acceptodds
Under review as a conference paper at ICLR 2027

REGINA: Regularized Encoder with Latent Cycle-GAN for In-vitro Neural Cell Perturbation Approximation.

Abstract

Counterfactual inference from unpaired observational data is a fundamental challenge in machine learning. A prominent example arises in single-cell biology, where predicting cellular responses to genetic perturbations is inherently unpaired, as each cell is destroyed during measurement. Existing approaches address this challenge using heuristic assumptions, such as restricting the input space to highly variable genes or incorporating external biological priors. However, these methods frequently optimize for mean point-estimates, resulting in catastrophic output variance collapse and a failure to capture biological heterogeneity. We introduce REGINA, a fully data-driven framework that formulates perturbation prediction as an unpaired distribution matching problem in a learned latent space. By combining structured latent representation learning with adversarial, cycle-consistent mappings between control and perturbed populations, REGINA captures non-linear, population-level transformations without requiring paired observations. A prompt-based latent conditioning mechanism further enables robust compositional generalization to novel perturbation combinations. Empirically, REGINA maintains competitive local prediction accuracy on key differentially expressed genes while generating biologically realistic single-cell distributions (evaluated via MMD and Wasserstein distance).

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Representation Learning, Generative modelling, Single-cell biology, Latent Cycle-Gan

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 2 reviews, above 27% of the others.
Reviewer_7WKROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

REGINA predicts perturbation responses from unpaired control and perturbed single-cell populations. It combines a Transformer autoencoder, regularization based on cell-state labels, and a latent CycleGAN conditioned on a prompt derived by replacing the target gene’s expression with −1. The paper evaluates local prediction and distributional metrics on the Adamson, Replogle, and Norman datasets, and reports competitive differential-expression accuracy and improved distribution matching on selected benchmarks.

Critical strengths

1 The paper tackles an important setting in which matched cell-level outcomes are unavailable, and evaluates distributional properties alongside differential-expression accuracy.

2 Evaluation spans datasets and perturbation types, with baseline comparisons, ablations, a split-stability analysis, and a reported code release.

Critical weaknesses

1 The paper motivates REGINA as counterfactual prediction for individual cells, but unpaired distribution matching and cycle consistency do not identify a unique cell-to-cell mapping. The reported metrics assess aggregate distributions and gene-level effects, not whether a particular control cell’s predicted response is correct. The claims should be framed more explicitly around population-level matching unless additional evidence supports cell-level correspondence.

2 The distributional results are mixed across datasets and metrics. On Adamson, REGINA is worse than PerturbNet on MMD (1.53 vs. 0.56), Wasserstein distance (0.255 vs. 0.210), and variance correlation (0.43 vs. 0.89; Table 2). On Norman, it has the best MMD, but CPA is better on Wasserstein distance, variance correlation, and energy loss (Table 4). The small MMD gap on Norman is also modest relative to the reported variation. These results do not support a general improvement in distributional fidelity.

Questions for authors

Is the intended target a population-level distribution, or a valid counterfactual for each individual cell? Please clarify which claims rely on cell-level correspondence.

For Table 6’s “0/2 Seen” cases, what exactly is held out, and how many combinations are evaluated? Are both constituent single-gene perturbations absent from training?

Are the comparisons in Tables 2 and 4 paired on the same held-out perturbations? Please clarify how the reported variability affects the claimed advantages.

2 comments

We thank the reviewer for the comment and agree that the original phrasing caused confusion. The metrics used for evaluation indeed calculate the population-level distribution similarity between the target and predicted cell populations and the individual effects were evaluated during the training phase. We have clarified this in the revised manuscript.

For Table 6, we evaluated our model's performance on perturbation combinations. The 0/2 seen means that neither perturbed gene was seen during the training (in neither phase). We used 139 conditions for training, where condition means any individual gene perturbation or combined perturbation. For validation we used 31 conditions and for testing we used 107 conditions.

Regarding Tables 2 and 4, the held-out perturbations were exactly the same across all models for a given dataset (Adamson, Replogle, and Norman). The table grouping is based on precision versus distributional metrics, and single-cell versus multi-gene perturbation datasets.

To address the variability in the distributional results: our core claim is not that our model is strictly superior across all distributional metrics. Rather, we emphasize that REGINA achieves high precision while maintaining competitive (and sometimes superior) distributional fidelity, whereas other models often excel in one aspect but fail significantly in the other.

I thank the authors for their response. I will keep my score.

Reviewer_CWK6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The aim of the study is to develop a framework that enables counterfactual inference and generalization to unseen and combined perturbation predictions. The paper presents REGINA, a framework for perturbation prediction using unpaired single cell sequencing data. The framework is based on regularized autoencoder architecture that includes encoder and latent space regularization followed by perturbation transition in latent space, and finally Cycle-GAN decoder to generate perturbed outcome. REGINA is evaluated on four datasets (Adamson, Replogle, Norman, Dixit) and compared to GEARS (Adamson, Replogle, Norman, Dixit), PerturbNet (only Adamson), and CPA (only Norman). REGINA results in best precision metrics in Adamson dataset, but has mixed outcome compared to GEARS in Replogle though the variance in these metrics are large and the difference is not statistically significant. PerturbNet seems to outcome better scores in distributional metrics than REGINA. In the Norman dataset REGINA is among the top-2 methods in all metrics. Based on the ablation study, the Latent Classifier and Latent Prompting components are essential in REGINA as the framework completely fails if either of those are removed. The authors recognize that REGINA is limited by dataset perturbation diversity that is common to all methods aiming for generalization without prior information.

Critical strengths
  • The paper is clearly written and easy to understand. Each component is presented separately and their inclusion in the framework are justified.
  • REGINA learns only from the training dataset and performs in similar accuracy as the compared method that incorporates prior biological knowledge.
Critical weaknesses
  • The paper is missing relevant benchmarks. At least one foundation model should be benchmarked (such as scGPT, scFoundation etc.) and linear baselines included (check Ahlmann-Eltze et al. (2025), Nat Methods, doi: 10.1038/s41592-025-02772-6). The authors limit foundation models out from comparisons and focus on task-specific models, but I would advise to compare foundation models as baselines to evaluate whether a task-specific model is able to improve outcomes of general models. Also a flow-based method could be compared as well (one example from the developer of scGen and CPA mentioned or compared in the paper, doi: 10.48550/arXiv.2601.11827).
  • The loss hyperparameter space should be explored. In the current implementation these are given without further information.
Questions for authors
  • Have you considered what other than a naive approach for tokenization could be evaluated?

Minor comments:

  • The paper could benefit from final polishing. For instance there are no references to Figures 1-3 and Tables 1-4 in the text. Appendix 7.2 paragraph is cut in the middle of the sentence. Tables present different numbers of significant digits.
  • Any thoughts why generalization when 1/2 genes are seen is worse than a completely novel combination (Table 6 lower part)?
2 comments

Thank you for the reviewer's comments on how to make the paper better. We have incorporated the suggested revisions to the manuscript and also evaluated a control baseline prediction on every dataset.

Regarding foundation models, existing literature already demonstrates that they are typically outperformed by task-specific models on these specific problems, which is why we focused our comparisons on specialized baselines.

For the loss hyperparameter space, our configuration uses the following weights: reconFactr (1.5), centerFactr (0.5), lossClassFactr (1.0), lossLatentFactr (1.0), lossCycleFactr (5.0), and adversarialFactr (3.5).

We had experiments to not use tokenization at all, but process the whole pipeline with a perceptron model, but that approach introduced computational overhead and the preliminary results were not promising either. We were considering Highly Variable Gene (HVG) ordering, but for thatit introduced non-trivial challenges in modeling gene relations.

We assume that the 1/2 seen genes case is a typical shortcut learning scenario, where the model rely more on the seen gene.

References

Atti, S., & Subramaniam, S. (2025). Fundamental Limitations of Foundation Models in Single-Cell Transcriptomics. bioRxiv.

Csendes, G., Sanz, G., Szalay, K. Z., & Szalai, B. (2025). Benchmarking foundation cell models for post-perturbation RNA-seq prediction. BMC Genomics, 26(1), 393.

I thank the authors for their response. Unfortunately, I cannot see that any changes were made in the paper.

Comments on the paper

2 comments

We appreciate the reviewer's comments. We may not have emphasized this clearly in the manuscript, but our goal was not to outperform other models in every single metric. Rather, it was to build a model that achieves strong precision while maintaining competitive distributional metrics, avoiding the severe trade-offs seen in some other models.

Yes, held-out cells were selected at the very beginning and were strictly excluded from preprocessing, auxiliary-label fitting, and both training phases. Cell labels were generated computationally based on the expression profiles of well-established cellular stress markers. As clarified in the revised manuscript, our claims evaluate population-level distributions, not individual counterfactual transitions.

In the 0/2 seen case, none of the cells were seen during training, meaning neither the individual components nor their combination were exposed. In the 1/2 seen case, only one of the two components was seen during training, while the other and the combination were held out. In the 2/2 seen case, both components were seen individually during training, but not as a combination.

We have added an extra evaluation for control-predictor baseline, but that didn't achieve better results than the best performing models. Since GEARS is always predicting a mean expression profile we couldn't compute variance correlation.

The reported standard deviations summarize the variability of metrics across the test dataset.

For Pearson we measured the correlation between the predicted and the targeted expression profiles, more precisely what is their difference from the mean of the control expression profile.

For NMSE (Normalized Mean Squared Error), we calculated the mean squared error between the predicted and true perturbation delta vectors (the difference from the control mean). To provide a scale-invariant measure, this error was normalized by the mean squared magnitude of the true perturbation delta.

For MMD (Maximum Mean Discrepancy), we quantified the overall distributional discrepancy between the predicted and true populations. We calculated the pairwise squared Euclidean distances between all true and predicted cells in a batch. These distances were then passed through a Gaussian kernel to capture complex statistical similarities. The final MMD score simply represents the difference between the average similarities within the same populations (true-to-true and predicted-to-predicted) and the cross-similarities between the true and predicted populations.

For the Wasserstein distance, we measured the overall geometric distance required to reshape the predicted cellular distribution so that it matches the true distribution. To manage the high-dimensional gene expression data efficiently, we utilized the Sliced Wasserstein approach, which mathematically projects both cell populations onto numerous random 1D axes and averages the distance across these simplified 1D projections using SciPy's wasserstein_distance implementation.

The top5% vs top20 DEG discrepancy was a labeling artifact and we corrected it.

For variance correlation we calculated the standard deviation for each gene across both the predicted and true cell populations, and then computed the Pearson correlation between them.

For implementation details see: https://anonymous.4open.science/r/single_cell-4F06/eval_metrics.py

During the training, we always take a batch of perturbed cells and a batch of control cells. For forward pass we only need the control cells and the corresponding perturbation label, which we want match. For the backward pass we always take the batch of control cells, that were obtained from that specific training step and calculate the latent prompt based on them. We thank the reviewer for pointing out the typo, it is correct, the transition applied in equation (6) notes transition applied twice, and we corrected it.

Thank you for the revision. Some of my concerns have been addressed. However, one important concern that the incremental benefit over the existing methods is still insufficiently established. I retain my rating of 2 (weak rejection).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.