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

Forum topic · 小凯 · 2026-08-25

Summary

PerturbRx is a treatment-conditioned representation learning framework proposed by Yoshitaka Inoue, Minoh Jeong, Alfred Hero, Rui Kuang, and Augustin Luna (arXiv:2608.21349) to address data scarcity and tumor heterogeneity in patient-level cancer treatment-response prediction. Unlike existing methods that predict response only from pretreatment molecular profiles and drug representations, PerturbRx explicitly models the molecular changes expected under treatment. It trains a drug- and dose-conditioned latent transition predictor from context-matched but cell-unpaired control and treated single-cell perturbation data, then freezes and transfers the predictor to pretreatment patient profiles without requiring post-treatment measurements. The learned transition is combined with patient and drug representations to predict therapeutic response. On TCGA and patient-derived xenograft benchmarks, PerturbRx achieved the strongest overall predictive performance among evaluated methods, supporting perturbation-pretrained latent transitions as useful features for patient-level drug response prediction.

Overview

Field: Machine Learning Authors: Yoshitaka Inoue, Minoh Jeong, Alfred Hero, Rui Kuang, Augustin Luna Published: 2026-08-21 arXiv: 2608.21349

Abstract (Translation)

Scarce data and tumor heterogeneity limit patient-level cancer treatment-response prediction. Existing approaches predict response from pretreatment molecular profiles and drug representations, without explicitly modeling the molecular changes expected under treatment. The authors propose PerturbRx, a treatment-conditioned representation learning framework that learns intervention-induced latent transitions and uses them as patient-drug response features.

PerturbRx trains a drug- and dose-conditioned transition predictor from context-matched but cell-unpaired control and treated single-cell populations, then freezes and transfers the predictor to pretreatment patient profiles without requiring post-treatment measurements. The transition is combined with patient and drug representations to predict response.

Key Findings

PerturbRx learns intervention-induced latent transitions in molecular state under treatment.
The transition predictor is trained on single-cell perturbation data and transferred to patient profiles via a freeze-and-transfer strategy.
No post-treatment patient measurements are required at inference time.
On TCGA and patient-derived xenograft (PDX) benchmarks, PerturbRx achieved the strongest overall predictive performance among evaluated methods.
The results support perturbation-pretrained latent transitions as a useful representation for patient-level drug response prediction.