MGDM: Molecular generation using a multinomial diffusion model.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• De novo drug design faces challenges in accurate molecular structure analysis and rapid valid molecule generation.
• Existing methods struggle to efficiently produce diverse and novel molecular structures.

Purpose of the Study:

• To introduce the Multinomial Generated Diffusion Model (MGDM) for advanced molecular generation in drug design.
• To demonstrate MGDM's capability for both unconditional and conditional molecular generation.
• To evaluate MGDM's performance against state-of-the-art methods in terms of validity, novelty, and diversity.

Main Methods:

• Developed a novel Multinomial Generated Diffusion Model (MGDM) utilizing a multinomial diffusion framework for discrete data.
• Implemented a classifier-free guidance strategy for efficient conditional molecular generation.
• Validated the model using the Molecular Sets (MOSES) dataset, assessing unconditional and conditional generation capabilities.

Main Results:

• MGDM successfully generates valid molecular structures by progressively denoising from a uniform noise distribution.
• The model demonstrated strong performance in unconditional generation, expanding the compound library.
• Conditional generation experiments showed effective property-specific molecule synthesis.
• MGDM achieved superior or comparable novelty and diversity compared to existing state-of-the-art methods on the MOSES dataset.

Conclusions:

• The proposed MGDM is an effective framework for de novo molecular generation, addressing key challenges in drug design.
• MGDM offers a powerful approach for generating diverse, novel, and valid molecules with potential applications in drug discovery.
• The classifier-free guidance strategy enhances the model's utility for targeted molecular design.