Few-shot graph and smiles learning for molecular property prediction

Abstract

Molecular property prediction can be applied to discover new drugs, which has attracted significant attention from both chemists and machine learning researchers. However, the existing methods for predicting molecular properties have the following limitations. The molecular representions learned from 2D structures will lose some important information contained in the molecular representions learned from 1D SMILES strings and vice versa; Second, the few-shot learning for molecular property prediction ignores the great differences between properties of different moleculars. To address the above problems, we propose a molecular property prediction model based on few-shot learning, called FewGS in this paper. We present a molecular graph-string combination method to fuse SMILES and graph represention, which can exploit a small amount of available molecular information to capture the hidden feature of moleculars. To promote the few-shot learning for molecular property prediction, we also propose a sample space transformation strategy to effectively eliminate bias between meta-train data and meta-test data. In addition, we construct a loss function based on graph-string combination, and adjust the weights through a self-attentive mechanism to achieve accurate prediction. We evaluate FewGS on molecular datasets and experimental results show that FewGS achieves state-of-the-art performance.