Evaluation of AlphaFold2 Structures for Hit Identification across Multiple Scenarios

Area of Science:

• Computational Chemistry and Molecular Modeling
• Drug Discovery and Development
• Structural Bioinformatics

Background:

• AlphaFold2 (AF2) has revolutionized protein structure prediction, impacting drug discovery research.
• Accurate protein structures are crucial for effective virtual screening (VS) in drug design.
• Previous studies on AF2 structures in VS require further comprehensive evaluation.

Purpose of the Study:

• To evaluate the performance of AF2 structures in virtual screening (VS) across various drug discovery scenarios.
• To compare VS performance using AF2 structures against holo and apo structures.
• To assess the utility of AF2 structures when only they are available for a target.

Main Methods:

• Evaluated AF2, apo, and holo structures for VS in three distinct scenarios.
• Employed physics-based Glide and deep-learning RTMscore for compound ranking.
• Utilized DUD-E, DEKOIS 2.0, and DECOY datasets for validation.

Main Results:

• VS performance on AF2 structures is comparable to apo but inferior to holo structures.
• When only AF2 structures are available, using holo structures from related protein subtypes yields better results than AF2 structures alone.
• Differences in VS performance are not attributable to variations in biological assembly composition between AF2 and holo structures.

Conclusions:

• Docking-based VS using solely AF2 structures may not achieve optimal outcomes in drug discovery.
• Alternative strategies, such as utilizing homologous holo structures, should be considered when AF2 structures are the only option.
• The study highlights the continued importance of high-quality experimental structures for reliable virtual screening.