Methods, data and limits

What the GEM Autocurator computes at each stage, the reference data behind it and at which release, and what each result cannot support.

How a model moves through it

Everything runs in your browser. The model file is read by JavaScript in this tab, curated in memory, and written back out by the same tab. It is never uploaded, and there is no server that could store it. The one exception is described under network.

Each stage produces proposals, not edits. A proposal only reaches the exported model if you approve it. Approving a finding does one of three things, and the report says which for every finding:

• Written into the model. Identifier renames, merges of duplicate entities, proton and water balance fixes, and bound changes on reactions the thermodynamics contradict.
• Carried with the model as a curation note. Findings that mark a reaction for attention without changing its chemistry, such as a reaction that cannot carry flux.
• Recorded in the report only. Findings where the fix is a modelling decision this tool does not make, such as gap-filling a dead-end metabolite.

What each stage checks

1. Identifier mapping

Every metabolite and reaction is looked up in a structure-clustered reference built from MetaNetX MNXref, which groups compounds across BiGG, KEGG, ModelSEED, ChEBI, MetaCyc, Rhea and HMDB by InChIKey. A cluster that contains a BiGG identifier lends that identifier to the whole cluster. A cluster with no BiGG identifier is given one stable BiGG-style identifier, minted once and frozen, so the same compound gets the same identifier in every model that passes through this page.

Every proposal carries one evidence tier, strongest first:

1. InChIKey or a database cross-reference. The entity matches on its structure, or on an identifier it already carries.
2. Matched by name only. The names agree and no structure or cross-reference does. A name cannot distinguish a stereoisomer, a salt form or a synonym group from the compound itself, and a name match is not used to merge two entities.
3. Minted from a name. Nothing in the reference matches, and a BiGG-style identifier is constructed from the entity's own name. It is unique within your model and recognised by no database.
4. No evidence. Nothing matches and nothing can be derived; the original identifier is kept.
5. Cross-references disagree. The cross-references in the file, or the file's own compound name, point at different compounds. Nothing is renamed, and the row names every candidate with the reference behind it.

How a cross-reference match is decided. Every cross-reference an entity carries is looked up. A full InChIKey settles the match, being the structure itself. A candidate whose heavy atoms contradict the model's own formula is excluded, and the entity's own name in the model has to corroborate the identifier it would be renamed to. Where the candidates still disagree the rename is withheld, and the row lists each candidate with the reference that supports it. A match that changes the recorded formula or charge is stated on the row and decided as its own class.

Each identifier is carried by at most one entity. Where a proposed identifier is already in use, the rename is withheld and reported, and both entities are kept.

2. Discrepancies

Two identifiers for the same compound in the same compartment are proposed for merging. A compartment suffix is not a discrepancy: cytosolic and extracellular glucose are two entities and stay two entities. Two reactions are proposed for merging only when their stoichiometry matches. Two reactions with similar names but different chemistry are a conflict, listed for review and not merged.

3. Structural QC

Dead-end metabolites (produced but never consumed, or the reverse), reactions with no gene association, and reactions that cannot carry flux even with every exchange reaction opened. That last test is a flux-variability sweep over the internal reactions with every exchange open, so a reaction it returns is blocked by the network itself and not by the medium loaded.

4. Mass and charge

Element and charge balance are judged at the model's own reference charge states, and are independent of pH. A reaction that balances at its reference charges is not a defect at any pH; the fractional net charge you would get by titrating it is proton buffering, not an error.

The pH control is an explorer: it recomputes metabolite charge states at the chosen pH from Henderson-Hasselbalch over ModelSEED pK_a sites. It raises no findings.

5. Thermodynamics

Reaction directionality is compared against a directionality call assembled from more than one source, each database casting one vote on the reversible/irreversible class. A bound is only proposed when the model's own equation could be matched to the reference orientation and a signed Δ_rG is available for the reaction as the model writes it; a reaction that crosses a membrane never qualifies, because the available free energies are computed without compartments. Everything else is review-only. Cycles that generate ATP or another energy carrier from nothing are detected by closing every uptake and asking the solver whether an energy carrier can still be made. A forced-open internal reaction, such as a maintenance reaction with a non-zero lower bound, makes that problem infeasible, so those bounds are relaxed for the duration of the test and the ones relaxed are named on the panel.

6. Lab validation

Your organism is resolved from the model name or an accession, looked up in GrowthDB, and simulated on the medium that experiment used, pulled from MediaDB. Measured and predicted growth rates are shown side by side. An exchange rate taken from an experiment is labelled measured; every other rate, including one derived from a medium recipe, is labelled a preset with no experimental backing. Each record states whether it is for the exact strain named in the model or for another strain of the same species.

Reference data and releases

Release and coverage of every reference in this build.

Limits

• No gap-filling. Dead ends and blocked reactions are reported, and can be carried into the exported model as notes. No reaction is added to repair them; which reaction to add is a modelling decision for a biologist.
• No reaction is deleted to repair a defect. The one approval that collapses entities is a merge of duplicates, and the survivor keeps the widest bounds and every copy's gene rule.
• The directionality call covers the sources named in the table above, one vote each. Rhea and MetaCyc directionality are not among them. ModelSEED publishes two directionality columns; they are two fields of one database, not two opinions, so they are combined into a single vote and where they contradict each other ModelSEED abstains. Where the sources disagree, or none has an opinion, the call is unknown, and unknown from a tie is reported separately from unknown for want of any source at all.
• Δ_rG′ values are estimates. They come from group contribution and component contribution methods, which estimate a reaction's free energy from the groups its compounds are built from. Their uncertainty is often the same order as the threshold that calls a reaction irreversible, so a reaction inside that band is undecided.
• Δ_rG′ values are standard-state. A transformed standard free energy is defined at a fixed pH, a fixed ionic strength and 1 mM concentration for every participant. A reaction that is irreversible under those conditions can run either way at cellular metabolite concentrations, so a directionality finding cannot settle the direction a reaction takes in vivo.
• No Δ_rG′ is offered for a reaction that crosses a membrane. Both sources compute free energy in a compartment-free namespace, where h_e and h_c are the same species and cancel. For a transporter or a chemiosmotic reaction the value that comes back therefore prices the same chemistry with the transport step deleted, and carries no proton-motive force or membrane-potential term: quoted for ATP synthase it reads as bare ATP synthesis, and says the physiological direction is uphill. Those values are withheld rather than shown, and no bound is proposed from one, so every transmembrane reaction on the Thermodynamics stage is review-only. A reaction is judged transmembrane on its BiGG equation where there is one (7,939 reactions) and on BiGG's transport-identifier grammar where the equation source has no entry for it (70 more). 2,368 of the 27,172 reactions in the reference are in neither position: their compartment span is not established, and they keep their Δ_rG′. Your own model is checked separately and on its own compartments, so a reaction you write across a membrane is treated as transmembrane whatever the reference says.
• pK_a values are estimates too, from ModelSEED's group-contribution table. A compound with no pK_a record is shown as not titratable.
• Flux balance analysis assumes steady state and an objective. Neither is a fact about your organism. A model that predicts growth shows that the network permits it, not that the culture will do it.
• Flux balance analysis gives a theoretical maximum. A model that over-predicts growth is unconstrained, which is not the same as wrong. A comparison with laboratory data is evidence, not a verdict.
• The experimental data favours organisms people study. GrowthDB and MediaDB are built from published measurements, and publication favours model organisms, defined media and aerobic growth. Every validation number carries the number of records behind it: the same grade over three records is weaker evidence than over three hundred.
• Not every organism has a record. Where GrowthDB holds nothing for the strain in the model, the stage reports no record and takes a condition you enter yourself.
• A minted BiGG-style identifier is not a BiGG identifier. It is unique and stable within this reference, and no database resolves it. It carries that label on every row and in the report.
• A name match is an evidence tier, not a resolution. It is labelled on the row, is not used to merge two entities, and is outside the batch action that accepts structurally resolved changes.
• MATLAB files saved with -v7.3 are not read. Those are HDF5 containers; re-save with -v7, or export to JSON. Level 5 .mat files, which COBRApy, scipy and the COBRA Toolbox write by default, are read. A file that is not a metabolic model is refused with a message naming what was missing.
• The .mat cannot carry every namespace. Every namespace on every metabolite, reaction and gene (BiGG, KEGG, ChEBI, MetaNetX, ModelSEED, BioCyc, HMDB, SABIO-RK, Reactome, LipidMaps, InChIKey, UniProt, SBO and anything else the file declares) is written into the JSON and the SBML, and a JSON model's annotation keys come back spelled as the file wrote them. The COBRA MATLAB convention names one column per namespace, so a namespace it does not name cannot be read back from a .mat. InChIKey is the case worth naming: the convention defines no column for it, this page writes metInChIKey, and COBRApy reads none of them.
• The export can add a cross-reference. A namespace is written only from a value shaped like an identifier of that namespace. An entity whose own identifier is a database identifier keeps that identifier as a cross-reference after a rename, which is the only case in which the export adds one. The report lists every namespace, its count, which of the three formats carries it, and how many references the export added.
• The reference has a build date. A compound or reaction added to an upstream database after the date in the table above does not resolve, and is given a minted identifier.
• Some computations run under a search budget. The substrate sweep scores a bounded sample of growth conditions, the exchange picker shows a bounded number of matches, and the exported dashboard's finding table is bounded. Each states the number used and the number available, on screen and in the exported report.
• A cancelled flux scan is a partial one. Its counts cover the reactions solved before the cancellation, and the remainder is reported as untested.

What leaves your browser

Your model does not. It is read from disk by this tab, held in memory, and written back to disk by this tab.

The single outbound request the application makes, beyond downloading its own code and reference files from this site, is organism resolution: if you paste a genome accession or a taxonomy identifier into the lab-validation search, that identifier string alone is sent to the NCBI Datasets API to look up the organism name. Nothing from your model is included in that request. A failed lookup is reported on the stage and falls back to the search box. If you never use accession lookup, no request is ever made.

The page loads no fonts, analytics, trackers or third-party scripts of any kind.

Licence and terms on the data

The code is MIT licensed. The page bundles GLPK, a GPL-3.0 linear programming solver compiled to WebAssembly, which is what solves every flux problem here. The reference data is derived from MetaNetX MNXref and the ModelSEED Biochemistry database, and is redistributed under their terms. The full statement, with the version, checksum and licence of every redistributed file, is in THIRD_PARTY.md, and the licence itself is in LICENSE. GLPK's own GPL-3.0 text is served beside the binary, at vendor/GLPK-LICENSE.

The page is set in IBM Plex Sans and IBM Plex Mono, copyright 2017 IBM Corp. with the reserved font name "Plex", under the SIL Open Font License 1.1. The latin subsets are served from this site at assets/fonts/ with the licence text beside them, at assets/fonts/OFL.txt. No font is fetched from a third-party origin.

If you use a curated model from this page in published work, cite the underlying resources as well: MetaNetX for the identifier reference, ModelSEED for the thermodynamic and pK_a data, and BiGG for the identifier namespace. CITATION.cff has the machine-readable form.

If something looks wrong

The identifier reference is shared, so an incorrect cluster affects every model resolved against it. Report one as an issue on the repository, with the identifier and what you expected. Rejecting the proposal keeps it out of your own model.