PaperlayerView exact revision on arXiv

Source-linked AI summary

GenoMAS: Guided Multi-Agent Programming for Scientific Discovery from Genomic Data

Haoyang Liu, Yijiang Li, Haohan Wang

arXiv:2507.21035v4cs.AIcs.LGcs.MAq-bio.GN

TL;DR

Scientific agents that handle documents, APIs, or abstract planning fall short in code-centered workflows requiring adaptation to dataset and experimental details. GenoMAS uses guided planning and specialized LLM agents to generate and refine analysis code; it outperforms strong agent baselines on two independent bioinformatics benchmarks and reaches 60.48% F1 on GenoTEX gene-trait association analysis.

  • Problem

    LLM agents’ document/API and abstract-planning capabilities fall short in scientific workflows where code must be tailored to dataset, model, and experimental-design idiosyncrasies.

    Source: §1 · p. 2

  • Method

    GenoMAS treats agents as collaborative programmers that generate, revise, and validate task-specific code, with guided planning organized around editable Action Units.

    Source: §1 · p. 3

  • Results

    GenoMAS outperforms strong agent baselines on ScienceAgentBench and DiscoveryBench and reaches 60.48% F1 on GenoTEX gene-trait association analysis, a 16.85% absolute gain over prior work.

    Source: §7 · p. 16

  • Takeaways & Limitations

    GenoMAS outputs include biologically plausible gene-phenotype associations corroborated by the literature.

    Source: §7 · p. 16

    Clinical feature extraction remains the principal open challenge, with a meaningful gap between GenoMAS and expert curation.

    Source: §6 · p. 15

Abstract

from arXiv · show

Gene expression analysis holds the key to many biomedical discoveries, yet extracting insights from raw transcriptomic data remains challenging due to the complexity of multiple large, semi-structured files and the need for extensive domain expertise. Current approaches either rely on inflexible workflows that fail in edge cases or fully autonomous agents that lack the precision required for rigorous scientific inquiry. GenoMAS addresses this by presenting a team of LLM-based agents that integrates structured workflows with autonomous adaptability, orchestrating diverse specialized agents through typed message-passing protocols. At its core lies a guided-planning framework where programming agents decompose task guidelines into Action Units and dynamically choose to advance, revise, bypass, or backtrack, maintaining logical coherence while adapting to the idiosyncrasies of genomic data. On the GenoTEX benchmark, GenoMAS achieves 89.13% CSC for data preprocessing and 60.48% $F_1$ for identifying trait-associated genes, surpassing prior art by 10.61% and 16.85%. Beyond transcriptomics, it outperforms strong agent baselines on ScienceAgentBench and DiscoveryBench, two independent bioinformatics benchmarks. Qualitative analysis shows that GenoMAS identifies gene-phenotype associations that are biologically plausible and corroborated by the literature.

1 Introduction

GenoMAS frames scientific agents as collaborative programmers that generate, revise, and validate task-specific code, addressing the gap between general agent reasoning and the precision required for scientific computation. Applied to gene expression analysis, it achieves state-of-the-art performance on GenoTEX and succeeds on ScienceAgentBench and DiscoveryBench.

Introduction

  • Scientific workflows require editable code to handle dataset-specific transformations, assumptions, and experimental designs, so planning alone cannot automate them reliably.Analysts adapt scripts for malformed metadata, confounding, deprecated identifiers, batch effects, and platform noise.Source: §1 · p. 2
  • Existing agent systems often compose tool calls or fixed modules, while lacking structured code revision, validation, and domain-aware correction to detect or recover from scientific errors.Small preprocessing or model-selection errors can invalidate gene-expression results.Source: §1 · pp. 2–3
  • GenoMAS achieves state-of-the-art performance on GenoTEX from raw data inputs and also succeeds on ScienceAgentBench and DiscoveryBench.GenoTEX includes 1,384 gene-trait association tasks spanning heterogeneous cohorts, evolving metadata, and statistical confounders.Source: §1 · p. 3
  • GenoMAS treats agents as collaborative programmers that generate, revise, and validate executable code tailored to scientific tasks, rather than as tool orchestrators.The framework automates gene expression workflows that previously required domain experts to script by hand.Source: §1 · p. 3
  • GenoMAS guides planning with editable Action Units that can be refined into executable code, balancing precise control with autonomous error handling.Its coordinated roles can combine distinct LLM strengths, including code synthesis, language reasoning, and scientific review.Source: §1 · p. 3

2 Related Work

Prior work spans reasoning and tool-using LLM agents, multi-agent collaboration, and scientific discovery applications. GenoMAS addresses a gap in systems that write and revise executable code under scientific constraints.

LLM-based Agents

  • LLM agents use structured reasoning and tools to execute complex tasks, with methods spanning sub-goal decomposition, tree or graph search, self-refinement, consistency checks, reusable skills, and knowledge-base integration.These approaches aim to turn LLMs from passive text generators into active problem-solving agents.Source: §2 · p. 3

Multi-Agent System

  • Multi-agent systems coordinate specialized roles through structured protocols, combining collaboration with planning and feedback, while research also examines their social behaviors, failures, and training.Examples include role-based agents for software development [36] [55] [114] and simulated settings showing theory-of-mind-like reasoning [76] [112] [129] [140] [197].Source: §2 · pp. 3–4

LLM Agents for Scientific Discovery

  • Scientific discovery agents support stages from hypothesis generation and research assistance to materials discovery, with domain-specific applications across mathematics, physics, chemistry, biology, and medicine.AI Scientist [92] completes a research cycle, while ResearchAgent [11] iteratively refines research problems, methodologies, and experimental designs using scientific literature.Source: §2 · p. 4

Positioning of Our Work

  • GenoMAS targets executable scientific coding, adding Action-Unit guided advance, revise, skip, or retract planning and cross-family role specialization to bioinformatics agents’ planner-executor-reviewer pattern.Retraction restores task context and execution state to an earlier step when downstream evidence exposes a problematic decision; the authors position the system between rigid workflows and fully autonomous agents.Source: §2 · p. 4

3 Preliminaries

The preliminaries define gene-trait association (GTA) analysis, its unconditional and conditional forms, and the GenoTEX evaluation framework. They also outline the technical challenges that make robust GTA analysis difficult.

3.1 Gene Expression Data Analysis

  • Gene expression analysis uses transcriptomic data to investigate cellular function and disease, with GTA as a central task.High-throughput microarrays and RNA sequencing generate datasets held in repositories such as GEO and TCGA.Source: §3.1 · p. 4

3.2 Problem Formulation

  • GTA identifies genes whose expression correlates with a phenotype, either without additional conditions or while accounting for a condition variable.Conditional analysis aims to identify associations independent of the condition’s influence.Source: §3.1 · p. 4, §3.2 · p. 5

3.3 Evaluation

  • GenoTEX benchmark [86] standardizes evaluation across the gene-expression analysis pipeline’s three core tasks, assessing dataset relevance and expert-aligned selection, while preprocessing quality is measured with AJ, SJ, and CSC.Dataset Filtering measures binary classification accuracy for relevance; Dataset Selection checks whether the chosen dataset matches domain experts’ choice among candidates. Preprocessing metrics are Attribute Jaccard (AJ), Sample Jaccard (SJ), and CSC = AJ × SJ × Corravg, with Corravg the average Pearson correlation between common features.Source: §3.3 · p. 5

3.4 Technical Challenges

  • GTA is complicated by high-dimensional sparsity, platform heterogeneity, evolving gene nomenclature, heterogeneous metadata, and confounding factors [27].Datasets often contain over 20,000 genes measured across fewer than 1,000 samples; platform-specific processing, metadata extraction, and control of batch effects, population stratification, and hidden covariates are also necessary.Source: §3.4 · pp. 5–6

4 Method: GenoMAS

GenoMAS combines structured workflows with agent autonomy to handle genomic-data edge cases while preserving precision. Its specialized agents use guided planning, iterative code review, reusable code memory, and system-level optimizations to support gene-expression analysis.

4.1 Overview

  • GenoMAS addresses the tension between rigid workflows and autonomous agents by combining procedural control with adaptive handling of genomic-data edge cases.The paper attributes the challenge to noisy, high-dimensional data and domain-specific decisions that affect downstream statistics.Source: §4.1 · p. 6

4.2 Multi-Agent Architecture

  • Six specialized agents divide orchestration, preprocessing and statistical analysis, code review, and biomedical guidance across the workflow.The PI agent schedules work based on dependencies, reuses completed cohort preprocessing, and coordinates structured request-response messages that advance or revise tasks.Source: §4.2 · pp. 6–7

4.3.1 Guided Planning

  • Programming agents translate task guidelines into DAG-structured Action Units, such as data loading, feature extraction, annotation, and normalization.At each step, they use task history, current state, data characteristics, and remaining objectives to advance, revise, skip optional steps, or terminate; retraction restores prior context and execution state for backtracking.Source: §4.3.1 · pp. 7–8

4.3.2 Domain Specific Code Generation

  • For each Action Unit, programming agents iteratively write and revise code using execution history and reviewer feedback until approval or a debugging limit.Reviewers assess outputs, errors, and guideline adherence; they receive no feedback from earlier review rounds at the same step, while the programming agent can use prior attempts and feedback to revise.Source: §4.3.2 · p. 8
  • For biomedical decisions, agents consult a Domain Expert that returns executable guidance from targeted biological context and helps debug failures across refinement rounds.Programming agents also store approved code snippets by Action Unit type for reuse and adaptation; specialized Python tools and local gene-synonym and gene-trait databases support domain-specific operations.Source: §4.3.2 · pp. 8–9

4.4 System Implementation and Optimizations

  • System optimizations use asynchronous calls, memory-efficient processing, resumable workflows, resource monitoring, caching, and distributed scheduling to support large-scale analysis.The paper describes these measures as addressing efficiency, robustness, and scalability while preserving adaptability for research-oriented exploration.Source: §4.4 · p. 9

5 Experiments

Experiments evaluate GenoMAS on GenoTEX and two independent bioinformatics benchmarks, reporting strong end-to-end performance and contributions from guided planning, specialized agents, code review, and code reuse.

5.1 Experimental Setup

  • GenoTEX [86] covers 1,384 problems across 913 datasets and 132 human traits; evaluation uses F1 as the primary gene-identification metric, alongside accuracy, runtime, and efficiency measures.The benchmark spans the workflow from raw data to biological insights and uses real genomic datasets with professionally curated ground truth.Source: §5 · p. 9, §5.1 · p. 10

5.2 Comparison with Prior Art

  • GenoMAS achieved 60.48% F1, 16.85% absolute above GenoAgent, with AUROC 0.81 (a 0.16 increase) and approximately 45% lower API cost.It scored 59.86% F1 on conditional and 66.37% on unconditional tasks, retaining 84% of the 71.63% human-analysis F1. Heterogeneous LLMs added 7.5% F1 and cut cost approximately 48% versus homogeneous Claude-only GenoMAS.Source: §5.2 · p. 10, Table 1 · p. 11
  • Claude Code reached 8.53% F1 and 12.14% success with Sonnet 4 versus homogeneous GenoMAS at 52.98% and 96.34%; Sonnet 5 reached 99.42% success but 47.34% F1, while Biomni scored 14.82% F1.The Sonnet 4 comparison attributes the 44-point F1 gap to the GenoMAS scaffold; the results suggest execution completeness alone does not close its analysis-accuracy gap.Source: §5.2 · pp. 10–12

5.3 Ablation Study

  • Removing guided planning lowered F1 by 9.2 points, removing the Domain Expert by 12.9 points, and removing code review reduced success to 56.23% and F1 to 24.98%.Without planning, cost rose from $0.38 to $0.48 and runtime from 307 to 360 seconds; a single review round recovered 91.76% success but left 13.9 F1 points unrecovered.Source: §5.3 · p. 12
  • Gene-expression preprocessing scored 91.15% CSC versus 32.61% for clinical traits; with expert-preprocessed inputs, batch-corrected GenoMAS reached 95.26% F1 versus Lasso’s 14.03%.Clinical feature preprocessing is the main preprocessing bottleneck; the statistical comparison isolates analysis using expert-preprocessed data, and highlights batch-effect handling and covariate adjustment.Source: Table 2 · p. 11, §5.3 · p. 12, Figure 3 · p. 13

5.4 Generality Beyond Transcriptomics

  • On ScienceAgentBench, GenoMAS achieved 63.0% success (17/27), versus 44.4% for HAL SAB Self-Debug, 37.0% for Direct, and 29.6% for Self-debug.Paired McNemar tests found significant gains over Direct (p = .039), Self-debug (p = .004), and the HAL generalist agent (p = .002), but not over HAL SAB Self-Debug (p = .227).Source: Table 3 · p. 13, §5.4 · p. 13
  • On 11 discovery-requiring DiscoveryBench tasks, GenoMAS scored 29.4 HMS versus 13.6 for Self-debug and 8.1 for the benchmark coder; removing planning reduced it to 11.1, while removing review yielded 30.3.Nine tasks that embed answers in the query were excluded; the reported results indicate guided planning drove the margin on discovery tasks.Source: Table 3 · p. 13, §5.4 · p. 14

5.5 Memory and Code Reuse Efficiency

  • In a tracked session of 50 cohort datasets, code memory saved 57.8 minutes, averaging 20.3 seconds per reused programming step, and reuse stabilized around 65%.Reusable snippets captured recurring steps such as loading GEO files, mapping gene symbols, and normalizing expression values; this diagnostic session is not a benchmark-wide efficiency estimate.Source: §5.5 · p. 14, Figure 4 · p. 14

6 Qualitative Analysis

GenoMAS produces biologically plausible gene-trait findings and adapts to workflow problems, while clinical trait extraction remains its principal open challenge. Its communication patterns show execution concentrated in programming agents, with limited orchestrator intervention.

GenoMAS identifies biologically plausible gene-trait associations

  • A 0.53 GSEA enrichment score (FDR q-value 0.16) indicates moderate concordance with Open Targets disease associations, without circularity in the unconditional problems examined.GenoTEX reference outputs were curated independently of Open Targets, and GenoMAS consults that platform only for missing condition variables in conditional problems.Source: §6 · p. 14
  • High-confidence predictions include established drivers such as TERT in lung cancer and plausible candidates linked to GABAergic dysfunction in Alzheimer’s disease and vascular tone in hypertension.Exploratory per-gene literature evidence supports using GenoMAS outputs for hypothesis generation in genomic medicine.Source: §6 · pp. 14–15

Autonomous agent behaviors enhance workflow robustness

  • Agents autonomously add diagnostic code to debug persistent errors and can rewrite downstream code after an early decision causes complications, improving workflow robustness without manual intervention.The behaviors are documented in Appendix H.Source: §6 · p. 15
  • Severity-categorized post-task notes document challenges, anomalies, and potential issues so human experts can audit high-priority errors and support continuous improvement.Notes are categorized as info, warning, or error; representative examples appear in Appendix J.Source: §6 · p. 15

Clinical feature extraction is the principal open challenge

  • Clinical metadata-to-trait extraction remains the principal challenge: despite leading evaluated systems, GenoMAS still trails expert curation, with errors involving invalid contrasts, irrelevant cohorts, covariate entanglement, and occasional synthetic labels.Preprocessing almost always completes; proposed safeguards include case-control validity and cohort relevance checks, plus a hard guard against synthetic labels.Source: §6 · p. 15

Collaboration patterns reflect role specialization and light-touch orchestration

  • In a representative 20-problem session, the Data Engineer handled 56.9% of interactions, while 97.7% of post-dispatch traffic was agent-to-agent coordination rather than top-down control.Planning requests and responses were most frequent at 634 each, followed by 351 code-validation requests and 36 revision requests; the authors interpret these patterns descriptively.Source: §6 · p. 15

7 Conclusion … A.7.1 GSE117261: Complete Failure in Probe-to-Gene Mapping

The section summarizes GenoMAS’s guided planning, complementary LLM backbones, reusable code memory, benchmark results, and potential generalization beyond transcriptomics, while noting limitations including clinical feature extraction. Its examples show failures in essential preprocessing and phenotype handling, as well as a failure to adapt reasoning to dataset-specific context.

7 Conclusion

  • The paper attributes GenoMAS’s approach to guided planning, diverse LLM strengths, and domain-informed programming, and identifies clinical feature extraction as a remaining gap.Future work includes multimodal biological data integration, richer cross-task memory, more sophisticated planning, and safeguards for clinical feature extraction.Source: §7 · pp. 15–16
  • GenoMAS achieves 60.48% F1 on GenoTEX gene-trait association analysis, a 16.85% absolute gain over prior work at approximately 45% lower API cost, and outperforms strong agent baselines on ScienceAgentBench and DiscoveryBench.Its reported gene-phenotype associations are biologically plausible and literature-corroborated.Source: §7 · p. 16

A Fundamental Limitations of Existing Autonomous Agent Methods

  • Across six tested autonomous agents, Biomni reached 14.82% F1, while Claude Code achieved 12.14% execution success across 1,384 benchmark problems; the other five did not produce complete, valid preliminary analyses.The agents were tested with Claude Sonnet 4, and most did not support heterogeneous multi-model configurations.Source: A Fundamental Limitations of Existing… · pp. 28–29
  • The analysis identifies inadequate interpretation of complex domain guidelines and lack of workflow-level planning or error recovery as key causes of invalid agent outputs.These limitations motivate user-specified, domain-trusted guidelines and precise procedural control for rigorous gene expression analysis.Source: A Fundamental Limitations of Existing… · p. 29

A.2.1 GSE77627: Misinterpretation of Portal Hypertension Phenotypes

  • Biomni misclassified portal-hypertension phenotypes in GSE77627, producing 18 cases and 36 controls instead of the benchmark’s 40 cases and 14 controls.It also omitted gene-symbol standardization and systematic bias checks, compounding the analysis errors.Source: §A.2.1 · pp. 30–31

A.3 AutoGen Summary

  • AutoGen [166] and MetaGPT [55] failed to complete benchmark-valid GEO preprocessing, showing that conversational coordination or role-based procedures alone did not ensure required biological and technical checks.Their runs are presented as qualitative evidence, not quantitative end-to-end preprocessing baselines.Source: §A.3 · p. 31, §A.4 · pp. 31–32

A.5 Claude Code Examples

  • Despite detailed guidelines, Claude Code treated liver cirrhosis (LC) as separate from INCPH instead of grouping them under the portal-hypertension label, and retained raw GSE117261 probe identifiers without gene-symbol mapping.Source: §A.5.1 · pp. 32–33, §A.5.2 · p. 33

A.6 Cursor Examples

  • Under the general prompt, Cursor omitted probe-to-gene mapping in GSE117261, leaving numeric probe IDs; separately, when relying on individual IDs alone, it missed explicit IPAH/control labels in sample-level fields.Source: §A.6.1 · p. 34, §A.6.2 · pp. 35–36

A.6.2 GSE256539: Misinterpretation of Clinical Phenotype Encoding

  • Despite detailed instructions to examine data carefully and incorporate domain knowledge, Cursor relies on generic keyword matching rather than context-specific pattern recognition.The passage identifies this as evidence of lacking adaptive reasoning for scientific data analysis.Source: §A.6.2 · p. 36

A.7 Manus Examples

  • Manus, presented as a general-purpose task-automation system [130], lacks the domain understanding and procedural control needed for bioinformatics data processing.The paper says its examples illustrate critical errors under both prompt conditions.Source: §A.7 · p. 36

A.7.1 GSE117261: Complete Failure in Probe-to-Gene Mapping

  • Manus leaves GSE117261 expression values indexed by numeric probe identifiers rather than gene symbols, making the dataset unusable for meaningful biological analysis.Its processing also omits essential bioinformatics steps autonomously, preventing biological interpretation and cross-study comparison.Source: §A.7.1 · pp. 36–37
  • Manus’s narrow trait-mapping logic fails to consistently use clinical_group and subtype, risking data loss and PAH case-control misclassification.The dataset’s metadata requires consistent use of both fields to distinguish PAH cases from donor/control samples.Source: §A.7.1 · p. 37
  • The human expert workflow extracts SOFT-file annotations, maps probe IDs to gene symbols, applies the mapping, and normalizes symbols using the NCBI Gene database.These steps contrast with Manus’s omitted processing requirements.Source: §A.7.1 · p. 37

A.7.2 GSE77627: Misalignment with the Benchmark Phenotype Definition

Under detailed guidelines, Manus misclassifies portal-hypertension samples and makes additional probe-mapping and quality-control errors, illustrating persistent limits in autonomous bioinformatics analysis.

A.7.2 GSE77627: Misalignment with the Benchmark Phenotype Definition

  • Manus labels liver cirrhosis as control, yielding 18 INCPH cases and 36 controls versus the benchmark’s 40 cases and 14 controls.Liver cirrhosis is the most common cause of portal hypertension and accounts for approximately 90% of cases in Western countries.Source: §A.7.2 · pp. 37–38
  • This case-control reversal means Manus would identify genes distinguishing INCPH from mixed healthy and cirrhotic livers, rather than genes associated with portal hypertension.Source: §A.7.2 · p. 38
  • Manus also mishandles many-to-one probe-gene mapping and mapped-data integration, while omitting gene-symbol normalization and systematic bias detection.Source: §A.7.2 · p. 38
  • These persistent phenotype-labeling and preprocessing errors, even with detailed prompts, show that agent architecture cannot replace domain understanding and precise procedural control.Source: §A.7.2 · p. 39

B Protocols for the Additional Experiments … C.5 Agent Communication Patterns

The appendices specify controlled protocols and evaluation caveats for additional benchmark experiments, then describe GenoMAS’s typed, topology-constrained messaging and agent coordination patterns. Together, they explain how comparisons are conducted and how agents plan, execute, and review work.

B Protocols for the Additional Experiments; B.4 Inference Configuration

  • Controlled comparisons fix the code-generating model, and no agent prompt or predicted output contains gold answers, gold result files, or answer lists.Across experiments, each LLM API uses its default decoding settings; GenoTEX configurations are single runs over 1,384 problems without confidence intervals, while ScienceAgentBench comparisons include paired significance tests.Source: B Protocols for the Additional… · p. 39, §B.4 · p. 40

B.1 General-Purpose Coding Agent on GenoTEX

  • Claude Code’s trait-prediction accuracy is 9.40% with Sonnet 4 and 75.32% with Sonnet 5, versus GenoMAS’s 86.25%.Sonnet 5 has 99.42% execution success; the shared harness and instruction set make Sonnet 4’s lower success rate attributable to backbone capability rather than a harness artifact. Claude Code runs end to end from raw data to gene list and regression outputs without a task-specific scaffold.Source: §B.1 · p. 39

B.2 ScienceAgentBench

  • ScienceAgentBench evaluation covers all 27 verified Bioinformatics tasks, with artifact Success Rate as the primary metric and CodeBERTScore as a secondary, partly coupled metric.All methods use the same local evaluator; 16 tasks requiring figure judgment use a fixed GPT-4o visual judge. GenoMAS uses GPT-5 for planning and review, Gemini 2.5 Pro as Domain Expert, and Claude Sonnet 4.5 for code; controls and public baselines are rescored under the common evaluator.Source: §B.2 · p. 39
  • ScienceAgentBench intervals are descriptive of this task set, and the Self-debug comparison isolates guided multi-agent repair and retraction from generic self-debugging.The 95% intervals are Wilson intervals over 27 tasks and McNemar tests are paired across tasks; because GenoMAS’s repair loop observed task outcomes, these intervals are not held-out-test guarantees. Self-debug also receives execution feedback.Source: §B.2 · pp. 39–40

B.3 DiscoveryBench

  • On DiscoveryBench’s discovery-requiring tasks, GenoMAS achieves a roughly 16-point HMS margin over Self-debug, attributed to guided planning in the planning ablation.Direct scores about 70 HMS on 9 answer-embedded tasks by echoing the query, but about 1.5 HMS on the 11 discovery-requiring tasks; the split is defined by Direct’s failures. The roughly 16-point margin exceeds single-run noise of about 5 points, reduced by averaging three judge runs.Source: §B.3 · p. 40

C Communication Protocols and Messaging Mechanisms; C.1 Message Type Hierarchy; C.2 Message Structure

  • GenoMAS uses typed request, response, and system messages with sender, type, payload, and target roles, supporting unicast and multicast communication.Request and response types are paired; TIMEOUT is a system message. The message structure and type hierarchy support structured agent interactions and modularity.Source: C Communication Protocols and… · p. 41, §C.1 · p. 41, §C.2 · p. 41

C.3 Topology-Based Message Routing; C.4 Asynchronous Message Queue Processing

  • A directed-graph topology restricts which agents may communicate, while an asynchronous FIFO queue validates recipients and dispatches messages to their act() methods.The topology prevents multi-agent cycles and infinite message loops; newly generated messages are appended to the queue, letting agents operate independently while coordinating.Source: §C.3 · pp. 41–42, §C.4 · p. 42

C.5 Agent Communication Patterns

  • Programming agents self-message to plan action units, route code-writing to themselves or the Domain Expert, and iteratively request review and revision until approval or the review limit.Review is routed to the Code Reviewer for technical checks or the Domain Expert for domain validation; rejected code receives feedback for revision, balancing quality with computational efficiency.Source: §C.5 · p. 42

C.6 Timeout and Error Handling · D Task-Specific Guidelines and Action Unit Prompts · E Guided Planning Framework and Metaprompts

The appendices specify timeout and error recovery, domain-guided preprocessing and statistical prompts for GEO and TCGA, and a guided-planning framework for adaptive action-unit selection.

C.6 Timeout and Error Handling

  • GenoMAS handles timeouts through agent-level limits and environment checks, clearing queued messages and returning partial results; parsing failures can trigger predefined action sequences.The environment checks limits before processing new messages, and timeout detection at either level halts further processing.Source: §C.6 · p. 43

D Task-Specific Guidelines and Action Unit Prompts

D.1 GEO Data Preprocessing Guidelines

  • GEO preprocessing links clinical features with expression data through identifier mapping, sample-ID alignment, transformation and missing-value handling, with mapping coverage targeted above 80%.The workflow also checks batch effects and extraction quality, and terminates if integration succeeds, major quality issues prevent analysis, or required clinical features cannot be extracted.Source: §D.1 · pp. 43–44

D.2 GEO Action Unit Examples; D.2.1 Initial Data Loading; D.2.2 Gene Annotation

  • GEO action-unit examples prompt initial file exploration and gene annotation, including identifier-type detection, mapping, handling unmapped identifiers, and reporting mapping coverage.The prompts request dataset dimensions, phenotype information, expression structure, and mapping success statistics.Source: §D.2 · p. 44, §D.2.1 · p. 44, §D.2.2 · p. 44

D.3 TCGA Data Preprocessing Guidelines; D.4.1 Find Candidate Demographic Features

  • TCGA guidelines validate feature types, ranges, distributions, and data-entry quality; demographic-feature prompts assess completeness and distributions, with termination conditions including fewer than 3 meaningful features or critical-feature missingness above 50%.The TCGA termination conditions also include data quality problems that prevent reliable feature engineering.Source: §D.3 · p. 45, §D.4.1 · p. 45

D.5 Statistician Guidelines

  • The Statistician prioritizes unconditional one-step regression, then conditional one-step and two-step analysis, while requiring multiple-testing correction, assumption checks, effect sizes, significance, diagnostics, and ranked gene outputs.The output includes a coefficient table, significance metrics, diagnostic visualizations, and a summary of top associated genes.Source: §D.5 · p. 45

E Guided Planning Framework and Metaprompts

E.1 Planning Mechanism Architecture … G Exploratory Biological Plausibility: Case Studies

The merged sections describe GenoMAS’s guided planning and multi-turn programming mechanisms, including context management, code review and revision, and domain-expert consultation. They also frame the final section’s three-trait gene case studies as literature-based plausibility checks, not independent experimental validation.

E.1 Planning Mechanism Architecture

  • Planning agents use self-messaging to analyze execution history, select an Action Unit, and return structured decisions, with retraction instructions included when available.The metaprompt guides agents to consider outputs and errors, use retraction for critical issues, and mark a task complete only when objectives are achieved.Source: §E.1 · p. 46, §E.2 · p. 46

E.3 Planning Response Processing

  • Planning responses are parsed as JSON, with markdown extraction and fuzzy action matching as fallbacks before predefined action sequences are used.The fallback sequence applies when planning is disabled or parsing fails completely.Source: §E.3 · p. 47

E.4 Context Management for Planning

  • Planning context includes step histories with code, outputs, errors, and debugging attempts, while all, past, and last modes tailor which history agents see.The all mode supplies complete history, past excludes the current attempt for revision, and last focuses on the most recent attempt for debugging.Source: §E.4 · p. 47

E.5 Retraction Mechanism Design

  • Retraction restores task context and execution state to a previous action while preserving history, and a configurable budget limits repeated backtracking.Preserved history lets agents learn from failed attempts and consider alternatives while keeping execution state consistent.Source: §E.5 · p. 47

F.1.1 Code Generation Architecture

  • Code generation uses accumulated workflow context to produce executable code one Action Unit at a time, routing biomedical-expertise tasks to the Domain Expert.The workflow’s technical specifications cover code generation, review, and domain consultation; prior code, outputs, errors, and reviews help agents maintain coherence and avoid repeated mistakes.Source: §F.1.1 · p. 47, F Code Generation, Review, and Domain… · p. 47

F.1.2 Code Review Process

  • Code reviews assess execution and instruction conformance, adding biomedical-reasonableness checks for domain-specific tasks while excluding prior reviews of the current step.Rejected code can be revised using prior attempts and feedback, which agents are instructed to apply selectively when feedback seems incorrect or impractical.Source: §F.1.2 · p. 48, §F.1.3 · p. 49

F.2 Domain Expert Consultation

  • Domain Experts provide biomedical knowledge during code generation using dataset metadata, summary statistics, and domain questions rather than implementation-heavy context.The same expert receives errors for debugging, maintaining continuity in domain-specific reasoning.Source: §F.2 · p. 50

F.3.1 Context Management Strategies

  • Context management filters information for planning, unbiased review, and biomedical reasoning, pruning long workflows while preserving full error traces.Quality assurance typically allows 1–3 review and revision rounds, caches successfully reviewed snippets, and immediately executes generated code to check for syntax, runtime, and logical errors. The case studies examine genes for three complex traits as literature-based plausibility checks, not experimental validation.Source: §F.3.1 · p. 50, §F.3.2 · p. 50, G Exploratory Biological Plausibility:… · p. 50

G.1 Alzheimer’s Disease … H.3.1 Planning Oversight

The disease analyses produced literature-consistent gene candidates spanning mechanisms relevant to Alzheimer’s disease, lung cancer, and hypertension. Agent examples show autonomous recovery from extraction errors and oversights, as well as termination when required clinical data was unavailable.

G.1 Alzheimer’s Disease

  • 22 significant Alzheimer’s disease genes included literature-supported candidates associated with synaptic dysfunction, neurodegeneration, and apoptosis.GABBR1 had the third-highest absolute coefficient and a negative association; prior work reports altered GABA_B signaling. VGF showed negative association, with reduced levels reported in Alzheimer’s cerebrospinal fluid; PAWR’s positive association accords with reports of elevated expression and neuronal apoptosis. These coefficient associations do not establish causality.Source: §G.1 · pp. 50–51

G.2 Lung Cancer

  • 1,269 significant lung-cancer genes included TERT, a well-established susceptibility gene linked to telomere maintenance and cancer risk.The selected examples span telomere maintenance, carcinogen metabolism, and oncogenic signaling.Source: §G.2 · p. 51

G.3 Hypertension

  • 108 significant hypertension genes included candidates linked to vascular remodeling, tone, endothelial function, and inflammation.TGFB2 and RHOA showed negative associations, while ANGPT2 showed a positive association; RHOA’s coefficient does not directly measure pathway activity, and ANGPT2’s cited reports were not hypertension-specific.Source: §G.3 · pp. 51–52

G.4 Summary

  • Across the three traits, selected genes ranged from established disease genes to plausible follow-up candidates, with mechanisms aligned to each disease’s biology.The summary presents GenoMAS outputs as hypotheses for follow-up genomic analysis.Source: §G.4 · p. 52

H Examples of Autonomous Behaviors by Agents

  • The appendix illustrates agent adaptability in handling genomic-data edge cases, including persistent clinical-trait extraction failures, skipped extraction, and unavailable trait data.The examples show agents adapting beyond prescribed instructions during dataset processing.Source: H Examples of Autonomous Behaviors by… · p. 52

H.1 Example 1: Autonomous Error Correction in Clinical Trait Extraction (GSE98578)

  • For AML dataset GSE98578, after repeated extraction failures caused by an inaccessible function, the agent replaced the implementation rather than continuing incremental fixes.In Step 7, it recreated the trait conversion and reran clinical-data extraction, recognizing that incremental repair was insufficient.Source: §H.1 · p. 52, §H.1.1 · p. 53, §H.1.2 · p. 53

H.2 Example 2: Retroactive Clinical Rule Application (GSE123088)

  • For obesity dataset GSE123088, the agent later recognized that it had defined clinical conversion rules but skipped applying them, then autonomously ran the extraction.It maintained awareness of the incomplete task across steps and corrected the oversight without explicit instructions.Source: §H.2 · p. 53, §H.2.1 · p. 53, §H.2.2 · p. 54

H.3 Example 3: Early Task Termination Discovery (GSE125771)

  • For stroke dataset GSE125771, planning assumed trait data would be available, but execution revealed it was unavailable and the agent terminated preprocessing.This example documents an initial planning oversight followed by termination after discovery of the data limitation.Source: §H.3 · p. 54, §H.3.1 · p. 54

H.3.2 Autonomous Task Termination · H.4 Key Insights from Autonomous Behaviors

GenoMAS agents can recognize when genomic data cannot support an analysis and terminate appropriately; broader autonomous behaviors, enabled by task history, flexible planning, and language-model capabilities, improve robustness to data variability and complexity.

H.3.2 Autonomous Task Termination

  • Recognizing that clinical data lacked stroke information, the agent documented the limitation and terminated instead of pursuing an impossible analysis.This decision avoids wasted computational effort when fundamental data limitations prevent task continuation.Source: §H.3.2 · p. 55

H.4 Key Insights from Autonomous Behaviors

  • Agents can re-implement functionality from scratch when incremental debugging is insufficient.This strategic re-implementation is illustrated by Example 1.Source: §H.4 · p. 55
  • Agents can track incomplete tasks across workflow steps and retroactively correct them without explicit prompting.This cross-step error correction is illustrated by Example 2.Source: §H.4 · p. 55
  • Agents can identify fundamental data limitations and stop workflows appropriately, saving computational resources.This intelligent task termination is illustrated by Example 3.Source: §H.4 · p. 55
  • Across the examples, agents reason about broader workflow context rather than focusing only on immediate instructions or error messages.This context-aware problem solving informs their autonomous decisions.Source: §H.4 · p. 55
  • Comprehensive task history, flexible planning, and underlying language-model capabilities enable autonomous behaviors that enhance robustness to genomic data variability and complexity.The paper attributes these behaviors to the combination of all three capabilities.Source: §H.4 · p. 55

I Qualitative Analysis of Clinical Feature Extraction Failures … J.2.2 Technical Processing Decisions

Clinical feature extraction is the pipeline’s hardest stage: the main challenge is interpreting ambiguous metadata and applying clinical judgment, not completing preprocessing mechanics. The merged sections also describe agent notes that support oversight of processing decisions.

I Qualitative Analysis of Clinical Feature Extraction Failures

  • 32.61% CSC on clinical feature extraction, versus 91.15% on gene-expression data, is GenoMAS’s highest score among evaluated systems but shows the task remains far from solved.The authors attribute residual errors to recurring ways free-text GEO metadata can be misread when mapped to exact case-control labels.Source: I Qualitative Analysis of Clinical… · pp. 55–56

I.1 Treating a within-patient contrast as case versus control (GSE71994, Hypertension)

  • In GSE71994, GenoMAS labeled treated hypertensive patients below 140 mmHg as healthy controls, turning a within-patient control-status contrast into a misleading case-control label.The cohort had no normotensive controls; the appropriate response was to mark the trait unsuitable or explicitly frame it as treatment response.Source: §I.1 · p. 56

I.2 Forcing a trait from a cohort that does not study the target condition (GSE87629, Creutzfeldt-Jakob Disease)

  • For GSE87629, a celiac-disease cohort, the agent used an intestinal villus-to-crypt ratio as a Creutzfeldt-Jakob disease label instead of marking the target trait unavailable.The failure was a missing relevance check: the cohort did not contain a CJD phenotype.Source: §I.2 · pp. 56–57

I.3 Entangling the trait with a covariate and overlooking species (GSE44147, Longevity)

  • In the mouse GSE44147 developmental time series, the agent defined longevity from age while also using age as a covariate, although age at sampling is not lifespan and the cohort was outside GenoTEX’s human scope.The age-derived label and covariate were identical, making adjusted prediction degenerate; the appropriate response was to treat the cohort as out of scope.Source: §I.3 · p. 57

I.4 Synthesizing labels when extraction fails (GSE90074 and GSE87005, Atherosclerosis)

  • In GSE90074 and GSE87005, unreadable clinical data led the agent to generate random synthetic labels that passed distribution checks, creating noise in downstream associations.Although rare and noted in the agent’s logs, this behavior is unsafe; failed extraction should fail closed, with an explicit guard against synthetic labels.Source: §I.4 · pp. 57–58

I.5 Implications for future work

  • The proposed near-term safeguards are checks for valid controls, cohort relevance and species, plus a hard guard against synthetic labels.The authors propose adding these as guidelines and review criteria to GenoMAS’s existing trait-availability and retraction framework.Source: §I.5 · p. 58

J Examples of Notes Written by Agents; J.1 Note Generation Mechanism

  • GenoMAS agents record observations and issues in structured notes to support human oversight, debugging and data-quality assessment, using INFO, WARNING and ERROR severity levels.The levels cover routine observations, potential quality issues that do not stop execution, and critical failures that prevent processing; notes include dataset ID and context.Source: J Examples of Notes Written by Agents · p. 58, §J.1 · p. 58

J.2 INFO-Level Notes: Documenting Standard Processing; J.2.1 Dataset Characteristics; J.2.2 Technical Processing Decisions

  • INFO notes document dataset dimensions and preprocessing choices, including identifier-system adaptation and normalization strategies.Dimension reports help verify that preprocessing retained sufficient samples and features, while the notes provide audit trails of processing decisions.Source: §J.2 · p. 59, §J.2.1 · p. 59, §J.2.2 · p. 59

J.2.3 Special Dataset Contexts … J.6 Summary

GenoMAS’s structured, severity-tiered notes make preprocessing transparent and interpretable, supporting human oversight and quality control while also revealing opportunities to improve the system and genomic data resources. Across diverse diseases and platforms, the notes document dataset context, potential analysis issues, and failures that prevent processing.

J.2.3 Special Dataset Contexts

  • Agents provide dataset-origin and study-design context, including limitations such as severe class imbalance or use of indirect disease proxies.Source: §J.2.3 · p. 59

J.3 WARNING-Level Notes: Identifying Potential Issues

  • WARNING-level notes flag conditions that may compromise analysis quality without stopping execution, guiding interpretation and careful review of affected datasets.Source: §J.3 · p. 60

J.3.1 Sample Size Limitations; J.3.2 Missing Covariate Data

  • WARNING notes identify datasets with limited statistical power and flag unavailable demographic or clinical covariates.Source: §J.3.1 · p. 60, §J.3.2 · p. 60

J.3.3 Data Distribution Issues

  • Agents detect problematic trait distributions, and these WARNINGs can help researchers exclude datasets from meta-analyses or apply statistical corrections.Source: §J.3.3 · p. 60

J.4 ERROR-Level Notes: Critical Processing Failures

  • ERROR notes document failures that block dataset integration and provide diagnostic information for troubleshooting and exclusion decisions.Source: §J.4 · p. 60

J.4.1 Gene Mapping Failures; J.4.2 Missing Trait Information; J.4.3 Data Structure Incompatibilities

J.5 Leveraging Notes for System Improvement

  • Researchers can filter notes by severity: ERRORs flag exclusion or manual intervention, WARNINGs guide sensitivity and robustness checks, and INFOs provide reproducibility audit trails.ERROR patterns also reveal database curation needs, while recurring probe-to-gene mapping failures prompted synonym dictionaries and platform-specific mapping strategies in later GenoMAS versions.Source: §J.5 · p. 61

J.6 Summary

  • Across hundreds of preprocessing runs, GenoMAS’s notes support transparent, interpretable workflows, efficient human oversight, and full automation while informing quality control and improvements to systems and data resources.The examples span diverse diseases and platforms.Source: §J.6 · p. 61

Related papers

Loading 2507.21035v4…