Establishment of interpretable cytotoxicity prediction models using machine learning analysis of transcriptome features

Cytotoxicity, usually represented by cell viability, is a crucial parameter for evaluating drug safety in vitro. Accurate prediction of cell viability/cytotoxicity could accelerate drug development in the early stage. In this study, by using machine learning algorithms on cellular transcriptome and cell viability data, highly accurate prediction models of 50% and 80% cell viability were developed with AUROCs of 0.90 and 0.84, respectively, which also showed good performance on diverse cell lines. With respect to the characterization of Feature Genes employed, the models can be interpreted, and the mechanisms of bioactive compounds with narrow therapeutic indices can also be analyzed. In summary, the models established in this study have the capacity to predict cytotoxicity highly accurately across cell lines and can be used for high safety substances screening efficiently. Moreover, the Cytotoxicity Signature genes from interpretability analysis is valuable for studying the mechanisms of action, especially for substances with narrow therapeutic indices. Overall design: A549, HepG2 and HT-29 cells were seeded in 6-well plates and incubated with the test compounds or 0.1% v/v DMSO for 48 h. HT-29 cells (colon) were treated with cyclosporine A (CAS No: 59865-13-3 ) at a final concentration of 10 µM; A549 and HepG2 cells were treated with triptolide (CAS No: 38748-32-2) at final concentrations of 30 nM and 10 nM, respectively.