Crosswalk to IMG's core database
model/crosswalks/img-core-v400.sssom.tsv maps
slots of the feature model to columns of IMG's img_core_v400 database, for
https://github.com/turbomam/feature-table-corpus/issues/47. It's an
SSSOM table: one row per slot and column, with
skos:closeMatch where the two hold the same thing in a different form, and skos:relatedMatch
where they only overlap. No row is skos:exactMatch, because the source doesn't state coordinate
bases or value spellings.
Each subject is a slot, typed rdf property, with its SSSOM subject_label naming the class it
is used on, such as SequenceFeature.start or AlignmentTarget.target_start; a slot keeps its URI whichever
table it maps. A list of structs, such as SequenceFeature.attributes, maps through its items' slots (Attribute.key,
Attribute.value), not as one column, and a table that maps any slot of a struct maps all of
that struct's required slots. A hit table's gene_oid maps to both seqid and parent, because the model
requires both on a protein-coordinate SequenceFeature.
SSSOM maps one column to one slot, so a value built from two columns is two relatedMatch
rows: SequenceFeature.dbxref is gene_ext_links.db_name, a colon, then gene_ext_links.id. A table
with target positions also maps AlignmentTarget.target_id, which the model requires.
Source
IMG's code isn't public, but its database schema is readable as a LinkML view in
https://github.com/cmungall/bridge-schemas, file src/bridge_schemas/schema/jgi/img_core_v400.linkml.yaml,
read at commit ba25724947002e40a0393dba624a6c200d9ef603 on 2026-09-29. That view has 117 tables,
and every column named in the crosswalk exists in it (checked the same day). The view isn't
vendored here, since the repository has no license. The IMG files this repository vendors come
from the same database: 2708743150.gff and its tables are a img_core_v400 taxon export, and
scripts/img_taxon_bundle_map.py maps them. So the hit-table rows follow what that mapping does.
How the model lines up
geneis a SequenceFeature on contig coordinates:start_coord,end_coord,strand,locus_type,product_name, andscaffoldandtaxonas integer keys to contig and genome tables.gene_feature_tagsisAttribute: an open tag/value table.- The per-method hit tables (
gene_pfam_families,gene_cog_groups,gene_ko_termsand others) are protein-coordinate Features on the gene:query_startandquery_endarestartandend,bit_scoreisscorewithscore_type: bit_score, andsubj_startandsubj_endare the alignment target's positions. gene_sig_peptidesholds the SignalP sites the TMHMM/SignalP mapping reads from GFF.
What doesn't line up
- Gene locations aren't mapped yet.
gene_frag_coordsholds each part's start and end, but aLocationPartalso needs a seqid, strand and endpoint status, andis_partial_wdoesn't say which end is partial: https://github.com/turbomam/feature-table-corpus/issues/119. - The view has no scaffold or taxon table, so
Contig.length_bp,topologyandContigCollectionhave no column here.docs/genes.mdin bridge-schemas joinsgeneto ataxontable, so the database has one; the view leaves it out. evalue,percent_identityandalign_lengthon the hit tables have no typed slot; they are Attributes today. Typed attribute values are https://github.com/turbomam/feature-table-corpus/issues/42.gene.is_pseudogeneis a flag on a gene row, where GFF, and so the model, writes typepseudogene.gc_percent,dna_seq_lengthandaa_seq_lengthcan be derived and have no slot.gene_tigrfamskeeps its positions as text lists (sfstarts,sfsends), so one row can hold several spans.
What the IMG website queries, and so which of these columns matter most, is application code. That still needs someone at JGI.