An m/z- and Intensity-Based HRMS Clustering Algorithm Targeted toward Single-Cell Metabolomics

Single-cell (SC) metabolomics holds great potential in the development of novel diagnostic tools and mechanistic insights into cell biology. Using high-resolution mass spectrometry (HRMS), the masses of a single cell's constituents can be determined with an accuracy high enough to derive their respective elemental compositions. Using a molecule's mass and its MS fragmentation pattern, in many cases a molecular structure can be assigned or looked up in databases. Due to the small measurement cell volume of an Orbitrap HRMS instrument, samples need to be scanned multiple times, which necessitates across-scan clustering per sample, and across-sample alignment of m/z values. However, existing HRMS data processing software is not designed to process SC HRMS data, as it typically requires liquid chromatography retention times or reference spectra for m/z clustering and alignment. Herein, a novel, robust SC HRMS m/z clustering and alignment algorithm is presented and compared with two commercially available and industry standard algorithms used by Sciex MarkerView and Thermo FreeStyle. Furthermore, output is compared with clustering results from DBSCAN and MaldiQuant binning. Our algorithm, Global Clustering unTargeted Analysis (GCTA), enforces a strict maximum on the cluster size, thereby reducing the chance of peak aggregation. Furthermore, by design, GCTA enables noise filtering based on intensity and number of peaks. Across-scan clustering and across-sample alignment were contrasted for accuracy in finding peaks identified by commercial software output and peaks with known m/z values corresponding to standards and HMDB and LipidMaps database hits. Comparisons are made based on data recorded for quality control samples containing standard mixes as well as SC HRMS data recorded for two different cell lines. This work shows that the presented algorithm is comparable in accuracy with respect to MarkerView and FreeStyle, reliably identifies compounds, is less prone to peak splitting than MaldiQuant binning while providing similar levels of error in clustering peaks, and successfully filters noise. Furthermore, it is shown to be competitive with DBSCAN, MaldiQuant binning and MarkerView when compared to theoretical m/z values based on database hits. GCTA encompasses both m/z clustering and reference-free alignment, which makes it pivotal to further development of untargeted SC HRMS metabolomics.