Statistical contributions to the analysis of 2D NMR spectra in metabolomics studies : from pre-processing workflows to 2D biomarker discovery
(2019)
Files
MAIN3.pdf
Open Access - Adobe PDF
- 6.72 MB
Details
- Authors
- Supervisors
- Abstract
- Along with Mass spectrometry data, metabolomics studies are commonly based on proton 1D 1H-NMR spectral multivariate data sources, typically characterized by a high dimensionality. Very powerful and complex pre-processing and analytical tools are adapted to these data, well-mastered and have demonstrated their potential. Nevertheless, when the studied medium is very rich in metabolites, some limitations quickly occur: spectral peaks overlapping, multiplicity of peaks for a same product, difficulty of peak interpretation… 2D spectral data sources (as COSY, for COrrelation SpectroscopY) can overcome these limitations by construction via the additional information contained in correlation cross peaks. If this information is really relevant, 2D peaks help identify and quantify with more accuracy and certainty metabolites of interest. But COSY spectra are not yet widely used and there is not an agreed upon best practice for handling, pre-processing and assessment of 2D NMR spectra in metabolomics studies. In this thesis, a whole process is proposed to promote a more massive use of 2D data in such studies. First, for handling, grouping in a global object all the individual spectra stemming from a design and for pre-processing via the novel Global Peak List (GPL) and Vectorization workflows. Then, for assessing their quality in terms of signal capture via the Metabolomic Informative Content (MIC) concept, which is based on unsupervised clustering quality measures. MIC indexes are proposed to compare 1D and 2D spectral matrices, and different 2D data sets pre-processed with various scenarii. Finally, the biomarker issue is deeply considered (using PCA, PLS, OPLS, sPLS and ensemble trees approaches) and the orthogonality and sparsity issues are addressed. A novel sparse algorithm is proposed: L-sOPLS which can combine orthogonality, sparsity and efficient predictions. This process was convincingly applied on six real experimental designs.
- Affiliations
Citations
Feraud, B. (2019). Statistical contributions to the analysis of 2D NMR spectra in metabolomics studies : from pre-processing workflows to 2D biomarker discovery. https://hdl.handle.net/2078.5/53803