A physics-informed graph neural network to approximate docking-based binding affinity for DYRK2 in Alzheimer’s drug repurposing

TL;DR

This paper introduces a physics-based machine learning model to predict drug binding to a protein linked to Alzheimer's, using computational docking data as a reference.

Contribution

The novel contribution is a physics-informed graph neural network (PhysDual-GCN) that approximates docking scores for DYRK2 in Alzheimer's drug repurposing.

Findings

PhysDual-GCN achieved low absolute errors (MAE = 0.31 kcal/mol; RMSE = 0.44 kcal/mol) compared to docking scores.

The model correctly identified strong binders like donepezil and brexpiprazole.

Integrating physical interaction terms improves interpretability and efficiency compared to traditional docking methods.

Abstract

Alzheimer’s disease (AD) requires the discovery of new therapeutic targets, but traditional molecular docking methods for virtual screening are often computationally expensive. This study introduces PhysDual-GCN, a physics-informed graph neural network designed to approximate docking-derived binding affinity scores for DYRK2, an understudied yet biologically relevant target in Alzheimer’s disease (AD). The model jointly processes ligand molecular graphs and a sequence-based graph representation of DYRK2, while explicitly incorporating Coulomb and Lennard–Jones interaction terms as analytical physical energy components. Because no experimentally measured binding affinities are available for DYRK2-drug pairs, all reference labels used for evaluation were obtained exclusively from widely used classical docking tools (AutoDock Vina, Smina, QVina, CB-DOCK). These tools exhibit an inherent uncertainty of approximately ± 0.5–1.5 kcal/mol, which constrains the interpretability of absolute deviations. PhysDual-GCN was trained solely on docking-derived scores and evaluated using a strict ligand-level separation to avoid circularity during model development. Due to the limited number of ligands (n = 4 FDA-approved AD drugs: brexpiprazole, donepezil, galantamine, rivastigmine), the results should be viewed as agreement with computational references rather than generalizable predictive performance. The model achieved low absolute errors (MAE = 0.31 kcal/mol; RMSE = 0.44 kcal/mol) relative to the reference docking scores and correctly identified stronger binders such as donepezil (− 10.8 kcal/mol) and brexpiprazole (− 10.0 kcal/mol). These findings demonstrate that integrating physical interaction terms into a GNN framework can enhance interpretability while providing a computationally efficient surrogate for classical docking workflows. Overall, PhysDual-GCN offers a biologically meaningful and explainable approximation tool for DYRK2 interaction scoring. While the present results are constrained by the small number of compounds and the absence of 3D protein features, the approach establishes a foundation for future large-scale, experimentally validated studies in AD drug repurposing.

Figures