Unsupervised Hierarchical Symbolic Regression for Interpretable Property Modeling in Complex Multi-Variable Systems.

Area of Science:

• Artificial Intelligence
• Cheminformatics
• Analytical Chemistry

Background:

• AI models excel at chemical analysis prediction but often lack interpretability.
• Thin-layer chromatography (TLC) is vital for analyzing molecular polarity.
• Explainable AI is needed to build trust in predictive chemical models.

Purpose of the Study:

• Introduce Unsupervised Hierarchical Symbolic Regression (UHSR) as an interpretable AI solution.
• Develop a model that maintains competitive predictive performance.
• Demonstrate UHSR's ability to derive chemically intuitive insights.

Main Methods:

• UHSR automatically distills retention indices from TLC data.
• UHSR discovers explainable equations linking molecular structures to chromatographic behavior.
• The model's adaptability to other property prediction tasks was assessed.

Main Results:

• UHSR successfully derived concise and accurate equations for polarity prediction from TLC data.
• Expert chemists expressed greater trust in UHSR compared to traditional models.
• The method showed adaptability beyond molecular polarity prediction.

Conclusions:

• UHSR provides a powerful and interpretable alternative for chemical predictive modeling.
• Explainable AI in chemistry can enhance model trust and utility.
• UHSR has broad applicability in cheminformatics and analytical chemistry.