KGDiff: towards explainable target-aware molecule generation with knowledge guidance.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Designing molecules with high binding affinity to protein targets is essential for drug discovery.
• Existing target-aware methods often overlook binding affinity, limiting their effectiveness.
• Accurately modeling 3D atomic interactions between molecules and proteins remains a challenge.

Purpose of the Study:

• To develop an explainable diffusion model for generating 3D molecules with high protein-ligand binding affinity.
• To explicitly incorporate chemical knowledge of binding affinity into the molecule generation process.
• To enhance the performance of target-aware molecule design methods.

Main Methods:

• An explainable diffusion model was proposed for molecule generation.
• An SE(3)-invariant expert network was developed to integrate Vina scoring functions, distilling chemical domain knowledge.
• Guidance for atom coordinates and types was achieved using the expert network's gradient.

Main Results:

• The proposed method demonstrated superior performance on the CrossDocked2020 benchmark dataset.
• The model successfully generated molecules with high binding affinity to target proteins.
• Atom-level explanations for generated molecules were provided, linking them to domain knowledge.

Conclusions:

• The explainable diffusion model effectively generates high-affinity molecules by incorporating chemical binding knowledge.
• This approach improves upon existing target-aware methods by explicitly considering binding affinity.
• The method offers atom-level insights, enhancing the interpretability of AI-driven drug design.