Equivariant 3D-Conditional Diffusion Model for De Novo Drug Design.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Artificial Intelligence

Background:

• De novo drug design accelerates discovery but faces challenges with 3D protein structures and physical consistency.
• Existing methods often underutilize target protein geometry or generate molecules in non-physical orders.

Purpose of the Study:

• To develop an advanced computational method for de novo drug design.
• To improve the utilization of 3D geometric information from target proteins.
• To generate novel pharmaceutical compounds efficiently and accurately.

Main Methods:

• Proposed an equivariant 3D-conditional diffusion model named DiffFBDD.
• Integrated full atomic information of protein pockets using an equivariant graph neural network.
• Developed a diffusion approach for generating ligand fragments tailored to specific protein pockets.

Main Results:

• DiffFBDD demonstrated superior performance in generating ligands with strong binding affinity compared to state-of-the-art models.
• Achieved significant reductions in computational resources and generation time (65.98%–96.10% lower).
• Maintained high validity, uniqueness, and novelty in generated compounds.

Conclusions:

• DiffFBDD effectively leverages 3D geometric information for efficient de novo drug design.
• The model shows significant potential for exploring drug-like chemical space and accelerating pharmaceutical development.