Diffusion-based generative drug-like molecular editing with chemical natural language.

Area of Science:

• Computational chemistry
• Artificial intelligence in chemistry
• Drug discovery informatics

Background:

• Diffusion models are effective for molecular design but primarily focus on graphs or 3D structures.
• Molecular sequence diffusion models are less explored, despite the linguistic nature of chemical nomenclature.
• International Union of Pure and Applied Chemistry (IUPAC) names represent chemical structures as natural language.

Purpose of the Study:

• To develop a diffusion model capable of translating IUPAC names into Simplified Molecular Input Line Entry System (SMILES) strings.
• To investigate the transferability of pre-trained diffusion model performance to chemical natural language processing.
• To enable controllable molecular editing from chemical natural language to chemical language.

Main Methods:

• Proposed DiffIUPAC, a conditional diffusion model for molecular editing.
• Utilized IUPAC names as input to generate corresponding SMILES strings.
• Evaluated model performance against existing methods and analyzed chemical space and scaffold diversity.

Main Results:

• DiffIUPAC demonstrated superior performance compared to existing methods.
• The model successfully learned and applied semantic rules of both IUPAC and SMILES.
• Generated molecules exhibited scaffold diversity while adhering to input constraints.

Conclusions:

• DiffIUPAC effectively converts chemical natural language (IUPAC) to chemical language (SMILES) using diffusion models.
• The model shows potential for applications in drug design, including functional group editing, analogue generation, and linker design.