Differentiating Shigella from E. coli using hierarchical feature selection on MALDI-ToF MS data

Shigella is a genus of pathogenic enteric micro-organisms, which in some cases pose a lethal threat in a human host. It is closely related to E. coli, another pathogenic enteric micro-organism. Although in a clinical observation they cause different symptoms, on a genetic level they are very similar (Brenner, Fanning, Miklos & Steigerwalt, 1973). This poses a problem when the need arises to differentiate the Shigella and E. coli genetically. A possible way to accomplish this, is through a MALDI-ToF MS analysis. Differentiating Shigella from E. coli is important since Shigella is one of the sources to infectious diarrhea that cause a problem in both developing and developed countries worldwide, that has a potential lethal result (Cheng, McDonald & Thielman, 2005). Our work is based on MALDI-ToF MS data containing a comprehensive analysis of various Shigella species and E. coli phylotypes. We propose an approach which leads to a proper differentiation between both Shigella, E. coli and their respective species or phylogroups. We make use of the elastic net method (Zou & Hastie, 2005) to extract the most important features that allow this differentiation without losing any correlation between the features. We further extend the elastic net method to be build-up in using multiple models abiding a hierarchical structure in order to increase prediction performance. We compare two hierarchical structures, one based on the evolutionary phylogeny and one on pathotype (Y. Zhang & Lin, 2012). Using a hierarchy we can increase predictive performance and it allows us to find features specific for each hierarchy level, without having any added noise of unrelated hierarchy levels. Furthermore we compare the impact in terms of predictive performance, feature selection and feature stability, by using two approaches to preprocessing the raw MALDI-ToF MS data. In both approaches we use common data-preprocessing techniques, but in one case we leave out the aggressive data reduction steps, smoothing and binning.