HMS-PLI: A multi-task deep learning framework for protein-ligand interaction prediction with hierarchical multi-scale encoding

Protein–ligand interactions (PLIs) are fundamental to understanding cellular signaling, drug efficacy, and molecular recognition, making their accurate prediction essential for computational drug discovery. However, traditional experimental methods for identifying PLIs are often costly, time-intensive, and limited in scalability, highlighting the need for efficient computational alternatives. In this study, we propose HMS-PLI, a novel deep learning framework that integrates a Hierarchical Multi-Scale (HMS) encoding strategy with an advanced multi-task architecture to enhance PLI prediction. Our HMS encoding integrates four complementary strategies: one-hot sequence identity, physicochemical properties, sinusoidal positional encoding, and multi-scale context features, transforming variable-length sequences into rich numerical representations that capture information across multiple biological organizational scales. The deep learning architecture incorporates three core innovations: Adaptive Multi-Head Attention (AMHA) for dynamic feature weighting, a Hierarchical Feature Fusion Network (HFFN) to model multi-scale interactions, and Task-Adaptive Gating (TAG) to handle classification and regression tasks within a unified model. Evaluated across five benchmark datasets, Human, C. elegans, and DUD-E for binary interaction classification, and BindingDB Ki and PDBbind 2016 for binding affinity regression, HMS-PLI achieves competitive performance relative to established computational methods. On the DUD-E classification benchmark, the model attains an AUC of 0.999, surpassing GanDTI (0.997) and PLM-PLI (0.986). On the BindingDB Ki regression task, it achieves an RMSE of 0.446 and a Pearson correlation of R = 0.901, outperforming all compared baselines including GanDTI (RMSE = 0.721) and DeepDTA (RMSE = 0.782). These findings indicate that HMS-PLI provides a computationally efficient, low-parameter framework for sequence-level pattern recognition within standardized benchmark spaces. While the model achieves high internal classification and regression scores using only raw protein sequences and ligand SMILES strings, these metrics reflect performance bounded strictly by the text regularities and curation styles of the evaluation databases, rather than a demonstration of broad, uncurated external generalization.