Tier 0 · open

Corpus atlas

The honest inventory: not only what the corpus can do, but what it cannot express. Coverage gaps are reported as findings, because a gap is a research target rather than a failure.

• 15,664
  organism models
• 1,865
  canonical exchange metabolites (MetaNetX)
• 1,534
  median reactions per model
• 9,911
  redistributable CC-BY (stated, contestable basis)

Sources

Models come from seven model sources (plus one internal synthetic biomass-composition reference, used only as a stand-in and not a real organism): published collections, automated reconstructions from genome databases, and the APOLLO-595 pan-species set (pan-genome reconstructions of human-microbiome organisms, predominantly bacteria). Source mix is declared because it sets the floor on what any discovery result can claim.

What the corpus can do

Capability is computed from model exchange reactions on automated, gap-filled draft reconstructions, not asserted from literature. Read every count as an upper bound, not a curated or measured capability. Gap-filling can add transporters or carbon-fixation reactions an organism does not actually have, which especially inflates autotrophic counts: growth on syngas or formate in a corpus dominated by gut and host-associated bacteria is largely a gap-fill artifact, so treat those rows as “where to look,” not a claim that the organism is an autotroph. The counts show how many models can carry flux to biomass on a minimal defined feedstock; the rest need rich media.

Anaerobic-capable (FBA on gap-filled models, an upper bound): 11,638 models (74%) can carry flux to biomass with oxygen uptake blocked, a computed capability, distinct from the models annotated as anaerobes in the Oxygen niche table below.

Oxygen niche

O2 annotation covers 5,256 of 15,664 models; the remainder have no annotation in the source metadata.

Hazard uptake capability

Models carrying an exchange (uptake) reaction for a known environmental hazard: a necessary but not sufficient indicator, and an upper bound inflated by gap-filling. An exchange reaction is boundary transport, not a demonstrated degradation pathway: a genuine degrader also needs the downstream catabolic reactions, which this count does not require. Read these as an upper bound on where to look, not a claim that the organism degrades the compound.

Quality and coverage

Quality is graded by canonical-identifier coverage: how many exchange metabolites map to a MetaNetX universal ID. Full coverage (>99%) enables cross-model comparisons with no disambiguation step.

Biosafety

BSL annotation is from DSMZ / ATCC metadata where available. Models without a verified annotation are listed as unverified. A missing annotation is not an implicit BSL-1 clearance.

Of 15,664 models, 1,737 carry a human-pathogen annotation. Discovery runs flag these automatically and exclude them from unsupervised proposals.

Licensing

License status is tracked at the model level. Redistributable models may be published as part of a derived corpus; link-only models require fetching from the source repository. The Data page has the full breakdown by source with citation links.

What this corpus cannot express

Gap-filled reconstructions. Most models are auto-reconstructed with gap-filling, which adds transporters and reactions to make the model biomass-viable on rich media. This inflates apparent metabolic capacity: a gap-filled transporter is a computational assumption, not a measured fact. Discovery results are candidates, not confirmations.

Scope: predominantly bacteria and archaea. The corpus is predominantly bacterial and archaeal. The discovery-live set counted on this page (v0.2.2) excludes 79 out-of-scope models (77 PARADIGM protozoan-parasite GEMs and 2 mammalian cell-line models), which are retained in the full database but not counted here and never offered for consortium discovery. (The per-source table above sums to that discovery-live total.) Kingdom is not yet stamped on every model (a portion carry no kingdom label and are bacterial by source), so this is a source-level statement, not a per-model guarantee. Plant, algal, and fungal coverage is in development; rhizobia-legume nitrogen fixation is the near-term priority.

Defined-medium viability. Most models require organic growth factors that do not exist in a minimal mineral medium. Obligate growth-factor cross-feeding cannot be confirmed with gap-filled models. The thermodynamic gate is only a necessary-condition, class-level check on energetic reactions; it does not adjudicate amino-acid or vitamin exchange, and it is not a per-candidate proof. These dependencies remain hypotheses for the bench.

Environmental-anaerobe coverage is thin, and the thinnest part is the archaea. A taxonomic census of the discovery-live set puts numbers where this page previously gave an impression, and the numbers run in both directions. The corpus is dominated by gut and host-associated bacteria: 15,431 Bacteria against 200 Archaea and just 17 eukaryotes, with 16 models still carrying no taxonomy. Those figures were 10,220, 14 and 5,390 until 2026-09-23, when taxonomy was backfilled from genome accessions; the archaeal count did not grow, it became visible. Further passes on 2026-09-25 and 2026-09-26 took the unlabelled remainder from 925 to 16, and three of the organisms they recovered are cyanobacteria, which is to say the corpus held phototrophs this very census could not see, because an unlabelled organism reads as absent to any question asked by taxonomy. A fourth is a diatom, and it is counted nowhere in the range below, because these bounds were drawn over bacteria and archaea only; eukaryotic phototrophs are reported separately rather than folded in, since doing so silently would move a published figure. The 16 that remain unlabelled are mostly organisms science has not formally named, such as phytoplasmas and undescribed species from metagenome assemblies, and they stay unlabelled rather than guessed. See the changelog. By taxonomy, between 485 and 1,080 of 15,664 models fall in a chemolithotroph, autotroph or phototroph guild, so between 3.1 and 6.9 percent. Two numbers rather than one because neither alone is honest: the lower figure counts a curated list of genera, which turns out to miss most of the corpus’s cyanobacteria, and the upper adds whole phyla whose members are predominantly but not exclusively of that guild. The truth lies between them, and both are taxonomic bounds rather than verified function, since genus or phylum membership is not the same as a retained pathway.

Methanogens are the sharpest gap. Twelve models carry a methane exchange at all. Eight of those are named to species and are four species, each a gut or rumen organism duplicated across AGORA and AGORA2: Methanobrevibacter ruminantium, M. smithii, Methanosphaera stadtmanae and Methanomassiliicoccus luminyensis. The other four are not named to species, so the species count is at least four and at most six and this corpus cannot yet say which. Three of those four do carry a domain, and only one, a coarse lumped methanotroph, carries no taxonomy at all. A domain does not name a species, so it does not narrow that range. A functional pass corrected the sharper half of this. The environmental hydrogenotrophic methanogens are not absent from the corpus: CarveMe reconstructions of Methanococcus maripaludis, Methanothermobacter marburgensis, Methanosarcina baltica, Methanoculleus chikugoensis and Methanothrix thermoacetophila are all present. They carried no taxonomy when this was first measured on 2026-09-22, which is why a taxonomy census could not see them then; the 2026-09-23 backfill stamped all five, and each now carries a domain and a genus. That makes the gap sharper rather than smaller: all five now sit in methanogen genera, which is what the census classifies on, and the models still carry no methane chemistry, so what is absent is the function and not the label. Every one of those five carries no metabolite with the formula CH4 and no metabolite named for any of the four methanogenesis cofactors, at 42 to 46 percent gap-filled and roughly 2,000 reactions each, against 690 to 753 reactions for the curated gut pair that do work. A heterotroph-biased universal database gave each organism reactions it does not need and dropped the pathway that defines it. This page previously described the corpus as dominated by hydrogenotrophic methanogens, which overstated a pool of four gut species; the wording is corrected and the change is logged on the changelog.

Two guilds are better covered than this page used to say. Acetogens reach 23 species across Acetobacterium, Sporomusa and Clostridium, and dissimilatory sulfate reducers reach 18 species of Desulfovibrio alone, 24 models across the sulfate and sulfur reducing genera. Carboxydotrophs and reverse-beta-oxidation chain-elongators remain genuinely thin, and anammox is absent entirely. For the conversions that depend on a thin guild a “hit” is typically a gap-fill artifact rather than a credible route, and the corresponding reports say so.

The cause is source composition rather than a choice. AGORA and AGORA2 are human gut model collections, so the only archaea they contribute are gut archaea, and no environmental or anaerobic-digester archaeal collection is represented. Closing that gap is an acquisition question, and it is open.

Corpus version

• Versionv0.2.2
• Content hash58a8f20e2aba...
• Released2026-10-05
• Models15,664
• Redistributable9,911
• Zenodo DOI10.5281/zenodo.22019863