Combining strong sparsity and competitive predictive power with the L-sOPLS approach for biomarker discovery in metabolomics

In the context of metabolomics analyses, Partial Least Squares (PLS) represents the standard tool to perform regression and classification. OPLS, the Orthogonal extension of PLS which has proved to be very useful when interpretation is the main issue, is a more recent way to decompose the PLS solution into predictive components correlated to the target Y and components pertaining to the data X but uncorrelated to Y . This predominance of (O)PLS can raise the question of the awareness of alternative multivariate regression and/or classification tools able to find biomarkers. Actually, the search for biomarkers remains a key issue in metabolomics as it is crucial to very accurately target discriminating features. Most of the time, (O)PLS methods perform well but a drawback often occurs: too many variables can be selected as potential biomarkers even using adapted statistical significance tests. However, for final users (in medical studies for instance), it can be advantageous to deal with only a small number of easily interpretable biomarkers. This drawback is approached in this paper via the use of sparse methods. The sparse-PLS (sPLS), an extension of PLS which promotes an inner variable/feature selection, is an interesting existing solution. But a new intuitive algorithm is proposed in this paper to combine sparsity and the advantages of an orthogonalization step: the "Light-sparse-OPLS" (L-sOPLS). L-sOPLS promotes sparsity on a previously optimized deated matrix which implies the removal of the Y -orthogonal components. A discussion around the compromise between sparsity and predictive modelling performances is provided and it is shown that L-sOPLS produces convincing results, illustrated principally on the basis of 1HNMR spectral data but also on genomic RTq-PCR data. The L-sOPLS algorithm allows to reach better predictive performances than (O)PLS and sPLS while taking into account only a very small number of relevant descriptors