/target-prioritization

API endpoints used by target-prioritization

All endpoints are free and require no API key.

UniProt REST

• Base: https://rest.uniprot.org/uniprotkb/search
• Query format: (gene:SYMBOL) AND (organism_id:9606) AND (reviewed:true)
• Fields requested (comma-separated): accession, gene_names, protein_name, cc_subcellular_location, protein_existence, keyword, ft_topo_dom, ft_transmem, ft_signal, cc_function
• Rate: ~100 req/sec; we sleep 0.1s between calls to be polite.
• Docs: https://www.uniprot.org/help/api_queries

OpenTargets GraphQL

• Endpoint: https://api.platform.opentargets.org/api/v4/graphql
• Two queries:
  1. search(queryString: SYMBOL, entityNames: ["target"]) to resolve symbol → Ensembl ID
  2. target(ensemblId: ENSG…) to get tractability, drugAndClinicalCandidates (returns rows with maxClinicalStage like APPROVAL, PHASE_3, etc.), associatedDiseases (integrates GWAS Catalog + other genetics evidence sources — so we do NOT call GWAS Catalog directly), depMapEssentiality (per-tissue + per-cell-line CRISPR geneEffect — we do NOT call the DepMap portal API directly), and geneticConstraint (gnomAD-derived LOEUF / oe_lof / upperBin — we do NOT call the gnomAD GraphQL directly, which is also useful for avoiding its WAF).
• Stage tokens are uppercase: APPROVAL (=approved), PHASE_4..PHASE_1, EARLY_PHASE_1, PRECLINICAL, WITHDRAWN, UNKNOWN. Mapping lives in PHASE_MAP inside fetch_opentargets.py.
• Tractability modality codes: SM (small molecule), AB (antibody), PR (protein degrader/other), OC (other clinical). Each modality has ordered tier labels — we pick the highest tier with value=True.
• Rate: generous; 0.2s sleep between targets to avoid hammering.
• Docs: https://platform-docs.opentargets.org/data-access/graphql-api

Human Protein Atlas

• Symbol → Ensembl resolver: https://www.proteinatlas.org/api/search_download.php?search=SYMBOL&format=json&columns=g,eg&compress=no
• Per-entry JSON: https://www.proteinatlas.org/<ENSG>.json (≈100 fields)
• Subcellular fields are not consumed here — UniProt is authoritative. HPA contributes:
  • RNA tissue specificity (tag like Tissue enriched / Group enriched / Tissue enhanced / Low tissue specificity / Not detected) and RNA tissue specific nTPM (dict tissue→nTPM)
  • RNA single cell type specificity (analogous tag) and RNA single cell type specific nCPM (dict cell-type→nCPM)
  • Single cell expression cluster (label of HPA's UMAP cluster)
  • Cancer prognostics - <cancer name> per-cancer dicts — aggregated into n_prognostic_cancers + top-3 cancers with prognostic type + p_val, plus the global RNA cancer specificity tag
• FOCUS_CELL_TYPES in scripts/aggregate.py must match HPA's exact cell-type strings (e.g. "T-cells", "Hepatocytes", "Microglial cells").
• No documented rate limit; fetcher sleeps 0.15s between calls.
• Docs: https://www.proteinatlas.org/about/help/dataaccess

PubMed E-utilities

• Endpoint: https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi
• Three queries per gene: total, focus_disease, cell_context. The focus_disease and cell_context query templates live in scripts/fetch_pubmed.py::CONTEXTS and can be re-pointed to any domain (oncology, neurodegeneration, metabolic, etc.) by editing the string.
• Rate: 3 req/sec without API key; we sleep 0.34s between requests. With an NCBI API key (env NCBI_API_KEY), up to 10 req/sec — not implemented here, easy to add.
• Maturity bins (configurable thresholds):
  • <5 = uncharted
  • 5-29 = novel
  • 30-99 = moderate
  • 100-499 = well_studied
  • ≥500 = saturated

ChEMBL REST

• Two HTTP calls per gene:
  1. https://www.ebi.ac.uk/chembl/api/data/target/search.json?q=SYMBOL&limit=10 — filter results client-side to organism == "Homo sapiens" and target_type == "SINGLE PROTEIN"; prefer exact pref_name / synonym match.
  2. https://www.ebi.ac.uk/chembl/api/data/activity.json?target_chembl_id=ID&standard_type=IC50&pchembl_value__gte=7&limit=5&order_by=-pchembl_value — pulls the 5 most-potent IC50 assays with pIC50 ≥ 7 (≈ 100 nM).
• Output is dossier-only: top compound ID, pIC50, IC50 (nM), optional molecule_pref_name (named tool compounds like MOBOCERTINIB).
• No API key, ~5 req/sec friendly; fetcher sleeps 0.2s/gene.
• Docs: https://www.ebi.ac.uk/chembl/api/data/docs

Adding a new source

To plug in another evidence source (e.g. DGIdb, OMIM, OpenTargets Genetics, Reactome):

1. Write scripts/fetch_<source>.py accepting --genes SYM,SYM,… and --output PATH.json, writing {gene: {fields}}.
2. Add the (name, script_path) tuple to FETCHERS in orchestrate.py.
3. Add a component term to compute_components in aggregate.py.
4. Add corresponding weights: and (if needed) flags: keys to weights.yaml.
5. Surface the new fields in the markdown skeleton in aggregate.py.

No re-fetch needed for downstream re-runs — the raw JSON cache makes re-scoring effectively free.