Transfer Learning Empowers the Development of Better-Performing Models for In Vivo Cardiotoxicity by Learning from In Vitro hERG Binding Data

Abstract

Evaluating chemical hazards, such as cardiotoxicity, is critical for sound chemical management but is challenging due to the labor-intensive and time-consuming nature of in vivo testing. In silico models can efficiently and cost-effectively screen potential cardiotoxic compounds from a large number of chemicals. However, the available small-scale in vivo toxicity data sets prevent models from achieving superior prediction accuracy. Herein, a transfer learning-based graph neural network (TL-GNN) model for predicting in vivo cardiotoxicity was developed by fine-tuning from a pretrained model on a data set regarding the human ether-a-go-go-related gene binding, a molecular initiating event that may ultimately cause cardiotoxicity. The TL-GNN model not only outperformed the GNN without TL and classic machine learning models, but also those fine-tuned from GNN models pretrained on a nonmechanistic octanol–water partition coefficient data set and Bidirectional Encoder Representations from Transformers models pretrained on a data set with over a million unlabeled compounds. A structure–activity landscape-based method was adopted to characterize the applicability domains of the TL-GNN model. The TL-GNN model with defined applicability domains was applied to screen approximately 600,000 chemicals across diverse categories, identifying over 25,000 as cardiotoxic. The robustness of the TL-GNN model suggests the potential of the modeling strategy for complex in vivo toxicity by learning from adverse outcome pathway (AOP)-related in vitro targets.