Multitask Learning Approach to Predict Hepatotoxicity Phenotypes of Chemicals in Clinical Assays

Hepatotoxicity is a key concern for developing new drugs since it is difficult to anticipate in the early stages of drug discovery due from pharmacologically unrelated drug effects (i.e., drug accumulation, metabolism, and pathway interference). 1,2 Despite that, the reliable prediction of hepatic adverse effects using traditional single-task learning (STL) approaches is a substantial challenge in clinical development. Here, we developed a multitask learning (MTL) strategy based on inductive transfer of knowledge to simultaneously predict LOAEL values of untested compounds for several clinical hepatotoxicity endpoints. Initially, a dataset comprising 1083 compounds with hepatic adverse effects in humans was collected from various sources. The complex pathology was clustered into 15 clinical hepatotoxicity endpoints. Then, an MTL model was developed using a Feedforward Neural Network (FNN) architecture and molecular fingerprints (ECFP4). The best MTL model significantly outperforms past STL methods, with acceptable average RMSE values (train ~0.24, validation ~0.62, test ~0.65).