Optimizing interactions to protein binding sites by integrating docking-scoring strategies into generative AI...

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Lead molecule identification is crucial for drug discovery.
• Artificial intelligence (AI) generative methods aid in designing molecules with specific constraints.
• Current methods often lack direct integration of protein 3D structural information.

Purpose of the Study:

• To incorporate protein 3D structural information directly into generative molecular design.
• To develop an automated structure-based drug design workflow.
• To evaluate AI generative models using flexible docking and a novel scoring function.

Main Methods:

• Derived a protein-ligand scoring function, RFXscore, using the PDBbind database and internal data.
• Employed flexible docking to integrate protein 3D information into generative design.
• Compared different AI generative workflows, including ligand-based and structure-based approaches.

Main Results:

• The proposed structure-based generative design workflow demonstrated promising results, particularly for generating diverse molecules fitting protein binding sites.
• The optimal workflow involved docking followed by the RFXscore function.
• Combining this approach with other metrics, like target-activity machine learning models, enhanced generation of "drug-like" molecules.

Conclusions:

• Directly incorporating protein 3D information via flexible docking and RFXscore significantly improves structure-based generative design.
• The developed workflow is effective for exploring diverse molecular structures tailored to specific protein targets.
• This approach advances automated structure-based drug design, offering a powerful tool for identifying novel drug candidates.