Chemical classification program synthesis using generative artificial intelligence.

Area of Science:

• Cheminformatics and Bioinformatics
• Computational Chemistry
• Artificial Intelligence in Chemistry

Background:

• Accurate chemical structure classification is vital for diverse applications, from drug discovery to environmental monitoring.
• Manual classification is inefficient for large datasets, while existing automated methods lack explainability or rely on predefined rules.
• Deep learning models offer high performance but often function as black boxes, hindering interpretability.

Purpose of the Study:

• To develop an AI-driven approach for automatically generating chemical classifier programs.
• To create an explainable and computable ontological model for chemical nomenclature.
• To address the limitations of existing automated chemical classification methods.

Main Methods:

• Utilized generative artificial intelligence to write classifier programs for ChEBI database classes.
• Developed the ChEBI Chemical Class Program Ontology (C3PO) for explainable, deterministic classification of SMILES structures.
• Validated the approach against the ChEBI database and compared it with deep learning and SMARTS-based classifiers.

Main Results:

• The C3PO approach generated programs capable of deterministic run-time classification with natural language explanations.
• C3PO demonstrated superior performance compared to a naive SMARTS pattern classifier.
• While not matching state-of-the-art deep learning performance, C3PO offers significant advantages in explainability and reduced data dependency.

Conclusions:

• The C3PO approach provides an explainable and computable ontological model for chemical classification.
• This method complements deep learning by offering transparency and can be used in tandem for validated classifications.
• C3PO programs can aid ontology development and be refined by human curators for improved accuracy and utility.