Retention prediction in reversed-phase liquid chromatography using XGBoost-based quantitative structure-retention
Ronald Carrasquillo-Flores1, Sarah C Rutan2, Trevor Kempen3
1Chemical Process Development, Bristol Myers Squibb, 1 Squibb Dr., New Brunswick, NJ 08903, USA.
Abstract:
Reversed-phase liquid chromatography (RPLC) method development remains resource-intensive due to the iterative optimization required to achieve fit-for-purpose separations. While machine learning-enabled quantitative structure-retention relationships (QSRRs) offer predictive capabilities, current models suffer from poor stereoisomer discrimination, uncertain applicability domains, and limited practical integration. This work develops an XGBoost-based QSRR model using 43,329 retention measurements from 86 solutes on 13 stationary phases measured with isocratic mobile phases in the range of 2-90% organic content. The model combines standard 2D Mordred molecular descriptors with 37 custom stereochemical and geometric descriptors that distinguish R/S and E/Z configurations from SMILES strings. Feature selection reduced the descriptor set to 29, achieving a test-set root mean square error (RMSE) of 0.12 (ln selectivity scale), with the addition of custom isomer descriptors providing 2.5-fold improvement in stereoisomer prediction accuracy. The model demonstrated reliable interpolation (RMSE ≤ 0.05, selectivity scale) when molecular descriptor values are within the training-set bounds but showed degraded performance during extrapolation (RMSE 0.11-0.20, selectivity scale). Paired-solute analysis across the 13 stationary phases linked column-dependent prediction failures to descriptors omitted or pruned from the model, (e.g., acid-base character, halogen content, H-bond capacity). This computationally accessible framework provides systematic approaches for assessing prediction reliability across mobile phase compositions and stationary phase chemistries, establishing clear boundaries for trustworthy model application.
Related Concept Videos
Silica Gel Column Chromatography: Overview
Polar components tend to bind strongly to the silica gel, causing them to move slowly through the column. In contrast, nonpolar compounds...
Chromatographic Methods: Terminology
High-Performance Liquid Chromatography: Introduction
In HPLC, two phases play a critical role in the separation process:
Size-Exclusion Chromatography
Silica particles offer advantages such as rigidity,...
Residual Plots
When the residual values are plotted against the variable x, it is called a residual...
Thin-Layer Chromatography (TLC): Overview
To begin the analysis, a mixture of compounds is spotted on the starting line on the TLC plate using a thin capillary. The bottom of the...
