Dynamic-GLEP: a dynamics-informed deep learning framework for ligand efficacy prediction in representative Class A...

Area of Science:

• Computational chemistry
• Pharmacology
• Structural biology

Background:

• G protein-coupled receptors (GPCRs) are key drug targets, but predicting ligand efficacy, beyond binding affinity, remains challenging.
• Current computational methods often fail to capture the conformational dynamics influencing GPCR functional selectivity.
• Understanding ligand efficacy is crucial for developing targeted therapeutics.

Purpose of the Study:

• To introduce Dynamic-GLEP, a framework integrating molecular dynamics and graph neural networks for predicting ligand efficacy in GPCRs.
• To develop a method that captures conformation-dependent interactions driving functional selectivity.
• To provide a reliable and interpretable platform for drug discovery targeting GPCRs.

Main Methods:

• Utilized molecular dynamics (MD) to generate conformational ensembles of GPCR-ligand complexes.
• Employed transfer learning on equivariant graph neural networks (EquiScore model) to identify interaction features.
• Constructed multi-conformation complexes to distinguish agonists from nonagonists.

Main Results:

• Dynamic-GLEP achieved AUCs of 0.74 (cross-validation) and 0.71 (external FDA dataset) for the 5-HT1A receptor.
• Holo-based models excelled in scaffold optimization, while Apo-derived ensembles showed better adaptability to diverse ligands.
• High performance (AUC > 0.85) was observed for the adenosine A2A receptor, demonstrating robustness and transferability.

Conclusions:

• Dynamic-GLEP offers a reliable and interpretable platform for predicting ligand efficacy in Class A GPCRs.
• The framework effectively integrates conformational dynamics into computational drug discovery.
• Dynamic-GLEP has broad potential for virtual screening, candidate prioritization, and mechanism-driven drug design.