henhaobio
Sign in

bioRxivSystems BiologyPreprint

METASPACE: A community-populated knowledge base of spatial metabolomes in health and disease

Theodore AlexandroviD, Katja Ovchinnikova, Andrew PalmeriD, Vitaly Kovalev, Artem TarasoviD, Lachlan StuartiD, Renat Nigmetzianov, Dominik FayiD, Key METASPACE contributors, Mathieu GaudiniD, Cristina López, Marina Vetter, John G. SwalesiD, Mark T. BokhartiD, Mario KompaueriD, James S. McKenzieiD, Luca RappeziD, Dušan VeličkovićiD, Régis Lavigne, Guanshi ZhangiD, Dinaiz Thinagaran, Elisa Ruhland, Marta SansiD, Sergio TrianaiD, Denis Abu SammouriD, , Sheerin K. Shahidi-LathamiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Metabolites, lipids, and other small molecules are key constituents of tissues supporting cellular programs in health and disease. Here, we present METASPACE, a community-populated knowledge base of spatial metabolomes from imaging mass spectrometry data. METASPACE is enabled by a high-performance engine for metabolite annotation in a confidence-controlled way that makes results comparable between experiments and laboratories. By sharing their results publicly, engine users continuously populate a knowledge base of annotated spatial metabolomes in tissues currently including over 3000 datasets from human cancer cohorts, whole-body sections of animal models, and various organs. The spatial metabolomes can be visualized, explored and shared using a web app as well as accessed programmatically for large-scale analysis. By using novel computational methods inspired by natural language processing, we illustrate that METASPACE provides molecular coverage beyond the capacity of any individual laboratory and opens avenues towards comprehensive metabolite atlases on the levels of tissues and organs.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (13 Feb 2019) · CC BY 4.0

An open-source cloud engine annotates imaging mass spectrometry data at controlled false discovery rate; publicly shared results reached over 3200 datasets and 6 million metabolite images by January 2019.

Imaging mass spectrometry (imaging MS) can map metabolites and lipids across tissue sections, but the authors note that its use was held back by a lack of methods linking ion images to molecules. METASPACE runs their earlier false discovery rate (FDR) controlled annotation method as a scalable cloud engine and pools users' public results, so spatial metabolomes from different laboratories can be compared, searched and mined together.

Key results

  1. ShownThe community-populated collection held over 3200 public annotated datasets from over 140 contributors in over 70 laboratories. It covers several organisms, several organs and the major high-resolution imaging MS platforms.

    As of January 2019: over 6 million metabolite images from human, mouse, rat and other samples; MALDI, DESI and IR-MALDESI sources with Orbitrap or FTICR analyzers; 101 public datasets from seven pharma companies; stated growth of 1000 datasets a year; 10.000-25.000 monthly pageviews, not counting the team. · Knowledge base; Web app; Screening for drugs and their metabolites; Discussion

  2. ShownThe pooled data measure typical imaging MS performance: up to 10^3 annotations per dataset at FDR 10% (most often in the low hundreds) and a dynamic range of up to 5 orders of magnitude (typically around 10^3).

    Coverage was counted as annotations per dataset at FDR 10%; dynamic range as the ratio of the most to the least intense annotation in each dataset. How often a molecule is detected can also be queried: glutathione at FDR 10% returns 519 annotations from 18 laboratories (one of the most prevalent molecules), against 26 for creatinine. · Knowledge base; Figure legends

  3. ShownWith datasets modelled as tf-idf vectors of their annotated formulas, ion polarity was the metadata factor best predicted by molecular content, ahead of laboratory and organ. Group-averaged tf-idf scores gave candidate markers for brain, kidney, liver, lung and eye.

    Cohen's kappa, nearest-neighbour vs random dataset pairs: 0.95 polarity, 0.85 laboratory, 0.77 organ; the similarity network (edges at cosine > 0.35) splits by ion mode. Brain markers, from data of 26 laboratories (27 in the Figure 4 legend), included 3-O-sulfogalactosylceramide (d36:0), PE(P-38:1), PS(36:1) and GM1 gangliosides; kidney markers included L-carnitine, betaine and cholesterol sulfate. · Large-scale analysis of spatial metabolomes; Quantifying the influence of metadata parameters…; Data similarity network; Characteristic molecules

  4. SuggestedThe authors imaged an organs-mimicking section to check organ-level patterns seen in contributed mouse and rat whole-body sections: glutathione (GSH) was highest in liver and LysoPE(22:6) was highest in brain.

    Mimicking section: gelatine wells of homogenized mouse liver, heart, muscle, lung, kidney and brain. In whole-body sections GSH was also seen in salivary gland, lung, intestinal wall and testis. Oxidized glutathione (GSSG) appeared only in the liver well. At 20 μm in rat liver, GSH and GSSG had complementary patterns. Higher-resolution images from several laboratories placed LysoPE(22:6) in grey matter. · Exploring spatial metabolomes on the levels of…; Figure legends

  5. SuggestedRe-annotating a published rat liver drug-imaging dataset reproduced the reported erlotinib and olanzapine images and found olanzapine metabolites that the original study had not reported.

    Metabolites were searched via DrugBank. Isomeric 7-hydroxyolanzapine/2-hydroxymethylolanzapine and 4'-N-desmethylolanzapine were detected in tissue from dosed animals, yet were absent from the standards-spiked mimetic section, at intensities one and two orders of magnitude below olanzapine. Erlotinib was seen in cassette-dosed animals, olanzapine in both cassette- and discrete-dosed ones. · Screening for drugs and their metabolites; Figure legends

Methods, data and reproducibility

An Apache Spark engine on Amazon Web Services annotates centroided imzML imaging MS data at the molecular-formula level with an estimated FDR, using databases such as HMDB and LIPID MAPS; public results and their required metadata are served through a web app and a GraphQL API. The collection was mined by representing datasets as term frequency-inverse document frequency (tf-idf) vectors of annotated formulas.

  • Public imaging MS datasets in the METASPACE knowledge base (tissue sections, whole-body animal sections, human cancer cohorts, other samples) · multiple (human, mouse, rat and others) · n = over 3200 public datasets (January 2019)
  • Organs-mimicking section with gelatine wells of homogenized tissue from six mouse organs, prepared and imaged by the authors · mouse · n = 6 organs
  • Mouse and rat whole-body sections contributed by pharma and academic laboratories (e.g. Servier, Genentech) · mouse; rat
  • High-resolution (20 μm) rat liver section used for GSH and GSSG localization · rat
  • Published rat liver drug-imaging dataset (control, cassette- and discrete-dosed at 2 h and 6 h, plus liver mimetic tissue with drug standards), contributed by AstraZeneca · rat · n = five animals plus one mimetic section
  • Ovarian cancer cohorts from the Takats (Imperial College London) and Eberlin (University of Texas Austin) laboratories, used for the NAA observation · n = 2 cohorts
  • Brain datasets forming the brain subnetwork of the similarity network · n = 26 laboratories (27 in Figure 4 legend)
Methods & tools
METASPACE annotation engine, Apache Spark / pySpark, Amazon Web Services (c4.2xlarge instances, S3), HMDB, ChEBI, LIPID MAPS, SwissLipids, gensim v2.3.0, Cytoscape, ClassyFire, DrugBank, PostgreSQL / ElasticSearch
Software & environment
The engine uses Apache Spark via pySpark. The web app uses Vue.js and TypeScript. Storage is PostgreSQL plus ElasticSearch, with RabbitMQ for queuing. gensim v2.3.0 computes tf-idf and co-localization. No containers or requirement files are described.
Assay conditions & compute
The engine usually runs on three AWS c4.2xlarge instances (8 vCPU, 15 GiB memory each); a 10 GB dataset is annotated in under 10 minutes on the default cloud cluster. ElasticSearch was estimated at least 10x faster than PostgreSQL for this data.
Reproducibility gaps
Dataset and image 'links' mentioned in the text have no printed URLs. Supplementary tables (marker lists, kappa table, enrichment results) are not in the text. No repository is named for the tf-idf, network and enrichment analysis code.
Limitations (6)
  • Metadata factors are confounded in the collection (e.g. lymph node data come from one laboratory only), so factor-level and organ-level findings need cautious interpretation. · stated by the authors
  • Annotation is at the molecular-formula level (Metabolomics Standards Initiative Level 2). Isomers are not resolved, so the engine lists every candidate isomer in the screened database. · stated by the authors
  • Apart from a few organs such as brain, there are not yet enough samples per organ for comprehensive organ-specific atlases; this depends on further growth of the collection. · stated by the authors
  • Many signals in contributed datasets remain unannotated ('dark matter'); the authors list exploring them as future engine work. · stated by the authors
  • The proposed link between N-acetylaspartic acid (NAA) and ovarian cancer rests on its detection in ovarian cancer cohorts from two laboratories and its absence from their other cancer cohorts. The text reports no statistical test, counts or independent validation; it cites only agreement with a prior report. · noted in this summary
  • The engine speed (10 GB in under 10 minutes) and the ElasticSearch speed-up are stated without benchmark data in the text. · noted in this summary

Read this if you generate or reuse imaging mass spectrometry data and want a shared FDR-controlled annotation service, or if you want to mine many spatial metabolomics datasets across laboratories, organs and drugs.

Citations
89

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISWilliams '26Mekonnen '26Ozturk '25Wadie '24Clarke '23Saharuka '24
67% semantic matchBioinformatics
Eleanor C Williams, Heather Hulme, Aleksandr Zakirov+10 more·bioRxiv·Apr 29, 2026