How accurately can one predict drug binding modes using AlphaFold models?

Area of Science:

• Computational biology
• Structural bioinformatics
• Drug discovery

Background:

• Protein structure prediction is crucial for drug discovery.
• Machine learning, exemplified by AlphaFold 2 (AF2), has advanced protein structure prediction accuracy.
• The utility of AF2 models for predicting drug-protein interactions remains to be fully elucidated.

Purpose of the Study:

• To evaluate the accuracy of AlphaFold 2 models for predicting drug binding poses.
• To assess the performance of AF2 models in drug discovery targeting G-protein-coupled receptors (GPCRs).

Main Methods:

• Comparative analysis of AF2 models versus traditional homology models.
• Assessment of binding pocket accuracy in AF2-predicted protein structures.
• Computational docking simulations to predict ligand-binding poses on AF2 models and experimental structures.

Main Results:

• AF2 models provide significantly more accurate representations of GPCR binding pockets than homology models.
• Despite improved pocket accuracy, computational docking to AF2 models does not yield significantly better ligand-binding pose predictions compared to homology models.
• Docking accuracy to AF2 models is substantially lower than docking to experimentally determined protein structures.

Conclusions:

• AlphaFold 2 excels at predicting protein binding site structures but does not directly translate to improved drug-binding pose prediction accuracy.
• Current computational docking methods may require further development to fully leverage the accuracy of AF2-generated protein models for drug discovery.
• These findings highlight limitations in using predicted protein structures for drug discovery and necessitate careful consideration of model utility.