Compared — LinkML source (20)
Not compared — no LinkML source (yet)
No machine-readable schema
JBEI Ganymede · JGI GOLD (API emits JSON, no schema) · JGI Proposal DB
Standards without LinkML artifacts
AmeriFlux flux-processing · BADM · GeoJSON · OGC O&M / EDR / Features · ISO 19115 · FGDC · STAC · ODM2 · USGS GeMS · EDX / DCAT · CORAL · GeoParquet · JGI GFF
Not publicly available
AmSC Open Metadata · IDeA-schema
Represented via a LinkML rendering
MIAPPE checklist → miappe-linkml, linkml-test · ESS-DIVE RFs → geochem-rf-linkml · basin3d → basin3d-linkml ×2
Different abstraction level
lakehouse DB extractions (GOLD / IMG / plant-db / SMC physical tables)
A concept-level comparison of 20 BER-relevant LinkML schemas, scored for structural overlap.
Same concepts, different spellings
Every schema defines its own Sample, Study, QuantityValue. NMDC writes has_numeric_value, lambda-ber writes numeric_value — same meaning, zero literal overlap.
Schemas from across BER
NMDC, KBase CDM, EMSL/MONet, ESS-DIVE reporting formats, structural biology, plant phenotyping, and the dataset portals — contributions from LBNL, PNNL, ORNL, ANL, and SLAC.
Every class, every slot
All 20 schemas were reduced to the same comparable form: classes with their full induced slot sets, straight from each repo's LinkML source.
Extract
SchemaView dump of classes + induced slots, one per schema repo
Map
Curated concept → class mapping, every class name validated against the extracts
Score
Pairwise slot-overlap (Jaccard) per concept, plus curated confidence + notes
Publish
47 concepts × 20 schemas, as a browsable, filterable matrix
Broadest concepts (schemas defining it, of 20)
Each row also carries a max pairwise slot-Jaccard score, a confidence rating, and curated notes.
Alignments that already exist
BASALT reuses NMDC conventions
Identical class names (ChromatographyConfiguration) and has_* value slots across the MONet model.
BRC ↔ OSTI transforms exist
brc Dataset carries 21 E-Link legacy slots; the repo includes transforms in both directions.
aims-leaf ⊂ lambda-ber
Near-verbatim subset; many class pairs have identical slot sets (1.00 Jaccard).
MIAPPE has two disjoint renderings
miappe-linkml and linkml-test model the same MIAPPE entities with zero shared slot names.
AValue classes & scaffolding
nmdc · kbase-cdm · lambda-ber · BERtron · BASALT · nmdc-lakehouse
Basis: two naming conventions (plain vs has_*) over near-identical structures.
Dependency: underlies groups B–E; changes propagate to all of them.
BSamples, sites & feature-of-interest
nmdc · BASALT · lambda / aims-leaf · MIAPPE · basin3d · geochem-rf · wastewater
Basis: 10-schema coverage; MIxS slots shared by nmdc, BASALT, wastewater.
Difficulty: members are categorially disjoint — specimen, site, observation unit.
CObservations & variables
basin3d ×2 · geochem-rf · MIAPPE · kbase-cdm · BASALT · BERtron
Basis: 9-schema coverage; OGC O&M / SOSA as external anchor.
Difficulty: slot names nearly disjoint across families; mapping-heavy.
DGenome features & annotation
nmdc · kbase-cdm · lambda / aims-leaf · BASALT · JGI GFF (in Dremio)
Basis: GFF core (start/end/strand/type/phase) shared; GFF data already loaded in the lakehouse.
Difficulty: contig/MAG classes share no slots; annotation granularity differs per schema.
EDataset & credit metadata
brc · osti · data-sheets · model-card · kg-registry · kbase-credit · BERtron
Basis: DataCite / DCAT as target frame; BRC↔OSTI transforms exist.
Difficulty: portal metadata only; no instrument or field data.
Sequencing — which two concepts first?
Candidate criteria: schema coverage, existing structural overlap, science priority, production readiness of the schemas involved.
Anchors — who takes A–E?
Each group needs one modeler per schema family it touches. Two groups can share a person.
The deliverable — one shape for all five groups
Proposal: per concept, one shared LinkML module plus per-schema mappings (SSSOM), contributed back to the matrix repo so the scores update.
Tracking
The matrix regenerates from the schema repos; adding or updating a schema is a pull request and a rerun.
First milestone per group: a drafted target class and mappings for its top-coverage concept. Two weeks?
One standing sync across groups on the value-class layer (group A).