Baptiste Féraud1,2, Justine Leenders3, Estelle Martineau4,5, Patrick Giraudeau4,6, Bernadette Govaerts7, Pascal de Tullio3. 1. Institute of Statistics, Biostatistics and Actuarial Sciences (ISBA), Université Catholique de Louvain (UCL), Voie du Roman Pays 20, bte L1.04.01, 1348, Louvain-la-Neuve, Belgium. baptiste.feraud@uclouvain.be. 2. Machine Learning Group, Université Catholique de Louvain (UCL), Louvain-la-Neuve, Belgium. baptiste.feraud@uclouvain.be. 3. Center for Interdisciplinary Research on Medicines (CIRM), Metabolomics group, Université de Liège (ULg), Liege, Belgium. 4. EBSI Team, Chimie et Interdisciplinarité, Synthèse, Analyse, Modélisation (CEISAM), CNRS, UMR 6230, Université de Nantes, Nantes, France. 5. Spectromaîtrise, CAPACITES SAS, 26 Bd Vincent Gâche, 44200, Nantes, France. 6. Institut Universitaire de France, 1 rue Descartes, 75005, Paris Cedex 5, France. 7. Institute of Statistics, Biostatistics and Actuarial Sciences (ISBA), Université Catholique de Louvain (UCL), Voie du Roman Pays 20, bte L1.04.01, 1348, Louvain-la-Neuve, Belgium.
Abstract
INTRODUCTION: The pre-processing of analytical data in metabolomics must be considered as a whole to allow the construction of a global and unique object for any further simultaneous data analysis or multivariate statistical modelling. For 1D 1H-NMR metabolomics experiments, best practices for data pre-processing are well defined, but not yet for 2D experiments (for instance COSY in this paper). OBJECTIVE: By considering the added value of a second dimension, the objective is to propose two workflows dedicated to 2D NMR data handling and preparation (the Global Peak List and Vectorization approaches) and to compare them (with respect to each other and with 1D standards). This will allow to detect which methodology is the best in terms of amount of metabolomic content and to explore the advantages of the selected workflow in distinguishing among treatment groups and identifying relevant biomarkers. Therefore, this paper explores both the necessity of novel 2D pre-processing workflows, the evaluation of their quality and the evaluation of their performance in the subsequent determination of accurate (2D) biomarkers. METHODS: To select the more informative data source, MIC (Metabolomic Informative Content) indexes are used, based on clustering and inertia measures of quality. Then, to highlight biomarkers or critical spectral zones, the PLS-DA model is used, along with more advanced sparse algorithms (sPLS and L-sOPLS). RESULTS: Results are discussed according to two different experimental designs (one which is unsupervised and based on human urine samples, and the other which is controlled and based on spiked serum media). MIC indexes are shown, leading to the choice of the more relevant workflow to use thereafter. Finally, biomarkers are provided for each case and the predictive power of each candidate model is assessed with cross-validated measures of RMSEP. CONCLUSION: In conclusion, it is shown that no solution can be universally the best in every case, but that 2D experiments allow to clearly find relevant cross peak biomarkers even with a poor initial separability between groups. The MIC measures linked with the candidate workflows (2D GPL, 2D vectorization, 1D, and with specific parameters) lead to visualize which data set must be used as a priority to more easily find biomarkers. The diversity of data sources, mainly 1D versus 2D, may often lead to complementary or confirmatory results.
INTRODUCTION: The pre-processing of analytical data in metabolomics must be considered as a whole to allow the construction of a global and unique object for any further simultaneous data analysis or multivariate statistical modelling. For 1D 1H-NMR metabolomics experiments, best practices for data pre-processing are well defined, but not yet for 2D experiments (for instance COSY in this paper). OBJECTIVE: By considering the added value of a second dimension, the objective is to propose two workflows dedicated to 2D NMR data handling and preparation (the Global Peak List and Vectorization approaches) and to compare them (with respect to each other and with 1D standards). This will allow to detect which methodology is the best in terms of amount of metabolomic content and to explore the advantages of the selected workflow in distinguishing among treatment groups and identifying relevant biomarkers. Therefore, this paper explores both the necessity of novel 2D pre-processing workflows, the evaluation of their quality and the evaluation of their performance in the subsequent determination of accurate (2D) biomarkers. METHODS: To select the more informative data source, MIC (Metabolomic Informative Content) indexes are used, based on clustering and inertia measures of quality. Then, to highlight biomarkers or critical spectral zones, the PLS-DA model is used, along with more advanced sparse algorithms (sPLS and L-sOPLS). RESULTS: Results are discussed according to two different experimental designs (one which is unsupervised and based on human urine samples, and the other which is controlled and based on spiked serum media). MIC indexes are shown, leading to the choice of the more relevant workflow to use thereafter. Finally, biomarkers are provided for each case and the predictive power of each candidate model is assessed with cross-validated measures of RMSEP. CONCLUSION: In conclusion, it is shown that no solution can be universally the best in every case, but that 2D experiments allow to clearly find relevant cross peak biomarkers even with a poor initial separability between groups. The MIC measures linked with the candidate workflows (2D GPL, 2D vectorization, 1D, and with specific parameters) lead to visualize which data set must be used as a priority to more easily find biomarkers. The diversity of data sources, mainly 1D versus 2D, may often lead to complementary or confirmatory results.
Authors: Manuela J Rist; Alexander Roth; Lara Frommherz; Christoph H Weinert; Ralf Krüger; Benedikt Merz; Diana Bunzel; Carina Mack; Björn Egert; Achim Bub; Benjamin Görling; Pavleta Tzvetkova; Burkhard Luy; Ingrid Hoffmann; Sabine E Kulling; Bernhard Watzl Journal: PLoS One Date: 2017-08-16 Impact factor: 3.240