How good are AlphaFold models for docking-based virtual screening?

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• High-quality protein structures are essential for structure-based drug discovery.
• Homology modeling is traditionally used when experimental structures are unavailable.
• AlphaFold (AF) has emerged as a powerful AI tool for predicting protein structures with high accuracy.

Purpose of the Study:

• To evaluate the accuracy of AlphaFold models for docking-based drug discovery.
• To compare the performance of AlphaFold models against experimental protein data bank (PDB) structures in high-throughput docking (HTD).

Main Methods:

• Utilized a benchmark set of 22 protein targets.
• Performed high-throughput docking (HTD) using AlphaFold models and corresponding experimental PDB structures.
• Employed four different docking programs and two consensus techniques for performance evaluation.

Main Results:

• AlphaFold models consistently demonstrated inferior performance in HTD compared to experimental PDB structures.
• The accuracy of AlphaFold in predicting protein architecture does not directly translate to reliable docking performance.
• Four docking programs and two consensus techniques all showed worse performance with AF models.

Conclusions:

• AlphaFold's current capabilities may not be sufficient for reliable docking-based drug discovery without further optimization.
• Post-modeling refinement strategies are crucial for enhancing the utility of AlphaFold models in HTD.
• Further research is needed to bridge the gap between predicted and experimentally validated structures for drug discovery applications.