PerturbRx: Learning Treatment-Conditioned Latent Transitions for Patient Drug Response Prediction

Scarce data and tumor heterogeneity limit patient-level prediction of cancer treatment response. Existing approaches predict response from pretreatment molecular profiles and drug representations, without explicitly modeling treatment-induced molecular changes. We propose PerturbRx, a treatment-conditioned representation learning framework that learns drug- and dose-conditioned latent transitions from context-matched but cell-unpaired control and treated single-cell populations. We freeze the learned transition predictor and transfer it to pretreatment patient profiles, where we combine its predicted transition with the patient state and drug representation to predict response without post-treatment measurements. Across two patient-disjoint TCGA cohorts and an independent model-disjoint PDX benchmark, PerturbRx achieves competitive or leading out-of-fold performance relative to static and transition-based controls. Its gains are strongest in the larger clinical cohort and remain evident in within-treatment analyses and cross-domain PDX transfer, although performance is treatment-dependent and simple transition-geometry measures do not explain the observed gains.