acceptodds
Under review as a conference paper at ICLR 2027

PerturbSB: Learning Single-Cell Perturbation Dynamics with Control-Informed Schrödinger Bridge

Abstract

Single-cell perturbation data are central to understanding cellular mechanisms and applications such as drug discovery. Yet perturbation effects vary across cellular states, making it difficult to predict responses in states without treated observations. We propose PerturbSB, a Schr\"odinger-bridge-inspired framework for transferring perturbation responses across cellular contexts. Its stochastic dynamics separate a learned intrinsic drift, shared by control and treated cells, from a state-dependent perturbation drift. Target control cells provide the initial distribution, while training matches predicted treated endpoints to observed treated populations in source contexts. Across three single-cell benchmarks, PerturbSB delivers competitive response predictions and outperforms established baselines on several evaluation measures. These results suggest that separating intrinsic and perturbation dynamics can help predict responses in unseen cellular states.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Schrödinger Bridge, Single-cells, Perturbation

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_Auc9OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors introduce PerturbSB where the idea is to add to a shared reference drift a perturbation correction. It performs well on held-out Kang and Haber cell types, with mixed results on Hagai.

Critical strengths

The paper is honest in reporting cases where the method does not perform the best, including cases where simpler baselines are better.

Critical weaknesses

The ablation ablates on endpoint matching and energy but it does not remove the shared drift or diffusion. Also the energy barely changes the response correlation (cf Section 4.4), so the experiments do not show that the proposed stochastic design drives the gains, and this is also admitted in the same section. On the Hagai mouse experiment, a simple global shift has better results. Could the authors provide some intuition on why their method does not perform well in this setup? From a method's perspective, the novelty is also quite low.

Questions for authors

Could you run deterministic / no-reference-drift / fixed-diffusion baselines with the same configurations ( e.g. encoder, endpoint loss)?

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

The paper addresses the transfer of perturbation responses to cellular contexts for which only control cells are available. It models the response as a stochastic differential equation in a PCA latent space, with a reference drift shared by control and treated cells, a perturbation correction that is zero for controls, and a shared learned diffusion. Target control cells are evolved under the perturbed branch, and training matches the predicted endpoint to treated cells from source contexts using minibatch Wasserstein matching, regularized by a Girsanov path-KL energy on the correction. On the Kang, Haber, and Hagai benchmarks, the method is compared with identity, a global mean shift, scGen, and CellOT, and reports the best response correlation and mean-response error on Kang and Haber, with mixed results on the cross-species Hagai task.

Critical strengths
  1. Modeling perturbation effects is an important biological problem.
Critical weaknesses
  1. Key implementation details remain unverified, limiting reproducibility. Appendix C describes the listed architecture and optimization settings as "planned settings, not verified resolved configurations of the completed benchmark", and states that the batch sampling, the inclusion of target controls in source batches, the energy normalization, and the solver-gradient implementation "must be checked against code". Under these conditions, the reliability of the reported numbers cannot be assessed.

  2. The reference–perturbation decomposition lacks independent supervision. The objective matches only treated endpoints, without a separate control-branch endpoint loss. Consequently, the reference drift is not independently anchored to observed control dynamics, and its separation from the perturbation correction is not sufficiently identified by the data. Setting the correction to zero for controls does not establish that the reference drift captures unperturbed dynamics. Additional constraints or control-branch validation are needed to support the interpretation and practical value of this decomposition.

Questions for authors
  1. Can the authors provide the exact configurations and code used to reproduce the results in this paper?
Reviewer_eBe5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the transfer of a known perturbation response to a held-out cellular context for which only control cells are available. PerturbSB represents cells in a PCA latent space and models the control-to-treated transition with a stochastic differential equation. Its drift is decomposed into a learned reference drift shared by control and perturbed branches and a state- and time-dependent perturbation correction that is gated to zero for controls. The two branches also share a learned diagonal diffusion. Training matches generated treated endpoints to observed treated populations in source contexts using minibatch squared-Wasserstein assignment, while a diffusion-weighted path-energy term regularizes the perturbation correction relative to the learned reference process.

The method is evaluated on held-out cell types in Kang and Haber and held-out species in Hagai, against identity, a global mean shift, scGen, and CellOT. PerturbSB performs well on mean-response metrics in Kang and Haber and on the rat target in Hagai. However, its advantage is not uniform: CellOT or the global shift performs better on several distributional metrics, the mouse result is comparatively weak, and the predicted mouse population exhibits substantial variance contraction. The paper also studies full-dimensional versus sliced Wasserstein matching and the inclusion of the path-energy term.

Critical strengths
  1. A coherent control-informed formulation for an important transfer setting. The paper studies a practically relevant problem in which a known perturbation must be transferred to a cellular context with control cells but no treated observations. PerturbSB introduces a clean inductive bias by sharing a reference drift and diffusion across control and perturbed branches while using a state- and time-dependent perturbation correction that is exactly zero for controls. The diffusion-weighted relative path energy provides a mathematically motivated regularizer for this correction. Although the decomposition is not biologically identified, it is a coherent way to exploit target controls and model state-dependent responses from unpaired populations.

  2. Strong diagnostic empirical evaluation. PerturbSB obtains the best reported mean-response correlation and errors on both Kang and Haber, with particularly strong results on Kang and correct induction directions for all ten predefined interferon-response genes in the CD4 T-cell diagnostic. The evaluation also goes beyond favorable aggregate metrics by including simple baselines and explicit reporting of failures such as variance contraction on the Hagai mouse target. This combination of competitive results and transparent failure analysis makes the empirical evidence more informative than a benchmark table alone.

Critical weaknesses
  1. The central reference/perturbation decomposition is neither identified nor empirically isolated. Training supervises only the perturbed endpoint, without a separate control-endpoint loss, so the prediction depends on the joint effect of the reference drift, perturbation correction, and learned diffusion. The reference drift may absorb perturbation-specific motion, while the diffusion rescales the path-energy penalty. The current ablation varies endpoint geometry and energy weight but does not compare against an endpoint-only predictor, deterministic dynamics, or fixed/removed reference drift and diffusion. Since the energy term has only small, metric-dependent effects, the experiments do not establish that the proposed stochastic decomposition is responsible for the gains.

  2. The comparative, generalization, and reproducibility evidence is insufficient to establish an advantage over current alternatives. The baselines omit both the closest SB-based methods, such as Departures and direction-constrained diffusion SBs, and recent population-level predictors such as STATE or PerturbDiff. Thus, it remains unclear whether the improvements arise from PerturbSB’s shared-reference design or from generic endpoint distribution modeling.

Questions for authors
  1. Could the authors provide ablations with an endpoint-only predictor, fixed/removed reference drift, and fixed/removed diffusion? The current ablation does not establish whether the proposed stochastic decomposition causes the gains.

  2. Could the authors provide controlled comparisons with (i) at least one recent Schrödinger-bridge-based perturbation method and (ii) one modern population-level model, such as STATE or PerturbDiff, under the same data and split protocol?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.