Dynamic Graph Prior-Guided Latent Adaptation for Generalizable Single-Cell Perturbation Response Prediction

Predicting transcriptional responses to drug perturbations in unseen cellular contexts is a central challenge in single-cell perturbation modeling. This challenge creates a need for models that combine flexible state transitions with structured biological prior knowledge. Existing data-driven models often struggle to generalize across unseen cell types, while many prior-based models rely on static biological graphs without explicitly conditioning prior influence on perturbation-specific target support.

We propose DGP-LA, a dynamic graph-prior latent adaptation framework that refines biological priors and injects them as gated latent corrections within a cell-state-conditioned transition model. DGP-LA couples a transition backbone with perturbation-conditioned sparse graph refinement. It propagates gene-aligned target support over the refined graph and selectively injects the resulting topology-aware signal into the hidden representation. On Tahoe and NeurIPS, DGP-LA improves Pearson Delta relative to State by 8.7% and 35.8% for generalization across unseen cell types. In zero-shot drug perturbation response prediction on unseen cell types, it improves Pearson Delta relative to CRISP by 79.1% and 67.6%. The model also improves DE-gene recovery, response discrimination, and Hallmark pathway recovery, supporting adaptive latent integration of biological priors for more generalizable and biologically interpretable perturbation modeling.

DGP-LA's source code and reproducibility materials are available at https://github.com/bieyl/DGP-LA and archived on Zenodo at https://doi.org/10.5281/zenodo.22155398.

Supplementary data are available at Bioinformatics online.