acceptodds
Under review as a conference paper at ICLR 2027

scBridge: Stochastic Distribution Transport for Single-Cell Drug Response Prediction

Abstract

Single-cell drug response prediction is limited by the scarcity of labeled scRNA-seq data, motivating knowledge transfer from large-scale bulk pharmacogenomic datasets. However, substantial distribution shifts between bulk and single-cell transcriptomes make such transfer challenging. Existing approaches are largely representation-centric, aligning the two domains in a shared latent space without explicitly modeling the transition between their distributions. To address this limitation, we formulate bulk-to-single-cell adaptation as an unpaired stochastic distribution transport problem and propose scBridge, a diffusion-based framework for single-cell drug response prediction. scBridge first learns compact latent representations and performs coarse covariance alignment to provide a stable cross-domain initialization. Drawing on Diffusion Schrödinger Bridge, scBridge then models a continuous stochastic transition between bulk and single-cell latent distributions through coupled forward and backward dynamics, optimized alternately via iterative proportional fitting with cycle-consistency regularization. Confidence-aware pseudo-label refinement further adapts the transported representations for downstream prediction. Across 12 scRNA-seq datasets, scBridge achieves a mean AUROC of 0.87, with a 9.33% relative improvement over the second-best model. The learned transport reduces domain separability by 56.8%–77.9% and reveals heterogeneous cell-level transport trajectories with distinct temporal correction dynamics across datasets. Model attribution further identifies biologically relevant mechanisms associated with BRAF inhibitor resistance. These results support stochastic distribution transport as an effective perspective for single-cell drug response prediction across modalities.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell drug response prediction, Stochastic distribution transport, Diffusion Schrödinger Bridge, Cross-modal transfer learning, Domain adaptation

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_G94tOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

scBridge transfers drug response supervision from bulk transcriptomics to the unlabeled single cell level via separate beta-VAEs, covariance alignment, bidirectional stochastic transport, cycle consistency and pseudo-label refinement. Across 12 datasets the average AUROC/AUPR/F1 are 0.867/0.878/0.770 compared to scDEAL's 0.793/0.802/0.758.

Critical strengths
  • Major problem and promising results. Transferring bulk supervision is a solution to the problem of the lack of single cell response labels. The reported improvements in AUROC and AUPR across drugs and cancer types are very encouraging.

  • Useful supporting analyses (component ablations, dataset level results, subsampling and noise experiments, biological attribution).

Critical weaknesses
  • The predictor has unresolved inconsistencies, and preservation of response correspondence is not established. Equation 7 defines the backward prediction as zt−Δt−ztΔt\frac{z_{t-\Delta t}-z_t}{\Delta t}, but Algorithm 1 subtracts it. Under this convention, the update should add it. The prediction head is trained with a fine-tuned bulk encoder, while transport uses a separate frozen encoder. Their compatibility is not explained.

  • Matching distributions doesn't guarantee that sensitive cells map to sensitive cells. You can have separate encoders with independently selected genes, covariance alignment and cycle consistency and still end up with mappings that swap class correspondence.

  • Noise free inference keeps the learned drift but does not generally preserve the distributions of the stochastic process. A matched deterministic baseline is required in order to isolate the benefit of stochastic training.

Questions for authors
  • How does Equation 7 satisfy Algorithm 1? Which encoder features are used to train the prediction head and how are they kept in line with the transport space?

  • Please provide the standard deviations across the five runs, the split identifiers and the exact model/baseline settings. Were the baselines evaluated on the same splits? How were the hyperparameters, checkpoints and classification thresholds selected in the context of the stated target-label restriction?

  • How are the bulk labels and drug specific thresholds defined? How are the single cell response labels actually constructed in the case of GSE241225? What is the positive class and what is the F1 averaging convention?

  • How is noise-free inference justified? Does the gain hold up against matched deterministic transport?

  • What actually tests the preservation of response correspondence beyond just distribution matching? A controlled correspondence experiment or one based on shared gene representation? How sensitive are the results to encoder initialisation?

Reviewer_7b9tOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes scBridge, a cross-modal single-cell drug response prediction framework based on the diffusion Schrödinger Bridge. By combining β-VAE, CORAL coarse alignment, bidirectional stochastic distribution transport, and confidence-aware pseudo-label optimization, this method transfers drug response knowledge from bulk RNA-seq to single-cell data. Across 12 scRNA-seq datasets, scBridge achieved strong AUROC, AUPR, and F1 scores.

Critical strengths
  1. The research question is of practical significance. This paper focuses on the distribution differences between bulk RNA-seq and scRNA-seq and achieves cross-modal knowledge transfer through random distribution transfer; the research motivation is well-founded.

  2. The experimental results are competitive. scBridge achieves an average AUROC of 0.867 and an AUPR of 0.878 across 12 datasets, significantly outperforming various existing domain adaptation methods.

  3. The experimental analysis is thorough. The authors provide ablation studies of key modules, visualizations of cross-domain distributions, and analyses of data sparsity and robustness to noise, along with biological interpretations based on BRAF inhibitor resistance.

Critical weaknesses

1. Random stratification at the cellular level may lead to an overestimation of generalization performance.

According to Appendix A.3.1, the authors performed stratified random stratification based on response labels within each scRNA-seq dataset, using 80% of the cells for training and the remaining 20% for testing. However, cells from the same patient, biological sample, or experimental batch may appear in both the training and testing sets. This splitting fails to adequately validate the model’s generalization ability to independent patients or experimental samples and may also allow the model to exploit sample-specific transcriptomic features. We recommend supplementing the evaluation with splits based on patients, independent samples, or experimental batches to further validate cross-sample generalization.

2. The theoretical implementation of the Diffusion Bridge and its innovation relative to existing methods require further clarification.

This paper adopts the existing Diffusion Schrödinger Bridge framework and achieves distribution transport by alternately updating the forward and backward networks. However, Eqs. (6)–(7) primarily use the finite-difference velocity of the reference trajectory for mean-squared error regression; they do not fully explain how this training process corresponds to the path distribution optimization objective of the Schrödinger Bridge, nor how it ensures endpoint distribution constraints. Furthermore, deterministic Cycle Consistency does not directly guarantee the preservation of the cell’s biological semantics or local topological structure. We recommend providing additional theoretical explanations and conducting further comparisons with relevant Schrödinger Bridge methods to clarify the independent methodological contributions of this paper.

3. Some performance improvements are modest and lack sufficient verification of statistical significance.

Although scBridge shows significant improvements in average AUROC and AUPR, its average F1 score is 0.770, compared to 0.758 for the second-best method, indicating a relatively limited improvement. Appendix Tables A.2–A.4 also show that scBridge still lags significantly behind existing methods on some datasets, such as GSE149383. Although the authors state in Appendix A.3.1 that they conducted five independent runs, the main experimental tables do not provide corresponding standard deviations or confidence intervals, making it difficult to assess the statistical stability of the performance advantages. We recommend supplementing the paper with complete results from multiple random seed runs and paired significance analyses, and further discussing variations in model performance across different datasets.

4. The open-source code link is empty.

The authors claim that the source code and data are publicly available at the provided URL, but the repository contains only a README file and no code, data, or environment specifications. It is recommended that the authors share the code along with comprehensive data and specifications, as well as the saved models, to facilitate replication and wider adoption.

Questions for authors

See in Weeknesses. If the author addresses the issue I raised, I am willing to update my rating.

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

scBridge transfers drug-response supervision from labeled bulk transcriptomic data to unlabeled single-cell data. It combines separate β-VAE representations, CORAL alignment, bidirectional stochastic bridge transport, and confidence-filtered pseudo-label refinement. Experiments on 12 single-cell datasets report mean AUROC, AUPR, and F1 scores of 0.867, 0.878, and 0.770, respectively. The paper also includes component ablations, robustness tests, transport-path analyses, and a biological case study.

Critical strengths
  1. The paper studies an important transfer problem. Single-cell drug-response labels are scarce, while labeled bulk data are more readily available. A method that can use bulk supervision without paired bulk and single-cell samples would be useful.
  2. The empirical evaluation is broad for this task. It covers 12 single-cell datasets and reports gains over several bulk-to-single-cell transfer methods. The component ablations, robustness tests, and transport analyses provide useful supporting evidence.
Critical weaknesses
  1. The central stochastic bridge needs more implementation detail for reproduction. For example, the drift-network architecture and the coupling of unpaired bulk and single-cell latents are not clearly specified. The sign convention between Equation (7) and the backward updates in Algorithm 1 would also benefit from clarification.
  2. The experiments do not fully isolate the advantage of stochastic bridge transport. Removing the complete bridge module is useful, but it does not distinguish stochastic path modeling from a more general learned nonlinear mapping. A matched deterministic transport control would make the methodological contribution clearer.
Questions for authors
  1. Could the authors clarify the main implementation details of the drift networks, how the unpaired latent samples are coupled, and the sign convention used in the backward updates?
  2. Could the authors include a matched deterministic transport control, or clarify how the existing OT baseline differs from scBridge beyond the stochastic bridge formulation?
  3. Could the authors clarify the intended scope of generalization under the random cell-level split? In particular, does the evaluation only cover held-out cells within each adapted dataset, or also unseen biological samples or cohorts?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.