P07-34

Molecular Properties Prediction by Contrastive Learning Using Graph Neural Network

Koshiro AOKI *, Apakorn KENGKANNA, Masahito OHUE

School of Computing, Institute of Science Tokyo
( * E-mail: aoki.k.as@m.titech.ac.jp )

Molecular properties are the chemical, biological and physical characteristics of a compound. Being able to predict the properties is useful in the search for the new drugs and machine learning models with molecular representation have been evolved. Particularly, molecular graph representations are the more naturalistic representation of compounds and molecular representation learning with GNN which can use graph representations as the input has achieved success. Despite this progress, it is difficult to learn all of the vast chemical space with machine learning due to the lack of labeled data.
In recent years, to address this problem, self-supervised learning has been investigated with large unlabeled data. In this study, we constructed the GNN contrastive learning model and verified the effects of augmentation strategies in contrastive learning.
We attempted to improve the performance of molecular properties prediction by using four augmentation strategies in contrastive learning. Prediction performance was evaluated on both classification and regression tasks. In addition to MoleculeNet, which are the molecular properties prediction benchmarks and widely used, performance was also evaluated on other biological data sets.