Towards Autonomous Mechanistic Reasoning in Virtual Cells

Abstract: 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.

• The paper introduces a formal framework that represents mechanistic reasoning as a DAG using ontologically-anchored action primitives for structured, verifiable explanations.
• The methodology employs a two-stage multi-agent pipeline that integrates biomedical NER, multi-source data retrieval, and specialized verifiers to ensure evidence-backed mechanistic traces.
• The system enhances downstream tasks such as gene expression prediction and aligns closely with expert evaluations, demonstrating practical benefits for virtual cell modeling.

Autonomous Mechanistic Reasoning for Virtual Cells: Structured Explanations and Verification with VCR-Agent

Introduction: The Challenge of Mechanistic Biological Reasoning

Mechanistic prediction and interpretability are central but unresolved challenges in the development of virtual cell models for biology and pharmacology. While LLMs have demonstrated emergent reasoning in formal domains, their application to open-ended, ambiguous biological phenomena is limited by data scarcity and intractable verification. This work introduces a formalism and computational framework, VCR-Agent, to autonomously construct and validate mechanistic reasoning in simulated cellular contexts, addressing the need for structured, biologically plausible explanations that are subject to systematic scrutiny.

Structured Mechanistic Reasoning Formalism

The proposed framework represents biological reasoning as the inference of a directed acyclic graph (DAG), where nodes encode discrete, biologically grounded action primitives (e.g., binding, regulation, phenotype induction) with schema-constrained arguments, and edges represent mechanistic dependencies (causal or correlative links). This action-space formalism delineates permissible reasoning steps, enforces interpretability, and enables rigorous logic checking as all steps are tied to ontologically anchored entities or concepts.

Figure 1: Example mechanistic reasoning traces converting a perturbation-context pair into actions and dependencies, forming a DAG representing stepwise mechanistic logic.

Confining mechanistic reasoning within a finite, pre-defined set of action primitives enables compatibility with fact-checking tools and external biological databases, supporting falsifiability of claims. For instance, the "binds_to" action requires explicit actors (small molecule, protein) and molecular context; the "regulates_expression" action must specify regulators and gene sets with directionality. This explicit, modular approach is fundamentally distinct from free-form rationales and prevents hallucinations prevalent in unconstrained LLM outputs.

Multi-Agent Reasoning Architecture: VCR-Agent

VCR-Agent operationalizes mechanistic reasoning as a two-stage multi-agent pipeline:

• Report Generator: Uses biomedical NER to extract entities from the perturbation-context input, performs document and relational retrieval from external structured sources (e.g., StarkPrimeKG, Harmonizome, PubMed, Wikipedia), and synthesizes relevant evidence into a comprehensive, context-dependent report.
• Explanation Constructor: Consumes this synthesized report and, via prompted generation, produces the structured mechanistic DAG, populating arguments for each action primitive with specifics grounded by the evidence summary.

A suite of specialized verifiers then acts on the structured trace, quantitatively scoring or flagging actions for plausibility and internal logic. Only explanations that pass the full set of verifications are retained for downstream use.

Figure 2: Overview of VCR-Agent, integrating knowledge retrieval, report synthesis, mechanistic explanation construction, and verifier-driven filtering.

Verifier-Based Filtering for Trace Reliability

Central to the framework is automated verification by biology-specific verifiers. The two primary verifiers in this work are:

• DTI Verifier: Assesses the physical plausibility of a "binds_to" event using a state-of-the-art structure-based binding predictor (Boltz-2), providing a high-precision, continuous confidence score for each predicted molecular interaction.
• DE Verifier: Validates claims of gene expression changes using empirical differential expression profiles from large-scale resources (Tahoe-100M).

Verification proceeds at two levels: per-action (filtering or correcting individual steps with low support) and per-trace (entire explanations with unresolved contradictions or unsupported events are discarded). This pipeline systematically reduces the prevalence of unsupported, spurious, or contradictory mechanistic claims.

Figure 3: Example of the verifier filtering process: raw mechanistic traces are pruned by action-level and global logical verifiers, increasing evidence alignment.

Quantitative Evaluation: Explanation Quality and Downstream Utility

Explanation Structural and Factual Quality

VCR-Agent was evaluated on structured traces for 18,950 compound-context pairs from Tahoe-100M. Metrics include trace format validity, verifiability (fraction of arguments mapped to valid biomedical entities), DTI and DE verifier scores, and LLM-judge scores for scientific accuracy, logical consistency, and mechanistic clarity. VCR-Agent achieves maximal syntactic and format validity (1.0), and superior DTI and DE scores when compared to strong LLM baselines—even prior to any verifier-based filtering. Filtering further excludes 28.2% of DTI claims and corrects 87.3% of DE actions, concentrating label precision.

Impact on Supervised Learning for Gene Expression Prediction

The VC-Traces dataset of verified mechanistic explanations serves as high-quality supervision for downstream tasks. On the TahoeQA differential expression classification benchmark, LLMs fine-tuned with explicit mechanistic reasoning traces (either as direct input context or by prompting for autonomous generation-and-prediction) significantly outperform conventional SFT and even transcriptomic foundation models trained on raw data.

Figure 4: TahoeQA task performance. Models incorporating structured mechanistic explanations—either via SFT-prompt or SFT-generate—outperform statistical, foundation model, and LLM-only baselines.

The improvement is especially pronounced for out-of-distribution and few-shot gene perturbations, demonstrating the transfer benefits of mechanistic inductive biases.

Expert and Automated Evaluation Alignment

Human expert ratings on explanation plausibility show strong Pearson correlation ($r>0.7$) with the automated LLM-judge scoring criteria, validating the choice of evaluation protocol.

Figure 5: Scatterplot of human expert versus LLM-judge scores for trace plausibility, demonstrating high correlation between expert and automated assessment along key explanation quality axes.

Ablation and Robustness Analyses

Ablations confirm that:

• Full multi-source retrieval provides complementary information; no single database is sufficient for optimal mechanistic generation.
• Claude 4 outperforms other LLMs as the backbone for structured explanation fidelity and format adherence.
• The explicit two-stage pipeline (separate report generation and explanation construction) nearly doubles verifier alignment over direct one-step generation.
• Verifier-based filtering, while yielding modest surface-level LLM-evaluator gains, is essential for factually consistent, expert-grade traces insensitive to superficial plausibility.

Theoretical and Practical Implications

• Theoretical: Formalizing biological reasoning as action graphs constrained by ontological schemas enables automated verification and hypothesis pruning, providing a rigorous alternative to correlational deep learning predictions. This paradigm bridges the gap between tractable symbolic reasoning and empirically driven life sciences.
• Practical: Integration of multi-source retrieval and multi-agent mediated verification serves as a scalable template for automated scientific hypothesis generation, evaluation, and dataset curation in other open-world domains.
• Future Directions: Extensions might include additional verifier types (e.g., phenotypic, localization, metabolic), incorporation of reinforcement learning with surrogate biological feedback, and generalization to multi-modal (omics, imaging, text) reasoning. Improved grounding, richer ontologies, and tighter coupling with experimental knowledge graphs are critical for advancing model reliability.

Conclusion

This work establishes a tractable methodology for autonomous mechanistic reasoning in simulated cellular systems by unifying structured action formalism, multi-source knowledge retrieval, modular reasoning agents, and domain-specific verifiers. The VCR-Agent system, and its released dataset VC-Traces, demonstrate strong empirical benefits for both explanation fidelity and downstream supervised learning. This principled architecture offers a path toward reliable, scalable, and auditable biological reasoning for both virtual cell modeling and other high-stakes scientific applications.