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

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Drug Discovery
• Neuroscience

Background:

• Alzheimer's disease (AD) necessitates novel therapeutic targets and efficient screening methods.
• Traditional molecular docking for virtual screening is computationally intensive.
• DYRK2 is an understudied but biologically relevant target in Alzheimer's disease.

Purpose of the Study:

• To introduce PhysDual-GCN, a physics-informed graph neural network (GNN) for approximating binding affinity scores.
• To serve as a computationally efficient surrogate for traditional molecular docking.
• To provide a biologically meaningful and explainable tool for DYRK2 interaction scoring in Alzheimer's disease research.

Main Methods:

• Developed PhysDual-GCN, a GNN integrating ligand molecular graphs and DYRK2 sequence-based graphs.
• Incorporated Coulomb and Lennard-Jones interaction terms as analytical physical energy components.
• Trained and evaluated the model using docking-derived scores from established tools (AutoDock Vina, Smina, QVina, CB-DOCK).

Main Results:

• Achieved low absolute errors (MAE = 0.31 kcal/mol, RMSE = 0.44 kcal/mol) compared to reference docking scores.
• Successfully identified potent binders like donepezil (-10.8 kcal/mol) and brexpiprazole (-10.0 kcal/mol).
• Demonstrated agreement with computational references, despite limitations due to a small ligand set (4 FDA-approved AD drugs).

Conclusions:

• Integrating physical interaction terms into GNNs enhances interpretability and computational efficiency for drug discovery.
• PhysDual-GCN offers a viable, explainable approximation for classical docking workflows targeting DYRK2.
• The approach lays the groundwork for future large-scale, experimentally validated studies in Alzheimer's disease drug repurposing.