Molecular Machine Learning (ML) has been widely investigated for efficient molecule property prediction. However, there are three challenges in implementing ML methods for molecules. (1) Acquisition of molecular properties is costly. Obtaining labels of molecules usually requires sophisticated and time-consuming simulations or wet-lab experiments. Consequently, the number of labels in most molecular learning benchmarks is far from adequate. (2) The magnitude of chemical space is gigantic. For example, it is estimated to have $10^{60}$ potential pharmacologically active molecules. ML models trained on limited labeled data find it difficult or even infeasible to generalize to novel molecules from the huge chemical space. (3) The topology of molecules is hard to embed. For instance, string-based descriptors, like SMILES, cannot directly embed the graphical structure of molecules. The recent development of Graph Neural Networks (GNNs) preserves topological information of molecules through message aggregation, facilitating learning on molecule graphs.<br/>To address the challenges, we propose MolCLR: Molecular Contrastive Learning of Representations via Graph Neural Networks (GNNs), a self-supervised learning framework that leverages large unlabeled data (~10M unique molecules). In MolCLR pre-training, we build molecule graphs and develop GNN encoders to learn differentiable representations. Three molecule graph augmentations are proposed: atom masking, bond deletion, and subgraph removal. A contrastive estimator maximizes the agreement of augmentations from the same molecule while minimizing the agreement of different molecules. By this means, GNN models are pre-trained on large unlabeled data in a self-supervised learning manner, which bears promise for learning generic representations on the large chemical space and reduces the effort to collect molecular labels.<br/>Experiments show that our contrastive learning framework significantly improves the performance of GNNs on various molecular property prediction benchmarks including both classification and regression tasks. Benefiting from pre-training on the large unlabeled database, MolCLR outperforms other supervised and self-supervised baselines on multiple benchmarks. Additionally, further investigations demonstrate that MolCLR learns to embed molecules into representations that can distinguish chemically reasonable molecular similarities.

Efficient representations of molecules learned with contrastive learning and graph neural networks

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.