DTI Prediction Network Based on Protein Language Models and Graph Neural Networks

The prediction of drug-target interactions (DTI) is of paramount importance in the field of drug discovery and development. On one hand, it has dramatically expedited the pace of new drug development; on the other hand, it opens new avenues for drug repositioning. With the rapid advancements in deep learning, data-driven DTI prediction methods have gradually become mainstream. However, existing approaches still face significant challenges in feature extraction and information fusion. For this problem, we have constructed an innovative DTI prediction model that integrates a protein language model with a graph neural network, aiming to enhance the accuracy of interaction prediction between drugs and targets. Our model leverages a Graph Feature Learning Module (G-FLM) to process molecular graph structures derived from SMILES sequences, extracting informative drug representations. For proteins, we employ the ESM2 language model to generate feature representations from amino acid sequences, followed by a self-attention mechanism to further refine these features. To assess the capability of the model, metrics such as ACC, Pre, and Rec are employed, and we validated its effectiveness on three public datasets. Experimental results reveal that the proposed model is significantly superior to existing baseline and achieves excellent prediction accuracy in DTI tasks.