Background: Extracting relevant biological information from large data sets is a major challenge in functional genomics research. Different aspects of the data hamper their biological interpretation. For instance, 5000-fold differences in concentration for different metabolites are present in a metabolomics data set, while these differences are not proportional to the biological relevance of these metabolites. However, data analysis methods are not able to make this distinction. Data pretreatment methods can correct for aspects that hinder the biological interpretation of metabolomics data sets by emphasizing the biological information in the data set and thus improving their biological interpretability.

Classifying groups of individuals based on their metabolic profile is one of the main topics in metabolomics research. Due to the low number of individuals compared to the large number of variables, this is not an easy task. PLSDA is one of the data analysis methods used for the classification. Unfortunately this method eagerly overfits the data and rigorous validation is necessary. The validation however is far from straightforward. Is this paper we will discuss a strategy based on cross model validation and permutation testing to validate the classification models. It is also shown that too optimistic results are obtained when the validation is not done properly. Furthermore, we advocate against the use of PLSDA score plots for inference of class differences.

Partial Least Squares-Discriminant Analysis (PLS-DA) is a PLS regression method with a special binary ‘dummy’ y-variable and it is commonly used for classification purposes and biomarker selection in metabolomics studies. Several statistical approaches are currently in use to validate outcomes of PLS-DA analyses e.g. double cross validation procedures or permutation testing. However, there is a great inconsistency in the optimization and the assessment of performance of PLS-DA models due to many different diagnostic statistics currently employed in metabolomics data analyses. In this paper, properties of four diagnostic statistics of PLS-DA, namely the number of misclassifications (NMC), the Area Under the Receiver Operating Characteristic (AUROC), Q2 and Discriminant Q2 (DQ2) are discussed. All four diagnostic statistics are used in the optimization and the performance assessment of PLS-DA models of three different-size metabolomics data sets obtained with two different types of analytical platforms and with different levels of known differences between two groups: control and case groups. Statistical significance of obtained PLS-DA models was evaluated with permutation testing. PLS-DA models obtained with NMC and AUROC are more powerful in detecting very small differences between groups than models obtained with Q2 and Discriminant Q2 (DQ2). Reproducibility of obtained PLS-DA models outcomes, models complexity and permutation test distributions are also investigated to explain this phenomenon. DQ2 and Q2 (in contrary to NMC and AUROC) prefer PLS-DA models with lower complexity and require higher number of permutation tests and submodels to accurately estimate statistical significance of the model performance. NMC and AUROC seem more efficient and more reliable diagnostic statistics and should be recommended in two group discrimination metabolomic studies.Electronic supplementary materialThe online version of this article (doi:10.1007/s11306-011-0330-3) contains supplementary material, which is available to authorized users.

Multiblock and hierarchical PCA and PLS methods have been proposed in the recent literature in order to improve the interpretability of multivariate models. They have been used in cases where the number of variables is large and additional information is available for blocking the variables into conceptually meaningful blocks. In this paper we compare these methods from a theoretical or algorithmic viewpoint using a common notation and illustrate their differences with several case studies. Undesirable properties of some of these methods, such as convergence problems or loss of data information due to deflation procedures, are pointed out and corrected where possible. It is shown that the objective function of the hierarchical PCA and hierarchical PLS methods is not clear and the corresponding algorithms may converge to different solutions depending on the initial guess of the super score. It is also shown that the results of consensus PCA (CPCA) and multiblock PLS (MBPLS) can be calculated from the standard PCA and PLS methods when the same variable scalings are applied for these methods. The standard PCA and PLS methods require less computation and give better estimation of the scores in the case of missing data. It is therefore recommended that in cases where the variables can be separated into meaningful blocks, the standard PCA and PLS methods be used to build the models and then the weights and loadings of the individual blocks and super block and the percentage variation explained in each block be calculated from the results. © 1998 John Wiley & Sons, Ltd.