Development of Data Augmentation Method for Graph Data by Perturbating Feature Vectors

Abstract

In drug discovery and material design, efficient methods to search for novel molecules with desired properties are needed. The use of graph neural networks (GNNs) as a quantitative structure-property relationship model enables virtual screening of candidate structures with better prediction performance than that of conventional feature extraction methods. However, previous studies have used a large amount of structure and property data for the training. In molecular design, where a large amount of data are difficult to collect, GNNs may fail to predict properties accurately. In this study, we designed a GNN model called Perturbating Message Passing Neural Network (PMPNN), which is based on MPNN, to augment graph data by adding perturbations to feature vectors during message passing operation. We compared PMPNN with MPNN on the QM9 dataset, verified the effectiveness of the proposed method, and discussed the effect of the perturbation on predictions. It was also shown that the proposed method could achieve the same level of prediction performance with about half of the dataset, and suggested that it can extract features successfully even with a small amount of graph data.