Domain-Aware Geometric Multimodal Learning for Multi-Domain Protein-Ligand Affinity Prediction

TL;DR

This work tackles the challenge of protein-ligand affinity prediction for multi-domain proteins by introducing DAGML, a domain-aware geometric multimodal framework that explicitly models domain modularity and inter-domain interfaces. DAGML combines a pre-trained protein language model with a domain-aware geometric encoder and a motif-centric ligand encoder, using cross-modal attention to align pharmacophoric motifs with inter-domain clefts. The authors curate a geometry-stratified multi-domain affinity benchmark and demonstrate substantial improvements over baselines, with the full DAGML-ID model achieving lower MSE and higher correlation metrics, especially for interface binders. They also provide comprehensive ablations to dissect the contributions of domain semantics, inter-domain message passing, and feature modalities, and discuss limitations related to linker noise and potential future gains from dynamic graph representations.

Abstract

The accurate prediction of protein-ligand binding affinity is important for drug discovery yet remains challenging for multi-domain proteins, where inter-domain dynamics and flexible linkers govern molecular recognition. Current geometric deep learning methods typically treat proteins as monolithic graphs, failing to capture the distinct geometric and energetic signals at domain interfaces. To address this, we introduce DAGML (Domain-Aware Geometric Multimodal Learning), a hierarchical framework that explicitly models domain modularity. DAGML integrates a pre-trained protein language model with a novel domain-aware geometric encoder to distinguish intra- and inter-domain features, while a motif-centric ligand encoder captures pharmacophoric compatibility. We further curate a specialized multi-domain affinity benchmark, classifying complexes by binding topology (e.g., interface vs linker binders). Extensive experiments demonstrate that DAGML achieves a 21% reduction in MSE and a Pearson correlation of 0.726 compared to strong baselines. Ablation studies reveal that explicit modeling of domain interfaces is the primary driver of this improvement, particularly for ligands binding in the clefts between structural units. The code is available at https://github.com/jiankliu/DAGML.

Figures (5)

• Figure 1: The DAGML framework. (A) Overview of the complete affinity prediction pipeline. The model processes ligand and protein inputs through respective encoders, followed by an Inter-Domain Modeling module and a Cross-Modal Multihead Attention mechanism that feeds into an MLP Predictor. (B) The protein feature processing stage. (C) Detailed Inter-Domain Modeling module.
• Figure 2: Examples of geometric interface classifications. (a) Single-Domain: A ligand binds to a protein consisting of one domain. (b) Single-Domain Binder: In a multi-domain architecture, the ligand interacts only with residues of a single domain. (c) Interface Binder: The ligand occupies the cleft between two distinct domains. (d) Linker Binder: The ligand interacts only with residues within a flexible linker region.
• Figure 3: Relative performance gain analysis across binding topologies. SD: Single-Domain, SDB: Single-Domain Binder, LB: Linker-Binder, IB: Interface-Binder.
• Figure 4: Impact of feature representation modalities. FP (blue): replacing Molecular Motif Learning with Morgan Fingerprints; SF (pink): sequential feature replacing ESM-GearNet with ESM-2; DAGML-ID (rose): full structure-and motif-aware framework.
• Figure 5: Impact of linker filtering on geometric noise. +Linker (purple) represents that the restrictions on flexible linker residues were relaxed during graph construction. MSE and RMSE are inverted so that larger areas indicate better performance.