Global and local integrated gradient-based diffusion model for de novo drug design

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Deep learning models are crucial for navigating chemical space in de novo drug design.
• Diffusion models show promise for generating target-binding molecules but face challenges in multi-objective optimization and computational cost.
• Existing methods struggle to balance binding affinity and drug-like properties efficiently.

Purpose of the Study:

• To develop a novel diffusion-based model for efficient and simultaneous optimization of binding affinity and drug-like properties.
• To address the limitations of slow denoising processes and computational expense in current generative models.
• To generate high-quality, stable drug candidates with optimal binding positions.

Main Methods:

• Introduction of the Global and local integrated gradient-based Diffusion Model (GlintDM).
• Implementation of a significantly faster denoising process using skip transitions, leveraging global and local gradients.
• Integration of position refinement, candidate evaluation, and ligand resampling phases within the generation process.

Main Results:

• GlintDM demonstrates superior performance on the CrossDocked and Binding MOAD datasets compared to existing methods, evidenced by Vina-related scores.
• The model successfully identifies optimal binding positions for target proteins.
• Generated molecules satisfy multi-objective molecular properties, confirmed by assessments of steric clash and geometric properties.

Conclusions:

• GlintDM offers a computationally efficient and effective approach for de novo drug design.
• The model generates stable, high-quality molecules with optimized binding and desirable drug-like properties.
• GlintDM advances the application of diffusion models in generating molecules for drug discovery.