DiffMC-Gen: A Dual Denoising Diffusion Model for Multi-Conditional Molecular Generation.

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery

Background:

• Designing drug molecules with specific properties is challenging.
• Deep learning, particularly denoising diffusion models, shows promise for molecular generation.
• Existing methods struggle with multi-property optimization due to limitations in capturing molecular geometry.

Purpose of the Study:

• To develop a novel deep learning model for multi-conditional molecular generation.
• To enhance the simultaneous optimization of multiple drug properties.
• To improve the perception of 3D molecular structures in de novo drug design.

Main Methods:

• Developed a dual denoising diffusion model (DiffMC-Gen) integrating discrete and continuous features.
• Employed a multi-objective optimization strategy for properties like binding affinity, drug-likeness, synthesizability, and toxicity.
• Utilized graph neural networks capable of capturing both topological and geometric information.

Main Results:

• DiffMC-Gen achieved state-of-the-art novelty and uniqueness in both 2D and 3D molecular generation.
• Generated molecules demonstrated comparable drug-likeness and synthesizability to existing methods.
• Predicted molecules showed favorable biological activity and druglike properties for LRRK2, HPK1, and GLP-1 receptor targets.

Conclusions:

• DiffMC-Gen effectively addresses limitations in multi-property drug molecule optimization.
• The model's ability to perceive 3D structures enhances molecular design.
• DiffMC-Gen shows significant potential for practical applications in accelerating drug discovery.