ACR-GNN: Adaptive Cluster Reinforcement Graph Neural Network Based on Contrastive Learning

It has been amply proven that the graph neural networks (GNNs) are effective at various graph-level tasks. The chemical molecule properties prediction based on GNNs have achieved outstanding results in drug discovery. However, many existing methods, when performing qualitative calculations, fail to notice the effect of critical atomic clusters in a chemical molecule. At the same time, inadequate utilization of high- and low-frequency information in a chemical molecular graph and the poor interpretability of graph models limit the application of GNNs in the field of chemical molecular research. This paper proposes a two-stage chemical molecular property prediction model, called the ACR-GNN, based on a contrastive learning architecture. By adaptively selecting critical atomic clusters and using high- and low-frequency signals simultaneously when performing graph convolution, the ACR-GNN can yield accurate predictions and insights. The experimental results of a comparison with other methods show that the ACR-GNN can effectively predict chemical molecular properties, and the test accuracy performance of ACR-GNN reaches over 96% in the determination of whether a molecule is toxic. In particular, the ACR-GNN has excellent interpretability in chemical molecular properties because of its use of the Get Key Clusters (GKC) layer, which finds the vital atomic clusters by adaptively learning the global features of a molecule. This contribution helps in the discovery and analysis of some of the potential functional groups of a molecule.