Wrapper- and Ensemble-Based Feature Subset Selection Methods for Biomarker Discovery in Targeted Metabolomics

摘要

The discovery of markers allowing for accurate classification of metabolically very similar proband groups constitutes a challenging problem. We apply several search heuristics combined with different classifier types to targeted metabolomics data to identify compound subsets that classify plasma samples of insulin sensitive and -resistant subjects, both suffering from non-alcoholic fatty liver disease. Additionally, we integrate these methods into an ensemble and screen selected subsets for common features. We investigate, which methods appear the most suitable for the task, and test feature subsets for robustness and re-producibility. Furthermore, we consider the predictive potential of different compound classes. We find that classifiers fail in discriminating the non-selected data accurately, but benefit considerably from feature subset selection. Especially, a Pareto-based multi-objective genetic algorithm detects highly discriminative subsets and outperforms widely used heuristics. When transferred to new data, feature sets assembled by the ensemble approach show greater robustness than those selected by single methods.