DynPerturb:Dynamic Perturbation Modeling for Spatiotemporal Single-Cell Systems

Abstract

Perturbation responses in multicellular systems are inherently spatiotemporal, governed by temporal dynamics, perturbation intensity, and tissue context. However, existing approaches rarely capture explicit spatiotemporal perturbation dynamics, limiting the ability to reconstruct how perturbations propagate over time, spread across space, and induce feedback at multiple biological scales. Here we introduce DynPerturb, a dynamic inference framework for single-cell and spatial transcriptomics that systematically models the "when-how strong-where" dimensions of perturbations. DynPerturb incorporates temporal encoding, spatial adjacency, and memory modules to capture nonlinear signal propagation and cross-regional feedback with mechanistic interpretability. Across temporal benchmarking datasets, DynPerturb consistently outperformed existing methods in regulatory link prediction, demonstrated zero-shot generalization to unseen nodes, and preserved accuracy under varying sparsity conditions and temporal resolutions. Applied to representative biological systems, DynPerturb uncovers dynamic regulatory rewiring in kidney disease, highlighting a mid-age sensitivity window and potential reversibility with early intervention; delineated a temporal fate-switching boundary between megakaryocyte-erythroid progenitors (MEPs) and granulocyte-monocyte progenitors (GMPs) during hematopoiesis; and revealed spatiotemporal interchamber signaling in murine heart development, where Igf2 perturbations induced spatial feedback shaping chamber proportioning. Collectively, these results establish DynPerturb as a unified and versatile system for dissecting dynamic perturbation effects and guiding precision intervention strategies in multicellular systems.