Departures: Distributional Transport for Single-Cell Perturbation Prediction with Neural Schrödinger Bridges
Changxi Chi, Yufei Huang, Jun Xia, Jiangbin Zheng, Yunfan Liu, Zelin Zang, Stan Z. Li · 2025
Read Departures: Distributional Transport for Single-Cell Perturbation Prediction with Neural Schrödinger Bridges by Changxi Chi, Yufei Huang, Jun Xia, Jiangbin Zheng, Yunfan Liu, Zelin Zang, Stan Z. Li (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-13124.txt.
Introduction
Changxi Chi and six co-authors predict how a single cell's gene expression will shift under a drug or genetic perturbation, without ever seeing the same cell twice. Sequencing destroys the cell, so data are inherently unpaired: control cells and perturbed cells are different cells. The authors model the trip from one population to the other with a Schrödinger bridge. The entropy-regularized optimal transport process that finds the most plausible stochastic path between two distributions, conditioned on each perturbation.
Earlier bridge approximations inferred the coupling between source and target by modeling both directions, which is ill-posed. This work pairs cells with minibatch optimal transport instead, directly guiding bridge learning, and trains two bridges jointly, one for discrete gene activation states and one for continuous expression levels. The pairing makes the approximation scalable and more precise.
Catalog record
| CREATOR | Changxi Chi, Yufei Huang, Jun Xia, Jiangbin Zheng, Yunfan Liu, Zelin Zang, Stan Z. Li |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-13124
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Departures: Distributional Transport for Single-Cell Perturbation Prediction with Neural Schrödinger Bridges Changxi Chi, Yufei Huang, Jun Xia, Jiangbin Zheng, Yunfan Liu, Zelin Zang, Stan Z. Li Abstract Predicting single-cell perturbation outcomes directly advances gene function analysis and facilitates drug candidate selection, making it a key driver of both basic and translational biomedical research. However, a major bottleneck in this task is the unpaired nature of single-cell data, as the same cell cannot be observed both before and after perturbation due to the destructive nature of sequencing. Although some neural generative transport models attempt to tackle unpaired single-cell perturbation data, they either lack explicit conditioning or depend on prior spaces for indirect distribution alignment, limiting precise perturbation modeling. In this work, we approximate Schrödinger Bridge (SB), which defines stochastic dynamic mappings recovering the entropy-regularized optimal transport (OT), to directly align the distributions of control and perturbed single-cell populations across different perturbation conditions. Unlike prior SB approximations that rely on bidirectional modeling to infer optimal source-target sample coupling, we leverage Minibatch-OT based pairing to avoid such bidirectional inference and the associated ill-posedness of defining the reverse process. This pairing directly guides bridge learning, yielding a scalable approximation to the SB. We approximate two SB models, one modeling discrete gene activation states and the other continuous expression distributions. Joint training enables accurate perturbation modeling and captures single-cell heterogeneity. Experiments on public genetic and drug perturbation datasets show that our model effectively captures heterogeneous single-cell responses and achieves state-of-the-art performance. Code — https://github.com/ChangxiChi/Departures Introduction Genetic perturbation and drug perturbation constitute the two primary avenues for studying and modeling single-cell responses to perturbations. Genetic perturbations, predominantly based on CRISPR technologies ( Barrangou and Doudna 2016 ; Lino et al. 2018 ) , target specific genes to reveal their functional roles, enabling systematic dissection of gene regulatory mechanisms at single-cell resolution. Drug perturbations, by contrast, modulate molecular pathways through small-molecule compounds ( Peidli et al. 2024 ) , offering complementary insights into cellular responses and pharmacological effects. Together, these two perturbation strategies form the foundation of single-cell perturbation studies, providing mechanistic and translational perspectives on cell transitions under diverse experimental conditions. However, due to the high cost and limited scalability of single-cell experiments, it is impractical to measure all combinations of perturbations and cell types. This limitation highlights the need for computational models that can predict cellular responses under unseen perturbation conditions. Figure 1: Single-cell perturbation data are unpaired as RNA-seq is destructive. A major challenge in this task is that single-cell perturbation data are inherently unpaired ( Bunne et al. 2023 ; Chi et al. 2025a ) . This arises from the destructive nature of RNA sequencing, which requires cell lysis to release RNA content ( Mortazavi et al. 2008 ) , as illustrated in Fig. 1 . As a result, it is impossible to observe the same cell under both control and perturbed conditions. Some studies ( Bunne et al. 2023 ; Cao et al. 2024 ) attempt to address this unpaired nature, but their use of unconditional models limits generalization to unseen perturbation conditions. In contrast, ( Chi et al. 2025a ) incorporates gene regulatory network (GRN) priors to model the effects of perturbations, providing a biologically grounded representation of cellular responses. It then aligns control and perturbed cell populations using dual conditional diffusion models with a shared prior space, enabling accurate prediction under unseen perturbation conditions. However, by aligning distributions indirectly through the prior space, it bypasses the underlying physical energy landscape, leading to transition paths that deviate from energy-optimal trajectories and fail to capture ideal optimal transport (OT). In this paper, we purpose Departures ( D istributional Transport for Single-C e ll P erturb a tion P r edic t ion with Ne u ral Sch r öding e r Bridge s ), a generative framework designed to predict cellular gene expression under diverse genetic and molecular perturbation conditions. By directly aligning the distributions of control and perturbed cells, the framework captures population-level transitions without relying on potentially noisy explicit cell pairing or restrictive latent space alignment. This method reduces information loss, improves robustness to biological variability and noise, and provides a principled, scalable solution to the challenges posed by unpaired single-cell perturbation data. While classical Schrödinger Bridge methods typically rely on alternating forward and backward updates ( Rüschendorf 1995 ; Shi et al. 2023 ) to iteratively refine the source-target node pairing between marginal distributions, the backward process is often ill-defined in conditional settings (Fig. 3 ) ( Tamir, Trapp, and Solin 2023 ; Tang et al. 2024 ) . Moreover, training two coupled models significantly increases computational cost and reduces scalability. To address this, we build upon Minibatch OT ( Fatras et al. 2021 ; Tong et al. 2023 ) to directly enhance source-target sample coupling during training, without requiring alternating bidirectional model updates for pairing consistency. Built upon this alignment, we adopt bridge matching ( Liu et al. 2022 ; Peluchetti 2023 ) to learn the Markovian projection ( Gyöngy 1986 ) , yielding a tractable approximation to the SB between control and perturbed distributions. To faithfully capture single-cell perturbation outcomes, we learn two SB models, the first models discrete gene activation states, and the second learns continuous expression dynamics. Joint training of these bridges enhances both biological fidelity and generative robustness. The main contributions of our work are as follows: • We introduce Departures , a Schrödinger Bridge-based generative model that directly aligns the distributions of control and perturbed cells. To better approximate the Schrödinger Bridge, we use minibatch OT to compute source-target sample pairings that guide bridge matching training, eliminating the need for bidirectional iterative updates. • To model single-cell responses to perturbations, we design two Schrödinger bridge models, one for discrete gene activation and one for continuous expression, which are trained jointly to improve fidelity and robustness. • We demonstrate the superiority of Departures over existing methods on publicly available genetic and molecular perturbation datasets. Related Works and Preliminaries Existing Perturbation Prediction Model There are numerous methods designed to predict single-cell perturbation responses, encompassing both genetic and drug-induced perturbations. These methods typically employ either generative models ( Lotfollahi, Wolf, and Theis 2019 ; Hetzel et al. 2022 ; Wu et al. 2022 ; Bereket and Karaletsos 2024 ; Cui et al. 2024 ) or regression-based approaches ( Roohani, Huang, and Leskovec 2024 ; He et al. 2024 ; Chi et al. 2025b ) . However, many of these models overlook the inherently unpaired nature of single-cell perturbation data and fail to explicitly model the relationship between control and perturbed samples. While a few studies have addressed the unpaired nature of the data ( Bunne et al. 2023 ; Cao et al. 2024 ) , their reliance on unconditional models limits their ability to generalize to unseen perturbation settings. ( Chi et al. 2025a ) addresses this by incorporating gene regulatory network (GRN) priors to model perturbation effects, and aligning control and perturbed cell populations via dual conditional diffusion models that share a common prior space. Additionally, a regression-based mask model is introduced to predict gene expression status, thereby improving generation quality. However, this implicit alignment bypasses the true physical energy landscape, resulting in transitions that