AUTHOR=Bujak Renata , Daghir-Wojtkowiak Emilia , Kaliszan Roman , Markuszewski Michał J. 

TITLE=PLS-Based and Regularization-Based Methods for the Selection of Relevant Variables in Non-targeted Metabolomics Data

JOURNAL=Frontiers in Molecular Biosciences

VOLUME=3

YEAR=2016

URL=https://www.frontiersin.org/journals/molecular-biosciences/articles/10.3389/fmolb.2016.00035

DOI=10.3389/fmolb.2016.00035

ISSN=2296-889X

ABSTRACT=<p>Non-targeted metabolomics constitutes a part of the systems biology and aims at determining numerous metabolites in complex biological samples. Datasets obtained in the non-targeted metabolomics studies are high-dimensional due to sensitivity of mass spectrometry-based detection methods as well as complexity of biological matrices. Therefore, a proper selection of variables which contribute into group classification is a crucial step, especially in metabolomics studies which are focused on searching for disease biomarker candidates. In the present study, three different statistical approaches were tested using two metabolomics datasets (<italic>RH</italic> and <italic>PH study</italic>). The orthogonal projections to latent structures-discriminant analysis (OPLS-DA) without and with multiple testing correction as well as the least absolute shrinkage and selection operator (LASSO) with bootstrapping, were tested and compared. For the <italic>RH study</italic>, OPLS-DA model built without multiple testing correction selected 46 and 218 variables based on the VIP criteria using Pareto and UV scaling, respectively. For the <italic>PH study</italic>, 217 and 320 variables were selected based on the VIP criteria using Pareto and UV scaling, respectively. In the <italic>RH study</italic>, OPLS-DA model built after correcting for multiple testing, selected 4 and 19 variables as in terms of Pareto and UV scaling, respectively. For the <italic>PH study</italic>, 14 and 18 variables were selected based on the VIP criteria in terms of Pareto and UV scaling, respectively. In the <italic>RH</italic> and <italic>PH study</italic>, the LASSO selected 14 and 4 variables with reproducibility between 99.3 and 100%, respectively. In the light of PLS-based models, the larger the search space the higher the probability of developing models that fit the training data well with simultaneous poor predictive performance on the validation set. The LASSO offers potential improvements over standard linear regression due to the presence of the constrain, which promotes sparse solutions. This paper is the first one to date utilizing the LASSO penalized logistic regression in untargeted metabolomics studies.</p>