PerturbRx: Single-cell perturbations for patient drug response prediction

Can drug-induced changes learned from single-cell experiments help predict how a patient's tumor will respond to treatment?

Yoshitaka Inoue, Minoh Jeong, Alfred Hero, Rui Kuang, and Augustin Luna

What PerturbRx predicts

PerturbRx uses untreated and treated cell populations from Tahoe-100M to learn how drugs shift gene-expression profiles in a latent space. For each patient–drug pair, it estimates a treatment-induced shift from the pretreatment tumor profile. A response model combines that shift with patient and drug features to predict response without needing a post-treatment sample.

Where we tested it

  • Tahoe-100M provided the single-cell perturbation profiles for pretraining.
  • TCGA-508 included 508 treatment episodes from 462 patients across five drugs. An additional TCGA benchmark included 186 patient–drug records.
  • Patient-derived xenografts (PDX) provided an independent preclinical benchmark with 791 treatment episodes from 177 models across 10 drugs.

What we found

Adding the learned transition improved aggregate prediction over patient and drug features alone in the reported benchmarks. The contribution varied by drug.

Aggregate mean across five held-out folds, arXiv v1
BenchmarkAUROCAUPRC
TCGA-5080.6920.787
PDX0.6690.608

These retrospective benchmarks measured response prediction. Validating the predicted molecular changes requires matched pre- and post-treatment tumor profiles.

Citation

Inoue, Y., Jeong, M., Hero, A., Kuang, R., & Luna, A. (2026). PerturbRx: Learning Treatment-Conditioned Latent Transitions for Patient Drug Response Prediction. arXiv:2608.21349.