Attention-based method to predict drug-target interactions across seven protein classes

Abstract

Most approved drugs bind with proteins to modulate their activity for treating a diverse range of diseases. Unfortunately, drug development is a long and costly process. Computational methods seek to accelerate drug discovery by predicting drug-target interactions, thus facilitating compound screening and drug repurposing. This thesis presents a deep learning approach for predicting interactions between compounds and proteins categorized into seven classes. The models utilize self-attention found in transformer neural networks to learn continuous interaction values from multimodal compound-protein representations. The models were evaluated in three test settings of increasing difficulty. Most models showed competitive predictive capabilities in the two easier settings, with the most difficult test rendering them ineffective. In particular, the kinase model demonstrated state-of-the-art performance in the bioactivity imputation task when compared against other methods, while leaving room for improvement in the new compound scenario. Furthermore, conformal predictors with uncertainty estimates displayed equivalent performance to contemporary methods and provided directions for future research.