RAPTORGraph: Graph-Based Pathway Modeling for Causal Discovery in Single-Cell Perturbations

Experiments involving perturbation of single cells are central to understanding cellular mechanisms and accelerating therapeutic discovery, however the space of combinatorial perturbations is intractably large. Causal representation learning has shown great promise for predicting unseen combination of perturbations, but existing methods often suffer from mean collapse and intervention spillover, which violate the theoretical requirements for identifiability guarantees. We introduce RAPTORGraph, an end-to-end VAE framework that addresses these issues through: i) a preconditioned GraphPathway encoder that enforces intervention-guided mappings to causal meta-pathways, enabling clean single latent-node interventions needed for causal identifiability, and ii) optimal-transport alignment that aligns control and perturbed populations to stabilize conditional generation. Empirical results on Perturb-seq datasets demonstrate that RAPTORGraph improves the trade-off between reconstruction fidelity and distributional matching, and yields improved performance on predicting non-additive combinatorial effects. Finally, we show that RAPTORGraph recovers biologically meaningful latent programs and a causal graph over meta-pathways, providing an interpretable bridge between generative quality and mechanistic insight.