Investigation of supervised and unsupervised discovery–based chemometric tools to expand the scope of multidimensional gas chromatography

Abstract

Comprehensive two-dimensional gas chromatography (GCÃ GC) has emerged as a powerful multidisciplinary tool for the separation of complex mixtures since its inception in 1991. When coupled to a time-of-flight mass spectrometer (TOFMS), GCÃ GC analysis can confidently identify up to thousands of analytes. The wealth of information provided within this three-dimensional (3D) data cube is too great for manual scrutinization, hence necessitating non-targeted chemometric analysis. Within non-targeted analysis, both supervised and unsupervised algorithms exist, which differ by the presence and absence of sample class labels, respectively. This dissertation describes several studies aimed at critically investigating non-targeted chemometric tools such as tile-based Fisher ratio (F-ratio) analysis and principal component analysis (PCA) for numerous multidimensional GC platforms, including GCÃ GC-FID, GCÃ GC-TOFMS, and comprehensive 3D GC (GC3)-TOFMS. First, the utility of data binning prior to PCA is demonstrated for a GCÃ GC-FID dataset of five diesel fuels, wherein the optimum bin size maintains chemical selectivity and improves the signal-to-noise ratio (S/N). Next, the advantage of ranking F-ratio hitlists using the top F-ratio m/z is demonstrated for low concentration comparisons of spiked and un-spiked JP8 jet fuel and the “limit of discovery” is identified as the limit of quantification (LOQ). Tile-based F-ratio analysis is then coupled with an offline one-way analysis of variance (ANOVA) to characterize the geographical differences between five Sicilian wines. An unsupervised algorithm known as tile-based variance rank initiated-unsupervised sample indexing (VRI-USI) is developed to identify sample-to-sample relationships in GCÃ GC-TOFMS data. Upon application of VRI-USI to a complex multi-fuel dataset, patterns in the resulting k-means clustering index assignments for each hit in the hitlist correctly revealed the presence of classes and sub-classes, which holds promise for future studies requiring chemical characterization of unknown samples. Finally, the first application of PCA to GC3 -TOFMS data is reported herein, wherein the 3D loadings are used to distinguish four jet fuels.