Advancing In Silico Drug Design with Bayesian Refinement of AlphaFold Models.

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Virtual screening is key for structure-based drug discovery, relying on accurate protein structures.
• AlphaFold2 provides accurate protein structure predictions but often lacks detail in binding pockets.
• Molecular dynamics simulations refine structures but don't integrate deep learning insights.

Purpose of the Study:

• To develop an integrative method, bAIes, combining physics-based and data-driven approaches.
• To enhance the accuracy of virtual screening using improved protein structural models.
• To overcome limitations of current protein structure prediction and refinement methods.

Main Methods:

• Developed bAIes, integrating physics-based force fields with data-driven predictions via Bayesian inference.
• Utilized AlphaFold2 predictions and molecular dynamics refinement as components.
• Applied bAIes to virtual screening campaigns for binder/non-binder discrimination.

Main Results:

• bAIes significantly outperformed both AlphaFold2 and molecular dynamics-refined models in virtual screening.
• The method demonstrated superior discrimination between binding and non-binding molecules.
• bAIes enhances AlphaFold2 model usability without extensive resources.

Conclusions:

• bAIes offers a novel solution for improving protein structure quality in drug discovery.
• This integrative approach accelerates structure-based drug design by enhancing virtual screening.
• bAIes facilitates more efficient identification of potential drug candidates.