Unlocking the application potential of AlphaFold3-like approaches in virtual screening

Area of Science:

• Computational Biology
• Structural Biology
• Drug Discovery

Background:

• Protein-ligand complex structure prediction has been revolutionized by AlphaFold3 (AF3).
• The application of AF3-like models for structure-based virtual screening (VS) is an emerging area requiring systematic evaluation.
• Existing virtual screening methods often face limitations in accuracy and efficiency.

Purpose of the Study:

• To systematically assess the efficacy of AlphaFold3-like approaches for structure-based virtual screening.
• To compare the performance of AF3, Protenix, and Boltz-2 in virtual screening tasks.
• To identify the strengths and limitations of these advanced models in drug discovery pipelines.

Main Methods:

• Utilized AlphaFold3, Protenix, and Boltz-2 as representative models for virtual screening.
• Benchmarked performance on established datasets like DEKOIS2.0 and challenging custom datasets.
• Evaluated intrinsic confidence metrics and third-party scoring schemes for compound ranking.
• Assessed pose generation accuracy and conformational plausibility of predicted structures.

Main Results:

• AlphaFold3 demonstrated exceptional screening capability, primarily driven by its intrinsic confidence metrics.
• Both AF3 and Protenix served as robust pose generators, outperforming conventional docking tools in most scenarios.
• Performance declined in challenging cases, including chemically similar actives exclusion, novel GPCR datasets, and experimentally verified inactive compounds.
• Predicted poses generally adopted physically plausible conformations with minor structural artifacts.

Conclusions:

• AF3-like methods show significant promise for structure-based virtual screening in drug discovery.
• Intrinsic confidence metrics are key drivers of AF3's screening performance.
• Current limitations necessitate careful consideration for deployment in complex drug discovery scenarios.
• These models offer valuable insights and often surpass traditional docking accuracy, despite current constraints.