Algorithmic Learning for Auto-deconvolution of GC-MS Data to Enable Molecular Networking within GNPS

Abstract

Gas chromatography-mass spectrometry (GC-MS) represents an analytical technique with significant practical societal impact. Spectral deconvolution is an essential step for interpreting GC-MS data. No public GC-MS repositories that also enable repository-scale analysis exist, in part because deconvolution requires significant user input. We therefore engineered a scalable machine learning workflow for the Global Natural Product Social Molecular Networking (GNPS) analysis platform to enable the mass spectrometry community to store, process, share, annotate, compare, and perform molecular networking of GC-MS data. The workflow performs auto-deconvolution of compound fragmentation patterns via unsupervised non-negative matrix factorization, using a Fast Fourier Transform-based strategy to overcome scalability limitations. We introduce a “balance score” that quantifies the reproducibility of fragmentation patterns across all samples. We demonstrate the utility of the platform with breathomics analysis applied to the early detection of oesophago-gastric cancer, and by creating the first molecular spatial map of the human volatilome.

Exploring correlations between MS and NMR for compound identification using essential oils: A pilot study

“…The total run time was 62 min and MS transferline temperature was kept at 250 C. GC-MS data were fully processed using MZMine2 13 under the ADAP 14 methods for peaks and spectral deconvolution. The feature list was then submitted to MN using the GC-MS-based workflow 15 and Cytoscape 16 for visualisation and handling (specific parameters for the MZMine2 and GNPS are shown in Tables S2 and S3). The same feature list was imported to Matlab for the SHY calculations.…”

Network-based strategies in metabolomics data analysis and interpretation: from molecular networking to biological interpretation

“…Another interesting tool recently incorporated was the algorithmic learning for auto-deconvolution of GC-MS data to enable molecular networking within GNPS [73]. The initial molecular networking strategies were developed with a focus on LC-MS 2 data, which is considerably different from traditional GC-EI-MS data.…”

“…The new functionality introduced in GNPS uses a machinelearning algorithm based on unsupervised non-negative matrix factorization to automatically select parameters for the deconvolution of GC-EI-MS data returning a 'balance score' that can be used to access how consistent the spectra is across the dataset. The strategy adopted makes data processing computationally efficient and more accessible to less experienced users, while also offering the direct integration with molecular networking available through GNPS [73].…”

“…Molecular networking, for instance, has been proven to be of great utility for metabolite annotation; however, little has been done in terms of analyzing the structure of the resulting metabolic networks. Moreover, the incorporation of extra layers of information within these networks have the potential to reveal unanticipated features, as it has been observed with the specificity of certain classes of compounds detectable by GC-MS in regard to frogs, cheese, and beer samples [73], revealed simply by the integration of multiple datasets and addition of metadata. The incorporation of FBMN into GNPS represents a great step forward in this sense, adding a layer of quantitative information in the networks.…”

Fungal–bacterial interaction selects for quorum sensing mutants with increased production of natural antifungal compounds

“…From PLS-DA, we found that 2heptanone was the most discriminating compound produced in greater amount post-mutation only by B. subtilis, and not by the fungus. Interestingly, by performing a GC data molecular network analysis 24 within the Global Natural Products Social (GNPS) platform 22 , we could discriminate a cluster composed of a family of ketones, remarkably, all 2-ketones ( Fig. 3a).…”