Feature Selection in metabolomics with PLS-derived methods

Martin, Manon [UCL] Govaerts, Bernadette [UCL]

Current feature selection in metabolomics rely essentially on t-tests, (O)PLS and their sparse counterparts, 60 but their validation as feature selection technique in this research field is lacking as the true biomarkers location is usually unknown. In this work, we use a set of semi-artificial 1H NMR spectra from rat urine where artificial alterations are known id advance to evaluate the features accuracy and stability, among other measurements, of the supervised methods. Several conditions are stressed with different levels of noise in the data, true feature intensities and sample sizes. Results showed that PLS65 vip with a threshold of 1 was the overall best method for our dataset. Nevertheless, all approaches have advantages and drawbacks that should be considered prior to their use for feature selection.