RESUMEN
Using raw GC/MS data as the X-block for chemometric modeling has the potential to provide better classification models for complex samples when compared to using the total ion current (TIC), extracted ion chromatograms/profiles (EIC/EIP), or integrated peak tables. However, the abundance of raw GC/MS data necessitates some form of data reduction/feature selection to remove the variables containing primarily noise from the data set. Several algorithms for feature selection exist; however, due to the extreme number of variables (10(6)-10(8) variables per chromatogram), the feature selection time can be prolonged and computationally expensive. Herein, we present a new prefilter for automated data reduction of GC/MS data prior to feature selection. This tool, termed unique ion filter (UIF), is a module that can be added after chromatographic alignment and prior to any subsequent feature selection algorithm. The UIF objectively reduces the number of irrelevant or redundant variables in raw GC/MS data, while preserving potentially relevant analytical information. In the m/z dimension, data are reduced from a full spectrum to a handful of unique ions for each chromatographic peak. In the time dimension, data are reduced to only a handful of scans around each peak apex. UIF was applied to a data set of GC/MS data for a variety of gasoline samples to be classified using partial least-squares discriminant analysis (PLS-DA) according to octane rating. It was also applied to a series of chromatograms from casework fire debris analysis to be classified on the basis of whether or not signatures of gasoline were detected. By reducing the overall population of candidate variables subjected to subsequent variable selection, the UIF reduced the total feature selection time for which a perfect classification of all validation data was achieved from 373 to 9 min (98% reduction in computing time). Additionally, the significant reduction in included variables resulted in a concomitant reduction in noise, improving overall model quality. A minimum of two um/z and scan window of three about the peak apex could provide enough information about each peak for the successful PLS-DA modeling of the data as 100% model prediction accuracy was achieved. It is also shown that the application of UIF does not alter the underlying chemical information in the data.
RESUMEN
There is a fundamental difference between data collected in comprehensive two-dimensional gas chromatographic (GCxGC) separations and data collected by one-dimensional GC techniques (or heart-cut GC techniques). This difference can be ascribed to the fact that GCxGC generates multiple sub-peaks for each analyte, as opposed to other GC techniques that generate only a single chromatographic peak for each analyte. In order to calculate the total signal for the analyte, the most commonly used approach is to consider the cumulative area that results from the integration of each sub-peak. Alternately, the data may be considered using higher order techniques such as the generalized rank annihilation method (GRAM). Regardless of the approach, the potential errors are expected to be greater for trace analytes where the sub-peaks are close to the limit of detection (LOD). This error is also expected to be compounded with phase-induced error, a phenomenon foreign to the measurement of single peaks. Here these sources of error are investigated for the first time using both the traditional integration-based approach and GRAM analysis. The use of simulated data permits the sources of error to be controlled and independently evaluated in a manner not possible with real data. The results of this study show that the error introduced by the modulation process is at worst 1% for analyte signals with a base peak height of 10xLOD and either approach to quantitation is used. Errors due to phase shifting are shown to be of greater concern, especially for trace analytes with only one or two visible sub-peaks. In this case, the error could be as great as 6.4% for symmetrical peaks when a conventional integration approach is used. This is contrasted by GRAM which provides a much more precise result, at worst 1.8% and 0.6% when the modulation ratio (MR) is 1.5 or 3.0, respectively for symmetrical peaks. The data show that for analyses demanding high precision, a MR of 3 should be targeted as a minimum, especially if multivariate techniques are to be used so as to maintain data density in the primary dimension. For rapid screening techniques where precision is not as critical lower MR values can be tolerated. When integration is used, if there are 4-5 visible sub-peaks included for a symmetrical peak at MR=3.0, the data will be reasonably free from phase-shift-induced errors or a negative bias. At MR=1.5, at least 3 sub-peaks must be included for a symmetrical peak. The proposed guidelines should be equally relevant to LCxLC and other similar techniques.