Structure-based drug design with equivariant diffusion models.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Structure-based drug design (SBDD) traditionally focuses on creating molecules with high affinity for specific protein targets.
• Current generative SBDD methods often require task-specific models, necessitating extensive data curation and retraining for each application.
• This limits the adaptability and efficiency of computational approaches in drug development.

Purpose of the Study:

• To introduce a versatile, single pretrained diffusion model for a wide array of SBDD tasks.
• To demonstrate the application of this model for property optimization, explicit negative design, and inpainting-based molecular design.
• To present DiffSBDD, an SE(3)-equivariant diffusion model for conditional ligand generation based on protein pocket structures.

Main Methods:

• Formulating SBDD as a three-dimensional conditional generation problem.
• Developing DiffSBDD, a diffusion model with SE(3)-equivariance for generating ligands conditioned on protein pockets.
• Incorporating additional constraints to refine generated drug candidates based on computational metrics.

Main Results:

• A single pretrained diffusion model successfully addresses multiple SBDD challenges, including property optimization and negative design.
• DiffSBDD generates novel ligands conditioned on protein pockets, demonstrating its capability in de novo design.
• The model's performance can be enhanced by applying specific constraints, leading to improved drug candidates.

Conclusions:

• DiffSBDD offers a unified and flexible framework for structure-based drug design, overcoming limitations of task-specific models.
• This approach streamlines the generation of optimized drug candidates by leveraging a single, adaptable AI model.
• The findings pave the way for more efficient and broader applications of generative AI in accelerating drug discovery.