name: STRING Reference Ingest Guide # Information about the Source of the ingest source_info: infores_id: infores:string name: Search Tool for the Retrieval of Interacting Genes/Proteins (STRING) description: >- STRING (Search Tool for the Retrieval of Interacting Genes/Proteins) is a database of known and predicted protein–protein interactions hosted by SIB / EMBL / CPR. STRING aggregates evidence from seven channels — genomic neighborhood, gene fusion, phylogenetic co-occurrence, homology, co-expression, experiments, and text mining — plus curated databases, and combines them into a single ``combined_score`` per interaction pair. (STITCH, STRING's sister database for protein–chemical interactions, was developed alongside this ingest but split out for a later, properly-scoped effort — preserved on the ``stitch-ingest`` branch. See CHANGELOG 2026-05-28.) citations: - Szklarczyk D, Kirsch R, Koutrouli M, et al. The STRING database in 2023. Nucleic Acids Res. 2023;51(D1):D638–D646. doi:10.1093/nar/gkac1000 terms_of_use_info: terms_of_use_url: https://string-db.org/cgi/access terms_of_use_description: >- STRING data is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0). license_name: CC BY 4.0 license_url: https://creativecommons.org/licenses/by/4.0/ data_access_locations: - STRING downloads: https://stringdb-downloads.org/ - STRING main site: https://string-db.org/ data_provision_mechanisms: - file_download - other # conformance: was 'web_interface' - api_endpoint # conformance: was 'web_service_api' data_formats: - tsv # conformance: was 'TSV (gzip), space-delimited' data_versioning_and_releases: >- STRING uses dotted version numbers (currently v12.0). The version is published at https://string-db.org/api/json/version and is embedded in download URLs (e.g. ``protein.links.v12.0/``). ``get_latest_version()`` resolves the version dynamically from this endpoint. # Information about Ingested Content ingest_info: ingest_categories: - primary_knowledge_provider # conformance: was 'Primary Knowledge Provider' utility: >- STRING provides a broad protein–protein interaction graph derived from multiple evidence channels, supplying Translator reasoning with channel-specific "these proteins interact / are coexpressed / are genomic neighbors" signals across human, mouse, and rat. scope: >- Protein–protein interactions for three Translator-target species: human (NCBITaxon:9606), mouse (NCBITaxon:10090), and rat (NCBITaxon:10116). Protein identifiers are kept in their native ENSEMBL form (ENSP / ENSMUSP / ENSRNOP) — no mapping to gene-level identifiers as canonical IDs; NCBIGene equivalents are attached via ``equivalent_identifiers``. Per-channel evidence subscores drive per-channel predicate emission (one edge per high-confidence channel); the aggregate ``combined_score`` gates row inclusion. Future iterations may extend to additional species, surface per-channel subscores as edge properties, and add the STITCH protein–chemical layer — see ``future_considerations`` below. relevant_files: - file_name: 9606.protein.links.full.v12.0.txt.gz location: https://stringdb-downloads.org/download/protein.links.full.v12.0/ description: >- Human protein–protein interaction file, full variant (~154 MB gzipped, ~2x the 3-column .links file). 16 space-delimited columns: protein1, protein2, 7 evidence channels (neighborhood, fusion, cooccurence, homology, coexpression, experiments, database, textmining), 6 orthology-transferred variants of those channels, and combined_score. One row per directed pair (each unordered pair appears twice). We use the .full variant rather than the 3-column .links variant because per-channel subscores drive per-channel predicate emission (see edge_type_info below). - file_name: 10090.protein.links.full.v12.0.txt.gz location: https://stringdb-downloads.org/download/protein.links.full.v12.0/ description: >- Mouse (Mus musculus) protein–protein interaction file, full variant (~140 MB gzipped). Same 16-column shape as human. - file_name: 10116.protein.links.full.v12.0.txt.gz location: https://stringdb-downloads.org/download/protein.links.full.v12.0/ description: >- Rat (Rattus norvegicus) protein–protein interaction file, full variant (~139 MB gzipped). Same 16-column shape as human. All three PPI files drive the ``string_ppi`` tagged reader. - file_name: all_organisms.entrez_2_string.tsv location: https://string-db.org/mapping_files/entrez/ description: >- Universal STRING ↔ Entrez gene-ID mapping (~285 MB uncompressed, all species in one file). Three tab-separated columns: NCBI taxid, Entrez gene ID, STRING protein ID. Loaded once at transform start via ``on_data_begin`` and used to populate ``equivalent_identifiers`` on Protein nodes with their NCBIGene equivalents. Filtered to ``SUPPORTED_TAXA`` at load time so resident memory stays bounded. Not iterated by the reader; loaded as an auxiliary lookup file via ``koza.input_files_dir``. Same source file used by Monarch's STRING ingest. included_content: - file_name: "{taxid}.protein.links.full.v12.0.txt.gz (9606 / 10090 / 10116)" included_records: >- Rows where combined_score > 500 (STRING's medium-confidence cutoff). Symmetric duplicates (each unordered pair appears in both p1→p2 and p2→p1 form) are deduplicated on the sorted-pair key so each pair is emitted once. Same logic applies uniformly across all three per-organism files. fields_used: >- protein1, protein2 (parsed into ENSEMBL protein CURIEs — ENSP for human, ENSMUSP for mouse, ENSRNOP for rat — after stripping the ``{taxid}.`` taxon prefix); the 7 evidence channels (neighborhood, fusion, cooccurence, coexpression, experiments, textmining, plus homology and database for KL/AT) drive per-channel predicate and knowledge-level/agent-type selection; combined_score gates row inclusion. filtered_content: - file_name: "{taxid}.protein.links.full.v12.0.txt.gz (all three taxa)" filtered_records: >- Rows with combined_score <= 500 are excluded. rationale: >- Below medium confidence, STRING's predictions are too noisy for Translator reasoning. Matches the default confidence cutoff offered on the STRING web interface. - file_name: "{taxid}.protein.links.full.v12.0.txt.gz (all three taxa)" filtered_records: >- Symmetric duplicate rows (the second occurrence of any unordered protein pair). rationale: >- STRING lists each interaction in both directions. We emit one undirected edge per pair (per predicate). future_considerations: # Per-channel predicates AND per-channel KL/AT have shipped (the transform # reads ``.full.v12.0``, fires up to 6 channel-specific predicates with a # fallback above 750, and derives KL/AT from the dominant channel). # Remaining axes below. - category: edge_content consideration: >- Add the STITCH protein–chemical layer. Developed alongside this ingest as a ``stitch_pcl`` tagged reader but split out for a later, properly-scoped effort — the complete implementation (ChemicalEntity → interacts_with → Protein, ``CIDm``/``CIDs`` → PUBCHEM.COMPOUND parsing, human/mouse/rat) is preserved on the ``stitch-ingest`` branch. The reintegration should land mode-of-action predicates from ``actions.v5.0.tsv`` (activation, inhibition, binding, catalysis, …) from the start rather than the bare ``interacts_with`` form, since no production STITCH KP exists to merely match (confirmed: Automat's STRING-DB graph has no STITCH interaction edges — its only Protein↔ChemicalEntity edges are Ubergraph ontology backbone, ``primary_knowledge_source: infores:ubergraph``). relevant_files: "{taxon}.protein_chemical.links.v5.0.tsv.gz, {taxon}.actions.v5.0.tsv.gz" - category: edge_content consideration: >- Extend to additional Translator-relevant organisms beyond human/mouse/rat (zebrafish, fly, worm, yeast). The transform accepts arbitrary taxa via the ``SUPPORTED_TAXA`` dict in ``string.py``; adding a species is a one-line code change plus a download URL and a reader file entry. Note that non-vertebrate species may use protein ID schemes other than ENSEMBL (e.g. yeast uses ORF names like ``YAL001C``); the parser would accept them but downstream NodeNormalizer coverage should be probed first. relevant_files: "{taxon}.protein.links.full.v12.0.txt.gz" - category: edge_property_content consideration: >- Surface per-channel subscores as edge properties (not just predicate-driving signals). The channel scores currently drive predicate + KL/AT selection and are then discarded; downstream consumers don't see the actual COEXPRESSION/EXPERIMENTS/TEXTMINING numbers. Attach them as ``biolink:Attribute`` instances on each edge, named after the channel. The data is already in the ``.full`` file we read; this is a transform-only change. relevant_files: "{taxid}.protein.links.full.v12.0.txt.gz (already ingested)" - category: node_property_content # conformance: was 'node_content' consideration: >- Ingest ``protein.aliases.v12.0.txt.gz`` to populate node ``name`` and ``synonym`` properties. Currently nodes carry only the ENSEMBL ID as their identifier; biolink validation flags this with INFO-level "missing recommended name" warnings on every node. relevant_files: 9606.protein.info.v12.0.txt.gz, 9606.protein.aliases.v12.0.txt.gz additional_notes: >- Current scope: STRING protein–protein interactions for human, mouse, and rat. STRING emits 6 per-channel predicates (physically_interacts_with, coexpressed_with, interacts_with, genetic_neighborhood_of, gene_fusion_with, genetically_interacts_with) plus a fallback ``physically_interacts_with`` for rows that pass the combined-score gate without any high-confidence channel. knowledge_level / agent_type are derived per row from the dominant evidence channel (ORION-aligned). The predicate-selection rule matches ORION's STRING parser to maintain downstream compatibility with the existing Translator KP ecosystem (Automat STRING-DB, RoboKOP). HOMOLOGY channel is deliberately excluded from predicate emission because STRING's HOMOLOGY score means "inferred via orthologs in another species" not "A is homologous to B." The STITCH protein–chemical sibling ingest was split out for a later effort (preserved on the ``stitch-ingest`` branch); see ``future_considerations`` and CHANGELOG 2026-05-28. Other growth axes: per-channel subscores as edge properties, additional non-mammalian species, and node names from protein.aliases. # Information about the Graph Output by the ingest target_info: infores_id: infores:string edge_type_info: # STRING per-channel edges (``string_ppi`` tag). # # The transform reads STRING's ``protein.links.full.v12.0.txt.gz`` (16 # columns: 7 channels + 6 _transferred variants + combined_score) and # emits one edge per evidence channel whose score exceeds the # high-confidence per-channel threshold (>750). Rows are gated by # ``combined_score > 500``. If no channel exceeds 750 but the row # passes the combined-score gate, a single fallback # ``physically_interacts_with`` edge is emitted (matches ORION's # behavior). Up to 6 edges per pair; typically 1–2. # # The HOMOLOGY channel is intentionally excluded from predicate # emission: STRING's HOMOLOGY score means "interaction inferred via # orthologs in another species", not "A is homologous to B" — so # mapping it to ``biolink:homologous_to`` would be a semantic # misread (per ORION's comment). DATABASE is similarly excluded. # # Common attributes across all 6 edge types: subject/object categories # = biolink:Protein; species_context_qualifier ∈ {9606, 10090, 10116}; # primary_knowledge_source = infores:string. # # knowledge_level / agent_type are a ROW-level property, derived from the # dominant (max-score) evidence channel rather than from the predicate # (matches ORION's STRING parser). So any edge type can carry any of the # values below — the enumeration is the same across all 6 types: # EXPERIMENTS/DATABASE dominant → knowledge_assertion / manual_agent # COEXPRESSION/COOCCURENCE dominant → statistical_association / data_analysis_pipeline # NEIGHBORHOOD/FUSION dominant → prediction / data_analysis_pipeline # HOMOLOGY dominant → prediction / computational_model # TEXTMINING dominant → not_provided / text_mining_agent # ≥2 channels above 750 → knowledge_assertion / manual_agent (or # data_analysis_pipeline if no curator-backed channel is high-conf) # # Per-edge differences: predicate, biolink class, and has_attribute # (only physically_interacts_with carries MI:0915). - subject_categories: - biolink:Protein predicates: - biolink:physically_interacts_with object_categories: - biolink:Protein qualifiers: - property: species_context_qualifier value_enumeration: - NCBITaxon:9606 - NCBITaxon:10090 - NCBITaxon:10116 knowledge_level: # row-level, from dominant channel (see header note) - knowledge_assertion - statistical_association - prediction - not_provided agent_type: - manual_agent - data_analysis_pipeline - computational_model - text_mining_agent - not_provided edge_properties: # PSI-MI MI:0915 (physical association) attached only here — the # semantically matched predicate. Other per-channel edges below # omit has_attribute. - has_attribute: [MI:0915] primary_knowledge_sources: [infores:string] ui_explanation: >- Driven by the EXPERIMENTS channel (when score > 750) and as the fallback predicate when no channel exceeds 750. Represents direct experimental evidence of physical binding between two proteins. source_files: - 9606.protein.links.full.v12.0.txt.gz - 10090.protein.links.full.v12.0.txt.gz - 10116.protein.links.full.v12.0.txt.gz - subject_categories: - biolink:Protein predicates: - biolink:coexpressed_with object_categories: - biolink:Protein qualifiers: - property: species_context_qualifier value_enumeration: - NCBITaxon:9606 - NCBITaxon:10090 - NCBITaxon:10116 knowledge_level: # row-level, from dominant channel (see header note) - knowledge_assertion - statistical_association - prediction - not_provided agent_type: - manual_agent - data_analysis_pipeline - computational_model - text_mining_agent - not_provided primary_knowledge_sources: [infores:string] ui_explanation: >- Driven by the COEXPRESSION channel (score > 750). Indicates that the two proteins' coding genes are transcriptionally co-expressed across conditions. Emitted as a ``biolink:GeneToGeneCoexpressionAssociation`` rather than ``biolink:PairwiseMolecularInteraction`` because biolink does not classify coexpression as a molecular interaction. source_files: - 9606.protein.links.full.v12.0.txt.gz - 10090.protein.links.full.v12.0.txt.gz - 10116.protein.links.full.v12.0.txt.gz - subject_categories: - biolink:Protein predicates: - biolink:interacts_with object_categories: - biolink:Protein qualifiers: - property: species_context_qualifier value_enumeration: - NCBITaxon:9606 - NCBITaxon:10090 - NCBITaxon:10116 knowledge_level: # row-level, from dominant channel (see header note) - knowledge_assertion - statistical_association - prediction - not_provided agent_type: - manual_agent - data_analysis_pipeline - computational_model - text_mining_agent - not_provided primary_knowledge_sources: [infores:string] ui_explanation: >- Driven by the TEXTMINING channel (score > 750). Text-mining co-mention evidence is loose semantically (the proteins co-occur in literature) so the predicate is the generic ``biolink:interacts_with`` rather than something stronger. source_files: - 9606.protein.links.full.v12.0.txt.gz - 10090.protein.links.full.v12.0.txt.gz - 10116.protein.links.full.v12.0.txt.gz - subject_categories: - biolink:Protein predicates: - biolink:genetic_neighborhood_of object_categories: - biolink:Protein qualifiers: - property: species_context_qualifier value_enumeration: - NCBITaxon:9606 - NCBITaxon:10090 - NCBITaxon:10116 knowledge_level: # row-level, from dominant channel (see header note) - knowledge_assertion - statistical_association - prediction - not_provided agent_type: - manual_agent - data_analysis_pipeline - computational_model - text_mining_agent - not_provided primary_knowledge_sources: [infores:string] ui_explanation: >- Driven by the NEIGHBORHOOD channel (score > 750). Indicates that the coding genes of the two proteins are genomic neighbors (commonly co-located in prokaryotic operons or conserved gene clusters across species). source_files: - 9606.protein.links.full.v12.0.txt.gz - 10090.protein.links.full.v12.0.txt.gz - 10116.protein.links.full.v12.0.txt.gz - subject_categories: - biolink:Protein predicates: - biolink:gene_fusion_with object_categories: - biolink:Protein qualifiers: - property: species_context_qualifier value_enumeration: - NCBITaxon:9606 - NCBITaxon:10090 - NCBITaxon:10116 knowledge_level: # row-level, from dominant channel (see header note) - knowledge_assertion - statistical_association - prediction - not_provided agent_type: - manual_agent - data_analysis_pipeline - computational_model - text_mining_agent - not_provided primary_knowledge_sources: [infores:string] ui_explanation: >- Driven by the FUSION channel (score > 750). Indicates that the two proteins' coding genes appear as a single fused gene in at least one other genome — a strong indicator of functional association. source_files: - 9606.protein.links.full.v12.0.txt.gz - 10090.protein.links.full.v12.0.txt.gz - 10116.protein.links.full.v12.0.txt.gz - subject_categories: - biolink:Protein predicates: - biolink:genetically_interacts_with object_categories: - biolink:Protein qualifiers: - property: species_context_qualifier value_enumeration: - NCBITaxon:9606 - NCBITaxon:10090 - NCBITaxon:10116 knowledge_level: # row-level, from dominant channel (see header note) - knowledge_assertion - statistical_association - prediction - not_provided agent_type: - manual_agent - data_analysis_pipeline - computational_model - text_mining_agent - not_provided primary_knowledge_sources: [infores:string] ui_explanation: >- Driven by the COOCCURRENCE channel (score > 750). Indicates that the two proteins' coding genes show a similar phylogenetic distribution across organisms — present together or absent together — a phylogenetic-profile signal of functional coupling. source_files: - 9606.protein.links.full.v12.0.txt.gz - 10090.protein.links.full.v12.0.txt.gz - 10116.protein.links.full.v12.0.txt.gz node_type_info: - node_category: biolink:Protein # Source identifier types reflect the canonical post-normalization identifier # set on Protein nodes in the merged Translator graph (as seen on Automat's # STRING-DB KP). The ingest itself emits ``ENSEMBL`` as the node id and # populates ``equivalent_identifiers`` with ``NCBIGene`` CURIEs from the # entrez_2_string mapping file. UniProtKB and PR equivalents are filled in # downstream by the Translator NodeNormalizer (~98.5% of human ENSPs resolve; # the remainder are retired ENSEMBL IDs that STRING v12.0 still ships). source_identifier_types: - ENSEMBL - UniProtKB - PR node_properties: - in_taxon - equivalent_identifiers additional_notes: >- Protein nodes use STRING's native identifiers — ENSEMBL protein IDs that vary by species: ENSP (human), ENSMUSP (mouse), ENSRNOP (rat). The leading ``{taxid}.`` prefix is stripped and the result is prefixed with ``ENSEMBL:``. Each node carries ``in_taxon`` derived from the row's taxon prefix, and (when a STRING ↔ Entrez mapping exists) ``equivalent_identifiers`` listing one or more ``NCBIGene:`` CURIEs. This explicit equivalence is the hybrid design choice that aligns with both Monarch's gene-endpoint ingest pattern (Kevin Schaper, 2026-05) and the Translator NodeNormalizer's expectations, without forcing a gene-centric reshape of the graph. NodeNormalizer resolution verified at 98.5% (human), 100% (mouse), 98% (rat) on sample probes. # Information about how the ingest was specified and performed provenance_info: contributions: - Kenneth Huellas-Bruskiewicz: ingest owner, code author - Richard Bruskiewich: data modeling, code co-author artifacts: - Ingest Survey: https://docs.google.com/spreadsheets/d/13Q4uJ90UFCXyXZMuhgtm1nREGtIIMsfDEgb4jpkgRIA/edit?gid=0#gid=0 - Ingest Ticket: https://github.com/NCATSTranslator/translator-ingests/issues/377 - Monarch STRING reference: https://github.com/monarch-initiative/string-ingest/blob/main/src/protein_links.py - ORION STRING reference: https://github.com/RobokopU24/ORION/blob/master/parsers/STRING/src/loadSTRINGDB.py - Automat STRING-DB knowledge provider (RENCI): https://automat.renci.org/string-db/