Learning Joint 2-D and 3-D Graph Diffusion Models for Complete Molecule Generation.

Area of Science:

• Computational chemistry
• Deep learning for drug discovery
• Molecular modeling

Background:

• Deep generative models advance drug discovery and material science by generating novel molecules.
• Current models often focus on either 2-D bonding graphs or 3-D geometries, limiting generation quality.
• Jointly modeling both 2-D and 3-D molecular representations is crucial for improved molecular design.

Purpose of the Study:

• To propose a novel deep generative model, JODO, for joint 2-D and 3-D molecular graph generation.
• To develop a diffusion graph transformer (DGT) to effectively capture correlations between 2-D and 3-D molecular data.
• To enable inverse molecular design targeting specific quantum properties.

Main Methods:

• Developed a joint 2-D and 3-D graph diffusion model (JODO) for generating complete molecular representations including atom types, charges, bonds, and 3-D coordinates.
• Introduced a diffusion graph transformer (DGT) with a relational attention mechanism to model dependencies between 2-D graphs and 3-D geometries.
• Enabled concurrent propagation and update of scalar attributes and geometric vectors within the DGT.

Main Results:

• JODO significantly outperforms baseline models in unconditional joint generation on QM9 and GEOM-Drugs datasets.
• The model demonstrates strong performance in few-step fast sampling for molecular generation.
• JODO shows effectiveness in inverse molecular design and molecular graph generation tasks.

Conclusions:

• JODO represents a significant advancement in generative models for molecular design by integrating 2-D and 3-D information.
• The proposed diffusion graph transformer effectively captures inter-dependencies crucial for high-fidelity molecular generation.
• JODO offers a versatile platform for both unconditional and conditional (inverse) molecular design, accelerating drug discovery and material science.