HPDAF: A practical tool for predicting drug-target binding affinity using multimodal features

Abstract

Accurate prediction of drug-target binding affinity is crucial for efficient drug discovery and design, enabling researchers to better understand molecular interactions and accelerate the identification of promising drug candidates. Despite recent advances, existing computational methods often face difficulties in effectively combining detailed structural information from drug-binding sites with broader molecular features. Here, we introduce HPDAF, a practical multimodal deep learning tool designed to improve the accuracy of drug-target binding affinity predictions. HPDAF uniquely integrates three types of biochemical information: protein sequences, drug molecular graphs, and structural interaction data from protein-binding pockets. Each of these data types is carefully processed using specialized modules that capture essential molecular characteristics. A novel hierarchical attention-based mechanism then effectively combines these diverse features, enabling the model to dynamically emphasize the most relevant structural and sequential information. Extensive evaluations using widely recognized benchmark datasets, including CASF-2016 and CASF-2013, demonstrate that HPDAF consistently achieves superior predictive performance compared to current state-of-the-art methods. The practical applicability and enhanced accuracy of HPDAF highlight its potential as a valuable computational tool for medicinal chemists involved in drug design and virtual screening efforts.