Similarity-Aware BERT Model for Drug-Target Interaction Prediction

Abstract

Modern drug discovery relies on accurate computational methods for predicting drug-target interactions. When trained on hundreds of millions of protein sequences, transformer-based methods such as ProteinBERT demonstrate high accuracy in predicting protein function and structural properties, including disordered regions and secondary structure. However, these models are limited in their ability to accurately predict associations between proteins and other biomedical entities, such as drugs. This is in part due to the lack of explicit regularization that incorporates biological and chemical similarity structures. In this study, we propose a model, SimBERT-DTI, that incorporates sequence similarities among proteins and structural similarities among drugs to enhance predictions of drug-target interactions (DTIs). The results achieved by this prototype model demonstrate that the general framework used in SimBERT-DTI can be utilized to develop more accurate and robust DTI prediction methods.