A reference-guided large language model workflow for mobile phase selection in thin-layer chromatography enabled by a...

Area of Science:

• Analytical Chemistry
• Chromatography
• Artificial Intelligence

Background:

• Thin-layer chromatography (TLC) mobile phase selection often relies on empirical trial-and-error.
• Current machine learning methods for TLC lack predictive performance due to manual descriptor engineering and data sensitivity.

Purpose of the Study:

• To propose a novel reference-guided large language model (LLM) workflow for recommending TLC mobile phases.
• To enhance the efficiency and accuracy of TLC method development.

Main Methods:

• A polarity-space tetrahedron strategy was developed using three polarity descriptors (molecular refractivity, topological polar surface area, n-octanol-water partition coefficient) to select reference compounds.
• An LLM integrated structural similarity, polarity descriptors, and analogical inference for mobile phase recommendation.
• A similarity-weighted tetrahedron strategy was compared against other reference selection methods.

Main Results:

• The similarity-weighted tetrahedron method demonstrated the best performance, achieving 81.48% availability and a score of 0.8436.
• Independent experimental validation confirmed a 92.16% success rate for the LLM-based recommendations.
• The framework provided interpretable outputs, including rationales and experimental suggestions.

Conclusions:

• Combining a locally constrained reference construction strategy with an LLM offers a practical and interpretable tool for TLC mobile phase optimization.
• This approach overcomes limitations of traditional methods and large labeled datasets, enabling efficient method development without task-specific model training.