Designing high-affinity 3D drug molecules via geometric spatial perception diffusion model.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Designing high-affinity molecules for protein targets is crucial for drug discovery.
• Current 3D molecular design methods struggle to accurately represent ligand position in Euclidean space.
• Understanding atomic interactions in 3D is key to improving molecular design.

Purpose of the Study:

• To develop a diffusion model that enhances molecular binding affinity to protein targets.
• To improve the representation of ligand molecular position in 3D space.
• To design molecules with improved binding affinity and drug-like properties.

Main Methods:

• Utilized a diffusion model based on SE(3)-equivariant graph neural networks.
• Incorporated a long-range and distance-aware attention head mix for enhanced affinity.
• Implemented a molecular geometry feature enhancement strategy to improve spatial perception.

Main Results:

• The proposed model outperforms state-of-the-art methods on the CrossDocked2020 dataset across various affinity metrics.
• Achieved superior performance in designing ligand molecules with macrocyclic structures.
• Preserved essential drug-like properties and offered moderate interpretability of binding interactions.

Conclusions:

• The novel diffusion model significantly enhances the design of high-affinity molecules for drug discovery.
• The model's ability to capture 3D spatial information improves molecular design accuracy.
• This approach offers a promising direction for developing effective therapeutics with interpretable binding mechanisms.