acceptodds
Withdrawn from ICLR 2027

CellWright: Mechanism-Conditioned Transport for Verifiable Virtual Cells

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Virtual cell models promise to forecast cellular responses to unseen perturbations in silico, yet recent audits show that most deep models barely beat mean or linear baselines, and even strong systems answer a narrower question than biologists ask: they regress a control-referenced expression shift, under-model real population heterogeneity, and offer no account of *why* the response occurs. We argue that the right target is the full conditional distribution of perturbed cells *together with* the mechanism that generates it. The obstacle is supervision: mechanisms carry no ground-truth labels at scale, so mechanistic reasoning cannot be trained by imitation and readily hallucinates. CellWright recasts this as a verification problem. First, it replaces the latent delta with a *mechanism-conditioned latent Schrödinger bridge*, trained by conditional flow matching over an optimal-transport coupling, that transports an entire control population to the perturbed population and composes velocity fields with an explicit epistatic residual for unseen combinations. Second, an LLM agent emits machine-readable mechanism graphs over the TFDNARNAprotein axis and is optimized by reinforcement learning with verifiable rewards: an energy-based verifier trained by noise-contrastive estimation against hallucinated negatives supplies edge-decomposable rewards, calibrated confidence and a test-time search signal. On held-out perturbations from Norman and Replogle-K562, CellWright reduces energy distance by and over the strongest prior method and improves Top-20 DEG -correlation, while cutting mechanism hallucination from to and expected calibration error from to relative to a free-text LLM reasoner. We also report where it does not win: a neural optimal-transport baseline retains the best energy distance on the chemical corpus, the mean baseline retains the best MSE, and the gain on the Systema-corrected score is not significant. Ablations show that distributional transport drives accuracy, while the verifier drives honesty. Our code is available at [https://anonymous.4open.science/r/CellWright-D4B2/](https://anonymous.4open.science/r/CellWright-D4B2/).

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Virtual Cell, Perturbation Response Prediction, Flow Matching, Mechanism Reasoning, Reinforcement Learning with Verifiable Rewards

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_PKwMOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CellWright, which formulates perturbation response as mechanism-conditioned distributional transport. An LLM agent generates machine-readable mechanism graphs, while an energy-based verifier, trained by noise-contrastive estimation against hallucinated negatives, supplies reinforcement-learning rewards, calibrated consistency scores, and a test-time search signal. The framework also incorporates an epistatic residual for combinatorial perturbation prediction. The authors evaluate predictive performance on genetic and chemical perturbation datasets and further report zero-shot cellular-context generalization, compositional multi-gene prediction, mechanism quality, and module ablations.

Critical strengths

1. A compelling perspective on connecting perturbation prediction with mechanistic reasoning. Formulating perturbation response as mechanism-conditioned distributional transport provides a useful framework for jointly modeling perturbed populations and their proposed mechanisms. In particular, evaluating whether a mechanism graph is coherent with an expression outcome offers a concrete entry point for learning mechanistic reasoning when direct annotations are scarce.

2. An interesting use of reinforcement learning for mechanism learning. Combining an energy-based verifier with GRPO provides a practical route to making mechanistic reasoning trainable without human-annotated mechanism labels. The edge-decomposable rewards offer a structured learning signal for improving the mechanism reasoner, making this a promising response to the supervision challenge identified in the paper.

Critical weaknesses

1. Dependence on the LLM backbone and the contribution of GRPO are insufficiently characterized. MECHREASON incorporates retrieval, structured graph generation, and GRPO, but is evaluated only with Qwen2.5-7B-Instruct (Table 6). It remains unclear how mechanism quality and downstream prediction depend on the backbone’s capabilities. The free-text LLM baseline and removal of MECHREASON do not isolate the contribution of GRPO under otherwise identical conditions. The relationship to related LLM-based perturbation reasoning approaches, such as SUMMER[1], VCWorld[2], and AROMA[3], also warrants clearer discussion.

2. The central evidence for improved mechanism reliability lacks independent validation. Appendix A.2 states that RgroundR_{\mathrm{ground}}, sign accuracy, hallucination labels, and the correctness labels underlying ECE rely on the same grounding index. Consequently, the reported improvements do not sufficiently distinguish better mechanism–outcome coherence from greater conformity to the reward source. The authors acknowledge this bias, but do not provide an independent mechanism reference or a sufficiently controlled comparison under matched retrieval conditions. Moreover, Section 4.4 compares the temperature-calibrated verifier with an uncalibrated score in the ablation, further confounding the attribution of ECE improvements.

3.The incremental contribution of mechanism conditioning to distributional prediction is not sufficiently isolated. Related distributional models including CellFlow and scDFM are cited but not evaluated. More importantly, the ablations in Tables 4 and 11 do not report the central distributional metric, E-dist. Removing MECHREASON entirely also does not distinguish the contribution of meaningful graph structure from that of additional gene and pathway information.


[1] Menghua Wu, Russell Littman, Jacob Levine, Lin Qiu, Tommaso Biancalani, David Richmond, and Jan-Christian Huetter. Contextualizing biological perturbation experiments through language, 2025

[2] Zhijian Wei, Runze Ma, Zichen Wang, Zhongmin Li, Shuotong Song, and Shuangjia Zheng. Vcworld: A biological world model for virtual cell simulation, 2026.

[3] Zhenyu Wang, Geyan Ye, Wei Liu, and Man Tat Alexander Ng. Aroma: Augmented reasoning over a multimodal architecture for virtual cell genetic perturbation modeling, 2026.

Questions for authors

See Weaknesses

Reviewer_KEPAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CellWright predicts single-cell perturbation responses together with signed mechanism graphs. It combines modality-specific perturbation encoders, a control-population encoder, a conditional flow model, an LLM reasoner, and an energy-based verifier. The verifier rewards reasoner training and selects mechanism–prediction pairs at inference. The paper reports improvements in response prediction and mechanism consistency metrics, with additional chemical, context-transfer, and regulatory evaluations beyond the headline genetic benchmarks.

Critical strengths
  1. A useful goal and an interesting connection between prediction and explanation.

Returning a predicted cell population together with an inspectable mechanism graph could be useful for hypothesis generation. Sharing a verifier across reasoner training, confidence scoring, and inference search gives the system a coherent organizing idea.

  1. Promising predictive results with informative ablations and limitations.

Tables 1 and 4 report gains over the included baselines and show that transport and mechanism conditioning both contribute. The paper also reports metrics on which simpler baselines perform better, explores inference cost, and explicitly acknowledges that verifier coherence does not certify biological truth.

Critical weaknesses
  1. The current result misses some of the closest existing approaches. By September 2026, relevant alternatives extended beyond STATE and TxPert. CellFlow (Klein et al., 2025) and scDFM (Yu et al., 2026) are particularly relevant to the distributional prediction claim. Both papers have been cited in Section 2 but absent from the result comparisons. Fawkes et al. (2026) is relevant to the knowledge-graph and RL-trained reasoning component but not discussed. This leaves unclear what CellWright adds over the approaches most closely related to its design.

Secondary: the benchmark presentation does not fully support the breadth of the modeling claims. Table 1 focuses on Norman and Replogle-K562, while sciPlex and Tahoe are put into Appendix Table 8 (why is this?). The paper includes broader evaluations and cellular context transfer, but these do not establish transfer between genetic and chemical interventions. Given the shared perturbation representation, a clearer discussion of what generalizes across intervention types would make the broader contribution easier to assess.

  1. The pipeline combines several established ideas, but it is unclear which capability requires their particular coupling. Table 4 shows losses when components are removed, which is useful evidence. However, removing the reasoner also removes mechanism-graph conditioning, and replacing the flow model with an MLP changes several modeling choices at once. These comparisons do not distinguish learned mechanism reasoning from the benefit of additional biological information, also they cannot show why we need the verifier and search beyond simpler graph-informed conditioning. A comparison using the same resources with retrieved graphs or a simpler conditional flow model would help identify the contribution of the integration.
Questions for authors
  1. Positioning relative to prior work: What is the main advance over flow-based predictors such as CellFlow/scDFM and knowledge-graph reasoning approaches? Which capability demonstrated in this paper is missing from those approaches? A comparison with the closest applicable methods is encouraged.

  2. Value of the full pipeline: Table 4 shows that removing components reduces performance, but removing MechReason also removes graph conditioning, while the MLP replacement significantly changes the predictor. What can the existing results tell about the benefit of learned mechanism reasoning and verifier-guided search beyond access to biological priors?

  3. Since Equation 9 scores each graph against its own conditioned prediction, what evidence shows that a low energy pair is supported by the real response rather than merely internally consistent? How should the confidence score be interpreted given that Appendix A.2 evaluates calibration against indexed edge support? I would suggest testing plausible, and resource grounded graphs that conflict with observed response evidence.

Reviewer_1sQhOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes CellWright, a method with the goal of generating perturbation responses for virtual cells while also verifying them. To this end, CellWright combines:

  1. Protein, DNA and RNA embeddings models via mixture-of-experts to learn a general perturbation embedding.
  2. JEPA-style loss on a set-encoder to learn an embedding over cell population.
  3. RL finetuning with GRPO and LoRA above Qwen2.5-7B-Instruct to provide with a mechanism graph against bioinformatics databases such as GO / Reactome as verifiable rewards.
  4. Diffusion Schrodinger Bridge model which jointly generates both perturbed gene-expression and pathway activity conditioned on the perturbation embedding, set-encoder embedding and mechanism graph.
  5. Energy-based soft-verification and verified guided-search at inference time to find the generated population that most aligns with the mechanism graph.

In order to both predict the gene-expression of sample under perturbation and verify it's biological cogency.

Critical strengths

I found this paper extremely hard to follow and parse, with many moving parts and few very enlightening details on any of them. The strengths of this paper are only hypothetical, as potentially some of it's parts, evaluated independently and described with clarity, could be of interest to the community. I found the use of bio-informatics data-bases to post-train LLMs as interesting, and so is the idea of generating while verifying. However, the current presentation and evaluation make it difficult to assess the value of these components independently.

Critical weaknesses

The main weakness is a lack of methodological focus and clarity. Cellwright takes an “everything but the kitchen sink” approach, combining numerous modeling choices and training procedures without sufficiently explaining what each component does, how the components interact, or why they are biologically motivated. These techniques are introduced in rapid succession, making it difficult to distinguish the paper’s particular methodological novelty from the assembly of existing approaches.

The manuscript needs to establish the specific problem each component addresses and the evidence that justifies its inclusion. As written, the number of moving parts obscures the central scientific contribution and makes it difficult to assess whether the added complexity is necessary and evaluate the scientific contribution of this approach.

Questions for authors
  1. What is the central novelty? Which components or interactions constitute the main contribution, and what specific biological or modeling problem motivates each?
  2. How are the mechanism graphs constructed and used throughout training?
  3. What does GRPO post-training contribute? How does it compare with the same LLM using retrieval?

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

Related papers

Discussion (0)

Sign in to comment.