Incorporating Neighboring Protein Features for Enhanced Drug-Target Interaction Prediction: A Comparative Analysis of Similarity-Based Alignment Methods.

Abstract

Drug-target interaction (DTI) prediction is a fundamental computational task in drug discovery. Despite recent advancements, existing approaches often suffer from data sparsity and fail to capture the intricate nature of molecular interactions, limiting predictive performance. To address these challenges, we propose a novel DTI prediction framework that enhances both accuracy and interpretability by incorporating features from highly similar protein neighbors. Our framework extracts chemical and physicochemical features from drug-target binding affinity data and integrates interaction features from highly similar protein neighbors to enrich representation. To identify these neighbors, we employ a range of protein similarity alignment algorithms, including BLAST, MUSCLE, MAFFT, Clustal Omega and Foldseek. Experiments on the Davis and KIBA data sets demonstrate that incorporating features from high-similarity neighbors substantially improves prediction accuracy. Further analysis reveals that top-ranked neighbors contribute the most to performance gains, underscoring the importance of similarity-based feature augmentation. Additionally, comparisons among alignment methods highlight their robustness in neighbor selection, and case studies confirm the biological relevance of shared targets among closely related proteins. Overall, our framework presents a novel solution to data sparsity, improves predictive performance, and enhances model interpretability. This work lays a solid foundation for precise DTI prediction and provides valuable insights for advancing computational methods in drug discovery.