Feature selection from mass spectra of bacteria for serotyping Salmonella

• Published in: Journal of analytical and applied pyrolysis Vol. 124; pp. 393 - 402
• Main Authors: Slavov, Svetoslav, Alusta, Pierre, Buzatu, Dan A., Wilkes, Jon G.
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.03.2017
• Subjects: algorithms; bacteria; discriminant analysis; Feature selection; ions; Outbreak recognition; Outbreak traceback; pathogens; pyrolysis; Pyrolysis mass spectrometry; Rapid Salmonella serotyping; Salmonella; serotypes; Outbreak recognition; Feature selection; Rapid Salmonella serotyping; Outbreak traceback; Pyrolysis mass spectrometry
• ISSN: 0165-2370; 1873-250X
• DOI: 10.1016/j.jaap.2016.10.014

•Feature selection from Py-MS of intact Salmonella cells facilitates serotyping.•Without feature selection to yield only 19 ions, serotyping isn’t obtained.•Selected ions are apparently derived from 3 plausible biochemical classes.•Feature selected Py-MS enables epidemiological concordance.•Spectra can be obtained in only 10s elapsed between analyses. Algorithms applied to mass spectra of Salmonella isolates were used to achieve accurate and rapid serotyping. Two classification strategies were developed, one using all ions and the other, only 19. Without feature selection, classification based on all ions (m/z 70–700) could distinguish similar strains, but samples so classified did not form serotype similarity super-clusters: i.e., different strains of the same serotype did not yield replicate spectra similar to each other when subject to pattern recognition and analysis. The first feature-selection strategy used all available ions, grouped training set samples by serotype (four groups), and passed the Principal Component eigenvectors into Linear Discriminant Analysis. By emphasizing the relative intensities of serovar-related ions, this approach provided implicit feature selection. The best model was used to predict the serotype of seven external samples (representatives from each of the four serotypes), producing correct serovar assignments for three of seven. The second strategy involved explicit feature selection. The five most statistically significant ions for distinguishing each pair were nominated. Redundancy among nominees reduced their number from 30 to 19, ions that consistently differentiated samples based on the six pairs among four classes. The resulting patterns were assessed by a novel pattern recognition approach, in which calculated ratios of average ion intensity for each ion pair combination were analyzed by Tanimoto Similarity and used to classify spectra with respect to serovar by k Nearest Neighbors. Six of the seven external samples were correctly serotyped. Ion redundancy observed among nominees for distinguishing particular serovar pairs suggested the approach can be extended to more than four serovars. Biochemical sources for serovar-correlated ions were provisionally associated with cell surface constituents prominent in Gram-negative pathogens, which also suggested the method could be used to distinguish other, perhaps many other, serovars.