acceptodds
Under review as a conference paper at ICLR 2027

PerturbBridge: Direct Phenotypic Supervision for Cellular Perturbation Generation

Abstract

Virtual cell models should capture phenotypes induced by specific perturbations. However, image-based models are typically trained and evaluated to match real-image distributions, which alone does not explicitly prioritize perturbation-specific biological responses. Recent approaches add biological signals through reinforcement-learning post-training or inference-time steering, raising a central question: where should biological response information enter the generative pipeline? We introduce PerturbBridge, a stochastic data-to-data bridge model that directly supervises perturbation-specific phenotypes in predicted perturbed images during training. PerturbBridge jointly optimizes bridge regression and differentiable biological response losses, with biological gradients passing through a frozen image encoder. We further introduce perturbation-response evaluations based on downstream classification and retrieval. Across three microscopy benchmarks for perturbation profiling and subcellular imaging, PerturbBridge surpasses CellFlux in biological fidelity, image quality, and sampling efficiency. On BBBC021, mechanism-of-action classification improves by 5.66 and 7.5 percentage points with morphological and learned visual features, respectively. On JUMP and AllenCell, retrieval Hit@10 and Hit@1 improve by 1.77 and 7.15 percentage points, respectively. On BBBC021 and JUMP, PerturbBridge achieves lower FID with 80% fewer model evaluations than CellFlux. Overall, these results show that directly supervising perturbation responses during training improves phenotypic fidelity while enabling efficient cellular image generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Virtual Cells, Perturbation Modeling, Image Generation, Diffusion Bridges

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_eEsgOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PerturbBridge, a stochastic bridge model for generating perturbation-induced cellular images from control images. Building on an ECSI-based bridge formulation, the method introduces two additional objectives—condition-moment matching and biological contrastive loss—to directly supervise generated phenotypes in a pretrained image embedding space. The authors evaluate PerturbBridge on three microscopy datasets using image quality, mechanism-of-action (MoA) classification, and perturbation retrieval.

Critical strengths
  1. Practical and biologically relevant problem. The paper addresses an important challenge in virtual cell modeling: generating perturbation-specific phenotypes beyond matching overall image distributions.

  2. Comprehensive empirical evaluation. The paper evaluates PerturbBridge across three microscopy benchmarks using complementary metrics, including image quality, MoA classification, and perturbation retrieval.

  3. Improved sampling efficiency. PerturbBridge achieves better image quality than CellFlux with substantially fewer model evaluations (20 vs. 100), demonstrating improved sampling efficiency.

  4. Clear presentation. The manuscript is well written and easy to follow, with clear conceptual figures illustrating the motivation and proposed method.

Critical weaknesses
  1. Limited methodological contribution. PerturbBridge largely builds upon existing stochastic bridge formulations and model architectures, with its main contribution being the addition of condition-moment matching and biological contrastive objectives. Both rely on established techniques, and the methodological novelty beyond integrating these losses into bridge training remains unclear.

  2. Insufficient justification of the biological objectives. Given that the two additional losses constitute the main method contribution, their motivation is insufficiently justified. While the authors motivate the proposed loss with FID ranking reversal (Appendix A.1), they provide no empirical evidence that it occurs on the evaluated datasets or how often it affects existing generative models. Moreover, the rationale for condition-moment matching remains unclear: the theoretical analysis considers feature differences between individual images rather than the population-level statistics used in the proposed loss.

  3. Potential overlap between training objectives and evaluation metrics. The condition-moment matching loss minimizes differences in feature means and variances using Inception-v3, which is also used to compute FID. Since FID measures differences in feature means and covariances, improvements in FID are partly expected. Similarly, the contrastive loss explicitly encourages MoA separation, which closely aligns with the MoA classification evaluation. Although the authors also evaluate using DINOv2 and CellProfiler features, it remains unclear to what extent the reported gains reflect broader biological fidelity rather than properties directly encouraged during training.

  4. Insufficient justification of the feature encoder choice. The proposed biological objectives rely on a frozen Inception-v3 encoder, originally designed for natural images. It is unclear why the authors chose this encoder over alternatives such as DINOv2, v3, or domain-specific models trained on Cell Painting data (e.g., Phenom1 MAE [1] or any other unsupervised approaches). The authors should clarify this design choice and assess how sensitive the results are to the encoder choice.

Reference

[1] Kraus et al. Masked Autoencoders for Microscopy are Scalable Learners of Cellular Biology, CVPR 2024

Questions for authors
  1. How do the two biological objectives interact with the bridge regression loss, and how do they influence the denoising process?

  2. What motivates the use of Mahalanobis distance rather than standard cosine similarity for perturbation retrieval? How sensitive are the retrieval results to this choice of distance metric?

9 comments

Response to Reviewer eEsg

Thank you for recognizing the importance of the problem, the breadth of our experiments, and the sampling efficiency of PerturbBridge. Your comments helped us explain the methodological contribution more clearly and prompted us to run additional experiments on the biological objectives and the choice of feature encoder.

W1. Methodological contribution

Thank you for raising this concern. We would like to clarify the methodological contribution of our paper. The stochastic bridge formulation, the backbone, and the individual losses all build on existing techniques. PerturbBridge introduces a training framework that directly supervises perturbation-specific phenotypes in the perturbed endpoint image predicted by the model. Our contribution is to bring phenotypic response learning and generative training together by applying both to the same endpoint prediction. We use an existing bridge formulation and existing individual loss functions.

Existing methods introduce biological information at different stages. CellFluxRL uses biological rewards to update a pretrained generator during reinforcement learning (RL) post-training. SimuSDB uses phenotypic scores to select candidate transitions during sampling. PerturbBridge uses the CellFlux backbone and incorporates biological response objectives directly into bridge training. Bridge regression and phenotypic supervision act on the same predicted perturbed image. This lets the generator learn perturbation-associated phenotypes during training without running a biological evaluator or selecting candidates during sampling.

Our ablations support this approach. On BBBC021, adding biological supervision to bridge regression increased MoA accuracy on unseen compounds from 50.7% to 59.0% under the Inception-based evaluation protocol. We also compared the three ways of introducing supervision on the standard split using DINOv2-L/14 features. PerturbBridge does not use these features for phenotypic supervision.

MethodStage of biological supervisionMoA accuracy (%) ↑
CellFluxRLRL post-training68.1
SimuSDB + SCG (our reproduction)Sampling69.6
PerturbBridgeGenerative training74.9

These results support directly supervising the predicted phenotype during training. On this biological endpoint, this approach achieved higher accuracy than the post-training and sampling-time methods we evaluated. We hope that this clarification and the experimental results will help you reconsider your assessment of our methodological contribution.

W2. Motivation and empirical evidence for the biological objectives

Thank you for asking us to provide empirical evidence for the ranking reversal discussed in Appendix A.1. In both CellFluxRL and our own BBBC021 experiments, FID and MoA classification accuracy give different model rankings.

An example where FID and biological performance give different rankings. In CellFluxRL, RL post-training increased MoA classification accuracy from 61% to 66%, while FID worsened from 20.36 to 24.01 (Table 1 and Section 5.2). The two metrics therefore favor different models under that study’s evaluation protocol. These values come from Table 1 and Section 5.2 of CellFluxRL.

We observed the same ranking reversal on BBBC021.

Training objectiveFID ↓MoA accuracy ↑
Condition-moment matching16.33367.4%
Condition-moment matching + biological contrastive supervision16.41372.4%

Condition-moment matching alone gave a lower FID, while adding the biological contrastive loss improved MoA accuracy by 5.0 percentage points. FID alone therefore cannot tell us whether a model generates the phenotype associated with the requested perturbation.

Other microscopy studies have also found that a difference in FID need not correspond to a similar difference in biological performance. In Biological Congruence, two models had FID scores of 16.131 and 20.853, while their cell-type classification accuracies were close, at 85.1% and 84.4%, respectively. This also supports reporting biological measures alongside image-distribution metrics. See Table 1 of Biological Congruence.

Why match population moments? Our motivation comes from the unpaired nature of perturbation data. Control and treated measurements usually come from different cells, so we do not observe the same cell before and after treatment. CellOT and CellFluxRL explicitly discuss this setting (see CellOT and Section 3 of CellFluxRL). We therefore use population feature statistics under each perturbation condition as biological supervision targets, without requiring cellwise correspondence. Matching means constrains the average population phenotype. Matching the variance of each feature also constrains how much that feature varies across cells. These two simple, differentiable constraints describe some properties of the population, although they do not fully characterize its distribution. Bridge regression still uses sampled image pairs to construct training states. These pairs do not represent observed trajectories of the same cell.

You are right that the individual-image analysis in Appendix A.2 does not derive the population-moment objective. That analysis explains why we need supervision that reflects differences in features. The motivation for population-level constraints comes from the unpaired nature of the data. We will clarify the role of each in the revision.

W3. Potential overlap between training objectives and evaluation metrics

Thank you for raising this concern. To test whether the improvement extends beyond the Inception-v3 features used for supervision, we computed FDD using DINOv2 features on BBBC021.

ModelDINOv2-FDD ↓
CellFlux207.17
PerturbBridge173.76

Computing FDD requires neither MoA labels nor Inception features. The lower FDD shows that the distribution of images generated by PerturbBridge is also closer to the real image distribution in an independent feature space.

We also tested whether the generated images reproduce the morphological changes observed in real cells under different perturbations relative to DMSO controls on the same plate. We used nuclear roundness, nuclear area, and total nuclear DNA signal as direct measures of the biological effect. PerturbBridge had lower DMSO ratio errors than CellFlux on all three measures.

Individual morphology metric ↓CellFluxPerturbBridge (ours)Difference (PerturbBridge − CellFlux)
Nuclear roundness / boundary regularity: DMSO ratio error0.03570.0160−0.0197
Nuclear area: ratio error relative to DMSO0.12900.0905−0.0385
Total nuclear DNA signal: DMSO ratio error0.14910.1142−0.0349

Following your suggestion, we also trained a new microscopy encoder from scratch on real BBBC021 training images, using compound identity as the classification label. We then froze the encoder and used its morphological features to supervise the generator. We continued to use DINOv2 and CellProfiler as independent representations for evaluation.

W4. Choice of feature encoder and sensitivity to that choice

We initially chose Inception-v3 because it already serves as a reference representation in cell-image generation. CellFlux [1] uses Inception features for MoA evaluation, and MorphDiff [2] uses them to evaluate image distributions. This choice allowed us to follow existing evaluation pipelines without assuming that Inception-v3 is optimal for microscopy images. To test how sensitive the results are to the choice of supervising encoder, we trained a ResNet-18 from scratch on drug-treated cell images from the BBBC021 training split.

[1] Zhang, Y., et al. CellFlux: Simulating Cellular Morphology Changes via Flow Matching. ICML, 2025.
[2] Wang, X., et al. Prediction of cellular morphology changes under perturbations with a transcriptome-guided diffusion model. Nature Communications, 16, 8210, 2025.

Each image was labeled with the compound used to treat the cell. We trained the encoder with cross-entropy loss to predict which of the 26 compounds had been applied, based on the cell’s morphology. For each compound, we held out approximately 10% of its training images for validation and used the rest to train the encoder. Test images were excluded from training and checkpoint selection. Training augmentations included flips, rotations by multiples of 90°, and small changes in channel intensity.

After training, we froze the encoder. Following the training procedure in the PerturbBridge paper, we used features from this newly trained encoder to supervise the model on the BBBC dataset. The results are shown below.

ModelDINOv2-L/14 MoA accuracy ↑Full CellProfiler MoA accuracy ↑
CellFlux67.4%72.2%
Original PerturbBridge74.9%77.8%
PerturbBridge with the new microscopy encoder74.4%81.0%

The additional evaluations used 5,120 samples. The classifiers were fitted on real training data and held fixed. Compared with the original PerturbBridge, the new encoder improved MoA accuracy with the full CellProfiler feature set, while DINOv2 accuracy remained similar. This shows that switching to a more suitable encoder can improve some aspects of the model’s performance. Thank you for this insightful suggestion. We will include the encoder’s training data, architecture, feature extraction procedure, and results in the revised manuscript.

Q1. How do the two biological objectives interact with the bridge regression loss, and how do they influence denoising?

Bridge regression trains the denoiser to recover the perturbed cell image in pixel space. Both auxiliary losses act on that same prediction. Condition-moment matching brings the feature means and variances of generated cells closer to those of real cells under the same perturbation. Prototype contrast draws the predicted features toward real prototypes with the same MoA. During training, gradients from both losses pass through the frozen encoder to update the denoising network.

Standard BBBC021 split. We evaluated 5,120 generated images per variant, using Inception-v3 features for MoA classification.

Training objectiveFID ↓MoA accuracy ↑Macro-F1 ↑Weighted-F1 ↑
Bridge regression + condition-moment matching16.367.4%43.1%67.2%
Bridge regression + prototype contrast17.469.4%44.4%69.4%
Bridge regression + condition-moment matching + prototype contrast16.472.4%46.9%72.0%

Condition-moment matching alone gave the lowest FID (16.3), while prototype contrast alone gave higher MoA accuracy (69.4% vs. 67.4%). With both losses, FID remained close to the lowest value (16.4), and the model achieved the highest MoA accuracy (72.4%) and both F1 scores.

Reviewer_pKNSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PerturbBridge incorporates perturbation-specific phenotypic supervision directly into a stochastic data-to-data bridge. It predicts the final perturbed image from an intermediate state and jointly optimizes weighted image regression, condition-level feature-moment matching, and an InfoNCE-style prototype loss through a frozen Inception-v3 encoder. Biological evaluation during sampling is unnecessary. Experiments on BBBC021, JUMP, and AllenCell report improved MoA classification, perturbation retrieval, and most distributional metrics relative to CellFlux, including unseen-compound MoA evaluation. Bridge-only and CellFlux+bio ablations support contributions from both the bridge and response supervision. An ECSI/DBIM sampler gives favorable FID at 20 logical evaluations, compared with CellFlux at 100.

Critical strengths
  1. The biological gains extend beyond the representation used for supervision. Although training uses frozen Inception-v3 features, BBBC021 MoA accuracy improves over CellFlux from 72.2% to 77.8% with CellProfiler and from 67.4% to 74.9% with DINOv2. Agreement across these complementary evaluation representations makes improved perturbation fidelity more convincing than improvement in the supervising feature space alone, and is a substantive reason to consider acceptance.

  2. Direct supervision of predicted perturbed images has a useful empirical benefit. Adding biological supervision to the bridge raises unseen-compound MoA accuracy from 50.7% to 59.0%, while the complete model also improves retrieval on JUMP and AllenCell. This integration produces measurable gains beyond bridge regression alone and requires no biological evaluator during sampling. These results support the practical value of the proposed training approach.

Critical weaknesses
  1. The implemented predictor does not match the conditioning used in the bridge derivation. The score formula requires a target expectation conditioned explicitly on the control endpoint, whereas the network receives only the intermediate state, time, and perturbation. Appendix B.4 acknowledges that equivalence is unestablished. Without testing explicit endpoint conditioning, it remains unclear whether this approximation compromises source information or sampling reliability, weakening the methodological justification.

  2. The central claim about where biological supervision should enter is not fully established. Table 3 includes CellFluxRL, and Table 4 reports published RL and SimuSDB distribution scores, but biological comparisons do not consistently match backbone and total compute across supervision stages. Because the method combines established components, this missing controlled comparison materially limits the demonstrated advance over alternative uses of biological information and weighs against acceptance.

Questions for authors
  1. What changes when the control endpoint is supplied explicitly to the predictor? Please compare this variant with the current model to assess whether the conditioning approximation in Appendix B.4 affects perturbation fidelity or source-information preservation.

  2. Can training-time supervision, RL post-training, and inference-time steering be compared using a common backbone and matched total compute on the same biological endpoint, preferably one not based on the supervising Inception representation?

  3. What are the separate contributions of condition-moment matching and prototype contrast? Please report moment-only and contrast-only ablations under the same training and sampling settings, particularly for unseen-compound MoA classification.

  4. Table 4 contains missing results for CellFluxRL and SimuSDB. Could the authors clarify whether available checkpoints or generated samples could support completing key comparisons under the same evaluation protocol, particularly the DINOv2-based metrics? If this is infeasible, please explain the practical limitations and clarify how these gaps constrain conclusions about the relative advantages of training-time supervision, RL post-training, and inference-time steering.

5 comments

Response to Reviewer pKNS

Thank you for reading our paper carefully and for recognizing the gains in biological metrics across feature spaces and the value of directly supervising perturbation phenotypes during training. Your question about endpoint conditioning prompted us to change how the predictor receives the control image. Your comments also prompted us to compare three ways of using biological supervision and examine the contribution of each auxiliary objective.

W1 / Q1. What changes when the control endpoint is supplied explicitly to the predictor?

Thank you for pointing out the difference between the predictor’s inputs and the conditioning variables in the bridge derivation. In the original model, the control image enters through the intermediate bridge state and the sampling updates. Following your suggestion, we trained and evaluated a variant that also supplies the control endpoint directly to the predictor.

We encode the control image xTx_T as a global feature vector and add it to the predictor’s conditioning embedding. We use this input during both training and sampling, giving the predictor direct access to a representation of the control image at every denoising timestep.

The table shows the results on the standard BBBC021 split. All three mechanism-of-action (MoA) classification metrics use Inception-v3 features and are reported as percentages.

ModelFID ↓MoA accuracy ↑Macro-F1 ↑Weighted-F1 ↑
CellFlux18.864.9%42.1%63.8%
Original PerturbBridge16.472.4%46.9%72.0%
PerturbBridge with explicit global conditioning on xTx_T15.475.1%49.0%74.6%

Explicit endpoint conditioning lowered FID and improved all three MoA metrics. These results show that giving the predictor direct access to the control image representation can further improve perturbation phenotype fidelity on BBBC021. Thank you again for this insightful suggestion, which helped us further improve our model’s performance.

W2 / Q2 / Q4. Comparing training-time supervision, RL post-training, and inference-time steering

Thank you for suggesting that we compare these methods in terms of backbone, compute budget, and a common biological endpoint independent of the features used for supervision.

When we wrote the paper, CellFluxRL had not released a checkpoint, and SimuSDB had not released its code or a checkpoint. We therefore lacked results for some metrics. For this response, we reproduced the results using the CellFluxRL code and the method description in the SimuSDB paper. PerturbBridge and CellFlux use the same 56.2M-parameter backbone. CellFluxRL also builds on CellFlux. Because the SimuSDB code was unavailable to us, we reproduced the method following the architecture and sampling configuration described in its paper. Our reproduction uses a conditional U-Net with approximately 56.2M parameters.

The three methods use biological information at different stages. PerturbBridge directly supervises the predicted perturbed image during bridge training. CellFluxRL uses biological rewards to update the generator during RL post-training. SimuSDB uses a phenotypic cost to select candidate transitions during sampling. PerturbBridge requires no biological evaluator or candidate selection at inference.

Following the convention in Figure 5 and Appendix B.4 of our paper, we count each classifier-free guidance (CFG) prediction as one logical model evaluation. PerturbBridge uses 20 evaluations per sample, and single-sample CellFluxRL uses 40 (two per Heun step). Our reproduction of SimuSDB + SCG uses 100 CFG predictions and 765 additional conditional U-Net evaluations for candidate selection. Counting the conditional and unconditional branches separately for all three methods gives 40, 80, and 965 U-Net evaluations per output, respectively. CellFluxRL therefore uses 2 times as many generator evaluations as PerturbBridge, and SimuSDB + SCG uses 24.1 times as many. SimuSDB + SCG also uses a separate classifier to score 816 candidates.

MoA classification with DINOv2-L/14 and CellProfiler features

As you requested, we compared all four models on BBBC021 using the same DINOv2 MoA classification protocol. The results for CellFluxRL and for our reproduction of SimuSDB + SCG are from the new evaluation. The last column gives MoA accuracy with CellProfiler features for all four models.

ModelDINOv2-L/14 accuracy ↑DINOv2-L/14 Macro-F1 ↑DINOv2-L/14 Weighted-F1 ↑CellProfiler accuracy ↑NFE ↓
CellFlux67.4%43.4%66.6%72.2%200
CellFluxRL68.1%44.2%67.3%72.1%80
SimuSDB + SCG (our reproduction)69.6%45.3%70.1%65.0%965
PerturbBridge (ours)74.9%46.9%74.3%77.8%40

These results show that, with matched backbones, PerturbBridge outperformed the other baselines and had the fastest sampling speed.

Q3. What are the separate contributions of condition-moment matching and prototype contrast?

Thank you for suggesting that we evaluate the two losses separately. On the standard BBBC021 split, we compared three variants using condition-moment matching alone, prototype contrast alone, or both auxiliary losses together. We generated 5,120 images per variant using the same sampling and scoring protocol. All three variants retain the bridge regression loss. The table reports MoA classification using Inception-v3 features.

Training objectiveFID ↓MoA accuracy ↑Macro-F1 ↑Weighted-F1 ↑
Bridge regression + condition-moment matching16.367.4%43.1%67.2%
Bridge regression + prototype contrast17.469.4%44.4%69.4%
Bridge regression + condition-moment matching + prototype contrast16.472.4%46.9%72.0%

Condition-moment matching constrains the feature distribution of cells under the same perturbation. It gave the lowest FID when used as the only auxiliary loss. On BBBC021, prototype contrast treats condition prototypes that share the same MoA as positives and draws the predicted phenotype toward those prototypes. Prototype contrast alone gave better MoA classification results than condition-moment matching alone. Combining the two kept FID close to the moment-only result and gave the highest MoA accuracy and both F1 scores. On the standard split, these ablations support complementary roles for the two objectives.

The unseen-compound split follows the CellFlux evaluation protocol used in Table 3 of the paper. The two middle rows are the new single-loss ablation results.

Unseen-compound settingMoA accuracy ↑Macro-F1 ↑Weighted-F1 ↑
Bridge-only50.7%37.4%46.1%
Bridge regression + condition-moment matching54.2%41.8%50.4%
Bridge regression + prototype contrast56.0%43.7%52.5%
Full PerturbBridge59.0%47.3%55.2%

Thank you again for the constructive feedback. We will include the training input, comparison protocol, and ablation experiments you suggested in the final version. We are glad to discuss and address any remaining question, and if you find your concerns resolved, we would appreciate if you could reflect this in your evaluation.

Dear Reviewer pKNS,

Just a friendly reminder as the discussion deadline approaches. We hope our responses have addressed your concerns.

Following your suggestion, we gave the predictor direct access to the control image features. This lowered FID and improved MoA classification. Thank you for the constructive suggestion. It helped us improve the model’s performance.

We also added the missing comparisons with CellFluxRL and SimuSDB. PerturbBridge achieved higher MoA accuracy with both DINOv2 and CellProfiler features and used fewer model evaluations per sample. Regarding your question about the losses, we also reported the individual contributions of moment matching and prototype contrast.

We look forward to hearing your thoughts and would be happy to answer any remaining questions. Thank you again for your time!

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

PerturbBridge generates perturbed-cell images from control images using a stochastic bridge with direct phenotypic supervision during training. The authors evaluate the generated images through MoA classification, perturbation retrieval, and image-distribution metrics on three datasets.

Critical strengths

Conditional on perturbation image generation being a useful task, this appears to provide a useful alternative strategy. The biological supervision improves several reported endpoints, including evaluations with features that were not used for supervision. The method could potentially serve as a component in computational workflows, although I have not inspected the supplied code to assess ease of reuse.

Critical weaknesses

My main reservation concerns the biological value added by image synthesis. If the intended application is perturbation prediction, the method should be compared with simple approaches that directly predict the relevant profiles or labels from the same available inputs. The comparisons with other image generators establish improvements within image generation, but leave unclear what useful predictive capability generating images adds.

Questions for authors

Can you demonstrate a biological prediction task where synthesizing images improves over a simple baseline that directly predicts the relevant profiles or labels, using the same information and evaluation split?

2 comments

Response to Reviewer Aomu

W1. What biological value does image synthesis add?

Thank you for raising this question. It is central to image-based virtual cell modeling [1]. Direct prediction of a profile or label requires choosing the readout in advance, and that readout summarizes information in the underlying images. Once a cell image has been generated, researchers can analyze the same image with different feature extractors and image-analysis pipelines, including tools that were not used during training. Images can also capture differences between individual cells and spatial features such as cell shape and subcellular organization. Biologists can inspect these features directly. Recent work, including IMPA [2] and MorphDiff [3], has explored image generation for modeling cellular responses to perturbations and for phenotypic screening. The two approaches therefore have overlapping uses, and image generation allows researchers to perform several morphological analyses on the same prediction.

[1] Bunne, C., et al. How to build the virtual cell with artificial intelligence: Priorities and opportunities. Cell, 187, 7045–7063, 2024.
[2] Palma, A., Theis, F. J., and Lotfollahi, M. Predicting cell morphological responses to perturbations using generative modeling. Nature Communications, 16, 505, 2025.
[3] Wang, X., et al. Prediction of cellular morphology changes under perturbations with a transcriptome-guided diffusion model. Nature Communications, 16, 8210, 2025.

For this flexibility to be useful, the generated images must preserve the phenotype induced by the requested perturbation. This is the problem PerturbBridge addresses. Table 1 of the paper reports results on the standard BBBC021 split. With CellProfiler features, mechanism-of-action (MoA) classification accuracy on generated images increased from 72.2% for CellFlux to 77.8% for PerturbBridge. With DINOv2 features, it increased from 67.4% to 74.9%. Neither representation was used for phenotypic supervision. The gains in both feature spaces support using the generated images for downstream analyses with different feature extractors.

We therefore see image synthesis as complementary to predictors designed for specific profiles or labels. A single generated image can serve as the input to several downstream morphological analyses. PerturbBridge supports this use by better preserving perturbation-specific phenotypes across multiple evaluation representations that were not used for supervision.

Q1. Comparison with direct biological prediction

Thank you for this insightful suggestion. We trained an MLP for 30 epochs to predict the CellProfiler profile of a perturbed cell directly from the 2,073-dimensional CellProfiler profile of a DMSO control cell, the compound embedding, and the dose. For the image-generation models, we extracted CellProfiler profiles from the generated images. We used held-out real perturbed cells as the reference for this comparison. The MLP outputs only profiles, so its predictions cannot be evaluated directly using DINOv2-based MoA classification.

MethodMoA accuracy ↑Macro-F1 ↑Weighted-F1 ↑Profile RMSE ↓Effect cosine ↑
Direct MLP82.9%66.7%79.6%0.2830.631
PerturbBridge77.8%73.0%79.3%0.2410.760
CellFlux72.2%65.8%73.2%0.1800.826

These results provide evidence that image-mediated prediction can better predict perturbation-induced changes in CellProfiler profiles than the direct MLP under this protocol, with a lower profile RMSE (0.241 vs. 0.283) and a higher effect cosine (0.760 vs. 0.631). Together with the flexibility discussed in W1, this demonstrates a concrete predictive benefit of image synthesis in the evaluated setting.

Thank you again for the constructive feedback. We will include this analysis in the final version. We are glad to discuss and address any remaining question, and if you find your concerns resolved, we would appreciate if you could reflect this in your evaluation.

Dear Reviewer Aomu,

Just a friendly reminder as the discussion deadline approaches. We hope our response has addressed your concerns.

In our response, we explained how image synthesis supports different uses in virtual cell modeling. Researchers can inspect cell shape and subcellular organization directly and analyze the same generated image with different tools, including ones not used during training. We also added the direct prediction baseline you suggested and found that profiles extracted from our images more closely matched those of real perturbed cells.

We look forward to hearing your thoughts and would be happy to answer any remaining questions. Thank you again for your time!

Comments on the paper

3 comments

G3 – Training-time supervision, RL post-training, and sampling-time guidance (pKNS, eEsg)

We evaluated CellFluxRL using its released code and reproduced SimuSDB + SCG following its published architecture and sampling configuration. All three methods used backbones of comparable size and the same BBBC021 MoA evaluation protocol. DINOv2-L/14 and CellProfiler were independent of the encoder used to supervise PerturbBridge. Sampling cost counts conditional and unconditional U-Net forward passes separately (Table 1 and Figure 5).

MethodStage of biological supervisionDINOv2 MoA accuracyCellProfiler MoA accuracyU-Net passes per image
CellFluxRLRL post-training68.1%72.1%80
SimuSDB + SCG (our reproduction)Sampling69.6%65.0%965
PerturbBridgeGenerator training74.9%77.8%40

G4 – What does generating an image add to biological prediction? (Aomu, eEsg)

Following Aomu’s suggestion, we trained an MLP to predict a treated cell’s CellProfiler profile directly from a control-cell profile, compound embedding, and dose. The MLP achieved higher MoA accuracy (82.9% vs. 77.8% for PerturbBridge). Profiles extracted from PerturbBridge images were closer to those of held-out real treated cells (RMSE 0.241 vs. 0.283) and better captured the direction of the drug effect (cosine similarity 0.760 vs. 0.631).

The same generated images also let us measure individual morphological changes. PerturbBridge reproduced the increase in nuclear area after etoposide treatment, the decrease in nuclear roundness after AZ138 treatment, and the decrease in total nuclear DNA fluorescence after lovastatin treatment. At 10 μM etoposide, nuclear area was 957.50 pixels² in DMSO controls, 1,308.67 pixels² in real treated cells, and 1,305.00 pixels² in generated cells. Table 4 in the main paper reports all three drug conditions.

Finally, following eEsg’s suggestions, we tested the supervising encoder and retrieval metric. Replacing Inception-v3 with a frozen ResNet-18 trained on BBBC021 microscopy images raised CellProfiler MoA accuracy from 77.8% to 81.0%, while DINOv2 accuracy remained similar (74.9% and 74.4%). When we repeated JUMP cross-plate retrieval with cosine similarity, PerturbBridge continued to outperform CellFlux on Hit@1, Hit@5, Hit@10, and mean reciprocal rank.

We thank all three reviewers again for the time they devoted to the paper. We have incorporated the additional baseline comparisons, ablations, batch analysis, and nuclear measurements into the revised manuscript, with changes highlighted in blue.

Global response to reviewers

Dear Area Chair and Reviewers Aomu, pKNS, and eEsg,

Thank you for your careful reviews and for the follow-up discussion. We appreciate your recognition of the gains under feature representations that were not used for supervision and of the reduced sampling cost. Your questions prompted us to test where biological supervision helps, what generating an image adds over predicting a profile directly, and whether the generated cells reproduce measurable drug responses.

G1 – Motivation and methodological contribution (pKNS, eEsg)

Given a control-cell image, a drug, and a dose, what morphological changes should we expect? A generated cell can look realistic even when its most prominent features reflect staining or imaging conditions. Following eEsg’s question about FID, we compared two groups of 3,000 real BBBC021 DMSO control images. With the same treatment, sample size, and evaluation procedure, FID was 6.24 when we pooled batches and split the images at random, and 24.93 when the groups came from separate experimental batches (Table 6). Batch differences alone can therefore change FID substantially, even when both groups contain real cells under the same treatment.

This motivates our choice of what to supervise and when. PerturbBridge predicts the treated-cell image from each intermediate bridge state, so image regression and phenotypic supervision act on the same prediction. To our knowledge, this is the first stochastic bridge framework for cellular perturbation image generation to apply biological supervision directly to the predicted treated-cell image and optimize it jointly with image regression in a single training stage. Sampling requires no biological evaluator.

G2 – What does biological supervision contribute? (pKNS, eEsg)

Control and treated images usually come from different cells. Condition-moment matching therefore supervises the population response by matching feature means and variances under the same perturbation. Prototype contrast helps separate perturbations with different mechanisms of action (MoAs) by drawing predicted features toward prototypes with the same MoA and away from those with other MoAs. On unseen BBBC021 compounds, MoA accuracy was 50.7% with bridge regression alone, 54.2% with condition-moment matching added, 56.0% with prototype contrast added, and 59.0% with both (Table 10). On the standard split, PerturbBridge achieved 6.1 percentage points higher MoA accuracy than CellFlux+bio, which applies the same losses to images reconstructed from velocity predictions (Table 3).

pKNS also asked what changes when the predictor receives the control image explicitly. Supplying a global feature vector from xTx_T at every denoising step lowered FID from 16.4 to 15.4 and raised Inception-v3 MoA accuracy from 72.4% to 75.1% on the standard split (Table 8).

AuthorsOpenReview ↗

PDF revision

We are grateful to the reviewers for their thoughtful suggestions. We have revised the main text and appendix to include additional experiments and analyses discussed in our rebuttal. All revisions are highlighted in blue for ease of reference.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.