Learning Latent Dynamical Causal Processes for Single-Cell Perturbation Prediction

2026-05-25

AI Summary

This work addresses the challenge in single-cell perturbation prediction of simultaneously modeling the implicit mechanisms of perturbations and their temporal dynamics, which limits generalization to unseen interventions. The authors propose CITE-VAE, an implicit dynamical causal generative model that formalizes perturbation effects as causal mechanisms exhibiting both latent structure and temporal evolution. CITE-VAE jointly models latent cellular programs, perturbation conditions, and time-dependent dynamics, and provides theoretical identifiability guarantees for latent variables within standard equivalence classes. Efficient inference is achieved through a variational autoencoder framework. Experiments demonstrate that the model validates its identifiability on the Causal-3DIdent synthetic benchmark and significantly outperforms existing methods on real CRISPR-based single-cell perturbation data, achieving superior out-of-distribution generalization performance.

Abstract

Single-cell perturbation prediction aims to infer how cells respond to unseen interventions and to achieve out-of-distribution (OOD) generalization, providing a computational route to understanding how perturbations reshape cellular programs over time. Existing machine learning methods have made important progress, but typically capture only one side of the response. Latent causal approaches seek mechanisms that support generalization and interpretation, yet often treat perturbation effects as static outcomes. Temporal models describe how gene expression changes across time, but usually do not explicitly recover the latent causal generative mechanisms driving these changes. In practice, perturbation effects are both latent and dynamical: interventions act through unobserved cellular programs, whose states evolve over time and give rise to observed expression profiles. Motivated by this view, we propose a latent dynamical causal generative model for single-cell perturbation data that jointly captures latent cellular programs, perturbation-conditioned mechanisms, and temporal evolution. We further provide an identifiability analysis showing that, under suitable conditions, the latent causal variables are recoverable up to standard equivalence classes. Guided by this analysis, we develop CITE-VAE, a learning framework for recovering latent cellular programs and their perturbation-driven dynamics from single-cell sequencing data. Experiments on Causal-3DIdent validate the theoretical results and the effectiveness of the proposed method in controlled settings. Additional experiments on real-world CRISPR-based single-cell perturbation data show improved generalization to unseen perturbations compared with state-of-the-art baselines, highlighting the practical robustness of our approach.

Problem

• single-cell perturbation prediction
• latent causal processes
• dynamical systems
• out-of-distribution generalization
• temporal gene expression

Innovation

• latent causal processes
• dynamical systems
• single-cell perturbation prediction
• out-of-distribution generalization
• identifiability