AI-guided competitive docking for virtual screening and compound efficacy prediction.

Area of Science:

• Computational biology
• Drug discovery
• Structural bioinformatics

Background:

• Machine learning (ML) has advanced protein structure and interaction prediction.
• The application of ML in drug discovery is an evolving field.
• Accurate prediction of protein-ligand interactions is crucial for identifying potential drug candidates.

Purpose of the Study:

• To evaluate the efficacy of denoise diffusion-based co-folding methods for protein-ligand interaction prediction.
• To introduce and validate a novel strategy, pairwise competitive docking, for ranking candidate molecules in drug discovery.
• To demonstrate the potential of ML in accelerating structure-based drug design.

Main Methods:

• Utilized denoise diffusion-based co-folding methods (e.g., AlphaFold3, Boltz-1/2) for predicting protein-ligand interactions.
• Developed and applied pairwise competitive docking to rank molecules based on relative binding affinity.
• Validated the method across 17 diverse protein benchmark systems.

Main Results:

• Denoise diffusion models achieved high accuracy in predicting protein-ligand interactions and distinguishing active from inactive compounds.
• Pairwise competitive docking generated rankings consistent with experimental trends, with concordance indices ranging from 0.52 to 0.89.
• The method demonstrated strong agreement with existing affinity prediction tools (Boltz-2) and accelerated hit identification in large chemical libraries.

Conclusions:

• Modern ML models, including pairwise competitive docking, offer a faster, more reliable, and cost-effective approach to structure-based drug design.
• The developed method serves as a practical alternative for prioritizing potential inhibitors.
• Pairwise competitive docking can guide the de novo design of potent inhibitors, improving drug discovery workflows.