acceptodds
Withdrawn from ICLR 2027

DynamicVC: A Trajectory-Ground Biological World Model For Virtual Cell Simulation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Building predictive virtual cells promises to uncover the mechanisms governing cellular behavior. Existing language model-based virtual cell systems rely largely on retrieved knowledge, while continuous generative models are typically evaluated only through their endpoint predictions. We introduce DynamicVC, a framework that converts conditional flow matching trajectories into structured evidence for perturbation reasoning. DynamicVC summarizes control-relative expression changes, vector field patterns along sampled ODE trajectories, endpoint effects, and directional agreement across stochastic initializations in a SimContext (Dynamic Simulation Context). It combines model-derived evidence with retrieved biological context to predict whether a drug induces differential expression of a target gene and to generate an evidence-grounded explanation.We evaluate DynamicVC on 55,400 queries involving held-out drugs across five cell lines. It achieves 79.61% AUROC and 64.42% Macro F1, improving over the strongest baseline for each metric by 12.82 pp and 11.06 pp, respectively. Comparisons across reasoning backbones and numerical evidence sources further show that context-specific response predictions consistently improve language model-based perturbation reasoning. DynamicVC is available at https://anonymous.4open.science/r/DynamicVC-50BF.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Virtual Cell, Flow Matching, Dynamic Evidence, Large Language Model

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_KJo8OpenReview ↗
Rating
1
Confidence
3
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

DynamicVC extends VCWorld with SimContext, which converts scDFM-predicted expression states, generative-path features, and consistency information into evidence for an LLM to predict differential gene expression under drug perturbations. Experiments cover held-out drugs across five cell lines, different LLM backbones, and alternative evidence configurations. The authors report improvements over VCWorld, but the biological interpretation and independent contribution of the dynamic evidence remain insufficiently established.

Critical strengths
  1. The paper presents a concrete framework for combining numerical perturbation predictions with LLM-based reasoning. SimContext makes simulation outputs accessible to the LLM through structured statistics, rules, and text templates.

  2. The experiments cover multiple LLM backbones and cell lines, and compare different evidence configurations. The reported improvements suggest that incorporating simulation-derived information can benefit prediction.

Critical weaknesses
  1. The biological meaning of the generative trajectory is unclear. scDFM generates expression states from noise, so intermediate states do not necessarily represent a biological transition from control to perturbed cells. Features such as “delay” and “reversal” may reflect the generative process rather than actual response dynamics. The paper does not establish why these features should be interpreted as biological evidence, weakening the central motivation of trajectory-grounded reasoning.

  2. The contribution of dynamic evidence is not convincingly isolated. The endpoint-only versus dynamic comparison is reported for scDFM, but not for CellFlow. The results therefore do not establish whether intermediate states provide consistent gains beyond endpoint predictions or whether the reported improvement depends on the particular generative model.

  3. Inconsistencies in reported metrics raise concerns about result reliability. The class-specific metrics in Appendix Table C.1 imply balanced-accuracy values for STATE and X-Cell that differ from those reported in Table 1 under the stated aggregation convention. These discrepancies are not explained, making the affected baseline comparisons difficult to verify and raising concerns about reproducibility.

Questions for authors
  1. What biological information do the intermediate generative states and features such as “delay” and “reversal” capture beyond the final predicted expression state? What assumptions justify interpreting them as evidence of perturbation dynamics rather than artifacts of the generation process?

  2. Could the authors provide matched endpoint-only versus dynamic comparisons for CellFlow and, if feasible, other generative models? Without these comparisons, it is difficult to determine whether the benefit of dynamic evidence generalizes beyond scDFM.

  3. Could the authors reconcile the balanced-accuracy discrepancies between Table 1 and Appendix Table C.1, clarify the metric definitions and aggregation procedures, and provide corrected results if necessary? Do these corrections affect the paper’s conclusions?

Reviewer_zbaZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

DynamicVC adds numerical evidence from a cell prediction model to an LLM-based perturbation prediction system. Its main contribution, SimContext, summarizes intermediate flow states, endpoint predictions, and consistency across random initializations. The LLM combines these summaries with retrieved biological knowledge to predict whether a drug induces differential expression of a queried gene.

Critical strengths
  1. Using numerical predictions to check biological hypotheses and support an LLM is a reasonable idea. Combining these predictions with the LLM’s pretrained knowledge could bring substantial gains.

  2. SimContext offers a reasonable way to organize and use evidence from a simulator. This approach could also be useful in other fields that combine LLMs with scientific prediction models.

Critical weaknesses
  1. The paper needs to explain why numerical predictions help and whether they introduce additional errors. The analyses and ablations address some of my concerns through the reported results. Still, I would like a more detailed analysis of where the gains come from, or at least case studies.
  2. The writing also needs improvement. It shows strong signs of AI-generated text, such as repeated defensive wording. Sections 4.2–4.3 largely repeat numbers without adding insight. This writing style puts a heavy burden on readers and further weakens the paper’s credibility.
Questions for authors
  1. Could you provide a more concrete quantitative analysis of when the simulator corrects the LLM and when it leads the LLM astray?
  2. The main table and appendix report inconsistent results. Specifically, the BAcc values calculated from Eq. (8) and Table C.1 do not match Table 1. This gives the impression that the baseline scores may have been deliberately understated. It makes the experimental results less trustworthy.
Reviewer_ACKGOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a new quite thought provoking idea to combine LLMs and perturbation generative models for perturbation prediction. The authors suggest to use the trajectories of a single cell perturbation flow matching model as context to a language model to improve performance. The authors test their idea on multiple cell lines and show improvement over existing generative model baselines.

Critical strengths
  • Combining generative models with large language model approaches for perturbation prediction is an interesting and promising direction.
  • The authors show strong empirical improvement on the perturbation prediction task showing that combining generative-based approaches and language models can improve performance.
Critical weaknesses
  • Conditioning on the trajectories of the flow matching model does not make any sense to me. A flow matching model is not trained to have particular trajectories, it is trained to match a given probability distribution at the endpoint. In fact you can achieve the same distribution with a wide variety of interpolants. So reading into the intermediate trajectories does not make any biological or mathematical sense.

  • A very important comparison in the paper is the ablation in Table 3. However, this comparison is currently misleading as the endpoint lacks important information that is not related to the intermediate dynamics. The authors should include stochastic consistency, cell-population dispersion, effect magnitude categories, and additional natural-language descriptions of the endpoint, not only the mean. As of now, it's not clear whether the ablation is meaningful.

Questions for authors

Can the authors make a case as for why the trajectories of the flow matching model should be informative at all for this biological prediction task ?

Can the authors clarify in great details how the ablation of Table 3 is constructed and ensure the comparison is fair and non-biased ?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.