Molecular contrastive learning of representations via graph neural networks

Molecular Machine Learning (ML) bears promise for efficient molecule property prediction and drug discovery. However, labeled data can be expensive time-consuming to acquire. Due the limited data, it is a great challenge supervised-learning ML models generalize giant chemical space. In this work, we present MolCLR: Contrastive of Representations via Graph Neural Networks (GNNs), self-supervised learning framework that leverages large unlabeled (~10M unique molecules). MolCLR pre-training, build graphs develop GNN encoders learn differentiable representations. Three graph augmentations are proposed: atom masking, bond deletion, subgraph removal. A contrastive estimator maximizes agreement from same while minimizing different molecules. Experiments show our significantly improves performance GNNs on various molecular benchmarks including both classification regression tasks. Benefiting pre-training database, even achieves state-of-the-art several challenging after fine-tuning. Additionally, further investigations demonstrate learns embed molecules into representations distinguish chemically reasonable similarities.