Manifold-constrained nucleus-level denoising diffusion model for structure-based drug design

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery

Background:

• AI models excel at generating high-affinity ligands for drug design.
• Existing models often neglect the physical constraint of minimum atomic distance, leading to collisions.

Purpose of the Study:

• To introduce NucleusDiff, an AI model designed to mitigate atomic collisions in structure-based drug design.
• To improve ligand binding affinity by enforcing physical priors.

Main Methods:

• NucleusDiff enforces spatial distance constraints using auxiliary mesh points around atomic nuclei.
• The model approximates van der Waals boundaries to prevent atomic collisions.
• Evaluation involved the CrossDocked2020 dataset and a COVID-19 therapeutic target.

Main Results:

• NucleusDiff reduced atomic collision rates by up to 100%.
• The model enhanced ligand binding affinity by up to 22.16%.
• Results surpassed state-of-the-art structure-based drug design models.

Conclusions:

• NucleusDiff effectively reduces atomic collisions in AI-driven drug design.
• The method improves binding affinity, offering a significant advancement.
• Qualitative analysis confirmed the model's visual effectiveness in optimizing molecular structures.