config.yaml

# Information about Babel
babel:
  github_url: "https://github.com/NCATSTranslator/Babel"

# Build information. Currently unstructured -- you can use this to write down any notes on what is going on with
# this specific build of Babel.
build:
  branch: babel-1.18.1
  workarounds:
  - Continued with the downloads from 2026jul15 under babel-1.18, but redownloaded CHEBI given many parsing changes.
  - DrugBank downloads are currently disabled https://github.com/cthoyt/drugbank-downloader/issues/23 -- I used the previously downloaded 5-1-13.
  - HMDB is now behind a Cloudflare bot challenge, so the pipeline can't download it (see docs/sources/DownloadPatterns.md). Unlike DrugBank, this is not a reused old download -- I downloaded it fresh in a browser on my laptop and copied it to the HPC.

# Versions that need to be updated on every release.
biolink_version: "4.4.3"
umls_version: "2026AA"
rxnorm_version: "07062026"
drugbank_version: "5-1-13" # The latest is 5-1-21, but since downloads are currently not allowed, we need to reuse our most recent download.
# GARD has no version number to bump, only a link that goes stale: this is the "GARD Rare Disease
# List <Mon><Year>.csv" link on https://rarediseases.info.nih.gov/about (under "Which rare diseases
# are included in GARD?"); NCATS publishes no stable data URL. The Salesforce `ids=068...` value is
# a ContentVersion id, i.e. ONE uploaded version of the file, so a new upload gets a new link and
# this one may stop resolving. Re-check the About page before each release (the network test
# tests/datahandlers/test_gard.py::test_gard_download_url_is_current does it for you) and repoint
# it; get_gard fetches it via src.datahandlers.gard.pull_gard(), and an expired link fails the rule
# rather than silently reusing the old file. Currently: "GARD Rare Disease List Jun2026.csv".
gard_download_url: "https://ncats.file.force.com/sfc/dist/version/download/?oid=00Dt00000004XG2&ids=068SJ00001HZAaEYAX&d=%2Fa%2FSJ00000BC4Xl%2FUj7U9WuHII571Akz5AUBLe6WSCeelaMBynbjWybmhuA&asPdf=false"

# Release naming and the previous-release comparison (see docs/RunningBabel.md).
# release_name is the date-style name for THIS build; it becomes the "name" in the generated
# prefix_report.json and names the directory it is archived into, releases/<release_name>/. Set it to
# the build you are about to make BEFORE starting a run.
release_name: "2026jul22"
# previous_release pins which committed baseline the end-of-run comparison diffs against: it is the
# release immediately BEFORE release_name, never release_name itself -- a build that diffs its own
# baseline produces a well-formed all-zeros report that reads as "nothing changed", so
# generate_prefix_comparison() raises rather than write one. Moving this on to the release just
# archived under releases/ is the deliberate "new baseline" action, done in the same
# commit that moves release_name to the next build; a unit test fails if the two drift apart.
previous_release: "2025sep1"
# Thresholds for the "Notable changes" section of reports/tables/prefix_comparison.md (informational
# only -- large drifts are flagged, never fail the build).
prefix_comparison_warn_abs: 100000 # flag any prefix whose absolute CURIE change is >= this
prefix_comparison_warn_pct: 25 # flag any prefix whose |percent change| is >= this

# Overall inputs and outputs.
input_directory: input_data
download_directory: babel_downloads
intermediate_directory: babel_outputs/intermediate
output_directory: babel_outputs
# Where DuckDB spills larger-than-memory intermediates. setup_duckdb() gives each job its own
# duckdb-$SLURM_JOB_ID subdirectory here. Override per run with the BABEL_DUCKDB_TEMP_DIR
# environment variable; see slurm/README.md ("Temporary Scratch Space") for the trade-offs.
tmp_directory: babel_downloads/tmp

# HTTP settings.
http: {}

# SPARQL query settings.
sparql:
  # Total number of attempts (initial try + retries) before giving up; must be >= 1.
  # Example: max_attempts=3 → one initial try, then up to two retries.
  max_attempts: 3
  # Base delay in seconds between attempts; must be >= 0. Actual delay grows exponentially
  # (attempt 1 → 1 s, attempt 2 → 2 s, attempt 3 → 4 s, …). Accepts fractional values.
  retry_base_delay_seconds: 1

# Maps Python compendium names (as used in src/createcompendia/ and tests) to the
# directory names that Snakemake uses under intermediate_directory.
# Compendia not listed here use their own name as the directory name.
# Update this whenever a new semantic type uses a shortened or different directory name.
compendium_directories:
  diseasephenotype: disease
  processactivitypathway: process

#
# SHARED
#

# DuckDB settings for use in all DuckDB connections.
# write_buffer_row_group_count=1: flush Parquet row groups to disk eagerly. The default of 5
# means DuckDB holds 5 row groups × threads × ~4 GiB in write buffers before flushing (~76 GiB
# peak at 128 GB allocation). Setting it to 1 keeps write-buffer RAM proportional to one row
# group, at no cost to compression ratio or query performance.
#
# enable_external_file_cache=false: DuckDB's external file cache keeps Parquet file blocks mmapped
# across queries, which is one source of memory mappings. The cross-compendium report rules were
# hitting the kernel's per-process mapping limit (vm.max_map_count, default 65530) and failing with
# `bad allocation` on a small allocation despite free RAM -- an address-space (mmap-count) limit.
# Disabling this cache removes it as a mapping source, but note it was NOT sufficient on its own:
# the dominant source is DuckDB's buffer pool (its allocator retains ~1.3 MB mappings up to the
# query's peak memory), so the report rules also cap memory_limit far below their natural peak. The
# definitive fix is for the cluster to raise vm.max_map_count (issue #846); until then we keep this
# off (it only costs a re-read from disk for our read-once/scan-twice queries, no correctness change).
duckdb_config:
  write_buffer_row_group_count: 1
  enable_external_file_cache: false

#
# UMLS
#
umls:
  subset: "full"
  # Replace with "level-0" to get only the level 0 subset (https://www.nlm.nih.gov/research/umls/licensedcontent/umlsknowledgesources.html).
  # subset: "level-0"

#
# PROTEINS
#

# Chris Bizon prepared a list of UMLS/UniProtKB mappings which we download and use.
UMLS_UniProtKB_download_raw_url: "https://raw.githubusercontent.com/cbizon/UMLS_UniProtKB/refs/heads/main/outputs/UMLS_UniProtKB.tsv"

#
# The rest of these configs need to be cleaned up.
#

ncbi_files:
  - gene2ensembl.gz
  - gene_info.gz
  - gene_orthologs.gz
  - gene_refseq_uniprotkb_collab.gz
  - mim2gene_medgen

generate_dirs_for_labels_and_synonyms_prefixes: [GO, CL, NCIT, UBERON, EMAPA, CHEBI, HP, MP, MONDO, PR]

# TODO: This doesn't appear to be used anywhere -- but it might be neat to use this list to make sure we got all of them.
ubergraph_ontologies: [UBERON, CL, EMAPA, GO, NCIT, ECO, ECTO, ENVO, HP, MP, UPHENO, BFO, BSPO, CARO, CHEBI, CP, GOREL, IAO, MAXO, MONDO, PATO, PR, RO, UBPROP]

mods: [WormBase, FB, MGI, ZFIN, RGD, SGD]

common:
  labels: [ubergraph/labels]
  synonyms: [ubergraph/synonyms.jsonl]
  descriptions: [ubergraph/descriptions.jsonl]

anatomy_prefixes:  [UBERON, GO, CL, EMAPA, UMLS, MESH, NCIT, SNOMEDCT]
anatomy_ids:       [UBERON, GO, CL, EMAPA, UMLS, MESH, NCIT]
anatomy_concords:  [UBERON, GO, CL, EMAPA, UMLS, WIKIDATA]
anatomy_outputs:   [AnatomicalEntity.txt, Cell.txt, CellularComponent.txt, GrossAnatomicalStructure.txt]
# Prefixes whose identifiers are treated as globally unique within a clique: if merging two
# cliques would produce a clique containing more than one identifier with one of these prefixes,
# the merge is blocked.  UBERON and GO are authoritative enough that two distinct terms from
# either ontology should never be collapsed into the same equivalence set.  If you add a new
# anatomy source whose concord file links to UBERON or GO, verify that it doesn't introduce
# spurious merges before removing either prefix from this list.
#
# EMAPA is deliberately NOT listed, even though it is an anatomy ids source.  122 UBERON terms
# cross-reference more than one EMAPA term (116 of them more than one *live* EMAPA term), and
# those are mostly genuine 1:n mappings (a mouse developmental partonomy resolving finer than the
# adult UBERON structure), not merge errors.
# Restricting EMAPA would block all 122 merges and would silently delete any EMAPA CURIE that
# lost the resulting contest without having its own ids-file row.  See
# docs/sources/EMAPA/README.md, which flags this for SME review.
anatomy_unique_prefixes: [UBERON, GO]

gene_labels:   [HGNC, NCBIGene, UMLS, OMIM]
gene_ids:      [ENSEMBL, HGNC, NCBIGene, UMLS, OMIM, ZFIN, WormBase, FB, MGI, RGD, SGD]
gene_concords: [NCBIGene, NCBIGeneENSEMBL, medgen, UMLS, UMLS_NCBIGene]
gene_outputs:  [Gene.txt]

protein_labels:   [UniProtKB, PR, UMLS]
protein_synonyms: [PR, UMLS]
protein_ids:      [ENSEMBL, MESH, UniProtKB, PR, UMLS]
protein_concords: [UniProtKB, PR, NCIT_UniProtKB, NCIT_UMLS, UMLS, UMLS_UniProtKB]
protein_outputs:  [Protein.txt]

disease_labelsandsynonyms: [MONDO, DOID, Orphanet, GARD, HP, MP, MESH, NCIT, UMLS, SNOMEDCT, EFO]
disease_ids:               [MONDO, DOID, Orphanet, GARD, HP, MP, MESH, NCIT, UMLS, OMIM, EFO]
# MONDO_GARD is MONDO's GARD hasDbXrefs, the one hasDbXref exception in this pipeline: MONDO
# asserts its ~15,930 GARD mappings only as hasDbXref, never as skos:exactMatch, so the MONDO
# concord (exact matches only) sees none of them. Kept as its own file so the exception is visible
# here and can be filtered without touching MONDO's exact matches -- docs/sources/MONDO/README.md.
#
# ORDER IS LOAD-BEARING: MONDO_GARD must precede DOID. 148 of DOID's 2,208 GARD xref rows put a
# GARD id on a different MONDO clique than MONDO itself does (e.g. DOID:0050120 -> GARD:6589, which
# MONDO maps to MONDO:0015541; docs/sources/GARD/scripts/count_doid_mondo_gard_disagreements.py). MONDO is a unique prefix, so glom() cannot merge the two cliques and the
# id stays with whichever concord claimed it first; listing MONDO's own mapping first is what makes
# MONDO, not DOID, the authority. A unit test pins the relative order.
#
# GARD_label must be LAST. It is the only concord in this pipeline derived from labels rather than
# from an asserted mapping, so it should decide nothing another concord has an opinion about; and
# its safety guard is "skip any GARD id another concord names", which is defined by the rest of
# this list. A unit test pins its position.
disease_concords:          [HP, MP, MONDO, MONDO_GARD, UMLS, DOID, EFO, Manual, GARD_label]
disease_outputs:           [Disease.txt, PhenotypicFeature.txt]

# The vocabularies build_gard_label_concord() may match a GARD label against, in PRIORITY ORDER:
# the first entry holding an identifier with the GARD term's label wins, and the term is skipped if
# that vocabulary holds two. MONDO/DOID/Orphanet lead because a curated disease ontology is a better
# thing to join than a Metathesaurus concept; UMLS is last of the large ones because it is the
# noisiest. Each entry names both an ids file (intermediate/disease/ids/<entry>, the authoritative
# identifier set for this pipeline) and a labels file (babel_downloads/<entry>/labels), so an entry
# must appear in BOTH disease_ids and disease_labelsandsynonyms -- a unit test checks that.
#
# HP and MP are deliberately absent. split_mutually_exclusive_cliques() keeps phenotype and disease
# cliques disjoint on purpose and disease_gard_ids types every registry term biolink:Disease, so
# matching a GARD term onto an HP or MP term asserts an identity this pipeline is built to refuse.
# OMIM is absent for a duller reason: it has an ids file but no babel_downloads/OMIM/labels.
# SNOMEDCT is absent because it has no disease ids file.
disease_gard_label_match_prefixes: [MONDO, DOID, Orphanet, MESH, NCIT, UMLS, EFO]

# Prefixes this pipeline ships even though the Biolink Model does not register them for the clique's
# class. write_compendium() otherwise drops an unregistered prefix silently, *after* glom() has
# already merged cliques on it -- so the merge effect ships and the identifier does not.
#
# KEYED BY BIOLINK CLASS, because extra_prefixes is a per-class allowlist and build_compendium
# writes several classes from one call. A single flat list passed to every class grants each of
# them an exemption that was argued for one of them; spelling the map out per class is what makes
# that unspellable. Look up each class's registered prefixes with
# get_biolink_model_toolkit(biolink_version).get_element(<class>).id_prefixes.
#
# EVERY ENTRY IS TEMPORARY AND EVERY ENTRY LINKS TO THE ISSUE THAT REMOVES IT. Once Biolink
# registers a prefix, its entry becomes a no-op that still suppresses the CURIE from leading its
# clique (extra prefixes are appended *after* the registered ones, so they can never win the
# preferred-CURIE contest) -- a stale entry is worse than dead config. Adding a prefix here without
# an issue link leaves nothing to close the loop. https://github.com/NCATSTranslator/Babel/issues/1061
# tracks the test that will fail when an entry goes stale.
#
# ICD0 is deliberately NOT here under any class: ICD-O codes are tumour *morphology* codes and may
# not be disease equivalents at all, so emitting them would assert something we have not decided.
# See https://github.com/NCATSTranslator/Babel/issues/1037.
disease_extra_prefixes_by_biolink_class:
  biolink:Disease:
    # MONDO's spelling of ICD-10 (DOID/EFO/HP write ICD10:), ~2,030 near-perfectly 1:1 xref rows.
    # Biolink registers ICD10 for biolink:Disease but not ICD10CM, so the fix is to normalise onto
    # the registered spelling rather than to ask for a second one; until then, shipping the rows
    # beats discarding curated MONDO mappings we would have to re-derive.
    # Retire with https://github.com/NCATSTranslator/Babel/issues/1060 (blocked on #1033).
    - ICD10CM
    # The NCATS rare-disease registry, unregistered for every Biolink class. Without this ~16k rare
    # diseases vanish from Disease.txt. GARD asserts no cross-references of its own; its ids reach
    # cliques through MONDO's hasDbXrefs (MONDO_GARD, ~15.9k), DOID's xrefs (~2.2k), and -- for the
    # ~277 terms neither maps -- the label matches in GARD_label. Same situation as GTDB for
    # biolink:OrganismTaxon (PR #978).
    # Retire with https://github.com/NCATSTranslator/Babel/issues/1051.
    - GARD
  biolink:PhenotypicFeature:
    # GARD again, and for the same reason: it is unregistered for *every* class, so this exemption
    # is not one earned on disease grounds and lent to phenotypes. disease_gard_ids types every
    # registry term biolink:Disease, but a handful name concepts HP also names, and the clique type
    # vote (classify_disease_clique) rightly trusts HP -- five terms today, "Chilblains" among them.
    # Without this entry those GARD ids are deleted rather than moved, so build_gard_label_concord()
    # had to refuse to link them at all and shipped a duplicate single-identifier Disease clique
    # beside the phenotype clique naming the same thing. One concept, one clique, wherever the
    # clique lands.
    # Retire with https://github.com/NCATSTranslator/Babel/issues/1051, the same issue as above.
    - GARD

# How each disease source spells *other* vocabularies' prefixes in its cross-references, and the
# Babel prefix each one becomes. Applied by babel_utils.norm() while the source's concord is built,
# keyed on the UPPER-CASED source prefix. Values are checked against src/prefixes.py when the map is
# loaded, so a typo fails the build rather than silently renaming nothing.
#
# A *missing* entry is not an error and that is the danger: the un-renamed CURIE still reaches
# glom(), joins nothing (no ids file carries that spelling, and write_compendium drops the prefix as
# unregistered), yet still fuses every subject that cites it into a single clique. That silent
# failure is why these maps sit here, in one reviewable block, instead of inline beside each build
# function -- reviewing "which prefixes does this source need renamed" is the whole job.
#
# A key may name the *stem* of a release-stamped prefix: SNOMEDCT_US matches DOID's
# SNOMEDCT_US_2025_09_01 and the six other dates it currently emits, without pinning a date list
# that goes stale on the next DOID release (norm() retries with a trailing _YYYY_MM_DD stripped).
disease_xref_prefixes:
  DOID:
    ICD10CM:     ICD10
    ICD9CM:      ICD9
    ICDO:        ICD0
    NCI:         NCIT
    SNOMEDCT_US: SNOMEDCT
    UMLS_CUI:    UMLS
    KEGG:        KEGG.DISEASE
    # DOID spells OMIM "MIM:" -- see https://github.com/NCATSTranslator/Babel/issues/321, which
    # tracks making MIM the canonical spelling once the Biolink Model registers it. Until then both
    # standardize to OMIM. 332 of these 6,483 rows are phenotypic series ("MIM:PS303350"), which
    # Babel spells OMIM.PS:303350 -- the "PS" belongs to the prefix, not to the local id. YAML
    # cannot express a rename that depends on the local id, so this value is resolved to a callable:
    # see LOCAL_ID_DEPENDENT_RENAMES in src/createcompendia/diseasephenotype.py.
    MIM:         OMIM
    # DOID's spelling of Orphanet, 2,321 rows. MONDO and HP already write orphanet:, so before this
    # entry none of DOID's Orphanet mappings could join theirs.
    ORDO:        orphanet
    # Not a prefix rename: the prefix stays GARD, but the local id loses the registry's zero
    # padding (GARD:0006038 -> GARD:6038; 28 of DOID's GARD xrefs are padded, the rest are not).
    # Resolved to gard.normalize_gard_curie via LOCAL_ID_DEPENDENT_RENAMES, like MIM above. Every
    # source that emits GARD xrefs needs this entry or its ids split from the registry's.
    GARD:        GARD
  HP:
    MSH:         MESH
    SNOMEDCT_US: SNOMEDCT
    SNOMED_CT:   SNOMEDCT
    # MONDO writes Orphanet mixed case ("Orphanet:2822"); the key is matched upper-cased, so this
    # one entry catches every casing and maps it to Babel's lower-case orphanet: prefix.
    ORPHANET:    orphanet
    ICD-9:       ICD9
    ICD-10:      ICD10
    ICD-0:       ICD0
    ICD-O:       ICD0
  MONDO:
    ORPHANET:    orphanet
    # MONDO writes every GARD id zero-padded; unpadded via LOCAL_ID_DEPENDENT_RENAMES (see DOID's
    # GARD entry). Feeds the MONDO_GARD concord.
    GARD:        GARD
# Taxon assigned to each phenotype ontology's identifiers in the compendia. Every HP term we
# ingest (the HP:0000118 "Phenotypic abnormality" subtree) describes a human (Homo sapiens)
# phenotype; every MP term describes a mammalian (Mammalia) phenotype. These are written to
# babel_downloads/<PREFIX>/taxa from each prefix's disease_ids file (the authoritative set of
# ingested identifiers) and read by TaxonFactory to populate the per-identifier `t` field. HP
# and MP are kept disjoint (see split_mutually_exclusive_cliques), so no clique ever carries
# both taxa in practice -- each clique's taxon set comes from whichever one of HP/MP (if
# either) is present. Keyed by prefix so the mapping stays adjacent to disease_ids, which
# defines the identifier set these taxa annotate.
disease_phenotype_taxa:
  HP: NCBITaxon:9606 # Homo sapiens
  MP: NCBITaxon:40674 # Mammalia

process_labels:   [GO, REACT, RHEA, EC, SMPDB, PANTHER.PATHWAY, UMLS]
process_ids:      [GO, REACT, RHEA, EC, SMPDB, PANTHER.PATHWAY, UMLS]
process_concords: [GO, RHEA, UMLS]
process_outputs:  [Pathway.txt, BiologicalProcess.txt, MolecularActivity.txt]

unichem_datasources: [CHEMBL.COMPOUND, DRUGBANK, GTOPDB, CHEBI, UNII, HMDB, PUBCHEM.COMPOUND, DrugCentral]  # KEGG.COMPOUND removed from UniChem — https://github.com/NCATSTranslator/Babel/issues/834

chemical_labels:    [CHEMBL.COMPOUND, GTOPDB, CHEBI, UNII, HMDB, PUBCHEM.COMPOUND, DrugCentral, UMLS, DRUGBANK]  # KEGG.COMPOUND removed from UniChem — https://github.com/NCATSTranslator/Babel/issues/834
chemical_synonyms:  [GTOPDB, CHEBI, UNII, HMDB, PUBCHEM.COMPOUND, UMLS, DRUGBANK]
chemical_concords:  [wikipedia_mesh_chebi, PUBCHEM_MESH, mesh_cas, mesh_unii, PUBCHEM_CAS, GTOPDB, CHEBI, UMLS, DrugCentral, RXNORM]
chemical_ids:       [CHEMBL.COMPOUND, GTOPDB, CHEBI, UNII, HMDB, PUBCHEM.COMPOUND, DrugCentral, DRUGBANK, MESH, UMLS, RXNORM]  # KEGG.COMPOUND removed from UniChem — https://github.com/NCATSTranslator/Babel/issues/834

# Food.txt holds DRUGBANK food entries retyped from ChemicalEntity to Food (issue #828); it must stay in
# this list so those cliques flow into DrugChemical conflation and the exports like any other chemical
# output. (The processed extracts among them go to the already-present ComplexMolecularMixture.)
chemical_outputs:   [MolecularMixture.txt, SmallMolecule.txt, Polypeptide.txt, ComplexMolecularMixture.txt, ChemicalEntity.txt, ChemicalMixture.txt, Drug.txt, Food.txt]

# Chemical clique type precedence, most preferred first. create_typed_sets() (src/createcompendia/
# chemicals.py) uses it in two places: as the tie-break when a clique's members vote for several types
# with equal counts, and to rank food/extract evidence against the clique's voted type (issues #828,
# #935). It does NOT decide the ordinary case -- the vote is majority-count first, and a clique whose
# PubChem members agree short-circuits the vote entirely -- so this list only settles ties.
#
# The ranking reflects Babel's chemical-tree design: identify as many biolink:SmallMolecule cliques as
# possible, and treat the vaguer types as fallbacks for cliques we could not resolve any better.
#
# biolink:Drug sits BELOW biolink:ChemicalEntity on purpose. It comes almost entirely from RxNorm drug
# formulations -- a level of detail neither Babel nor Translator needs -- so we keep it only for cliques
# where we could not merge the formulation with its active small molecule or mixture. Ranking it above
# ChemicalEntity puts RXCUI CURIEs ahead of chemical entities like UNII:PVI5M0M1GW "Filgrastim".
#
# biolink:Food sits below every structure-bearing type -- including the mixture types, which assert a
# composition a whole food does not -- so food evidence never demotes a defined molecule (NCIt
# classifies water, riboflavin, isoleucine and beta carotene as foods). It sits above only the
# uninformative ChemicalEntity it exists to improve on. Note that it therefore outranks Drug: a clique
# carrying food evidence that votes Drug is typed Food, which is mildly wrong -- a drug formulation is
# not a food. That is an ACCEPTED tradeoff, not an oversight: it does not occur in any build so far
# (scripts/replay_type_vote.py reports the count), and both a reorder and a special case cost more
# than the error does. Revisit only if Food starts appearing where a drug formulation belongs.
#
# The types marked "not produced today" never appear in intermediate/chemicals/partials/types, but a
# type missing from this list raises ValueError mid-build, so they are ranked here in advance. Counts
# are identifiers, not cliques, from the babel-1.18 build.
chemical_type_order:
  - biolink:SmallMolecule                 # 115,171,614 identifiers
  - biolink:Polypeptide                   # 15,282
  - biolink:NucleicAcidEntity             # not produced today
  - biolink:MolecularEntity               # not produced today
  - biolink:ComplexMolecularMixture       # 244
  - biolink:ChemicalMixture               # 675
  - biolink:MolecularMixture              # 11,924,009
  - biolink:ProcessedMaterial             # not produced today; the eventual home for extracts (#929)
  - biolink:FoodAdditive                  # not produced today
  - biolink:Food                          # from food evidence only, never from partials/types
  - biolink:EnvironmentalFoodContaminant  # not produced today
  - biolink:ChemicalEntity                # 862,541
  - biolink:Drug                          # 361,554 -- see above for why this is last

# A bunch of food related-configuration items.
#
# NCIt subtrees whose members are foods: NCIT:C1949 "Food" plus NCIT:C73913 "Seed" (nuts/grains like
# almond and wheat that NCIt files under Seed, not Food). NCIt types a food by what it is rather than what
# it came from, so this covers non-plant foods (scallop, venison, beef liver) too. These are facts about
# NCIt, not about any one consumer; today the only consumer is the DRUGBANK food-and-extract retype
# (issue #828), where a structureless entry counts as food material if its UNII is classified under either
# root OR its UNII carries a botanical-database flag (PLANTS/GRIN/MPNS, which reaches the plant materials
# NCIt has no food class for). Such material becomes biolink:Food, or biolink:ComplexMolecularMixture when
# it is an extract (see drugbank_extract_markers).
food_ncit_roots:
  - NCIT:C1949 # "Food" -- https://ncithesaurus.nci.nih.gov/ncitbrowser/ConceptReport.jsp?dictionary=NCI_Thesaurus&code=C1949
  - NCIT:C73913 # "Seed" -- https://ncithesaurus.nci.nih.gov/ncitbrowser/ConceptReport.jsp?dictionary=NCI_Thesaurus&code=C73913

# NCIt subtrees whose members are NEVER food, however botanical they are (issue #828). A botanical flag
# says "plant material", not "food", so on its own it must not overrule an NCIt class that says this is a
# drug: without this veto DRUGBANK:DB00965 "Ethiodized oil" (an iodinated poppy-seed-oil contrast agent)
# and DRUGBANK:DB05051 "BZL101" (an antineoplastic) are typed biolink:Food. A vetoed entry stays
# biolink:ChemicalEntity — we make no claim — and an explicit food_ncit_roots classification still wins, so
# a food that is also a diagnostic agent (inulin) is unaffected. Keep these roots NARROW: NCIT:C1909
# "Pharmacologic Substance" is deliberately absent because NCIt files cocoa butter, garlic oil, rice bran
# and green tea under it, and vetoing on it would throw out real foods.
nonfood_ncit_roots:
  - NCIT:C1966 # "Imaging Agent" -- https://ncithesaurus.nci.nih.gov/ncitbrowser/ConceptReport.jsp?dictionary=NCI_Thesaurus&code=C1966
  - NCIT:C274 # "Antineoplastic Agent" -- https://ncithesaurus.nci.nih.gov/ncitbrowser/ConceptReport.jsp?dictionary=NCI_Thesaurus&code=C274

# Case-insensitive name/synonym substrings that mark a DRUGBANK food material as a processed
# *extract*: such rows become biolink:ComplexMolecularMixture (an interim type — the eventual home is
# biolink:ProcessedMaterial once issue #929 adds that output) rather than biolink:Food. "allergen" was
# deliberately dropped as a marker (too broad); "extract" is the reliable signal (issue #828).
drugbank_extract_markers: [extract]

drugchemicalconflated_synonym_outputs: [DrugChemicalConflated.txt]
geneproteinconflated_synonym_outputs:  [GeneProteinConflated.txt]

taxon_labels:   [NCBITaxon, MESH, UMLS]
taxon_synonyms: [NCBITaxon, UMLS]
taxon_ids:      [NCBITaxon, MESH, UMLS]
taxon_concords: [NCBI_MESH, UMLS]
taxon_outputs:  [OrganismTaxon.txt]

cell_line_outputs: [CellLine.txt]

genefamily_labels:  [PANTHER.FAMILY, HGNC.FAMILY]
genefamily_ids:     [PANTHER.FAMILY, HGNC.FAMILY]
genefamily_outputs: [GeneFamily.txt]

umls_outputs: [umls.txt]

macromolecularcomplex_outputs: [MacromolecularComplex.txt]

ubergraph_iri_stem_to_prefix_map:
  "https://identifiers.org/ncbigene/": NCBIGene
  "http://www.ncbi.nlm.nih.gov/gene/": NCBIGene
  "http://www.genenames.org/cgi-bin/gene_symbol_report?hgnc_id=": HGNC
  "http://www.informatics.jax.org/marker/MGI:": MGI
  "http://www.pombase.org/spombe/result/": PomBase
  "http://www.wormbase.org/species/c_elegans/gene/": WormBase
  "http://flybase.org/reports/": FB
  "http://www.ecogene.org/gene/": ecogene
  "http://zfin.org/action/marker/view/": ZFIN
  "http://birdgenenames.org/cgnc/GeneReport?id=": cgnc
  "http://www.ensemblgenomes.org/id/": ENSEMBL
  "http://www.ensembl.org/id/": ENSEMBL
  "http://purl.obolibrary.org/obo/AISM_": AISM
  "http://purl.obolibrary.org/obo/BCO_": BCO
  "http://purl.obolibrary.org/obo/CDNO_": CDNO
  "http://purl.obolibrary.org/obo/CLAO_": CLAO
  "http://purl.obolibrary.org/obo/GNO_": GNO
  "http://purl.obolibrary.org/obo/HAO_": HAO
  "http://purl.obolibrary.org/obo/LEPAO_": LEPAO
  "http://purl.obolibrary.org/obo/MMO_": MMO
  "http://purl.obolibrary.org/obo/MRO_": MRO
  "http://purl.obolibrary.org/obo/MmusDv_": MmusDv
  "http://purl.obolibrary.org/obo/OARCS_": OARCS
  "http://purl.obolibrary.org/obo/OBA_": OBA
  "http://purl.obolibrary.org/obo/PCL_": PCL
  "http://purl.obolibrary.org/obo/PECO_": PECO
  "http://purl.obolibrary.org/obo/PPO_": PPO
  "http://purl.obolibrary.org/obo/TS_": TS
  "http://purl.obolibrary.org/obo/UO_": UO
  "http://purl.obolibrary.org/obo/PCO_": PCO
  "http://rgd.mcw.edu/rgdweb/report/gene/main.html?id=": RGD

publication_outputs: [Publication.txt]

geneprotein_outputs: [GeneProtein.txt]

drugchemical_outputs: [DrugChemical.txt]

preferred_name_boost_prefixes:
  biolink:ChemicalEntity: [DRUGBANK, DrugCentral, CHEBI, MESH, GTOPDB]

ensembl_datasets_to_skip:
#  - elucius_gene_ensembl
#  - hgfemale_gene_ensembl
#  - charengus_gene_ensembl
#  - otshawytscha_gene_ensembl
#  - aocellaris_gene_ensembl
  - omykiss_gene_ensembl        # 2026jun5: keeps getting stuck with this dataset for some reason.

# Labels longer than this limit are demoted (not used as preferred label if a shorter alternative exists).
# Keyed by Biolink type; types not listed here are never demoted. Uses ancestor traversal, so
# biolink:ChemicalEntity applies to all chemical subtypes (SmallMolecule, Drug, etc.).
# See https://github.com/NCATSTranslator/Babel/issues/597
demote_labels_longer_than:
  biolink:ChemicalEntity: 25