HMRLBA: A hierarchical multi-scale representation learning model for predicting protein-ligand binding affinity

Abstract

Accurately assessing protein-ligand binding affinity is a cornerstone of drug discovery. Despite advances in computational affinity prediction, challenges persist in database quality, input representations, and architectural design. Integrating pre-trained large language models (LLMs) with deep neural networks presents a promising avenue to improve prediction accuracy while reducing computational costs. Here, we introduce HMRLBA, a hierarchical multi-scale representation learning model for predicting protein-ligand binding affinity. It employs a unique cascaded multichannel fusion strategy for protein representation learning, wherein sequence embeddings guide structural and surface feature learning. Evaluated across four PDBbind datasets, the external Metz benchmark, and the MUV active-decoy benchmark, HMRLBA consistently surpasses established SOTA methods across both affinity prediction and virtual screening paradigms. In virtual screening experiments, HMRLBA successfully identified EGFR inhibitors from a library of 2616 FDA-approved small-molecule drugs, with 11 of 13 drugs ranked within the top 30%. Furthermore, HMRLBA exhibited superior predictive accuracy in identifying novel drug candidates for Dot1L, a therapeutic target for leukemia, highlighting its potential to expedite early-stage drug discovery. These findings underscore HMRLBA's capability to enhance understanding of protein-ligand interactions and advance virtual screening pipelines.