MG-DIFF: A novel molecular graph diffusion model for molecular generation and optimization.

Area of Science:

• Artificial Intelligence
• Computational Chemistry
• Drug Discovery

Background:

• Denoising diffusion models have advanced generative tasks in various domains.
• Existing diffusion models for molecule generation lack optimization for molecular features, hindering performance and conditional optimization.
• There is a need for novel approaches to enhance molecular generation and optimization using diffusion models.

Purpose of the Study:

• To introduce MG-DIFF, a novel diffusion model specifically designed for molecular generation and optimization.
• To address the limitations of existing models in capturing complex molecular structures and enabling conditional optimization.
• To improve the quality and efficiency of generating and optimizing molecules.

Main Methods:

• Proposed a mask and replace discrete diffusion strategy for enhanced molecular structure representation.
• Introduced a graph transformer model with random node initialization to overcome limitations of traditional graph neural networks.
• Developed a graph padding strategy for conditional generation and molecule optimization via atomic group addition.

Main Results:

• MG-DIFF achieved state-of-the-art performance on several molecular generation benchmarks.
• Demonstrated significant potential for conditional molecular optimization tasks.
• The proposed strategies effectively improved the quality of generated molecules and the capabilities of optimization.

Conclusions:

• MG-DIFF represents a significant advancement in diffusion-based molecular generation and optimization.
• The model's novel strategies offer improved performance and flexibility compared to previous methods.
• MG-DIFF shows great promise for applications in drug discovery and materials science.