Semi-supervised learning-based virtual adversarial training on graph for molecular property prediction

The prediction of molecular properties is a crucial task for drug screening and new drug discovery. It is remains a huge challenge to find small molecule drugs with good pharmacological, toxicological and pharmacokinetic properties in the field of drug discovery, that will accelerate the process of drug discovery and save a lot of resources. Most current methods carry out property prediction based on supervised learning pattern. However, the limited label data hinders the related research. Manually labeling data is expensive and will bring many noise. A large number of unlabeled data contains valuable information. The neural network can learn more knowledge by learning the unlabeled data. In this paper, we propose GVAT, a novel Virtual Adversarial Training method on Graph based on unlabeled data. Specifically, we apply GVAT to the semi-supervised learning tasks on molecular property prediction. Comparing with current graph-based and conventional machine learning-based models, GVAT achieves competitive performance with only a small amount of label data. GVAT has better robustness and can defend graph neural networks against adversarial attacks, although it does not exceed the performance of supervised learning methods. Ablation study demonstrates the effectiveness of our proposed GVAT.