Comparison study of dominant molecular sequence representation based on diffusion model.

Area of Science:

• Artificial Intelligence in Chemistry
• Computational Drug Discovery
• Machine Learning for Molecular Design

Background:

• Large language models (LLMs) and generative AI are revolutionizing AI for science, particularly in drug discovery.
• Molecular representation learning and generation are key tasks in computational chemistry.
• Existing molecular representations like SMILES, SELFIES, SMARTS, and IUPAC have varying strengths and information granularity.

Purpose of the Study:

• To investigate and compare the characteristics of four mainstream molecular representation languages (SMILES, SELFIES, SMARTS, IUPAC) within a unified diffusion model framework.
• To evaluate the impact of molecular representation choice on generative molecular set training and performance.
• To provide insights for selecting appropriate molecular representations in AI-assisted drug design.

Main Methods:

• Representing single molecules using four distinct methodologies: SMILES, SELFIES, SMARTS, and IUPAC nomenclature.
• Training a denoising diffusion model with identical parameters across all four representations.
• Generating and analyzing thirty thousand molecules for each representation to assess performance metrics.

Main Results:

• The four molecular representations exhibit distinct attribute and spatial distributions, with SELFIES/SMARTS showing high similarity and IUPAC/SMILES showing significant differences.
• IUPAC excels in generating novel and diverse molecules.
• SMILES performs best on QEPPI and SAscore metrics, while SELFIES and SMARTS lead in QED metric performance.

Conclusions:

• The choice of molecular representation significantly impacts generative model performance in AI drug design.
• SELFIES and SMARTS offer a good balance, particularly for the QED metric.
• Understanding these differences is crucial for optimizing AI-driven drug discovery pipelines and enhancing development efficiency.