HMDB 5.0: What the Human Metabolome Database Contains and How to Use It
The Human Metabolome Database is a free, curated reference of small molecules found in the human body, and version 5.0 (released January 2022) is the current major release. It holds 217,920 metabolite entries. Each entry carries the following.
- chemical structure
- synonyms
- predicted and experimental physicochemical properties
- reference mass spectra
- known biological pathways
- reported concentrations in blood, urine and other biofluids
- cross-references to ChEBI, KEGG, PubChem, DrugBank and UniProt
You use it for two things in practice: turning a mass or an MS/MS spectrum into a candidate compound name, and turning a compound name into biological and reference-range context. It is a knowledgebase, not a measurement, and the distinction matters more than most users assume.
What is in HMDB 5.0, and how much of it was measured
The headline count of 217,920 entries hides an important split. A minority of entries are compounds that have been detected and quantified in a human sample by someone, somewhere, with a citation. The majority are predicted or expected metabolites. Many of them are computationally enumerated derivatives, lipids, and microbial or xenobiotic transformation products. HMDB tags each entry with a status, and the ones you care about for annotation are “detected and quantified” and “detected but not quantified”.
This structure is deliberate and dates back to the original releases, which framed HMDB as a knowledgebase linking chemical, clinical, and biochemical data for each compound rather than a spectral library alone.1 The project grew out of the Human Metabolome Project, which set out to catalogue and quantify the measurable human metabolome and to build the supporting software and reference spectra around it.2 Version 5.0’s notable additions are a much larger set of predicted MS/MS and NMR spectra, expanded lipid and microbial metabolite coverage, and ChemFOnt, a chemical functional ontology that assigns machine-readable functional roles to compounds.
For work on microbially derived metabolites specifically, the same group maintains MiMeDB, the Human Microbial Metabolome Database. It separates gut-derived compounds and links them to the taxa and enzymes that produce them.3 If your question is about the gut, start there and use HMDB for the human-side chemistry.
Which files to download
Do not scrape the website. Everything useful is in the downloads section at hmdb.ca as bulk files.
The primary file is hmdb_metabolites.zip, which unzips to a single XML document of a few gigabytes. Parse it with a streaming parser, not a DOM parser, or you will exhaust memory. In Python:
import xml.etree.ElementTree as ET
NS = "{http://www.hmdb.ca}"
for event, elem in ET.iterparse("hmdb_metabolites.xml", events=("end",)):
if elem.tag == f"{NS}metabolite":
accession = elem.findtext(f"{NS}accession")
mono = elem.findtext(f"{NS}monisotopic_molecular_weight")
inchikey = elem.findtext(f"{NS}inchikey")
status = elem.findtext(f"{NS}status")
# emit row, then free memory
elem.clear()
The elem.clear() call is not optional. Without it you hold the whole tree.
For mass-based search you only need a flat table that holds the following.
- accession
- name
- formula
- monoisotopic mass
- InChIKey
- SMILES
- status
- the biofluid concentration blocks
We build that once into SQLite or Parquet and index on rounded monoisotopic mass. Query times drop from minutes to milliseconds.
The structure files (structures.sdf) are what you want for cheminformatics. Load them with RDKit and compute your own descriptors rather than trusting the predicted properties in the XML, which come from tools of varying vintage.
Experimental MS/MS spectra ship as a separate archive of individual XML files, one per spectrum, with collision energy and adduct in the metadata. Convert them to MSP or MGF before feeding a matching tool such as MZmine or matchms. The predicted spectra archive is far larger and should be kept in a separate index, because mixing predicted and experimental spectra in one search silently inflates your apparent identification rate.
How to use it for annotation without fooling yourself
Mass alone does not identify a metabolite. At 5 ppm tolerance on a 300 Da monoisotopic mass, your window is ±0.0015 Da, and HMDB will typically return several to dozens of formula-compatible candidates. Isomers are indistinguishable by mass at any instrument resolution.
So treat HMDB lookup as candidate generation and then apply orthogonal evidence. Retention time is the cheapest and most informative addition. Recent work has shown that combining mass-to-charge ratio with retention time in a machine learning classifier meaningfully improves lipid class assignment in untargeted data.4 MS/MS fragmentation is the next layer, and a match to an experimental HMDB spectrum at the same adduct and a comparable collision energy is worth far more than a match to a predicted one.
Beyond per-feature evidence, the strongest current approaches use network context. A two-layer method that combines a knowledge layer of known biochemical relationships with a data layer of observed co-behaviour among features improves annotation accuracy over spectrum matching alone.5 The intuition is simple. A candidate that sits one enzymatic step away from another confidently identified feature in the same sample is more plausible than an isolated structural match. Structure-based function prediction with graph neural networks is a complementary direction, assigning likely functional roles to compounds whose biology is uncharacterised.6
We would run a pipeline in this order. First a formula-constrained mass search against detected-status HMDB entries, then MS/MS matching against experimental spectra. After that comes retention-time consistency against your own in-house standards, then network plausibility. Report a confidence level per feature and never let a level-3 putative annotation graduate to a compound name in a figure legend without a standard run on the same instrument.
Where HMDB will mislead you
Three failure modes recur.
First, the reference concentration ranges. HMDB aggregates published concentrations from many studies, platforms, and cohorts. The ranges are wide, the units vary, and the age and sex breakdowns are sparse or absent for most compounds. They are useful for order-of-magnitude sanity checking. They are not suited to deciding whether your own value is abnormal. Clinical reference intervals come from the laboratory that ran your sample, on its own platform. Any interpretation of whether your number means something about your health is a conversation with a physician.
Second, predicted entries polluting search results. If you search the full 217,920 entries with a 10 ppm window, you will find a name for nearly every feature, including noise peaks and in-source fragments. Filter on status before you believe anything.
Third, staleness of cross-references. KEGG and ChEBI identifiers in an XML dump from 2022 will include retired accessions. Re-map through UniChem or InChIKey rather than trusting stored IDs.
Pathway and network context
Once you have identifications, the interesting work is placing them in metabolic context. Metscape, a Cytoscape plug-in, builds compound-reaction-enzyme-gene networks around a metabolite list and remains a practical way to view metabolomic data against human metabolic networks.7 For studies that also generated shotgun metagenomic data, PredCMB predicts likely changes in microbial metabolites from gene-level metagenome shifts, which gives you a hypothesis to test against measured metabolites.8 Harmonized cross-study resources for gut microbiome-metabolome associations are now emerging as well, which helps when a single cohort is too small to support a finding.9
HMDB’s most direct clinical use has been in inborn errors of metabolism, where a specific metabolite accumulation maps to a specific enzyme deficiency, and the database collects those marker compounds with their associated disorders.10 That is the clearest case of a metabolite carrying a definite interpretation. Most metabolites are not like that.
Questions people also ask
What is the human metabolome?
The human metabolome is the complete set of small-molecule chemicals present in the human body: endogenous products of our own enzymes, compounds produced by the gut microbiome, and everything absorbed from food, drugs, and the environment. Estimates of its size depend on where you draw the line on food-derived and microbial compounds, which is why HMDB’s count has grown by roughly two orders of magnitude across its releases.
What is the 80% rule in metabolomics data preprocessing?
The 80% rule is a filter for missing values: keep a metabolite feature only if it was detected in at least 80% of the samples in at least one study group. It removes features that are absent or below detection in most samples, where imputation would invent most of the data. Apply it per group rather than across the whole dataset, otherwise you discard features that are genuinely present in one condition and absent in another.
What is the best database for metabolomics research?
There is no single best one, and we use three. We use HMDB for human biological and concentration context, and MassBank or the NIST library for experimental spectral matching. For lipids we use LIPID MAPS, whose nomenclature and isomer handling are better served there. For gut-derived compounds, MiMeDB is the more appropriate starting point.3
Is HMDB free to use?
Yes for academic and personal use, with bulk downloads available directly from hmdb.ca. Commercial use requires a licence, so check the terms before you embed the data in a product.
Footnotes
-
D. S. Wishart, C. Knox, A. C. Guo, et al. HMDB: a knowledgebase for the human metabolome. Nucleic Acids Research, 2009. https://doi.org/10.1093/nar/gkn810 ↩
-
David Wishart. Systems Biology Resources Arising from the Human Metabolome Project. Genetics Meets Metabolomics, 2012. https://doi.org/10.1007/978-1-4614-1689-0_11 ↩
-
David S Wishart, Eponine Oler, Harrison Peters, et al. MiMeDB: the Human Microbial Metabolome Database. Nucleic Acids Research, 2022. https://doi.org/10.1093/nar/gkac868 ↩ ↩2
-
Christelle Colin-Leitzinger, Yonatan Ayalew Mekonnen, Isis Narvaez-Bandera, et al. A machine learning framework for classifying lipids in untargeted metabolomics using mass-to-charge ratios and retention times. Metabolomics, 2025. https://doi.org/10.1007/s11306-025-02343-y ↩
-
Haosong Zhang, Xinhao Zeng, Yandong Yin, et al. Knowledge and data-driven two-layer networking for accurate metabolite annotation in untargeted metabolomics. Nature Communications, 2025. https://doi.org/10.1038/s41467-025-63536-6 ↩
-
Tancredi Cogne, Mariam Ait Oumelloul, Ali Saadat, et al. Structure-based metabolite function prediction using graph neural networks. Bioinformatics Advances, 2024. https://doi.org/10.1093/bioadv/vbaf174 ↩
-
Jing Gao, V. Glenn Tarcea, Alla Karnovsky, et al. Metscape: a Cytoscape plug-in for visualizing and interpreting metabolomic data in the context of human metabolic networks. Bioinformatics, 2010. https://doi.org/10.1093/bioinformatics/btq048 ↩
-
Jungyong Ji, Sungwon Jung. PredCMB: predicting changes in microbial metabolites based on the gene–metabolite network analysis of shotgun metagenome data. Bioinformatics, 2024. https://doi.org/10.1093/bioinformatics/btaf020 ↩
-
Shivangi Verma, Nalin Arora, Chandra Prakash Ajay, et al. HuMMANet: A Harmonized Cross-Study Resource for Integrative Analysis of Human Gut Microbiome–Metabolome Associations. 2026. https://doi.org/10.64898/2026.08.24.746727 ↩
-
Rupasri Mandal, Danuta Chamot, David S. Wishart. The role of the Human Metabolome Database in inborn errors of metabolism. Journal of Inherited Metabolic Disease, 2018. https://doi.org/10.1007/s10545-018-0137-8 ↩
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.