Velodrome is a method of drug response prediction that takes gene expression of cell lines (labeled) and patients (unlabeled) as input and predicts Area Above dose-response Curve as output. The model learns a predictive invariant representation using a standard supervised loss on cell lines, an alignment loss on cell lines and patients, and a consistency loss on patients.

Our experimental results on cell lines, PDX samples, and clinical trial data suggest very promising performance. We demonstrated generalization to well- and under-represented tissue types as well as completely new tissue types w.r.t tissue types available in training domains.


Code and data:

November 2021 UPDATE! Velodrome is now published in Nature Machine Intelligence: