Are Deep Learning Structural Models Sufficiently Accurate for Free-Energy Calculations? Application of FEP+ to...

Area of Science:

• Structural biology
• Computational chemistry
• Drug discovery

Background:

• AlphaFold2 predicts protein structures with high accuracy, exciting the scientific community, especially drug hunters.
• The utility of AlphaFold2 structures for predicting ligand binding sites in drug discovery remains under investigation.

Purpose of the Study:

• To assess the accuracy of AlphaFold2-predicted protein structures for supporting computational drug discovery.
• To determine if AlphaFold2 structures are suitable for physics-based methods like free-energy perturbation (FEP) calculations.

Main Methods:

• Performed FEP calculations on protein-ligand complexes using AlphaFold2-predicted structures.
• AlphaFold2 predictions excluded templates with >30% identity to the target protein.
• Compared FEP results from AlphaFold2 structures with those from crystal structures.

Main Results:

• Free-energy perturbation (FEP) calculations using prospective AlphaFold2 structures yielded ΔΔG values comparable in accuracy to those from crystal structures.
• The accuracy was maintained even when template information was limited during AlphaFold2 prediction.

Conclusions:

• AlphaFold2-modeled protein structures are sufficiently accurate for physics-based computational methods in drug discovery.
• These structures can be reliably used in lead optimization stages when appropriate conditions are met.