Skip to main content
PMC home page
Bioinformatics logoLink to Bioinformatics
. 2024 Sep 30;40(10):btae584. doi: 10.1093/bioinformatics/btae584

Combined LC-MS/MS feature grouping, statistical prioritization, and interactive networking in msFeaST

Kevin Mildau 1,2,3,✉, Christoph Büschl 4, Jürgen Zanghellini 5, Justin J J van der Hooft 6,7,✉
Editor: Pier Luigi Martelli
PMCID: PMC11471276  PMID: 39348165

Abstract

Summary

Computational metabolomics workflows have revolutionized the untargeted metabolomics field. However, the organization and prioritization of metabolite features remains a laborious process. Organizing metabolomics data is often done through mass fragmentation-based spectral similarity grouping, resulting in feature sets that also represent an intuitive and scientifically meaningful first stage of analysis in untargeted metabolomics. Exploiting such feature sets, feature-set testing has emerged as an approach that is widely used in genomics and targeted metabolomics pathway enrichment analyses. It allows for formally combining groupings with statistical testing into more meaningful pathway enrichment conclusions. Here, we present msFeaST (mass spectral Feature Set Testing), a feature-set testing and visualization workflow for LC-MS/MS untargeted metabolomics data. Feature-set testing involves statistically assessing differential abundance patterns for groups of features across experimental conditions. We developed msFeaST to make use of spectral similarity-based feature groupings generated using k-medoids clustering, where the resulting clusters serve as a proxy for grouping structurally similar features with potential biosynthesis pathway relationships. Spectral clustering done in this way allows for feature group-wise statistical testing using the globaltest package, which provides high power to detect small concordant effects via joint modeling and reduced multiplicity adjustment penalties. Hence, msFeaST provides interactive integration of the semi-quantitative experimental information with mass-spectral structural similarity information, enhancing the prioritization of features and feature sets during exploratory data analysis.

Availability and implementation

The msFeaST workflow is freely available through https://github.com/kevinmildau/msFeaST and built to work on MacOS and Linux systems.

1 Introduction

Untargeted metabolomics deals with the comprehensive characterization of the composition of small chemicals, or metabolites, in biological samples. Typically, high-resolution Liquid Chromatography Tandem Mass Spectrometry (LC-MS/MS) workflows are used to provide comprehensive snapshots of the metabolome (Wolfender et al. 2018). However, despite recent advances in computational metabolomics, the reliable annotation of ms/ms spectral data remains a challenge (de Jonge et al. 2022). Hence, further complementary analyses, laborious manual annotations, and detailed validations of putative structure hypotheses remain a necessity (Wolfender et al. 2018, Beniddir et al. 2021). To assist with exploratory data analyses, tools like molecular networking provide with mean to organize and prioritize features for in-depth evaluations (Watrous et al. 2012, Nothias et al. 2020). From a technical perspective, molecular networking comprises a combination of data clustering and exploratory data visualization. Large datasets are subdivided into smaller, more manageable feature groups via a network topology approach based on spectral similarities (Nothias et al. 2020, Mildau et al. 2024). Those feature groups are further presented to the user as subnetworks (known as mass spectral molecular families) for exploration of relationships across features within them (Nothias et al. 2020). While advantageous as a first analysis step, relationships between subnetworks are lost completely (Olivon et al. 2018, Mildau et al. 2024), parameter setting for clustering and visualization are time-consuming and their evaluation opaque, and further manual customization of resulting molecular networks in Cytoscape is usually necessary to highlight statistical features for subnetwork or feature prioritization (Pakkir Shah et al. 2024).

While statistical data can often be manually integrated into Cytoscape networks, there is currently a lack of workflows in untargeted metabolomics that provide both spectral clustering and statistical data integration on the spectral group level aimed at feature prioritization (McLuskey et al. 2021). As a notable exception, the PALS (Pathway Activity Level Scoring) workflow makes use of molecular families, i.e. feature groupings representing “pathways” via linking together chemical analogues, for latent variable summation-based statistical comparisons across treatment groups (Tomfohr et al. 2005, McLuskey et al. 2021). PALS has two primary limitations however. First, it is based on mass spectral molecular networking groups and thus inherits the time-consuming and opaque clustering approach based on network topology. Second, it is exclusively based on the first component of a latent variable projection of group data, potentially introducing a loss of information. In addition, PALS stops short of visually integrating such analysis results back into the network representations it is based on. The field is thus still lacking an integrated approach providing spectral data clustering, statistical analysis at the group level, and integrated visualization of the results of both of these features for streamlined exploratory data analysis.

To fill this gap, we developed Mass Spectral Feature Set Testing (msFeaST), a comprehensive workflow for integrated analysis and visual exploration. We draw inspiration from gene-set testing and pathway enrichment analyses, which shift the focus from the often large and cluttered feature space to aggregated groups such as gene ontologies or pathways. This approach provides the basis group-based comparisons of experimental conditions that can then be further explored in detail (Rosato et al. 2018, Chong et al. 2020, Maleki et al. 2020). To achieve a similar approach in untargeted metabolomics, where genes are replaced by metabolite features with unknown pathway membership, we make use of k-medoids clustering on pairwise similarity data to provide homologous subdivisions of the data (Schubert and Rousseeuw 2021, Mildau et al. 2024). Testing is then performed for each cluster using globaltest, providing high power to detect small concordant treatment-specific effects across cluster members via a single group-wise statistical test (Goeman et al. 2004). In addition, msFeaST draws visualization inspiration from both specXplore and MetGem making use of a t-SNE embedding to represent the entire spectral dataset in one overview (Maaten 2008, Olivon et al. 2018, Mildau et al. 2024). This embedding overview, the clustering, and associated cluster-based testing are integrated into an interactive dashboard inspired by EdgeMaps and specXplore, allowing the user to interactively explore and trace pathway relationships between feature nodes within and across clusters in a threshold independent manner (Dork et al. 2011, Mildau et al. 2024) (Fig. 1).

Figure 1.

Figure 1.

Open in a new tab

View of the visualization panel of the msFeaST dashboard with overview embedding, side panels, group-wise specific P-value heatmap below. A differential feature is selected, highlighting its top-10 neighbors as set, as well as highlighting its set members in color. Node size represents a linear mapping of univariate, feature-specific P-values, showing clear areas of enrichment in the t-SNE map.

2 Methods and implementation

The msFeaST workflow is a combination of data clustering, statistical testing at the cluster level, and interactive visualization using overlays of 2D embeddings with an ego-network, i.e. a node-centric neighborhood graph (Fig. 1). An overview of the Python and R-based processing pipeline can be seen in Fig. 2.

Figure 2.

Figure 2.

Open in a new tab

msFeaST workflow overview and dashboard network visualization example. (A) Feature-based molecular networking data is translated into the expected msFeaST input data format. Once loaded, a msFeaST pipeline instance is created. Using msFeaST pipeline methods, spectral similarities are computed, clustering is performed (python package kmedoids), t-SNE embedding is performed (using scikit-learn), feature-set testing is performed (via an embedded R script, globaltest), and data is integrated into a json format compatible with the interactive javascript based visualization (visjs, plotly). The user only needs to run a sequence of commands while intermediate data structures are handled by the pipeline object. (B) msFeaST network visualization example run on the illustrative example mushroom data using ms2deepscore as the scoring approach (Huber et al. 2021b; de Jonge et al. 2024). Nodes represent individual features, while node size is derived from a linear mapping of statistical properties to node size (e.g. univariate P-values, or log-2 fold changes) to visually highlight differential features. The example shows a clicked feature within a feature set (member nodes highlighted in color) that shows differential abundance across fruiting bodies of Pleurotus eryngii cultivated using 0% and 80% olive mill solid waste mixed into their substrate. The selected node is shown with its top 30 neighbors, connecting to other clusters (nodes not highlighted in color) within the local t-SNE area.

2.1 Data processing for msFeaST

The msFeaST pipeline requires MS/MS spectral data, a feature quantification table, and statistical metadata for two or more sample groups. When working with feature based molecular networking data, pre-processing functionalities are available to load the data accordingly (Schmid et al. 2023, Nothias et al. 2020). Once the data are loaded, the msFeaST pipeline method is initialized and handles all intermediate data structures. The workflow guides the user through the data processing steps in the form of spectral similarity computations using the desired similarity score (Huber et al. 2020, 2021a,b, de Jonge et al. 2024), k-medoid clustering based on the resulting distance matrix (Schubert and Rousseeuw 2021, Schubert and Lenssen 2022), and t-SNE embedding using the same structure (Maaten 2008, Gove et al. 2022, Lause et al. 2024). To set the number of clusters k in clustering, the user may base their decision Silhouette scores for a comprehensive grid of K values, where high values of K are an indicator of better data subdivision (Tibshirani et al. 2001). We do note that it is ultimately up to the user to decide on the optimal subsetting of their dataset that best connects with their research question, as any measure such as the Silhouette scores can only be used as a guide (Jain 2010). Similarly, the choice of t-SNE embedding is assisted by a tuning grid over perplexity values with associated distance preservation scores and embedding plotting capabilities for evaluation. We refer to Supplementary Sections S1.1–S1.4 for further details. Here, we will focus specifically on the statistical testing and interactive visualization components.

2.2 Statistical testing via globaltest

Within this step of the pipeline the clustered data are passed to a R script for group-wise testing using globaltest, a so-called self-contained null hypothesis test (Goeman et al. 2004, Goeman and Bühlmann 2007, Fridley et al. 2010, Maleki et al. 2020, Goeman and Oosting 2023). The test was originally developed for genetic analyses, but has since found use in the field of targeted metabolomics as well (Rosato et al. 2018, Chong et al. 2020, Pang et al. 2021). The test is based on a generalized linear model which aims to assess whether any of the feature-specific effects in a pre-specified group of features have predictive utility in differentiating treatment groups, a task that is closely related to testing whether treatment group-specific effects are not null (for further details see Supporting Information Section S4). This test does not require concordance of effects but is considered powerful at detecting even small concordant effects (Goeman et al. 2004). Within msFeaST, each feature-set and contrast combination, that is control versus (treatment group 1, treatment group 2, etc.), is tested using this model, and P-values for the group as a whole, as well as feature-specific univariate P-values, are extracted. Multiple testing correction is applied to group-wise P-values using the Bonferroni method at the family size given by the number of groups times the number of contrasts. This correction is usually substantially smaller than the correction needed if univariate testing was performed, resulting in less stringent cutoff adjustments. However, special attention is needed when using globaltest with small sample sizes as results may be unreliable (Maleki et al. 2019).

2.3 Visual integration via interactive dashboard

The results of the msFeaST workflow are integrated into a .json formatted file that can be read by the msFeaST visualization dashboard. The visualization dashboard is run as a local browser-based javascript tool and makes use of the plotly and visjs libraries (Plotly-Technologies-Inc. 2015, VISJS Community 2024) (https://plot.ly/ and https://visjs.org/). Upon opening the website in a modern browser (e.g. Firefox, Chrome, Safari, Edge, etc.), the user may load the generated json file and start their analysis.

Loading the data initiates the interactive visualization tab of the dashboard (Fig. 2). This page contains t-SNE embedding-based overview representation of the spectra data, which serves as both a global data view and interactive exploratory analysis platform. Nodes represent individual features, with their sizes encoding linear mappings to pixel width of feature-specific selectable statistical mappings. Two mappings are available: absolute log-2 fold change for the feature, or feature-specific P-values. Both mappings are contrast specific and the user can toggle between contrasts using dropdown menus.

Users may hover over nodes to receive node information, e.g. feature identifiers and group identifier for the group the feature belongs to. In addition, they may click on nodes to receive feature details including feature-specific and group-specific statistics results. The last-clicked node and all members in its k-medoid cluster will be highlighted in color. Clicking on a node further prompts the overlay of local topology via top-K edge drawing, giving a glimpse into connectivity within and across groups of the feature in a similarity threshold independent manner. This approach allows exploring neighborhood for any node regardless of nearest neighbor similarity value. To allow for quick assessments of similarity, edge overlays make use of a discrete similarity mapping to give a qualitative indication of connection strength, while quantitative similarity score labels provide more precise information.

Clicking successively on different nodes adds additional edge overlays until a click on the empty canvas is used to reset edge overlays. In this way, local node neighborhood topology can be explored easily without causing computational bottlenecks or visual overload. In addition to the t-SNE embedding, the tool allows the user to make use of iterations of force-directed layout stabilization to mitigate overlapping nodes sometimes caused by the t-SNE embedding.

Group-based prioritization in the network view is supported by a connected heatmap representation of group-wise P-values which provides a quick reference to groups with statistically significant deviations for the selected contrast. Clicking on a group entry in the heatmap highlights the respective group in the t-SNE embedding, and vice-versa.

3 Illustrative example

To showcase msFeaST we make use of the high-resolution LC-MS/MS data from the study of Kathib et al. (2024). Specifically, we make use of the contrast of the Pleurotus eryngii fruiting body samples grown on 0% olive mill solid waste substrate against those grown on 80% olive mill solid waste substrate. While each treatment group contained only three samples, substantial differences in the metabolome can be observed with msFeaST, where many metabolites show elevated log-2 fold changes, and a lower number showing corresponding statistical enrichment patterns. Using the node size encoding, the t-SNE overview representation thus allows to spot in a straightforward manner the features showing differential abundance trends. Feature subsets of potential interest are easily spotted and their local topology can be explored (see Supplementary Fig. S1). For example, the differential feature in Fig. 2B is visualized with its top 30 neighbor nodes, highlighting relatively strong connectivity of this feature beyond its small feature cluster highlighted in color. The displayed edges encode not only connectivity, but also pairwise similarity, providing a means of checking connectivity strength between connected features. The heatmap of globaltest feature-set testing P-values provides an overview of cluster-wise results, and a means to highlight promising clusters in the embedding overview 1.

The processed data files are available on github with the respective processing notebooks and can be used to inspect the results within the visualization dashboard. No installation is required for this, only the pre-processed .json file and the html dashboard bundle that can be used using web-browsers are needed (see github readme quickstart). In addition to files for the illustrative example shown here, additional runs using the modified cosine score on the above contrast, as well as a modified cosine score comparison between the different mushroom types are included. The latter show large differential intensity patterns resulting in wide scale feature highlighting.

4 Limitations and future work

Our msFeaST workflow combines feature-clustering, feature-set-based statistical analysis, and interactive visualization into one workflow. Each one of these steps comes with potentially impactful choices on the conclusions drawn. We note that k-medoid clustering, its comparison to alternatives, and the formal setting of the parameter k have not been studied in detail. To move beyond the guidelines set and shared in this paper, more research is needed to compare clustering methods and evaluate potential optimality criteria in the context of untargeted metabolomics. We note, however, that strict and clear-cut guidelines are unlikely to be realized. Unsupervised learning problems are generally ill-specified and do not provide singular answers (Jain 2010), nor are untargeted metabolomics mass spectral data or their underlying chemical structures easy to categorize in a universal sense (e.g. Djoumbou Feunang et al. 2016, Kim et al. 2021).

We make use of globaltest to perform feature-set-based testing. This method is used in pathway enrichment analyses, with the main difference between our tool and existing methods lying in the usage of data-driven feature clustering rather than grouping according to a-priori pathways (Chong et al. 2020). A comprehensive study of clustering methods and their interaction with testing (including multiple testing adjustments for feature-sets) and general prioritization metrics would be valuable future research contributions.

We note here that transparency in the use of statistical methods is important for scientific reporting. While we strongly recommend against re-running testing for various k with the aim of increasing the number of significant features, any researcher doing so for exploratory purposes should report the impact of the settings on their final results and conclusions (Steegen et al. 2016, Gelman and Hennig 2017). We believe that flexibility of exploratory analyses is critical to the scientific process, and best accommodated with transparency (Tukey 1980, Steegen et al. 2016, Gelman and Hennig 2017, Thompson et al. 2020). Stricter approaches can and should be used in confirmatory analyses (Wicherts et al. 2016).

Finally, we note that visualization strategies employed in msFeaST work for certain ranges of settings, and cannot accommodate all types of data or settings equally well. For example, a large number of clusters, including high significant fraction, may interfere with the heatmap’s capability to pinpoint interesting feature groups. In such a case, most features would be interesting for a follow-up through their strong differential patterns. Extending the visualization capabilities beyond the prioritization and network exploration dashboard is an important area for future developments.

5 Conclusion

The msFeaST workflow separates visual design considerations from data subdivision and clustering criteria. Clustering is dealt with separately, using its own optimality criteria, and different clustering methods could in principle be used with the pipeline. It thus allows for a more principled approach to subdividing spectral data that is not impacted by network visualization settings aimed at limiting visual overload (such as edge thresholds, top-K edge limits on each node, maximum cluster sizes). This separation is made possible by making use of a network visualization approach using “interactive details on demand” design. Here, interactive neighborhood explorations using top-K edge overlays provide a robust means of exploring and visualizing data. The visualization and clustering workflows can handle different similarity scores without requiring extensive visual or topological fine tuning. The msFeaST workflow thus seamlessly combines spectral similarity-based feature clustering, feature cluster prioritization using statistical contrast information, and local topological neighborhood exploration in interactive exploratory analysis context. We believe that msFeaST can assist researchers in better understanding their untargeted metabolomics data and identifying relevant chemistry by leveraging both spectral similarity and statistical information.

Supplementary Material

btae584_Supplementary_Data
btae584_supplementary_data.pdf (4.8MB, pdf)

Acknowledgements

The authors would like to thank Soliman Khatib for providing the mushroom data used in the illustrative example. K.M. would like to thank Henry Ehlers and Mara Meisenburg for their valuable feedback on the project.

Contributor Information

Kevin Mildau, Bioinformatics Group, Department of Plant Sciences, Wageningen University & Research, Radix Building, Droevendaalsesteeg 1, Wageningen, 6708PB, the Netherlands; Department of Analytical Chemistry, University of Vienna, Vienna 1090, Austria; Doctoral School in Chemistry (DOSCHEM), University of Vienna, Vienna 1090, Austria.

Christoph Büschl, Department of Agrobiotechnology, Institute of Bioanalytics and Agro-Metabolomics, University of Natural Resources and Life Sciences, Konrad-Lorenz-Straße, Lower Austria 3430, Austria.

Jürgen Zanghellini, Department of Analytical Chemistry, University of Vienna, Vienna 1090, Austria.

Justin J J van der Hooft, Bioinformatics Group, Department of Plant Sciences, Wageningen University & Research, Radix Building, Droevendaalsesteeg 1, Wageningen, 6708PB, the Netherlands; Department of Biochemistry, University of Johannesburg, Johannesburg, Gauteng Province 2006, South Africa.

Author contributions

K.M. and J.J.J.v.d.H. conceived the tool. K.M. implemented code for the tool, wrote the initial manuscript, and implemented the illustrative example. J.J.J.v.d.H, C.B., and J.Z. are part of the supervision team of K.M. and reviewed and edited the manuscript.

Supplementary data

Supplementary data are available at Bioinformatics online.

Conflict of interest

J.J.J.v.d.H. is member of the Scientific Advisory Board of NAICONS Srl., Milano, Italy, and consults for Corteva Agriscience, Indianapolis, IN, USA. All other authors declare to have no competing interests.

Funding

None declared.

References

  1. Beniddir MA, Kang KB, Genta-Jouve G. et al. Advances in decomposing complex metabolite mixtures using substructure- and network-based computational metabolomics approaches. Nat Prod Rep 2021;38:1967–93. 10.1039/d1np00023c [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Chong J, Liu P, Zhou G. et al. Using microbiomeanalyst for comprehensive statistical, functional, and meta-analysis of microbiome data. Nat Protoc 2020;15:799–821. 10.1038/s41596-019-0264-1 [DOI] [PubMed] [Google Scholar]
  3. de Jonge N, Joas D, Truong L-J. et al. Reliable cross-ion mode chemical similarity prediction between ms2spectra. bioRxiv, 2024.
  4. de Jonge NF, Mildau K, Meijer D. et al. Good practices and recommendations for using and benchmarking computational metabolomics metabolite annotation tools. Metabolomics 2022;18:103. 10.1007/s11306-022-01963-y [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Djoumbou Feunang Y, Eisner R, Knox C. et al. Classyfire: automated chemical classification with a comprehensive, computable taxonomy. J Cheminform 2016;8:1–20. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Dork M, Carpendale S, Williamson C.. Visualizing explicit and implicit relations of complex information spaces. Inf Vis 2011;11:5–21. 10.1177/1473871611425872 [DOI] [Google Scholar]
  7. Fridley BL, Jenkins GD, Biernacka JM.. Self-contained gene-set analysis of expression data: an evaluation of existing and novel methods. PLoS One 2010;5:e12693. 10.1371/journal.pone.0012693 [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Gelman A, Hennig C.. Beyond subjective and objective in statistics. J R Stat Soc Ser A Stat Soc 2017;180:967–1033. [Google Scholar]
  9. Goeman J, Oosting J. Globaltest: testing association of a group of genes with a clinical variable. Bioinformatics 2023;20:93–9. [Google Scholar]
  10. Goeman JJ, Bühlmann P.. Analyzing gene expression data in terms of gene sets: methodological issues. Bioinformatics 2007;23:980–7. 10.1093/bioinformatics/btm051 [DOI] [PubMed] [Google Scholar]
  11. Goeman JJ, van de Geer SA, de Kort F. et al. A global test for groups of genes: testing association with a clinical outcome. Bioinformatics 2004;20:93–9. 10.1093/bioinformatics/btg382 [DOI] [PubMed] [Google Scholar]
  12. Gove R, Cadalzo L, Leiby N. et al. New guidance for using t-sne: alternative defaults, hyperparameter selection automation, and comparative evaluation. Vis Inf 2022;6:87–97. 10.1016/j.visinf.2022.04.003 [DOI] [Google Scholar]
  13. Huber F, Verhoeven S, Meijer C. et al. matchms – processing and similarity evaluation of mass spectrometry data. JOSS 2020;5:2411. 10.21105/joss.02411 [DOI] [Google Scholar]
  14. Huber F, Ridder L, Verhoeven S. et al. Spec2vec: improved mass spectral similarity scoring through learning of structural relationships. PLoS Comput Biol 2021a;17:e1008724. 10.1371/journal.pcbi.1008724 [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Huber F, van der Burg S, van der Hooft JJJ. et al. Ms2deepscore: a novel deep learning similarity measure to compare tandem mass spectra. J Cheminform 2021b;13:84. 10.1186/s13321-021-00558-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Jain AK. Data clustering: 50 years beyond k-means. Pattern Recognit Lett 2010;31:651–66. 10.1016/j.patrec.2009.09.011 [DOI] [Google Scholar]
  17. Khatib S, Pereman I, Kostanda E. et al. Olive mill solid waste induces beneficial mushroom-specialized metabolite diversity: a computational metabolomics study. bioRxiv, 10.1101/2024.02.09.579616, 2024, preprint: not peer reviewed. [DOI]
  18. Kim HW, Wang M, Leber CA. et al. Npclassifier: a deep neural network-based structural classification tool for natural products. J Nat Prod 2021;84:2795–807. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Lause J, Kobak D, Berens P. The art of seeing the elephant in the room: 2D embeddings of single-cell data do make sense. PLoS Comput Biol 2024;20:e1012403. [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Maaten L. Visualizing data using t-sne. J Mach Learn Res 2008;9:2579. [Google Scholar]
  21. Maleki F, Ovens K, McQuillan I. et al. Size matters: how sample size affects the reproducibility and specificity of gene set analysis. Hum Genomics 2019;13:42. 10.1186/s40246-019-0226-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Maleki F, Ovens K, Hogan DJ. et al. Gene set analysis: challenges, opportunities, and future research. Front Genet 2020;11:654. 10.3389/fgene.2020.00654 [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. McLuskey K, Wandy J, Vincent I. et al. Ranking metabolite sets by their activity levels. Metabolites 2021;11:103. 10.3390/metabo11020103 [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Mildau K, Ehlers H, Oesterle I. et al. Tailored mass spectral data exploration using the specxplore interactive dashboard. Anal Chem 2024;96:5798–806. 10.1021/acs.analchem.3c04444 [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Nothias L-F, Petras D, Schmid R. et al. Feature-based molecular networking in the gnps analysis environment. Nat Methods 2020;17:905–8. 10.1038/s41592-020-0933-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Olivon F, Elie N, Grelier G. et al. Metgem software for the generation of molecular networks based on the t-sne algorithm. Anal Chem 2018;90:13900–8. 10.1021/acs.analchem.8b03099 [DOI] [PubMed] [Google Scholar]
  27. Pakkir Shah AK, Walter A, Ottosson F. et al. Statistical analysis of feature-based molecular networking results from non-targeted metabolomics data. Nat Protoc 2024. [DOI] [PubMed] [Google Scholar]
  28. Pang Z, Chong J, Zhou G. et al. Metaboanalyst 5.0: narrowing the gap between raw spectra and functional insights. Nucleic Acids Res 2021;49:W388–96. 10.1093/nar/gkab382 [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Plotly-Technologies-Inc. Collaborative data science. 2015.
  30. Rosato A, Tenori L, Cascante M. et al. From correlation to causation: analysis of metabolomics data using systems biology approaches. Metabolomics 2018;14:37. 10.1007/s11306-018-1335-y [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Schmid R, Heuckeroth S, Korf A. et al. Integrative analysis of multimodal mass spectrometry data in mzmine 3. Nat Biotechnol 2023;41:447–9. 10.1038/s41587-023-01690-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Schubert E, Lenssen L.. Fast k-medoids clustering in rust and python. JOSS 2022;7:4183. 10.21105/joss.04183 [DOI] [Google Scholar]
  33. Schubert E, Rousseeuw PJ.. Fast and eager k-medoids clustering: O (k) runtime improvement of the pam, clara, and clarans algorithms. Inf Syst 2021;101:101804. 10.1016/j.is.2021.101804 [DOI] [Google Scholar]
  34. Steegen S, Tuerlinckx F, Gelman A. et al. Increasing transparency through a multiverse analysis. Perspect Psychol Sci 2016;11:702–12. [DOI] [PubMed] [Google Scholar]
  35. Thompson WH, Wright J, Bissett PG.. Open exploration. Elife 2020;9:e52157. 10.7554/elife.52157 [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Tibshirani R, Walther G, Hastie T.. Estimating the number of clusters in a data set via the gap statistic. J R Stat Soc Ser B Stat Methodol 2001;63:411–23. 10.1111/1467-9868.00293 [DOI] [Google Scholar]
  37. Tomfohr J, Lu J, Kepler TB.. Pathway level analysis of gene expression using singular value decomposition. BMC Bioinformatics 2005;6:225. 10.1186/1471-2105-6-225 [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Tukey JW. We need both exploratory and confirmatory. Am Stat 1980;34:23–5. 10.2307/2682991 [DOI] [Google Scholar]
  39. VISJS Community. A dynamic, browser based visualization library. 2024.
  40. Watrous J, Roach P, Alexandrov T. et al. Mass spectral molecular networking of living microbial colonies. Proc Natl Acad Sci USA 2012;109:E1743–52. 10.1073/pnas.1203689109 [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Wicherts JM, Veldkamp CLS, Augusteijn HEM. et al. Degrees of freedom in planning, running, analyzing, and reporting psychological studies: a checklist to avoid p-hacking. Front Psychol 2016;7:1832. 10.3389/fpsyg.2016.01832 [DOI] [PMC free article] [PubMed] [Google Scholar]
  42. Wolfender J-L, Nuzillard J-M, van der Hooft JJJ. et al. Accelerating metabolite identification in natural product research: toward an ideal combination of liquid chromatography—high-resolution tandem mass spectrometry and nmr profiling, in silico databases, and chemometrics. Anal Chem 2018;91:704–42. 10.1021/acs.analchem.8b05112 [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

btae584_Supplementary_Data
btae584_supplementary_data.pdf (4.8MB, pdf)

Articles from Bioinformatics are provided here courtesy of Oxford University Press

RESOURCES