Disentangled generative modeling of single-cell perturbation responses across dynamic cellular states

Abstract

Single-cell perturbation studies are increasingly conducted across complex, multi-condition biolog-ical systems, offering a powerful lens into how interventions reshape cell states and molecular pro-grams. However, standard computational models compress single-cell profiles into unified repre-sentations where treatment effects are obscured by cell division, lineage progression, donor varia-tion, and cellular stress. Here, we present Disentangled Variational Inference for Perturbation Re-sponse (PertVI), a deep generative model that separates baseline mitotic phase, cellular state, and batch effects from treatment-induced shifts while isolating aberrant off-manifold responses in or-thogonal anomaly space. This structured design tackles two intertwined tasks: (1) quantifying treatment-induced changes in cell-state composition, and (2) inferring state-gated transcriptional changes in response to treatment. Across simulated benchmarks, PertVI accurately disentangles coupled biological axes and recovers weak treatment signatures. In conditional chemical and genet-ic screens of primary human CD8+ and CD4+ T cells, PertVI separates memory-precursor adju-vants from toxic phenocopies, resolves co-clustered regulators into directional kinetic trajectories and biphasic rebounders obscured by standard workflows, and reveals that treatment-induced phe-notypic shifting is decoupled from broad transcriptional perturbation responses and instead gov-erned by state-specific impulses. Altogether, PertVI brings unique mechanistic granularity to sin-gle-cell perturbation response modeling, serving as a proof of principle for biology-informed, struc-ture-guided variance disentanglement in dynamic cellular systems.