Marker discovery in volatolomics based on systematic alignment of GC-MS signals: Application to food authentication

Starting with an experiment to authenticate walnut oils based on GC-MS analysis of the volatolome, this paper aims to demonstrate the relevance of a two-step alignment-based strategy for the systematic research of VOC markers. The first step of the treatment consists of roughly reducing the time shifts with efficient, known warping techniques like COW (Correlation Optimized Warping). The second step relies on an accurate peak apex alignment in order to refine residual local misalignments and to enable further systematic marker research through univariate or multivariate data treatments. This two-step strategy was implemented on 117 GC-MS analyses of the volatolome of three vegetable oils with very similar composition. During the analysis campaign, the GC-MS system was intentionally subjected to instrumental drifts in order to generate realistic signal shifts. The first part of this study aims to assess the efficiency of the warping-based strategy in terms of signal alignment and sample discrimination. Whereas no distinction between the three oils was possible with unaligned raw GC-MS data, the application of COW enabled a significant but insufficient improvement of both reduction of temporal drifts and between-group separation with 79% of samples being well-classified according to Linear Discriminant Analysis (LDA). Applying the peak apex alignment procedure to COW-treated signals resulted in a suitable correction of the remaining local distortions and improved the proportion of well-classified samples in LDA to 100%. The second part of this study assesses the robustness of the discriminant markers highlighted in this approach by: (i) discussing the relevance of the best markers involved in the LDA model, where a close review of literature confirmed the consistency for two of them, and (ii) validating highlighted makers by retrieving the set of the 23 markers previously determined by manual processing among those automatically found. The third part shows the potential of the systematic approach for untargeted detection of 184 highly significant relevant markers from the oil volatolome. Finally, the fourth part presents a comparison of our hybrid alignment strategy with two reference alignment methods (iCoshift and STW) in order to assess quality alignment of the GC-MS data and to show the three methods' abilities to detect discriminant markers.