M2Mol: Multi-view Multi-granularity Molecular Representation Learning for Property Prediction

Molecular property prediction has drawn considerable attention in drug discovery, material design, and biological system analysis in recent years. Particularly, artificial intelligence-driven computational methods hold great promise in expediting the molecular prediction process. While sequence-based methods and graph-based methods have significantly enhanced molecular property prediction, they still exhibit limitations in capturing intricate multi-dimensional molecular information and fully exploiting multi-level structural units. We present a Multi-view Multi-granularity Molecular representation learning framework, M2Mol. Specifically, M2Mol jointly models the molecular sequential arrangement and structural topology with the sequence view and graph view, respectively. In each view, M2Mol captures atom-level attributes and motif-level semantics to intensify the perception of the molecular multi-granularity structural units. In addition, we design inconsistency loss to promote the alignment between the sequence and graph views, and mutual information loss to capture the complementarity between the atom and motif levels. Finally, extensive experiments conducted on six real-world datasets demonstrate the superior effectiveness of the proposed model across both classification and regression tasks.