Metabolomics data repositories and how to pull data from them
If you want public metabolomics data, there are two repositories worth learning and a handful worth knowing about. Metabolomics Workbench, run by the NIH Common Fund, holds several thousand studies under ST accessions and exposes everything through a REST API that returns JSON 1. MetaboLights, run at EMBL-EBI, holds studies under MTBLS accessions in ISA-Tab format with raw instrument files on a public FTP server 2. Workbench is easier to query programmatically. MetaboLights is usually better if you want the raw files and the experimental metadata that go with them.
This page covers how to get data out of each and what the files contain. It also covers how to reprocess raw spectra and the specific ways public metabolomics data will disappoint you.
What each repository holds
Metabolomics Workbench stores study metadata and the processed metabolite intensity matrix. For most recent studies it also stores raw data files. The native metadata format is mwTab, a flat text format with sectioned headers such as #SUBJECT, #COLLECTION, #TREATMENT, #ANALYSIS, and #MS_METABOLITE_DATA. The Workbench also maintains RefMet, a standardized metabolite naming system that maps the ragged vocabulary of submitted data onto a consistent hierarchy 1. RefMet is the single most useful thing the Workbench provides, and we use it even when the underlying data came from somewhere else.
MetaboLights stores studies as ISA-Tab. A study holds one i_Investigation.txt file and one or more s_*.txt sample tables. It also holds a_*.txt assay tables that link each sample to each raw file, and m_*.tsv metabolite assignment files. The assay tables record chromatography column, ionization mode, and instrument model per run. That level of per-file metadata is what makes MetaboLights studies reprocessable 2.
Beyond these two, GNPS/MassIVE holds untargeted LC-MS/MS spectra and molecular networking results. A newer entrant, MB-POST, is built specifically around reanalysis rather than archival deposition. The repository is structured so that stored data can be re-searched under new parameters and new spectral libraries 3. The general direction of the field is toward repositories designed for computation over stored data rather than storage alone 4.
Pulling data from Metabolomics Workbench
The REST API follows a fixed grammar: https://www.metabolomicsworkbench.org/rest/<context>/<input_item>/<input_value>/<output_item>/<format>. The contexts you will use are study, refmet, compound, and moverz.
# Study summary
curl -s https://www.metabolomicsworkbench.org/rest/study/study_id/ST000061/summary
# Analysis records (one per MS or NMR run configuration)
curl -s https://www.metabolomicsworkbench.org/rest/study/study_id/ST000061/analysis
# Named metabolites for one analysis
curl -s https://www.metabolomicsworkbench.org/rest/study/analysis_id/AN000095/metabolites
# The intensity matrix
curl -s https://www.metabolomicsworkbench.org/rest/study/analysis_id/AN000095/data
Note the distinction between a study ID and an analysis ID. A single ST study often carries four or more AN analyses: positive and negative ionization, reversed-phase and HILIC columns, sometimes targeted panels alongside untargeted runs. Intensities are not comparable across analyses. Merging them into one matrix without accounting for that is the most common mistake we see in reuse of Workbench data.
For bulk work, install the mwtab Python package, which parses mwTab files into nested dictionaries and validates them against the schema.
pip install mwtab
python -c "import mwtab; [print(f['STUDY']['STUDY_TITLE']) for f in mwtab.read_files('ST000061')]"
RefMet name resolution is one call per name, so cache it. Map everything to a RefMet name plus an InChIKey before you attempt any cross-study comparison.
Pulling data from MetaboLights
MetaboLights study files sit at a predictable FTP path, which makes mirroring straightforward:
wget -r -np -nH --cut-dirs=6 \
ftp://ftp.ebi.ac.uk/pub/databases/metabolights/studies/public/MTBLS1/
Parse the ISA-Tab tables first, before downloading anything large. The assay file tells you the file name for each run and the parameters used. Use it to select only the runs you need. A single MetaboLights study can exceed 100 GB in vendor raw format, and most of that will be irrelevant to your question.
The m_*.tsv metabolite assignment file is the processed result: feature identifiers, m/z, retention time, putative annotation, and per-sample intensities. If you trust the original authors’ processing, start here. If you do not, go to the raw files.
Reprocessing raw files
Vendor formats (.raw for Thermo, .d for Agilent and Bruker, .wiff for SCIEX) need conversion to the open mzML format first. ProteoWizard’s msconvert does this:
msconvert *.raw --mzML --zlib --64 \
--filter "peakPicking vendor msLevel=1-" \
--filter "titleMaker <RunId>.<ScanNumber>"
Use vendor peak picking where the vendor DLLs are available. The centroiding algorithms shipped by instrument makers are better than the generic alternatives, and the difference shows up in feature counts.
For untargeted LC-MS, we use XCMS in R through the Spectra backend. A starting parameter set for a 15-minute UPLC run on a high-resolution Orbitrap:
cwp <- CentWaveParam(ppm = 15, peakwidth = c(4, 25),
snthresh = 10, prefilter = c(3, 500),
noise = 1000, mzdiff = -0.001)
xd <- findChromPeaks(data, param = cwp)
xd <- adjustRtime(xd, param = ObiwarpParam(binSize = 0.6))
pdp <- PeakDensityParam(sampleGroups = groups, bw = 5,
minFraction = 0.5, binSize = 0.025)
xd <- groupChromPeaks(xd, param = pdp)
xd <- fillChromPeaks(xd, param = ChromPeakAreaParam())
Tune ppm to your instrument’s measured mass accuracy. Do not take the value from its specification sheet. Tune peakwidth from the observed full width at half maximum of a few known peaks. Leaving these at defaults is the main cause of the split and merged peaks that show up as duplicate features downstream. The broader toolchain for peak detection, alignment, and annotation has been surveyed repeatedly. The choices you make at this stage propagate through everything after 5.
Where public metabolomics data fails
Annotation confidence is the first problem. The Metabolomics Standards Initiative defines four levels, and Level 1 requires matching both retention time and MS/MS fragmentation against an authentic standard run on the same instrument. Most repository annotations are Level 2 or 3, meaning a spectral library match without local retention time confirmation. Treat any repository metabolite name as a hypothesis.
Batch and site effects are the second. Signal drift within a run sequence is routine, and intensities from different instruments are not on a common scale. Studies that deposited pooled quality control samples let you correct for this with QC-based signal correction. Studies that did not leave you guessing. Check the sample table for QC injections before committing to a dataset.
Missing values are the third. Untargeted matrices are typically 20 to 50 percent missing, and missingness is informative because low-abundance features fall below detection. Imputing with the column mean destroys that structure. We impute left-censored values with a small fraction of the per-feature minimum, or model them explicitly.
Metadata quality is the fourth and the most common reason a reanalysis dies. Sample timing, fasting state, and storage conditions are often absent. Reuse of public metabolomics data works well when the original deposit was complete, and it fails silently when it was not 6.
Multiple omics and repeated measures
Metabolomics is most informative next to other layers. Pathway-level integration of metabolite abundance with transcript and protein data recovers biology that a single layer misses, and automated frameworks for doing that integration are now mature 7. Integration with DNA methylation has been used to relate metabolic signatures to regulatory state in tumor cohorts 8. Combined genetic and metabolomic analyses have been used to nominate candidate biomarkers in type 2 diabetes cohorts 9. Those are research findings about groups of people, and translating any of them to an individual is a question for a clinician.
If you have repeated measurements on yourself, the analysis problem changes shape. Cross-sectional tools treat each sample as independent, which throws away the within-person structure that makes longitudinal data valuable. The MeTime R package implements mixed-effects modeling for longitudinal metabolomics with reproducible preprocessing and time-course visualization 10. Its central idea generalizes: fit a random intercept per subject and test the time term, so that between-person variation does not swamp within-person change.
Questions people also ask
Which repository should I start with?
Metabolomics Workbench if you want processed matrices fast and a clean API. MetaboLights if you need raw files with trustworthy per-run metadata 1 2.
Can I compare metabolite intensities across two different studies?
Only after normalizing within each study and mapping names to a shared identifier such as RefMet or InChIKey. Absolute intensities from untargeted LC-MS are instrument-specific and should not be pooled directly.
What file format will I get?
Processed data arrives as tab-delimited text or mwTab from Workbench, and as ISA-Tab plus m_*.tsv from MetaboLights. Raw data arrives as vendor formats or mzML.
Are repositories being built for reanalysis rather than archiving?
Yes. MB-POST is designed so deposited data can be re-searched with updated libraries and parameters 3, and the field is moving toward computation directly over stored public data 4.
Can repository data tell me anything about my own health?
No. Public datasets describe cohorts under specific protocols, and mapping a group-level association onto one person requires clinical context. Bring any result you find to a physician.
Footnotes
-
Manish Sud, Eoin Fahy, Dawn Cotter, et al. Metabolomics Workbench: An international repository for metabolomics data and metadata, metabolite standards, protocols, tutorials and training, and analysis tools. Nucleic Acids Research, 2015. https://doi.org/10.1093/nar/gkv1042 ↩ ↩2 ↩3
-
Christoph Steinbeck, Pablo Conesa, Kenneth Haug, et al. MetaboLights: towards a new COSMOS of metabolomics data management. Metabolomics, 2012. https://doi.org/10.1007/s11306-012-0462-0 ↩ ↩2 ↩3
-
Yushi Takahashi, Akiyasu C. Yoshizawa, Takato Kiuchi, et al. Reviving Data Potential─The Reanalysis-Driven Metabolomics Data Repository MB-POST. Analytical Chemistry, 2026. https://doi.org/10.1021/acs.analchem.6c00697 ↩ ↩2
-
Ewy A Mathé, Justin JJ van der Hooft, Haley Chatelaine, et al. Computational metabolomics at scale: from open data to insight. Current Opinion in Biotechnology, 2026. https://doi.org/10.1016/j.copbio.2026.103557 ↩ ↩2
-
Masahiro Sugimoto, Masato Kawakami, Martin Robert, et al. Bioinformatics Tools for Mass Spectroscopy-Based Metabolomic Data Processing and Analysis. Current Bioinformatics, 2012. https://doi.org/10.2174/157489312799304431 ↩
-
Michael Witting. (Re‐)use and (re‐)analysis of publicly available metabolomics data. PROTEOMICS, 2023. https://doi.org/10.1002/pmic.202300032 ↩
-
Tao Huan, Amelia Palermo, Julijana Ivanisevic, et al. Autonomous Multimodal Metabolomics Data Integration for Comprehensive Pathway Analysis and Systems Biology. Analytical Chemistry, 2018. https://doi.org/10.1021/acs.analchem.8b00875 ↩
-
Fadhl M. Alakwaa, Masha G. Savelieff. Bioinformatics Analysis of Metabolomics Data Unveils Association of Metabolic Signatures with Methylation in Breast Cancer. Journal of Proteome Research, 2019. https://doi.org/10.1021/acs.jproteome.9b00755 ↩
-
Hajra Yaqoob, Zunaira Abid Khan, Attique Ahmed, et al. Integrated Bioinformatics Analysis Identifies Novel Genetic and Metabolomic Risk Biomarkers for Type 2 Diabetes in Humans. Journal of Pharma and Biomedics, 2026. https://doi.org/10.56810/jpbm.004.02.0200 ↩
-
Bharadwaj Marella, Patrick Weinisch, Lara Vehovec, et al. MeTime: an R package for reproducible longitudinal metabolomics data analysis. Bioinformatics Advances, 2026. https://doi.org/10.1093/bioadv/vbag252 ↩
Pothos builds longitudinal molecular profiles of individuals: whole-genome sequencing, single-cell RNA sequencing, and plasma proteomics, delivered as a detailed report, every data file, and stored cells. Build your profile.