acceptodds
Under review as a conference paper at ICLR 2027

CellBridge: Biologically Faithful Cell Morphology Prediction via Source-Conditioned Schrodinger Bridges

Abstract

Cellular Morphology Perturbation Prediction (CMPP) is a fundamental capability toward building predictive virtual cells, aiming to simulate how cellular phenotypes change in response to chemical or genetic interventions. Such predictive models have the potential to accelerate therapeutic discovery by reducing experimental perturbation screening and enabling systematic exploration of cellular response mechanisms. However, existing generative approaches insufficiently account for the source-dependent nature of cellular state transitions and the biological mechanisms which connect perturbations to phenotypic responses. To fill these gaps, we propose CellBridge, a source-conditioned Schrödinger Bridge framework for modeling cellular morphological responses to chemical and genetic perturbations. CellBridge consists of two key components: i) modeling perturbation response as a bridge from an observed control cell state to its perturbed phenotype, with the control image supplied to the network at every step, and ii) incorporating structured compound-target-pathway conditioning to connect molecular perturbations with their downstream biological programs. For the challenging generalization setting, where structurally distinct compounds can induce similar phenotypes through shared biological mechanisms, the pathway representation enables CellBridge to move beyond scaffold similarity and capture mechanism-related responses. Comprehensive experiments on BBBC021, RxRx1, and JUMP demonstrate that CellBridge consistently improves perturbation-specific phenotype recovery across various settings including chemical compound and genetic perturbations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
cell morhology, cell perturbation, schrodinger bridge, diffusion model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_8ncSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CellBridge, a source-conditioned Schrödinger bridge framework for Cellular Morphology Perturbation Prediction (CMPP). The model generates perturbed cellular morphologies from same-plate control images through a stochastic bridge, explicitly conditioning on the observed control image at every reverse step. It further incorporates perturbation embeddings based on Morgan fingerprints or Gene2Vec, together with a structured compound–target–pathway prior, to model chemical and genetic perturbation responses.

Critical strengths

1. A well-motivated generative formulation. Cellular Morphology Perturbation Prediction (CMPP) is an important task in virtual-cell modeling. Formulating perturbation-induced morphological changes as source-conditioned Schrödinger bridge transport provides an interesting approach to modeling heterogeneous cellular responses from unpaired control and treated populations.

2. Promising empirical results across multiple benchmarks. CellBridge achieves improvements in FID, KID, and perturbation recovery on BBBC021, RxRx1, and JUMP. These results demonstrate the potential of the framework for generating perturbation-specific cellular morphologies.

Critical weaknesses

1. The claim of biologically faithful perturbation prediction is insufficiently substantiated. Despite emphasizing biologically faithful prediction in the title, the evaluation primarily relies on FID, KID, perturbation recovery, MoA classification, and DINOv2-based EffectSim. While these metrics assess image fidelity and perturbation-specific phenotypic similarity, they do not directly establish whether the generated morphological changes faithfully reflect the underlying biological mechanisms. Moreover, the structured compound–target–pathway prior on BBBC021 closely aligns with the MoA classes, and all five held-out compounds used to demonstrate novel-scaffold generalization have targets represented in training. Thus, the improvements in MoA accuracy primarily support the effectiveness of transferring known target/pathway information, rather than demonstrating that the model captures biologically faithful, mechanism-driven morphological responses. Overall, the current evidence supports knowledge-guided phenotype generation, but does not sufficiently substantiate the stronger claim of biological faithfulness.

2. The central claim of source-conditioned cellular transitions lacks direct validation. A key contribution of CellBridge is to model source-dependent morphological transitions by conditioning on the observed control image at every reverse step. However, Section 3.2 states that each treated image is randomly paired with a control image from the same plate, without any true cell-level correspondence. Under this independent endpoint coupling, the training data cannot identify the actual transition of an individual control cell under perturbation. The model may primarily exploit plate-level technical variation carried by the source image rather than learning biologically meaningful cell-specific responses.

Questions for authors

What evidence supports the claim that CellBridge generates mechanistically faithful morphological responses, beyond perturbation recovery and image-level similarity metrics?

Reviewer_2o8iOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

CellBridge attempts to solve the cell painting-based virtual cell problem using two components, a diffusion bridge and a biological prior. The goal is to enable generalisation to unseen compounds through a shared mechanism with seen compounds rather than fingerprint similarity. They evaluate on multiple datasets and report mainly FID and KID, which are arguable uninformative metrics for this task when taken alone. They also use MoA classification which is more suited. Their results show superior performance to other models, expecially in terms of MoA classification. While this is a very interesting and important task, I believe a lot more needs to be done to ensure the biological reliability of such models.

Critical strengths
  1. Very interesting and important task.
  2. Perturbation recovery (a biological metric) as the main evaluation, with reasonable controls on the pathway prior. Testing whether the applied perturbation can be recovered from generated images goes beyond FID/KID, and the paper argues for it. Still, I think more could be done. The prior’s ablation includes a content-free token control with matched capacity (Table 13). That control opens its gates as wide as the full model but recovers far less (26.6% vs 67.3%), which is decent evidence that the gain comes from the pathway information rather than added parameters.
  3. The source-swap experiment (Table 15). It shows the network actually uses the control image: swapping it collapses MoA accuracy to 25%. This supports the central architectural claim.
Critical weaknesses
  1. Missing trivial baseline: The headline result is perturbation recovery, but the paper lacks the two baselines the virtual-cell literature has shown are needed to read such numbers. I know that this is a different modality (imaging), but the principals remain.
  • Control passthrough (a no-change baseline): return the paired same-plate control unchanged and score it with every metric. Each treated image is paired with a control from its own plate, and plate layout is tied to perturbation identity (in RxRx1 each plate carries a fixed subset of siRNAs; BBBC021 compounds are tied to particular plates and batches). A DINOv2 probe could therefore recover perturbation identity partly from plate signal carried in from the control, without the model producing any real perturbation effect. All recovery numbers should be reported relative to this floor. See Ando et al. 2017 where accuracy for DMSO mock treatments is higher thhan random chance. Also see Ahlmann-Eltze, Huber & Anders 2025 (Nature Methods) to see why it is important to include these baselines.
  • Perturbation mean effect: compute the mean feature shift between each perturbation’s treated cells and the controls, add it to a new control, and score the result. Feature space is enough for EffectSim, retrieval and the probes; pixel space also works if image metrics are wanted. This simple non-generative baseline often performs surprisingly well, and without it the gains over PhenDiff, IMPA and CellFlux are hard to interpret. Note that in distribution it would score near-perfect EffectSim almost by construction, which itself says something about that metric. See Viñas Torné et al. 2025 (Nature Biotechnology) and Roohani et al. 2025 (Cell). Again, I know these references are for single-cell sequencing data, but the problem transfers.
  1. Generated images classify better than real ones, and the paper treats this as a positive. On BBBC021, CellBridge reaches 82.9–85.5% MoA accuracy against 73.7% for real treated images (Table 8). A generative model that is more diagnostic than the data it imitates is overshooting, not simply succeeding. Likely explanations are that it strips biological heterogeneity or exaggerates the features the classifier relies on. The authors’ defence (A.8: spread 0.86 of real, less nuisance variation) does not settle this. Reduced variation along the directions the classifier uses is exactly what would inflate accuracy, and a total covariance trace can stay high while those directions are compressed. I am not convinced this reflects real biology, and it weakens the use of MoA accuracy as the main measure of fidelity.

Less critical, but important:

  • Key results are in the appendix. The submission limit is 9 pages, and the main text ends just past the top of page 9. Yet the most important results sit in the appendix: MoA accuracy (Table 8, which the text itself calls the result that separates methods most sharply), the identity probes (Tables 5–6), EffectSim/retrieval (Table 9), and the in-distribution prior comparison (Table 10). Table 8 at least should be in the main paper.

  • Metric definitions. FID and KID should be defined in the paper, including which feature backbone is used (Inception? DINOv2?) and at what resolution. Per-perturbation FID on small image sets is also strongly biased by sample size, so the number of images per condition should be reported.

  • Unclear statements. Line 94, “inherit the response of training compounds that share its target rather than that of its nearest scaffold”: please make this precise. It is a design intention, not something shown at that point. Lines 223–226, the meaning of “Part (i) recovers the Gaussian corruption kernel…” and the flow-matching correspondence: what does each limit add in practice? Line 369, “Success tracks chemical similarity more closely than MoA-class coverage”. I was confused by this. I could have misunderstood, so please clarify.

  • Annotation sources. “Curated pharmacology supplies the inputs” (line ~376) needs specific sources and citations for the binding annotations and mechanism signs (ChEMBL? DrugBank? manual curation?).

  • Generalisation claim. “Generalises to novel perturbations” rests on 5 compounds from a single split, and one of them (epothilone B) fails.

  • Broken and missing content. There are broken cross-references (“Section ??” around line 161, “equation ??” at line 259, “Section ??” at line 887 and in A.13). A.14 is empty (“The table behind Section 5.2.”).

Questions for authors
  1. What do control passthrough and the perturbation mean-effect baseline score on FID/KID, the MoA classifier, the DINOv2 identity probes, EffectSim and retrieval, on all three datasets?
  2. How much perturbation identity can a probe recover from the paired control images alone? This would quantify plate-level leakage.
  3. How do you explain generated images beating real images on MoA classification? Can you show per-class confusion matrices, plus within-class spread along the directions the classifier uses (not just the total covariance trace)?
  4. Which feature extractor is used for FID/KID, and how many images go into each per-perturbation FID?
  5. What are the sources for the compound - target annotations and mechanism signs, and how were the 11 BBBC021 pathways chosen given how closely they line up with the MoA classes?
  6. How is “chemically related analog” defined in Section 4.3 (line 369)?
Reviewer_PurWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper generates images of how cells look after a drug or gene perturbation, starting from an untreated control image. It uses a diffusion bridge from the control image to the treated image, plus an optional prior that tells the model each drug's target and pathway. It reports better image quality and better perturbation recognition than prior methods on three datasets, and better results on unseen drugs.

Critical strengths
  • Generated images are much easier to match to the correct perturbation than those of prior methods (Table 5).
Critical weaknesses
  • On unseen drugs, the prior essentially gives the model the drug's mechanism class, which is also what the evaluation classifier predicts (Table 12). The gain may come from this.
  • The bridge method is very similar to the existing I²SB (Liu et al., 2023), which is not cited.
Questions for authors
  1. Does the gain on unseen drugs remain if a baseline (e.g., CellFlux) gets the same prior, or if the annotations are shuffled?
  2. What is new in the bridge compared with I²SB?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to reviewers

Dear Reviewers,

Thank you again for your time and constructive feedback. We have now posted responses to all of your comments, together with the additional analyses and results. As the discussion period is drawing to a close, we would be grateful if you could take a look at our responses at your convenience and let us know whether they address your concerns, or whether any points would benefit from further clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.