- 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
-
Practical and biologically relevant problem. The paper addresses an important challenge in virtual cell modeling: generating perturbation-specific phenotypes beyond matching overall image distributions.
-
Comprehensive empirical evaluation. The paper evaluates PerturbBridge across three microscopy benchmarks using complementary metrics, including image quality, MoA classification, and perturbation retrieval.
-
Improved sampling efficiency. PerturbBridge achieves better image quality than CellFlux with substantially fewer model evaluations (20 vs. 100), demonstrating improved sampling efficiency.
-
Clear presentation. The manuscript is well written and easy to follow, with clear conceptual figures illustrating the motivation and proposed method.
Critical weaknesses
-
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.
-
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.
-
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.
-
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
-
How do the two biological objectives interact with the bridge regression loss, and how do they influence the denoising process?
-
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
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 objective | FID ↓ | MoA accuracy ↑ |
|---|---|---|
| Condition-moment matching | 16.333 | 67.4% |
| Condition-moment matching + biological contrastive supervision | 16.413 | 72.4% |
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.
| Model | DINOv2-FDD ↓ |
|---|---|
| CellFlux | 207.17 |
| PerturbBridge | 173.76 |
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 ↓ | CellFlux | PerturbBridge (ours) | Difference (PerturbBridge − CellFlux) |
|---|---|---|---|
| Nuclear roundness / boundary regularity: DMSO ratio error | 0.0357 | 0.0160 | −0.0197 |
| Nuclear area: ratio error relative to DMSO | 0.1290 | 0.0905 | −0.0385 |
| Total nuclear DNA signal: DMSO ratio error | 0.1491 | 0.1142 | −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.
| Model | DINOv2-L/14 MoA accuracy ↑ | Full CellProfiler MoA accuracy ↑ |
|---|---|---|
| CellFlux | 67.4% | 72.2% |
| Original PerturbBridge | 74.9% | 77.8% |
| PerturbBridge with the new microscopy encoder | 74.4% | 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 objective | FID ↓ | MoA accuracy ↑ | Macro-F1 ↑ | Weighted-F1 ↑ |
|---|---|---|---|---|
| Bridge regression + condition-moment matching | 16.3 | 67.4% | 43.1% | 67.2% |
| Bridge regression + prototype contrast | 17.4 | 69.4% | 44.4% | 69.4% |
| Bridge regression + condition-moment matching + prototype contrast | 16.4 | 72.4% | 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.
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.
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.