acceptodds
Under review as a conference paper at ICLR 2027

DeMixPert: Decomposed Response Modeling with Gaussian Mixtures for OOD Single-Cell Perturbation Prediction

Abstract

Accurate prediction of transcriptome-wide responses to unseen genetic perturbations requires modeling both perturbation-specific transcriptional shifts and heterogeneous cellular responses. Existing methods often entangle deterministic response structure with stochastic population-level variation, causing dominant shared patterns to mask weaker perturbation-specific signals and impair distributional modeling. To address these challenges, we propose DeMixPert, an approach for Decomposed response Modeling with Gaussian Mixtures for Out-Of-Distribution (OOD) single-cell Perturbation prediction. DeMixPert decomposes perturbation-induced changes into a basal-state-dependent systematic response, a perturbation-specific response, and population-level variation. The systematic component is derived from the basal state encoded from control-cell expression, whereas the perturbation-specific component is inferred from pretrained target embeddings for unseen-target generalization. DeMixPert models population-level variation using a Gaussian prototype Invertible Network and adaptively combines reusable Gaussian prototypes according to the basal state and perturbation condition. The resulting mixture is mapped to a condition-specific variation distribution. Sampled variations are integrated with the systematic and perturbation-specific components, followed by joint decoding with the basal state to reconstruct perturbed-cell gene expression. Experimental results show that DeMixPert effectively captures heterogeneous single-cell perturbation responses and achieves superior performance across unseen-perturbation settings. The source code will be made publicly available upon publication.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell perturbation prediction, Out-of-distribution generalization, Response decomposition, Gaussian mixture models, Invertible neural networks

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_vZHgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DeMixPert predicts single-cell perturbation responses by separating basal-state-dependent effects, perturbation-specific effects, and stochastic population variation. Target-gene embeddings support prediction for held-out perturbations. A conditional Gaussian mixture and an invertible network model the residual variation before the components are combined to reconstruct expression. The evaluation covers four genetic perturbation datasets, five seeds, and differential-expression, distributional, and perturbation-discrimination metrics. Single-gene conditions are held out; combinatorial tests use new combinations of targets seen individually (0/2 unseen). Ablations and mean/additive baselines support the predictive gains, although the biological interpretation of the components is less clear.

Critical strengths
  1. The prediction results are convincing on several metrics. Table 1 shows strong differential-expression performance and the lowest energy distance among the listed learned methods on all four datasets. Condition-disjoint splits, validation-based checkpoint selection, and repeated seeds strengthen these comparisons.

  2. The evaluation goes beyond mean-expression error. Distributional metrics, perturbation discrimination, retrained ablations, and the Training-mean/Additive baselines in Appendix N give a useful view of both gains and limitations. This breadth matters for a model intended to capture cellular heterogeneity.

Critical weaknesses
  1. My main concern is that the Gaussian-mixture contribution has not been isolated. The “w/o GM” ablation removes the mixture, invertible network, and stochastic residual together (PDF p. 18). It shows that the full branch helps, but not that mixtures are preferable to another expressive stochastic model. The K-sensitivity analysis starts at two components, leaving this comparison open.

  2. The distributional results do not yet make a clear case for improved population heterogeneity. The energy-distance gains are useful evidence, and evaluating a held-out training metric is reasonable. However, Wasserstein distance in Table 1 does not favor DeMixPert over scGPT. On Norman, the simple Additive predictor also has better Wasserstein distance and PDS (PDF p. 25). Since PDS measures centroid ranking, it does not establish recovery of multimodality or subpopulation proportions. The predictive gains remain credible, but the heterogeneity claim needs more support.

  3. The decomposition is easier to justify as a predictive model than as a biological explanation. Held-out representation separability in Appendix L.1 supports perturbation discrimination, but it does not uniquely identify the systematic and perturbation-specific terms. The prototype–residual analysis instead uses training conditions and model-derived residuals (Appendix L.2, PDF pp. 23–24). Independent annotations or constraints would be needed to interpret these components as distinct biological sources of variation.

Questions for authors
  1. Is there a matched comparison with a unimodal conditional flow, retaining comparable capacity, losses, and training budget? Simply setting K = 1 would also restrict the prototype-based shift function, so it may not isolate the role of mixtures.

  2. Do the existing held-out predictions show better recovery of covariance or independently defined subpopulation proportions? How do the Wasserstein and Additive results affect the intended claim about heterogeneity?

  3. What constraints or independent annotations distinguish the learned components from other decompositions with similar prediction error? Do the prototype–residual associations in Appendix L.2 also hold on unseen perturbations?

Reviewer_TdgYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces a novel method, named DeMixPert, that focuses on disentangling systematic, perturbation-specific, and population-level effects en route to perturbation prediction. The method leverages a separate encoder for each of these constituent components and incorporates a Gaussian mixture model-based prototype network to model population-level variance. The final training objective in this framework consists of four distinct loss components, covering perturbation response alignment, reconstruction, likelihood estimation from the Gaussian mixture model, and distributional alignment for the expression prediction. The authors benchmark DeMixPert against five single-cell foundation models across multiple Perturb-seq datasets, and they show that DeMixPert frequently outperforms the baseline models in predicting the perturbation responses of unseen perturbation conditions. The paper also features ablation analyses as well as qualitative visualization results that highlight the ability of their framework to separate out global/systematic axes of variation to disentangle more nuanced perturbation-specific effects.

Critical strengths
  • The paper aims to tackle the important challenge of decoupling global, perturbation-specific, and population-level effects in perturbation modeling. The problem is well-formulated, and the model architecture appears to align well with this goal, with separate encoders for each of these components.
  • The qualitative results in Figure 4 alongside its accompanying clustering metrics appear to effectively demonstrate the utility of DeMixPert in learning representations that can decouple global/systematic axes of variation from perturbation-specific effects.
Critical weaknesses
  • The paper’s explanations of both the methodology and the rationales behind modeling choices are often unclear, making it difficult to assess the merits of the proposed framework. For example, the paper mentions coupling layers and the use of binary masks to partition the feature dimensions, but it is not clear why the authors chose this design other than that this specific formulation involves layers that are bijective.
  • Related to the above point, the Figure 2 diagram of the model architecture does not appear to cohere with the description of the model in the main text. For example, the systematic response representation gi=Esys(si)g_i = E_{sys}(s_i) is stated in the main text to be dependent on the basal state representation si=Es(xj0)s_i = E_s(x_j^0), but the visualization represents them as independent entities.
  • The DeMixPert model does not appear to consistently outperform the other baseline methods in Table 1, although it does yield the best and second-best performance for many of the metrics.
  • A few of the non-standard evaluation metrics are not defined, including the Common-Differentially Expressed Genes (C-DEGs) and Differential Expression Score (DES).
  • The ablation analyses in Table 2 appear to only show a subset of the evaluation metrics from Table 1.
  • The Appendix sections in the paper are not numbered in the order they appear in the main text.
  • The paper’s Results section is primarily focused on re-iterating metrics that appear in the Results tables and there is limited discussion or interpretation of the implications of the findings and/or their implications.
Questions for authors
  • What is the rationale behind the design of the coupling layers and the binary masks? Can you please provide more intuition for these modeling choices?
  • What are the implications of the findings presented in Table 1 as well as the ablation study results (Table 2)? Does DeMixPert possess properties that enable it to perform particularly well on some tasks/metrics, and which components of the model architecture most contribute to this performance?
  • Can you please help clarify the significance of the analysis performed in Figure 5? Why was the analysis performed on training conditions rather than held out test conditions, and is this result aimed at highlighting the per-perturbation condition relationship between the prototype weights and the corresponding observed expression residual profiles?
Reviewer_A2dkOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper predicts single-cell responses to unseen genetic perturbations. It splits the response into a basal term, a perturbation-specific term, and a residual variation term, with a Gaussian mixture plus invertible network for the residual. It evaluates on four perturbation datasets and beats scGPT on DEG overlap, energy distance, and perturbation discrimination, but not on Wasserstein distance.

Critical strengths
  1. The decomposition is well motivated and easy to inspect.

  2. The evaluation is broad, the splits are clean, and the method shows real gains on perturbation discrimination and DEG overlap.

Critical weaknesses
  1. The reported E-Dist is not independent of training: the loss L_dist (Eq. 11) is the same energy-distance objective, so the E-Dist win is expected. W-Dist, the one distribution metric not optimized for, is worse than scGPT on all four datasets. This matters because the paper’s central claim is distributional, and the independent evidence points the other way.

  2. The baseline comparison is thin. MixFlow is the closest prior: a descriptor-conditioned Gaussian mixture transported by flow matching. It is evaluated on the same datasets, but it is never run here; CPA and PerturbNet are also absent. A plain additive baseline also does as well or better on several distribution and centroid metrics. I cannot tell whether the method actually improves over the nearest prior work.

Rubbi, Andrea, et al. “Shortest-Path Flow Matching with Mixture-Conditioned Bases for OOD Generalization to Unseen Conditions.” arXiv preprint arXiv:2601.11827 (2026).

Questions for authors

W-Dist is worse than scGPT on all four datasets, yet the paper claims distributional fidelity. How do you reconcile this, and why is W-Dist not reported in the results section?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.