediction while reducing consultation cost.
Builds a 4,421-case interactive diagnosis benchmark and a multi-agent consultation system that gathers demographics, questions the patient and refines candidate diagnoses over turns. It gained 10.3 percent accuracy over the baseline, closing about half the gap to a full-information oracle.
ClinicalMC is a benchmark designed to evaluate large language models' performance in multi-course clinical decision-making scenarios, covering patient care from admission to discharge. The system includes a multi-agent evaluation framework and assesses various LLMs across different settings, demonstrating the complexity and evolving nature of medical decision-making processes.
The paper describes a multi-agent system that uses large language models to support researchers throughout the clinical trial life cycle, from preparation to translation. The system demonstrates how an agentic AI approach can manage and streamline complex clinical research workflows with human oversight.
The paper presents a genetic neuro-symbolic LLM system that integrates multiple AI agents for managing cholangitis, demonstrating superior accuracy compared to conventional AI models and human experts in a multi-center study. The results suggest potential for advanced AI architectures in clinical decision support for complex medical conditions.
DynamiCare is a dynamic multi-agent framework designed to simulate interactive and iterative clinical decision-making processes, using MIMIC-Patient dataset for patient-level simulations. The system demonstrates the feasibility of modeling complex medical scenarios through iterative queries and adaptive strategies by specialist agents.
The MAP framework uses specialized clinical agents to support decision-making for patient admission, diagnosis, and treatment planning in inpatient settings. Experiments showed that MAP outperforms existing large language models in aligning with operational protocols.
ClinNoteAgents is an LLM-based multi-agent system designed to predict heart failure 30-day readmission risk from clinical notes, transforming free-text into structured representations and clinician-style abstractions. The system demonstrates high extraction fidelity for clinical variables and key risk factor identification while preserving predictive signal with significant text reduction.
ClinEnv is an interactive benchmark that evaluates large language models as attending physicians in real inpatient admissions, focusing on their decision-making process and information gathering. The system reveals significant gaps between the quality of discharge diagnoses and management actions made by these models.
DrugClaw is a multi-agent system designed to answer drug information questions by querying a registry of drug and pharmacovigilance skills, ensuring answers are grounded in primary regulatory or peer-reviewed sources. It outperforms other systems on the DrugAudit benchmark, demonstrating high accuracy and faithfulness in its responses.
The study evaluates the ability of clinical agents to detect and handle patient identity inconsistencies in electronic health records. It finds that these agents often fail to identify subtle changes in patient identity, highlighting a critical risk for misbinding during routine clinical workflows.
SURGENT is a surgical multi-agent assistant with a tree-of-thought planner, department-specific agents, guideline and literature retrieval, and separate long-term and working memory over patient histories. It was evaluated on five perioperative tasks including case analysis, plan simulation, safety monitoring, risk assessment and rehabilitation.
The paper introduces the Causal Sensitivity Score (CSS) to evaluate how clinical AI systems adjust their recommendations based on changes in patient inputs, revealing differences not captured by traditional coverage-based metrics. It highlights a universal safety blind spot where all models fail to update recommendations when surgery status changes.
HetMedAgent is a multi-agent framework that integrates generalist large language models and domain-specific specialist models to enhance clinical decision-making. The system demonstrates improved performance in real-world tasks by leveraging the strengths of both types of models, highlighting the continued importance of specialist models for precise analysis.
Treats agent improvement as edits to a bounded library of procedural skills, admitting a candidate only if it nets improvement on a held-out probe under a hard regression budget. On FHIR-based MedAgentBench it lifted gpt-oss-120b from 40.6 to 88.8 percent and beat the best of five self-improvement baselines by 21 points.
MediHive is a decentralized multi-agent system designed for medical reasoning tasks, utilizing large language models to collaboratively analyze and resolve complex medical questions. The system demonstrates superior performance compared to single-LLM and centralized MAS approaches on standard medical QA datasets.
The paper presents a multi-agent AI framework for de-identifying clinical narratives from low- and middle-income countries, addressing the lack of tools that can handle diverse linguistic characteristics. The system uses a hybrid approach combining LLMs with rule-based modules to detect and substitute personally identifiable information, achieving high F1-scores across various document types.
Runs repeated stochastic rollouts of an LLM reasoning over 550 MIMIC-IV sepsis trajectories across six timesteps and measures divergence, entropy and temporal consistency. Individual outputs looked plausible, but intervention agreement across runs was near zero and instability grew over time.
Agents combine patient context with external drug knowledge and a safety verification step to recommend medication sets at the fourth ATC level. On MIMIC-III and MIMIC-IV it improved fine-grained accuracy while limiting interactions, contraindications and set size.
LLM agents decompose fetal ultrasound queries into plane recognition, segmentation and measurement subtasks run by visual tools, and an arbitration step reconciles LLM reasoning with tool evidence stored in a retrieval bank. On a new 3,205-question fetal ultrasound VQA benchmark it beat the strongest MLLM baseline by more than 25 percent.
The system uses a multi-agent language model pipeline to detect and classify delusion-related content in naturalistic audio diaries, demonstrating the effectiveness of detailed diagnostic prompts and majority voting for accurate classification. The study highlights the importance of careful prompt design and evaluation methods in automated mental health diagnosis.
MDIA is a multi-agent system designed for clinical reasoning tasks, implemented as a 7-node specialty-routed graph that significantly outperforms ChatGPT on the HealthBench Professional benchmark. The study highlights how architectural design and context preservation can enhance performance in clinical decision support systems.
PathNavigate is a training-free pathology agent designed to address whole-slide image visual question answering (WSI-VQA) by navigating gigapixel slides efficiently. It uses a surprise-guided scan and shared slide memory to locate sparse, high-resolution evidence, improving answer accuracy and interpretability.
The Ventilator Decision Support System (VDSS) is a human-in-the-loop multi-agent system designed to assist clinicians in making ventilator decisions by adapting to individual clinician preferences using contextual bandit learning. Retrospective evaluations with expert review indicate that VDSS improves the acceptability of recommendations and reduces interaction rounds compared to traditional methods.
Gives LLMs 44 tools covering gene and variant utilities, pathogenicity and phenotype databases, expression atlases, orthologs and literature, with the model inferring which to call and in what order. On 100 expert-curated questions a 20B model reached a 94 percent pass rate with the tools versus 41 percent without.
Contrasts a single LLM with a four-step pipeline separating data interpretation, guideline checking and explanation when narrating XGBoost ICU mortality predictions on the eICU demo set. The pipeline avoided outcome leakage and grounded explanations in guidelines better, while the standalone LLM aligned more closely with SHAP attributions.
VLM diagnostic agents record pairwise differential-diagnosis notes from their own failures and reflective re-diagnosis, then retrieve these notes when reading new radiology cases. Across 11 radiology subspecialties it raised accuracy by up to 7 points over several models.
ClinSeekAgent is an automated framework that enables clinical agents to actively seek and integrate multimodal evidence from patient records for better decision support. It demonstrates significant improvements in performance over models using curated inputs, showing the effectiveness of its dynamic evidence acquisition approach.
The DeepLens Diagnosis Agent is a structured pipeline that enhances diagnostic accuracy by guiding a small medical reasoning model through five stages of clinical extraction, retrieval, generation, triangulation, and decision-making. It demonstrates significant performance gains over the same model without the pipeline on the DiagnosisArena benchmark, highlighting the importance of workflow design in complex tasks like medical diagnosis.
Splits medical question answering among an interpreter agent, an explorer agent that keeps retrieving until evidence is sufficient, and an arbiter agent that adjudicates the answer. Across five benchmarks and five backbones it beat the strongest baseline by 6.46 accuracy points on average.
The paper explores the use of large language models to simulate individualized behavioral responses to pandemic interventions, demonstrating their ability to generate mobility changes in different venues without fitting to observed data. The study highlights the potential and limitations of LLM-based agents for modeling population compliance with non-pharmaceutical interventions during a pandemic.
EchoAgent is a framework that uses large language models to control specialized vision tools for echocardiography, enabling structured and interpretable automation. It outperforms existing medical visual-language models in video-level reasoning tasks by accurately measuring anatomical structures and providing clinically grounded interpretations.
SepsisAgent is a large language model augmented with a clinical world model to simulate patient responses and recommend sepsis treatment decisions. The system demonstrates superior performance in value-based decision-making while maintaining safety compared to traditional reinforcement learning and LLM-based methods.
The paper introduces a reinforcement learning (RL) post-training approach for an agent to improve multi-turn reasoning and answer clinically meaningful questions over FHIR data. The RL-trained agent significantly enhances the accuracy of responses while adhering to data integrity constraints, demonstrating superior performance compared to prompt-based methods.
ProtoMedAgent is a framework that uses iterative, zero-gradient optimization to generate clinically interpretable reports from multimodal data while maintaining patient privacy. It outperforms standard retrieval-augmented generation methods in terms of faithfulness and reduces the risk of membership inference attacks.
RealICU is a benchmark designed to evaluate large language models' ability to understand and reason about long-context ICU data, using hindsight annotations from senior physicians. The study reveals that existing LLMs struggle with clinical decision-making in realistic ICU conditions, highlighting issues like recall-safety tradeoffs and anchoring bias.
Wraps each stage of a retrieval-based screening pipeline as a policy-governed agent that is tuned against proxy metrics and locked or rolled back, coordinated by an orchestrator agent. A transcript-based depression detection proof of concept converged to stable configurations while avoiding regressions.
Evaluates LLM agents on clinical risk prediction from EHR time series, images, radiology reports and notes, in unimodal and multimodal settings and in single versus multi-agent configurations. Single agents outperformed naive multi-agent systems, handled multimodal inputs better and were better calibrated.
A deliberative council of Gemini agents, with class-specific advocates scoring evidence under clinical rules, labels defense mechanisms in emotional-support dialogues, and a builder-critic-guard agent team selects overrides from fine-tuned models. The system placed second of 64 teams in the shared task, though the council was often wrong on minority classes.
DeepTumorVQA is a hierarchical benchmark for evaluating medical vision-language models and AI agents on tumor diagnosis tasks using 3D CT scans. It decomposes the diagnostic process into stages, identifies measurement as a critical bottleneck, and demonstrates that tool augmentation can significantly improve performance.
CodeClinic is a benchmark designed to evaluate the ability of large language model agents to automate clinical reasoning tasks, such as ICU monitoring and patient state tracking. It demonstrates that these agents can synthesize reusable clinical skills more effectively than zero-shot code generation methods, improving consistency while reducing computational costs.
The AI co-clinician is a conversational medical AI system that uses audio-visual data to assist in real-time clinical decision-making during telemedicine consultations. It demonstrated performance comparable to primary care physicians and significantly better than text-only approaches, highlighting the importance of multimodal input for effective medical consultation.
The study presents a human-in-the-loop framework using multi-agent large language models to identify cutaneous immune-related adverse events from clinical notes, showing improved accuracy and inter-rater agreement compared to manual review alone.
The study evaluates an agentic system combining GPT-5.0 and Gemini 2.0 Flash in critical care clinical decision-making, showing superior performance compared to standard LLMs and human experts in acid-base disorder interpretation and sepsis management using text-based vignettes.
WoundScribeAI's Scribe AI uses a multi-agent system to transcribe and generate educational content for patients, enhancing communication in clinical settings. The study demonstrated the model's ability to produce accurate SOAP notes, visit summaries, and engaging patient education materials without hallucinations.
The paper presents a multi-agent framework using large language models to automatically extract demographic information from clinical vignettes in medical education materials. The system demonstrates high accuracy and reliability, with improvements through autonomous error correction.
MedExAgent is a trained large language model agent designed to perform clinical diagnosis by interacting with patients and ordering medical exams, formalized as a Partially Observable Markov Decision Process (POMDP). The system demonstrates comparable diagnostic accuracy to larger models while maintaining cost-efficient examination strategies through extensive experiments.
Ingenix BioResearcher is a scenario-guided multi-agent system designed for translational medicine, which integrates over 30 specialized tools and machine-learning endpoints to perform evidence synthesis from heterogeneous biomedical sources. The system demonstrates superior performance in unit-level capabilities, open-ended biomedical reasoning, and end-to-end clinical discovery benchmarks compared to existing baselines.
TheraAgent is a self-improving therapeutic agent that iteratively generates, judges, and refines treatment plans to ensure they are precise and comprehensive. It outperforms existing methods on the HealthBench dataset and achieves an 86% win rate against physicians in expert evaluations.
CuraView is a multi-agent framework designed to detect and explain faithfulness hallucinations in discharge summaries generated by large language models, using a GraphRAG-based knowledge graph for evidence retrieval. The system demonstrates improved factual reliability of clinical documentation with an F1 score of 0.831 on the safety-critical E4 metric.
ADAPTS is a framework that uses a mixture-of-agents LLM architecture to automate the rating of depression and anxiety severity from clinical interviews. It decomposes long-form interviews into symptom-specific tasks, providing auditable justifications while maintaining temporal alignment. The system demonstrated improved accuracy compared to human ratings on high-discrepancy interviews.
One hundred physician-reviewed tasks drawn from real consultations run inside an EHR environment accessed through commercial-style APIs, each needing about 27 tool calls and graded against 670 execution-verified checkpoints. The best of 13 agents passed only 46 percent of tasks and open models at most 19 percent.
MediCARE trains a graph neural network on patient graphs for medication recommendation, explains it with XAI methods, and has collaborating LLM agents discuss and revise reports of the predictions. On MIMIC-III it produced accurate reports and improved prediction precision.
ALEX is a multi-agent framework that uses machine learning to identify important subgroup treatment effects from clinical trials and then employs large language models to generate natural-language explanations for these effects. The system was evaluated by specialist physicians who found the generated explanations clinically relevant, indicating its potential in enhancing precision medicine.
Seven specialized LLM agents with DSM-5/ICD-11 retrieval extract symptoms, match criteria, consider differentials, check timing, reason counterfactually, vote and arbitrate on schizophrenia interview transcripts. On 71 diagnosed cases it reached 95.8% positive agreement, above single-agent baselines with and without retrieval.
MedAgentBrief is an AI system that generates hospital course summaries using a large language model, which was evaluated in a clinical setting for safety and utility. The study found that the system was frequently used by physicians with minimal risk of harm and associated with reduced physician burnout.
AERO is an AI agent designed to refine and optimize clinical trial eligibility criteria for better real-world applicability. It demonstrated the importance of systematic refinement in generating accurate real-world evidence by emulating a specific clinical trial with Mayo Clinic data.
The system uses a federated multi-agent architecture to harmonize public health datasets into OMOP and FHIR standards, enabling interoperable data integration. It demonstrated high mapping accuracy in integrating MIMIC-IV, ELSA, and synthetic datasets, showing the feasibility of scalable clinical decision-making support.
The Virtual Speech Therapist (VST) is an AI-driven system that supports clinicians in assessing stuttering and creating personalized therapy plans. It integrates deep learning for stutter detection with large language models to generate, critique, and refine therapy plans iteratively, ensuring clinical interpretability and feasibility.
The paper presents a lightweight agent architecture that integrates a memoryless clinical risk encoder over time to produce continuous concern trajectories, enabling early detection of rising risks in clinical settings. The study demonstrates that second-order dynamics can reveal gradual increases in concern before critical thresholds are reached, enhancing the transparency and usability of LLM agents for human monitoring.
The paper presents a Dual-Stream Memory Architecture for health coaching agents that separates patient narratives from structured clinical records to detect discrepancies, ensuring safe and accurate longitudinal healthcare management. The system was evaluated on real-world data, demonstrating its capability to identify 84.4% of designed clinical discrepancies with high safety-critical recall.
SCOPE is a multi-LLM planner-based framework designed to reason over clinical trial data, breaking down tasks into row selection, structured planning, and execution phases. It demonstrates improved accuracy for reasoning-based questions compared to direct prompting or heavier agentic baselines.
AgentClinic is a multimodal benchmark designed to evaluate large language models in simulated clinical environments, focusing on sequential decision-making and tool usage. The system demonstrates that existing LLMs struggle with complex clinical tasks when evaluated under realistic conditions, highlighting the need for further development in their ability to use tools effectively.
A neurosymbolic multi-agent system with a LangGraph orchestrator and eight specialized agents answered 30 guideline-based COPD and asthma vignettes. Its accuracy was compared against 30 board-certified pulmonologists, allergists and emergency physicians.
The paper introduces a method for evaluating clinical AI systems using case-specific rubrics authored by clinicians, validated through automated scoring by large language models (LLMs). It demonstrates that LLM-generated rubrics can approximate clinician agreement in ranking the quality of AI outputs, enabling more efficient and cost-effective evaluation compared to traditional expert review methods.
An agentic system navigates tens of thousands of clinical documents and lab values from 811 myeloma patients to answer treatment-history questions, compared with single-pass RAG, iterative RAG and full-context input. It reached 79.6 percent concordance with oncologist labels, above both baselines, though more of its errors were clinically significant than expert disagreements.
Pathology, molecular and clinical agents produce schema-constrained evidence reports, and a governance agent applies executable guideline rules from a knowledge graph, with hard overrides such as POLE status. On TCGA-UCEC it reached 0.943 accuracy with under 1 percent logic violations, and 0.842 on CPTAC versus under 0.31 for neural baselines.
DxChain is a clinical reasoning framework that uses large language models to assist in diagnosing diseases by emulating the cognitive process of clinicians. It demonstrates improved diagnostic accuracy and logical consistency compared to existing methods on real-world benchmarks.
The paper describes the development and evaluation of a DeepSeek AI agent for dental communication training, which uses contextually adaptive dialogue generation to enhance doctor-patient interaction skills. The study found that students trained with the AI agent showed significantly higher scores in simulated patient consultations compared to those using traditional peer-to-peer role-play practice.
DxDirector-7B is an autonomous large language model designed to guide the entire clinical diagnostic process, from initial ambiguous patient complaints to final diagnosis. It demonstrates superior diagnostic accuracy and efficiency compared to existing models while significantly reducing physician involvement in routine tasks.
The paper introduces a multi-agent system for personalized physiotherapy that uses generative AI and computer vision to create patient-specific exercise videos and provide real-time pose correction. The system aims to improve at-home physiotherapy compliance by offering dynamic, tailored feedback.
The system uses a two-tier approach for glaucoma screening, combining a semi-supervised classifier with a multi-agent deliberation system. It demonstrates improved diagnostic accuracy on ambiguous cases by routing uncertain predictions to agents for further analysis.
DeepSeek V3 is a large language model used for clinical tasks, particularly in generating synthetic patient cases and ranking differential diagnoses. The study shows that when embedded within an agentic workflow, DeepSeek V3 can reduce explicit bias in diagnostic tasks compared to its standalone version.
Agents translate clinical guideline recommendations into symbolic logic that a SAT solver checks for redundancies and conflicts arising in patients with multiple conditions. On 12 SGLT2 inhibitor guidelines it found most conflicts were comorbidity-specific and detected them with an F1 of 0.861, where plain LLMs failed.
PKGPT is a recursive large language model agent designed to automate the development of NONMEM-based population pharmacokinetic models. It showed promise in producing executable and converging models, but still requires human oversight for physiological plausibility and clinical interpretability.
StoryMI is a multi-agent framework that generates therapeutic dialogues using large language models, with situational context and dynamic strategy control. The system demonstrates improved adherence to motivational interviewing standards and clinical plausibility through structured dialogue generation.
RadAgent is a tool-using AI agent designed to generate stepwise and interpretable chest CT reports, providing clinicians with an inspectable trace of the reasoning process. The system demonstrates improved clinical accuracy, robustness under adversarial conditions, and introduces a new metric called faithfulness compared to its 3D VLM counterpart.
The paper presents an agentic AI system that automates the generation of pharmacogenomic recommendations using large language models and structured evidence from biomedical literature. The system was evaluated by experts who found it to outperform other LLMs in clinical clarity and guideline concordance.
Adds a memory of past case episodes, reflection-updated diagnostic rules and per-tool reliability tracking to a tool-orchestrating chest X-ray agent, letting it learn across cases at test time without training. On ChestAgentBench it raised accuracy from 0.68 to 0.79 with GPT-5-mini and from 0.76 to 0.87 with Gemini-3 Flash.
The paper explores fairness in a multi-agent system within a hospital triage scenario, where two agents negotiate patient allocation based on ethical frameworks and demographic biases. It demonstrates that fairness emerges through the interaction of these agents, leading to more equitable outcomes than those achieved by individual agents alone.
Agents plan the diagnostic workup, consult a knowledge base and interpret multimodal eye images through structured debate across four modalities and 11 datasets. It averaged 85 percent retinal diagnosis accuracy, above leading multimodal LLMs, and improved accuracy by up to 20 percent for 311 ophthalmologists in a global reader study.
The Clinical Agentic Research Intelligence System (CARIS) automates clinical research workflows, including planning and documentation, using large language models and modular tools. It was evaluated on three datasets with high completeness scores in both automated and human evaluations.
CARE-ECG is a causally structured framework that uses large language models to interpret ECGs and provide clinical decision support. It improves diagnostic accuracy and explanation faithfulness by encoding multi-lead ECGs into latent biomarkers, performing causal graph inference, and supporting counterfactual analysis.
TrajOnco is a multi-agent framework that uses temporal reasoning over longitudinal EHR data to predict cancer risk and generate patient-level summaries. It demonstrated comparable performance to supervised machine learning methods in predicting lung cancer, with the added benefit of better interpretability and scalability.
The paper evaluates a neurosymbolic multi-agent system designed for clinical decision support in diagnosing and managing ileus and volvulus. The system, comprising specialized agents for perception, synthesis, and verification, outperforms general-purpose LLMs in diagnostic accuracy and reduces hallucinations and critical safety errors.
G-Health turns routine health examination reports into preventive actions using aligned LLMs, 20 quantitative risk models and guideline retrieval in a multi-agent framework. Its models ranked best on 13 benchmarks, and 79 blinded medical assessors preferred its reports to those of other LLMs and senior physicians.
Builds 300 cases with an interactive patient simulator that withholds concerns and tracks whether a dialogue agent elicits and addresses them over multiple turns. No model dominated across metrics, and human clinicians remained stronger on several aspects of the interaction.
ASDAgent is a framework that includes DoctorAgent and ChildAgent to provide high-fidelity dialogue synthesis and clinical decision support for Autism Spectrum Disorder intervention. It demonstrates strategic consistency with human experts in real autism interventions and enhances the therapeutic capabilities of small language models through synthetic data generation.
SEA is a self-learning diagnostic agent designed to improve clinical decision support by integrating a dual-memory module and reinforcement learning for reasoning and memory management. The system demonstrates superior performance in standard evaluation metrics and long-horizon tasks, with expert evaluations confirming its clinical correctness and practical usefulness.
GraphDx is a multi-agent framework designed for sequential diagnosis that integrates medical knowledge into a structured graph to enhance diagnostic accuracy while minimizing costs. The system demonstrates significant improvements in diagnostic success rates and cost reduction compared to existing large language model approaches.
CCD-CBT is a multi-agent framework designed to simulate Cognitive Behavioral Therapy (CBT) more accurately by dynamically reconstructing the Cognitive Conceptualization Diagram and implementing information-asymmetric interactions. The system demonstrates improved counseling fidelity and positive-affect enhancement compared to existing methods, as validated by clinical scales and expert therapists.
LLM extraction and reasoning agents work with an oncology knowledge graph, a recommendation engine and human review to screen patients for trials over 157,367 clinical pages. In a prospective evaluation of 3,804 patients it achieved an F1 of 0.82 against dual-oncologist labels, above GPT-4 prompting baselines.
A GPT-5.2 agent calls 23 tools over Model Context Protocol servers to extract DICOM metadata, preprocess images, run Monte Carlo simulation, segment organs and write dosimetry reports. On 597 PSMA-PET/CT studies it ran without failures and agreed closely with OpenDose3D organ doses.
A controller agent routes patient questions through safety-constrained generation and separate clinical-safety and hallucination risk assessments, refining answers until both risk scores fall below a threshold. It outperformed baseline models on PatientSafeBench, MedSafetyBench and MedHallu.
OGA-AID is a clinician-in-the-loop system that uses multiple specialized agents to draft structured reports for post-stroke rehabilitation based on multimodal gait analysis data. The study demonstrates the effectiveness of OGA-AID in reducing errors compared to baseline methods and highlights its potential as an AI-assisted tool complementing human clinical judgment.
The paper introduces a Latent Diagnostic Trajectory Learning (LDTL) framework for sequential clinical diagnosis, using two LLM agents to model the acquisition of diagnostic evidence under uncertainty. The system demonstrates improved diagnostic accuracy and efficiency compared to existing methods on the MIMIC-CDM dataset.
Med-ICE has several peer LLM agents generate and review answers over repeated rounds until a semantic consensus monitor detects agreement, without an external arbiter. It outperformed single-model generation and self-refinement on medical question-answering benchmarks.
Simulates 100 users with multi-year device streams, exam records and life events with known per-event effects, and asks 100 queries each about lookup, trend, comparison, anomaly and explanation. Database-native agents scored 48 to 58 percent against 30 to 38 percent for memory-augmented retrieval, with the gap on multi-hop attribution questions.
DILIConsult is a multi-agent large language model framework designed to evaluate drug-induced liver injury (DILI) in ICU settings by intelligently parsing clinical and drug information. The system demonstrated improved performance in extracting DILI characteristics and analyzing suspected cases compared to conventional approaches, with evaluations conducted by clinician panels.
A proprietary LLM generates structured categories and transitions without seeing patient data, and a local LLM uses them to gather evidence and decide, on ICU cases where symptoms contradict signs. On the derived MIMIC-DOS dataset it outperformed single-pass LLMs and other agentic workflows while keeping data local.
VIBEMed is a self-evolving multi-agent framework designed for clinical decision support, integrating specialized agents for diagnosis and treatment planning. It demonstrates superior performance in complex clinical cases through its ability to learn from patient outcomes over time.
This study ran 4,800 trials of multi-agent LLM clinical pipelines across four model families and found they generate clinical assertions that no single agent produced. Between 30% and 56% of network assertions were emergent and judges rated most of them as clinically dangerous.
PDT screens patients for possible Castleman disease and then uses knowledge-enriched GPT-4o diagnostic and treatment agents to produce diagnoses and treatment plans. It was tested on 10 Castleman and 10 non-Castleman patients with results reviewed by clinicians.
CAMP is a multi-agent system designed for clinical prediction that dynamically assembles a specialist panel based on the complexity of each case. It demonstrates improved performance and decision transparency compared to existing single-agent and fixed-role multi-agent approaches.
Specialist agents edit clinical findings counterfactually, measure how diagnosis confidence shifts and use these signals to challenge hypotheses in multi-round discussion. Across three diagnostic benchmarks and seven LLMs it improved accuracy over prompting and earlier multi-agent methods, and humans rated its reasoning more interpretable.
The system uses a multi-agent architecture to provide personalized clinical support for Parkinson's disease, employing generation, critique, and synthesis agents. It achieved high accuracy in diagnosing and recommending treatments based on 50 representative clinical queries.
PubMed Reasoner is a biomedical question-answering system that iteratively refines PubMed queries and retrieves evidence to generate accurate, evidence-backed responses. It outperforms human experts on the PubMedQA dataset and demonstrates strong performance in clinical relevance and trustworthiness evaluations.
The Virtual MDS Panel (VMP) is a multi-agent AI system designed to provide precision care recommendations for myelodysplastic syndromes. It demonstrated superior accuracy and reduced major errors compared to general-purpose LLMs, achieving expert-level ratings in diagnosis, prognosis, and treatment.
An orchestrator runs a Monte Carlo tree search over diagnostic hypotheses, verifying evidence and backtracking, supported by a mutable patient-state memory and a retrieved store of guidelines and past cases. It outperformed single-agent and multi-agent baselines on diagnostic accuracy and explainability.
SkinGPT-X is a collaborative multi-agent system designed for dermatological diagnosis, featuring a self-evolving memory mechanism to enhance adaptability and transparency. The paper demonstrates SkinGPT-X's superior performance in diagnosing both common and rare skin diseases compared to existing large language models.
The system uses a multiagent framework with large language models to simulate team-based diagnosis for Alzheimer's disease, improving interpretability and generalizability. It achieved high accuracy in classifying different stages of cognitive impairment without fine-tuning on specific patient data.
AD-CARE is a large language model-based agent designed for real-world Alzheimer's disease diagnosis, capable of generating comprehensive diagnostic reports from incomplete and heterogeneous data. It demonstrated superior accuracy across multiple cohorts and improved decision-making efficiency in a reader study involving neurologists and radiologists.
WiseMind is a multi-agent framework designed for psychiatric diagnosis that integrates logical reasoning and empathetic communication to provide accurate and trustworthy mental health assessments. The system was evaluated using simulated patient interactions and real user sessions, demonstrating high diagnostic accuracy and clinical soundness.
The paper introduces a multi-agent framework for medical multiple-choice question answering that uses specialist agents to generate independent diagnoses and Two-Phase Verification to improve uncertainty calibration. The system demonstrates significant improvements in calibration metrics on filtered subsets of MedQA-USMLE and MedMCQA, suggesting enhanced reliability in clinical decision support.
MedClarify is an AI agent designed to assist in medical diagnosis by iteratively generating follow-up questions based on patient information, aiming to reduce diagnostic uncertainty. The system demonstrates a significant reduction in diagnostic errors compared to standard LLMs through its ability to engage in targeted questioning.
The paper evaluates the performance of clinical language models in scoring NIHSS under realistic, noisy conditions and finds that structured retrieval workflows significantly improve accuracy compared to non-agentic baselines. Lower-cost models benefit more from structured retrieval than larger models, suggesting a practical approach for deploying robust clinical decision support systems.
The AI-simulated patient tool, powered by GPT-4o and grounded in medical case reports and databases, was developed to enhance clinical training for family medicine. The study showed that the tool is highly rated by participants for its usability, educational value, and ability to provide realistic and evidence-based feedback.
The paper introduces RWE-bench, a benchmark for evaluating large language model (LLM) agents in generating real-world evidence from medical databases. It assesses the ability of LLMs to execute observational studies and generate coherent evidence bundles, finding that current models struggle with end-to-end task completion despite some success in individual steps.
The paper introduces a hybrid multi-agent framework for comprehensive epilepsy management, integrating diverse patient data and clinical guidelines to support decision-making. The system demonstrates improved reliability and comprehensiveness in managing epilepsy compared to existing task-specific models.
GSEM is a clinical decision-making agent that uses a graph-based memory system to store and retrieve past experiences, enhancing its ability to make accurate decisions. The system demonstrated superior performance on the MedR-Bench and MedAgentsBench benchmarks compared to other methods, showcasing its effectiveness in clinical reasoning tasks.
Cerebra is an interactive multi-agent AI system designed to assist in dementia diagnosis and risk assessment by integrating electronic health records, clinical notes, and medical imaging data. It demonstrated superior performance compared to single-modality models and language model baselines, with significant improvements in accuracy when used by experienced physicians.
The system uses a multi-agent architecture to automate BT-RADS scoring for brain tumor follow-up assessments, integrating clinical data extraction and decision logic application. It demonstrated higher accuracy compared to initial clinical assessments, particularly in detecting critical categories like BT-4.
Compares a deployed multi-agent LLM intake system's diagnoses and dispositions with those of blinded treating clinicians in 2,379 real primary-care telemedicine visits. Top-1 diagnosis matched in 91.3 percent of cases and 96.3 percent above a safety confidence threshold, with a 2.5 percent disposition error rate.
DispatchMAS uses AutoGen caller and dispatcher agents grounded in a curated taxonomy of chief complaints and caller identities to simulate emergency medical dispatch calls. Expert clinical assessment and automated metrics were used to evaluate the realism and protocol adherence of the simulations.
Agentic Cognitive Profiling (ACP) is an AI framework designed to realign automated Alzheimer's Disease screening with clinical protocol logic, focusing on construct validity. It decomposes standardized assessments into cognitive tasks and uses specialized language model agents for semantic understanding and deterministic function calling for quantification, achieving high accuracy in AD prediction while providing interpretable cognitive profiles.
CoNVict is an agentic AI system designed to prioritize copy number variations (CNVs) for rare disease diagnosis by integrating patient phenotype data and performing structured reasoning. The study demonstrates that CoNVict outperforms existing computational methods in identifying causal CNVs, particularly in cases involving uncertain or non-coding variants.
The study evaluates an agentic AI system with a two-agent architecture for auditing fairness in machine learning models used to detect early-onset colorectal cancer, focusing on demographic disparities. The results indicate that agents augmented with retrieval capabilities achieve higher semantic similarity to expert-derived statements compared to other configurations, suggesting potential for scaling fairness audits in clinical AI.
TheraAgent is a multi-agent framework designed for PET theranostics in precision oncology, addressing challenges of data scarcity and heterogeneous information integration. It demonstrates improved accuracy in predicting treatment outcomes compared to existing models, highlighting its potential for reliable clinical decision support.
The paper describes a multi-AI agent framework for neurosurgical education and evaluation, converting static vignettes into interactive conversations. The system includes specialized AI agents to simulate patient interactions and provide objective clinical data, demonstrating that traditional multiple-choice testing may overestimate LLM capabilities in realistic clinical scenarios.
The paper introduces an agentic operating system for hospitals that adapts large language model agents to manage clinical workflows safely and transparently. It evaluates a memory component using manifest-guided retrieval, demonstrating higher precision and efficiency in navigating longitudinal patient records compared to a metadata-filtered RAG baseline.
Combines a domain-tuned LLM with literature retrieval to propose diagnoses for 82 endocrinology case reports, and tests serial and concurrent collaboration with physicians. It matched senior specialists and beat residents, held accuracy on rare diseases, and helped physicians fix errors while introducing automation bias.
Integrates 11 multi-agent architectures under a common communication protocol and evaluates them on 11 clinical benchmarks spanning 473 diseases, using a VLM-based reasoning evaluator. Multi-agent systems reasoned more deeply but were fragile when moving between medical subdomains.
FetalAgents is a multi-agent system designed for fetal ultrasound image and video analysis, dynamically orchestrating specialized vision experts to enhance accuracy in diagnosis, measurement, and segmentation. The system demonstrates superior performance across multiple clinical tasks compared to existing models, providing an auditable solution aligned with clinical workflows.
The paper presents a framework that uses large language models to harmonize clinical trial data across studies, making it interoperable and accessible for exploratory analysis. The system enables natural language querying of the harmonized data, reducing manual effort and improving query accuracy in secondary analysis workflows.
Meissa is a lightweight, multi-modal medical large language model designed for offline deployment in clinical settings. It demonstrates competitive performance with larger models on various medical benchmarks while offering significant advantages in cost and latency.
The paper presents an agentic LLM workflow for optimizing the placement of spectroscopy volume-of-interest in brain tumors, enhancing clinical decision-making by generating and selecting from multiple candidate placements. The system demonstrates improved coverage of solid tumors and avoidance of necrosis based on user preferences compared to general expert placements.
The paper introduces SELSM, a framework that enhances the logical reasoning skills of medical agents to handle clinical tasks in FHIR-based electronic health records without requiring additional training data. It demonstrates significant improvements in task completion rates and success when tested on MedAgentBench.
DERM-3R is a framework of three collaborative agents designed for dermatologic diagnosis and treatment using traditional Chinese medicine, addressing fine-grained lesion recognition, multi-view representation, and holistic clinical reasoning. The system demonstrates strong performance in dermatologic tasks with limited data and computational resources, matching or surpassing larger general-purpose models.
A medical imaging agent mines its verified execution trajectories for recurring tool sequences, registers them as composite tools, and reinforces their use with an image-feature memory and a GRPO-style training loop. It improved multi-step orchestration accuracy and cross-domain generalization over prior agentic methods across imaging datasets.
MedCoRAG is a framework that uses multi-agent collaboration and retrieval-augmented generation to diagnose hepatic diseases, providing structured reasoning and interpretability. It outperforms existing methods in diagnostic accuracy and interpretability on real-world clinical cases.
Compiles expert protocols into step-wise alignment ratings that are accumulated along a diagnostic agent's trajectory and calibrated into correctness probabilities, triggering extra checks when uncertain. On three diseases from MIMIC-IV it beat the best baseline by 12 percent AUROC and halved Brier score, and clinicians judged it useful.
The paper describes an agentic AI system designed to detect and grade immune-related adverse events from clinical notes, showing high accuracy in both retrospective and prospective evaluations. The system also demonstrated significant time savings and improved consistency when used by clinical trial staff for annotation tasks.
CARE is a medical reasoning framework that enhances clinical accountability by decomposing tasks into specialized modules and using explicit evidence, outperforming existing models on standard benchmarks. It demonstrates the benefits of an agentic approach in medical AI for more accurate and trustworthy decision support.
The AI Liver Transplant Selection Committee is a multiagent system using large language models to simulate the decision-making process of a liver transplant selection committee. The study demonstrated high accuracy, sensitivity, and specificity in predicting patient survival post-transplantation and identifying contraindications for transplantation.
ChatCLIDS is a benchmark where LLM nurse agents with evidence-based persuasion strategies hold multi-turn conversations with expert-validated virtual patients who resist closed-loop insulin delivery. Larger reflective models adapted over time, but all models struggled against resistance and social pressure.
The Contextual Invertible World Model (CIWM) integrates a machine learning emulator with a large language model to predict drug responses in colorectal cancer. It revealed that mutant KRAS has a hierarchical dominance over the APC/Wnt-axis in driving resistance to 5-fluorouracil and identified a paradox where repairing PIK3CA increases chemoresistance.
TARSE is a system designed to enhance clinical decision-making by retrieving and adapting skills and experiences at test time. It demonstrates consistent improvements over existing methods on medical question-answering benchmarks, highlighting the effectiveness of its approach in aligning model reasoning with clinically valid logic.
MedCollab is a multi-agent framework that uses an Issue-Based Information System (IBIS) protocol and Hierarchical Disease Relation Chains to assist in clinical diagnosis by integrating patient evidence and medical knowledge. The system demonstrates improved diagnostic accuracy, evidence consistency, and clinical reasoning quality compared to existing large language models and baselines.
OPGAgent is a multi-tool agentic system designed for interpreting dental panoramic X-rays, combining specialized perception modules and a consensus mechanism to enhance accuracy and versatility. The system outperforms existing vision language models and medical agent frameworks in both structured-report generation and visual question answering tasks.
CXReasonAgent is a diagnostic reasoning agent designed to interpret chest X-rays by integrating large language models with clinical tools, ensuring evidence-grounded responses. The system demonstrates superior reliability and verifiability compared to vision-language models through evaluations using CXReasonDial, a dialogue benchmark for multi-step diagnostic tasks.
Multiple LLM agents with different objectives negotiated Medicaid care plans under a Nash bargaining scheme, compared with compute-matched sequential self-critique. In a pre-registered paired study of 200 patients, Nash orchestration gave a small but significant composite gain in safety, efficiency and equity scores.
The paper evaluates LangGraph v1.0, a neuro-symbolic large language model system for managing acute cholecystitis, which outperforms human experts in diagnostic accuracy and treatment recommendations based on the Tokyo Guidelines 2018. The study demonstrates the system's superior performance across various clinical scenarios through automated benchmarking and expert review.
Separates clinical summarization into context compression, drafting, detection of weakly grounded spans via attention signals and targeted revision under supervisory control. On two public datasets it beat vanilla LLMs and other baselines under reference metrics, LLM judges and human evaluation.
Pairs AI therapy agents with simulated alcohol use disorder patients that have dynamic cognitive-affective states, scoring 369 sessions against a care-quality and risk ontology. It found failures such as validating delusions and not de-escalating suicide risk, and stakeholders found the audit dashboard useful.
Decomposes natural-language CT questions into sequences of measurement, radiomics and segmentation tool calls exposed through Model Context Protocol servers, trained on a new 3D CT tool-use benchmark. It improved diagnostic accuracy by 41 percent over baselines with a 95 percent tool invocation success rate.
HevaDx is an AI system designed to diagnose and recommend treatments for vascular anomalies, using a multi-agent architecture that separates visual perception from clinical reasoning. The system demonstrated high diagnostic accuracy (94.8%) and treatment recommendation accuracy (83.3%), showcasing its potential in providing reliable support for managing complex medical conditions.
GPAS is an AI system that integrates rapid pathogen identification with semantic interpretation using a large language model, aiming to reduce technical barriers in clinical diagnostics. It demonstrated accurate pathogen detection and provided clinically instructive interpretations of immune dysregulation effects on the respiratory microbiome.
The paper introduces Biomni, an autonomous LLM agent that performs end-to-end trial emulation from electronic health records to support clinical decision-making. It demonstrates Biomni's ability to calibrate EHR-based estimates with RCT results using a Bayesian hierarchical model, significantly reducing discrepancies and improving the reliability of evidence translation.
NutriOrion is a hierarchical multi-agent framework designed for personalized nutrition intervention in patients with multiple chronic conditions. It decomposes complex nutritional planning into specialized tasks handled by isolated agents, ensuring clinical validity and interoperability, and outperforms existing models in reducing drug-food interactions and improving dietary recommendations.
DeepRare is a multi-agent LLM system that uses over 40 tools and knowledge sources to rank rare disease diagnoses from phenotypes, free text and genetic results with linked evidence. Across nine datasets it reached 57.18% average Recall@1 on HPO tasks, beat Exomiser on multimodal cases, and experts agreed with most of its reasoning.
The study evaluates two AI agent systems, OpenManus and Manus, for clinical decision tasks using various benchmark datasets. The results show modest performance gains with high computational costs and significant hallucination rates despite advanced tool integration.
DAVP is a hierarchical agentic LLM pipeline that pre-filters genes, writes knowledge-graph-grounded variant reports and runs iterative LLM tournaments to rank variants for rare disease patients. On 709 cases it reached 85.5% top-3 recall, and on a new Turkish cohort it placed the causal variant in the top 3 for 76.2% of patients versus 47.6% for Exomiser.
The paper describes an AI agent designed to improve the detection of radiologist follow-up recommendations in electronic health records. The system uses a pretrained large language model to review clinical impressions, extract key details for follow-ups, and integrate these findings into patient outreach workflows, achieving high accuracy and significantly outperforming existing macro-based systems.
A modular pipeline of sequential specialized LLM agents extracts 229 structured fields from breast cancer synoptic pathology reports. Seven LLMs scored 93.8-99.0% on synthetic tests but recall fell to 61.8-87.7% on 90 real reports.
MARL-Rad is a multi-modal multi-agent reinforcement learning framework designed for generating radiology reports, which optimizes agents within the context of chest X-ray interpretation. The system demonstrates improved clinical efficacy and report accuracy compared to existing methods, as validated by both automated metrics and clinician evaluations.
The paper presents a system that compiles patient data into structured artifacts to reduce hallucination in clinical AI systems. It shows that retrieval-augmented generation increases hallucination rates, while structured representation with explicit provenance tracking significantly reduces unsupported claims and improves factual accuracy.
Resp-Agent is an autonomous system designed to generate multimodal respiratory sounds and diagnose diseases by addressing inherent information loss and data scarcity issues. It demonstrates improved diagnostic robustness under challenging conditions through extensive experiments.
Three fine-tuned LLM families are combined in a pipeline where a reasoning agent explains, a retrieval agent grounds answers in PubMed and a refinement agent improves them, with uncertainty and bias checks. DeepSeek R1 scored best on MedQuAD-derived data and clearly beat BioGPT zero-shot.
The Mixed-Vendor Multi-Agent Conversation (MAC) framework uses agents from different vendors to improve clinical diagnosis, showing better recall and accuracy compared to single-vendor systems. The study reveals that mixed-vendor teams can pool complementary biases to identify diagnoses missed by individual models or homogeneous teams.
Vital Trace is a multi-agent framework designed for clinical reasoning in intensive care units, using a compact patient-state memory and protocol-constrained agents to predict future medical interventions. The system demonstrates improved temporal consistency and predictive performance compared to free-form agent systems.
LATCH is an LLM-assisted framework that converts natural language clinical hypotheses into structured EHR data analyses, automating cohort construction and statistical analysis. It demonstrated the ability to reproduce findings from published studies efficiently while also enabling new insights through hypothesis modifications.
EC2Seq2Sql is a two-stage framework that converts clinical trial eligibility criteria written in natural language into executable SQL queries for patient screening from electronic health records. The system demonstrated high accuracy in generating valid and coherent SQL statements, achieving an exact match accuracy of 0.84 and execution accuracy of 0.91 after normalization.
A video reasoning model interleaves reasoning with coarse-to-fine tool calls that retrieve and verify temporally localized evidence in long surgical and clinical videos, trained with a grounding-aware GRPO variant on a new video suite. It reached state-of-the-art results on in-domain and out-of-domain clinical video benchmarks.
The paper introduces Differential Reasoning Learning (DRL), a framework that enhances clinical decision support agents by learning from reasoning discrepancies. DRL uses reference rationales and agent-generated chain-of-thought to improve the reliability of clinical decision-making, as demonstrated through evaluations on medical QA benchmarks and internal clinical data.
LingxiDiagBench is a multi-agent framework that evaluates large language models in Chinese psychiatric consultation and diagnosis, using a dataset of synthetic dialogues. It highlights the limitations of current LLMs in handling complex mental health diagnoses and dynamic consultations.
FHIR-AgentEval is a modular evaluation sandbox designed to assess the performance of large language model agents in handling clinical tasks using HL7 FHIR standards. The system demonstrates that memory augmentation significantly improves task success rates and reduces strategic errors, suggesting potential for more reliable clinical applications.
Fine-tuned specialist agents each see only their partition of clinical data, and their outputs are combined by a deterministic embedding projection that estimates each agent's marginal contribution. It performed well on tumor board datasets from North American and East Asian hospitals and on public oncology benchmarks.
Describes how signals from over 115 million live patient calls and clinician testing are used to train and evaluate an orchestrated patient-facing voice agent system with independent safety checks. Deployed over 10 million calls, it reported a 99.9 percent clinical safety score and halved speech recognition errors.
SynthAgent is a multi-agent system that simulates high-fidelity patient scenarios for complex diseases like obesity with mental health comorbidities. The study shows that GPT-5 and Claude 4.5 Sonnet outperform other models in generating realistic patient simulations, highlighting the potential of SynthAgent for clinical decision support and knowledge discovery.
The system integrates document retrieval, reranking, and evidence grounding to support multi-step medical reasoning, outperforming standalone language models on several medical question-answering benchmarks. It demonstrates the effectiveness of combining tool-augmented and evidence-grounded strategies for reliable and scalable AI in clinical decision support.
MedAgentBrief is an AI system that generates hospital course summaries using a large language model, aiming to reduce clinician documentation burden. The study found that the system was frequently used by physicians with minimal risk of harm and significantly reduced physician burnout.
MedSAM-Agent is a framework that uses multi-turn agentic reinforcement learning to enhance interactive medical image segmentation, enabling more efficient and accurate decision-making. The system demonstrates state-of-the-art performance across various medical imaging modalities by integrating expert-curated trajectory generation and end-to-end outcome verification.
RE-MCDF is a multi-expert framework designed to enhance clinical diagnosis by integrating logical constraints and medical knowledge. It demonstrates superior performance in complex diagnostic scenarios compared to existing single-agent systems, highlighting the benefits of structured multi-agent reasoning.
Has 110 dialysis nurses, ChatGPT-4 and the agent-based MAI-DxO system answer four validated hemodialysis scenarios, scored by nephrology nursing experts. The agent system scored highest, especially on structured justification and differential diagnosis, while nurses contributed contextual judgement.
LLMs propose candidate binary rules for unit-weighted clinical checklists while a deterministic, data-based loop verifies and selects them under deployability constraints. Across eight prediction tasks it beat existing score-generation methods and on two externally validated tasks exceeded established guideline scores.
An LLM planner writes FHIR-based discharge plans with confidence estimates, a rule-based auditor checks coverage and calibration, and failed cases are regenerated or buffered for replay. On MIMIC-IV-on-FHIR the self-improvement loop raised task coverage from 32 to 86 percent and improved calibration.
Trains medical reasoning verifiers that query external medical corpora repeatedly while judging a reasoning trace, using iterative reinforcement learning with an adaptive curriculum. Across four benchmarks it raised MedQA accuracy by 23.5 percent relative to the base generator with an eightfold smaller sampling budget.
An implantable epilepsy platform feeds EEG-derived biomarkers to an LLM agent that converses with patients through a phone app, logs their reports and retrains its detectors. In 13 inpatients, seizure detection precision improved over days without expert programming and patients rated the system usable.
The paper describes an agentic framework that uses large language models to iteratively prompt patients for a comprehensive clinical history, generating EHR-ready summaries and differential diagnoses. The system was evaluated using case reports and simulated patient interactions, showing high accuracy in capturing relevant history elements and aligning with physician-recommended investigations.
Q4Dx is a benchmark of synthetic patient cases at different levels of symptom disclosure in which LLM physician agents question simulated patient agents to reach a diagnosis. It measures how efficiently models choose diagnostic questions rather than only final accuracy.
SAGE is an LLM-based planning agent designed for automated stereotactic radiosurgery treatment planning, demonstrating comparable dosimetry to human planners while reducing cochlear dose. It shows systematic reasoning behaviors such as constraint verification and trade-off deliberation when prompted.
AgentsEval is a multi-agent framework that evaluates the clinical correctness and reasoning fidelity of automatically generated medical imaging reports by emulating radiologists' collaborative diagnostic workflow. It demonstrates robust evaluation capabilities under various perturbations, ensuring clinically aligned and semantically faithful assessments.
SycoEval-EM is a simulation framework that evaluates the robustness of large language models in clinical decision support scenarios, particularly under adversarial patient persuasion. The system revealed significant variability among LLMs in adhering to evidence-based guidelines and highlighted the need for multi-turn adversarial testing in AI evaluation.
GastroTCM is a specialized large language model assistant for TCM gastroenterology that addresses limitations of general LLMs by incorporating domain-specific knowledge and multi-turn dialogue optimization. It demonstrated superior performance in clinical accuracy and safety compared to baseline models, as validated through expert review.
SpAgents is a multi-agent system designed to assist primary care physicians in diagnosing axial spondyloarthritis early, integrating large language models and imaging analysis. The study shows that SpAgents significantly improves diagnostic accuracy and sensitivity compared to both junior rheumatologists and primary care providers.
The system automatically generates evaluation rubrics for medical large language models by decomposing retrieved medical evidence into atomic facts and synthesizing them with user interaction constraints. It significantly outperforms baseline methods in evaluating the clinical intent alignment of responses on HealthBench and LLMEval-Med benchmarks, demonstrating its potential to improve patient safety.
AgentEHR is a benchmark and framework designed to enable autonomous navigation of Electronic Health Records (EHRs) for complex clinical decision-making tasks. It introduces RetroSum, which enhances the performance of agents by preventing information loss through retrospective summarization and maintaining logical coherence in high-noise environments.
Treats prompt refinement as a partially observable decision process in which planner, teacher, critic, student and target agents iteratively improve instructions for recognizing depression, anxiety and trauma states in text. It improved diagnostic accuracy and robustness across domain-specific and stratified benchmarks.
OncoPainBot uses pain extraction, mechanism reasoning, treatment planning and safety check agents to assess and manage cancer pain. Comparing seven LLMs and three retrieval strategies on 516 real records, Claude-4 with retrieval performed best, and clinicians validated its outputs.
Two generator models answer clinical queries while two evaluator agents score them against AMA ethics principles and a five-tier safety protocol, looping until responses pass. Over 900 queries the loop cut ethical violations by 89 percent and downgraded risk in 92 percent of cases.
PAIR-SAFE is a paired-agent system designed to audit and refine AI-generated mental health support responses, ensuring they are clinically aligned. The system demonstrates significant improvements in key clinical dimensions of mental health interactions when guided by a supervisory Judge agent.
MedConsultBench is a comprehensive benchmark for evaluating medical consultation agents that covers the entire clinical workflow from history taking to follow-up. It introduces fine-grained metrics and Atomic Information Units to assess information gathering efficiency, diagnostic accuracy, and medication safety, revealing significant gaps between theoretical knowledge and practical application in large language models.
ClinAgent is a five-layer architecture that augments any AI coding agent with clinical trial statistical programming capabilities, enabling them to handle SAS datasets and ADaM specifications. The system was validated using artifacts from a real Phase 2 cardiovascular study, demonstrating high accuracy in log analysis and data validation.
ART is a benchmark designed to evaluate medical AI agents' performance on complex, multi-step reasoning tasks derived from real electronic health records. It highlights significant gaps in current models' abilities to handle aggregation and conditional logic, indicating the need for improved clinical decision support systems.
KT-LLM is a multi-agent system designed to bridge the gap between structured clinical data and text-based rules in kidney transplantation, ensuring decisions are evidence-grounded and auditable. The system demonstrated superior performance in predictive calibration and evidence attribution compared to baselines, while also enabling equitable subgroup analysis among Black recipients.
ECR-Agent is designed to mitigate the Einstellung Effect in medical LLMs by ensuring evidence-based reasoning through structured causal inference and iterative knowledge refinement. The system demonstrates improved diagnostic accuracy in atypical cases compared to existing models, as evaluated using MedEinst benchmark.
The paper presents an autonomous agentic workflow using large language models to detect cognitive concerns from clinical notes, demonstrating its effectiveness in early detection of cognitive impairment. The system showed comparable performance to expert-driven methods and maintained interpretability while offering scalable decision support.
A tumor board-style multi-agent system uses a coordinator to choose among 34 ASCO guidelines, specialist agents to extract text, table and figure content, and a reviewer to compose answers. It selected the right guideline 94% of the time and answered 90% of 100 questions correctly, well above single LLMs.
EvoMDT is a self-evolving multi-agent system designed to support structured clinical decision-making in multidisciplinary tumor boards for various cancers. It demonstrated improved guideline adherence and semantic alignment with expert plans, outperforming large language model baselines while maintaining interpretability and traceability.
RadDiff is a multimodal agentic system designed to describe differences between radiology image sets, aiding in clinical insights and AI interpretation. It uses an innovative framework that includes medical knowledge injection, multimodal reasoning, iterative hypothesis refinement, and targeted visual search, achieving significant accuracy on the RadDiffBench benchmark.
A multi-agent system synthesizes 5,200 patient cases, a fact-limited patient agent with hallucination correction plays them in dialogue, and expert-refined guideline-based rubrics score the doctor model's questioning. Current LLMs showed substantial gaps in information gathering and diagnosis.
The paper introduces a hierarchical agent reflection mechanism to train large language models for medical diagnosis, ensuring the model's reasoning aligns with human clinical practice. The method improves diagnostic transparency and interpretability by generating detailed clinical notes that capture clinicians' step-by-step reasoning processes.
Four Claude 3.5 Sonnet agents played ACLS team members following AHA guidelines while an emergency physician led simulated resuscitations for team leadership training. Three blinded emergency specialists rated ten scenario transcripts for completion of recommended steps.
Feature reasoning agents write and execute SQL over tables of interpretable cellular features, and a knowledge comparison agent checks the resulting measurements against pathology knowledge. On two pathology visual question answering datasets this produced executable evidence traces linking measurements to diagnoses.
MedAgentBench v2 evaluates medical LLM agents on clinical tasks using a FHIR-compliant EHR, introducing improvements such as memory components and enhanced system prompts. The agent demonstrated high success rates with these enhancements, showing adaptability to new tasks without explicit training data.
A smart home platform for post-stroke rehabilitation fuses insole, eye-tracking and ambient sensor data, and an embedded LLM agent issues reminders, adjusts the environment and notifies caregivers. In 20 users, mean satisfaction rose from 3.9 in conventional homes to 8.4 with the full system.
The study evaluates different harness designs for biomedical claim verification agents, finding that models formulating their own PubMed queries significantly improve accuracy compared to naive retrieval methods. The research highlights the importance of effective retrieval over reasoning in enhancing model performance and underscores the need for better safeguards against irrelevant citations.
The paper presents a method for guiding skill loading in AI agents using workflow topology to improve the quality of clinical trial data pipeline artifacts. The study shows that topology-guided context improves artifact quality compared to prompt-only approaches, especially when tasks require evidence not easily retrievable through lexical means.
The Pathways system uses an auditable pipeline to discover and maintain community healthcare directories by expanding queries across multiple dimensions and verifying extracted information. It demonstrates high accuracy in interpreting requests and restoring missing data, but emphasizes the need for further validation and clinical evaluation.
NeuroPilot is an agentic framework that uses a multimodal large language model and computational imaging modules to analyze longitudinal neural recordings, improving signal recovery and tracking in challenging conditions. It demonstrated the ability to decode behavioral states from neural population activity across different experimental settings.
The paper describes an agentic LLM pipeline that curates the accessible human surfaceome from literature, outperforming existing databases in accuracy and breadth. The system uses a manually curated benchmark to assess its performance and provides detailed annotations for each protein record.
CADS is a framework designed to measure the incremental contribution of large language models in guiding sequential biological screens, providing an error-controlled evaluation that separates the model's value from underlying algorithmic capabilities. The study reveals that while LLMs can contribute positively to discovery rates, their impact diminishes over time and may be overstated without proper control mechanisms.
BioMINTREL is an AI agent designed to discover and apply methods from various disciplines to biological data, enabling the exploration of new research questions. It outperforms existing models in a comprehensive benchmark by improving agreement with human experts and expanding coverage of potential applications.
ExpertTwins builds multi-agent scientific panels in which each seat retrieves evidence under an explicit policy and deliberation is checked with quotation verification and peer source review. In a nine-seat neuroblastoma panel, 438 of 439 claims passed quotation checks and seats defended or withdrew positions when challenged.
Brieflow-auto is an automated pipeline for optical pooled screening data analysis, utilizing a large language model to interactively tune parameters and validate configurations. It demonstrated improved accuracy in gene assignment compared to manual analyses and enhanced the reliability of the underlying codebase.
AutoOnboarder is a system that autonomously adapts large language model agents to perform specific tasks, such as submitting genomic sequencing prior authorization forms. It demonstrates the ability of these adapted agents to match or exceed baseline performance at significantly lower costs but highlights the variability in final agent performance.
ReproBAIT evaluates the ability of coding agents to design, implement, and train biological AI tools from limited descriptions. The study finds that some tools can be substantially reproduced by these agents, but complex ones are harder to recreate, suggesting caution in publicly releasing tool information.
Heterogeneous Agentic Coevolution (HACE) is a framework that evolves specialized foundation models and large language model agents through reciprocal learning, improving their performance in various domains including protein function prediction. HACE demonstrates superior performance compared to existing self-improving methods across multiple challenging tasks.
UMCR is a pipeline designed to extract and link medical concepts from various text types to any target ontology, addressing the challenge of adapting to different ontologies without retraining. The system demonstrates strong performance on the MedPath benchmark, showcasing its ability to enhance interoperability in healthcare by aligning diverse medical standards.
Claude Code instances in an autonomous harness surveyed reverse transcriptase loci across 1.9 billion protein clusters. They surfaced array-associated RTs, a new jumbo-phage RT family with repeat arrays that are expressed during Staphylococcus phage infection.
The GNQA system uses a combination of retrieval-augmented generation and querying a biomedical knowledge graph to answer expert-curated questions, demonstrating significant improvements in performance when using more advanced language models compared to local models.
LangPoli is a system designed to curate experimental data from scientific literature by using a policy agent that generates natural language action guidelines and actor agents that convert these into executable tools. The system demonstrated strong performance in paper classification and structured information retrieval, outperforming plain LLMs and other existing systems.
DRYLAB-Bench is a benchmark designed to evaluate the biosecurity risks posed by large language models (LLMs) in designing hazardous biological mutations. It assesses LLMs' capabilities in orchestrating specialized biological design tools to iteratively refine mutation designs, scoring them against empirical wet-lab measurements for risk quantification.
PaleoRigor is a bounded research agent that uses a language model to plan data analysis for paleomicrobiome studies, ensuring reproducibility and adherence to predefined constraints. The system demonstrated high accuracy in processing ancient microbial evidence while maintaining the integrity of the data through rigorous validation steps.
ARGUS is an evidence-constrained agentic framework that uses large language models and biological computation to interpret single nucleotide variants in noncoding regulatory regions. It demonstrates the ability to navigate complex genomic data, making decisions based on real experimental evidence to determine the functional impact of genetic variations.
CircuitATLAS is a knowledge graph and agentic framework for discovering therapeutic targets in neurological diseases by reasoning through neural circuits to molecular interventions. It successfully identified ATP1A3 as a control point on cortical excitability, which was validated through experimental studies.
DrugTargetWorld is a synthetic biobank framework that generates simulated worlds to train and evaluate AI agents for drug target discovery. The system evaluates nine agents across multiple episodes, revealing limitations in distinguishing causal from non-causal proteins despite high composite scores.
Enrich.AI is an AI platform that integrates statistical gene set enrichment analysis with automated biological interpretation, using a conversational interface to guide users through the process. The system was validated in a study involving 50 gene sets, demonstrating its ability to provide accurate and relevant biological interpretations.
AgentDiscover is an autonomous discovery framework that uses a large language model to plan and execute scientific searches, storing all attempts in a database for iterative improvement. The system demonstrates superior cost-efficiency compared to existing frameworks across various domains including biology, outperforming human competitors in past contests.
The paper evaluates AI agents' ability to perform scientific analyses by assessing their reliance on provided data versus prior knowledge. It introduces a method to test evidence grounding, revealing that many agents can answer correctly without using the given dataset.
TEMPLAR is an AI system that uses multi-agents and neuro-symbolic architecture to induce and evolve standardized radiology reporting templates from large clinical datasets. It outperforms existing methods in terms of coverage, information fidelity, and diagnostic accuracy.
Prompt-as-a-Protocol (PaaP) converts natural-language research questions into executable analytical protocols, which are then executed by an agentic Generative Research Automation layer. This system autonomously identified significant biological findings in single-cell transcriptomics and biosensor analytics, demonstrating its capability to produce novel peer-reviewed results across different biomedical domains.
ddkg.skill is an agent skill with a controller, references, validated query patterns and a routing table that helps LLMs write correct Cypher queries for the Data Distillery Knowledge Graph. In a nine-test evaluation against a live instance, five of seven tests on unseen sources returned correct results, and failures were documented.
The paper evaluates multi-agent systems in scientific hypothesis generation, focusing on the impact of self-critique rounds and equivalence rules used to measure diversity. It shows that increasing critique rounds significantly diversifies hypotheses initially but has diminishing returns thereafter, and different equivalence rules yield vastly different assessments of mechanism similarity.
GAMA is a multi-agent framework designed to enhance biomedical named entity recognition by inducing and verifying annotation rules from labeled data, generating schema-constrained entities, and refining predictions. The system demonstrates superior performance compared to existing LLM-based methods across multiple datasets.
MR-MAS is a multi-agent LLM system that runs Mendelian randomization end to end, from data acquisition and causal estimation to interpretation and manuscript drafting, backed by an MR knowledge base for code repair. In test scenarios it executed workflows stably, reproduced literature-supported causal signals and gave consistent estimates across repeated runs.
The paper describes a workflow that uses an AI agent to generate hypotheses for combination therapies and constructs quantitative systems pharmacology models to evaluate them. The system identified one testable interaction from seven generated hypotheses, highlighting the importance of strict evaluation criteria in filtering out unsupported hypotheses.
Collaborative Theme Identification Agents (CoTI) is a multiagent system designed to automate thematic analysis in qualitative research, generating supporting excerpts and themes for human review. The study showed that CoTI's outputs were more aligned with those of senior investigators compared to junior researchers or baseline models, indicating its potential as an efficient tool for scalable qualitative research.
CGEx is a system that translates natural-language questions into Cypher queries to retrieve and explain information from multiple heterogeneous biomedical knowledge graphs. The evaluation showed significant improvements in query accuracy with user feedback, demonstrating the utility of CGEx for uncovering meaningful insights across diverse data sources without requiring users to learn Cypher.
OSWorld-Science is a benchmark designed to evaluate visual language models (VLMs) on their ability to perform scientific tasks using specialized software interfaces. The paper demonstrates that while current VLMs are proficient in certain areas, they still struggle with complex scientific workflows and require further development.
PhenoAIR is a reliability-aware multi-agent framework designed to predict mechanisms of action from Cell Painting profiles by reasoning over uncertain evidence. It outperforms existing methods in various MOA prediction settings, demonstrating its capability to handle noisy and misleading data.
gemot is an agentic framework that proposes explicit mechanistic models of biological systems and scores them against data through an MCP evaluation layer that keeps hypotheses auditable. On 18 published problems in molecular biology, epidemiology and immune differentiation its models matched or exceeded reference models in fit and parsimony.
BIABench is a benchmark designed to evaluate AI agents' ability to perform real-world bioimage analysis tasks end-to-end, including handling various imaging modalities and dimensions. The evaluation reveals that while agents can handle routine 2D tasks well, they struggle with more complex analyses involving 3D volumes or time-lapse sequences.
The paper evaluates large language models' decision-making processes in biochemical domains using Bayesian Optimization, revealing that while LLMs show a sincere intent to explore new data, they often overreact and fail to effectively balance exploration with exploitation. The study highlights the need for decoupling priors from incoming data to improve LLM-driven scientific discovery.
Introduces a stage-wise admissibility test that refuses comparisons between literature-review agents given unequal evidence or budgets, plus a benchmark of 2,042 biomedical articles as claim graphs. It refused 11 of 21 comparisons that conventional statistics would have decided, and a stage-reward-trained agent earned a certified top rank interval.
BioDyad is a system that integrates biomedical discovery and machine learning engineering through two hierarchies within Monte Carlo graph search, enabling the iterative construction, execution, and validation of candidate programs. It demonstrated superior performance on the BioXArena benchmark compared to other agent methods and one-shot baselines.
The Genome Skeptic framework evaluates the performance of AI models in orchestrating biological analyses and making judgments based on those analyses. The study found that while GPT-5.6 Sol excelled at efficiently directing follow-up analyses, it faltered when tasked with interpreting evidence to make accurate gene annotation decisions.
LabFactory is a framework that enables the creation of task-specific AI labs for various scientific tasks, integrating models, knowledge resources, and tools. The system demonstrated its effectiveness by successfully executing 10 constructed labs across different domains, surpassing reference values in all subtests under independent evaluation.
FiberPro 1.0 is a multiagent AI framework that guides the high-throughput production and conformal deposition of functional protein micro/nanofibers using focused rotary jet spinning, demonstrating its ability to optimize formulations within two iterations on average. It successfully designed an antibacterial zein-based active packaging system for food matrices, showcasing real-world applicability in extending shelf life.
Llemy is an LLM-based system designed to assist users in navigating and understanding complex molecular interaction maps. The paper details a user-driven development process involving domain experts, highlighting the system's ability to summarize pathways and its potential for future enhancements based on user feedback.
Six agents for planning, hybrid retrieval, evidence distillation, data interrogation, review and synthesis link literature with omics data, and a module distills past trajectories into reusable rules. It scored 84 of 100 on a plant stress benchmark, well above baseline LLMs, and a prioritized rice gene, OsPP2a, showed salt-response phenotypes when disrupted.
The paper evaluates Biomni and two other agentic coding systems on constructing a bioinformatics pipeline for CUT&Tag data analysis. While the generated pipelines appeared plausible, they did not fully satisfy the requested analysis requirements, highlighting the need for expert review in complex workflow construction.
scACORN is a system that uses specialized small language models and an orchestrating agent to interpret single-cell transcriptomic data, addressing the challenge of heterogeneous tasks in single-cell analysis. The system demonstrates improved performance metrics such as macro-F1 scores and recall rates, while reducing unsupported gene citations.
WetRobo is a robot kit designed for biological laboratories that enables coding agents to perform tasks without the need for local teleoperation training data or neural network retraining. The system demonstrated successful task execution across different laboratory environments, showcasing its adaptability and practicality in real-world settings.
HypoEvolve is a system that uses genetic algorithms to coordinate specialized large language model agents for scientific hypothesis discovery. It demonstrates superior performance in drug repurposing tasks across 34 cancer types, outperforming baseline methods by generating more selective and evidence-supported hypotheses.
BRACED is a system designed to improve multi-hop reasoning in biomedical question answering by using reinforcement learning to optimize the decomposition of complex queries into simpler sub-queries. It demonstrates significant improvements over existing methods on the BioHopR benchmark, highlighting its effectiveness in retrieving accurate and verifiable answers.
The paper describes the use of a large language model agent to iteratively develop an RNA 3D structure prediction model, QuickFold. The model was developed through a feedback-driven process where the agent proposed and implemented changes over multiple iterations, eventually matching leading open-source baselines in performance.
Karenina is an open-source framework that evaluates biomedical AI systems by incorporating expert knowledge across multiple dimensions, going beyond simple scoring to ensure trustworthy decision-making. The system illustrates its utility in evaluating question-answer pairs, multi-turn conversations, and autonomous data analysis tasks.
This BioASQ submission pairs modular BM25 and dense retrieval with an agent quorum in which several differently prompted LLM agents debate and converge on a consensus biomedical answer. It also uses an LLM-as-a-judge step and adaptive document retention for answer generation.
VERGE is an agentic workflow designed to extract early-onset colorectal cancer symptoms and family history risk status from clinical notes. It uses a verification-refinement cycle to enhance accuracy, reducing false positives while maintaining detection of true cases.
ELISA is a hybrid AI agent designed for single-cell genomics that integrates expression embeddings with semantic retrieval and large language model interpretation to facilitate interactive discovery. It outperforms existing methods in cell type retrieval tasks and can generate biologically plausible hypotheses, bridging the gap between transcriptomic data exploration and mechanistic biological understanding.
An agentic system following a human annotation manual decomposes systematic review tasks, self-debugs and validates its outputs over tens of thousands of trial abstracts, linking structured results to guideline treatment pathways. Review accuracy was 95.1 to 97.2 percent against human annotations and the platform gave more comprehensive evidence than general AI search tools.
The paper evaluates the impact of reference composition biases on automated cell annotation in xenotransplantation single-cell atlases using Biomni, a biomedical agent. It reveals that biased references can lead to widespread mislabeling and obscure biological signals, highlighting the need for auditing reference compositions when interpreting AI-assisted annotations.
EvoSCM keeps a population of structural causal model hypotheses, uses them to pick discriminating experiments, and distils contradictions into correction rules that revise the models. Evaluated on physics, chemistry and biology tasks, it explained and predicted better than baseline agents while using experimental budget more efficiently.
The Agentic Retrobiosynthesis Framework uses large language models to select which molecules to expand next in a deterministic biochemical search, demonstrating that policy-based frontier selection can enhance the efficiency of retrosynthetic pathway discovery. The framework's performance was evaluated using various benchmarks, showing improvements over traditional methods like MCTS and direct prompting.
NMRAgent is a large language model-based agent designed to interpret complex NMR spectra for molecular structure elucidation, combining database retrieval and de novo generation with evidential reasoning. It demonstrated superior accuracy and transparency compared to existing methods on benchmarks involving novel scaffolds and real-world applications.
This work extends the Gemini-based Co-Scientist from in silico hypothesis generation to closed-loop work with real instruments across materials science, biology and computer science. In biology it predicted emergent behaviors and ran lab-in-the-loop cycles, alongside reactor-driven materials syntheses.
Agents propose hypotheses, edit and debug code in an ESMFold codebase, train and evaluate variants, and store successful and failed interventions in memory under a tree-search resource policy. Exploring about 80 variants with roughly 5,000 GPU-hours it improved best lDDT by 7.5 percent over independent proposals and random search.
BixBench3 is a benchmark designed to evaluate AI agents' ability to perform computational biology tasks at the scale of complete research studies, from raw data processing to generating scientific results. The study reveals that leading large language models (LLMs) vary significantly in their capacity to handle sequential analysis steps and manage large datasets across different biological domains.
The paper evaluates the performance of autonomous agents in protein characterization tasks, finding that while language models like Opus provide high accuracy, cheaper deterministic policies can achieve near-frontier accuracy with better reproducibility. Federation across institutional boundaries imposes a negligible penalty on performance.
FrontierChallenge is a benchmark for evaluating scientific workflow completion across various domains, assessing models' ability to produce complete research artifacts. The study reveals that high partial scores and confident claims of task completion do not reliably indicate full delivery, underscoring the need for comprehensive evaluation of end-to-end workflows.
BioCheck Agent fact-checks biomedical claims by issuing Boolean PubMed searches, appraising the retrieved literature and writing a structured report rather than a bare label. Reinforcement learning is used to improve its search and reasoning for public health claim verification.
Biomni is a general biomedical agent whose environment was built by mining tools, databases and protocols from publications across 25 domains, combined with retrieval-augmented planning and code execution. It generalized across tasks such as gene prioritization, drug repurposing, rare disease diagnosis and cloning, and was shown in case studies running analyses and lab protocols.
The Little Scientist is an LLM-based agent that follows the scientific method to discover new algorithms and ensemble strategies. It demonstrated its capability by outperforming existing methods in protein fitness prediction and DNA motif discovery.
BioFirewall is a middleware system designed to screen genome-writing plans for biosecurity risks, operating between the foundation model and DNA synthesizer stages. It demonstrated high accuracy in detecting hazardous sequences while maintaining low false-refusal rates, showcasing its potential as an effective governance tool during the design stage of agentic AI systems.
The system uses a 'Union-then-Filter' framework to prioritize cardioprotective drugs for anthracycline-induced cardiotoxicity by combining predictions from multiple graph learning models and filtering with evidence synthesis from PubMed abstracts. It successfully identified novel drug candidates that were validated in zebrafish models, demonstrating the pipeline's potential for translational research.
Research Assistant is an internal large language model-based system at AstraZeneca designed to assist scientists and clinicians in exploring biomedical questions across various data sources. The system provides a chat-style interface that supports both direct question answering and multi-step research tasks, demonstrating its utility in supporting day-to-day R&D workflows.
OntoCodex is a multi-agent framework that uses large language models and ontologies to semi-automatically enrich biomedical knowledge bases, improving precision, recall, and F1 scores in concept extraction across various chronic diseases. The system demonstrated enhanced accuracy in standardized terminology mapping for medications but showed lower performance in laboratory test mapping.
CASCADE is an agentic framework that predicts downstream transcriptional effects of gene perturbations using precomputed regulatory networks. It demonstrates strong concordance with real patient data for MYC knockdown predictions, outperforming baseline methods and existing knowledge.
ANCHOR-RE is an agentic framework for biomedical relation extraction that uses retrieval augmentation and error-driven verification to enhance the reliability of knowledge graph construction and hypothesis generation. The system demonstrated improved performance across various benchmarks, showing its capability to generalize well even with newly published literature.
Frames oncology clinical development as an offline decision task where an agent predicts a drug program's next trial portfolio from date-gated public records, and compares offline-trained policies with frontier LLM agents using a shared retrieval scaffold. Reward-weighted behavioral cloning beat the best tool-using LLM agent on indication-level and strict F1.
The Gemma curation agent is an automated software tool that uses large language models and traditional methods to curate biological sample data for genomics studies, achieving near-human performance at a fraction of the cost and time. The system's evaluation includes benchmarking against human-curated datasets and preliminary error triage methods to identify potential inaccuracies.
RelAgent is a multi-agent system designed to bridge the semantic-structural gap in molecular relationship grounding, improving tasks like patent interpretation and structural analysis. It demonstrates substantial improvements over existing models on the MolGround benchmark, showcasing its effectiveness in entity extraction, substructure localization, and ontology-guided relationship reasoning.
MDArena evaluates the performance of coding agents on realistic molecular dynamics tasks, finding that while some configurations show promise with partial successes, significant challenges remain for complex and detailed scientific workflows. The benchmark provides a platform to track progress in automating biomolecular simulation research.
EMBL AI Librarian is a knowledge layer designed to enhance the Europe PMC interface for AI agents, enabling them to query in natural language and receive precise evidence. It demonstrates improved performance across various benchmarks, including literature synthesis and claim verification tasks.
CyberNeuro is an AI-driven workbench designed to automate and streamline neuroimaging data analysis, reducing the need for manual intervention. It demonstrated significant improvements in accuracy and efficiency compared to existing methods while maintaining clinical-grade privacy.
Plato-Bio extends the Plato/Denario research agent architecture for biology with explicit workflow states, provenance records, citation checks, claim-to-evidence links and publication gates. A source audit fixed three evaluation defects, and in a frozen pre-1986 rediscovery task it ranked the later-confirmed fish oil and Raynaud link first.
EnzGFM-Agent is a protein language model specifically designed for enzymes, which uses hierarchical pre-training to predict catalytic properties from amino acid sequences. The system demonstrates improved performance and efficiency in predicting enzyme properties compared to general-purpose models, with experimental validation showing its potential in reducing the need for extensive wet-lab screening.
The paper introduces an agentic calibration method using a large language model to optimize grey-box simulation models, showing it achieves lower error rates with fewer evaluations compared to traditional methods under unconstrained conditions. Under constrained conditions, the method performs comparably while being more efficient and offering explainability through textual rationale.
ReCo is a self-configuring and self-extending agentic framework for biomedical research that uses a large language model to interact with various computing tools and servers. It successfully operationalized five heterogeneous systems, demonstrating its capability in configuring, documenting, and integrating diverse biomedical frameworks.
BioPathfinder links publications with patient single-cell data in a provenance-aware knowledge resource and uses role-specialized LLM agents to generate, review and prioritize mechanistic hypotheses. Applied to CAR-T studies, it nominated NKG2A, and blocking it improved CAR-T antitumor activity in vivo.
Retrieves background knowledge, fixes admissible variables and mechanistic constraints, proposes candidate differential equations for biological dynamical systems, then fits and evaluates them. It recovered correct state variables and structures in retrieval and extrapolation tasks, with ablations showing the knowledge-based constraints were essential.
EcoXAI is a containerized multi-agent system with bioinformatics agents and explicit knowledge graph integration that runs staged analyses and produces grounded explanations. In an Alzheimer's drug repurposing case it evaluated 103 candidates and flagged 79 novel ones, including maraviroc, with later literature support.
The paper evaluates an AI-assisted pipeline for revising Delphi statements on RNA-based medicines, comparing its performance to human reviewers. The AI system, comprising multiple agents for citation detection and summarization, achieved comparable consensus rates with human oversight, demonstrating potential for accelerating expert consensus while maintaining methodological rigor.
OCellus is a nine-billion-parameter language model fine-tuned on various biological tasks, enabling it to perform spatial reasoning, perturbation prediction, and multi-task queries without specialized architectural modifications. It demonstrates superior performance across multiple benchmarks compared to existing models, showcasing its capability in natural-language explanations of predictions.
Ensemble QSP is a multi-agent framework with hierarchical memory architecture designed for long-horizon computational modeling tasks in biology and biomedicine. It demonstrated improved parameter recovery, robust prompt interpretation, and independence from proprietary LLMs across various scientific domains.
SpaCellAgent is a self-evolving multi-agent framework that automates trajectory inference analysis in computational biology, reducing manual intervention and improving analytical efficiency. It demonstrates significant performance gains over existing methods on diverse datasets, showcasing its potential to democratize advanced spatiotemporal modeling.
BioMaster is a multi-agent system designed to automate complex bioinformatics workflows, integrating planning, execution, error recovery, and validation. It demonstrated superior performance in completing more complex workflows compared to existing automated systems across various tasks.
The paper presents a multi-agent framework using large language models to analyze sleep disorders by integrating various lifestyle, behavioral, and physiological factors. The system identifies key drivers of sleep disorders through an interpretable approach that enhances evidence-based reasoning in clinical research.
GeneGenie is a multi-agent framework that enhances biomedical question-answering by orchestrating query planning, retrieval from curated databases, and execution of bioinformatics tools. The system demonstrates superior performance compared to standalone large language models in terms of accuracy, completeness, and factual grounding.
Agentomics is an autonomous system that uses large language models to develop machine learning solutions for biomedical datasets, implementing various modeling strategies and validation checkpoints. It demonstrated superior performance compared to other agentic systems and generated state-of-the-art models for several benchmark datasets.
The paper describes an AI-driven framework that integrates large language models with laboratory automation to validate and refine scientific hypotheses in systems biology. The system identified novel biological interactions, such as the effect of glutamate on cell growth under specific conditions, demonstrating its potential for reliable machine-driven discovery.
AgentODE is a framework that uses an LLM to propose ODE structures and iteratively refines parameter distributions using population-level summary statistics. It demonstrates the ability to recover functionally consistent ODE structures in sparse, noisy data settings for rare diseases like RDEB.
GeneBench-Pro is a benchmark designed to evaluate AI agents' ability to perform multi-stage reasoning in genomics and translational biomedicine. It includes complex, real-world problems reviewed by domain experts and demonstrates that current models like GPT-5.6 Sol struggle with long-horizon biological reasoning.
The paper presents Clair3-Connect, a client-server rearchitecting of the variant caller Clair3 to enable efficient agent-driven analysis. It demonstrates that this design reduces token usage and improves reliability compared to traditional shell-driven methods.
Octopus couples local LLM agent swarms with deterministic modeling engines to generate therapeutic hypotheses from CRISPR dependency data, trace causal cascades and test them against in vivo and survival data. An unsupervised sweep of colorectal cancer transcriptomes nominated IGF2 in 5-fluorouracil resistance, which held up statistically and in an independent mouse cohort.
Exposes the Virtual Fly Brain knowledgebase of Drosophila neuroanatomy, connectome and molecular data to LLMs through Model Context Protocol tools. On 30 neuroscience tasks the MCP-equipped LLM gave precise, verifiable answers on 25, versus 14 with web search and 2 for a bare LLM.
The Process-Reward Tactic Evolution framework trains an agent to execute complex bioinformatics workflows in Galaxy, a platform for managing and executing such workflows. The system demonstrates improved completion rates, biological correctness, and execution efficiency compared to baseline methods.
The paper introduces a dual-agent framework for translating natural-language biological experiment protocols into executable commands for robotic laboratory platforms, addressing the semantic gap between human-readable instructions and machine-executable code. The system was evaluated through automated benchmarks involving various parsers and validators, demonstrating its effectiveness in accurately converting ELISA protocols and enabling autonomous execution of protein quantification assays.
FlowAgent is a modular, provider-agnostic framework for evaluating agentic bioinformatics systems by separating planning, fault recovery, and biological interpretation. The study reveals that while models excel at generating valid workflow plans from named toolchains, they struggle with inferring appropriate toolchains from biological intent alone and handling data-grounded interpretation.
Orion is an AI agent designed for biomedical image analysis and interpretation, capable of automating computational workflows in laboratory settings. It demonstrated the ability to generate research reports from experimental imaging data and produce biologically plausible hypotheses that were prioritized by human scientists.
ProvenanceGuard is a source-aware factuality verification system designed to ensure that claims made by large language model agents are correctly attributed to their sources. It evaluates the provenance of answers generated through the Model Context Protocol, demonstrating high accuracy in detecting and correcting cross-source conflation errors.
The paper presents Model Context Protocol (MCP) servers that enable AI agents to rapidly prototype multicellular mechanistic models through natural language interactions with bioinformatics tools. The system demonstrated the ability to construct a multiscale cancer cell fate model using three different tools, showcasing its potential for accelerating biological research and hypothesis exploration.
DeepRoot is a multi-agent system that uses historical medical texts to discover potential drug treatments by building and utilizing a verified knowledge graph. It outperforms other systems in recovering known compound-disease pairs and maintaining reasoning coherence without hallucination.
DeepTCM1.0 is a multi-expert AI agent framework that integrates classical TCM theory with modern scientific research to analyze the mechanisms of Chinese herbal formulae, demonstrating robust reliability and validity in its evaluation through multiple independent assessments.
PathLab is an autonomous framework that translates natural-language research objectives into executable computational pathology workflows, reducing the need for programming expertise. It demonstrated non-inferior performance compared to expert implementations across various tasks and significantly reduced the time required for designing and executing computational pathology studies.
The paper evaluates OpenClaw, an AI agent augmented with medical research skills, in a non-small cell lung cancer biomarker task. It shows that skill-augmented outputs are directionally higher in quality compared to native AI models but the difference is not statistically significant due to high variability and noise.
ABC-Bench is a benchmark suite designed to evaluate the biosecurity-relevant capabilities of large language model agents in biological research tasks. The system demonstrated that LLM agents can outperform human experts on various biology and software tasks, including scripting for liquid handling robots and designing DNA fragments, with some limitations in novel bioinformatics reasoning.
The paper evaluates the performance of Biomni and other AI systems in retrieving viral genome data from NCBI's Virus database, finding significant variability in accuracy. It introduces gget virus, a deterministic query framework that enhances retrieval accuracy and stability for these systems.
The system automatically generates OMOP concept sets from natural language descriptions, combining various vocabulary tools and ontologies. It demonstrated high precision and recall in cohort-level validation against EHR data and outperformed unrestricted SQL access and PHOEBE 2.0 on a benchmark of AD/ADRD studies.
CodeCytos is an AI-assisted framework that uses code-driven automation to analyze spatial molecular imaging data, enabling dynamic and customizable interaction with tissue images. The system demonstrates improved efficiency and flexibility in exploring custom spatial cellular features across various tissue types through case studies and evaluations by bioscientists.
CP-Agent is a context-aware multimodal large language model designed to interpret cellular morphological changes under chemical perturbations, generating human-interpretable rationales for mechanism-of-action inference. It demonstrates high accuracy in distinguishing treatment effects and mechanisms with an F1-score of up to 0.896.
A ReAct-style agent selects among biomedical knowledge bases, vector retrieval over 44 million documents, external APIs and NER tools to answer scientists' questions. It led on long-form accuracy and faithfulness, was competitive on short-form questions, and serves over 1,600 users in production.
BIOGEN is a multi-agent framework designed to interpret RNA-seq data in the context of antimicrobial resistance, integrating biomedical retrieval and structured interpretation. It demonstrated strong biological coherence and evidence-grounded reasoning across multiple datasets compared to LLM-only baselines.
The paper explores the use of large language models and agentic systems for creating interactive multiview visualizations in genomics, addressing challenges such as data heterogeneity and complexity. It evaluates different configurations of these systems and finds that agentic iteration significantly improves visualization quality compared to baseline methods.
PRAXIS is a framework for biological research agents that uses literature learning and case distillation to create structured long-term memory, enabling it to support various biocomputational tasks. The system demonstrates improved method selection, error suppression, and workflow organization in complex biological research through object validation and public benchmarks.
NanoCortex is a unified agentic system for nanopore sequencing analysis that bridges the gap between raw signal basecalling and biological interpretation, using a multi-agent architecture. The system demonstrates significantly higher usability compared to general-purpose large language models in complex analytical tasks.
MARICL is a framework that uses LLM agents to analyze residuals from base models and iteratively refine explanations for scientific data, particularly in biomedical contexts. The system demonstrates its ability to generalize beyond batch-specific noise by successfully applying learned formulas across different experimental conditions.
S2F-Agent is a human-in-the-loop framework that orchestrates sequence-to-function models for genome interpretation, translating biological requests into reliable execution and rigorous downstream analysis. It demonstrated superior performance in various genomic analyses compared to general-purpose language models, showcasing its reliability and scalability.
The EOS AI Agent integrates large language models with laboratory orchestration to enable scientists to create and monitor automated lab protocols using natural language. It was evaluated in three simulated labs, achieving high success rates in protocol generation and significantly reducing interface actions.
BiomniBench is a process-level evaluation framework for large language model agents in biomedical research, scoring their full trajectory against expert-designed rubrics. It reveals that while models can reliably ground claims in real sources, they struggle with method selection and scientific reasoning.
MechAInistic is a multi-agent system that generates auditable mechanistic drug hypotheses by grounding language model outputs in computational evidence. It demonstrated the ability to identify specific metabolic pathways and drugs for therapeutic hypothesis generation in immune-cell diseases, outperforming general-purpose LLM systems in terms of auditability.
ClaroAI-Bench is an evaluation suite designed to measure AI agents' ability to reproduce computational findings from real biomedical research papers. It demonstrates that full-capability agents can achieve a 60.6% success rate in reproducing results, highlighting the importance of accessible data and code for successful reproduction.
The paper presents an expert-in-the-loop LLM workflow for extracting data from literature on targeted protein degradation, achieving high accuracy in expanding molecular glue and PROTAC databases. The system demonstrates the potential of AI-assisted scientific curation to enhance predictive models by providing essential experimental context.
AssayBench is a benchmark for predicting phenotypic screen outcomes using large language models and agentic systems, formulated as gene rank prediction tasks. It demonstrates that zero-shot generalist LLMs outperform specialized biology-focused models in this task.
Open-Rosalind is a biomedical agent system that emphasizes evidence-grounded outputs and tool-mediated execution, designed to enhance accountability in scientific research. The study demonstrates the importance of structured workflows and explicit tool usage for improving accuracy and reliability, though it also highlights challenges in external validity.
MechAInistic is a multi-agent system that uses large language models to convert natural-language biological questions into executable workflows for metabolic model analysis. It demonstrated its utility in generating therapeutic hypotheses by identifying potential drug targets and repurposing candidates from paired immune-cell metabolic models.
BioProVLA-Agent is a multi-agent system designed for biological laboratory manipulation, using protocols as the task interface and integrating visual state verification with embodied execution. The system demonstrates improved robustness in executing complex tasks under challenging visual conditions compared to existing methods.
Starling is a multi-agent system that autonomously extracts nuanced biomedical knowledge from PubMed, producing large-scale datasets with high precision and recall. It demonstrates the potential of AI-driven therapeutic design by outperforming existing curated databases in terms of accuracy and scale.
SPARROW is a multimodal framework that integrates genomic, imaging, clinical and cognitive data for Parkinson's disease subtyping. It uses chain-of-thought reasoning to provide transparent subtype decisions, highlighting contributions from each data source, and demonstrated superior classification performance on the PPMI dataset.
SciResearcher is an automated framework that constructs data for frontier scientific reasoning by synthesizing diverse tasks and eliciting sophisticated computational and reasoning capabilities. It demonstrates superior performance on benchmarks like HLE-Bio/Chem-Gold, SuperGPQA-Hard-Biology, and TRQA-Literature compared to larger proprietary agents.
The paper evaluates the performance of advanced language model agents on a curated bioinformatics benchmark, demonstrating near saturation accuracy with skill augmentation. It highlights that access to high-quality scientific skills and external resources significantly improves agent performance.
NEURA is a framework that automates neuroimaging research workflows by converting free-text questions and datasets into executable analysis plans, validated outputs, and structured reports. It demonstrates improved accuracy in planning tasks compared to direct LLM queries and effectively detects hallucinations without false positives.
Separates governance, observability and orchestration layers from pluggable agent teams for phenotyping, study design and analysis over OMOP CDM databases, enforcing deterministic validation outside agent reasoning. A natural-language-to-SQL agent team on Synthea, MIMIC-IV and an NHS dataset reached reliability scores of 0.84 to 0.94 and blocked all adversarial queries.
MSAgent is an autonomous framework that uses large language models to integrate over 50 mass spectrometry tools, enabling automated and evidence-grounded molecular identification from raw data. It demonstrated significant improvements in retrieval performance and reliability across various benchmarks compared to existing methods.
GeneBench is a benchmark designed to assess AI agents' performance in multi-stage scientific data analysis tasks relevant to genomics and quantitative biology. It evaluates the ability of these systems to handle complex, real-world challenges such as data cleaning, exploratory analysis, model selection, and producing actionable conclusions.
Extracts claims and citations from a submitted pneumonia quality measure, retrieves further PubMed evidence and rates the strength and agreement of evidence for each claim with justifications. It judged 64 claims and 355 claim-evidence pairs in about five hours, agreeing with a subject matter expert on 69 percent of claims.
The study evaluates the reliability of large language model agents in annotating protein sets using a controlled task, finding that while coverage is complete, reproducibility and biological correctness vary. The research highlights the need for deterministic evidence tables and expert review to ensure high-quality annotations.
SIMBA is an AI-driven platform that integrates single-molecule localization, spectral processing, and deep learning-based denoising into a unified framework. It demonstrates the ability to support standard imaging workflows while enabling advanced analysis of multi-dimensional imaging data through user-guided interactions.
MetaMuse is a multi-agent AI system that autonomously curates and harmonizes unstructured biomedical metadata from public repositories like GEO. It demonstrated high accuracy in curation tasks, achieving over 95% across key fields on a gold-standard dataset.
U-Probe is a universal probe design platform for spatial-omics that uses AI agents to facilitate natural language-based workflows and supports diverse probe architectures. It was validated in various scenarios, demonstrating its capability to drive MiP-seq panel design, genome tiling DNA-FISH, and novel RCA-based ligation probes.
GPCR-Nexus combines a GPCR-ligand knowledge graph with literature retrieval and a multi-agent architecture for query planning, evidence retrieval, validation and synthesis. It returns citation-backed answers about receptor interactions, mechanisms and disease links, shown through example queries.
VCR-Agent is a multi-agent framework that integrates biologically grounded knowledge retrieval and verification to generate and validate mechanistic reasoning in virtual cells. The system demonstrates improved factual precision and effectiveness in gene expression prediction when trained with verified explanations from the VC-TRACES dataset.
Pipette is a multi-agent AI framework that uses a literature-derived Skill Graph to guide the execution of bioinformatics workflows through natural language interaction. It was benchmarked across various biological domains and showed superior performance in generating coherent, biologically valid workflows compared to LLMs without such constraints.
The paper introduces a framework where researchers can describe their microscopy problems to an LLM agent, which then autonomously designs and trains deep learning models without further human intervention. The system was validated across various microscopy modalities, demonstrating its ability to achieve high accuracy in tasks such as nuclear segmentation and histopathology classification.
MedResearchBench evaluates six AI research systems on clinical epidemiology tasks using NHANES data, revealing that citation accuracy is crucial for system reliability. The study also demonstrates the effectiveness of a multi-agent pipeline in improving citation integrity and overall performance.
The Medical Concept Mapping (MCM) system uses language models to rephrase ambiguous biomedical text into standardized descriptions before linking them to concepts in terminologies and ontologies. This approach improves the accuracy of concept mapping, especially for underrepresented or abbreviated terms, outperforming existing methods on multiple benchmarks.
The paper evaluates three AI agents (Codex, Biomni Lab online, DeerFlow 2) on their ability to retrieve coccolithophore calcification-related proteins from UniProt. Codex showed the best balance between sensitivity and specificity, while also demonstrating high repeatability across runs.
Benchmarks reasoning and non-reasoning LLMs on screening abstracts and extracting functional evidence for genomic variants from ClinGen curations, then builds an end-to-end pipeline that expands identifiers, retrieves literature, filters, fetches PDFs and drafts evidence reports. The reasoning model classified full-text evidence with 96 percent accuracy and much higher specificity.
The VaaS pipeline is a systematic approach to reduce hallucinations in AI-generated scientific content, particularly for rare disease gene curations. It demonstrates significant reduction in Type I and II errors through iterative refinement and validation, achieving near-zero error rates with minimal cost per review.
An autonomous research framework that turns surveyed literature into clinically grounded ideas, implements and runs experiments and drafts manuscripts under three levels of autonomy. Across 171 cases its ideas were rated above those of commercial LLMs and its experiments executed successfully more often.
The Judge Agent automates the validation of scientific simulation code generated by large language models, significantly reducing silent failure rates. It introduces a structured specification format and formalizes the boundary for reliable automated error certification.
Three tool-using agents extract evidence-table values from trial publications, an attribution verifier confirms each value on its cited page, and agent disagreement routes cells to human review whose feedback updates a curation knowledge base. On a clinician-annotated benchmark it reached 91.7 percent accuracy autonomously and 95.2 percent after reviewing 15.6 percent of cells.
The Virtual Study Group is an agentic AI framework that uses hierarchical feature selection to identify ageing-related Gene Ontology terms and generate biological knowledge. The system's findings were validated by reviewing existing literature, demonstrating its effectiveness in scientific discovery.
An LLM agent repeatedly checks its structured outputs against variable dependencies, the source text and retrieved domain knowledge, revising until they converge. In colorectal synoptic reporting, Ewing sarcoma CD99 patterns and lung cancer staging it raised accuracy, for example tumour stage from 0.680 to 0.833.
HARVEST is a multi-agent AI system that autonomously extracts and structures bioactivity data from pharmaceutical patents, significantly expanding the available chemical and biological knowledge. The system's performance was validated through comparison with human-curated databases and revealed limitations in existing models when dealing with novel scaffolds and uncharacterized protein targets.
The Samyama Graph Database constructs large-scale biomedical knowledge graphs from diverse public data sources, enabling cross-KG federation and AI agent access. It demonstrated high accuracy in answering pharmacology questions compared to standalone LLMs or schema-aware text-to-Cypher approaches.
stMCP is a Model Context Protocol framework that enables natural language-driven spatial transcriptomics analysis, executing tools locally to ensure data privacy and reproducibility. It demonstrates robust performance in biological discovery tasks, reducing the need for bioinformaticians by empowering biologists to independently explore their data.
EnsAgent is a multi-agent system designed for robust annotation in spatial transcriptomics, decoupling structural partitioning from semantic labeling to ensure reproducibility and accuracy. It demonstrates superior performance by effectively neutralizing batch effects and resolving subtle tumor microenvironment niches missed by single-paradigm approaches.
The system uses multi-agent large language models to automate the creation and optimization of protein purification protocols, reducing manual labor and analysis time. It demonstrated high accuracy in protocol extraction and generated expert-validated optimization recommendations.
ARMS is an AI agent that uses large language models to standardize legacy biomedical metadata by dynamically querying authoritative terminology services at runtime. The system demonstrated significant improvements in precision and recall compared to a prompt-based approach, particularly for ontology-constrained fields.
The paper describes a formalized scientific methodology that enables AI agents to conduct rigorous research across various domains, including population genomics. The approach involves decomposing the research process into procedural workflows and integrity disciplines, which are validated through end-to-end projects and controlled studies demonstrating improved auditability and evidence-backed outputs.
The Robust Biomedical Reasoning Agent (RBRA) enhances the robustness and accuracy of large language models in biomedical multi-hop reasoning tasks. It achieves this through hierarchical query decomposition, dynamic context filtering, iterative fact verification, and optimization for robust metrics, demonstrating superior performance on a challenging dataset compared to existing methods.
EnzySeek is an AI agent designed to assist in enzyme catalysis simulations by integrating a domain-specific knowledge base and custom Model Context Protocol interfaces. It demonstrates the ability to autonomously execute computational tasks such as molecular docking, MD simulations, and QM/MM calculations while continuously learning from new literature and being validated through human expert review.
HarmonyCell is an AI agent designed to automate single-cell perturbation modeling by addressing semantic and statistical heterogeneity challenges. It demonstrates high execution rates on heterogeneous datasets and matches or exceeds expert-designed models in out-of-distribution evaluations.
ProtRLSearch is a reinforcement learning-trained agent designed to perform multi-round, multimodal protein sequence and text searches for healthcare applications. It demonstrates improved accuracy in reasoning about disease-related variants and protein functions by integrating both textual and sequence data, as validated through the ProtMCQs benchmark.
ChatSpatial is a schema-enforced orchestration framework that uses an LLM to select versioned tools and parameters for spatial transcriptomics analyses, ensuring reproducibility and cross-platform compatibility. In cancer case studies, ChatSpatial successfully recovered analytical structures and conclusions from published research with higher execution success and output concordance compared to code generation alternatives.
Eubiota is an open-source modular agentic framework designed for autonomous discovery in the gut microbiome, using specialized agents for planning, execution, verification, and synthesis. It demonstrated its utility by identifying key DNA repair genes under inflammatory stress, designing a bacterial consortium to attenuate colitis severity, and discovering diet-associated metabolites that suppress NF-kB signaling.
PantheonOS is a multi-agent framework designed for automatic genomics discovery, enabling the evolution of code and algorithms to achieve super-human performance in tasks such as gene panel selection and batch correction. It demonstrated its capability by uncovering novel biological insights in early mouse embryo development and human fetal heart disease.
CELLSWARM uses large language models to drive autonomous cell agents in simulating tumor microenvironment dynamics, achieving performance comparable to traditional rule-based methods. It demonstrates capabilities such as cross-cancer generalization and sensing indirect genetic perturbations that go beyond the limitations of conventional simulation approaches.
The Web of Microbes (WoM) Agent uses a Bayesian Personalized Ranking model to predict bacterial substrate preferences, integrating it with growth models and large language models for autonomous analysis. It demonstrated accurate predictions in novel medium experiments and correctly identified bacteria enriched in soil metabolite spike-in tests.
The DAS red-teaming framework autonomously tests large language models in clinical settings for safety and reliability across various axes, revealing significant discrepancies between static benchmark performance and real-world robustness. This dynamic stress-testing approach highlights the need for continuous evaluation to ensure trustworthy medical AI deployment.
KG-Orchestra is an open-source multi-agent framework that enriches seed biomedical knowledge graphs by autonomously acquiring, validating, and integrating evidence to create specialized, directional BKGs. The system was evaluated in two contexts, demonstrating superior reasoning performance with Qwen 3 variants and enhanced evidence relevance through hybrid retrieval strategies.
Agents decompose constraints, extract provenance-aware records from literature and check consistency to build a knowledge graph of 34,215 living material records linking cells, matrices, functions and test conditions. In a prospective wound-healing task it prioritized a four-component design with strong in vivo performance.
This work trains a hierarchical multi-agent system with process-level rewards for gene-disease validity curation, where a supervisor agent calls other agents as tools. On ClinGen data, outcome-only rewards raised accuracy to 0.732 but gave poor process alignment, while adding process rewards improved adherence to valid curation pathways.
NeuroWeaver is an autonomous evolutionary agent designed to explore and optimize EEG analysis pipelines, incorporating neurophysiological knowledge. It demonstrated superior performance compared to existing methods on various benchmarks, achieving high accuracy with significantly fewer parameters.
CellMaster is an AI agent designed to annotate cell types in single-cell RNA sequencing data by leveraging large language models for on-the-fly annotation. It demonstrated significant improvements over existing tools, particularly in identifying rare and novel cell states.
scChat is a large language model-powered co-pilot designed to enhance the analysis of single-cell RNA sequencing data by integrating research context and supporting hypothesis validation, mechanistic interpretation, and experimental design. The system demonstrates high accuracy in cell type annotation while providing biologically relevant explanations and insights.
BioGuider is a multi-agent platform that evaluates and improves the documentation quality of bioinformatics software, treating documentation as a testable object. It demonstrates robust error detection and correction capabilities and shows a positive correlation between better documentation quality and higher citation frequency among widely used tools.
ToxMCP is a set of Model Context Protocol servers that wrap toxicology tools such as CompTox, ADMETlab, AOP services, the QSAR Toolbox and PBPK simulation as typed, auditable tools for LLM agents. The authors show natural-language risk questions compiled into tool calls and outline a protocol for evaluating reproducibility and utility.
PBio-Agent is a multi-agent framework designed to predict gene regulation responses under complex chemical perturbations in bulk-cell environments, addressing the limitations of existing models. It integrates specialized agents with biological knowledge graphs and iterative refinement processes, outperforming baselines on LINCSQA and PerturbQA benchmarks.
BOLE is a multi-agent AI framework designed to enhance genomic breeding analysis by integrating heterogeneous bioinformatics tools into coherent workflows. It demonstrated the ability to autonomously generate and execute complex analytical workflows, supporting tasks such as genome-wide association studies and heritability estimation, while ensuring reproducibility through artifact-level provenance.
The EpiCausalX Agent is an AI-powered tool designed to automate causal inference in epidemiology, integrating literature retrieval and causal reasoning. It demonstrated efficient automation of the end-to-end workflow, producing evidence-based conclusions with interactive visualizations accessible via a user-friendly interface.
LABBench2 is a roughly 1,900-task successor to LAB-Bench measuring what AI systems can do on realistic biology research tasks. Frontier models improved over LAB-Bench but scored 26-46% lower on the harder tasks, leaving a clear gap.
S1-NexusAgent is a self-evolving agent framework designed for multidisciplinary scientific research, capable of handling complex workflows and large-scale data through hierarchical planning and execution. It demonstrates state-of-the-art performance on authoritative scientific benchmarks involving long-horizon planning and tool orchestration.
The AI-Native Biofoundry integrates large language models and active learning to autonomously engineer enzymes, bridging the gap between computational design and physical experimentation. In a proof-of-concept study, it successfully evolved a DNA polymerase for improved sequencing accuracy with non-natural nucleotides.
VenusRAR is a two-stage agentic framework that uses multi-agent collaboration to predict protein mutations, improving the efficiency and accuracy of wet-lab validation. The system demonstrated significant improvements in predicting mutation fitness and identified novel mutants with enhanced activity through rigorous evaluation.
BioAgent Bench is an evaluation suite for AI agents in bioinformatics tasks, featuring curated end-to-end tasks and automated assessment. The system reveals that while advanced LLMs can reliably complete complex pipelines, they exhibit vulnerabilities under controlled perturbations.
The authors release a 16-million-abstract biomedical search corpus and a 60k-question factoid dataset, then train search agents with verifiable rewards to reason and retrieve over it. The trained agents beat non-reinforcement retrieval baselines and showed planning and self-verification behavior.
The paper evaluates the accuracy of large language models (LLMs) in generating biomedical data analysis code and finds a low overall accuracy, raising concerns about their reliability. The authors develop an AI agent that iteratively refines analysis plans before generating code, achieving higher accuracy and enabling users to complete over 80% of analysis code for three studies in a user study.
DSGym is a framework for evaluating and training data science agents, addressing limitations of existing benchmarks by providing a standardized, modular architecture. It includes tasks from various domains including bioinformatics and molecular prediction, demonstrating the ability to rigorously measure agents' capabilities in realistic scientific contexts.
MARBLE is a multi-agent system designed to automate the iterative process of bioinformatics model improvement, including hypothesis formulation and architectural redesign. It demonstrated sustained performance improvements across various bioinformatics tasks while maintaining robust execution stability.
GenomAgent is a multi-agent system designed to enhance genomic question answering by coordinating specialized agents for complex queries, surpassing the previous state-of-the-art GeneGPT. It demonstrates improved performance and flexibility across various scientific domains requiring expert knowledge extraction.
PFUA is a tool-augmented agent for protein function understanding that decomposes questions, calls biological tools for verifiable intermediate evidence and then answers. Across four benchmarks it outperformed text-only reasoning models, which the authors show mostly amplify keyword patterns.
The paper evaluates eight AI frameworks on their ability to autonomously conduct scientific research tasks, finding that while the systems can handle conceptual planning and summarization well, they struggle with robust implementation and produce hallucinations. The study highlights the need for human supervision in deploying these AI tools effectively.
HeurekaBench is a benchmarking framework designed to evaluate AI systems acting as co-scientists in biology, focusing on single-cell analysis. It demonstrates the effectiveness of adding critic modules to improve the performance of open-source LLM-based agents.
An LLM agent refines computational protein function predictions by querying ontology constraints, structured metadata and literature and documenting its evidence. It achieved the lowest Smin on all three Gene Ontology branches and the best Fmax for molecular function and cellular component.
MedHopper is a biomedical question-answering system that uses an agentic retrieval and language model pipeline to answer complex, multi-hop questions from a Wikipedia-derived corpus. It demonstrated improved performance on the MedHopQA task with exact match scoring, highlighting the effectiveness of its controlled state machine for routing queries and executing up to three hops of entity extraction.
Magnolia is an agent framework for computational chemistry with a three-tier memory that distils session observations into project entries and human-validated cross-project rules. Across four case studies, including protein-peptide docking and modified-DNA simulations, it tracked 317 runs and memory-enabled replays passed configuration checks that memory-free runs failed.
PRAXIS-MD is a long-term brain-enhanced agent framework that converts natural-language biological questions into auditable molecular dynamics workflows, using case memory and reusable skills. The system demonstrated improved reproducibility and interpretability in MD research when evaluated against benchmark tasks.
The paper describes an AI-driven pipeline for designing protein-binding aptamers that iterates between computational design and wet-lab testing. The system showed significant improvement in binding affinity and specificity of the designed aptamers over successive rounds, demonstrating the value of integrating empirical lab data into iterative molecular discovery processes.
Places an LLM among molecular tools, single-step prediction engines, a purchasable materials library, route memory and a multi-criterion evaluator, learning from feedback to backtrack and revise weak routes. With Claude Opus 4.8 a top-10 route passed audit for 86.7 to 95 percent of targets, and two routes were executed in the wet lab after expert revision.
Couples over 80 scientific operations with induction of conditional skills from execution trajectories, which a separate review agent verifies before they guide later tasks. It beat reported MolClaw results on MolBench, and a compound designed in a prospective HPK1 study reached a cellular IC50 of 14.19 nM.
Organizes about 1.6 million commercial scaffolds into a coarse-to-fine action space so an agent can jump to new scaffold regions and then edit candidates, keeping trajectory memory for long searches. Trained with agentic reinforcement learning, MolHop-9B beat large generative baselines on cross-scaffold success while producing more distinct scaffolds.
BioAgent-RAG is a multi-agent framework designed for drug repurposing that integrates knowledge graphs to enhance biological plausibility and reduce reasoning errors. The system demonstrates superior performance on various benchmarks, indicating its potential as an explainable computational tool for prioritizing drug candidates.
CLARA uses a language model to guide batch active learning for ligand discovery, making strategic decisions about chemical compound selection. It outperformed random sampling in most scenarios and showed promising results in terms of recall rates across different protein targets, demonstrating the feasibility of using AI agents for molecular drug design.
MADDIE is a multi-agent system designed for drug discovery that includes specialized agents for different stages of the process, such as hit identification and lead optimization. It demonstrates superior performance compared to existing systems by efficiently generating molecules with desired properties and synthesis routes across 50 protein targets.
The paper evaluates the performance of autonomous large language model (LLM) agents in molecular prediction and optimization tasks, showing that these agents can develop methods competitive with human-developed approaches across various drug discovery benchmarks. The study demonstrates the capability of LLMs to autonomously create high-performing models for ADMET property prediction, activity prediction, and practical molecular optimization.
MAGI is an open modular agent designed to coordinate molecular design tools for drug discovery, capable of generating molecules either directly or by delegating tasks. The study shows that MAGI's effectiveness in real-world projects is limited more by the accuracy and applicability domain of predictive models than by its own orchestration capabilities.
EASER is a reflective agent designed to bridge the gap between scientific reasoning and sequence manipulation in peptide design, using evidence-aware strategies. It demonstrates superior multi-objective performance in designing antimicrobial peptides by integrating literature evidence with iterative feedback mechanisms.
The paper introduces a lightweight framework where a compact language model coordinates 18 scientific tools for early-stage drug discovery, demonstrating its ability to generate parseable and schema-compliant plans. The evaluation shows high performance in tool selection and argument matching but highlights the challenge of compositional generalization.
ELMS is a multi-agent system that uses evaluator-in-the-loop Monte Carlo Tree Search to guide motif scaffolding in protein design, improving the generate-then-filter paradigm by incorporating structural evaluation feedback into iterative design processes. The system demonstrates superior performance compared to existing methods on standard benchmarks for single and paired-motif tasks.
M3OS is a multi-agent system that uses large language models and Monte Carlo graph search to optimize small molecules for drug design, maintaining evidence-traced optimization histories. It demonstrates higher success rates in molecular optimization benchmarks compared to baseline methods.
AnewDDE is an integrated drug discovery engine that combines various AI capabilities for biomolecular interaction modeling and design, including structure prediction, affinity estimation, molecular design, and experimental decision-making. The system demonstrates superior performance in identifying high-affinity binders and provides scientific reasoning and ADMET predictions, validated through internal benchmarks designed by pharmaceutical experts.
PFArena is a benchmark designed to evaluate the performance of various AI models in protein modification tasks, including single-mutant generation and multi-mutant ranking. The evaluation reveals that different model types excel under varying conditions but all face challenges as task complexity increases.
T-REX is an agentic framework that orchestrates multiple protein design tools to efficiently allocate computational resources and generate structurally distinct binder designs. It demonstrates adaptive orchestration's potential in complementing powerful protein design models by achieving higher throughput compared to baseline methods.
Contains 2,000 molecule generation and optimization tasks with implicit design narratives, explicit property constraints and infeasible cases, solved through 17 chemistry tools. The best frontier LLM succeeded on only about 43 percent, failing mainly on implicit constraints and infeasibility detection.
ChatT2 is a mentor-executor-evaluator agent system for research on bacterial type II polyketides that combines retrieval with bioinformatics and cheminformatics tools. The mentor clarifies user intent, the executor runs the tools and the evaluator checks outputs before they are returned.
TRACEDD is a multi-agent drug design framework in which LLM agents orchestrate tools for target validation, structure prediction, pocket finding, molecular generation, ADMET and retrosynthesis in a reason-act-observe loop. A JAK2 case study showed the workflow running end to end, including falling back to AlphaFold when no structure existed.
SpecOpt docks a compound against its intended target and known off-targets, compares residue-level contact differences and has an LLM propose structural edits that improve selectivity while keeping the scaffold. The authors built a ChEMBL-derived benchmark of intended and off-target activities to evaluate this.
A Claude Code agent prepared ligand libraries and ran Uni-Dock and DiffDock virtual screens on formyl peptide receptor 2 with either default choices or a Markdown skill file encoding expert protocol decisions. The skill-guided protocol raised Uni-Dock ROC AUC from 0.701 to 0.733.
Runs Kosmos, K-Dense, ToolUniverse, BioAgents and AI Scientist-v2 on reproducing antibody developability prediction and lipid-RNA molecular dynamics studies from industry. The frameworks caught some methodological issues but none matched the scope of the originals, and several hallucinated or failed badly.
ABLE benchmarks how well LLM agents use protein design tools such as ProteinMPNN and AlphaFold3 across structure retrieval, sequence generation and design validation. Of 15 frontier models, seven refused all tasks and the rest varied widely, with the best still inconsistent at planning compared with an expert baseline.
The Drug Discovery Agent (DDA) is a multi-agent framework designed for automated structure-based drug design, integrating various bioinformatics and molecular modeling tools. It was evaluated on the CrossDocked2020 benchmark, demonstrating superior performance with a higher screen-pass rate compared to specialized baselines.
DynaMate is an autonomous multi-agent system designed to automate protein-ligand molecular dynamics simulations, reducing setup time and improving efficiency. It successfully executed MD simulations on a diverse dataset of systems, demonstrating its adaptability and robustness in biomolecular modeling.
The paper introduces an evaluation framework for ChatInvent, a drug discovery assistant using large language models (LLMs), focusing on human alignment and output quality dimensions. It validates the LLM's performance as a judge through expert annotations and demonstrates improvements in aligning with human judgments.
The paper introduces a verification framework for AI-driven drug discovery in immune rejuvenation, featuring verifiers like PACE and CellQ to ensure predictions align with biological data. The system demonstrates self-correction capabilities by refining its scoring logic based on experimental results.
SABLE is a modular framework that uses an LLM to interpret user-defined goals and route tasks for hit-to-lead optimization in drug discovery. It enriches candidate sets by evaluating only a subset of the enumerated search space, demonstrating its effectiveness in prioritizing synthetically constrained analogs.
The GNPS2 Agentic AI system uses large language models to generate structural hypotheses for unknown drug metabolites from mass spectrometry data, integrating spectral alignment and machine learning-based spectrum prediction. The system successfully predicted and experimentally confirmed several novel metabolite structures, demonstrating its capability in generating testable hypotheses for previously uncharacterized compounds.
KnowMol couples LLM agents that generate knowledge-guided hypotheses about drug and protein features with machine learning validation in a closed loop to predict drug-target interactions. It outperformed prior methods on three interaction benchmarks and eight property datasets and was applied to PRMT3 inhibitors for NAFLD.
Pepti-Agent is an AI agent designed for the iterative design and optimization of therapeutic peptides, addressing multiple constraints such as solubility, hemolytic activity, and non-fouling properties. It demonstrates improved peptide candidate refinement through a closed-loop framework that integrates generation, property prediction, and mutation operations, offering a reproducible method for benchmarking and experimental validation.
Evolves populations of executable synthetic routes from purchasable building blocks, with an LLM choosing high-level search strategy while deterministic chemistry code builds, validates and scores routes. On a soluble epoxide hydrolase proxy task it beat single-pass LLM and deterministic controllers on binding score, synthesizability and route success.
AMREC is an AI system designed to recover the chemical validity of molecular structures generated by language models while preserving their structural identity. It demonstrates superior performance in recovering valid SMILES strings from invalid drafts compared to existing methods, as measured by various structural and exact-match metrics.
WMLLM is a self-evolving optimization agent that uses predict-then-act world modeling to improve the efficiency of black-box molecular optimization tasks. It demonstrates state-of-the-art performance in multi-objective molecular optimization under limited evaluation budgets.
Equips a pretrained protein language model with decoding that interleaves sequence generation with calls to structure prediction, stability and docking tools, trained by a trajectory-level preference method. Across enzyme design, antibody optimisation, thermostability and interface design tasks it improved on passive baselines, notably antibody top-10 percent hit rate.
The PROBE system uses an iterative multi-agent framework to optimize molecular structures for drug design, addressing the challenge of balancing binding affinity and druggability. It introduces diagnostic metrics that reveal common failure modes in existing LLM-based optimization pipelines and demonstrates improved performance on a benchmark dataset.
The PKFM is a grey-box Transformer framework that reconstructs sparse pharmacokinetic curves and initializes pharmacometric models, improving accuracy and stability in drug exposure inference. The system outperformed general-purpose programming tools on standard benchmarks but requires human confirmation for downstream use.
MolLingo is a multi-agent system that uses chemically meaningful representations to automate molecular design, outperforming existing LLMs and specialized methods in optimizing drug properties and target binding. It demonstrates the potential of LLMs as molecular design assistants when guided by biologically grounded structural context.
LipoAgent coordinates fine-tuned LLM agents for lipid nanoparticle design, predicting toxicity first and treating it as a precondition for efficiency prediction, with multi-agent verification and human review on disagreement. It improved mRNA transfection efficiency prediction by about 32% on average and its virtual screening rankings held up in wet-lab tests.
TRACE is an AI agent designed for molecular lead optimization in drug discovery, which uses trajectory-aware decision-making to select and sequence molecular optimization tools. The system demonstrated superior performance in optimizing ADMET properties compared to baseline models.
SMDD-Bench is a multi-turn benchmark of 502 solvable small-molecule drug design tasks across five task types and 102 protein targets, with a limited oracle call budget. Seven frontier LLM agents were tested, showing how far current planning and tool use fall short on real design problems.
DrugSAGE keeps a cross-task memory of verified skills, evidence about effective strategies and recurring errors so that it can reuse past searches when building drug discovery models. It ranked first among nine agents on 33 molecular property prediction tasks and scored 0.935 on 17 held-out tasks using accumulated memory without new search.
ToolMol is an evolutionary agentic framework that uses a multi-objective genetic algorithm and an LLM operator to iteratively generate drug-like molecules. It demonstrates superior performance in discovering ligands with higher predicted binding affinity and better Absolute Binding Free Energy scores compared to existing methods.
VibeProteinBench is a benchmark designed to evaluate the performance of language-interfaced models in protein design tasks, including recognition, engineering, and generation stages. The benchmark reveals that no current model excels across all three stages simultaneously, highlighting recurring failure patterns and the importance of tool integration for effective protein design.
BioDesignBench is a benchmark for evaluating large language model agents in protein design tasks, using expert-curated workflows and metrics. The study reveals that while the agents can select appropriate tools, their performance is limited by premature termination of candidate evaluation and lack of comprehensive metric coverage.
EvoSyn is a multi-agent framework that integrates large language model reasoning with domain expert knowledge to design molecules, addressing the challenge of hallucinations and chemical logic errors. The system demonstrates significant improvement over existing methods in generating bioactive and synthetically feasible molecules through extensive benchmark evaluations.
AgenticPosesRanker has GPT-5 reason over six deterministic structural analysis tools to rank protein-ligand docking poses. On a balanced benchmark of ten systems and 162 poses it matched the Smina scoring function at 50% best-pose accuracy, far above random, recovering some systems Smina failed on.
Open Intelligence Hub (OIH) is an autonomous platform that uses large language models to plan and execute multi-step workflows for designing protein binders and prioritizing antibody-drug conjugates. The system demonstrated high accuracy in classifying oncology targets and generating binders with improved scores compared to alternative methods, suggesting its potential utility in computational drug design.
RepurAgent is a hierarchical multi-agent system with supervisor, planning, research, prediction, data and report agents, episodic memory and human oversight, built on drug repurposing tools and procedures. In acute myeloid leukemia, blinded experts judged its candidates more novel and mechanistically credible than a plain LLM's, and it was also applied to experiment design and assay analysis.
ProtoCycle uses an LLM planner that calls protein design tools, reflects on their feedback and revises its plan over multiple rounds, trained with supervised trajectories and reinforcement learning. It improved alignment to natural-language functional requirements while keeping foldability, and ablations showed reflection drove much of the gain.
CACM is a drug discovery framework that uses precise set-level diagnosis and memory write-back to correct protocol violations, improving the success rate of language-based drug discovery agents. It demonstrates that reliable autonomous drug discovery benefits from more accurate failure localization and efficient planning context management.
DrugAgent is a multi-agent framework designed for virtual drug screening that ingests raw physicochemical property signals and performs closed-loop sensing and decision-making. It demonstrates superior performance in accuracy, interpretability, and robustness compared to existing models under various evaluation metrics.
Cirrina is an AI agent that uses large language models and pharmacology tools to evaluate preclinical CNS drug candidates, providing reasoning traces for decision-making. It demonstrated higher accuracy in identifying failure-prone drug candidates compared to traditional deterministic methods.
MDCrow is an agentic large language model assistant designed to automate molecular dynamics workflows for proteins, using a chain-of-thought approach with expert-designed tools. The system demonstrates robust performance across various tasks and complexity levels, showing the potential of LLMs in automating complex scientific processes.
The paper demonstrates the use of large language models (LLMs) to optimize and modernize computational chemistry tools, specifically improving the LEaP program in the AMBER molecular dynamics suite. The LLM was used to implement an optimized unit merge algorithm and 64-bit indexing, significantly reducing parametrization time for mid-sized systems and enabling the handling of multimillion-molecule systems.
Agent Rosetta is an AI agent designed to perform protein design tasks using the Rosetta software, demonstrating its capability in both canonical amino acid designs and non-canonical residue designs where traditional ML methods struggle. The study highlights the importance of environment design for integrating large language models with specialized scientific tools.
MAESD is a multi-agent framework that uses evolutionary algorithms and large language models to design protein sequences based on natural language instructions. It showed the ability to iteratively refine protein sequences under functional and structural constraints, reducing the need for expert-level biological inputs in traditional computational protein design.
A control plane enforces role-based tool isolation, constrained action spaces and reflective replanning, while a workflow plane runs drug discovery stages as stateful skill graphs with data contracts and human checkpoints. It outperformed baselines on orchestration accuracy on PharmaBench and was demonstrated in end-to-end therapeutic case studies.
Candidates for Alzheimer's drug repurposing from three computational predictors were pooled and an LLM agent synthesized literature evidence to filter them using explicit criteria. The prioritized list was checked against patient data, clinical trial registries and pharmacological reviews.
MAC-AMP is a multi-agent system designed for the creation of antimicrobial peptides (AMPs) that balances multiple objectives such as activity, toxicity, and novelty. The system uses an adaptive reinforcement learning framework to generate AMPs with optimized properties, demonstrating superior performance compared to existing models in terms of antibacterial activity, likelihood of being an AMP, compliance with toxicity standards, and structural reliability.
The paper evaluates six agentic systems for drug discovery across 15 task classes, identifying capability gaps such as lack of support for protein language models and peptide-specific predictions. It suggests that the limitations are architectural rather than due to knowledge deficits.
The paper introduces a modular framework using large language models (LLMs) to automate tasks in early-stage drug discovery, including data retrieval and molecular generation. The system demonstrated improved accuracy in answering biomedical questions and successfully guided iterative molecular refinement, increasing the number of molecules meeting drug-likeness criteria.
The system uses a knowledge-grounded framework to analyze drug safety cases, integrating disproportionality analysis and a novel execution planner that minimizes hallucination risk. It demonstrates how combining structured medical knowledge with statistical evidence can support reliable and explainable case assessments.
Conditions each molecular proposal on the task description, the full optimization trajectory and explainability-derived feedback about oracle scores. Across drug-discovery-relevant tasks it was more sample-efficient than existing optimizers and zero-shot LLM generation, and chemists can steer it in natural language.
PABLO is a hierarchical agentic system that uses large language models pretrained on chemistry and biology literature to iteratively refine biological candidates for tasks like molecular design and antimicrobial peptide optimization. It demonstrated state-of-the-art performance in sample efficiency and final objective values, with optimized peptides showing strong activity against drug-resistant pathogens.
ChemCRAFT is a framework that uses agentic reinforcement learning to enable small language models to interact with an external chemical sandbox for precise information retrieval, enhancing their ability in molecular design and synthesis tasks. Evaluations show ChemCRAFT outperforms cloud-based large language models in various drug design aspects while being cost-effective and privacy-preserving.
The authors analyzed how LLMs help or hurt Bayesian optimization on molecular optimization and four-residue protein binding-motif design. They propose an agentic workflow combining chemical tools with statistical BO for molecules, and found reasoning LLMs useful for hypothesis generation in the protein task.
Medea runs multi-step therapeutic analyses with biological tools, machine learning models and literature retrieval while verifying planning, execution and evidence synthesis so biological context is preserved. It was evaluated on 5,673 open-ended analyses covering cell-type-specific target nomination, synthetic lethality and related tasks.
ChatInvent is an agentic AI system developed by AstraZeneca to assist in molecular design and synthesis planning within their drug discovery pipeline. The paper highlights the evolution of ChatInvent from a single-agent proof-of-concept into a multi-agent architecture with a graphical user interface, emphasizing its integration challenges and future potential.
AutoBinder Agent is an agentic framework for end-to-end protein binder design that integrates multiple AI and bioinformatics tools under the coordination of a Large Language Model (LLM) using the Model Context Protocol (MCP). It demonstrated improved reproducibility, reduced manual overhead, and enhanced extensibility in drug design processes.
M^4olGen is a multi-stage molecular generation framework that uses a retrieval-augmented, fragment-level approach to generate molecules with precise numeric constraints over multiple physicochemical properties. The system demonstrates superior performance in generating valid and target-constrained molecules compared to existing large language models and graph-based algorithms.
PigGPA is an agent for porcine genomic breeding analyses that combines domain-constrained skills with hashed execution contracts checking that run parameters match declared ones. The authors showed that general agents were inconsistent across sessions and that skills improved scores from 73.44 to 89.06 out of 100.
CucurbitAgent is an AI research agent designed for functional genomics in cucurbits, integrating extensive multi-omics data and literature. It demonstrated superior performance compared to general-purpose LLMs on a benchmark curated by domain experts, showcasing its ability to provide accurate, complete, relevant, and traceable responses.
LLM agents with MCP-exposed tools locate published omics datasets, extract metadata, download data, run containerized quantification pipelines and synthesize results across studies. The system catalogued dataset references across thousands of PubMed Central articles and completed five end-to-end proteomics and RNA-seq reanalyses.
BioGAIP is an LLM-powered multi-agent system for automating bioinformatics tasks. No abstract was available, so details of its evaluation are not summarized here.
BaseCamp assigns six specialized agents to the judgment layer of DNA sequencing pipelines, choosing thresholds, configuring established tools, adjudicating borderline variants and flagging cases for review while the tools do the analysis. Evaluation found agent configurations concordant with expert practice and cross-stage checks that surfaced anomalies normal monitoring misses.
Agents survey literature, parse data, choose baselines, debate a model design with a critic, then write and run training code to return a trained perturbation predictor. On six transcriptomic, epigenomic and proteomic datasets its models were competitive with hand-built state of the art, and blinded raters preferred its research plans.
MErlin is a 17-module bacterial epigenomics toolkit also delivered as an agent skill that lets an LLM assistant select and run the audited modules with preflight checks and human-input rules. It was validated on synthetic data with planted signals and demonstrated on a Pseudoalteromonas methylome-transcriptome comparison.
Bioinfoysis treats each bioinformatics request as a persistent analysis run where a planner keeps an executable checklist and replans from structured handoffs after each worker step. A controlled runtime validates generated scripts, tables and figures before they are reused, binding conclusions to the evidence that produced them.
MetaboAnnotate is a web-based framework that uses a large language model to integrate and orchestrate multiple metabolite annotation tools for untargeted metabolomics, enabling users to submit MS/MS spectra without local installation or programming expertise. The system demonstrates improved accuracy in metabolite identification through consensus among integrated tools on benchmark datasets.
CyChat is a Cytoscape app that integrates a chat interface with an AI agent to facilitate no-code, reproducible network analysis. It was evaluated across various workflows and LLM providers, demonstrating high accuracy and efficiency compared to manual GUI workflows.
An LLM agent framework builds tissue-specific protein co-abundance networks, derives consensus clusters for tissue pairs and grounds them in atlases, interaction databases, pathways, disease catalogues and literature. Run over all 820 pairs of 41 human tissues and fluids it found 1,833 conserved clusters across 406 pairs.
GEOMeta is a multi-stage pipeline of task-specific LLM agents that retrieves, extracts, standardizes, ontology-maps and quality-checks GEO metadata. It annotated about 600,000 human bulk RNA-seq samples, and a comparison of 22 LLMs showed cheaper open models matching leading reasoning models.
HistAgent couples a visual-omics model trained on 2.23 million H&E-spatial transcriptomics spots with an agentic module that organizes inferred readouts into evidence cards for multi-turn analysis. It recovered literature-defined spatial findings and reproduced five standard spatial analyses at lower cost than conventional workflows.
ACCREDIT registers spatial omics images to H&E sections with deterministic pipelines, scores registration quality without a reference, and calls an LLM rescue agent to diagnose failures and pick recovery strategies. It treats registration as an iterative decision process that rejects plausible but incorrect alignments.
ASTA combines ensemble clustering, marker-based cell annotation and a GPT-4o-mini agent for autonomous spatial transcriptomics analysis. Clustering improved silhouette scores over single methods and the agent produced three biological hypotheses with validations on ovarian tissue.
The GrEBI system uses a multi-agent architecture to integrate and query multimodal biomedical data across different scales and organisms, grounded in biological knowledge from ontologies. It demonstrates the capability of large language model agents to perform cross-cutting queries on complex, heterogeneous datasets when provided with scientifically accurate grounding.
TCellAlign aligns study-specific T-cell population labels to standardized nomenclature using agents for literature retrieval, information extraction, nomenclature-guided alignment and evidence-based adjudication. The authors built a manually validated benchmark linking study labels, CZ CELLxGENE annotations and standard T-cell names across 44 curated studies.
MetaClaw maps a user request to registered metagenomic and multi-omics workflows, runs standardized upstream processing and study-specific downstream analyses in isolated containers, and records parameters and outputs per job. Across four published cohorts with 769 profiles it recovered most of the originally reported marker groups.
One hundred evaluations give agents only the raw sequencing data and context a human analyst would have, across seven task categories and several sequencing technologies, and grade structured answers deterministically. The best of sixteen model-harness pairs cleared only about half, with errors coming from choices of references, thresholds and normalization.
Guides a locally deployed small language model through single-cell analysis from raw counts to cluster annotation, planning steps, running standard tools, inspecting intermediate results and revising decisions. On GTEx it lifted an 8B model from 0.39 to 0.89 accuracy, near GPT-4o, and beat reference-, marker- and LLM-based annotators on three atlases.
Prompt-to-Paper is a multi-agent system that grounds manuscript claims in a retrieved corpus, runs real computational biology experiments through a coding agent, and scores drafts on eight quality dimensions. It is aimed at making generated bioinformatics papers verifiable rather than fabricated.
Tests AI scientist systems on cell type annotation across 15 expert-labeled datasets and on 193 discovery questions drawn from 41 published single-cell studies, with bioinformaticians as a human baseline. Current systems fell short of autonomous discovery but showed useful support for data-driven research.
IAN is an R package that runs pathway, regulatory and protein interaction enrichment analyses and has multiple LLM agents summarize and integrate the results into interpretations. It was demonstrated on human transcriptomic data, tested for robustness to LLM choice, and rated by human experts for clarity and relevance.
An LLM agent plans and executes isoform-resolved long-read single-cell RNA-seq workflows with stage-aware planning and persistent computational context. It recovered major published findings on real data and extended a differential transcript usage event into a sequence-based functional hypothesis.
The M3A framework lets LLM agents reason over persistent multimodal single-cell multi-omic data states in autonomous and human copilot modes. Using datasets from 11 cancer types, the authors assessed what current agents can and cannot do in open-ended biological discovery.
mLLMCelltype has several LLMs annotate single-cell clusters and then deliberate iteratively to reach a consensus label with a confidence score. Across 49 datasets it reached 77.2% mean accuracy, 15.7 points above the best single LLM.
PAG-Agent runs preprocessing, differential expression and pathway analyses on bulk and single-cell data, then interprets pathways in light of experimental context with literature support and writing help. In Alzheimer's disease case studies it consistently recovered neurodegeneration pathways and it beat six LLMs at citation retrieval.
Continuously scans NCBI submissions, parses host information with an LLM, predicts human infectivity with a fine-tuned long-context genome model, and then scores viral proteins against approved drugs for repurposing candidates. It was cross-validated on a curated set of over 33,000 complete viral genomes.
SpatialBench-Long asks agents to recover biological claims from raw spatial measurements without prescribed methods, with 24 evaluations across pancreatic cancer, glioblastoma organoids, lineage-traced lung tumors and aging optic nerve across many platforms. Claims were hardened by reproduction and scientist review and are graded deterministically.
AutoScientists is a decentralized team of agents that self-organize around hypotheses, critique proposals before spending compute and share failures across long-running computational experiments. On BioML-Bench it reached a 74.4% mean leaderboard percentile across 24 tasks and found an ACE2-Spike fitness prediction method 12.5% better than the prior best.
SpatialClaw is a memory-augmented autonomous ecosystem designed for spatial omics analysis, integrating 30 specialized skills to handle various aspects of data processing and analysis. It demonstrates superior performance in complex biological scenarios compared to standard large language models, showcasing its utility in dissecting intricate tumor microenvironments.
Claude Opus wrote variant-calling plans of increasing detail and six open-weight LLMs executed them as agents on desktop GPUs. qwen3.6:27b matched frontier accuracy on all plans and on an error-injection matrix, running on hardware under $2,000.
MethylCurate combines deterministic GEO retrieval, quality control and clock evaluation modules with LLM agents for metadata extraction, harmonization and data parsing in a dialogue-driven workflow. It turns user-specified DNA methylation datasets into standardized inputs for benchmarking epigenetic aging clocks.
Provides 244 expert-curated bioinformatics and data science tasks with reference answer files and structured file comparison for scoring agents. Biomni and ToolsGenie performed comparably, and all agents degraded sharply as task difficulty rose.
STAT is a multi-agent framework for conversational spatial transcriptomics analysis with a persistent session, an interactive tissue viewer and a staged skill-aware pipeline. Across eleven task categories on three platforms it outperformed a plain LLM and existing autonomous spatial analysis agents.
BioAgent is a LangGraph multi-agent workflow with a shared research state that coordinates literature retrieval, planning, data download from nine repositories, analysis, reporting and review with checkpointing and retry logic. It was evaluated on BRAF V600E melanoma, TP53 pan-cancer and PBMC single-cell case studies.
ROSIE combines deep-learning prediction of multiplex immunofluorescence from H&E slides with LLM agent logic that orchestrates biomarker inference and spatiotemporal reasoning. Applied to pancreatic neoplasia in transgenic mice, it profiled 10.44 million cells and identified three ordered immune-stromal states.
ORION is a multi-agent framework that combines statistics, machine learning and automated literature review to interpret PhIP-seq immune profiling data. On an APS-1 dataset it recovered the known autoantibody signature in about two hours, compared with one to two months of manual work.
STAnalyzer is an intelligent multi-agent framework designed to automate the analysis of spatial transcriptomics data, from raw data processing to hypothesis generation. It demonstrated its capability in dynamically translating natural language queries into bioinformatics workflows and ensuring analytical robustness through multimodal feedback and cross-validation with biological knowledge.
One hundred expert-curated tasks across genomics, transcriptomics, epigenomics, single-cell analysis, human genetics and machine learning workflows, built with scrambled metadata and synthetic data so each has one ground-truth answer. Leading coding agents reached 78 to 83 percent overall and 49 to 69 percent on the hardest questions from a bare environment.
Builds candidate-specific premises from tissue context and marker genes, then verifier agents turn them into reasoning trees and debate over bounded rounds before reaching a cell type label. With a local Qwen3-30B model it averaged 75.5 percent accuracy across five datasets, above chain-of-thought and scPilot baselines.
Describes an agentic system that reasons across biological datasets for drug target prioritisation, with LLM-guided evolutionary optimisation of scoring functions, plus an inference-time protocol that replaces gene names with codes to measure reliance on memorised knowledge. Blinding changed 16 percent of top-20 oncology targets while recovering the same validated targets.
OMICX-Agent is an AI-powered web platform that enables users to perform integrative analysis of multi-condition single-cell transcriptomics data through natural language interactions. The system demonstrates its utility in real-world case studies, showcasing its ability to integrate single-cell RNA-seq with bulk RNA-seq and GWAS summary statistics.
BioMedAgent is a self-evolving multi-agent LLM framework that learns to chain bioinformatics tools into executable workflows through exploration and memory retrieval. It reached a 77% success rate on the new 327-task BioMed-AQA benchmark and generalized to BixBench.
Recasts the LLM as a protocol-driven engine that loads per-database schema metadata files and separates entity resolution through REST APIs from schema-guided SPARQL generation over about 60 aggregated life-science databases. On 50 biologically grounded questions it improved substantially over an unaided LLM, with the schema files contributing most.
CellVoyager is an LLM agent that autonomously proposes and runs scRNA-seq analyses in Jupyter notebooks. On the 76-study CellBench it beat GPT-4o and o3-mini by up to 23% at predicting authors' analyses, and experts rated its new findings in three case studies as sound.
OmicClaw translates natural-language requests into traceable multi-omics workflows over the OmicVerse ecosystem using a registry-grounded, state-aware execution layer with iterative repair. It was evaluated on a benchmark of bulk, single-cell and spatial analysis tasks.
The paper presents a benchmarking system for evaluating AI agents in single-cell omics analysis, covering diverse tasks and metrics. It reveals that multi-agent frameworks outperform single-agent approaches by enhancing collaboration and execution efficiency through specialized role division.
The paper introduces 'vibe coding', an iterative process where AI agents generate executable code in real-time based on natural language prompts, significantly reducing the time and cost of developing custom data analysis platforms. It demonstrates this approach by creating a proteomics data analysis platform capable of performing standard tasks like normalization and differential expression testing within minutes.
Assesses image quality, retrieves reusable preprocessing skills and past experiences, plans a workflow with an LLM and refines it from label-free feedback, leaving deployed microbial colony detectors untouched. It raised mean detection mAP by 0.08 to 0.17 across three detector architectures on AGAR and an in-house dataset.
This audit examined how agents interact with 15 scientific tools in the ToolUniverse environment and catalogued cases where a call looks successful but returns incomplete data. Of 91 validated silent failures, most arose in the API or wrapper layer and could propagate into apparently valid scientific outputs.
Three cooperating agents simulate teaching dynamics to build a Chinese dataset of 6,400 clinical reasoning tutoring dialogues, used to fine-tune and reinforcement-train an instructing model. It beat eight baselines, exceeding GPT-4.1 by 10 to 15 points under LLM, expert and NLP evaluation, and was tested in a classroom.
Compares a plain LLM, single-agent, multi-agent and ReAct configurations that turn wearable data into daily workout plans on 50 matched user-days, scored by four blinded LLM judges on eight rubric criteria. Deterministic preprocessing before the LLM drove the gains, while differences among the scaffolded configurations were not significant.
P-Bench holds 425 open-ended hypothesis-testing tasks from economics, biology and medicine where an agent must choose a method, compute a p-value and conclude, and Fisher-R1 is an open-weight agent trained with reinforcement learning for this. Fisher-R1-14B improved substantially over its backbone and beat strong baselines, exposing inferential errors that benchmarks scoring only execution miss.
Defines a containerized task in which frontier agents, using only a bash shell, implement four attacks and compute a security score for clinical prediction models on breast cancer and MIMIC-IV data. Claude Sonnet 4.6 and GPT-4.1 completed all runs correctly, while GPT-4o finished 61 percent.
BioBrain is a multi-agent framework that turns natural-language goals into microscopy analysis pipelines assembled from validated methods and lab scripts, reporting every method and parameter. It reproduced expert results on TIRF and lattice light-sheet benchmarks, while frontier LLMs writing code made large quantitative errors.
Five subagents modeled on brain functional regions, with retrieval and character reasoning, jointly simulate standardized patients for medical students' communication practice. It outperformed single-LLM virtual patients on response quality, role-play, efficiency and perceived educational value.
StructAgent pairs a literature-grounded structural reasoning agent with an execution agent that runs local cryo-EM software, tracks state and recovers from failures, with expert approval gates for major changes. Case studies covered refitting a 64-chain proteasome, auditing 530 ribosomal metal sites and fitting an ambiguous ligand.
A tool in which an LLM agent takes natural-language goals and calls existing tabular synthesizers and privacy, utility and fairness metrics to build synthetic health datasets. The demo runs six synthesizers on a schizophrenia dataset with a locally hosted LLM orchestrating the workflow.
Encodes the regulatory order of clinical trial dataset derivation as a directed graph whose nodes are run by worker agents with pharmaverse context, validation gates and retries. On a benchmark built from the FDA CDISCPilot01 submission it matched all 49 ADSL variables in every run, where single-agent and flat multi-agent setups scored zero.
BenchGuard uses frontier LLMs to cross-audit the artifacts of execution-based agent benchmarks, optionally using agent solutions and traces as evidence. It found 12 author-confirmed issues in ScienceAgentBench and matched 83.3% of expert-identified problems on a BixBench subset for under USD 15 per 50-task audit.
Provides simulated EHR, payer portal and fax environments with 135 expert-defined prior authorization, appeals and equipment order tasks for computer-use agents. The best agent completed only 36.3 percent of tasks end to end despite high subtask success.
BloClaw is an agentic workspace for life-science research with a dual-track tool-call protocol, a sandbox that captures plots at runtime and a dynamic viewport interface. It was benchmarked on cheminformatics, ESMFold protein folding, docking and retrieval tasks, with far fewer serialization failures than JSON tool calling.
Turns sparse expert evaluations into global quality principles and procedural knowledge that guide an LLM agent editing medical texts, without training. Across three clinical text datasets and a real-world test it outperformed LLM baselines and memory-based adaptation methods.
BioTrouble is a multi-agent workflow that uses a troubleshooting knowledge base, small language models and cost-aware model routing to help fix failed PCR, qPCR, cloning and other experiments. Its recommendations were comparable to those of a single state-of-the-art LLM, and user feedback is stored to grow the knowledge base.
Generates ggplot2 code from natural language, fixes execution errors in a loop, and uses a vision-enabled agent to check the rendered figure's appearance for bioinformatics visualization. Against DeepSeek-V3 it raised code executability from 85 to 100 percent and scored higher on publication readiness.
Agents discuss memes from several aspects grounded in Cognitive Analytic Therapy competencies to identify depressive symptoms, alongside a new dataset with human-annotated explanations. It improved macro-F1 by 7.55 percent over more than 30 compared methods.
Generates patient cases from a disease knowledge graph and runs multi-turn consultations between an LLM doctor and an automated patient agent, scoring accuracy, questioning efficiency and rubric-based quality. It exposed clinical reasoning gaps in current LLMs that static question answering hides.
HARMON-E is a hierarchical agentic reasoning framework that uses large language models to extract structured clinical data from unstructured oncology notes. It demonstrated high accuracy in extracting over 100 oncology-specific variables, with critical variables achieving F1-scores above 0.95 on a dataset of 400,000 real-world clinical notes.
The paper presents an agentic AI framework for training general practitioner student skills, which includes evidence-based vignette generation, persona-driven patient dialogue, and standards-based assessment. The system was evaluated with medical students who reported realistic dialogues, appropriate difficulty levels, stable personality signals, useful feedback, and high usability.
The paper presents a multiagent AI framework for automating medical data analysis workflows, which outperforms nonagent-based approaches using large language models. The system successfully replicated primary outcomes from bone marrow transplantation studies with higher accuracy and fewer errors.
AIPatient is a simulated patient system powered by large language model-based AI agents and a knowledge graph, designed to enhance medical education through realistic clinical interactions. The study demonstrates AIPatient's high accuracy in electronic health record-based question answering, robustness, and positive user feedback from medical students.
MedEduChat is an EHR-integrated LLM agent designed to provide personalized prostate cancer education for patients. The study demonstrated that MedEduChat significantly enhances patient health confidence and receives high usability scores, indicating its potential in improving patient engagement and educational outcomes.
A router, planner, executor and manager architecture with speech input and retrieval from PubMed, trial registries and guideline sources delivers multilingual patient guidance in a kiosk. It is benchmarked on 500 clinician-verified simulated cases against GPT-4o, Claude 3 Opus, Mixtral and Gemma for accuracy, empathy and explainability.
Agents generate, validate and iteratively refine search strategies over a clinically verified hepatology knowledge graph in a retrieve-evaluate-refine loop before an LLM answers. On custom clinical questions it beat GPT-4, standard RAG and graph RAG on faithfulness, context recall and answer relevancy.
The Multi-Agent Medical Decision Consensus Matrix System uses seven specialized large language model agents to simulate realistic multidisciplinary team workflows in oncology, improving clinical decision-making quality and efficiency. The system demonstrates substantial gains over existing approaches with high accuracy and consensus rates, as well as strong ratings from expert reviewers for clinical appropriateness.
TxAgent is an AI system designed to assist in therapeutic decision-making by iteratively retrieving and executing biomedical functions. It demonstrated improved performance in the NeurIPS CURE-Bench competition through enhanced tool-retrieval strategies, highlighting the importance of accurate reasoning and tool usage for clinical decision support.
CP-Env is a simulated hospital environment that evaluates large language models on complex clinical pathways, including patient care transitions and multidisciplinary team interactions. The system reveals that most models struggle with pathway complexity and exhibit issues like hallucinations and loss of diagnostic details.
INFORM-CT is a framework that integrates large language models and vision-language models to automate the detection, classification, and reporting of incidental findings in abdominal CT scans. The system demonstrates improved accuracy and efficiency compared to existing approaches.
HopeAI is an agentic workflow-based system designed to provide evidence-based treatment recommendations in oncology. The study showed that HopeAI outperformed other LLMs and RAG systems in terms of accuracy, relevance, and comprehensiveness for clinical scenarios related to multiple myeloma.
MoMA is a multi-agent architecture that uses specialized language model agents to convert non-textual medical data into structured text, which is then aggregated and used for clinical predictions. The system demonstrated superior performance in clinical prediction tasks compared to existing methods using private datasets.
A hierarchical team of agents mirrors a multidisciplinary tumor board, interpreting endoscopy, radiology and lab data separately before combining them. Expert raters scored it 4.60 of 5, above a single multimodal model, especially for reasoning logic and medical accuracy.
CognAlign is a multi-agent system designed to enhance transparency and bias-awareness in medical triage using small language models. It demonstrated improved clinical accuracy, reduced errors by 24%, eliminated patient safety violations, and increased transparency compared to baseline systems.
MCP-AI is a protocol-driven framework for autonomous reasoning in healthcare that integrates Model Context Protocol (MCP) to enable secure, long-term state management and collaboration among AI agents. It was validated through clinical use cases involving Fragile X Syndrome with comorbid depression and remote coordination for Type 2 Diabetes and hypertension, demonstrating its capability to streamline clinical processes while ensuring regulatory compliance.
The Multimodal Oncology Agent predicts IDH1 mutation in low-grade glioma by combining a TITAN-based histology tool with reasoning over clinical and genomic inputs and searches of PubMed and OncoKB. On 488 TCGA-LGG patients it reached an F1 of 0.912 with histology fused in, above histology and clinical baselines.
CureAgent is a modular framework that uses specialized TxAgents for executing tasks and long-context foundation models for clinical reasoning, addressing the limitations of monolithic models. It demonstrates improved performance on CURE-Bench without requiring expensive end-to-end fine-tuning.
ClinTutor-R1 is a multi-agent system designed to facilitate one-to-many clinical education, addressing the challenge of balancing individual and collective learning. The system demonstrates improved performance over baseline models in maintaining instructional quality across diverse student groups through comprehensive evaluation methods.
The paper introduces a multi-agent system designed to address the unique challenges of neurological clinical reasoning, demonstrating its effectiveness through comprehensive benchmarking against various large language models. The system significantly outperforms individual LLMs and retrieval-enhanced models on complex neurological questions, showcasing the potential for structured AI approaches in specialized medical domains.
Radiologist Copilot is an AI system designed to assist radiologists in the complex task of creating accurate and standardized reports for medical images. It autonomously orchestrates various specialized tools to handle image localization, interpretation, template selection, report composition, and quality control, demonstrating superior performance compared to existing methods.
Agents are barred from changing positions and interact only through targeted verification of visual evidence, with one inquiry round to expose image-text mismatch. On four medical VQA benchmarks it reached 71.3 percent on PathVQA, six points above prior work, with 87.7 percent fewer tokens.
Sets up doctor agents advised by assistant agents, some of which collude adversarially, plus a verifier agent checking decisions against clinical guidelines. Collusion pushed harmful recommendation rates up to 100 percent, and the verifier restored full accuracy on 50 clinical questions.
The paper introduces a two-agent neuro-symbolic framework for forecasting HFMD cases, which integrates contextual information through an LLM-based Event Interpreter and generates probabilistic forecasts with a Forecast Generator. The system demonstrates competitive point accuracy and robust interval coverage while providing interpretable explanations of forecast changes.
The paper introduces a Byzantine fault-tolerant multi-agent system for healthcare that uses gossip protocols and cryptographic validation to ensure secure message propagation in adversarial environments. The system successfully maintains consensus accuracy even with up to 33% Byzantine nodes, demonstrating its capability to support collaborative medical decision-making securely.
PathReasoning is a multimodal reasoning agent designed to navigate whole-slide images for cancer diagnosis and prognosis, focusing on tumor microenvironment analysis. It demonstrates superior performance in selecting relevant regions of interest compared to existing methods, enhancing the accuracy of diagnostic reports.
MTBBench is a benchmark designed to evaluate the performance of multimodal large language models in simulating clinical decision-making processes within oncology, specifically focusing on Molecular Tumor Boards. The system highlights the limitations of current LLMs in handling complex, longitudinal, and multimodal data and demonstrates improvements through an agentic framework that enhances reasoning capabilities.
KOM is a multi-agent system that handles knee osteoarthritis evaluation, risk prediction and treatment prescription. It beat general LLMs on imaging analysis and prescription benchmarks, and a randomized simulation showed clinician-KOM collaboration reduced workload.
ClinPreAI is an autonomous AI system designed to predict early postpartum depression risk using multimodal electronic health record data. It demonstrated superior performance compared to traditional AutoML and commercial solutions, showcasing the potential of agentic AI in clinical settings for democratizing sophisticated predictive modeling.
A cloud benchmark of over 700,000 clinician-reviewed tasks across 24 specialties with separate tracks for LLMs, multimodal models and agents. Agents built on the same backbones scored a mean of 79.8 of 100, far above base LLMs at 54.1, with large gains on safety tasks.
GHAR is a hierarchical agentic retrieval framework designed to enhance healthcare predictions by addressing when and how to retrieve external knowledge. It demonstrates superior performance over existing methods in clinical decision support tasks, showcasing the potential of its dual-agent architecture.
The paper presents a multi-agent system that uses clinically validated flowcharts to guide patients through medical decision-making, ensuring accuracy and transparency. The system was evaluated using synthetic datasets and achieved high performance in both flowchart retrieval and navigation tasks.
The paper describes an AI agent framework for managing chronic gastritis that integrates large language models with retrieval-augmented generation and a search engine tool. The study demonstrates the superior performance of this AI agent in addressing complex clinical questions compared to standard LLMs, highlighting its potential for enhancing patient self-management and reducing physician workload.
EasyMED is a multi-agent framework designed to simulate standardized patients for medical education, offering stable and inquiry-conditioned patient behavior. The system was evaluated against human-standardized patients using SPBench, demonstrating superior performance in case consistency and controlled disclosure.
The paper evaluates the use of large language model agents integrated into electronic medical records to assess the appropriateness of blood cultures. The study highlights high sensitivity but low specificity in performance, influenced by prompt phrasing and other factors, indicating both potential benefits and limitations for clinical decision support systems.
Bio AI Agent is a multi-agent system designed to streamline CAR-T cell therapy development by integrating target discovery, toxicity prediction, and molecular design. The system demonstrated its capability in identifying high-risk targets and patent infringement risks, showcasing potential for accelerating the translation of immunotherapies from discovery to clinical use.
The Voice-Interactive Surgical Agent (VISA) is a hierarchical multi-agent system designed to assist surgeons in accessing and manipulating patient data during robotic surgery through voice commands. VISA demonstrates high accuracy and robustness in executing tasks such as retrieving clinical information and navigating 3D anatomical models, enhancing surgical workflow without interruptions.
Combines an LLM with a curated oncology trial knowledge base and retrieval to screen patients for trial eligibility with source-grounded explanations, sending uncertain cases to human reviewers. On 731 breast cancer patients it resolved 61.9 percent automatically, reached 98.6 percent accuracy, and cut review time per case from 20 minutes to 43 seconds.
Agents search embedded case reports and literature for evidence relevant to unfamiliar brain MRI findings and use it to reason about rare diagnoses without additional training. On the NOVA dataset of 280 rare diseases it improved several LLMs by up to 10.2 percent.
The MDAT framework uses multiple large language models to simulate a multidisciplinary tumor board, enhancing clinical decision-making in complex oncology cases. The study shows that fixed-size MDATs outperform single-agent approaches and are particularly effective for high-complexity questions.
OpenMedicine AI is an LLM-powered diagnostic agent that generates ranked differential diagnoses for complex medical cases. The system demonstrated superior accuracy and actionability compared to both AMIE and human physicians, significantly reducing diagnostic errors in high-complexity clinical scenarios.
The authors cleaned MedCalc-Bench, built a stepwise evaluation of formula selection, entity extraction and arithmetic, and an automatic error attribution checked against experts. Their MedRaC pipeline, combining retrieval with Python code execution, raised LLM accuracy on medical calculations from 16.35% to as much as 53.19% without fine-tuning.
MarIA is a GPT-3.5-powered conversational assistant designed to support patients with type 2 diabetes mellitus, demonstrating increased user engagement and richer contextual understanding through personalized interactions. The study highlights the importance of comprehensive patient profiling and safety mechanisms in AI-supported chronic care.
A multi-agent generative AI system processed clinical notes, neuropsychological tests and MRI to classify primary progressive aphasia by syndrome and underlying pathology. It was evaluated on 54 autopsy-confirmed cases at UCSF in open-ended and constrained diagnosis settings.
A fine-tuned multimodal LLM reassesses its own pressure ulcer stage prediction by reasoning over image features and encoded clinical knowledge before settling on an answer, with natural-language justifications. On a public pressure injury dataset it reached 85 percent accuracy, four points above earlier convolutional models, and was tested under live inference.
Trains an interactive diagnostic agent with multi-turn reinforcement learning inside a learned diagnostics world model built from electronic records, alongside a multi-center benchmark with physician rubrics. It beat 11 LLMs and two prompted agents, gaining 11.2 percent diagnostic accuracy and 17.6 points of examination recommendation F1.
AgentMD is a language-based system that curates and applies clinical calculators to predict medical risks, demonstrating superior accuracy compared to GPT-4. It also provides population-level risk insights for institutional management.
EBMChat is a large language model-based agent designed to support clinical decision-making by integrating Evidence-Based Medicine principles and contextual conversation capabilities. The system demonstrated superior performance in retrieving high-quality evidence and maintaining context during multi-turn dialogues compared to existing methods.
The study evaluates the ability of clinical agents to detect inconsistencies in patient identity across electronic health records. It finds that these agents often fail to notice such discrepancies, highlighting a critical risk for misbinding and emphasizing the need for explicit verification mechanisms before deployment.
TriAgent pairs automated biomarker discovery for emergency triage with a supervisor research agent that delegates literature queries to sub-agents for grounding and novelty checks. Candidate biomarkers are classified as already supported or novel with transparent justification for acute care risk stratification.
MedTrust-RAG requires every biomedical answer to cite retrieved documents and runs an iterative loop where a verification agent judges evidence sufficiency and rewrites queries after a medical gap analysis. It is designed to cut hallucinations caused by noisy retrieval in medical question answering.
MedKGEval is a framework that evaluates large language models in clinical settings through multi-turn patient interactions, using a knowledge graph to simulate realistic conversations. It demonstrates the ability to identify subtle flaws and safety risks in LLMs' responses by assessing each turn with fine-grained metrics.
MedARC is a multi-agent framework designed to enhance medical question answering by facilitating structured debates among LLM agents, improving accuracy and factual consistency. The system demonstrates significant performance gains over baseline methods in various medical QA tasks through automated benchmarks and human evaluations.
OpenPFx compared zero-shot, few-shot, multiple few-shot and agentic LLM workflows for writing patient-friendly explanations of incidental MRI findings. Over 407 outputs per workflow, the agentic workflow gave the highest ICD-10 accuracy and reached a sixth-grade reading level.
Answers text and image questions about myopia by calling an image classifier and retrieving from a knowledge base of literature, expert consensus and guidelines. It graded myopic maculopathy accurately and, in a randomized trial, raised patient satisfaction and education quality over printed leaflets.
The paper introduces a large language model (LLM) agent that performs automated radiotherapy treatment planning for IMRT without any prior training or fine-tuning. The LLM-generated plans were evaluated against clinical manual plans and showed comparable organ-at-risk sparing while improving hot spot control and conformity.
Confidence-aware doctor and pharmacist agents cross-check each other to connect diagnosis with medication decisions, evaluated on a new DrugCareQA benchmark. It reached 67.58 percent accuracy on both diagnosis and medication, about seven points above a single agent.
Worker agents read long EHR histories chunk by chunk, writing key events into a shared memory, and a manager agent uses the summaries and timeline to predict risk. In zero-shot one-year lung cancer prediction from five years of records it beat four classes of baselines.
Turns 1,762 physician-curated vignettes into interactive scenarios where agents must request history, examination findings and investigations before diagnosing. LLMs that diagnose well from full vignettes degraded sharply, showing anchoring, poor test ordering, premature closure and missed critical conditions.
Derives an expert-validated taxonomy of ten collaboration failures from 3,600 execution logs and applies automated probes to 14,400 cases across six multi-agent architectures and six datasets. Agents repeated their initial views in 98 percent of discussions, left unsupported observations in 16.6 percent of cases and showed authority bias in 28.8 percent.
The paper introduces an agentic system that allows natural language querying of TCGA clinical data, combining rule-based extraction and large language model capabilities. The system successfully handles a range of queries from simple patient lookups to complex cohort analyses across ten cancer types.
The paper presents a multi-agent framework called the 'Council of AIs' that enhances Large Language Models (LLMs) performance on knowledge assessments by facilitating collaborative decision-making among multiple GPT-4 instances. The Council achieved high consensus accuracy rates across different stages of the USMLE, demonstrating the potential for AI collaboration to improve consistency and correctness in clinical decision support.
The Cascade Framework is an LLM agent built on the Dify platform for hypertension education and clinical decision support. Adding it to GPT-4o, GPT-4o-mini and DeepSeek-V3 improved rated accuracy and comprehensiveness, and it was externally validated in patients with suspected hypertension.
An information collection agent normalizes clinical text and optional fundus or OCT images, and a diagnosis agent ensembles several LLMs and a CNN with uncertainty-aware fusion for neuro-ophthalmic screening. It was more robust than single-model baselines on a curated corpus.
The paper presents a multi-agent system that translates unstructured clinical narratives into structured imaging recommendations based on ACR Appropriateness Criteria. The system uses a fine-tuned dense retrieval model and LLM-based agents to achieve high accuracy in matching guidelines, demonstrating significant improvement over standalone models.
MedPAO is a protocol-driven agent designed to structure medical reports accurately by following established clinical protocols, ensuring verifiable reasoning and reducing hallucinations. The system demonstrated high reliability with an F1-score of 0.96 for concept categorization and received an average score of 4.52 out of 5 from expert radiologists and clinicians.
Doctor-R1 is an AI agent designed to excel in clinical decision-making and patient consultation, trained using a multi-agent interactive environment with a two-tiered reward system. It surpasses existing models on benchmarks like HealthBench and MAQuE, demonstrating superior performance in communication quality, user experience, and task accuracy.
GraphMed-LT is a system designed to enhance multi-turn medical conversations by maintaining a patient-specific graph memory that captures clinical evidence across conversation turns. It demonstrates improved performance in generating answerable follow-up questions and providing accurate final answers compared to existing methods.
LabscriptAI is an execution-aware agent framework that translates natural-language instructions into robot scripts for laboratory automation, addressing biosecurity and community standards. It demonstrated high accuracy in a liquid-handling benchmark and was used to screen sequences, characterize GFP designs, engineer enzymes with enhanced functionality, and prepare genetic parts for distribution.
CLARITY is an AI-driven platform that uses a hybrid architecture to facilitate patient-to-specialist routing and clinical consultations. It surpasses human-level performance in first-attempt routing precision and significantly reduces consultation duration compared to human interaction.
Three task-specific LLM agents extracted unstructured right heart catheterization variables across a 12-hospital network, cross-validated by two pulmonary hypertension experts, with machine learning filling missing pressure values. The resulting data supported automated detection of pulmonary hypertension.
EMR-AGENT is an automated framework that uses language models to dynamically extract and standardize clinical data from EMRs, replacing manual rule writing for cohort selection and feature extraction. The system demonstrates strong performance across multiple EMR databases, showing the potential of automating complex tasks previously requiring expert-driven design.
An LLM agent retrieves patient-specific structured and unstructured data, assesses the evidence iteratively and returns structured outcome labels. It was validated on treatment-detail retrieval and on labeling osteoradionecrosis and cancer recurrence in prostate and head-and-neck cohorts.
Provides 3,000 simulated patient agents with varied language, cognitive limits, emotions and reluctance to disclose, scored on task success, inquiry skill, dialogue quality, efficiency and patient experience. Leading LLMs left substantial room for improvement and were sensitive to realistic patient behavior.
MedMMV is a controllable multimodal multi-agent framework designed to enhance the reliability and verifiability of clinical reasoning. It stabilizes reasoning processes, grounds intermediate steps in structured evidence, and aggregates candidate paths to improve accuracy by up to 12.7% on medical benchmarks while ensuring truthfulness as confirmed by blind physician evaluations.
EHR-ChatQA is a benchmark designed to evaluate the performance of large language model-powered agents in handling Electronic Health Record data access, focusing on query clarification and value resolution. The study reveals that while top-performing agents achieve high accuracy rates under certain conditions, they struggle with consistency across multiple trials, highlighting the need for more robust agent development.
AMANDA is a training-free framework that uses large language model agents to augment medical knowledge for visual question answering tasks in low-resource settings. It demonstrates substantial improvements over existing models across multiple benchmarks, showing its effectiveness in both zero-shot and few-shot scenarios.
Pairs a small VLM with a SmolAgent orchestration layer that invokes tools, fuses speech and vision, builds scene graphs and uses hybrid retrieval for clinical video understanding in robotics. It was competitive with larger VLMs on Video-MME and more robust on a custom clinical dataset.
An LLM learns rules from tabular patient data by revising them after prediction errors and checking them against global data distributions. Across medical datasets, including a catheter-related thrombosis cohort, it matched or beat over 20 baselines while producing evidence-based explanations.
TissueLab is an agentic AI system designed for medical imaging analysis that integrates various tools and allows real-time expert feedback to refine analyses. It demonstrates state-of-the-art performance in clinical tasks, such as staging and prognosis, by continuously learning from clinicians.
An LLM agent interprets natural-language or structured requests and uses tool calls to retrieve and modify CT acquisition, reconstruction and post-processing protocol elements. It generated device-compatible protocol files and implemented requests faithfully, though validity suffered without a unified device API and on ambiguous requests.
MACD is a multi-agent framework that enables large language models to self-learn clinical knowledge through iterative consultations, improving diagnostic accuracy. The system demonstrates significant improvements in primary diagnostic accuracy and shows the potential for human-AI collaboration in clinical decision support.
The paper introduces an AI agent using a Virtual Patients Ensemble (VPE) approach to predict specific surgical complications from individual patient profiles, outperforming baseline methods. The system generates plausible variations of the original case to capture uncertainty and provides interpretable predictions for clinical decision-making.
The Medical AI Consensus framework uses a multi-agent reinforcement learning approach to generate and evaluate radiology reports, integrating large language models and vision models. It demonstrates the potential for fine-grained assessment of clinical reliability in automated report generation through iterative human-in-the-loop evaluation.
EHR-MCP is a framework that integrates large language models with hospital electronic health record databases to retrieve clinically relevant information. The system demonstrated near-perfect accuracy in simple tasks but faced challenges in more complex scenarios requiring time-dependent calculations.
KAMAC is a framework that uses large language models to dynamically form and expand expert teams for medical decision-making, addressing the limitations of static multi-agent systems. It demonstrates significant improvements over single-agent and advanced multi-agent methods in real-world medical benchmarks, particularly in complex clinical scenarios.
The paper presents a protocol for generating synthetic clinical documentation using multiple generative agents that are context-aware and role-specific, aiming to preserve privacy, linguistic accuracy, and information integrity. The authors developed an automated evaluation framework to compare synthetic EHRs with real ones, demonstrating the potential of this approach in advancing learning health systems.
Dr. CaBot is an AI system designed to emulate expert diagnostic reasoning by generating detailed case presentations from initial patient descriptions, demonstrating its ability to solve complex medical cases and outperform existing models in a benchmark based on real clinical scenarios. In evaluations, physicians were unable to distinguish between diagnoses generated by Dr. CaBot and those written by human experts.
Translates clinician instructions into guideline-constrained commands such as history retrieval, exam ordering, report writing, referral and treatment planning, implemented as Python, API and MCP tools. On 214 cases across 22 specialties it achieved high diagnostic accuracy and clinically sound exam requests.
Voice-based LLM agents systematically question patients and informants to collect narratives relevant to Alzheimer's disease and related dementias. In 30 adults with suspected dementia, agent-detected symptoms agreed well with blinded specialist interviews and users valued its patient questioning.
Routes each clinical premise-statement pair to one of four reasoning families and applies a specialised solver followed by explicit verification and targeted refinement. On a 200-item clinical trial NLI benchmark it raised mean accuracy from about 23 to 57 percent across four backbones.
FHIR-AgentBench is a benchmark designed to evaluate large language model agents on real-world clinical questions grounded in the HL7 FHIR standard, assessing their ability to retrieve and reason over complex health data. The benchmark highlights challenges such as intricate data retrieval from FHIR resources and reasoning difficulties, underscoring the need for robust LLMs in clinical applications.
MAVeRiC-AD is an agentic framework that dynamically selects the best model for classifying Alzheimer's disease from MRI images, improving robustness across different imaging protocols and demographics. It demonstrates enhanced performance by integrating a mixture-of-experts vision-language model with demographic information, achieving higher ROC-AUC scores compared to single-model approaches.
This paper presents BioLab, a multi agent system that integrates domain specific foundation models and computational tools to automate end to end biological research. BioLab outperforms leading large language models on biomedical reasoning benchmarks and autonomously executes complex pipelines, including de novo antibody design with experimental validation. The results demonstrate that AI native multi agent frameworks can tightly couple computation and experimentation to drive scalable scientific discovery.
CancerGUIDE is an AI system that uses large language models to generate guideline-concordant treatment trajectories for non-small cell lung cancer patients, reducing the time and expertise required. It demonstrates high correlation with expert-annotated benchmarks and includes a meta-classifier for verifying prediction accuracy.
The DeepSeek-R1 framework integrates radiomics, deep learning models, and large language model-based agents to generate personalized treatment recommendations for hepatocellular carcinoma. The system demonstrated improved prediction accuracy for critical pathological markers like microvascular invasion and capsule presence, with clinical relevance scores indicating its potential in precision oncology.
The Healthcare Agent Orchestrator (HAO) is a multi-agent system that uses large language models to generate comprehensive patient summaries for molecular tumor boards, aiding in clinical decision-making. It demonstrates high accuracy and completeness in capturing critical patient information through the TBFact evaluation framework.
Code Like Humans is a framework that uses large language models to map unstructured clinical notes to ICD-10 codes, supporting over 70K labels. It demonstrates superior performance in coding rare diagnosis codes compared to existing methods.
A multi-agent LLM system with critique, search, knowledge retrieval, medication verification and summarization agents suggests improvements to hospital order sets at Vanderbilt. Physicians rated 735 suggestions, and an LLM-as-a-judge filter was developed to align outputs with expert preferences.
RDguru is a conversational intelligent agent designed to assist in the diagnosis and management of rare diseases by integrating multiple knowledge sources and diagnostic tools. It demonstrates improved diagnostic accuracy compared to existing methods, as well as effective elicitation of relevant patient information through iterative questioning.
DR. INFO is an LLM-based medical assistant that uses the RAG architecture to provide situationally aware answers to clinical questions, outperforming other leading models on a complex benchmark called HealthBench. The system demonstrates strengths in communication and accuracy but shows areas for improvement in context awareness and response completeness.
A manager agent assembles specialist agents that debate hierarchically over the subjective and objective sections of SOAP notes to identify clinical problems. On 420 MIMIC-III notes it beat a single agent for heart failure, acute kidney injury and sepsis, though debates sometimes showed groupthink.
The Personal Health Agent (PHA) is a multi-agent system designed to provide personalized health recommendations by analyzing multimodal data from consumer wellness devices and personal health records. The PHA framework includes sub-agents for data analysis, expert insights generation, and user guidance, which were evaluated through both automated benchmarks and human-in-the-loop assessments involving extensive annotations and expert reviews.
Combines a safety taxonomy of clinical scenarios and failure modes, an LLM hazard-detecting evaluator and a simulated patient agent to stress-test clinical dialogue agents. Its evaluator matched expert hazard detection with an F1 of 0.96 and the framework benchmarked five agents over 2,100 dialogues across 14 hazard scenarios.
An EHR agent estimates confidence at each reasoning step so it can answer or abstain, evaluated with a new metric of accuracy under hallucination control. On MIMIC-III and eICU it improved accuracy at a 70 percent reliability threshold by 44 and 25 points where baselines failed.
AT-CXR is an uncertainty-aware agent designed for chest X-ray triage, which decides when to issue a decision or defer to human intervention based on per-case confidence and distributional fit. The system outperforms existing models in terms of accuracy and selective prediction performance while maintaining lower latency suitable for clinical use.
A voice interface uses collaborating LLM agents to run MRI scanner functions such as sequence execution, parameter changes and table positioning during interventions. In 90 tests of 18 tasks in a real scanner room it completed 93.3% of tasks with 5-10 second response times on a laptop.
Specialised LLM agents extract Alzheimer's-relevant signs and symptoms from longitudinal EHR notes and evaluate them by domain before a combined risk judgement. Predicting risk ten years before the first diagnosis code, it reached 0.53 accuracy versus 0.26 to 0.45 for single-model baselines.
The study evaluates the performance of a multi-agent system versus a single agent in handling clinical-scale workloads, demonstrating that the multi-agent setup maintains higher accuracy and efficiency under load. The research highlights the benefits of task isolation and reduced context interference in large language models for clinical applications.
JAMA Clinical Challenge cases were converted into multi-turn dialogues in which a clinical LLM questions a patient agent and a system agent holding exam, lab and imaging data. Diagnostic accuracy of O1, GPT-4o, LLaMA-3-70B and a DeepSeek distillate dropped in this conversational format compared with the original vignettes.
Trains an LLM end to end with reinforcement learning to query guideline-derived disease profiles, patient records and biomedical literature, evaluate evidence and refine searches until the diagnosis is supported. It beat GPT-4o, DeepSeek-R1 and medical frameworks by 22.7 percent on average and raised physicians' accuracy from 45.6 to 69.1 percent on 150 real cases.
MedResearcher-R1 is a medical deep research agent designed to address the limitations of general-purpose LLMs in clinical reasoning tasks. It uses a knowledge-informed trajectory synthesis framework and integrates specialized retrieval tools, demonstrating superior performance on complex medical benchmarks.
The paper introduces a framework that enhances large language model-based virtual patients for medical training by enabling non-textual interactions and providing real-time feedback. The system was evaluated with medical experts who found the simulated patient behaviors natural and coherent, and the feedback from virtual tutors useful and appropriate.
The EMRC framework uses multiple LLMs to enhance medical decision-making by dynamically selecting the best agents based on their expertise and integrating their responses through confidence scores and adversarial validation. The system demonstrates superior diagnostic performance compared to single-LLM approaches, showcasing its effectiveness in leveraging diverse expert capabilities.
This paper introduces RAG-HPO, a retrieval-augmented generation tool that improves automatic extraction and assignment of Human Phenotype Ontology (HPO) terms from clinical text. By combining LLM-based phrase extraction with real-time retrieval from a large vector database of phenotype–HPO mappings, RAG-HPO reduces hallucinations and avoids costly fine-tuning. Evaluated on 112 rare-disease case reports, it significantly outperforms existing HPO extraction tools in precision, recall, and F1 score, with most false positives being clinically relevant ancestor terms rather than hallucinations. The results demonstrate that RAG-based grounding enables accurate, reliable phenotype extraction for clinical genomics and rare disease diagnosis.
ADAM-1 is a multi-agent LLM framework that integrates microbiome profiles, clinical data and literature to classify Alzheimer's disease. Compared with XGBoost, it achieved a higher mean F1 score with lower variance.
CardAIc-Agents is a multimodal framework that uses adaptive reasoning and external tools to support cardiac care tasks, generating task-aware plans and delivering decisions based on clinical data. The system demonstrated efficiency in handling complex cardiac cases compared to existing vision-language models and agentic systems.
The system uses a consortium of fine-tuned vision-language models and an OpenAI reasoning LLM to provide decision support for the analysis of H-reflex EMG waveforms, demonstrating improved accuracy and standardization in neuromuscular diagnostics compared to traditional methods.
Gregory is a multi-agent system designed to assist in neurological diagnosis, demonstrating superior diagnostic accuracy and efficiency compared to human neurologists and base LLMs. The DiagnosticXchange platform used for evaluation revealed that Gregory employs targeted testing strategies, significantly reducing costs and time without compromising thoroughness.
This paper introduces CART-GPT, a transformer-based model trained on over one million CAR-T single-cell RNA-seq profiles to predict patient-specific treatment response and neurotoxicity risk in CAR-T therapy. The model achieves strong predictive performance while providing interpretable insights showing that outcomes are driven by coordinated subsets of T cell states rather than single cell types. By linking single-cell predictions to patient-level outcomes and releasing a large annotated atlas, the study demonstrates the value of foundation models for precision CAR-T treatment planning.
Implements shared mental models, leadership, trust networks and mutual monitoring as coordination mechanisms so small language models can do multi-step clinical reasoning. Across eight medical benchmarks it matched the accuracy of MDAgents, MedAgents, DyLAN and ReConcile at one to two orders of magnitude lower token cost, with the lowest cross-dataset variance.
An LLM mediator lets several open-source general and medical VLM agents exchange and reflect on each other's answers to medical image questions. On five medical VQA benchmarks the collaboration outperformed every individual model without any training.
The paper evaluates large language models as dynamic treatment planners for insulin dosing in Type 1 diabetes, showing that smaller LLMs can match or outperform small neural network-based agents with zero-shot prompts. However, the study also highlights significant limitations of LLMs in handling complex clinical scenarios and underscores the need for careful prompt engineering and hybrid approaches.
Tree-of-Reasoning is a multi-agent framework designed to enhance complex medical diagnosis by recording reasoning paths and clinical evidence in a tree structure, improving the consistency of decision-making through cross-validation. The system outperforms existing methods on real-world medical data, demonstrating its effectiveness in handling intricate diagnostic tasks.
ICARE is an evaluation framework for radiology report generation that uses large language model agents to generate clinically meaningful questions and assess the consistency of findings in generated reports. The system demonstrates significant alignment with expert judgment, providing a transparent and interpretable assessment method.
SNOW is a multi-agent language model system designed to generate clinical features from unstructured EHR notes, mimicking the iterative reasoning and validation workflow of human clinicians. The study shows that SNOW can achieve performance comparable to manual feature generation by experts and outperform other methods in predicting cancer recurrence and heart failure mortality.
This study assesses how well large language models can generate clinically accurate pharmacogenomic (PGx) recommendations aligned with CPIC guidelines. Using 599 curated gene–drug–phenotype scenarios and a novel expert-validated semantic metric (LLM Score), the authors show that general-purpose LLMs often produce incomplete or unsafe recommendations. In contrast, a domain-adapted, fine-tuned model achieves substantially better accuracy (LLM Score = 0.92) and faster inference. The results emphasize that fine-tuning and structured prompting are more critical than model scale alone for safe, effective AI-driven personalized medicine.
The paper describes the development and evaluation of an agentic retrieval augmented generation (ARAG) framework using open-source large language models for generating evidence-based Arabic patient education materials. The ARAG framework significantly improved the quality of generated patient education materials, especially when used with larger, context-aware models like AceGPT-v2-32B.
This Q4Dx benchmark uses synthetic patient narratives and simulated physician-patient dialogues to measure how efficiently LLMs choose diagnostic questions. It scores whether models reach the right diagnosis while keeping questions few and short.
GPT-RadPlan is an automated radiotherapy treatment planning system that uses GPT-4Vision to iteratively refine treatment plans based on textual feedback, aiming to balance multiple objectives. It demonstrated superior performance compared to human-generated clinical plans in prostate and head & neck cancer cases, showing better target coverage and reduced organ-at-risk doses.
ReasonAgent combines a multimodal image analysis module, knowledge retrieval and a diagnostic reasoning module for ophthalmic diagnosis and treatment planning. On 30 real cases its diagnostic accuracy was comparable to residents and its treatment plans were better than GPT-4o and residents.
P-MedAgent is a specialized AI agent designed for medical data analysis tasks, transforming them into code generation problems to predict and classify outcomes. The system demonstrates high accuracy in various medical diagnostic tasks, achieving performance levels comparable to human experts.
A Qwen-32B agent calls segmentation, counterfactual image generation and evaluation tools to analyse, explain and test hypotheses about medical images interactively. The work demonstrates the system as an explainability agent for chest imaging rather than reporting a large benchmark.
Doctronic is a multi-agent AI system designed to autonomously function as an AI doctor in urgent care settings. The study shows that Doctronic achieves high diagnostic and treatment plan concordance with human clinicians, indicating its potential to address healthcare workforce shortages.
Transplant-Agents combines LLM agents with machine learning in rule-governed iterative dialogue to select reproducible biomarkers and predict transplant rejection risk. On 683 kidney, liver and heart transplant patients from ImmPort it reached AUROCs of 0.93, 0.88 and 0.88.
Infherno is an end-to-end framework that uses large language model agents to convert free-form clinical notes into structured FHIR resources, enhancing interoperability in healthcare. The system demonstrates strong performance on synthetic and real clinical datasets compared to human baselines.
Dr. Copilot is a multi-agent system designed to improve the quality of written medical advice given by Romanian-speaking doctors, providing feedback on 17 communication-related axes. The system has been deployed in real-world telemedicine settings and shown measurable improvements in user reviews and response quality.
The study evaluates the use of a Large Language Model (LLM)-based multi-agent system for safer therapy recommendations in patients with multimorbidity, comparing its performance to single-agent approaches and real-world benchmarks. The results indicate that while LLM-based systems can provide correct recommendations addressing clinical goals, they often produce incomplete advice or unnecessary medications leading to potential conflicts.
DocCHA is a modular framework that uses large language models to assist in clinical diagnosis through iterative dialogue, improving accuracy and symptom recall. It outperforms existing prompting-based LLMs on real-world datasets by refining reasoning links and prioritizing informative clarifications.
MedScrubCrew is a medical multi-agent framework that automates appointment scheduling and triage classification in healthcare settings. The system integrates the Gale-Shapley stable matching algorithm, knowledge graphs, and large language model-based agents to enhance operational efficiency and contextual relevance.
Compares seven vision-language models under baseline, fine-tuned, multi-model consultation and literature-retrieval configurations for telemedicine dermatology questions. Fine-tuning often hurt, while the consultation and retrieval architectures reached up to 70 percent accuracy and held up on unseen data.
The paper introduces openCHA, an open-source framework for developing conversational health agents (CHAs) that leverage large language models. The system demonstrates its effectiveness in various healthcare tasks such as diabetic patient management and mental health chatbot evaluation, outperforming other LLMs like GPT-4.
A multilingual question-answering system applies agent-based multi-step retrieval over national and WHO vaccination guidelines and is delivered as a point-of-care mobile app. Agentic retrieval outperformed standard RAG, especially on multi-step or ambiguous questions.
KERAP is a multi-agent system that uses knowledge graphs to enhance large language models for accurate zero-shot medical diagnosis prediction. It demonstrates improved diagnostic reliability and scalability compared to traditional LLMs by iteratively refining predictions through structured knowledge extraction.
The LLM-Powered Agent is designed to analyze physiological time-series data, specifically for heart rate estimation from PPG signals. It demonstrates superior performance compared to benchmark models by integrating user interaction and analytical tools, achieving lower error rates in HR estimation.
The paper presents an automated radiotherapy treatment planning system that uses a reinforcement-learning framework with GPT-4V to optimize beam angles for better dose distribution. The study demonstrates the system's superior performance in prostate cancer cases compared to random selection and deep reinforcement learning baselines.
AI-HOPE is an AI-driven conversational agent that integrates clinical and genomic data to support precision medicine research through natural language interactions. It demonstrated its utility by identifying significant genetic associations in colorectal cancer data, aligning with established literature.
LA-CDM is a language agent designed to support clinical decision-making by iteratively requesting and interpreting relevant tests to converge towards an accurate diagnosis. The system demonstrates improved diagnostic performance and efficiency through a hybrid training paradigm that combines supervised and reinforcement learning.
Three agents consult reference tables and a long-context neuromuscular textbook to integrate patient history, complaints and electrodiagnostic findings into standardized reports. It classified tests as normal or abnormal with 92.2 percent accuracy versus 62.6 percent for one-shot LLM inference.
The paper presents a pre-screening pipeline for clinical trials involving complex liver diseases, using large language models to interpret semantically complex criteria from electronic health records. The system demonstrates high precision and good recall in patient recruitment, with different pathways excelling in specific aspects of reasoning and data extraction.
Tiered Agentic Oversight (TAO) is a hierarchical multi-agent system designed to enhance AI safety in clinical settings by routing tasks to specialized agents based on complexity. The system was shown to outperform single-agent and other multi-agent systems across multiple healthcare safety benchmarks, with human validation confirming its effectiveness in improving medical triage accuracy.
This study evaluates GPT-4 for automated therapy recommendations by comparing its suggestions with multidisciplinary tumor board decisions for gynecologic and breast cancer cases. GPT-4 generated coherent and clinically reasonable recommendations, though limitations in accuracy and completeness remained, particularly in surgical and systemic treatment choices. Iterative prompting substantially improved concordance with expert decisions, indicating that GPT-4 may serve as a supportive tool in clinical decision making rather than a replacement for physicians.
The MCP-FHIR framework integrates large language models with HL7 FHIR data to enhance clinical decision support and EHR insights, enabling real-time summarization and personalized communication. It is evaluated using synthetic EHR data from SMART Health IT sandbox for privacy and reproducibility.
This randomized trial examined factors associated with opening patient portal messages about cancer genetic services and initiating pretest genetic care. Engagement was strongly driven by prior interaction with the healthcare system, such as frequent portal use and primary care visits, while sociodemographic and clinical characteristics showed no significant effects. The findings suggest that improving access to healthcare and digital tools may be key to increasing uptake of cancer genetic services.
NordDRG-AI-Benchmark is a public benchmark designed to evaluate large language models' ability to perform Diagnosis-Related Group (DRG) reasoning, which is crucial for clinical decision support and hospital funding. The benchmark includes logic and grouper tasks that assess the models' capability in DRG coding and full grouper emulation with strict exact-match scoring.
A director agent coordinates glaucoma detection models and several role-specific LLM agents to produce diagnostic reports through an interface for clinical review and teaching. The authors report that the multi-agent design improves reliability and reduces hallucination compared with a single generalist agent.
HEAL-KGGen routes genetic biomarker questions from a general practitioner agent to genomics, transcriptomics, proteomics and clinical specialist agents grounded in an automatically built ontology-aligned knowledge graph. It improved question-answering accuracy across several LLMs, by up to 43.75% for Claude 3.5 Sonnet.
A GPT-4 agent combines vision transformers for MSI and KRAS/BRAF detection, MedSAM segmentation and searches of OncoKB, PubMed and Google for oncology decisions. On 20 multimodal cases it selected tools correctly 87.5% of the time and raised decision accuracy from 30.3% for GPT-4 alone to 87.2%.
ChatMyopia is an AI agent designed to provide personalized education and information about myopia, integrating image classification and a knowledge base. It demonstrated high accuracy and patient satisfaction in clinical settings compared to traditional educational materials.
AutoCT is a framework that uses large language models to predict clinical trial outcomes, combining reasoning capabilities with explainability through iterative refinement. It demonstrates superior or comparable performance to state-of-the-art methods while maintaining interpretability and cost-efficiency.
The paper describes a Patient Simulator that uses real-world EHR data to create realistic patient scenarios for training and testing AI agents in clinical decision support. The simulator was validated by expert clinicians who found it consistent with actual patient encounters, demonstrating its potential as an evaluation tool for conversational AI agents.
Agents analyse clinical assessment items in detail, retrieve similar cases and pass conclusions through a structured verification step to diagnose mood disorders, alongside a released synthetic dataset of 1,173 cases. Its baseline agent beat GPT-4o by 12.3 percent on real cases and the full multi-agent system improved further.
This paper introduces Biomni, a general purpose biomedical AI agent designed to autonomously execute complex research tasks across diverse biomedical domains. By integrating large language model reasoning with retrieval augmented planning and code based execution, Biomni demonstrates strong generalization across tasks such as gene prioritization, drug repurposing, and rare disease diagnosis without task specific tuning. The results suggest that agentic AI systems can substantially enhance biomedical research productivity and support human scientists in complex discovery workflows.
The CHATCBD system uses blockchain technology to enhance the reliability of large language models in personalized healthcare, addressing issues like hallucinations and ensuring patient safety. It includes a controllable AI agent framework and a self-audit mechanism that improves decision-making and audit professionalism.
Trains a triage agent to route cases to specialty agents and an attending agent to integrate their opinions, using curriculum-guided reinforcement learning on Qwen2.5-VL. On five medical VQA benchmarks it gained 23.6 percent on average over strong baselines.
An agent-based architecture plugs in domain-specific tools and keeps traceable intermediate reasoning for diagnosis across modalities. It reached 93.26 percent accuracy on Alzheimer's diagnosis, over four points above the prior best, and was also evaluated on chest X-rays and multimodal question answering.
The Consensus Mechanism is a clinical decision support system that uses an ensemble of specialized medical expert agents to improve accuracy and robustness. It demonstrated superior performance compared to single-model systems across multiple benchmarks, including differential diagnosis tasks.
CDR-Agent is a large language model-based system that autonomously selects and applies the most appropriate clinical decision rules based on unstructured clinical notes to enhance emergency department decision-making. The system demonstrated significant improvements in accuracy and efficiency compared to standalone LLMs, reducing unnecessary medical interventions while effectively identifying positively diagnosed cases.
The Virtual Patient Chatbot integrates large language model-based agents to simulate patient interactions for clinical anamnesis training, offering a flexible and scalable solution. Evaluations demonstrated its effectiveness in providing consistent and plausible responses, with high usability scores indicating its potential as a valuable educational tool.
Adds a role-specialised agent that challenges an emerging diagnosis, modulating how strongly it intervenes by case complexity and calibrating its tone. Across nine medical question answering and three medical VQA benchmarks it beat single- and multi-agent frameworks including GPT-4o and DeepSeek-R1.
FUAS-Agents is an autonomous system that uses large language models to assist in treatment planning for focused ultrasound ablation surgery, integrating patient data and MRI images with specialized medical tools. The system was evaluated by clinical experts who rated its generated plans highly in terms of completeness, accuracy, fluency, and compliance.
MRGAgents is a multi-agent framework designed to improve medical report generation by fine-tuning specialized agents for different disease categories, addressing the limitations of existing models that often overlook critical abnormalities. The system demonstrates improved comprehensiveness and diagnostic utility in generated reports compared to state-of-the-art methods.
Over 1,000 human-curated questions require agents to retrieve and reconcile multi-hop facts from live medical knowledge bases on trials, studies, regulation and cost. Frontier agentic systems scored as low as ten percent on some subsets.
Compares multi-agent collaboration with single LLMs and conventional models on medical QA and VQA, lay summaries, EHR prediction and clinical workflow automation. Multi-agent setups helped task completeness in workflow automation but did not consistently beat strong single LLMs or specialized conventional methods.
Task-specific agents handle retrieval, draft writing, visual analysis, refinement and synthesis in sequence to generate radiology reports. It outperformed a strong baseline on automatic metrics and LLM-based evaluation.
SAMIRA is an AI agent designed for medical image segmentation in virtual reality, assisting users with tasks such as localization and visualization of 3D medical concepts through speech-based interaction. The system was evaluated in a user study where it demonstrated high usability and potential for training radiologists in complex segmentation tasks.
This paper introduces SleepBert, a retrieval augmented model that integrates structured polysomnography features with unstructured clinical narratives for comprehensive sleep disorder analysis. By fine tuning ClinicalBERT and incorporating real time medical literature retrieval, SleepBert achieves substantially higher diagnostic accuracy than existing baselines. The system supports scalable and evidence based interpretation of complex sleep data to assist clinical decision making and research.
Tests GPT-4o on 1,062 Chilean licensure questions under single-agent prompting strategies and multi-agent strategies including voting, MedAgents and MDAgents. MDAgents scored highest at about 90 percent, though simpler strategies sufficed for many questions.
This study evaluates whether large language models (Llama-2-70B-chat and GPT-3.5) can generate plausible clinical vignettes, clinician–patient dialogues, and management plans for pediatric and adult presentations across 282 genetic conditions. Clinician grading showed that both models generally produced age-appropriate descriptions with good correctness and completeness for disease presentation, including for neonatal-onset metabolic disorders. However, both models performed substantially worse when generating management plans, with only moderate correctness and completeness. These findings suggest that while LLMs can capture age-dependent clinical features, their reliability for clinical decision-making and management remains limited.
mAIstro is an open-source, multi-agentic system that automates the end-to-end development of medical AI models using a natural language interface. It successfully executed various tasks across diverse datasets and imaging modalities, demonstrating its capability to unify data analysis, model development, and inference in healthcare applications.
The paper presents a framework for the automated design of medical agent architectures using large language models, aiming to improve diagnostic accuracy in clinical settings. The proposed method demonstrates effectiveness in evolving workflow structures and enhancing diagnostic performance through iterative self-improvement guided by feedback.
MD2GPS is a multi-agent debate system that uses language models to enhance the accuracy of diagnosing Mendelian diseases by integrating data-driven and knowledge-driven approaches. The system was evaluated on multiple datasets, demonstrating improved diagnostic accuracy and reduced time for identifying potential pathogenic genes.
This review summarizes how artificial intelligence is increasingly applied to the diagnosis and management of infective endocarditis, a clinically complex and high-morbidity disease. Machine learning models and AI-enhanced imaging and microbiological techniques show improved accuracy in diagnosis, risk stratification, pathogen identification, and outcome prediction compared with traditional approaches. Despite these promising results, challenges remain in data availability, interpretability, ethics, and clinical validation. The authors conclude that carefully integrated and validated AI, including generative AI as decision-support tools, could substantially improve patient outcomes and safety in infective endocarditis care.
AI-HOPE-PM is a conversational AI platform that integrates genomic, clinical, and social determinants of health data to enable users to conduct multi-dimensional cancer analyses through natural language interaction. The system successfully demonstrated its capability by identifying significant differences in survival outcomes influenced by genetic mutations and socio-economic factors.
MedAgentSim is a simulated clinical environment that evaluates and enhances large language models' performance in dynamic diagnostic settings through multi-agent interactions. It demonstrates the effectiveness of its approach by showing improved diagnostic strategies and context-aware interaction abilities in various simulated scenarios.
MedAgent-Pro is a hierarchical reasoning system that uses multi-modal inputs and clinical criteria to generate reliable diagnostic plans for medical conditions. It demonstrates superior performance in both 2D and 3D medical diagnosis tasks, providing precise and explainable diagnoses.
DEMENTIA-PLAN is a retrieval-augmented generation framework that uses large language models and multiple knowledge graphs to provide conversational support for dementia patients, addressing memory loss and emotional instability. The system includes a self-reflection planning agent that coordinates knowledge retrieval and integration across different types of knowledge graphs to optimize response generation.
This study designs and evaluates a GPT-4 powered chatbot to support the return of positive population wide genomic screening results in a large public health program. Expert evaluation shows that the chatbot provides clear, user friendly, and contextually appropriate communication, performing well in tone, usability, and boundary management, though program specific accuracy requires improvement. The findings demonstrate the feasibility of LLM based chatbots to augment genetic counseling services and improve the accessibility of genomic health care.
GPT-Plan uses GPT-4 agents acting as dosimetrist and physicist to iteratively adjust and evaluate radiotherapy plans, supported by retrieval of historical plans. On 12 lung IMRT and 5 cervical VMAT cases its plans were comparable to or better than ECHO auto-planning and a senior physicist.
A locally hosted LLaMA agent adjusts optimization objectives inside a commercial treatment planning system over repeated iterations, using retrieval of past attempts and reinforcement learning. On 18 prostate cancer patients the 70B model scored 16.4 percent higher than the 8B model and retrieval improved scores by 19.8 percent.
The Consultation Flow Multi-Agent Framework uses a multi-agent system to simulate clinical consultations, integrating multiple modalities of medical data for dynamic diagnosis. The framework demonstrates improved diagnostic accuracy and decision-making persistence compared to existing methods on a public benchmark.
The study evaluates the performance of large language models in performing medical calculations and finds that they are prone to errors. It demonstrates that integrating these models with task-specific calculation tools significantly improves their accuracy, suggesting a potential pathway for their safe integration into clinical workflows.
The MMA system is designed to integrate multi-source medical data for answering questions related to medical diagnosis, enhancing clinical decision-making. The system uses a multi-agent architecture tailored to handle various modalities of medical data and demonstrates improved performance in handling complex diagnostic tasks compared to existing methods.
Agentographer is a CT operation agent designed to autonomously perform chest CT lung cancer screenings, reducing the workload and error risk for radiographers. In a clinical trial involving 500 subjects, Agentographer successfully completed 95.2% of cases independently, demonstrating significant improvements in efficiency and radiation dose reduction.
TxAgent is an AI agent designed to generate personalized treatment recommendations by leveraging multi-step reasoning and real-time biomedical knowledge retrieval across a wide range of tools. It outperforms existing models in drug reasoning tasks, achieving high accuracy while ensuring that treatment recommendations align with clinical guidelines.
The MAC framework uses multiple conversational agents to enhance disease diagnosis in complex medical scenarios, outperforming single LLMs and other methods. It demonstrates higher diagnostic accuracy and consistency through evaluations on rare disease cases.
SurgRAW is a multi-agent system designed for zero-shot reasoning in robotic surgical video analysis, addressing the need for unified scene understanding and interpretability. The system demonstrates superior performance compared to existing vision-language models and supervised approaches through extensive experiments on SurgCoTBench.
This study evaluates four generative AI models for answering rare disease and genetic counseling questions, focusing on accuracy and safety. While all models generally provided reliable and professional information, performance varied, with ChatGPT scoring highest and Perplexity showing the most low quality responses. The results highlight the potential value of generative AI in supporting patients and families, while emphasizing the need for expert oversight to mitigate inaccuracies and ambiguity.
The paper introduces AMIE, an AI system optimized for clinical management dialogue that uses large language models to reason about disease progression and therapeutic responses. In a randomized study against primary care physicians, AMIE demonstrated non-inferior performance in management reasoning and better precision in treatment plans aligned with clinical guidelines.
This work explores the use of generative large language models for drug sensitivity prediction from structured pharmacogenomics data. With domain specific prompt engineering and fine tuning, GPT achieved strong performance and generalization across multiple datasets, outperforming established baselines in several settings. The results indicate that generative LLMs can serve as effective computational tools to support precision oncology research.
This study evaluates the ability of GPT-4o to perform differential diagnosis for rare genetic diseases across nine languages using nearly 5,000 structured clinical vignettes derived from HPO-annotated phenotypes. Using zero-shot prompting and ontology-based automated evaluation, the model achieved comparable performance across English and eight non-English languages, ranking the correct diagnosis first in about 17–21% of cases and within the top three in about 25–28%. The consistency of results across languages suggests that language barriers may not substantially degrade LLM-based diagnostic reasoning. Overall, the findings indicate that LLMs like GPT-4o have potential utility for differential diagnosis support in non-English clinical settings.
An orchestrator coordinates a history-taking simulator, a knowledge retrieval agent and a diagnosis strategy agent to refine differential diagnoses over multiple turns. On a new benchmark covering respiratory, skin and rare diseases it improved interactive diagnosis accuracy by over 10 percent across LLM sizes.
Adds specialised models for history-of-present-illness inquiry and initial diagnosis, trained on over 3.5 million medical records, to an LLM agent that must gather information interactively. It improved over baselines by more than 30 percent and approached the accuracy obtained with complete clinical data.
A Therapist agent drafts answers to insomnia patients' questions and a Supervisor agent critiques and refines them in an actor-critic protocol based on active inference. Blinded CBT-I therapists often rated the system's answers higher than appropriate therapist-written responses.
Deep learning has the potential to transform radiology and nuclear medicine by delivering faster and more reliable image analysis than humans, but the role of generative AI in clinical decision making remains controversial. AI lacks understanding, judgment, and empathy, so its outputs must be interpreted and applied by nuclear physicians who retain responsibility for patient care. Used in a complementary manner, AI can support precision theranostics by integrating radiomics, genomics, and clinical data while enabling physicians to focus on informed, patient-centered treatment decisions.
Automates construction and continual updating of a medical knowledge graph while retrieving external evidence from sources such as PubMed during reasoning. It reached 74.1 F1 on MedQA and 66.3 percent on MedMCQA, above comparable and much larger models without extra compute.
MedAgentBench is a FHIR-compliant virtual EHR environment with 300 physician-written tasks across 100 patient profiles for testing LLM agents. It provides a standard way to measure how well agents plan and use tools on clinical record tasks.
ChatExosome is an AI agent that uses deep learning to analyze Raman spectroscopy data of exosomes for diagnosing hepatocellular carcinoma, achieving high accuracy. It also includes a chatbot component based on large language models to enhance knowledge retrieval and user interaction.
Ask Patients with Patience (APP) is a large language model-based medical assistant designed to engage in empathetic dialogue and provide grounded, transparent diagnoses. The system was evaluated using patient agents simulating real-world consultations, demonstrating improved diagnostic accuracy and user engagement.
N-Power AI is an agentic framework that uses large language models to perform sample size and power analysis for clinical trial design. It consists of specialized agents for function identification, calculation, and reporting, achieving 100% accuracy in comparison with ground truths from statistical software.
This work presents **MedBioLM**, a domain-adapted biomedical question-answering model that combines fine-tuning with retrieval-augmented generation to improve factual accuracy, reasoning, and contextual depth in medical and biological queries. Evaluated across diverse biomedical QA benchmarks, MedBioLM shows substantial accuracy gains from fine-tuning and improved factual consistency through RAG. The results demonstrate the promise of domain-optimized LLMs for biomedical research, education, and clinical decision support.
MultiMedRes uses a learner agent that splits a multimodal medical question into sub-problems, queries domain expert models in repeated ask-answer rounds, and integrates the answers. It targets zero-shot medical image reasoning such as comparing disease severity and lesion sizes across images.
MedRAX is an AI agent designed to interpret chest X-rays and provide medical reasoning for complex queries, integrating advanced CXR analysis tools with large language models. It demonstrates superior performance compared to existing solutions on ChestAgentBench, a comprehensive benchmark of 2,500 medical queries.
JingFang is a multi-agent system designed for Traditional Chinese Medicine that uses large language models to provide personalized medical consultations and accurate syndrome differentiation. The system demonstrates significant improvements in diagnostic accuracy compared to existing TCM models and state-of-the-art LLMs.
The study developed a multi-agent AI workflow using LLaMA 3 8B to identify cognitive concerns in clinical notes, achieving high classification performance comparable to expert-driven methods. The agentic workflow showed improved specificity and efficiency, demonstrating potential for scalable and cost-effective detection of cognitive concerns.
This review surveys past, current, and emerging applications of artificial intelligence in clinical genetics, spanning diagnostics, patient management, therapeutics, and clinical support. It outlines key AI categories, including machine learning, deep learning, and generative AI, and discusses their evolving roles in genetic practice. The authors argue that AI is likely to substantially transform clinical genetics and emphasize the need for proactive preparation to balance benefits and risks.
ENTAgents is a framework that combines retrieval augmented generation and multi-agent systems to improve the accuracy and explainability of responses in otolaryngology. It demonstrates enhanced performance over other LLM models by refining answers with additional information and self-correcting in multiple-choice questions.
The MAIDS framework uses a multi-agent system to support in-patient diagnostic decision-making, improving disease assessment accuracy by 20.7% over existing large language models and showing better clinical compliance than human experts.
The paper presents CRAFT-MD, a framework for evaluating large language models in clinical settings through simulated patient interactions. It highlights the limitations of current LLMs in handling complex medical dialogues and proposes recommendations to improve future evaluations.
The paper evaluates the effectiveness of large language model-based systems in answering clinical questions, finding that retrieval-augmented generation and agentic systems outperform general-purpose LLMs. It highlights the potential for combining evidence summarization with novel evidence generation to enhance patient care.
This study benchmarks large language models for gene prioritization in rare disease diagnosis, addressing challenges from limited data and large candidate gene sets. Using multi-agent strategies, Human Phenotype Ontology–based case stratification, and a divide-and-conquer ranking framework, GPT-4 consistently outperformed other models in identifying causal genes while mitigating positional and literature biases. Although biases toward well-studied genes and input order sensitivity remain, the proposed framework significantly improves diagnostic efficiency for rare genetic disorders. This work proposes an autonomous mechanical design framework that combines a pretrained large language model with finite element method (FEM) analysis to iteratively generate, evaluate, and refine structural designs without domain-specific fine-tuning. Using 2D truss structures as a testbed, the LLM effectively navigates discrete, multi-objective design spaces and achieves faster convergence with fewer FEM evaluations than NSGA-II. The results demonstrate that reasoning-based, language-driven LLMs can serve as efficient optimizers for autonomous structural design and iterative refinement. -->
This randomized trial examines how prior healthcare system engagement influences patient uptake of digital tools for cancer genetic services. Findings suggest that existing patient-system interactions are key predictors of engagement with genetic testing workflows.
Alvessa is an evidence-grounded research assistant designed to integrate entity recognition, biological tool orchestration, and data-constrained answer generation with statement-level verification. It improves accuracy in answering functional genomics and drug target assessment questions compared to general-purpose language models while ensuring traceable outputs.
Queryome is a system of specialized large language model agents designed to perform iterative, evidence-grounded synthesis across biomedical literature. It demonstrated superior performance on the MIRAGE benchmark and excelled in constructing review articles compared to other systems.
Moremi Bio Nano, an agentic reasoning LLM, was given only the heavy and light chain sequences of an anonymized monoclonal antibody and asked to hypothesize, rank and check its antigen and pathogen targets. Four of ten blinded runs completed and three of those identified the antigen confirmed by wet-lab data.
The Agentic AI Model for NGS Downstream Analysis automates the interpretation and analysis of Next-Generation Sequencing data, providing biological insights and advanced analytical recommendations. It demonstrated its utility by identifying significant differentially expressed genes in a cancer dataset and linking them to clinical outcomes through literature-based validation.
GRASP is a multi-agent framework that uses graph reasoning to develop quantitative systems pharmacology models, making the process more accessible and rigorous for domain experts. It outperforms traditional methods in terms of biological plausibility, mathematical correctness, structural fidelity, and code quality when evaluated by human judges.
Open-source LLM agents read PubMed abstracts to confirm evidence for genes already in a gene set and to propose missing genes, producing context-specific sets. Applied to reconstructing Human Phenotype Ontology gene sets, it recovered and extended their membership from literature evidence.
Genie-CAT is an AI system that integrates large language models with domain-specific tools to accelerate hypothesis generation in protein design, focusing on metalloproteins. It demonstrates the ability to autonomously identify residue-level modifications affecting redox properties, reproducing expert hypotheses more efficiently.
MedSAM-3 is a medical image segmentation model that fine-tunes the Segment Anything Model (SAM) on annotated medical images to enable precise anatomical structure segmentation via text prompts. The system demonstrates superior performance across various imaging modalities compared to existing models, showcasing its potential for biomedical discovery.
BioSage orchestrates retrieval agents with query planning, cross-disciplinary translation agents and reasoning agents over biomedical, biosecurity, AI and data science literature with citations. On LitQA2, GPQA, WMDP and HLE-Bio plus a new cross-modal benchmark it beat plain and retrieval-augmented baselines by 13-21%.
LongevityLLM is an AI agent designed to perform real, reproducible analyses in aging biology and structural bioinformatics by mapping natural language queries to deterministic bioinformatics pipelines. It demonstrated its capability to retrieve scientific information from databases and literature, integrate multiple epigenetic clocks, predict the impact of mutations on protein structure, and conduct phylogenetic analysis.
OpenBioLLM reproduces and extends GeneGPT with open-source models and separate agents for tool routing, query generation and response validation over genomic databases. It matched or beat GeneGPT on over 90% of tasks, scoring 0.849 on GeneTuring and 0.830 on GeneHop.
Checks three KOSMOS-generated radiation biology hypotheses on public expression datasets against random-gene null models. One hypothesis was well supported, one uncertain and one false, showing the need for rigorous auditing.
Agentic Lab is an AI platform that integrates large language models and vision-language models to assist in cell and organoid experimentation, enabling real-time protocol design, data analysis, and error detection. It demonstrates its capability by autonomously managing the differentiation of human pluripotent stem cells into organoids, providing context-aware instructions and improving experimental efficiency through continuous learning.
BioVerge Agent is an LLM-based framework designed for biomedical hypothesis generation, utilizing a ReAct approach with iterative generation and evaluation modules. The study demonstrates that structured and textual data sources enhance the quality of generated hypotheses, while self-evaluation improves their novelty and relevance.
FLEX is an autonomous agent that uses a gradient-free learning paradigm to continuously evolve its capabilities through experience, improving performance in tasks like chemical retrosynthesis and protein fitness prediction. The system demonstrates scalable continuous evolution and inheritable improvements across different agents, showing significant performance gains on benchmark datasets.
UORCA is a tool that uses large language models and multi-agent systems to automate the reanalysis of public transcriptomic datasets, identifying differentially expressed genes. It demonstrated its utility by uncovering biological patterns not initially described in the literature and generating hypotheses relevant for drug design.
QSP-Copilot is an AI-augmented platform that accelerates the development of Quantitative Systems Pharmacology (QSP) models by automating tasks such as literature extraction and model structuring. It demonstrated high precision in extracting biological interaction pairs from peer-reviewed articles, significantly reducing manual curation efforts.
TEAM-PHI is a multi-agent framework that uses large language models to automatically evaluate and select the best PHI de-identification model, reducing reliance on expert annotations. It demonstrates consistent and accurate rankings of de-identification models through independent evaluations by multiple agents followed by LLM-based majority voting.
DeepEvolve is an AI agent that integrates deep research with algorithm evolution to discover new scientific algorithms. It demonstrates consistent improvement in generating executable and validated algorithms across various complex domains, showcasing its capability to bridge the gap between unguided evolution and theoretical research.
M-Reason assigns each agent to one biomedical evidence stream so retrieval, appraisal and synthesis run in parallel with full traceability from source to conclusion, focused on cancer research. An interactive interface lets researchers inspect the workflow, and evaluation showed gains in efficiency and output consistency.
Qimai is a multi-agent framework that predicts DNA-protein interactions by integrating deep learning predictions with biological evidence using an LLM as the reasoning engine. It demonstrated superior performance on unseen proteins compared to standalone models, highlighting its capability in generalizing across diverse datasets.
BioinfoMCP is a platform that converts bioinformatics tools into Model Context Protocol (MCP) compliant servers, enabling seamless integration with AI agents. The system was validated across diverse computational tasks and showed high reliability and versatility in executing complex workflows.
The Nano Bio-Agent (NBA) framework uses small language models to address genomics questions, integrating task decomposition and API access for efficient performance. It demonstrates superior accuracy compared to larger models on the GeneTuring benchmark while significantly reducing computational costs.
OpenLens AI is a multi-agent system designed for health informatics research, automating tasks such as literature review, data analysis, and manuscript preparation. It demonstrates the ability to generate publication-ready manuscripts with transparent workflows while incorporating medical visualization interpretation and quality control.
BCI-Agent repurposes vision-language models as few-shot classifiers of neuronal cell types from electrophysiological features and wraps this in an agent that tracks neurons, validates calls against molecular atlases and synthesizes literature. It was tested against optogenetically tagged datasets for subtype inference.
HypoGeneAgent is a large language model-driven framework that uses gene-set analysis and hypothesis generation to optimize the resolution selection and functional annotation of clusters in single-cell Perturb-seq datasets. The system demonstrated improved cluster resolution and functional annotation compared to traditional metrics, highlighting its potential for automated interpretation pipelines in multi-omics studies.
The AI co-scientist platform generated hypotheses about the mechanism of gene transfer in bacteria, which were experimentally validated. The study demonstrates that AI can effectively generate and test high-level scientific hypotheses, potentially accelerating discovery in biological research.
Paper2Agent converts research papers into interactive AI agents that can interpret genomic variants and perform single-cell and spatial transcriptomics analyses, demonstrating its ability to reproduce original results and answer novel queries. It showcases the potential of transforming static scientific literature into dynamic tools for knowledge discovery and exploration.
AuraGenome is an LLM-powered framework that generates multilayered circular genome visualizations, combining a semantic-driven multiagent workflow with interactive visual analytics. The system was validated through case studies and user evaluations, demonstrating its effectiveness in rapid, reusable, and scalable visualization of genomic data.
SciToolAgent is a scientific agent that integrates and orchestrates hundreds of computational tools across various domains using a knowledge graph, enabling both experts and nonexperts to automate complex workflows. The system demonstrates its effectiveness in handling tasks like protein engineering and chemical synthesis through extensive evaluations and case studies.
PRINCE is a cloud-hosted data integration platform that uses a multi-agent architecture to integrate and analyze safety study reports for preclinical drug development. It evolved from a simple search tool into a sophisticated research assistant capable of drafting regulatory documents, demonstrating the potential of AI in enhancing data accessibility and research efficiency while maintaining human accountability.
The paper introduces a self-explained thinking agent that uses multimodal large language models to identify image degradation in microscopy and select restoration strategies. The system demonstrates high accuracy and explainability, outperforming human-level performance across various scenarios.
miRKatAI is a multi-agent system that integrates and analyzes microRNA-target interactions using a comprehensive database (miRKatDB) and large language models. It provides researchers with an intuitive interface for querying complex biological data, enhancing exploratory analysis and hypothesis generation in miRNA research.
This study applies AI to uncover aging-related mechanisms underlying idiopathic pulmonary fibrosis by combining a pathway-aware proteomic aging clock with a domain-specific language model, ipf-P3GPT. The aging clock shows strong predictive performance and links accelerated aging to severe COVID-19, while molecular analyses reveal that IPF reflects dysregulated aging pathways rather than simple aging acceleration. These results highlight new connections between aging biology and IPF and demonstrate the value of AI-guided strategies for studying age-related diseases.
BioDisco is a multi-agent system designed to generate novel scientific hypotheses by integrating language model reasoning with dual-mode evidence retrieval from biomedical knowledge graphs and literature. The system demonstrates superior hypothesis novelty and significance through rigorous temporal evaluation methods, outperforming both ablated configurations and generalist agents.
BioRAGent is a biomedical assistant that uses retrieval-augmented generation and a multiagent system to facilitate natural language queries about genes, phenotypes, diseases, and their interrelationships. It demonstrates superior performance in handling complex multi-hop queries compared to existing models through automated benchmarks and user evaluations.
CellAtria is an agentic AI framework that automates the process of ingesting and standardizing single-cell RNA-seq data analysis through a chatbot interface. It demonstrated efficient and accessible data reuse by integrating large language models with tool-execution capabilities to handle metadata extraction, dataset retrieval, and downstream analysis.
This work presents CRISPR-GPT, an LLM-based agent system designed to support end-to-end CRISPR gene-editing experiment design and data analysis. By integrating domain knowledge, retrieval, external tools, and a fine-tuned model, CRISPR-GPT assists with experimental planning, guide RNA design, protocol drafting, and result interpretation. Experimental validations demonstrate its effectiveness as an AI co-pilot for genome engineering across different CRISPR modalities.
PrimeGen is a multi-agent system that uses large language models to streamline primer design for targeted next-generation sequencing, reducing labor and improving efficiency. The system demonstrates high amplification uniformity and minimizes dimer formation across various applications.
GeneAgent is an AI agent designed to analyze gene sets by generating functional descriptions and verifying them against biological databases, thereby reducing inaccuracies. It was shown to be more accurate than GPT-4 in analyzing diverse gene sets and provided valuable insights into gene functions when applied to novel datasets.
AgentMRI is an autonomous AI system that uses vision language models to detect and correct MRI image degradations, selecting the appropriate deep learning model for reconstruction without human intervention. It demonstrated high accuracy in both zero-shot and fine-tuned settings on a brain MRI dataset, showcasing its potential for scalable and intelligent MR image processing.
This study introduces an uncertainty-aware framework for protein–protein interaction prediction using fine-tuned large language models. By integrating LoRA ensembles and Bayesian LoRA for uncertainty quantification, the approach improves the reliability and reproducibility of LLM-based PPI analysis across multiple disease contexts. The results highlight the importance of confidence-aware modeling for trustworthy applications of LLMs in precision medicine and computational biology.
RDMA is an agentic framework that uses smaller quantized language models to mine rare diseases from unstructured clinical notes, resolving abbreviations and grounding phenotypes against Orphanet and HPO ontologies. It outperforms baselines without task-specific training, significantly reducing inference costs and expert annotation burden.
Manager, developer, critic and tool-creation agents update reasoning templates and grow their own tool collection, extending to robotic execution via vision-language-action models. It scored 4.01 of 5 on its tool-creation benchmark above GPT-5 and Biomni, identified BTN3A1 as a negative regulator of NK cell function confirmed by CRISPR knockout, and found an enzyme variant with doubled activity.
LLM agents acting as a multidisciplinary team iteratively propose, debate and settle on phenotypes and confounders for phenome-wide association analysis of heart and aorta imaging data. Its autonomously chosen phenotypes matched expert-selected ones on disease classification, with mean AUC differences near zero and better recall for six of nine diseases.
BioMARS is a multi-agent robotic system that integrates large language and vision models to autonomously design, plan, and execute biological experiments. It demonstrated the ability to match or exceed manual performance in cell culture tasks and supported context-aware optimization for differentiating retinal pigment epithelial cells.
STELLA is a self-evolving AI agent designed for biomedical research that uses a multi-agent architecture to autonomously improve its capabilities. It demonstrated superior performance on several biomedical benchmarks and showed significant improvement in accuracy with increased trials, indicating its potential for dynamic scaling of expertise.
AgentDistill is a framework for distilling knowledge from large language model-based agents to smaller models without training, using reusable task-solving modules called MCP boxes. The system demonstrates its effectiveness in biomedical and mathematical tasks by achieving performance comparable to advanced systems with minimal supervision.
The paper introduces a multi-agent AI system that uses large language models to curate high-quality metadata from biomedical datasets and associated publications, achieving both precision and scalability. The system outperforms existing automated methods in terms of recall on key metadata fields, demonstrating its potential for downstream applications like classifier model development.
OriGene is a self-evolving multi-agent system that integrates over 600 specialized tools to identify therapeutic targets by reasoning across diverse data modalities. It demonstrated superior performance in accuracy and robustness compared to human experts and other AI models, and nominated previously underexplored therapeutic targets for liver and colorectal cancer with significant anti-tumor activity.
KODA is an agentic framework that uses large language models and knowledge graphs to discover antimicrobial drug targets in the gut microbiome. It demonstrates its efficacy by identifying conserved essential genes as potential drug targets, thereby providing actionable insights for antimicrobial research.
The Potato Knowledge Hub is a platform that uses large language models and curated research papers to provide accurate gene information and specialized knowledge about potatoes. It includes a functional gene database and a retrieval-augmented generation knowledge base, which mitigates LLM hallucination by providing literature citations for answers.
BehaveAgent is an autonomous AI agent designed to analyze behavior from video inputs across various species without the need for retraining or manual intervention. It demonstrates its capability by autonomously identifying behavioral paradigms, tracking subjects and objects, executing targeted analysis, and generating comprehensive research reports.
Agents search MEDLINE through the PubMed API and communicate bidirectionally to draft literature reviews with verifiable citations across five medical fields. Citation accuracy was 99.82 percent, similar to Perplexity Sonar, and referencing accuracy was higher at 96.81 percent.
The study uses AI tools like ChatGPT and GitHub Copilot to assist in the analysis of functional protein networks in invasive colorectal cancer, showing that while these systems can improve efficiency, they require human oversight for complex tasks. The research highlights the limitations of current AI in fully automating biotech studies due to issues with data accuracy and system stability.
Intelliscope integrates knowledge graphs with large language models to facilitate biomedical discovery by identifying relevant pathways between entities and designing AI predictors for these pathways. The system also proposes wet-lab experiments for validation, demonstrating its capability in accelerating scientific discoveries across various therapeutic areas.
ARIEL is an evaluation and optimization framework that uses expert-vetted tasks to assess the capabilities of AI models in biomedical research, focusing on summarization and figure interpretation. The system demonstrates improvements through prompt engineering and fine-tuning, showing potential for generating testable hypotheses by integrating textual and visual information.
The paper introduces a multi-agent framework for distilling high-quality biomedical data from scientific literature, enhancing the training of large language models in biomedicine. The system demonstrates significant improvements in question-answering tasks compared to existing models and baselines.
The paper introduces P3GPT, an AI agent-driven framework that integrates a multi-omics transformer model with TCM compound-target interaction databases to identify potential geroprotectors from traditional Chinese medicine. It demonstrates the system's ability to bridge ancient therapeutic knowledge with modern pharmacology by identifying promising TCM formulations for age-related diseases.
TxGemma is a suite of large language models fine-tuned for therapeutic development tasks, demonstrating superior performance in predicting properties and adverse events compared to existing models. It also includes conversational agents that provide mechanistic reasoning and engage in scientific discussions, enhancing its utility in drug discovery.
scAgent is a large language model-based framework designed to annotate and discover novel cell types from single-cell RNA-seq data across various tissues. The system demonstrates superior performance in general cell-type annotation, novel cell discovery, and extensibility to new cell types compared to existing methods.
This paper introduces SpatialAgent, a fully autonomous AI agent designed to support end to end spatial biology research by integrating large language models with adaptive reasoning and dynamic tool use. Evaluated on large scale human and mouse spatial datasets, SpatialAgent outperforms existing computational methods and matches or exceeds human performance on key tasks. The study demonstrates a new paradigm for scalable and collaborative AI driven discovery in spatial biology.
OLAF is an open-source platform that uses large language models to enable researchers to perform bioinformatics analyses through natural language commands. It demonstrated the ability to execute complex bioinformatics workflows, including single-cell RNA-seq analysis and gene annotation, making computational biology more accessible to non-programmers.
This study introduces the PreCancerous Molecular Resource (PCMR), a large integrated database compiling multi-omics molecular profiles of precancerous lesions across diverse cancer types and tissues. By combining differential analysis with ChatGPT-based text mining, PCMR systematically identifies precancer–gene associations and validates their biological and clinical relevance. The resource provides a comprehensive foundation for studying early carcinogenesis and advancing precancer research and intervention.
GeneWhisperer is a system that uses large language models to assist in the manual annotation of gene functions, aiming to enhance human expertise and efficiency. It demonstrates improved accuracy and speed in genome annotation tasks through its conversational interface and integration with relevant bioinformatics tools.
The Enzyme Co-Scientist is an LLM-based system designed to extract enzyme kinetic data from scientific literature, significantly reducing the need for manual curation. It demonstrated high accuracy with an F1 score above 0.9 on expert-annotated datasets and successfully extracted over 91,000 entries from nearly 3,500 papers.
This paper introduces MRAgent, an automated LLM-driven agent designed to facilitate large-scale causal discovery using Mendelian Randomization. MRAgent autonomously mines scientific literature to identify candidate exposure–outcome pairs and then performs MR analyses using genome-wide association study data, reducing reliance on manually curated hypotheses. Evaluations across different LLMs and a proof-of-concept case study demonstrate its ability to execute end-to-end causal inference workflows. The results highlight MRAgent as a powerful tool for scalable, systematic causal analysis in complex disease research.
BixBench is a comprehensive benchmark designed to evaluate the performance of LLM-based agents in conducting complex bioinformatics analyses. The study reveals that even advanced models like GPT-4o and Claude 3.5 Sonnet struggle with nuanced biological data analysis, achieving only 17% accuracy on open-answer questions.
Co-Scientist is a multi-agent AI system designed to assist scientists in generating novel hypotheses for complex scientific problems. It demonstrated its effectiveness by identifying new drug repurposing candidates and synergistic combination therapies for acute myeloid leukemia, which were validated through in vitro experiments.
BioMaze is a dataset and evaluation framework for assessing large language models' ability to reason about complex biological pathways. It demonstrates that existing LLMs struggle with pathway reasoning, especially in perturbed systems, but shows promise when enhanced through interactive subgraph-based navigation.
LLM agents turn a free-form hypothesis into measurable implications, design and run falsification experiments on existing or new data, and aggregate results under sequential testing with Type-I error control. On biological hypotheses it matched human scientists' validation performance while taking about a tenth of the time.
Harmonia is a system that uses large language models and an interactive interface to automate the creation of data harmonization pipelines, making it easier for experts to integrate datasets from diverse sources. The paper demonstrates Harmonia's effectiveness in clinical data harmonization by creating reusable pipelines that standardize dataset formats.
ESCARGOT is an AI agent that integrates large language models with a dynamic graph of thoughts and biomedical knowledge graphs to enhance reasoning capabilities. It demonstrates improved reliability and reduced hallucinations compared to traditional retrieval-augmented generation methods, offering greater transparency in its reasoning process.
Informed by a user study with drug discovery researchers, it decomposes protein-protein interaction pathway analysis over a large knowledge graph into edge-level sub-tasks solved by an agent pipeline. It targets explainable exploration of therapeutic impacts for target identification.
BioAgents is a multi-agent system designed to facilitate complex bioinformatics tasks by leveraging small language models fine-tuned on genomics data and augmented with retrieval capabilities. The system demonstrates performance comparable to human experts in conceptual genomics tasks, indicating its potential for democratizing access to advanced bioinformatics analysis.
This study benchmarks GPT-3.5 and GPT-4 against professional human curation for genetic trait and QTL extraction. The results show that GPT-4 achieves high accuracy in document classification and trait retrieval, supporting a hybrid human–AI curation paradigm for biological databases.
The paper presents Foam - Agent, a multi - agent framework automating CFD workflows from natural language. It features unique retrieval, file - generation and error - correction systems, lowering expertise barriers.
The paper presents Foam - Agent, a multi - agent framework automating CFD workflows from natural language. It features unique retrieval, file - generation and error - correction systems, lowering expertise barriers. -->
This study aims to enhance large language model performance in genomics by incorporating large-scale variant annotation data using retrieval-augmented generation (RAG) and fine-tuning. By integrating 190 million curated variant annotations from five major datasets into GPT-4o, the system enables accurate, reasoning-supported variant queries and interpretations. Results show that while fine-tuning improves performance in some annotation fields, RAG is more effective overall in terms of scalability, accuracy, and cost-efficiency for injecting factual genomic knowledge. The work demonstrates the potential of LLMs as accessible and efficient tools for variant interpretation and specialized genomics applications.
This paper presents AutoMAxO, a semi-automated LLM-based workflow designed to scale the biocuration of the Medical Action Ontology (MAxO) for rare diseases, addressing the limitations of manual curation. The system uses LLMs to extract candidate treatment and medical action annotations from literature abstracts, then aligns them with MAxO, HPO, and MONDO ontology terms through post-processing and human-in-the-loop validation. Applied to abstracts covering 37 rare genetic diseases, AutoMAxO identified 958 novel treatment annotations incorporated into the MAxO dataset. The results demonstrate that LLM-assisted curation can substantially improve coverage and efficiency in clinical ontology development.
Prompt-to-Pill is a multi-agent system that integrates various large language models to automate drug discovery and clinical trial simulation from target identification to virtual patient recruitment. The study demonstrates the pipeline's capability by applying it to Dipeptidyl Peptidase 4 (DPP4), showcasing its potential for operationalizing drug development workflows in silico.
Trains LLM agents with reinforcement learning to edit and optimize molecules over multiple turns, calling tools for validity, property and similarity checks and using their feedback. MolEditAgent-7B made nearly all requested edits validly, and MolOptAgent-7B beat the best closed reasoning model on LogP.
The paper describes an AI-driven platform for discovering materials used in insulin delivery patches, which uses active learning and integrates literature mining, polymer SMILES generation, and molecular dynamics simulations. The system demonstrates the potential of large language models to orchestrate complex scientific workflows, accelerating material discovery in biomedical applications.
SelfAI turns a research intent into executable experiments, reasons over accumulated experiment trajectories to pick the next move and stops unproductive branches under explicit efficiency-diversity trade-offs. On real tasks from machine learning to drug discovery it found good solutions with fewer redundant trials than classical optimization and other LLM baselines.
The paper introduces a swarm intelligence framework using multiple large language model agents for de novo protein design, which iteratively proposes mutations to achieve tailored properties without fine-tuning or specialized training. The system's effectiveness is demonstrated through experimental validation on proteins with alpha helix and coil structures, showing efficient navigation of the protein fitness landscape.
MADD is a multi-agent system designed for drug discovery that builds and executes customized hit identification pipelines from natural language queries. It demonstrates superior performance compared to existing large language model-based solutions, pioneering AI-first drug design for several biological targets.
Coated-LLM has researcher, reviewer and moderator LLM agents generate and evaluate hypotheses about effective drug combinations for Alzheimer's disease. It outperformed knowledge-based methods (accuracy 0.74 vs 0.52), and one predicted combination reduced amyloid aggregation in vitro.
LIDDiA is an autonomous AI agent designed to navigate the drug discovery process, leveraging large language models for reasoning and molecule generation. It demonstrated its capability by generating molecules that meet pharmaceutical criteria for over 70% of clinically relevant targets and identified a promising candidate for treating prostate and breast cancers.
MolAgent is an agentic AI framework designed for high-fidelity modeling of molecular properties in early-stage drug discovery. It autonomously implements expert-level pipelines for property estimation, ensuring optimal accuracy and model robustness through automated feature engineering and comprehensive validation frameworks.
Combines structured tool use with a parameterized memory pool that standardizes heterogeneous drug data across multi-turn workflows, trained on a new drug instruction dataset. It completed 98.0, 93.5 and 64.0 percent of simple, multi-tool and multi-turn tasks, ahead of ReAct and LoT.
DELPHAI is an AI agent designed to predict drug responses and resistance mechanisms in patient-derived organoids, particularly for glioblastoma. It successfully identified the mechanism of action for temozolomide and suggested combination therapies that align with clinical trial outcomes, demonstrating its potential in precision medicine.
Moremi Bio Agent is an AI system designed to autonomously discover broad-spectrum antibiotics targeting Enterobacteriaceae. It demonstrated the feasibility of using dual-target strategies for antibiotic design, generating and evaluating over a thousand candidate molecules with promising drug-like properties.
FROGENT is a multi-agent system designed for full-process drug design, integrating target identification, small-molecule generation, peptide optimization, and retrosynthetic planning. It demonstrates superior performance across various benchmarks compared to existing methods, showcasing the potential of large language model-based systems in automating complex drug discovery pipelines.
BioScientist Agent is a framework that integrates a large biomedical knowledge graph with machine learning modules to facilitate drug repurposing and elucidate the mechanism of action for drug-disease pairs. The system surpasses existing methods in various metrics, providing mechanistic explanations consistent with literature.
The paper describes a hierarchical agent framework for automated molecular optimization in drug discovery, where agents collaborate to generate and optimize molecules. The study demonstrates that multi-agent systems excel at focused binding optimization compared to single-agent or unguided LLM approaches.
An LLM principal investigator agent runs research meetings with specialist LLM agents under light human direction, here building a nanobody design pipeline around a protein language model, AlphaFold-Multimer and Rosetta. The 92 designed nanobodies included functional binders, two with improved binding to recent variants while retaining ancestral spike binding.
NAMD-Agent is an automated pipeline that uses large language models and Python scripting to prepare, run, and analyze molecular dynamics simulations for proteins. It was evaluated across various biomolecular systems, demonstrating its ability to reproduce key metrics from published benchmarks while differing in efficiency and cost among different LLM orchestrators.
DrugMCTS combines retrieval, five specialized agents for molecular and protein evidence, and Monte Carlo tree search to reason about drug repositioning. On DrugBank and KIBA it achieved higher recall and robustness than general LLMs and deep learning baselines.
This paper introduces AutoPM3, an automated framework that extracts ACMG/AMP PM3 variant evidence from biomedical literature to support rare disease diagnosis. Using open-source LLMs combined with a Text2SQL-based variant extractor and a retrieval-augmented generation module, AutoPM3 efficiently processes both text and tables. Evaluated on the newly curated PM3-Bench dataset, the method achieves high accuracy and recall, outperforming approaches that rely on larger models. The results show that open-source LLMs can provide fast, cost-effective, and scalable solutions for literature-based variant interpretation.
This work introduces GOLF, a generative AI framework for predicting and interpreting the pathogenicity of MYOC OLF domain variants associated with glaucoma. By combining generative models with sparse autoencoders, the framework achieves accurate classification while revealing underlying biochemical features driving predictions.
MT-Mol is a multi-agent system designed for molecular optimization that integrates large language models with RDKit tools to enhance structured reasoning and interpretability. The system demonstrates superior performance on the PMO-1K benchmark, showcasing its effectiveness in generating high-quality molecules through collaborative interactions between specialized agents.
The authors prompted the Moremi Bio biodesign agent with its safety guardrails removed and examined the toxicity of what it produced, finding generated proteins computationally resembling known toxins. They use this dual-use assessment to argue for layered safeguards in LLM-enabled biodesign pipelines.
Integrating large language models with biomedical research tools raises challenges in transparency, reproducibility, and data provenance. This work introduces BRAD, an agent-based system that combines LLMs with external tools and databases to automate research workflows while maintaining transparent and reliable protocols. Applied to biomarker discovery and other tasks, BRAD enables contextualized interpretation and automation that exceed conventional bioinformatics tools.
The paper introduces the DO Challenge, a benchmark for evaluating AI agents' decision-making in drug discovery. The Deep Thought multi-agent system was developed to participate in this challenge and showed strong performance but still fell short of expert-designed solutions, highlighting both potential and limitations of AI-driven methodologies in molecular drug design.
Fleming is an agent for tuberculosis antibiotic lead discovery that combines trained inhibition classifiers, generative models, molecular optimization, ADMET prediction and literature search. It separated known Mtb leads from others better than a generic LLM agent and its predicted inhibitors showed an 83% in vitro hit rate.
PharmAgents is a virtual pharmaceutical ecosystem that uses large language model agents to streamline and accelerate the drug discovery process, from target identification to preclinical evaluation. The system demonstrates its potential by identifying therapeutic targets, discovering lead compounds, and conducting in silico analyses for toxicity and synthetic feasibility.
A fine-tuned worker LLM generates candidate molecules while a research agent searches literature for prior knowledge used to select the most promising ones. It reached an 82.3 percent success rate across six multi-objective tasks and was applied to improving A1R selectivity of XAC and saquinavir bioavailability.
CLADD is a retrieval-augmented generation (RAG)-empowered system that uses multiple LLM agents to dynamically retrieve and contextualize information from biomedical knowledge bases for drug discovery tasks. It demonstrates effectiveness in various drug discovery tasks, outperforming both general-purpose and domain-specific models as well as traditional deep learning approaches.
The Moremi Bio Agent was used to design monoclonal antibodies against the Plasmodium falciparum AMA1-RON2 invasion complex, generating 999 candidates and running structural and affinity evaluation. Binding analysis flagged 864 with successful target interaction and the top candidates showed very tight predicted binding and stable folds.
This work proposes an LLM-driven strategy to identify true negative drug candidates from clinical trial data, addressing a key bottleneck in supervised drug repositioning. By using GPT-4 to label failed drugs, the approach significantly improves prediction accuracy and enables large-scale, data-driven drug repurposing.
CASSIA is a multi-agent LLM system that annotates cell types in single-cell RNA-seq data with reasoning, quality scoring and confidence calibration. Across 970 cell types it improved annotation accuracy, including on complex and rare populations.
An LLM agent system processes differential expression signatures from the Tahoe-100M drug perturbation atlas and writes short summaries of each drug response. Blinded summaries were matched to their drug or mechanism of action better than chance, with results depending on the LLM and signature score.
A command-line agent built on gffutils and the Strands agent framework turns natural-language questions into executable queries over GFF genome annotations, using local or cloud LLMs. It supports coordinate and attribute queries, GO searches, hierarchy traversal, summaries and export without coding.
MS4MS is a multi-agent LLM system that runs LC-MS/MS small-molecule identification from raw spectra to annotated compounds, aimed at complex mixtures such as traditional Chinese medicine. It achieved the best molecular formula prediction on a public benchmark and identified compounds and isomers in herbal extracts.
kai is an agent that iteratively writes and runs single-cell analyses in computational notebooks so that human scientists can follow and critique its reasoning. Compared with one-shot generation it was more robust to code errors and sycophancy on cell type annotation problems.
The paper explores the use of lightweight LLM agents to improve automated annotation in spatial transcriptomics, integrating rule-based heuristics and multirole reasoning. Results show that these models match or slightly exceed deterministic baselines in cluster recovery while providing more spatially consistent predictions.
Agents generate competing cell type hypotheses from full expression data and study context, check them against external databases and iteratively self-evaluate. It outperformed reference-based and LLM-based annotators, and its confidence scores identified reliable annotations.
Scores candidate surfaceome genes across immunotherapy, inflammation and NF-kB contexts with structured prompting and takes consensus across three LLMs, checked with positive and negative control genes. It prioritized eight candidates and highlighted PTPRS, VCAN and MXRA5 as previously unrecognized stromal checkpoint-like regulators.
BRAINCELL-AID is a multi-agent system that annotates brain cell gene sets by combining free-text descriptions with ontology labels and refining predictions using retrieved PubMed literature. It produced correct annotations for 77% of mouse gene sets among its top predictions.
SPI-WSI iteratively generates, critiques with PubMed evidence and refines text prompts that a vision-language model uses to localize bulk RNA-seq pathway activity on whole-slide images, with pathologist guidance. Among the LLMs tested, Claude 4.0 gave the highest image-prompt similarity.
GenCellAgent routes cellular images to specialist segmenters or vision-language models through a planner-executor-evaluator loop with long-term memory and text-guided segmentation of uncovered organelles. Across seven benchmarks with 4,718 images it matched or beat the best single tool on every dataset and did markedly better on out-of-distribution organelles.
AI-HOPE-PM converts natural-language queries into Python statistical workflows that join genomic, clinical and social determinants of health data for cancer analysis. It was assessed on colorectal cancer cohorts from TCGA and cBioPortal for query interpretation accuracy and runtime.
Vibe coding is an iterative process where users describe their goals in natural language and AI agents generate, test, and refine executable code for omics data analysis applications. The paper demonstrates the rapid development of a proteomics data analysis website using vibe coding, showcasing its potential to democratize computational biology software development.
Specialised agents validate scope, resolve entities against authoritative resources, generate schema-aware SPARQL, refine failed queries and interpret results over mass spectrometry metabolomics knowledge graphs. It was validated on an expert-authored benchmark of natural-language questions about plant metabolites and bioactivities.
Given a task description and evaluation function, the agent organizes domain knowledge into a tree and iteratively diagnoses and improves candidate computational methods. It beat expert methods and other scientific agents on tasks including single-cell RNA-seq denoising, and on two genetics problems uncovered new associations.
BioML-bench evaluates AI agents on end-to-end biomedical machine learning tasks in protein engineering, single-cell omics, imaging and drug discovery, graded by standard metrics. Four open-source agents, including STELLA and Biomni, were tested and generally fell short of human baselines.
An agent for query-conditioned multi-table summarization uses entropy over table-selection rollouts and summary self-consistency as control signals, both in reinforcement learning and at inference. On multi-omics benchmarks it roughly tripled correct useful claims per summary and raised downstream survival prediction C-index from 0.32 to 0.63.
This version of AI-HOPE-TP53 pairs a fine-tuned LLaMA 3 model with harmonized cBioPortal cohorts to translate questions into mutation, survival and odds ratio analyses of TP53 alterations in early-onset colorectal cancer. It reproduced known associations and reported subgroup differences by treatment and population.
AI-HOPE-WNT is a conversational AI agent designed to enable natural-language, WNT-focused analysis of integrated clinical and genomic colorectal cancer data. The system accurately reproduces published findings on WNT pathway alterations while uncovering new associations related to survival, treatment response, mutation co-occurrence, and population-specific risk. By automating complex bioinformatics workflows, AI-HOPE-WNT lowers technical barriers and establishes a pathway-specific paradigm for precision oncology research.
K-Dense Analyst is a hierarchical multi-agent system that plans bioinformatics analyses and executes verified tasks in secure environments through a dual-loop design. On the open-ended BixBench it reached 29.2% accuracy, 6.3 points above GPT-5.
GenoMAS runs a team of programming agents that decompose gene expression analysis guidelines into action units and can advance, revise, bypass or backtrack while passing typed messages. On GenoTEX it reached 89.13% preprocessing composite score and 60.48% F1 for trait-associated gene identification, above prior systems.
AtlasAgent is a vision-language model and AI agent framework designed to evaluate large-scale single-cell data integration efficiently. It uses chain-of-thought reasoning and prompting strategies to assess batch correction quality, biological signal preservation, and overcorrection risks, significantly reducing evaluation time from hours to seconds while maintaining accuracy.
ARMOA is an autonomous system that integrates retrieval-augmented generation, large language models, and agentic AI to analyze multi-omics data for the PI3K/AKT pathway. It demonstrated high accuracy in predicting therapeutic outcomes and identifying drug repurposing candidates through external testing and cross-validation.
PromptBio is a modular AI platform that uses multi-agent systems to facilitate scalable and reproducible bioinformatics analysis, supporting both human-in-the-loop workflows and automated large-scale omics analysis. The system demonstrates its capability through the integration of specialized agents for data management, pipeline execution, and interactive summarization, alongside tools for machine learning and statistical interpretation.
Bio-Copilot is a multi-agent bioinformatics system with agent group management, human-agent interaction, a shared knowledge base and continual learning for large omics studies. It outperformed GPT-4o and other agents across bioinformatics tasks and reproduced and extended a human lung cell atlas.
DeepSeq is an AI system that uses web search and generative models to automatically label single-cell RNA sequencing data, achieving high accuracy. It demonstrates the potential of agentic foundation models in overcoming bottlenecks in biological data annotation by reducing reliance on manual curation.
PhenoGraph is a multi-agent LLM system that finds phenotype-annotated cohorts, aligns them with spatial transcriptomics data, and selects, runs and corrects analysis pipelines with knowledge graph support. It was demonstrated across several tumor microenvironment datasets and phenotype classes.
Agentomics-ML autonomously builds classification models for genomic and transcriptomic datasets by working through predefined machine learning steps via shell commands and reflecting on validation metrics. It produces the trained model plus files needed for reproducible training and inference.
MedAgentGym is a training environment of 72,413 coding tasks from 12 real biomedical data science scenarios in executable sandboxes with verifiable answers and trajectory generation. Benchmarking 29 LLMs showed large gaps, and its Med-Copilot gained over 43% from offline and online reinforcement learning.
This study introduces MEGA-GPT, a retrieval-augmented, AI-driven assistant designed to help users navigate the increasingly complex functionalities of the MEGA evolutionary genetics software through natural language queries. By integrating MEGA documentation and version-specific resources, MEGA-GPT provides step-by-step guidance, clarifies analytical settings, and recommends optimal workflows while significantly reducing hallucinations compared with standard ChatGPT. The tool improves usability for both novice and experienced users and demonstrates how customized RAG-based assistants can enhance complex scientific software.
A simulated team of GPT-4o agents was asked to write and run code reproducing findings from five highly cited Alzheimer's studies using the NACC dataset, given only abstracts, methods and the data dictionary. The agents approximately reproduced 53.2% of the 35 extracted findings, often differing in numeric values and statistical choices.
An LLM agent analyses Perturb-seq data, retrieves prior knowledge, runs statistical and predictive tools and proposes the gene panel for the next experimental round. It supports iterative screen design by prioritising perturbations under a fixed experimental budget.
A fine-tuned LLaMA 3 system turns natural-language questions into executable statistical pipelines over cBioPortal colorectal cancer data, building cohorts and running survival and mutation analyses. It reproduced known PI3K pathway associations in early-onset colorectal cancer and compared well with existing platforms.
AI-HOPE-TP53 is a conversational agent that turns natural-language questions into executable cBioPortal analyses of TP53 pathway alterations in colorectal cancer. It supports stratification by age, ancestry, MSI status, treatment and stage with mutation, co-mutation, survival and odds ratio analyses.
AI-HOPE-TGFbeta couples an LLM, a natural-language-to-code interpreter and a bioinformatics backend to analyze TGF-beta pathway alterations in colorectal cancer using harmonized clinical and genomic data. It enables cohort stratification and hypothesis testing aimed at underrepresented populations.
CellTypeAgent is an LLM agent that annotates cell types and verifies its calls against marker databases to limit hallucinated labels. It was tested on nine datasets covering 303 cell types from 36 tissues and was more accurate than existing methods.
This study investigates the genetic links between autoimmune processes and intracranial aneurysm (IA) by integrating transcriptomic data, functional enrichment analyses, and multiple machine-learning approaches. Two key autoimmune-related diagnostic genes, ADIPOQ and IL21R, were identified and used to build a neural-network model with strong predictive performance (AUC = 0.944). The results highlight immune infiltration and gene–miRNA–drug regulatory networks as important contributors to IA pathogenesis, suggesting potential therapeutic targets.
This work presents a comprehensive map of human subcellular architecture by integrating biophysical interaction data and immunofluorescence imaging for over 5,100 proteins. Using self-supervised multimodal learning and large language model–assisted annotation, the study identifies hundreds of molecular assemblies, reveals new protein functions, decodes cancer-associated alterations, and provides a public platform for structural and functional cell biology research.
This study applies an integrated multi-omic Mendelian randomization framework to identify causal plasma protein targets for immune-related bone diseases. By combining MR, SMR, Bayesian colocalization, LDSC, and downstream network and drug analyses, the authors systematically uncover disease-specific therapeutic targets across rheumatoid arthritis, multiple sclerosis, psoriatic arthritis, and Crohn’s disease–related arthritis. The results provide robust genetic evidence for key proteins underlying disease pathogenesis and offer a strong foundation for precision drug discovery in immune-associated bone disorders.
Converts scientific questions into structured queries against the CellDepot single-cell repository and drives a visualization platform to return violin plots, UMAP embeddings and heatmaps. It gives researchers without programming skills access to gene expression exploration across diseases, cell types and conditions.
An agentic platform lets users explore, simulate and analyse SBML kinetic models from the BioModels database through natural language instead of a fixed interface. Use cases cover precision medicine, epidemiology and emergent network properties for users without computational biology training.
This study evaluates large language models (LLMs) for automated R and Python code generation in predictive modeling tasks using standardized omics datasets from multiple DREAM challenges in reproductive health. Several LLMs successfully generated executable analysis pipelines, with OpenAI’s o3-mini-high performing best and achieving test performance comparable to or exceeding top DREAM challenge teams. The results demonstrate that LLMs can effectively automate key steps of omics-based predictive modeling, lowering technical barriers and accelerating exploratory biomedical research.
SAGE is an agentic auditing loop that injects soft prompt perturbations into genomic models such as ESM2, tracks behavior across checkpoints, computes risk metrics and writes narrative reports. It showed that variant effect prediction degrades measurably under targeted embedding-space attacks.
Plan, update, annotation and verification agents select codebooks, maintain retrieval databases, annotate transcripts with retrieval-guided prompting and check consistency. On 50 rheumatology and OB/GYN transcripts it reached an F1 of 0.928 against trained human coders.
A designer agent proposes coding workflows, a coder executes them on clinical notes, and a reflector critiques results while a memory archive stores prior designs for reuse. It outperformed earlier agentic and non-agentic baselines on medical coding benchmarks.
Evaluates open and closed LLM agents on iterative design of genetic perturbation and molecular property experiments against bandit and Gaussian process baselines. The agents ignored experimental feedback, performing the same with permuted labels, and a simple LLM-prior nearest-neighbour hybrid did better.
This study compares the performance of five Chinese generative AI models on a medical molecular biology exam. Results show that several models outperform undergraduate students, while underscoring the need for critical evaluation and domain-specific AI systems in education.
Compares an embedding-plus-retrieval pipeline with LLM agents that follow predefined mapping protocols and refine their queries, mapping Brazilian SIGTAP procedure and medicine codes to OMOP CDM concepts. Both reached similar F1, with the agent pipeline achieving better recall through dynamic query refinement.
SciGym is a dry-lab benchmark where LLMs iteratively design perturbation experiments on systems biology models encoded in SBML and interpret the simulated results. Six frontier models were tested on 137 systems, and all degraded sharply as system complexity grew.
A safety-evaluation framework in which an LLM agent with bioinformatics tools probes whether DNA generative models can be steered past their safety behaviour, scored automatically against a reference database. The audit found current models are vulnerable and that larger models carry greater dual-use risk, motivating stronger safety alignment and provenance tracing.
PhenoAssistant uses an LLM to orchestrate a curated toolkit for plant phenotype extraction, visualization and model training through natural language. It was validated on representative case studies and a set of evaluation tasks.
This work introduces STAgent, an autonomous multimodal AI agent that integrates large language models with specialized computational tools to automate complex spatial transcriptomics analyses. Applied to human stem cell–derived pancreatic tissues, STAgent rapidly uncovers spatial maturation patterns, cell–cell interactions, and underlying biological mechanisms while contextualizing findings with relevant literature. The results demonstrate a new agentic paradigm that dramatically reduces analysis time and expertise barriers in spatial biology research.
Agents conduct structured conversations to generate, refine and evaluate themes from interview transcripts, with cardiac experts giving feedback during the process. On interviews with parents of children with a rare coronary anomaly it achieved higher hit rate, coverage and distinctiveness than single-agent analysis.
This paper explores personality-based prompting as a mechanism to capture the diversity of human reasoning in LLMs. By combining psychological personality models with genetic algorithms, the approach enables LLMs to approximate human response distributions across intuitive and deliberative reasoning tasks.
Integrates Parsl into LangChain and LangGraph tool nodes so an agent's tool calls are dispatched concurrently to workers on a workstation or supercomputer. The agent was prompted to run molecular dynamics simulations of different protein structures and conditions, which executed in parallel on the available resources.
This survey provides a comprehensive review of foundation models (FMs) in medicine, covering their evolution, learning strategies, flagship models, applications, and challenges across healthcare. It systematically examines how models such as BERT and GPT are transforming clinical NLP, medical imaging, and omics research, and proposes a detailed taxonomy of FM-enabled medical applications. The survey also discusses key limitations, open research questions, and lessons learned to support responsible and effective deployment of FMs in healthcare.
This paper evaluates the ability of leading large language models to generate preschool-appropriate science explanations. Through expert teacher assessments, it highlights both the promise and limitations of LLMs in supporting early childhood science education.
This study explores attitudes toward using the Gia® genetic information chatbot to support APOL1 testing among living kidney donor candidates of African ancestry. Through focus groups and surveys involving donors, recipients, and community members, participants generally supported using the chatbot prior to clinical visits and expressed interest in APOL1 testing, though concerns about cost, personal preferences, and individual barriers emerged. Overall, participants were receptive to integrating the APOL1 chatbot into the donor evaluation workflow. The findings suggest that culturally adapted chatbots can facilitate genetic testing discussions and integration in clinical care.
This editorial compares OpenAI o1-Preview with ChatGPT-4 in healthcare applications, highlighting advances in reasoning capabilities and their implications for clinical decision support. The work discusses opportunities and limitations of next-generation generative AI models in medical reasoning contexts.
This study examines how facial expressions influence AI and human diagnosis of genetic syndromes using facial images. Both deep learning models and clinicians were more likely to correctly identify Williams and Angelman syndromes when individuals were smiling, revealing a bias linked to stereotypical expressions, while no such effect was observed for other syndromes. The findings highlight facial expression as an important confounder that should be considered and mitigated in AI driven genomic and clinical diagnostics.
This study compares the treatment recommendations of five publicly available large language models (GPT-4, GPT-3.5 variants, Llama2, and Bard) with those of a multidisciplinary tumor board for complex breast cancer cases. GPT-4 showed the highest concordance, particularly for invasive breast cancer and radiotherapy decisions, while all models performed poorly in genetic testing recommendations. Overall, the findings indicate that current LLMs are not yet sufficiently reliable for safe clinical use in breast cancer care, highlighting the need for improved prompting strategies, data control, and further clinical validation.
This study evaluates GPT-4’s ability to deliver genetic information on BRCA1, HFE, and MLH1 compared with GPT-3.5 using expert review by genetic counselors and clinicians. GPT-4 showed significantly improved accuracy and generally favorable relevance, but persistent errors and usability limitations remained. The findings highlight GPT-4’s promise in genetic education while emphasizing the need for continued refinement and ethical oversight in clinical use.
This review examines the role of AI, particularly ChatGPT, in advancing research, education, and clinical practice for rare and genetic disorders. It highlights the potential of AI to improve personalized patient care while critically discussing current limitations and risks. The study outlines future directions for developing AI tools that can better support healthcare professionals and patients in rare disease management.
This study evaluates multiple large language models for phenotype driven gene prioritization in rare disease diagnosis across different prompts, input types, and task settings. Although GPT 4 performed best, its accuracy remained substantially lower than that of traditional gene prioritization tools, with limited benefit from retrieval augmentation or few shot learning. The findings highlight both the potential and current limitations of LLMs in clinical genomic workflows, including biases toward well studied genes.
This paper argues that existing automatic karyotype analysis methods are limited by treating the task purely as visual object recognition, neglecting holistic information and interpretability needed for clinical use. The authors introduce Tokensome, a vision–language model that represents chromosomes through tokenization to enable explainable, cognitively driven karyotyping. By elevating analysis from visual perception to cognitive decision-making, Tokensome integrates domain knowledge and reasoning via knowledge graphs and LLMs. This approach improves interpretability and enhances the detection of chromosomal abnormalities.
This work presents GeneGPT, a framework that enables large language models to answer genomics questions by directly invoking NCBI Web APIs. By learning to generate and execute API calls through in-context learning and augmented decoding, GeneGPT substantially reduces hallucinations and achieves state-of-the-art performance on the GeneTuring benchmark. The results demonstrate that tool-augmented LLMs can reliably perform complex and multi-hop reasoning in genomics tasks.
This review examines the evolution of foundation AI models in healthcare, highlighting their growing impact on diagnostics, medical interventions, and precision medicine across domains such as imaging, genomics, and decision support. It discusses both the transformative potential of generative and traditional AI models and the challenges of slow adoption, ethics, bias, and governance in healthcare systems. The authors advocate for transparent, fair, and accountable AI frameworks, including the concept of General Medical AI, to enable scalable and responsible healthcare innovation.
This paper examines the limitations of using plain ChatGPT as a bioinformatics assistant, highlighting errors caused by incomplete knowledge, poor data retrieval, and hallucinations in scientific domains. To address these issues, the authors introduce NagGPT, a middleware that sits between LLMs and genomics databases to validate, correct, and mediate LLM-generated queries while guiding responses with authoritative, up-to-date data. Paired with a custom GPT that generates and executes Python code, the system enables dynamic retrieval and analysis across major genomics databases. The results demonstrate practical ways to improve factual accuracy, reduce hallucinations, and enhance the reliability of unmodified LLMs for bioinformatics tasks.
This paper introduces **RNA-GPT**, a multimodal RNA-focused language model designed to support RNA research by integrating RNA sequence encoders with large language models. RNA-GPT aligns RNA sequence representations with textual knowledge, enabling it to process user-provided RNA sequences and answer complex RNA-related queries accurately. The system is trained using RNA-QA, an automated pipeline that extracts and organizes RNA annotations from RNACentral to generate large-scale instruction-tuning data. Experiments show that RNA-GPT effectively facilitates RNA discovery, supported by a curated dataset of over 400,000 RNA samples for modality alignment and instruction tuning.
This work introduces GP-GPT, a specialized large language model fine-tuned on large scale genomics, proteomics, and medical genetics data for genetic–phenotype knowledge representation and relation analysis. GP-GPT outperforms general purpose LLMs on genomics information retrieval and relationship determination tasks, demonstrating strong domain specific understanding. The results suggest that tailored LLMs can more effectively support gene–phenotype research and biomedical knowledge discovery.
This work introduces VarChat, a generative-AI–based tool designed to address the growing challenge of genomic variant interpretation amid rapidly expanding scientific literature. VarChat automatically identifies, synthesizes, and summarizes variant-related research into concise, clinically relevant descriptions, linking genetic variants to protein function and health outcomes while providing traceable references. By streamlining literature review and interpretation, VarChat helps translate fragmented genomic knowledge into actionable insights for researchers and clinicians.
ChIP-GPT presents a fine-tuned LLM system for scalable and robust metadata extraction from biomedical databases. By iteratively prompting a customized GPT model, the framework accurately extracts ChIP-seq targets and cell lines from noisy or incomplete records, enabling expert-level reasoning over large-scale database entries.
This work introduces FAVOR-GPT, a retrieval augmented generative interface that integrates large language models with the FAVOR database to support interpretation of whole genome variant annotations. By translating complex functional annotations into user friendly summaries, FAVOR-GPT improves accessibility and usability for researchers without specialized expertise. The results demonstrate that RAG based LLM interfaces can effectively enhance genomic data interpretation while maintaining high annotation accuracy.
This study evaluates transformer-based language models for extracting protein–protein interactions from biomedical text. Bidirectional encoder models, especially BioBERT, achieved the strongest overall performance, while GPT-4 demonstrated comparable precision and recall despite not being specifically trained on biomedical corpora. The results indicate that large language models can effectively support automated PPI mining from the biomedical literature.
Designing transition metal complexes involves searching an enormous chemical space, which limits the effectiveness of traditional genetic algorithms. This work introduces an LLM driven evolutionary optimization framework that leverages pretrained chemical knowledge to guide single and multi objective optimization without supervised fine tuning. The proposed approach efficiently identifies top performing complexes, enables flexible natural language driven optimization, and demonstrates strong potential for chemistry and materials design.
Large language models show strong few-shot learning ability, but their capacity to generalize to complex biological problems remains underexplored. This study proposes a few-shot LLM-based approach for predicting drug pair synergy in rare cancer tissues by leveraging prior knowledge from text rather than structured data. Experiments across seven rare tissues demonstrate that the compact CancerGPT model achieves high accuracy with minimal or no samples, performing comparably to much larger fine-tuned GPT-3 models.
BioDiscoveryAgent is an LLM-based autonomous agent that designs and evaluates genetic perturbation experiments by reasoning over biological knowledge without requiring explicit optimization objectives or model training. Across multiple datasets, it substantially outperforms Bayesian optimization baselines in identifying relevant genes and gene combinations, including challenging non-essential perturbation tasks and unseen data. By integrating literature search, data analysis, and self-critique, BioDiscoveryAgent offers an interpretable and efficient paradigm for closed-loop biological experiment design.
This work introduces PheNormGPT, a large language model–based framework for extracting and normalizing phenotypic findings from unstructured clinical text to Human Phenotype Ontology concepts. Using fine-tuned GPT models and a novel adaptive few-shot selection strategy, PheNormGPT achieved state-of-the-art performance in the BioCreative VIII phenotype extraction challenge. The results demonstrate the effectiveness of LLM-driven approaches for accurate clinical phenotype normalization.
This paper introduces CellAgent, an LLM-driven multi-agent framework that fully automates single-cell RNA-seq data analysis by coordinating planner, executor, and evaluator agents with hierarchical decision making and self-iterative optimization. CellAgent autonomously selects appropriate tools and hyperparameters, executes end-to-end analyses, and evaluates results without human intervention. Benchmarks across diverse tissues and cell types show that CellAgent achieves robust, high-quality performance while substantially reducing analyst workload, advancing an “Agent for Science” paradigm.
This work introduces SteLLA, a structured grading framework that combines retrieval-augmented generation with large language models to improve the reliability of automated short answer grading. By grounding evaluation in instructor-provided reference answers and rubrics, SteLLA achieves strong agreement with human graders while producing detailed, point-by-point grades and feedback. The analysis shows that LLMs accurately capture factual content but may over-infer implications, offering insights into safe and effective deployment for educational assessment.
The paper proposes ERFSL, a framework that leverages large language models to search and refine reward functions for complex multi-objective reinforcement learning tasks. By combining LLM-driven semantic reasoning with evolutionary strategies, the method achieves efficient zero-shot reward design.
The paper characterizes age-related hematological and biochemical changes across three commonly used male rat strains. It highlights strain-specific physiological differences relevant to aging and geriatric research, emphasizing the importance of strain selection in experimental design.
This work introduces SeqMate, a large language model–powered tool that enables one-click RNA-seq and single-cell RNA-seq analysis for non-expert users. By automating data preparation, downstream analyses, and result interpretation with literature-linked reporting, SeqMate removes technical barriers inherent to command-line pipelines and lowers the entry threshold for biological data analysis.
This survey reviews the rapid expansion of ChatGPT applications in bioinformatics and biomedical informatics during 2023. It covers a broad range of areas, including omics and genetics, biomedical text mining, drug discovery, image understanding, bioinformatics programming, and education. The authors analyze both the strengths and limitations of ChatGPT in these domains and discuss key challenges. The survey also outlines promising directions for future development of large language models in bioinformatics.
This study introduces GP4NLDR, an explainable AI framework that integrates genetic programming–based nonlinear dimensionality reduction with a large language model–powered chatbot. By combining an XAI dashboard with LLM-generated explanations, the system provides intuitive, user-centered interpretations of high-dimensional data transformations. The results demonstrate the potential of LLMs to enhance the explainability and usability of genetic programming methods.
This article analyzes how U.S. and Chinese regulatory frameworks create barriers to collaboration in generative-AI–assisted genomic research, particularly around data sharing, security, and compliance. It highlights tensions between innovation and regulation in both countries and argues that misaligned legal requirements hinder responsible international collaboration. The authors recommend renewing and updating bilateral science and technology agreements to promote transparent, ethical, and secure sharing of human genomic data.
This study evaluates four large language models as assistants in precision oncology by comparing their treatment recommendations with those of human experts. While the models produced more candidate treatment options, their agreement with expert decisions was low due to limited supporting evidence, though some clinically useful and unique suggestions emerged. The findings suggest LLMs may support idea generation and literature screening but cannot replace expert judgment.
This study examines whether large language models (LLMs) can capture the full spectrum of human reasoning, including both fast, intuitive System 1 and slow, deliberative System 2 processes. Using a generalized Natural Language Inference task and crowdsourced human response distributions, the authors evaluate LLMs’ ability to reproduce not just majority answers but the full diversity of human judgments. They introduce personality-based prompting inspired by the Big Five model, optimized with genetic algorithms, to elicit diverse reasoning styles from LLMs. Results show that this approach significantly improves alignment with human response distributions, with open-source models outperforming proprietary ones, suggesting that modeling personality and suboptimal reasoning is key to more human-like AI cognition.
Large language models are accelerating and democratizing research by providing accessible expertise across disciplines, but they also risk lowering barriers to dual use technologies with serious biosecurity implications. An educational experiment showed that, with minimal prompting, chatbots could guide non experts through key steps related to pandemic scale agents, indicating that such capabilities may become broadly accessible. These findings highlight the urgent need for safeguards such as third party model evaluation, careful curation of training data, and comprehensive DNA synthesis screening to mitigate misuse risks.
This work demonstrates how GPT-4–based prompt engineering can streamline knowledge extraction from synthetic biology literature. By mining 176 publications on oleaginous yeasts, the authors converted unstructured text into structured datasets that enabled machine-learning models to accurately predict fermentation performance. The results show that generative AI can significantly reduce manual curation effort, support predictive modeling in biomanufacturing, and enable transfer learning to less-studied microbial systems.
This paper presents Coscientist, a GPT-4–driven AI system that autonomously designs, plans, and executes complex scientific experiments by integrating large language models with tools for literature search, code execution, and experimental automation. Across six representative tasks—including palladium-catalyzed reaction optimization—Coscientist demonstrates strong capabilities in semi-autonomous experimental design, execution, and reasoning. These results highlight the versatility, effectiveness, and interpretability of LLM-powered AI systems in accelerating scientific discovery.
The study investigates the genetic architecture of milk citrate, an early biomarker of negative energy balance in dairy cows, using large-scale mid-infrared spectral predictions and single-step GWAS. Results show high heritability and identify key genomic regions and candidate genes relevant for genomic selection.
This study explores the role of ChatGPT in discussing the future of personalized medicine and pharmacogenomics, which aim to tailor treatments based on individual genetic and molecular profiles. By querying ChatGPT about these fields, the authors found that the model portrays them as promising approaches with strong potential to improve patient outcomes through algorithm-driven, individualized therapies. However, the responses also highlight current limitations and challenges that must be addressed before such AI-guided personalization can be fully realized in clinical practice.
This work introduces AutoBA, an autonomous large language model–based agent that streamlines conventional omics data analysis with minimal user input. Validated by expert bioinformaticians, AutoBA demonstrates robust performance and adaptability across diverse modalities, including WGS, RNA-seq, single-cell, ChIP-seq, and spatial transcriptomics, while preserving data privacy through local execution. The system’s ability to dynamically design analysis workflows highlights its potential as a flexible alternative to predefined bioinformatics pipelines.
This genome-wide association study identifies SNPs and candidate genes linked to lactation traits in Chinese Holstein cows raised under high-temperature and high-humidity conditions. The findings provide genetic insights to support breeding programs adapted to challenging environmental stressors.
This study evaluates whether generative AI–created images can support medical education by helping pediatric residents recognize genetic syndromes. While AI-generated images performed comparably to real images, they were perceived as slightly less helpful, with real images remaining the preferred learning tool. The findings suggest that generative AI images can serve as a useful adjunct, particularly for teaching less familiar genetic conditions.
This study analyzes how different concentrate levels in early-lactation diets affect hepatic transcriptomic responses in dairy cows. Results reveal parity-specific metabolic and inflammatory responses, offering insights into nutritional strategies for mitigating negative energy balance.
This paper introduces *mergen*, an R-based system that uses LLMs to automate data analysis workflows from natural language descriptions. The framework combines prompt engineering, execution feedback, and workflow automation to enable non-experts to perform complex biological data analyses while exposing current limitations of LLM-generated code.
This work evaluates the symbol binding capabilities of LLMs on Vietnamese multiple-choice questions using a newly constructed LaTeX-based dataset. The study assesses zero-shot to few-shot performance across several models, providing insights into reasoning efficiency and multilingual evaluation under strict output constraints.
This study develops PhenoBCBERT and PhenoGPT, two large language model–based approaches for automated phenotype recognition in clinical notes by expanding Human Phenotype Ontology vocabularies. The models outperform traditional rule-based tools by identifying a broader and more nuanced range of phenotype concepts, including previously uncharacterized terms, and demonstrate strong performance across clinical text and biomedical literature. These results highlight the effectiveness of adapting general-purpose LLMs for domain-specific clinical and genetic research.
This article provides ten practical guidelines for using ChatGPT and GPT-4 in computational biology workflows, covering code generation, data analysis, scientific writing, and prompt engineering. It discusses both productivity gains and ethical considerations for bioinformatics applications.