SynthMol: A Drug Safety Prediction Framework Integrating Graph Attention and Molecular Descriptors into Pre-Trained Geometric Models.

Abstract

Drug safety is affected by multiple molecular properties and safety assessment is critical for clinical application. Evaluating a drug candidate's therapeutic potential is facilitated by machine learning models trained on extensive compound bioactivity data sets, presenting a promising approach to drug safety assessment. Here, we introduce SynthMol, a deep learning framework that integrates pre-trained 3D structural features, graph attention networks, and molecular fingerprints to achieve high accuracy in molecular property prediction. Evaluation of SynthMol on 22 data sets, including MoleculeNet, MolData and published drug safety data, showed that it could provide higher prediction accuracy than state-of-the-art model in most tasks. SynthMol achieved an ROC-AUC value of 0.944 in the BBBP data set, 2.61% higher than the next best model, and an ROC-AUC of 0.906 on the hERG data set, a 2.38% improvement. Validation of SynthMol in real-world applications with experimentally determined hERG toxicity and CYP inhibition data supported its capacity to distinguish functional changes for drug development. The implementation code and data are available at https://github.com/ThomasSu1/SynthMol.