Pattern recognition analysis of differential mobility spectra with classification by chemical family

Anal Chim Acta. 2006 Oct 2;579(1):1-10. doi: 10.1016/j.aca.2006.07.013. Epub 2006 Jul 14.


Differential mobility spectra for alkanes, alcohols, ketones, cycloalkanes, substituted ketones, and substituted benzenes with carbon numbers between 3 and 10 were obtained from gas chromatography-differential mobility spectrometry (GC-DMS) analyses of mixtures in dilute solution. Spectra were produced in a supporting atmosphere of purified air with 0.6-0.8 ppm moisture, gas temperature of 120 degrees C, sample concentrations of approximately 0.2-5 ppm, and ion source of 5 mCi (185 MBq) 63Ni. Multiple spectra were extracted from chromatographic elution profiles for each chemical providing a library of 390 spectra from 39 chemicals. The spectra were analyzed for structural content by chemical family using two different approaches. In the one approach, the wavelet packet transform was used to denoise and deconvolute the DMS data by decomposing each spectrum into its wavelet coefficients, which represent the sample's constituent frequencies. The wavelet coefficients characteristic of the compound's structural class were identified using a genetic algorithm (GA) for pattern recognition analysis. The pattern recognition GA uses both supervised and unsupervised learning to identify coefficients which optimize clustering of the spectra in a plot of the two or three largest principal components of the data. Because principal components maximize variance, the bulk of the information encoded by the selected coefficients is about differences between chemical families in the data set. The principal component analysis routine embedded in the fitness function of the pattern recognition GA acts as an information filter, significantly reducing the size of the search space since it restricts the search to coefficients whose principal component plots show clustering on the basis of chemical family. In a second approach, a back propagation neural network was trained to categorize spectra by chemical families and the network was successfully tested using familiar and unfamiliar chemicals. Performance of the network was associated with a region of the spectrum associated with fragment ions which could be extracted from spectra and were class specific.