Geometric Structure-Aware Diffusion Model with Self-Optimization Strategy for Molecular Generation

Area of Science:

• Artificial Intelligence
• Computational Chemistry
• Drug Discovery

Background:

• Existing generative models for molecular design struggle with geometric representation, leading to unstable and implausible molecules.
• Current methods often overlook critical quantum and drug-likeness properties essential for effective drug design.
• There is a need for advanced AI approaches that can generate molecules with both structural integrity and desired properties.

Purpose of the Study:

• To develop a novel Geometric Structure-Aware Diffusion Model (MolGD) for accelerated and improved molecular generation and optimization.
• To address limitations in geometric representation and incorporate quantum and drug-likeness properties into the molecular design process.
• To enhance the chemical plausibility, structural stability, and property regulation of generated molecules.

Main Methods:

• Designed a Geometric Structure-Aware Network (GSAN) incorporating a Molecular Graph Attention Network (MGAT) for integrated topological and geometric representation.
• Developed a Geometric Reconstruction Network (GRN) to update atomic positions for stable molecular structures.
• Integrated quantum attributes as conditional constraints and employed reinforcement learning (MolGD-RL) for drug-likeness optimization.

Main Results:

• MolGD demonstrates superior performance over existing methods in generating effective and stable molecules.
• The model successfully generates molecules with specific, desired quantum properties.
• MolGD achieves significant optimization for drug-likeness and synthesizability.

Conclusions:

• MolGD offers a powerful new approach for intelligent molecular design, overcoming key limitations of previous generative models.
• The model's ability to integrate structural, quantum, and drug-likeness properties advances drug discovery pipelines.
• This work provides valuable insights for developing next-generation AI-driven drug design tools.