How to Buy Metabolomics Services and Analyze the Data Yourself

If you want metabolomics run on your own samples, you have three practical options. The first is a targeted panel from a clinical or research lab, which gives roughly 100 to 600 named, absolutely quantified metabolites for $300 to $1,500 per sample. The second is an untargeted LC-MS discovery run from a commercial provider like Metabolon or a university core. That gives thousands of features, several hundred named, for $300 to $800 per sample at volume and more for small orders. The third is an NMR panel of about 250 quantified analytes including lipoprotein subfractions, which is cheap per sample but limited to abundant species. For a single person trying to understand their own biology, we recommend a targeted quantitative panel first and untargeted only if you have a specific question and multiple timepoints, because an untargeted run on one sample from one person produces a feature table you cannot interpret. The interpretability of metabolomics comes almost entirely from having a comparison, whether that is a reference population, a paired timepoint, or a perturbation.

What metabolomics measures and what the technology limits

Metabolomics is the measurement of small molecules (conventionally under 1,500 Da) in a biological sample. Those molecules are the following.

• amino acids
• organic acids
• acylcarnitines
• bile acids
• sugars
• nucleotides
• lipids
• the steroid and neurotransmitter precursors

Unlike the genome, the metabolome is a state variable. It shifts with the last meal, sleep, exercise in the preceding day, menstrual cycle phase, and time of collection, which is why a provider’s collection protocol matters more than their mass spectrometer.

No single platform sees the whole metabolome. Chemical diversity across the metabolome spans many orders of magnitude in polarity and concentration, so coverage is a function of how many orthogonal separations and ionization modes you run ¹. A serious untargeted service will run at least four injections per sample. These are reverse-phase C18 in positive mode, C18 in negative mode, HILIC in positive mode, and often a lipid-specific method. HILIC is hydrophilic interaction chromatography, for polar compounds that do not retain on C18. GC-MS after derivatization adds sugars, sugar phosphates, and small organic acids that ionize poorly by electrospray. If a quote mentions a single method, you are buying partial coverage.

The hard limit on untargeted work is annotation. A typical human plasma untargeted run yields 10,000 to 30,000 aligned features, of which a few thousand are real, non-redundant compounds after removing isotopes, adducts, and in-source fragments, and only several hundred to about 1,500 can be confidently named against spectral libraries. Everything else is a mass and a retention time. Commercial providers that charge a premium are largely selling their in-house library of authentic standards run on their own instruments, which is the part you cannot reproduce with open tooling.

Targeted versus untargeted, and which to buy

Targeted assays use multiple reaction monitoring on a triple-quadrupole instrument with stable-isotope-labeled internal standards for each analyte class. You get absolute concentrations in µM or ng/mL, coefficients of variation typically under 10 to 15% for most analytes, and numbers you can compare against a published reference interval or a later draw of your own. That comparability is the reason we start here.

Untargeted metabolomics means acquiring everything the instrument can detect without pre-specifying compounds, then finding features that differ between groups. It is a hypothesis-generating design, and it earns its cost when you have contrast. That contrast can be a case-control study or a treated-versus-untreated comparison. For an individual, it can be a dense time series across an intervention. In multi-omics work it is most useful as the layer that tells you whether a transcriptional change has a downstream chemical consequence. Studies that pair RNA sequencing with untargeted metabolomics routinely find that only a fraction of differentially expressed genes have a matching metabolite shift, which is precisely the discrimination you want ². The same pattern holds in stress-response work where transcriptomic and metabolomic layers are integrated to identify which pathways carry flux rather than merely regulation ³.

For questions about human disease mechanism, be careful about causal reading. Association between a serum metabolite and a condition is common and mostly non-causal, which is why genetic instrumental-variable designs have become the standard filter in that literature ⁴. On your own data, treat an unusual metabolite as a signal to repeat the measurement. If it persists, bring it to a physician who can order a validated clinical assay. Interpreting a metabolic abnormality in a person is clinical work, and the multi-omics literature is explicit that a clinician has to sit in that loop ⁵.

What to demand in the deliverable

Ask for these before you sign anything, because the difference between a usable and unusable deliverable is entirely in the file list.

Raw files in an open format. Vendor formats such as Thermo .raw, Agilent .d and Sciex .wiff are convertible with ProteoWizard msconvert. Even so, insist the provider either ships mzML or permits conversion. A reasonable command is msconvert *.raw --mzML --64 --filter "peakPicking vendor msLevel=1-2" --filter "zeroSamples removeExtra", using vendor centroiding because the vendor algorithms handle profile data better than the open reimplementations.

The processed feature table with retention time and m/z. A list of named compounds alone is not enough. Without m/z and RT you cannot re-annotate later or match features across batches.

Annotation confidence levels per compound, following the Metabolomics Standards Initiative convention. Level 1 means matched to an authentic standard on the same instrument for both MS/MS spectrum and retention time. Level 2 means library spectral match without an in-house standard. Level 3 is a compound class, and level 4 is an unknown with a mass. Any provider that hands you a flat list of names without levels is overstating what they know.

QC samples and their raw data. You want pooled QC injections every six to ten samples, extraction blanks, and a set of internal standards, so you can compute per-feature CV in the QCs and drop anything above about 30%. Also ask for the injection order and batch assignment, since that is the only way to correct signal drift.

MS/MS spectra for the features that matter to you, as an MGF or in the mzML at MS2 level. Without fragmentation data, structural annotation of anything not already in the provider’s library is impossible.

Processing the data yourself

If you receive raw mzML, the R pipeline we would use starts with xcms for peak detection and alignment. Next comes CAMERA or MsFeatures for grouping isotopes and adducts, then MetaboAnalystR or plain limma for statistics. For high-resolution LC-MS, run CentWaveParam(ppm = 10, peakwidth = c(5, 30), snthresh = 10, prefilter = c(3, 500)), tightening ppm to 5 on an Orbitrap or timsTOF. Align with ObiwarpParam(binSize = 0.6), then correspondence with PeakDensityParam(minFraction = 0.5, bw = 5, binSize = 0.025), and finish with fillChromPeaks() so that missing values are integrated from the raw signal rather than imputed.

The 80% rule applies at the filtering step: keep a feature only if it is detected in at least 80% of samples within at least one biological group. This exists because metabolomics missingness is largely left-censored, meaning the compound is below the detection limit rather than randomly absent, and imputing across a feature present in a quarter of samples manufactures structure. After that filter, replace remaining missing values with a small value such as one-fifth of the minimum observed for that feature, normalize by pooled-QC regression or probabilistic quotient normalization, log-transform, and Pareto-scale before any multivariate model.

For pathway-level interpretation, use MetaboAnalyst’s enrichment and pathway modules. Read the results with the knowledge that mapping metabolites to pathways is noisy, because a single compound like pyruvate or glutamate sits in dozens of reactions. Curated databases differ in scope and in how they resolve compound identifiers: MetaCyc provides experimentally supported pathways with enzyme annotation across thousands of organisms ⁶, while WikiPathways was built specifically to link metabolite identifiers to gene-centric pathway content so metabolomics and transcriptomics can be viewed in one diagram ⁷. Map to InChIKey or ChEBI rather than compound names, since name-matching failures silently drop 20 to 40% of a feature list.

For putative identification of an unknown, the workflow starts with mass to formula. Use a 7.5 ppm window and apply the seven golden rules for formula plausibility. Next comes formula to candidate structures in HMDB and PubChem, then in silico fragmentation with SIRIUS + CSI:FingerID against your MS2 spectrum. That yields a level 2 or 3 annotation. Getting to level 1 requires buying the standard and running it on the same column, which is why most reported identifications in the literature are not level 1 ⁸.

Questions people also ask

What does untargeted metabolomics mean? It means acquiring signal from every molecule the instrument can detect, without specifying a compound list in advance, then discovering which features differ across conditions. The tradeoff is relative rather than absolute quantification and the fact that most detected features remain chemically unidentified.

How much does metabolomics cost? Expect $300 to $800 per sample for commercial untargeted panels at reasonable volume, $300 to $1,500 for a targeted quantitative panel, and $150 to $400 per sample at an academic core that charges for instrument time and labor separately. Small orders of one to five samples carry setup costs that can double the per-sample price.

What are the 13 metabolic pathways? There is no canonical set of thirteen. The list people usually mean is the central pathways.

• glycolysis
• gluconeogenesis
• the pentose phosphate pathway
• glycogen synthesis and breakdown
• the TCA cycle
• oxidative phosphorylation
• fatty acid synthesis
• beta-oxidation
• ketogenesis
• the urea cycle
• amino acid catabolism
• nucleotide synthesis
• cholesterol and steroid synthesis

Curated resources catalogue thousands of pathways rather than a short list ⁶.

What are the types of metabolomics? By intent, targeted, untargeted, and flux analysis using isotope tracers. By platform, LC-MS, GC-MS, and NMR. Specialized subsets include lipidomics and exposomics, and integrated designs combine them with proteomics for broader chemical coverage of a sample ⁹.

What does metabolomics do for an individual rather than a study? On a single person it is most informative longitudinally: repeated draws under standardized conditions let you use yourself as the control and detect drift in a pathway over months. A single snapshot is mainly useful for the targeted analytes that have established reference intervals, and those readings belong in a conversation with a physician.

Footnotes

1. Simone Rochfort. Metabolomics Reviewed: A New “Omics” Platform Technology for Systems Biology and Implications for Natural Products Research. Journal of Natural Products, 2005. https://doi.org/10.1021/np050255w ↩

2. Shiping Su, Xinxin Liu, Caixia Song, et al. Multi-omics reveals glycolipid metabolism processes during the early development of largemouth bass (Micropterus salmoides). BMC Genomics, 2026. https://doi.org/10.1186/s12864-026-12692-0 ↩

3. Jiahui Jiang, Gaofeng Fan, Ruiqi Wang, et al. Multi-omics analysis of Populus simonii × P. nigra leaves under Hyphantria cunea stress. Frontiers in Plant Science, 2024. https://doi.org/10.3389/fpls.2024.1392433 ↩

4. Haohan Ye, Yuanheng Liu, Jun Tang, et al. Bidirectional Mendelian Randomization and Multi-Omics Uncover Causal Serum Metabolites and Neuro-Related Mechanistic Pathways in Acute Myeloid Leukemia. International Journal of Molecular Sciences, 2025. https://doi.org/10.3390/ijms262311307 ↩

5. Clara D. M. van Karnebeek, Saskia B. Wortmann, Maja Tarailo‐Graovac, et al. The role of the clinician in the multi‐omics era: are you ready?. Journal of Inherited Metabolic Disease, 2018. https://doi.org/10.1007/s10545-017-0128-1 ↩

6. R. Caspi, H. Foerster, C. A. Fulcher, et al. The MetaCyc Database of metabolic pathways and enzymes and the BioCyc collection of Pathway/Genome Databases. Nucleic Acids Research, 2007. https://doi.org/10.1093/nar/gkm900 ↩ ↩²

7. Denise N Slenter, Martina Kutmon, Kristina Hanspers, et al. WikiPathways: a multifaceted pathway database bridging metabolomics to other omics research. Nucleic Acids Research, 2017. https://doi.org/10.1093/nar/gkx1064 ↩

8. Akram Tayanloo-Beik, Masoumeh Sarvari, Moloud Payab, et al. OMICS insights into cancer histology; Metabolomics and proteomics approach. Clinical Biochemistry, 2020. https://doi.org/10.1016/j.clinbiochem.2020.06.008 ↩

9. Marianthi Sidira, Sofia Agriopoulou, Slim Smaoui, et al. Omics-Integrated Approach (Metabolomics, Proteomics and Lipidomics) to Assess the Quality Control of Aquatic and Seafood Products. Applied Sciences, 2024. https://doi.org/10.3390/app142210755 ↩