MDFCL: Multimodal data fusion-based graph contrastive learning framework for molecular property prediction

Abstract

Molecular property prediction is a critical task with substantial applications for drug design and repositioning. The multiplicity of molecular data modalities and paucity of labeled data present significant challenges that affect algorithmic performance in this domain. Nevertheless, conventional approaches typically focus on singular data modalities and ignore either hierarchical structural features or other data pattern information, leading to problems when expressing complex phenomena and relationships. Additionally, the scarcity of labeled data obstructs the accurate mapping of instances to labels in property prediction tasks. To address these issues, we propose the Multimodal Data Fusion-based graph Contrastive Learning framework (MDFCL) for molecular property prediction. Specifically, we incorporate exhaustive information from dual molecular data modalities, namely graph and sequence structures. Subsequently, adaptive data augmentation strategies are designed based on the molecular backbones and side chains for multimodal data. Built upon these augmentation strategies, we develop a graph contrastive learning framework and pre-train it with unlabeled data ($\sim$ 10M molecules). MDFCL is tested using 13 molecular property prediction benchmark datasets, demonstrating its effectiveness through empirical findings. In addition, a visualization study demonstrates that MDFCL can embed molecules into representative features and steer the distribution of molecular representations.