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

TL;DR - PerturbRx predicts patient-level cancer drug response by learning treatment- and dose-conditioned latent molecular transitions from single-cell perturbation data. It improves aggregate performance on TCGA and patient-derived xenograft benchmarks without requiring patients’ post-treatment measurements.

• Learns intervention-induced transitions from context-matched but unpaired control and treated single-cell populations.
• Freezes and transfers the pretrained transition predictor to pretreatment patient molecular profiles.
• Combines predicted transitions with patient and drug representations for response prediction.
• Addresses data scarcity and tumor heterogeneity by explicitly modeling expected treatment-driven molecular changes.