GSNet: Multimodal Deep Learning for Predicting Molecular Metabolic Stability

Accurate prediction of molecular metabolic stability is considered to be of significant practical value in the field of drug development, particularly in ensuring the safety and efficacy of drugs. The metabolic stability of drugs has been effectively predicted by existing deep learning methods through the analysis of the structural features of drug molecules. However, the propagation of molecular graph information is primarily focused on by most of the current methods, with limited consideration given to molecular sequence data. The Graph-Seq Network (GSNet) model, which is proposed in this paper, is aimed at predicting the metabolic stability of compounds. A deep learning framework has been designed to integrate the SMILES representation of compounds and molecular graph data for the extraction of molecular features. Specifically, for molecular sequence information, a bidirectional long short-term memory network (BiLSTM) is utilized to capture the context of symbols and to learn the sequence representation of molecules. For molecular graph data, a graph encoder is introduced, through which the graph structure representation is learned by capturing the consistency between global and local structural information. Finally, the sequence and structural information are combined by the model through fully connected layers for model training. It is demonstrated by experimental results that, in comparison to alternative methods, the GSNet model performs more effectively across a range of public datasets. The importance of both the sequence encoder and the graph encoder in the model is also confirmed by ablation studies.