PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Autonomous Mechanistic Reasoning in Virtual Cells

Yunhui Jang, Lu Zhu, Jake Fawkes, Alisandra Kaye Denton, Dominique Beaini, Emmanuel Noutahi

arXiv:2604.11661v3cs.LGcs.AI

TL;DR

Virtual-cell models need explanations that are both mechanistically grounded and reliable, but biological reasoning is difficult to curate and verify. The paper introduces structured mechanistic graphs and the VCR-Agent pipeline, releasing verified VC-Traces from Tahoe-100M. These explanations improve factual precision and downstream gene-expression prediction, while remaining intended for mechanistic plausibility rather than formal causal discovery or clinical implementation.

  • Problem

    Biological reasoning lacks scalable, reliable datasets and direct correctness verification, limiting factually grounded explanations for virtual cells.

  • Method

    The paper represents explanations as mechanistic action graphs and uses VCR-Agent to retrieve biological knowledge, construct structured traces, and filter them with verifiers.

  • Results

    Verified structured explanations improve factual precision and downstream gene-expression prediction, with VCR-Agent explanations outperforming baselines.

  • Takeaways & Limitations

    Structured, verified mechanistic explanations provide a scalable supervision signal for reliable virtual-cell reasoning and downstream biological tasks.

  • Takeaways & Limitations

    The framework targets mechanistic plausibility rather than formal causal discovery or clinical implementation, and external knowledge-base biases or gaps can affect validity.

Abstract

from arXiv · show

Large language models (LLMs) have recently gained significant attention as a promising approach to accelerate scientific discovery. However, their application in open-ended scientific domains such as biology remains limited, primarily due to the lack of factually grounded and actionable explanations. To address this, we introduce a structured explanation formalism for virtual cells that represents biological reasoning as mechanistic action graphs, enabling systematic verification and falsification. Building upon this, we propose VCR-Agent, a multi-agent framework that integrates biologically grounded knowledge retrieval with a verifier-based filtering approach to generate and validate mechanistic reasoning autonomously. Using this framework, we release VC-TRACES dataset, which consists of verified mechanistic explanations derived from the Tahoe-100M atlas. Empirically, we demonstrate that training with these explanations improves factual precision and provides a more effective supervision signal for downstream gene expression prediction. These results underscore the importance of reliable mechanistic reasoning for virtual cells, achieved through the synergy of multi-agent and rigorous verification.

1 Introduction

Virtual-cell reasoning is limited by unreliable biological explanations and difficult verification. The paper addresses this with structured mechanistic graphs, VCR-Agent, and verified VC-Traces that improve downstream gene-expression prediction.

  • Motivation: Biological explanations must move beyond correlation-based prediction while remaining biologically plausible and reliable.The paper identifies reliable mechanistic explanation generation as a critical bottleneck for virtual cells.
  • Motivation: Human annotations are expensive and non-scalable, whereas LLM-generated reasoning traces are often factually unreliable and prone to hallucination.These data-curation constraints make specialized biological reasoning datasets difficult to build.
  • Structured explanations: Structured explanations represent biological reasoning as discrete mechanistic actions connected by directed dependencies, making traces interpretable and falsifiable.Nodes encode actions, while edges encode mechanistic relationships such as preconditions or regulatory requirements.
  • VCR-Agent: VCR-Agent separates biological knowledge retrieval, structured explanation construction, and verifier-based factual filtering.The report generator grounds the process in external knowledge, the explanation constructor produces the formal representation, and the verifier filters traces.
  • Dataset and evaluation: VC-Traces is a structured mechanistic-explanation dataset generated from the Tahoe-100M atlas for explanation assessment and downstream gene-expression prediction.The dataset is evaluated both for explanation quality and as supervision for downstream biological tasks.
  • Results: Verified structured explanations improve LLM performance on gene-related downstream tasks and provide a stronger supervision signal for gene-expression prediction.The reported experiments compare models trained on verified traces against baselines.

2 Structured Mechanistic Reasoning for Virtual Cells

Structured mechanistic reasoning represents virtual-cell explanations as directed acyclic graphs of biologically grounded actions and dependencies. A finite action space with argument schemas enables interpretability and verifier-based falsification.

  • Problem Formulation: Virtual-cell reasoning infers how a perturbation affects cellular state through mechanistic actions triggered in a cellular context.The input is a perturbation–context pair, and the output is a reasoning graph capturing the resulting action chain.
  • Problem Formulation: The output graph G is a directed acyclic graph whose nodes are mechanistic actions and whose edges encode dependencies between them.An upstream action can enable or influence a downstream action, such as ligand–receptor binding preceding pathway modulation.
  • Action Spaces: Each action combines a predefined primitive with domain-specific arguments, constraining reasoning to biologically grounded operations.The action space contains twenty primitives grouped into seven categories spanning molecular interactions through phenotypic manifestations.
  • Action Spaces: The binds to primitive encodes a direct interaction between two biomolecules using identifiers, actor and target entities, and optional interaction attributes.Optional arguments include affinity, units, residues, interaction route, and confidence.
  • Action Spaces: Structured arguments allow verifiers to assess action reliability against curated biological databases and computational tools.The representation supports interpretability and downstream biological consequences while remaining distinct from formal interventional causal discovery.

3 LLM-Agent Framework for Reasoning

VCR-Agent generates structured mechanistic explanations through a two-stage pipeline that first retrieves and summarizes biological knowledge, then converts that report into a structured reasoning trace. This separation is intended to improve factual grounding and structural integrity.

  • Framework Overview: VCR-Agent uses a report generator and an explanation constructor to produce structured explanations from perturbation–context inputs.The report generator handles factual knowledge, while the constructor produces the structured reasoning format.
  • Report Generation: The report generator extracts biomedical entities, retrieves related information, and summarizes the retrieved facts into a comprehensive natural-language report.Entity extraction enables entity-based search instead of complex natural-language queries.
  • Knowledge Retrieval: Retrieval queries StarkPrimeKG, Harmonizome, PubMed, and Wikipedia to gather relational, gene-level, literature-based, and general biological information.StarkPrimeKG aggregates one-hop relational context, while Harmonizome enriches gene-specific information.
  • Explanation Construction: The explanation constructor transforms the knowledge-grounded report into the structured reasoning format used for verification and falsification.Claude 4 is used for explanation generation, with the report serving as the factual foundation.

4 Verifier-based Filtering and Quality Control

VCR-Agent applies specialized biological verifiers to generated actions and filters traces or arguments that fail plausibility thresholds. The pipeline targets hallucinations while retaining claims that do not directly contradict biological evidence.

  • Filtering Pipeline: The quality-control pipeline performs per-action verification followed by filtering to mitigate hallucinations and remove biologically conflicting reasoning traces.Verification scores can eliminate either an entire trace or partial arguments below plausibility thresholds.
  • Verifier Design: Four specialized verifiers cover drug–target interaction, differential expression, subcellular localization, and phenotype actions.DTI and DE are the primary verifiers used for filtering in this study.
  • DTI Verification: The DTI verifier uses Boltz-2 to estimate the physical plausibility of drug–protein binding and produce a continuous binding-probability score.Binding claims below a predefined confidence threshold are discarded.
  • DE Verification: The DE verifier checks whether a perturbation regulates a target gene in the predicted direction using ground-truth differential-expression datasets.It flags hallucinated gene targets that contradict established biological knowledge.
  • Filtering Results: The two primary filtering action types occur in 91.5% of VC-Traces explanation traces, providing broad filtering coverage.The filtering process excludes faulty DTI claims and refines DE actions to remove hallucinations.

5 Experiments

Experiments evaluate explanation quality and the downstream usefulness of VC-Traces for gene-expression prediction. VCR-Agent produces high-quality traces, while structured reasoning improves TahoeQA performance, especially for differential expression and novel compounds.

  • Experimental Setup: The experiments assess both the quality of VCR-Agent explanations and VC-Traces as supervision for downstream biological tasks.The dataset is derived from 18,950 compound perturbation–context pairs in Tahoe-100M, with experiments focused on five cell lines.
  • Explanation Quality: VCR-Agent consistently generates structurally valid, verifiable traces that align more closely with biological reference data than baseline models.Table 1 reports format-based and verifier-based evaluation before filtering.
  • Explanation Quality: 28.2% of faulty DTI claims were excluded and 87.3% of DE actions were refined during VC-Traces construction.These filtering results are reported separately from the raw generative metrics in Table 1.
  • TahoeQA: Structured reasoning significantly and consistently improves TahoeQA predictive accuracy, with SFT-Prompt achieving the strongest overall performance across cell types and tasks.SFT-Generate substantially surpasses all baselines on the differential-expression task.
  • TahoeQA: Training with explicit biological reasoning provides a more effective supervision signal than direct label prediction alone.The evaluation uses F1-score because the differential-expression labels are imbalanced.
  • TahoeQA: Structured reasoning improves generalization to novel compounds and supports performance in sparse-data and out-of-distribution settings.This contrasts with baselines such as STATE, which rely primarily on raw numerical representations.

6 Related Work

Prior biological reasoning systems face unreliable fact discovery, structurally ambiguous explanations, and difficult verification under causal uncertainty. VCR-Agent addresses these challenges by combining retrieved knowledge with structured mechanistic graphs and automated verification.

  • Biology reasoning traces are difficult to verify because they depend on fragmented literature rather than deterministic rules.
  • Existing models often rely on internal parametric knowledge, producing biologically implausible hallucinations without sufficient external grounding.
  • Unstructured rationales obscure mechanistic dependencies and prevent systematic verification.
  • Prior strategies focus mainly on gene-centric perturbations and neglect diverse modalities such as drug-induced responses.
  • VCR-Agent combines knowledge retrieval, structured reasoning generation, and verifier-based filtering to produce more reliable mechanistic traces.

7 Conclusion

The paper defines a structured, falsifiable reasoning format for virtual cells and introduces VCR-Agent to generate and validate explanations. Its verified traces improve downstream gene expression prediction and support more grounded mechanistic reasoning.

  • The work defines a structured, falsifiable reasoning format for virtual-cell explanations.
  • VCR-Agent separates knowledge retrieval from explanation construction and applies a verifier pipeline to validate generated explanations.
  • The generated explanations show factual accuracy and logical coherence while outperforming baselines.
  • The resulting dataset improves performance on the downstream TahoeQA gene expression prediction task.

Impact statement

The framework supports interpretable, verifiable mechanistic reasoning by replacing unstructured text with structured graphs and releasing VC-Traces for expert-augmented research. Its action space represents biological processes across molecular, regulatory, phenotypic, and cellular contexts.

  • Structured mechanistic graphs replace unstructured text to support systematic verification and falsification of biological hypotheses.
  • VC-Traces is a scalable, grounded dataset derived from Tahoe-100M that is intended to augment domain experts rather than provide a closed-loop solution.
  • The framework is intended for mechanistic plausibility, not formal causal discovery or direct clinical implementation.
  • Dependence on external knowledge bases means repository biases or incomplete data may affect validity, especially in specialized contexts.
  • The action space includes context initialization, metabolic conversion, activity modulation, protein regulation, transcriptional and translational regulation, interactions, perturbations, and phenotypic outcomes.
  • Action primitives encode mechanisms including binding, pathway modulation, expression regulation, genetic gain or loss of function, and phenotype induction or alleviation.

B.2 Generated report

The generated report describes binimetinib treatment in BRAF-mutant C32 melanoma cells as a mechanistic progression from MEK inhibition to cell-cycle arrest, metabolic stress, compensation, and apoptosis. The report links these stages to measurable transcriptomic, phenotypic, and viability outcomes.

  • Perturbation Description: Binimetinib is an ATP-uncompetitive allosteric MEK1/2 inhibitor targeting inactive MEK1 and MEK2 conformations.
  • Cellular Context: C32 cells harbor BRAF V600E, CDKN2A loss, PTEN loss, and TERT activation, producing hyperactive MAPK and PI3K/AKT signaling.
  • Mechanistic Analysis: MEK1/2 inhibition blocks the BRAF–MEK–ERK cascade, reducing ERK phosphorylation and nuclear translocation within the early treatment response.
  • Mechanistic Analysis: Loss of ERK-dependent transcription lowers immediate-early, cell-cycle, and S-phase gene expression, contributing to G1/S arrest.
  • Mechanistic Analysis: Cells compensate through receptor tyrosine kinase upregulation and PI3K/AKT hyperactivation beginning 12–24 hours after treatment.
  • Mechanistic Analysis: Prolonged treatment selects resistant subpopulations through genetic, receptor-amplification, epigenetic, and metabolic adaptations.
  • Final Phenotypic Outcome: 40–70% cell viability reduction by 96 hours accompanies mitochondrial dysfunction and caspase-dependent apoptosis.
  • Measurable Biological Endpoints: 60–80% of cells accumulate in G1 versus 40–50% of controls, while S-phase cells fall to 5–15% versus 30–40% of controls.

C.1 Detailed experimental results

The section specifies how perturbations are converted into detailed mechanistic reports and structured, falsifiable explanations, while comparing knowledge sources for biological coverage.

  • C.1 Detailed experimental results: The report generator describes perturbations, targets, affinities, causal chains, downstream changes, and measurable phenotypic effects.Reports distinguish causal from correlative steps and prioritize transcriptomic or phenomic end nodes.
  • C.1 Detailed experimental results: The report-generation and explanation-construction prompts require detailed biological reasoning that ends with measurable transcriptomic or phenomic effects.The report is intended to support hypothesis generation and falsification.
  • C.1 Detailed experimental results: The explanation constructor outputs private reasoning, a mechanism-of-action summary, structured actions, and a causal DAG.Structured explanations begin with context, use allowed primitives with unique identifiers, and end in measurable biological outputs.
  • C.1 Detailed experimental results: Structured actions encode dependencies between mechanistic nodes and distinguish causal edges from correlative associations.The representation supports mechanistic plausibility and downstream biological consequences while remaining distinct from formal interventional causal discovery.
  • C.1 Detailed experimental results: No single knowledge source approaches the performance of the full combination, indicating complementary contributions from four biological databases.StarkPrimeKG contributes relational context, Harmonizome gene-level information, PubMed literature evidence, and Wikipedia broad background knowledge.

D.2 Ablation study: Backbone LLM

The backbone ablation compares Claude with GPT-4.1 and Gemini-2.5-Flash under identical retrieval and constructor prompts, emphasizing validity, verifiability, and biological accuracy.

  • D.2 Ablation study: Backbone LLM: Claude achieves perfect trace validity and substantially outperforms GPT-4.1 and Gemini-2.5-Flash on drug-target interaction score.GPT-4.1 achieves a higher differential-expression score but produces lower verifiability because many outputs are structurally invalid.
  • D.2 Ablation study: Backbone LLM: The two-stage pipeline nearly doubles both drug-target interaction and differential-expression scores compared with one-step generation.It also achieves higher verifiability, supporting separation of knowledge synthesis from structured reasoning generation.
  • D.2 Ablation study: Backbone LLM: Identical retrieved reports preserve VCR-Agent’s highest validity and verifiability relative to baseline models.This comparison isolates the structured explanation pipeline from differences in retrieved context.

D.5 Ablation study: Verifier-based filtering

Verifier-based filtering modestly but consistently improves explanation-trace quality, although LLM-based evaluators may miss gains requiring specialized regulatory knowledge.

  • D.5 Ablation study: Verifier-based filtering: Verifier-based filtering evaluates no filtering, DTI-only, DE-only, and full-pipeline conditions across explanation-quality metrics.The study uses complete traces derived from 18,950 Tahoe-100M perturbation-context pairs.
  • D.5 Ablation study: Verifier-based filtering: Without verifiers, scientific accuracy is 0.641, logical consistency is 0.720, and mechanistic clarity is 0.725.The comparison provides the unfiltered baseline for the reported verifier-based results.
  • D.5 Ablation study: Verifier-based filtering: +DE&DTI achieves 0.645 scientific accuracy, 0.724 logical consistency, and 0.725 mechanistic clarity.These are the reported full-pipeline values in the verifier-filtering ablation.
  • D.5 Ablation study: Verifier-based filtering: Filtering improves overall trace quality, but LLM evaluators may not distinguish filtered from unfiltered traces when both appear superficially mechanistic.The stated limitation is that evaluators lack specialized regulatory knowledge used by the verifiers.

D.6 Human evaluation

Human experts and the LLM judge show strong agreement when evaluating explanation traces, supporting the judge’s use as a scalable assessment proxy.

  • D.6 Human evaluation: Ten traces sampled across five cell lines were independently scored by molecular-biology and pharmacology experts.Experts assessed scientific accuracy, logical consistency, mechanistic clarity, and binary plausibility without seeing verifier or LLM-judge outputs.
  • D.6 Human evaluation: Average Pearson correlation between human expert ratings and LLM-judge scores is r = 0.72 across the three evaluation criteria.The authors report consistently high average LLM-judge scores across the full dataset.
  • D.6 Human evaluation: Expert-plausible traces cluster in the upper-right score region, whereas expert-implausible traces receive lower scores from both evaluators.The figure uses teal for plausible traces, orange for implausible traces, regression lines, and 95% confidence intervals.
  • D.6 Human evaluation: The alignment supports using the LLM judge as a scalable proxy for expert review of biological soundness.This conclusion combines numerical correlation with agreement on which traces are biologically sound.
Loading 2604.11661v3…