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 <i>in vitro</i>. Accurate prediction of cell viability/cytotoxicity could accelerate drug development in the early stage. In this study, by integrating cellular transcriptome and cell viability data using four machine learning algorithms (support vector machine (SVM), random forest (RF), extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM)) and two ensemble algorithms (voting and stacking), highly accurate prediction models of 50% and 80% cell viability were developed with area under the receiver operating characteristic curve (AUROC) of 0.90 and 0.84, respectively; these models also showed good performance when utilized for diverse cell lines. Concerning the characterization of the employed Feature Genes, the models were interpreted, and the mechanisms of bioactive compounds with a narrow therapeutic index (NTI) can also be analyzed. In summary, the models established in this research exhibit superior capacity to those of previous studies; these models enable accurate high-safety substance screening <i>via</i> cytotoxicity prediction across cell lines. Moreover, for the first time, Cytotoxicity Signature (CTS) genes were identified, which could provide additional clues for further study of mechanisms of action (MOA), especially for NTI compounds.