Binding affinity prediction for protein-ligand complex using deep attention mechanism based on intermolecular interactions

Accurate prediction of protein-ligand binding affinity is important in that it can lower the overall cost of drug discovery in structure-based drug design. For more accurate prediction, many classical scoring functions and machine learning-based methods have been developed. However, these techniques tend to have limitations, mainly resulting from a lack of sufficient interactions energy terms to describe complex interactions between proteins and ligands. Recent deep-learning techniques show strong potential to solve this problem, but the search for more efficient and appropriate deep-learning architectures and methods to represent protein-ligand complexes continues. In this study, we proposed a deep-neural network for more accurate prediction of protein-ligand complex binding affinity. The proposed model has two important features, descriptor embeddings that contains embedded information about the local structures of a protein-ligand complex and an attention mechanism for highlighting important descriptors to binding affinity prediction. The proposed model showed better performance on most benchmark datasets than existing binding affinity prediction models. Moreover, we confirmed that an attention mechanism was able to capture binding sites in a protein-ligand complex and that it contributed to improvement in predictive performance. Our code is available at https://github.com/Blue1993/BAPA. Author summaryThe initial step in drug discovery is to identify drug candidates for a target protein using a scoring function. Existing scoring functions, however, lack the ability to accurately predict the binding affinity of protein-ligand complexes. In this study, we proposed a deep learning-based approach to extract patterns from the local structures of protein-ligand complexes and to highlight the important local structures via an attention mechanism. The proposed model showed good performance for various benchmark datasets compared to existing models.