Fragment-Driven Progressive Alternating Diffusion for De Novo Molecular Design.

Area of Science:

• Artificial Intelligence
• Computational Chemistry
• Drug Discovery

Background:

• AI-driven de novo molecule design aims for high reliability and creativity.
• Current methods face challenges in generating complex molecular structures efficiently.

Purpose of the Study:

• To propose a novel framework for de novo molecule design that improves reliability and creativity.
• To address combinatorial complexities in generating intricate molecular structures.

Main Methods:

• Developed a fragment-driven progressive alternating diffusion (FDPAD) framework.
• Modeled molecules as fragment-structured graphs with a progressive discrete diffusion process.
• Introduced conditioned fragment diffusion (CFD) and inter-fragment bond diffusion (IBD) processes.

Main Results:

• FDPAD mitigates combinatorial complexities by modeling molecules as fragment-structured graphs.
• CFD enhances fragment diversity using partially generated molecules as conditions.
• IBD improves prediction of intricate chemical bond connections among fragments.

Conclusions:

• The proposed FDPAD framework outperforms state-of-the-art algorithms in de novo molecular generation.
• FDPAD excels in generating novel and unique molecules, demonstrating enhanced creativity and reliability.