Geometry-complete diffusion for 3D molecule generation and optimization.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Generative deep learning, particularly diffusion models with equivariant graph neural networks (GNNs), has advanced 3D molecule generation.
• Existing methods struggle to learn crucial geometric properties, limiting the generation of valid large 3D molecules.

Purpose of the Study:

• To introduce the Geometry-Complete Diffusion Model (GCDM) for improved 3D molecule generation.
• To address the limitations of non-geometric GNNs in learning molecular geometry.
• To enhance the generation of valid and energetically stable large 3D molecules.

Main Methods:

• Developed the Geometry-Complete Diffusion Model (GCDM), a novel generative model for 3D molecules.
• Utilized a denoising diffusion framework incorporating geometric properties.
• Evaluated GCDM on QM9 and GEOM-Drugs datasets for conditional and unconditional generation.

Main Results:

• GCDM significantly outperforms existing 3D molecular diffusion models on both QM9 and GEOM-Drugs datasets.
• GCDM successfully generates a substantial proportion of valid and energetically stable large molecules from the GEOM-Drugs dataset.
• Previous methods failed to generate such valid large molecules due to limitations in learned features.

Conclusions:

• GCDM represents a significant advancement in 3D molecule generation, overcoming geometric learning limitations.
• The model demonstrates versatility, enabling design for specific protein pockets and optimization of existing molecules for stability and properties.
• GCDM showcases the potential of molecular diffusion models for broader applications in chemistry and drug discovery.