Search papers, labs, and topics across Lattice.
The paper introduces CITE-VAE, a latent dynamical causal generative model, to predict single-cell perturbation responses by jointly modeling latent cellular programs, perturbation-conditioned mechanisms, and temporal evolution. This approach addresses limitations in existing methods that either treat perturbation effects as static or fail to explicitly recover latent causal generative mechanisms. CITE-VAE demonstrates improved generalization to unseen perturbations on real-world CRISPR-based single-cell data compared to state-of-the-art baselines, and its theoretical identifiability is validated on synthetic data.
Unseen single-cell perturbation effects can be predicted more accurately by explicitly modeling the latent, dynamic causal processes that drive cellular response.
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.
