ABSTRACT
In metabolomics, retention prediction methods have been developed based on the structural and physicochemical characteristics of analytes. Such methods employ regression models, harnessing machine learning algorithms mapping experimentally derived retention time (tR) analytes with various structural and physicochemical descriptors, known as Quantitative Structure Retention Relationships (QSRR) models. In the present study, QSRR models have been developed by applying four Machine Learning regression algorithms, i.e. Bayesian Ridge Regression (BRidgeR), Extreme Gradient Boosting Regression (XGBR) and Support Vector Regression (SVR) using both linear and non-linear kernels, all tested and compared for their retention prediction ability on experimentally derived and on publicly available chromatographic data, using Molecular Descriptors to describe the physical, chemical or structural properties of molecules. Various configurations of the available datasets, in terms of the highly-correlated features levels (defined as the maximum absolute value of the Pearson's correlation coefficient calculated between any pair of features) they contained, were analyzed in parallel. This is the first study, to the best of our knowledge, of the effect of collinearity on the performance of QSRR predictive models. In the vast majority of cases studied there was no statistically significant difference in the performance of the generated QSRR predictive models among the specified dataset configurations, indicative of the ability of the selected regression algorithms to effectively handle collinearity. In terms of the individual performance of the selected regression algorithms, no pattern was found where one algorithm (or class of algorithms) stood out significantly relative to the others among the study datasets.
