Hierarchical Multi-Scale Modeling of Absolute Binding Affinity in Protein Complexes

Dongyun Kim, Soorin Yim, Sungjoon Park, Kiwoong Yoo, Kyungwook Lee, Doyeong Hwang, Soonyoung Lee, Jongseong Jang, Kiyoung Kim

Confirmation: I have read and agree with the workshop's policy on behalf of myself and my co-authors.

Track: tiny / short paper (2-4 pages excluding references; extended abstract format)

Keywords: protein-protein interactions, binding affinity prediction, hierarchical representation learning, graph neural networks

TL;DR: We propose HierBind, a hierarchical graph-based model that predicts protein binding affinity by integrating multi-scale structural features and employing a robust data-relabeling strategy.

Abstract: Predicting the absolute binding affinity ($\Delta G$) of a protein-protein complex from structure alone remains challenging: physics-based simulations are costly, and experimental affinity labels are noisy and heterogeneous. We propose a hierarchical architecture that represents each complex at multi-scale resolutions-atoms, residues, chains, and the whole complex. We use distance-weighted message passing so that closer atom/residue pairs contribute more strongly, and then pool information across levels to produce a single binding affinity score. The model further incorporates transferable physicochemical priors via pretrained representations. On PPB-Affinity dataset, our method improves rank correlation over a strong baseline (Spearman's rank correlation coefficient $0.659$ vs. $0.646$). Ablations show that distance-weighted message passing, multi-scale modeling, and physically grounded representations each contribute to model performance.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Submission Number: 82