acceptodds
Under review as a conference paper at ICLR 2027

scPertWGF:Wasserstein Gradient Flow for Single-Cell Perturbation Prediction

Abstract

Predicting how an unseen genetic or chemical perturbation reshapes a cell population is a central problem in functional genomics. Due to the destructive nature of single-cell sequencing, control and perturbed populations are strictly unpaired. While recent deep generative models strive to capture complex cellular heterogeneity, they often fail to outperform simple linear baselines on mean expression shifts. We propose **scPertWGF**, a **W**asserstein **G**radient **F**low framework for **s**ingle-**c**ell perturbation prediction that explicitly models **pert**urbation effects through changes in mean expression and distributional shape. To guide residual generation, a smoothed Wasserstein gradient flow objective aligns generated and observed populations, with flow matching providing complementary supervision of transport velocities. Among the evaluated baselines, scPertWGF achieves the lowest prediction error and highest Pearson delta correlation on four genetic perturbation screens and one drug screen, with the lowest energy distance on four, demonstrating its superior capability to capture both mean shifts and complex population shapes.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell perturbation prediction, Wasserstein gradient flow, Flow matching

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_VN6VOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose wasserstein gradient flows for perturbation modelling. during inference the generated cells are translated towards ridge predicted mean. this method is evaluated on 4 datasets. the main idea is to combine fitted ridge estimator and JKO inspired distributional endpoint objective.

Critical strengths

The only strength i could say is the separation of distribution population mean and centered variation. beyond this decomposition i do not find the methodological or empirical contribution as substantial.

Critical weaknesses

the whole paper is a extremely close to scDFM (ICLR2025) which means conditional flow matching directly in gene-expression space combined with a population-level alignment loss

the only thing authors have changed is to use WGF and ridge estimator. I would call this paper a true representation of A+B research.

no controlled comparison of MMD vs sinkhorn vs WGF endpoint losses on identical backbones and training setup.

missing the most crucial metric which is de-spearman log fold change. cell-eval 2 metrics have also been released by ARC institute, but the authors are still using cell-eval 1. all results are shown with a single seed.

inference-time mean guidance can directly improve pearson delta , L2, MSE/ MAE, and even energy distance without demonstrating better perturbation dynamics. so none of the reported gains show improvement of perturbation specific transcriptional programs

biological validation is missing too like pathway enrichment GSEA or regulator level agreement against resources or databases such as TRRUST.

it is difficult to determine whether the gains correspond to perturbation-specific transcriptional programs rather than improved population-mean calibration.

Questions for authors

authors need to show identical architecture, mean predictor, source distribution, training protocol across same 3 fixed seeds for MMD, Sinkhorn, WGF endpoint losses with statistical tests using Holm corrected values.

report all metrics from cell-eval 2 including de-spearman log fold change for all methods and all ablations. this is needed because the current mean guidance can improve both mean metrics and energy distance without changing covariance.

I truly believe the authors need to make a more impactful contribution to the field rather than taking 2 well established ML concepts and plugging it into perturbation modelling.

perturbation modelling is a crowded field and these incremental changes are not at the level of ML contribution expected at a conference like ICLR.

Reviewer_eA7DOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces scPertWGF as a framework for modelling perturbation predictions. The idea is to decompose responses into a population mean component and a residual component. Here, the mean is predicted with a simple ridge estimation while the residual variation is predicted by a more intricate neural flow model. The residual flow is trained using flow matching and an objective inspired by Wasserstein gradient flows. The manuscript validates scPertWGF using both genetic and chemical perturbation datasets, comparing against both simple and machine learning baselines. Across benchmarks, the method generally improves predictions. Ablations examine contributions of the mean predictor, flow objective, and guidance during inference.

Critical strengths

Strength 1. Generally a strong experimental design and baseline coverage.

Strength 2. Good ablations and explicit exploration of failure cases.

Critical weaknesses

Weakness 1. The paper argues that learning both mean and cellular variation "leads to severe optimization entanglement, making it exceptionally difficult for deep models to simultaneously preserve accurate mean shifts and rich, non-linear population variation". However, the authors do not directly measure the supposed optimization problem, and the presented evidence (Appendix A.4) does not convincingly establish the "optimization entanglement" as the reason for the observed differences.

Weakness 2. Generalization is only tested within the same datasets used in training and evaluation, making it very unclear how well, if at all, the method may transfer across contexts and harder settings.

Questions for authors

Question 1. The paper describes that the ridge head is modelling mean response while the flow is modelling residual variation around that mean. However, with z0=cx+σΣξ−μ^pz_0=c_x+\sigma\Sigma\xi-\hat\mu_p and z1=x−μ^pz_1=x-\hat\mu_p, the flow matching target z1−z0z_1 - z_0 is independent of μ^p\hat\mu_p. Could the separation between mean and residual components be clarified a bit more? The narrative of "mean prediction + residual variation" seems a bit stronger than what the equations suggest.

Question 2. Could the authors provide direct evidence that learning both mean and cellular variation gives "severe optimization entanglement"? An expressive end-to-end model could in principle itself learn how to allocate capacity between mean shifts and population variation. I would like to better understand why (or when) directly imposing a decomposition is preferable rather than just being beneficial in the reported experiments.

Question 3. During evaluation, the authors hold out entire perturbations, but the model remains within a fixed cellular context. How well do the authors expect the method to generalize across harder settings of unseen perturbation/cell contexts?

Minor:

The paper repeatedly describes the modelling to be in raw gene expression space -- however, the experimental section describes a log-normalized space. The log-normalization changes the distribution over the counts but may make modelling more easy (e.g., avoiding count distribution targets). I know it is mainly describing the dimensionality of the space, but could "raw gene expression space" be made clearer to avoid some potential confusion?

Line 314 mentions a "unified evaluation pipeline", but the evaluations seem reasonably method-dependent (Appendix A.3)? This makes results and comparisons a bit less straightforward than what a unified pipeline suggests, so it would be nice to be more explicit about this.

Reviewer_nWD9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes scPertWGF for predicting single-cell responses to unseen genetic and chemical perturbations. It combines a closed-form ridge estimator for population means with a conditional residual flow for cellular heterogeneity. Training uses flow matching, an adversarial distribution-matching objective, and a JKO-inspired proximal penalty. Inference-time guidance calibrates generated populations toward the predicted mean. Experiments across five screens report competitive mean-expression accuracy and improved energy distance on four screens. 15842-2.pdf

Critical strengths
  1. The paper combines a strong mean predictor with a flexible population generator to address the practical challenge of capturing both average responses and cellular heterogeneity. This is an interesting idea.

  2. Evaluation includes four genetic screens and one chemical screen, with simple statistical and recent deep-learning baselines.

  3. The appendices provide substantial implementation detail, same-head population controls, and an informative analysis of the method’s failure on Replogle. And the matched-guidance comparison in Appendix A.4 supports a distributional benefit from residual parameterization when predicted means are identical.

Critical weaknesses
  1. One key idea that conditional gene-space flow matching with population-level supervision already appears in scDFM, while neural JKO methods also predate this submission. The added value of the particular combination needs stronger isolation.

  2. I think we need one decisive experiment: with the same backbone, head, guidance, and training budget using WGF-inspired training versus MMD or energy-distance training. This would directly address the distinction from scDFM. At full mean guidance, all mean-based metrics depend exclusively on the ridge head. Consequently, results on Schmidt and SciPlex do not establish the flow’s contribution to mean accuracy.

  3. On Replogle, removing the adversarial component substantially improves distributional agreement and variance recovery. E.g., removing adversarial matching lowers energy distance from 1.73 to 1.03, while the generated-to-observed variance ratio improves from 0.54 to 0.78.

  4. Report raw and adapted baseline outputs, harmonize gene coverage, and add centered distributional metrics, covariance or variance recovery, and paired comparisons across runs. CMonge is also a relevant missing chemical-perturbation comparator.

[1]scDFM: Chenglei Yu, Chuanrui Wang, Bangyan Liao, and Tailin Wu. “scDFM: Distributional Flow Matching Model for Robust Single-Cell Perturbation Prediction.” ICLR, 2026. Paper. arxiv.org

[2]Neural JKO / WGF: Jaemoo Choi, Jaewoong Choi, and Myungjoo Kang. “Scalable Wasserstein Gradient Flow for Generative Modeling through Unbalanced Optimal Transport.” ICML, PMLR 235:8629–8650, 2024. Paper. proceedings.mlr.press

[3]CMonge: Alice Driessen, Dhruva Abhijit Rajwade, Benedek Harsanyi, Marianna Rapsomaniki, and Jannis Born. “Conditional Monge Gap enables generalizable single-cell perturbation modelling.” Nature Machine Intelligence 8:984–996, 2026. DOI.

Questions for authors
  1. At \alpha=1, all mean-based metrics depend exclusively on the ridge estimator. How much improvement does the residual flow provide over the identical head combined with simple population reconstruction, or competing generators calibrated to the same predicted mean?

  2. Does the proposed objective outperform MMD or energy-distance training under identical backbone, guidance, and compute settings? How do the adversarial and proximal components separately affect performance?

  3. Can you provide raw and postprocessed baseline results using matched gene sets, together with paired comparisons across multiple splits and training seeds, to establish that the reported gains persist beyond method-specific output adjustments?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.