Classification prediction of drug target binding affinity based on the MolrProtTrans model

Abstract

Predicting drug-target interactions is essential for virtual drug screening. While many models predict the binding affinity between small molecules and proteins, they often overemphasize molecular features while overlooking important protein characteristics, leading to biased predictions. Traditional deep learning models, such as transformerCPI, show limited performance in tasks like label inversion on G protein-coupled receptor (GPCR) datasets. To address this, this study proposes an enhanced transformer-based model that integrates both molecular and protein information. By leveraging Molr and ProtTrans networks for feature extraction, and incorporating a transposed attention mechanism with a triple-loss self-supervised learning approach, the model improves prediction accuracy. Experimental results show that the proposed model achieved observed Area Under the Curve (AUC) of 0.81 on the GPCR label-inversion dataset and 0.92 on a human target dataset, which are numerically higher than those reported for TransformerCPI and several baseline methods in our experiments. These observations indicate improved performance in our experimental setting, offering promising prospects for advancing virtual drug screening and drug discovery.