Graph Neural Network Model for Predicting Mitochondrial Toxicity via Complementary Molecular Features

Abstract

In this study, we introduce a hybrid Graph Neural Network model for predicting mitochondrial toxicity, a significant contributor to drug-induced liver injury. This model effectively captures both the structural and physicochemical properties of compounds by integrating graph embeddings and molecular fingerprints. After systematically optimizing the model architecture, we found that the combination of the GATv2Conv convolutional layer, 64 hidden neurons, and a single predictor layer was the most effective configuration for our purposes. When benchmarked against well-established 3D GNN models like DimeNet++ and ViSNet, our hybrid model consistently outperformed them, exhibiting improvements in AUC-ROC by up to 18.3% and 6.2%, respectively. These results emphasize the importance of achieving an optimal balance when integrating features and demonstrate the potential of our hybrid GNN model for computational toxicity prediction. This model could notably expedite drug discovery, reduce dependence on animal testing, and provide valuable insights into molecular toxicity.