Self-organizing maps: a versatile tool for the automatic analysis of untargeted imaging datasets

Introduction Mass Spectrometry Imaging (MSI) experiments constitute the ideal complement to metabolomics to investigate the spatial distribution of key metabolites. In spite of their caveats and limitations, they generate highly informative datasets, which are difficult to mine mainly due to their sheer size. In this contribution we illustrate how self-organising maps (SOMs) could be efficiently used to automatically analyze spatial information in MSI untargeted metabolomics datasets. In our approach, SOMs are used to identify a shortlist of m/z signals sharing a common, characteristic and interesting spatial distribution, thus labeling them as “biomarkers” for an area of the section. Additionally, the proposed algorithm can be used to process the raw data and extract high-resolution information on interesting ions. Methods Untargeted full scan (m/z 120 - m/z 700) imaging experiments were performed on apple (Golden Delicious) sections with a MALDI LTQ Orbitrap XL mass spectrometer with a resolution of 60.000. The CHCA matrix was deposited by using an ImagePrep station. Raw data were converted into the open CDF format and analyzed with a set of algorithms developed in R. Results The proposed algorithm has been applied to the imaging dataset collected on the apple section (Figure) to identify 42 characteristic spatial distributions. The one grouping the ions which show a high concentration in the region below the apple skin and in correspondence of one of the apple bundles is shown in Figure. The SOM algorithm associates to this spatial class a list of 35 ions. For 17 of these ions, it was possible to to associate them to secondary metabolites known to be present in apple in this specific area. Conclusions SOMs form a versatile tool for the untargeted analysis of high-resolution and high-accuracy MSI metabolomics datasets where they can be used to automatically identify spatial patterns and assess co-localization among different ions. This co-localization can be used to improve the chemical selectivity of imaging experiments, giving important tissue-specific information. Novel Aspect With the proposed algorithm, SOMs are used to associate the thousands of signals collected over the tissue to a limited number of characteristic spatial distributions. The ions belonging to the same spatial class are co-localized and they can be used in combination to mass spectra libraries and in-silico fragmentation engines to perform (partial) chemical annotation.