Variational Inference Driven Drug Protein Binding Prediction

The identification of drug-protein interactions (DPIs) is a key task in drug discovery, where drugs are chemical compounds and targets are proteins. Traditional DPI prediction methods are either time consuming (simulation-based methods) or heavily dependent on domain expertise (similarity-based and feature-based methods). Recent explorations involving deep learning either exploit 3D structure of the proteins and/or use GNNs (graph convolutions) to capture neighbour relationships. In this paper, we present a novel end-end deep learning architecture, VED-BI, that leverages variational inference based encoder and decoder along with GraphSage based approach to predict drug-protein interaction. Due to better generalization, our architecture is able to deliver better results on the drug protein binding prediction. Detailed experimental comparative analysis using Precision, Recall and AUC metrics, has been performed across all the approaches on metador and bindingDB datasets. Our proposed novel architecture performs better on bindingDB dataset as compared to key state-of-the-art results.