Disentangled diffusion model for 3D molecular generation with protein-ligand interaction priors.

Area of Science:

• Computational chemistry and drug discovery.

Background:

• Structure-based drug design (SBDD) is crucial for identifying drug candidates.
• Existing diffusion models struggle to integrate dynamic protein-ligand interactions, limiting molecular diversity and binding affinity.
• Capturing the dynamic interplay between protein pockets and ligand atoms is essential for high binding affinity.

Purpose of the Study:

• To develop a diffusion model that effectively incorporates dynamic protein-ligand interaction priors for 3D molecular generation.
• To improve molecular diversity and binding affinity in structure-based drug design.

Main Methods:

• Introduced DPDiff, a Disentangled Prior-Conditioned Diffusion model.
• Utilized two complementary interaction prior networks for geometry-based and sequence-based interactions.
• Employed dynamic extraction and adaptive fusion of interaction priors during generation.
• Implemented a disentangled denoising network to balance prior guidance and generative flexibility.

Main Results:

• DPDiff generated molecules with realistic 3D structures and state-of-the-art binding affinities.
• Achieved an average Vina Dock score of -8.58 and a 69.4% high affinity ratio.
• Outperformed existing methods in binding affinity while maintaining drug-likeness and synthetic accessibility.

Conclusions:

• DPDiff offers a significant advancement in protein-specific 3D molecular generation for drug discovery.
• The model's ability to dynamically integrate interaction priors leads to improved molecular design outcomes.
• DPDiff provides a flexible and effective framework for generating high-affinity drug candidates.