CFM-GP: unified conditional flow matching to learn gene perturbation across cell types.

Understanding how gene perturbations reshape cellular states across diverse contexts is fundamental to functional genomics and therapeutic discovery, yet experimental profiling across all perturbations and cell types remains infeasible. Computational approaches promise scalable inference but often rely on discrete mappings or per-cell-type models that fail to capture continuous and shared biological dynamics. We introduce CFM-GP, a conditional flow-matching framework that learns a continuous vector field transforming control expression profiles into perturbed states, explicitly conditioned on cell type. This unified design models both common regulatory programs and type-specific responses within a single architecture, removing the need to train separate models. Across five single-cell perturbation datasets, CFM-GP consistently outperformed existing methods in predictive accuracy, distributional alignment, and cross-species generalization. The inferred flow trajectories recovered canonical signaling pathways and context-dependent transcriptional cascades, demonstrating mechanistic interpretability. By coupling principled generative dynamics with biological conditioning, CFM-GP offers a scalable foundation for modeling cellular perturbation responses, enabling data-driven exploration of gene function and intervention strategies across heterogeneous cellular systems.