Retention prediction in reversed-phase liquid chromatography using XGBoost-based quantitative structure-retention...

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Chromatography

Background:

• Reversed-phase liquid chromatography (RPLC) method development is time-consuming.
• Existing machine learning models for quantitative structure-retention relationships (QSRRs) struggle with stereoisomer discrimination and defining applicability domains.

Purpose of the Study:

• To develop an enhanced XGBoost-based QSRR model for RPLC separations.
• To improve stereoisomer discrimination and prediction accuracy in RPLC.
• To establish a framework for assessing model reliability and applicability domains.

Main Methods:

• Developed an XGBoost QSRR model using 43,329 retention measurements across 13 stationary phases.
• Integrated 2D molecular descriptors with 37 custom stereochemical and geometric descriptors.
• Employed feature selection to reduce the descriptor set to 29.
• Evaluated model performance using root mean square error (RMSE) on a test set.

Main Results:

• Achieved a test-set RMSE of 0.12 on the ln selectivity scale.
• Custom isomer descriptors improved stereoisomer prediction accuracy 2.5-fold.
• Demonstrated reliable interpolation (RMSE ≤ 0.05) within training bounds but degraded performance during extrapolation (RMSE 0.11-0.20).
• Identified column-dependent prediction failures linked to omitted descriptors.

Conclusions:

• The developed QSRR model offers improved prediction accuracy and stereoisomer discrimination for RPLC.
• The framework provides systematic assessment of prediction reliability across different mobile and stationary phases.
• Clear boundaries for trustworthy model application are established, aiding efficient RPLC method development.