Prediction of Drug-Target Interactions Using BERT for Protein Sequences and Drug Compound

Abstract

The current drug development crucially depends on identifying potential relationships between medicines and targets. However, anticipating such relationships is difficult due to the limits of current computational techniques. Hence, the use of deep learning is essential for identifying potential therapeutic drug compounds and providing support throughout the entire drug development process. This study discusses the deep learning technique of using bidirectional encoder representations from the Transformers (BERT) model which helped to build representations using protein and drug SMILES (Simplified Molecular Input Line Entry System) dataset to enhance DTI prediction. We used the pretrained Protein BERT model and ChemBERT for protein sequences and drug SMILES data respectively for feature extraction and resulting features are concatenated together and fed into a random forest (RF) for classification. BERT model helps to use protein and drug datasets for feature extraction without using the descriptor dataset for finding the interaction between drugs and proteins. .