Towards Autonomous Mechanistic Reasoning in Virtual Cells

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

cs.LGApr 13, 2026

Towards Autonomous Mechanistic Reasoning in Virtual Cells

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

Organizations: Korea Advanced Institute of Science and Technology (KAIST) · Valence Labs · Recursion · University College London

Abstract

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.

Figures & tables

Figure 1: An Overview of the VCR-Agent Multi-Agent Framework. The Report Generator accepts the perturbation and cellular context, performing knowledge retrieval and synthesis to produce a comprehensive, biologically grounded report. The Explanation Constructor then translates this report into the formal structured mechanistic explanation. This generated structured explanation is subsequently evaluated by the Verifier for factual validation and filtering.

Panel of Figure 2 — Figure 2: An overview of structured reasoning. ( 2(a) ) Given an input $(p,c)=$ (Binimetinib, C32), the model generates mechanistic reasoning traces. Blue and light blue indicate the action primitives and the arguments, respectively. The elements within the <dag> tag represent the edge list defining the reasoning graph. ( 2(b) ) An example of DAG. Same color indicates the same action primitive.

Figure 3: An overview of action spaces. The sub-categories are represented with bold and action primitives with verifier are represented with purple . The argument schemes are in Appendix A .

Figure 4: An example of generated report. The input perturbation - cellular context pair follows the one in Figure 2(a) .

Figure 5: An example of verifier-based filtering process. The pipeline processes initial structured explanation (top) through verifiers (middle) to produce filtered output (bottom). Same colors link the action primitive to their corresponding verifiers.

		Format		Verifier
Model		Validity	Verifiability	Drug-target interaction	Differential expression
Open	DeepSeek-R1-8B	0.003 $\pm$ 0.004	1.000 $\pm$ 0.081	0.000 $\pm$ 0.000	0.000 $\pm$ 0.000
	Qwen3-30B	0.860 $\pm$ 0.016	0.965 $\pm$ 0.007	0.528 $\pm$ 0.004	0.272 $\pm$ 0.030
	Llama3.3-70B	0.841 $\pm$ 0.046	0.887 $\pm$ 0.017	0.322 $\pm$ 0.034	0.090 $\pm$ 0.010
Closed	Claude-Sonnet-4	1.000 $\pm$ 0.000	0.867 $\pm$ 0.008	0.657 $\pm$ 0.005	0.504 $\pm$ 0.064
Closed	VCR-Agent (Ours)	1.000 $\pm$ 0.000	0.945 $\pm$ 0.011	0.725 $\pm$ 0.017	0.528 $\pm$ 0.060

Table 1: Explanation quality performance. The best results are highlighted in bold . The standard deviation is computed across cell lines.

Figure 6: TahoeQA performance. Baselines are categorized by model type: statistical and gene foundation models are shown in shades of gray, LLM-based baselines in shades of blue, and our model with structured explanation in brown . Average denotes the mean F1-score across the five individual cell-line test sets while Union denotes the performance on a test set combining all five cell lines.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Sub-category	Action name	Arguments
system initialization		set_context	set_context ({cell_type, genotype, disease, prior_perturbation, extras})
metabolic		converts_substrate	converts_substrate (id, enzyme, substrate, product, {via, confidence})
regulation	activity modulation	modulates_molecule_activity	modulates_molecule_activity (id, target, direction, {via, confidence})
	activity modulation	modulates_pathway_activity	modulates_pathway_activity (id, pathway, direction, {via, confidence})
	protein regulation	modulates_complex	modulates_complex (id, members, complex, direction, {stoichiometry, via, confidence})
	protein regulation	post_translational_modification	post_translational_modification (id, protein, mod_type, site, direction, {via, confidence})

Appendix

Table Appx.1 : Argument schema for action spaces.

Task	Model	C32	HOP62	HepG2/C3A	Hs 766T	PANC-1	Average	Union
Differential expression	Random	0.269	0.254	0.232	0.239	0.301	0.259	0.286
	Mean	0.405	0.419	0.345	0.465	0.330	0.393	0.401
	K-neighbor	0.459	0.419	0.343	0.453	0.355	0.406	0.403
	STATE (ST)	0.273	0.280	0.199	0.219	0.313	0.257	0.251
	Zero-shot	0.363	0.423	0.316	0.332	0.399	0.367	0.338
	SFT	0.344	0.353	0.163	0.363	0.236	0.292	0.291

Appendix

Table Appx.2 : TahoeQA performance with total model. The best results are highlighted in bold and the improvement compared to the vanilla SFT is highlighted with teal .

	Format		Verifier
Retrieval source	Validity	Verifiability	Drug-target interaction	Differential expression
PubMed	1.000 $\pm$ 0.000	0.300 $\pm$ 0.099	0.432 $\pm$ 0.059	0.435 $\pm$ 0.050
StarkPrimeKG	1.000 $\pm$ 0.000	0.296 $\pm$ 0.089	0.382 $\pm$ 0.031	0.449 $\pm$ 0.054
Harmonizome	1.000 $\pm$ 0.000	0.297 $\pm$ 0.092	0.399 $\pm$ 0.013	0.457 $\pm$ 0.051
Wikipedia	1.000 $\pm$ 0.000	0.292 $\pm$ 0.091	0.469 $\pm$ 0.028	0.485 $\pm$ 0.063
All (Ours)	1.000 $\pm$ 0.000	0.601 $\pm$ 0.009	0.766 $\pm$ 0.008	0.487 $\pm$ 0.054

Appendix

Table Appx.3 : Ablation study on retrieval source. The best results are highlighted in bold . The standard deviation is computed across cell lines.

Backbone	Validity	Verifiability	Drug-target interaction	Differential expression
	Format		Verifier
GPT-4.1	0.970 $\pm$ 0.002	0.299 $\pm$ 0.105	0.399 $\pm$ 0.063	0.527 $\pm$ 0.076
Gemini-2.5-Flash	0.060 $\pm$ 0.016	0.963 $\pm$ 0.078	0.338 $\pm$ 0.036	0.207 $\pm$ 0.036
Claude (Ours)	1.000 $\pm$ 0.000	0.601 $\pm$ 0.009	0.766 $\pm$ 0.008	0.487 $\pm$ 0.054

Appendix

Table Appx.4 : Ablation study on backbone LLM. The best results are highlighted in bold . The standard deviation is computed across cell lines.

Step	Validity	Verifiability	Drug-target interaction	Differential expression
	Format		Verifier
One-step	0.999 $\pm$ 0.002	0.534 $\pm$ 0.067	0.329 $\pm$ 0.026	0.267 $\pm$ 0.051
Two-step (Ours)	1.000 $\pm$ 0.000	0.601 $\pm$ 0.009	0.766 $\pm$ 0.008	0.487 $\pm$ 0.054

Appendix

Table Appx.5 : Ablation study on one-step generation. The best results are highlighted in bold . The standard deviation is computed across cell lines.

	Format		Verifier
Model	Validity	Verifiability	Drug-target interaction	Differential expression
DeepSeek	0.046 $\pm$ 0.051	0.711 $\pm$ 0.269	0.185 $\pm$ 0.099	0.246 $\pm$ 0.145
Qwen3	0.937 $\pm$ 0.055	0.340 $\pm$ 0.173	0.081 $\pm$ 0.056	0.556 $\pm$ 0.152
Llama3.3	0.789 $\pm$ 0.105	0.468 $\pm$ 0.227	0.079 $\pm$ 0.098	0.231 $\pm$ 0.146
Two-step (Ours)	1.000 $\pm$ 0.000	0.601 $\pm$ 0.009	0.766 $\pm$ 0.008	0.487 $\pm$ 0.054

Appendix

Table Appx.6 : Ablation study given the same retrieved report. The best results are highlighted in bold . The standard deviation is computed across cell lines.

Without verifiers	0.641	0.720	0.725
	Scientific accuracy	Logical consistency	Mechanistic clarity
+DE	0.642	0.721	0.725
+DTI	0.643	0.721	0.725
+DE&DTI	0.645	0.724	0.725

Appendix

Table Appx.7 : Ablation study on verifier-based filtering. The best results are highlighted in bold .

	Scientific accuracy	Logical consistency	Mechanistic clarity
Human–LLM correlation ( $r$ )	0.72	0.69	0.76
Average LLM-judge score (full)	0.65	0.72	0.73

Appendix

Table Appx.8 : Human expert evaluation. Pearson correlation between human expert ratings and LLM-judge scores on a randomized subset of traces, alongside average LLM-judge scores on the full dataset.

Figure Appx.1 : Agreement between human expert and LLM-judge scores. Each point represents a single trace. Teal points indicate traces judged as plausible by experts; orange points indicate implausible traces. Dashed lines show the linear regression fit and shaded regions denote the 95% confidence interval. Pearson $r$ is reported per criterion.

Scientific Reasoning in Language Models

Scientific reasoning in language models studies how AI systems use evidence, domain knowledge, and tools to solve problems and support discovery.

See all 144 papers

OmniVCBench: Benchmarking Evidence-Grounded Multimodal Reasoning Towards AI Virtual Cells
Sep 29, 2026Manyu Li, Xunkai Li, Yongfu Xiong +3Visual Question Answering Multimodal Reasoning
Fudan University, Shanghai, China · Beijing Institute of Technology, Beijing, China · Chongqing Ant Consumer Finance Co,. Ltd
AI Scientists for Building Virtual-Cell Models
Aug 4, 2025Xiangru Tang, Jiapeng Chen, Daniel Shao +12AI Agents for Scientific Discovery Autonomous Research Agents
Yale University · University of Pennsylvania · Helmholtz Zentrum München +3
AblateCell: A Reproduce-then-Ablate Agent for Virtual Cell Repositories
Apr 21, 2026Xue Xia, Chengkai Yao, Mingyu Tsoi +10AI Coding Agents ML Reproducibility
Shanghai Artificial Intelligence Laboratory · The Hong Kong University of Science and Technology (Guangzhou) · University of California San Diego +4

Sep 29, 2026J/K move · Enter open · S save

OmniVCBench: Benchmarking Evidence-Grounded Multimodal Reasoning Towards AI Virtual Cells

Manyu Li, Xunkai Li, Yongfu Xiong, Yi Liu+2

Fudan University, Shanghai, China · Beijing Institute of Technology, Beijing, China · Chongqing Ant Consumer Finance Co,. Ltd

Visual Question Answering Multimodal Reasoning AI Agent Benchmarks Multimodal Model Evaluation

Open paper AskPDF arXiv

Artificial Intelligence Virtual Cells (AIVCs) are envisioned as scientific agents that simulate cellular responses, explain underlying mechanisms, and support hypothesis-driven discovery. Existing AIVC benchmarks, however, operate primarily at the simulation layer, motivating complementary evaluation of how models interpret experimental evidence and formulate biological hypotheses. We introduce OmniVCBench, a figure-centric, source-traceable benchmark for the interpretation component of an AIVC. It contains 6,077 curated single- and multi-subfigure question--answer pairs derived from figures and experimental contexts in the scientific literature. Guided by Bloom's taxonomy, we instantiate interpretation-layer counterparts of the AIVC Predict--Explain--Discover agenda through three scientific reasoning tasks. We further introduce AIVC-Judge, a task-conditioned MLLM-as-a-judge framework with category-specific, reference-aware rubrics for evaluating open-ended responses. A complementary Model-Derived Hard-Negative Mining (MDHNM) strategy converts plausible errors observed during model inference into MCQ distractors for lower-cost evaluation. Within the evaluated heterogeneous model pool, MCQ accuracy correlates positively with AIVC-Judge scores, providing a complementary view of performance alongside open-response evaluation. Code and data demo are available at https://anonymous.4open.science/r/OmniVCBench.

Abstract

Scientific Reasoning in Language Models

Explore similar work

OmniVCBench: Benchmarking Evidence-Grounded Multimodal Reasoning Towards AI Virtual Cells

AI Scientists for Building Virtual-Cell Models

AblateCell: A Reproduce-then-Ablate Agent for Virtual Cell Repositories