Deep learning-based screening and design of novel therapeutics that reverse cancer-associated transcriptional phenotypes

Identifying drugs that reverse expression of disease-associated transcriptomic features has been widely explored as a strategy for discovering drug repurposing candidates, but its potential for novel compound discovery and optimization remains largely underexplored. Here, we present a deep learning-based drug discovery platform, guided by transcriptomic features, that screens large compound libraries and optimizes lead compounds. We first develop a model that predicts gene expression changes solely from chemical structures and deploy it to infer the expression changes induced by compounds in large screening libraries. We then refine compound scoring and employ a Monte Carlo Tree Search method for multi-objective optimization. By incorporating Structure-Gene-Activity Relationships, we uncover drug mechanisms directly from transcriptomic data. To demonstrate the utility of the system, we identify and validate compounds for hepatocellular carcinoma (HCC). In HCC, we design a novel compound that improves the IC 50 from 4 µM to 0.5 µM, with increased in vitro selectivity, favorable pharmacokinetics and in vivo activity.