Sparse non-negative matrix factorization for the processing of mass spectrometry data in metabolomics

Overview & Synopsis

Mass spectrometry (MS) based metabolomics is a leading strategy for biomarker discovery. In particular, tandem mass spectrometry facilitates the structural characterization of metabolites through the acquisition of fragmentation spectra (MS2). Recent approaches to Data Independent Acquisition (DIA) such as SWATH (Sequential Window Acquisition of all THeoretical fragment ions) enable high-throughput fragmentation of all precursor ions. However, the SWATH DIA spectra obtained are hybrid, as they contain mixed fragments from all co-isolated compounds. This is why the development of strategies capable of separating these spectra in a completely untargeted manner is a major challenge for metabolite identification. To process SWATH DIA data rigorously and efficiently, we have developed a new approach based on sparse non-negative matrix factorization (NMF) named DIANMF. This is the first blind source separation method that simultaneously processes MS1 and MS2 data in a unified framework, without relying on predefined peak models or spectral libraries. We have implemented an NMF algorithm derived from non-negative generalized morphological analysis with soft thresholding (nGMCAˢ), which exploits non-negativity and parsimony (in a predefined known transformation domain) as well as morphological diversity to improve separation in heavily mixed linear data. We have shown through simulation that a high rank can be chosen a priori so that the spectra of all compounds present in the time window can be extracted simultaneously at each iteration, and we have developed criteria for selecting only the components corresponding to signal and assigning each to the precursor ions of the same molecule. The entire DIANMF workflow is implemented as the publicly available software library DIANMF (https://github.com/odisce/DIANMF). Application of the software to several real-world datasets representative of the main detection instruments has shown that DIANMF outperforms existing tools in terms of compound identification rates and spectral quality, particularly for low-abundance and co-eluting metabolites. Furthermore, our approach using simultaneous NMF on MS1 and MS2 data makes it possible for the first time to group all ion species associated with the same compound (adducts, isotopes, and neutral losses) into a chemically coherent spectrum. Taken together, these results demonstrate the value of DIANMF for the processing of SWATH DIA data and the identification of biomarkers in metabolomics.