Graph-guided flow matching enables biologically informed prediction at single-cell resolution

Abstract

Perturb-seq enables systematic interrogation of massively parallel genetic perturbations at single-cell resolution, yet exhaustive characterization of the full perturbation space remains impractical. Here, we present Graph-guided Flow Matching (GFM), a generative framework that integrates biological knowledge graphs with latent flow matching to predict post-perturbation single-cell transcriptomes. Across multiple large-scale Perturb-seq datasets, GFM outperforms a strong linear baseline and current deep-learning models in predicting unseen single-gene perturbations, while training up to 30x faster and remaining robust with limited training data. GFM also demonstrates competitive performance in the two-gene perturbation prediction task when neither perturbed gene is observed during training. Ablation experiments show that combining flow matching with knowledge graphs enables accurate prediction of perturbation responses. Finally, GFM accurately calculates transcriptional consequences of HSP90AA1 inactivation and identifies target candidates unseen by the model in chronic myeloid leukemia. Our study establishes GFM as a biologically informed generative framework that predicts single-cell perturbation responses for prioritization of therapeutic targets.