Improving AlphaFold2 Performance in Virtual Screens Targeting GPCRs by Enhancing Binding-Site Conformational...

Area of Science:

• Computational biology
• Structural biology
• Drug discovery

Background:

• Artificial intelligence, specifically AlphaFold2 (AF2), has revolutionized protein structure prediction, achieving near-experimental accuracy.
• However, AF2's focus on single-model generation limits its application in structure-based drug design due to its inability to capture protein conformational flexibility.

Purpose of the Study:

• To develop strategies for generating diverse ensembles of protein binding-site models for structure-based virtual screening.
• To enhance drug discovery efforts targeting G protein-coupled receptors (GPCRs) by addressing protein conformational flexibility.

Main Methods:

• Introduced AFsample2T, a novel approach utilizing multiple sequence alignment column masking within the receptor binding site.
• This masking technique reduces coevolutionary signals, promoting greater structural heterogeneity in the generated protein models.

Main Results:

• AFsample2T successfully generated ensembles of models that capture multiple relevant binding-site conformations.
• These ensembles reproduced experimentally observed conformational variability and significantly improved ligand enrichment in structure-based virtual screening when compared to single-model approaches.
• Docking simulations using diverse AF2-based models demonstrated enhanced performance in identifying active ligands over decoys.

Conclusions:

• Ensembles of diverse binding-site models generated using AFsample2T offer substantial improvements for structure-based virtual screening, particularly for GPCR targets.
• The AFsample2T approach provides valuable guidelines for utilizing AF2-based models in structure-based ligand discovery and is adaptable to other protein families.