Ensemble graph neural networks for structural classification of HIV inhibiting molecules

Graph Neural Networks (GNN) have proved to be extremely effective in molecular machine learning in the past few years. Identifying whether a molecule can inhibit the Human Immuno-Deficiency Virus (HIV) based on its structure is one such task that can be performed by harnessing the power of graph neural networks. This paper presents an ensemble graph neural network based approach to classify molecules as HIV inhibitors by using graph attention networks and graph convolutional networks. Graph representations are constructed from these molecules and are used for training the proposed GNN models. The ensemble GNN model achieves a test accuracy of 86.4% along with a high precision value for both the classes-HIV inhibitors and non-inhibitors. This paper further shows that ensemble models can be used in GNNs to overcome the problems related to biasing of models towards a particular class specifically for imbalanced classification tasks which are common in medical and bioinformatics datasets.