Performance comparison of nonlinear and linear regression algorithms coupled with different attribute selection...

Area of Science:

• Analytical Chemistry
• Chromatography
• Chemometrics

Background:

• Micellar liquid chromatography (MLC) involves surfactant addition to the mobile phase, altering solubilizing capacity and stationary phase properties.
• Predicting analyte retention in MLC is complex due to these system changes.
• Mixed Quantitative Structure-Retention Relationships (QSRR) offer a powerful approach for estimating analyte retention.

Purpose of the Study:

• To compare 48 mixed QSRR models for predicting the retention of aripiprazole and its impurities in MLC.
• To evaluate the predictive ability of models based on molecular structures and Brij-acetonitrile system factors.
• To identify the most effective regression algorithms and influential molecular descriptors for MLC retention prediction.

Main Methods:

• Developed 48 mixed QSRR models by combining six feature selection methods (PCA, NMF, ReliefF, MLR, Mutual Info, F-Regression) with eight regression algorithms (LR, Ridge, Lasso, ANN, SVR, RF, GBT, k-NN).
• Optimized model hyper-parameters and utilized a dataset of 78 cases generated from 13 experiments per analyte under varying Brij L23 concentration, pH, and acetonitrile content (Box-Behnken design).
• Incorporated 27 molecular descriptors (physicochemical, quantum chemical, topological, spatial) alongside chromatographic parameters as independent variables.

Main Results:

• Gradient Boosted Trees (GBT)-based models demonstrated the highest accuracy in predicting retention factors in MLC mode.
• Different regression algorithms exhibited significant variations in their pattern-learning capabilities, emphasizing the need for extensive algorithm testing.
• Steric factors and dipole-dipole interactions were identified as key contributors to the observed retention behavior.

Conclusions:

• Mixed QSRR modeling, particularly with GBT algorithms, provides a robust method for predicting analyte retention in micellar liquid chromatography.
• The choice of regression algorithm significantly impacts model performance, more so than minor variations in input variables.
• This study serves as a promising foundation for developing comprehensive MLC retention prediction strategies.