MV-MoCL: A Multi-View Molecular Contrastive Learning Framework by Integrating SMILES, 2D Graph, and 3D Geometry

Abstract

Machine learning-based molecular property prediction (MPP) has garnered significant attention in computer-aided drug discovery, with its primary goal being the accurate estimation of molecular properties using structural data to accelerate drug development. In recent years, multi-view-based molecular property prediction learning has aimed to integrate information from different molecular views to learn high-quality molecular representations. Furthermore, the three-dimensional geometric information of molecules encompasses a richer set of spatial features, which plays a critical role in enhancing the accuracy of molecular property prediction. However, existing models often overlook 3D information in molecules. To address this, we propose a contrastive learning model named MV-MoCL, which incorporates encoders for multiple molecular dimensions: a SMILES Transformer for the SMILES sequence view, a Graph Isomorphism Network (GIN) for 2D molecular graphs, and SchNet for 3D geometric conformations. By aligning representations from these views using a contrastive loss function, our approach captures rich, multi-faceted molecular features during pre-training, thereby improving performance on downstream molecular property prediction tasks and effectively mitigating the issue of scarce labeled data. The proposed model was evaluated on several benchmark datasets from MoleculeNet, and experiments demonstrate that MV-MoCL matches or surpasses existing models across multiple benchmarks.