Metabolomics Interpretation — Tools Landscape

Supporting survey for the
Metabolomics Interpretation with GO and GO-CAM project.
The goal is to map what existing metabolomics-interpretation tools do, where
they get their networks/identifiers, and which parts are reusable versus
where GO / GO-CAM can add something they currently lack.

The standard untargeted-metabolomics interpretation pipeline

raw spectra (LC-MS / NMR)
  → feature table (m/z, RT, intensity)
     → metabolite annotation  (formula/adduct → compound; ChEBI / HMDB / KEGG / RefMet)
        → statistics           (differential abundance: significant metabolite set)
           → interpretation     (pathway / set enrichment)   ← GO is absent here

The interpretation step is where this project intervenes. Today it is served by
metabolite-centric set databases (KEGG, SMPDB) and, for MS1 peak lists,
network methods (mummichog). GO is not used because GO annotates gene products,
not compounds — the bridge has to be built through reactions (Rhea/Reactome,
which are written in ChEBI).

Tool-by-tool

MAGI — Metabolite Annotation and Gene Integration (LBNL)

• What: Computes a metabolite↔gene association score by propagating
  evidence across a biochemical reaction network: a candidate compound
  identity is reinforced if a nearby gene's predicted enzyme activity consumes/
  produces it, and vice-versa. Improves both compound IDs and genome
  enzyme annotations by maximizing consensus.
• Provenance: LBNL (Trent Northen, Benjamin Bowen, Oliver Rübel, Markus de
  Raad). Validated on Streptomyces coelicolor A3(2). PMID:30896917.
• Why it matters here: MAGI is the closest existing analogue to what we
  want — it already lives in the metabolite-reaction-gene space. It does not
  attach GO semantics. Reuse: the consensus-scoring idea; add: map the
  reactions to GO MF via rhea2go/ec2go and lift genes to GO BP so the
  associations become GO-interpretable and multi-omics-joinable (Approach C).

mummichog (Emory; Li lab)

• What: For untargeted MS1 data, maps every plausible formula/adduct
  match of each peak onto a genome-scale metabolic network, then tests for
  local network enrichment. The trick: true metabolites cluster in the same
  pathway neighbourhood, false matches scatter randomly — so structure in the
  network reveals real activity without committing to single compound IDs
  first.
• Provenance: Li et al., PLoS Comput. Biol. 2013; re-implemented in R inside
  MetaboAnalyst's "MS Peaks to Pathways".
• Why it matters here: mummichog's permutation/null-model framework is
  directly reusable. Its network is reaction-only (no causal direction, no
  regulation). Add: substitute or augment with GO-CAM curated causal
  networks (Approach B) — same local-enrichment logic, but over edges that carry
  direction and regulation.

MetaboAnalyst (Xia / Wishart)

• What: The dominant web platform for metabolomics interpretation. Modules:
  ORA, MSEA (Metabolite Set Enrichment Analysis), pathway analysis over
  KEGG and SMPDB, "MS Peaks to Pathways" (mummichog + GSEA), and
  multi-omics integration. Knowledgebase: five curated genome-scale models plus
  ~21 KEGG-derived organism models; metabolite sets from HMDB/SMPDB.
• Why it matters here: It defines the UX and statistical expectations of
  the field. Reuse: enrichment statistics and reporting conventions;
  add: a GO-BP enrichment backend (Approach A) so GO results are comparable
  side-by-side with the SMPDB/KEGG outputs analysts already trust.

SMPDB & HMDB (Wishart group)

• SMPDB — Small Molecule Pathway Database: curated human small-molecule
  metabolic and signalling pathways; the set source behind MetaboAnalyst's
  pathway/MSEA modules.
• HMDB — Human Metabolome Database: ~exhaustive human metabolite catalogue
  with ChEBI/KEGG/PubChem cross-references and pathway links.
• Why they matter here: They are the incumbent GO-free interpretation
  resources — the baseline any GO-based method must be compared against. Their
  ChEBI/KEGG cross-refs are also part of the identifier bridge.

Metabolomics Workbench (NIH / UCSD)

• What: Public repository of metabolomics studies plus a Metabolite
  Database (~174k entries, aggregated from LIPID MAPS, ChEBI, HMDB, KEGG,
  PubChem, etc.) and RefMet, a standardized reference nomenclature spanning
  ~500k MS/NMR annotations at four levels of structural resolution.
• Why it matters here: RefMet → ChEBI is a clean, programmatic on-ramp
  for turning a real study's messy metabolite names into the ChEBI IDs the
  Rhea/GO bridge needs.

MetaboLights (EBI)

• What: EBI's open metabolomics repository (raw data + metadata,
  cross-species, cross-technique). During curation, metabolite names are
  assigned ChEBI identifiers (with new compounds submitted to ChEBI).
• Why it matters here: The most ChEBI-grounded public study source —
  ideal for a demonstration dataset, because the metabolites already carry the
  identifier the whole GO bridge keys on.

The identifier bridge (why ChEBI is the linchpin)

Every hop is keyed on ChEBI or on a mapping this repository already maintains
(rhea2go, ec2go; see the RHEA project). That is what makes a
GO-native metabolomics interpretation feasible without inventing new mappings.

Reuse vs build — summary

Sources

• MAGI — https://pubs.acs.org/doi/10.1021/acschembio.8b01107 (PMID:30896917)
• mummichog — https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3701697/
• MetaboAnalyst 6.0 (PMID:38587201) — https://academic.oup.com/nar/article/52/W1/W398/7642060
• MetaboAnalyst functional-analysis vignette —
  https://www.metaboanalyst.ca/resources/vignettes/Functional_Analysis_global_metabolomics.html
• Metabolomics Workbench / RefMet —
  https://www.metabolomicsworkbench.org/databases/refmet/index.php
• MetaboLights — https://pmc.ncbi.nlm.nih.gov/articles/PMC10767962/
• GO-CAM — https://pmc.ncbi.nlm.nih.gov/articles/PMC7012280/ (PMID:31548717)
• GO-CAM metabolic phenotypes (Genetics 2023, PMID:37579192) —
  https://academic.oup.com/genetics/article/225/2/iyad152/7242464