METHODOLOGIES FOR ESTIMATING REPEATABILITY AND REPRODUCIBILITY VARIANCES IN MULTIVARIATE DATABASES

Due to the huge amount of information available from spectra obtained from the analyses of biological samples using spectroscopic analytical techniques such as NMR or MIR/NIR multivariate analysis such as Principal Component Analysis (PCA) are required to understand the influence of major experimental factors. However, many experiments in these studies have more complexes variability structures than simply comparing several treatments: they may include time effects, biological effects such as diet or hormonal status, and other blocking factors or variability sources: samples stability, age of the individuals, pH of a buffer, days of acquisition, and so on. Analysis of these databases needs to extract from the spectral data matrix the variations linked to a change indicated in the factor of interest. However other sources of variability may impair this objective. This stresses the importance to discover the sources of variability of the spectral data using appropriate methodology.
Classically, to analyze such data analysis of variance (ANOVA) or multivariate ANOVA (MANOVA) is used. However direct application of these methodologies to spectrum obtained from structured experimental studies is inappropriate or impossible. More complex data analyses methodologies are required to understand the importance of the various factors implied in the experiments and to provide a measure of their variance components. Three related methodologies have been proposed to achieve this: ANOVA simultaneous component analysis (ASCA), ANOVA-PCA (APCA) and AComDim. The ASCA and APCA methodologies combine first an analysis of variance step (ANOVA) and then a PCA step. The AComDim one adds to the output of the ANOVA-PCA step a multi-block analysis. In addition, an extension of MANOVA is also available called 50-50 MANOVA and Principal variance component analysis (PVCA) has also been proposed.
In this work, the usefulness and applicability of these advanced techniques to data analysis of NMR metabolomic spectra and MIR spectra are given to highlight the increase of knowledge gained and the estimation of main sources of variability arising in an experimental setup. In addition another methodology is proposed which combines PCA and Multivariate linear mixed modeling (PCA-MLMM).