Skip to content

Acceptance tests by discipline

Who is likely to judge a new feature table format, the first test each would apply, and the number that would show whether it passes. Written 2026-10-02 to plan evidence before it is asked for. The tests are predictions, not requirements anyone has stated. The status column says what this repository has today, read from its own documents and code on 2026-10-02; nothing in this page was newly measured unless it says so.

Producers of feature data

discipline likely question acceptance test metric status here
Annotation pipeline developers (NCBI PGAP, JGI IMG, Prokka, Bakta) Can my output go in and come back out unchanged? Round trip of each pipeline's real files Share of rows and attribute keys preserved, per producer Byte-for-byte round trips for NCBI RefSeq (annotwriter), three NMDC file types (Pfam, Prodigal and CRT), BLAT BED12 and INSDC GenBank (conversion round trips); other NMDC file types are not round-tripped, and an older Functional Annotation GFF that repeats an ID is refused (schema notes on feature_id); and for IMG and Phytozome files through the dialect mappers (conversion profiles, tests/test_img_*_map.py, tests/test_phytozome_*_map.py). No Prokka or Bakta file is in the corpus; only their attribute key lists (columns and discretion)
Eukaryotic genome groups (Phytozome, MycoCosm) Does it hold exons, UTRs, isoforms and a representative transcript? Gene, several mRNAs and multi-part CDS load, validate and reconstruct Share of genes whose structure reconstructs exactly The Phytozome mapper loads gene, mRNA, exon, CDS and UTR rows and sets is_representative from longest; the 200-gene Populus excerpt (293 mRNAs, up to 10 per gene) round-trips byte for byte (tests/test_phytozome_gene_exons_map.py). CDS parts stay separate Features under one parent, not one joined location; that and MycoCosm are open: https://github.com/turbomam/feature-table-corpus/issues/48, https://github.com/turbomam/feature-table-corpus/issues/92
Archive submission curators (INSDC, GenBank) Can it produce a valid submission? Export to GenBank feature-table form that NCBI's validator accepts Validator errors per 1,000 features insdc-locations/1.0.0 reads GenBank flat files and writes them back, exact and reconstructed, checked with Biopython 1.85 (feature locations); no five-column submission table is written and no NCBI validator has been run
Natural-product and biosynthetic gene cluster researchers (SMC, antiSMASH, MIBiG) Can I store a cluster and ask what is next to what? Gene order and neighbor queries on a known cluster give the known answer Query results against a curated answer just bgc-check does this for the RefSeq actinorhodin cluster (BGC exercise); SMC is https://github.com/turbomam/feature-table-corpus/issues/166
Protein-domain annotators (Pfam, InterPro, KEGG) Are hits kept in protein coordinates, with e-value and bit score apart? A CDS with several domain hits keeps each hit's protein interval, target interval and score type Share of hits that convert back to the source rows NMDC Pfam through nmdc-pfam-protein/5.0.0 (protein-relative profile), bit score typed bit_score, e-values kept as text attributes; IMG Pfam, COG, KO/EC, TIGRFAM, SMART, SUPERFAMILY and CATH-FunFam hits through the per-method dialect (tests/test_img_per_method_map.py), where score_type stays unset. Only the NMDC profile fills the typed target interval, from model_start and model_end; the IMG per-method mapping keeps hits' protein intervals but no target, and the IMG taxon bundle keeps subj_start and subj_end only as generic attributes. No mapper fills Attribute.numeric_value

Consumers of the data

discipline likely question acceptance test metric status here
Microbial ecologists and metagenome users (NMDC) Does it scale to every sample and keep MAG membership? Load one full NMDC annotation run into BERDL; contig-collection queries work Load time and table size per million features; result counts and latency of the contig-collection queries Not measured at NMDC scale: https://github.com/turbomam/feature-table-corpus/issues/167 waits on loading (https://github.com/microbiomedata/nmdc-lakehouse/issues/388). The largest timed Parquet export is 4,439 IMG features, 10.78 s wall clock (Parquet export, 2026-09-25); the largest timed round trip is all 532,682 rows of the TAIR10 gene_exons file, 211 s (conversion profiles, 2026-09-28); a contig-collection join is tested on the 15-feature example (tests/test_feature_model.py)
Comparative genomics researchers Can I join the same gene across annotation versions and sources? Join features from two sources on stable identifiers Share of features with a usable cross-source identifier stable_identifiers exists on SequenceFeature; the IMG taxon bundle mapper fills it from the bundle's external-link table (scripts/img_taxon_bundle_map.py), NMDC files carry none (schema notes); no cross-source join has been measured
Genome browser developers (JBrowse, IGV) Can I get GFF3 or BED back out, by region, fast enough to draw? Region query returns nested features; exported GFF3 loads in a browser Region-query latency on a large table Overlap and neighbor queries exist in DuckDB (query examples); latency is not measured (https://github.com/turbomam/feature-table-corpus/issues/167), and no exported GFF3 has been loaded in JBrowse or IGV
Agent and LLM users (BERIL) Can an agent answer questions from the tables without being told the schema? A fixed question set gets correct SQL and correct answers Share of questions answered correctly Not started
Statisticians and machine-learning users Do nulls, types and counts mean the same thing across sources? Each column has one type and one meaning in every source's tables Columns whose types disagree across sources The Parquet export checks each column's type against the schema on every run; nmdc-lakehouse's feature_convert.py writes phase as int8 where the export writes int64

Infrastructure and governance

discipline likely question acceptance test metric status here
Data engineers and lakehouse architects (BERDL, Iceberg, Trino) Does it load without hand fixes, and fit a flat-only catalog? Load into Iceberg with no manual schema edits; the flat profile passes the BRIDGE catalog rules Load failures per table; query cost of flat against nested The actinorhodin export loaded into BERDL on 2026-10-02 (https://github.com/turbomam/feature-table-corpus/pull/165); a column null in every row lost its type when written through pandas and Iceberg v2 refused it, so the file's own Parquet schema had to be passed (observed that day, not yet filed). ARRAY columns are refused by the catalog profile (https://github.com/microbiomedata/nmdc-lakehouse/issues/342); see flat profile
Relational database designers (IMG, SMC) Does it map to our existing tables? A crosswalk from each database's columns to model slots, gaps listed Share of source columns mapped IMG core crosswalk; SMC waits on its DDL
Data modelers and standards groups (LinkML, BRIDGE, KBase CDM, NMDC) Does it import cleanly beside our modules and reuse what exists? Builds under the BRIDGE central schema with no name or prefix clash; mappings to CDM and NMDC resolve Name clashes; share of classes with mappings name clashes with the central schema's slot and is renamed display_name with alias: name in https://github.com/doe-bridge/bridge-central-schema/pull/5; numeric_value does not clash (class-local on Attribute; the root schema builds with the measurements slot beside it). The same PR renames the EDAM.DATA prefix to EDAM and redeclares MIXS and EDAM in the root. CDM mapping is https://github.com/turbomam/feature-table-corpus/issues/139
Ontologists (Sequence Ontology, EDAM) Are feature types and score types real terms? Every type resolves to a Sequence Ontology term or is marked as a database accession Share of type values mapped linkml/valuesets GenomeFeatureType matches 44.9% of the 77,959 rows in the 42 GFF and GTF files under corpus/sources exactly and 56.8% ignoring case (2026-10-02 comment on https://github.com/turbomam/feature-table-corpus/issues/131); type is a plain string with no accession marker, and 7 of 14 NMDC file types put an accession there (https://github.com/turbomam/feature-table-corpus/issues/8)
Bioinformatics tool developers Do existing validators and parsers agree with it? Independent GFF3 validators' verdicts on the corpus are recorded and explained Agreement between validators and expected labels GenomeTools and GFF3toolkit in just validity-check (independent validation); AGAT is blocked on macOS arm64
Data stewards and licensing (JGI data policy) Do license and citation travel with the data? Converted outputs carry the source's terms and citation Share of outputs carrying terms and citation Provenance is carried. Neither terms of use nor citations reach the Dataset or the Parquet export; the Populus excerpt files carry their citation as a # cite: line, which the Phytozome converters keep in their header JSON (https://github.com/turbomam/feature-table-corpus/issues/151)
Program managers and funders (DOE BER) Will facilities adopt it, and who maintains it? Data from two or more facilities load in production Facilities using it; effort to add a source format None in production. NMDC's nmdc-lakehouse writes the model's shape (feature_convert.py), and one 45-row demo export is in BERDL as of 2026-10-02 (https://github.com/turbomam/feature-table-corpus/pull/165, https://github.com/turbomam/feature-table-corpus/issues/167); no second facility

Unusual gene structures

https://github.com/cmdcolin/oddgenes (text CC0, checked 2026-10-02) is a curated list of gene annotations that break common bioinformatics assumptions, with literature links. Its cases are a ready source of tests for the representation, as distinct from the format. The rows below are a selection of the cases that touch this model, not all of them; the status comes from reading feature locations, columns and discretion and the schema, and none was tested for this page; where an existing test already covers a case, the row names it.

oddgenes case what the model must hold status here
Circular chromosomes A feature crossing the origin Supported: FeatureLocation.crosses_origin, tested on six origin-crossing joins in phiX174 through insdc-locations; gff3-contig refuses the phiX174 GFF3
Translational (ribosomal) frameshift One CDS whose parts overlap by one or more bases Refused as one structured location (parts must not overlap) and as a GFF3 discontinuous CDS (two rows with one ID, which gff3-contig refuses); as separate CDS-part Features under one parent nothing checks overlap. The full NC_003888.3 file has nine such CDS (18 rows marked exception=ribosomal slippage): seven whose parts overlap by one base and two with a two-base gap. None is in the converted excerpt
Trans-splicing of exons on different strands One feature with parts on both strands Refused as one structured location (one strand per location); separate part Features on two strands under one parent are not refused and not checked
Exon shared across different genes A feature with more than one parent Supported: SequenceFeature.parent is multivalued, and tests/test_conversion_profiles.py round-trips a GFF3 exon with two parents through gff3-contig; the corpus fixture corpus/fixtures/edge-cases/multiple_parents.gff3 is not run through it
0 bp exon A zero-length site between two bases Refused: between-base sites are refused
1 bp exon A part whose start equals its end A feature with start equal to end validates (test_legitimate_absence_and_boundary_values in tests/test_feature_model.py); a one-base part inside a joined location is not checked
Ribosome hopping One CDS whose parts skip a stretch of the mRNA Expressible as a joined location of nonoverlapping parts; not checked
Stop codon readthrough A CDS that continues past a stop codon A plain interval holds it, but no slot marks the read-through stop; not checked
Stop codon reassignment, alternative codon tables A genetic code per sequence, and exceptions within a gene Per contig through Contig.translation_table; a CDS translation_table attribute that differs from the contig's or from other CDSs is refused by scripts/validate_closed.py. A codon-level exception such as a GenBank transl_except qualifier would survive only as a generic attribute, since insdc-locations keeps every qualifier as one, with no typed slot or check; no corpus file has one
Overlapping genes, nested genes, antisense transcription Independent intervals that overlap, on either strand Plain intervals. Two real overlapping genes, SCO5087 and SCO5088 in the actinorhodin excerpt (4 bases), are tested by tests/test_neighbor_query.py. Nested genes are present in the phiX174 record that insdc-locations round-trips (B inside A, E inside D, all plus strand), but no test asserts the nesting; no antisense case is tested
Polycistronic transcripts and operons One transcript parent to several CDS Expressible through parent; not checked
Polyproteins and inteins Protein products cut from one CDS A contiguous product or an intein can be a protein-coordinate SequenceFeature with its CDS as seqid and parent (protein coordinates); a mature product that joins two exteins cannot, because structured locations require contig coordinates (scripts/feature_locations.py). No class for protein products; not checked
Very large introns, many exons, many isoforms Large part counts and isoform counts per gene Isoform counts measured once: 5,804 of 27,416 TAIR10 genes have more than one mRNA (schema notes on is_representative, 2026-09-25); the Populus excerpt round-trips a gene with 10 mRNAs. Intron lengths and exon counts per gene are not measured
0-based versus 1-based coordinates One coordinate convention The model is one-based inclusive; the BED12 profile converts (conversion profiles)

Keeping this page useful

When a test here gets run, replace the status with the result, the date and the command. When a discipline states its own test, quote it and say where it was stated.