RAPTORGraph: graph-based pathway modeling for causal discovery in single-cell perturbations

RAPTORGraph, a beta-variational autoencoder with a GraphPathway encoder, learns gene-to-gene interactions in pathways and improves unseen-perturbation downstream performance.

• Why it matters: Perturb-seq–style closed-world interventions suffer from intervention spillover and prior-pathway bias, limiting identifiable causal discovery and the ability to capture signals outside annotations.
• What they did: The method uses preconditioning to isolate perturbed genes into clean single-node latent interventions, and trains on optimal-transport aligned control–perturbed data to stabilize latent mappings.
• The result: It improves state-of-the-art results on analyses of unseen perturbations, including non-additive interactions, while remaining competitive on mean squared error and maximum mean discrepancy.

AI summary of the abstract · may contain errors · not medical advice · read the paper.