Metabolite Enrichment

GO annotates gene products, not chemicals. A metabolomics result reaches GO only through an intermediary, and the choice of intermediary decides what the enrichment can say. prepare-chem2go materializes three such bridges as ordinary genesets-rs annotation tables whose “genes” are canonical ChEBI ids, so the engine runs unchanged.

Bridge	Path	Reaches	Species
`logical_definition`	GO-plus axioms naming a ChEBI class (has_primary_input/output, has_participant, regulates_levels_of)	Chemical-named processes and transporters: fatty acid metabolic process, citrate transport	Independent
`rhea`	Rhea reaction participants → `rhea2go`	Catalytic molecular functions	Independent
`enzyme` (optional)	Rhea → UniProt enzyme → that enzyme’s GAF annotations (BP by default)	Pathway processes: tricarboxylic acid cycle, urea cycle	One organism’s GAF

The bridges are kept separate so they can be compared the way GOA evidence variants are (annotations.<bridge>.tsv, plus all, plus species_independent when the enzyme bridge is on).

Prepare

just chem2go-prep           # human GAF for the enzyme bridge
# or, species-independent bridges only:
uv run --project python/genesets-workflows genesets-workflows prepare-chem2go \
  --out-dir evals/chem2go/generated/current

Downloads (about 200 MB, cached under downloads/): GO-plus JSON, ChEBI lite, ChEBI cross-references, the Rhea flat file, rhea2go, Rhea’s pH 7.3 mapping, and with --gaf/--gaf-url also rhea2uniprot_sprot and the GAF.

Normalize, then enrich

genesets-workflows normalize-chemicals changed.txt  --prep-dir $PREP --out query.txt
genesets-workflows normalize-chemicals measured.txt --prep-dir $PREP --out background.txt
genesets-rs enrich --annotations $PREP/annotations.all.tsv \
  --terms $PREP/terms.tsv --closure $PREP/closure.tsv \
  --sample query.txt --background background.txt --min-overlap 2

normalize-chemicals accepts ChEBI in any spelling, HMDB (5- or 7-digit), KEGG COMPOUND, LIPID MAPS, MetaCyc, DrugBank and CAS. Its report says how each id mapped (canonical, merged, secondary, xref, xref_ambiguous, …) and whether the result carries any annotation. --extra-xrefs takes a two-column TSV for ids ChEBI does not cross-reference.

Design decisions

Collapse what metabolomics cannot distinguish. Conjugate acids/bases, tautomers and Rhea’s pH 7.3 mapping are merged, and the group is represented by its pH 7.3 major microspecies, the form GO and Rhea use. Citric acid, citrate(3-) and HMDB0000094 all become CHEBI:16947. Enantiomers are not merged.

Roll chemicals up the ChEBI hierarchy, but not through near-root classes. A chemical reaches a GO term whose axiom names it or any ChEBI is_a ancestor, so palmitate reaches fatty acid metabolism. Roll-up stops at classes with more than --max-rollup-class-size (default 2500) descendants; without the limit “chemical entity” links every chemical to dozens of transporter and homeostasis terms. Broad GO terms are still reached through the GO closure.

Drop currency metabolites as participants. Water, H+, ATP, NAD(P)H, CoA, SAM and common ions are excluded wherever a chemical is one participant among many (Rhea, has_participant). Primary inputs/outputs keep them, so ATP metabolic process still reaches ATP. currency.tsv lists them with their Rhea degree; --keep-currency disables this.

Default universe = measurable metabolites. Without --universe, only chemicals that a bridge links directly, Rhea participants, and chemicals with an HMDB, KEGG, LIPID MAPS or MetaCyc cross-reference are annotated (about 49k of ChEBI’s 220k classes; tripeptides and the like are left out).

Link kinds. logical_definition keeps metabolism, transport, catalysis and homeostasis terms by default. Binding, response to and signaling terms are dropped (--link-kinds): “response to glucose” describes the genes that respond, not what happens to glucose.

The background matters

The background must be the metabolites your platform measured, normalized with the same command. The annotated chemicals are not a valid background: untargeted platforms detect a biased fraction of metabolism, and using all annotated chemicals inflates significance for well-measured pathways (see Wieder et al. 2021, PLoS Comput Biol 17:e1009105).

What it recovers

On the smoke queries in evals/chem2go/ (TCA cycle and urea cycle intermediates, as mixed HMDB/KEGG/ChEBI ids):

logical_definition and rhea return transporters and catalytic activities (C4-dicarboxylate transport, arginase activity), not the cycles themselves. GO rarely links pathway processes to chemicals: tricarboxylic acid cycle lists only citrate as a participant.
enzyme puts tricarboxylic acid cycle and urea cycle in its top two, with some noise from multifunctional enzymes’ other BP annotations (e.g. DNA repair terms on TCA enzymes).

So the species-independent bridges answer “which chemistry is this”, and the enzyme bridge answers “which pathway”. Use all when you want both.

Known gaps

ChEBI’s cross-references are uneven: HMDB0000052 (argininosuccinate) is attached only to an unconnected 2-star stub, and carbamoyl phosphate’s HMDB id is missing. --extra-xrefs patches these; a UniChem or MetaNetX export is the general fix.
Lipid species reported at sum composition (“PC 34:1”) need mapping to a ChEBI class first; once mapped, roll-up handles them.
The enzyme bridge inherits GAF annotation biases, and one metabolite fans out to every pathway of every enzyme that touches it. Counting stays per metabolite, but expect broad terms.
There is no curated metabolite benchmark yet. Natural answer keys: inborn errors of metabolism (known pathway, known accumulating metabolites), leave-one-out over Reactome/Rhea pathways, and published MetaboLights studies.

Keyboard shortcuts