Model selection for within-batch effect correction in UPLC-MS metabolomics using quality control - Support vector regression

Abstract

Ultra performance liquid chromatography – mass spectrometry (UPLC-MS) is increasingly being used for untargeted metabolomics in biomedical research. Complex matrices and a large number of samples per analytical batch lead to gradual changes in the instrumental response (i.e. within-batch effects) that reduce the repeatability and reproducibility and limit the power to detect biological responses. A strategy for within-batch effect correction based on the use of quality control (QC) samples and Support Vector Regression (QC-SVRC) with a radial basis function kernel was recently proposed. QC-SVRC requires the optimization of three hyperparameters that determine the accuracy of the within-batch effects elimination: the tolerance threshold (ε), the penalty term (C) and the kernel width (γ). This work compares three widely used strategies for QC-SVRC hyperparameter optimization (grid search, random search and particle swarm optimization) using a UPLC-MS data set containing 193 urine injections as model example. Results show that QC-SVRC is robust to hyperparameter selection and that a pre-selection of C and ε, followed by optimization of γ is competitive in terms of accuracy, precision and number of function evaluations with full grid analysis, random search and particle swarm optimization. The QC-SVRC optimization procedure can be regarded as a useful non-parametric tool for efficiently complementing alternative approaches such as QC-robust splines correction (RSC).