feat: add eqtl-catalogue-region-fetch skill (cis-QTL region tabix fetcher)

[madaraviv]

What this PR adds

A standalone eqtl-catalogue-region-fetch skill: tabix-on-FTP region fetcher for EBI's eQTL Catalogue v7+. Returns harmonised per-variant cis-QTL summary stats (variant_id, chromosome, position, ref, alt, beta, SE, p-value, MAF, molecular_trait_id, study_id) for any one (study × tissue × quantification) dataset in the catalogue, suitable for downstream colocalization, fine-mapping, regional plotting, or Mendelian randomisation.

Coverage gap this fills

The current K-Dense bio/genomics coverage doesn't include a region-level cis-QTL fetcher. database-lookup covers point queries to GTEx and eQTL Catalogue (per-variant) but not the per-region pulls needed for coloc / fine-mapping / regional plots. This skill complements that — point lookups stay in database-lookup; window-level summary-stats slices come here.

Why a tabix-on-FTP path (not REST)

The eQTL Catalogue REST API at /api/v2/datasets/{id}/associations silently truncates region requests to one side of the TSS — the pos_min / pos_max filters are ignored, returning only the genomic-lower hits (verified May 2026). The tabix-indexed .all.tsv.gz files on EBI's FTP contain the full ±1 Mb of strand-aware cis-window per gene as designed; this skill uses that path. REST is kept only for dataset metadata.

Coverage of catalogue's quant methods

All expression-related QTL flavors the catalogue hosts: gene-level (ge), exon-level (exon), transcript-level (tx), transcript-event usage (txrev, txrevise tool), splicing (leafcutter), microarray, and allelic fold-change (aFC). pQTL data is not in this catalogue (UKB-PPP / deCODE are the canonical pQTL sources).

Layout

Follows the K-Dense convention (light SKILL.md frontmatter + workflow-doc body + references/ + scripts/ + tests/), modelled after pydeseq2/ and exa-search/:

scientific-skills/eqtl-catalogue-region-fetch/
├── SKILL.md                          # Overview, Trigger, Scope, Workflow,
│                                     # CLI Reference, Example Output,
│                                     # Gotchas, Safety, Agent Boundary
├── references/
│   ├── dataset_resolution.md         # how to resolve human description /
│   │                                 # OT studyId slug → QTD###### dataset_id
│   ├── quant_methods.md              # the 6 catalogue quant methods + per-
│   │                                 # method β interpretation + pitfalls
│   └── cis_window_biology.md         # strand-aware ±1 Mb of TSS, REST API
│                                     # truncation gotcha, FTP tabix path
├── scripts/
│   ├── eqtl_catalogue_region_fetch.py    # CLI: --input, --output, --demo,
│   │                                     # --list-demos. PEP 723 inline deps.
│   └── examples/
│       ├── sort1_gtex_minor_salivary_gland.json  # SORT1 1p13.3 LDL/CHD locus
│       ├── il6r_gtex_small_intestine.json        # IL6R Swerdlow 2012 MR-classic
│       └── irf5_gtex_adipose_visceral.json       # IRF5 Wang 2021 SLE eQTL
└── tests/
    └── test_eqtl_catalogue_region_fetch.py       # 18 stdlib unittest cases
                                                  # (pure helpers, row normaliser,
                                                  # demo loader, CLI plumbing) —
                                                  # no network, no extra deps

Cache writes to ~/.cache/eqtl_catalogue_region_fetch/ (overridable via EQTL_CATALOGUE_CACHE_DIR), so repeated calls skip the FTP round-trip.

Also adds an eQTL Catalogue Region Fetch entry under ## Scientific Databases & Data Access in docs/scientific-skills.md, next to Database Lookup.

Try it

# Run via uv (PEP 723 inline deps, no separate install)
uv run --with pysam,pandas,requests \
    scientific-skills/eqtl-catalogue-region-fetch/scripts/eqtl_catalogue_region_fetch.py \
    --demo sort1_gtex_minor_salivary_gland --output /tmp/sort1_run

# Or with an existing env that has pysam, pandas, requests
cd scientific-skills/eqtl-catalogue-region-fetch
python scripts/eqtl_catalogue_region_fetch.py --list-demos
python scripts/eqtl_catalogue_region_fetch.py \
    --demo sort1_gtex_minor_salivary_gland --output /tmp/sort1_run
ls /tmp/sort1_run/  # variants.tsv, manifest.yaml, report.md

Verifications

• Cisco AI Defense Skill Scanner (skill-scanner scan scientific-skills/eqtl-catalogue-region-fetch --use-behavioral): SAFE / 0 findings
• Unit tests: python -m unittest discover -s scientific-skills/eqtl-catalogue-region-fetch/tests -v → 18/18 passing (no network, stdlib only)
• Smoke test: 2,833 SORT1 cis-eQTL variants in minor salivary gland (±500 kb of TSS); 3,276 IRF5 variants in visceral adipose (±500 kb of the OT coloc lead at chr7:128,937,250)
• Network-only on first call per (dataset, region); subsequent calls hit the local cache
• Python 3.10+ compatible

Test plan

• Scripts run cleanly (--help, --list-demos, --demo against live FTP)
• Unit tests pass locally (python -m unittest discover -s tests -v → 18/18)
• Cisco scanner SAFE / 0 findings (skill-scanner scan ... --use-behavioral)
• No __pycache__/, .tbi, .tsv.gz, or run-output dirs tracked (.gitignore in skill root)
• Live smoke against the 3 bundled demos (reviewer step if desired)

Coming next: 3 more skills to enable coloc signal inspection

This is skill 1 of 4 in a suite enabling regional LocusCompare diagnostics — the visual sanity check that distinguishes a real shared causal variant from an LD artefact, and the gold standard for genetic target support in drug discovery and drug development. The other three (separate PRs):

• gwas-catalog-region-fetch — tabix on GWAS Catalog harmonised summary stats (disease/trait outcome side)
• ld-1000g-region-compute — plink2 r² against 1000G Phase 3 GRCh38 (LD coloring)
• locuscompare-region-render — orchestrator composing the three primitives + GENCODE gene track into the canonical 4-panel LocusCompare plot (Liu 2019 Nat Methods) + JSON manifest

This PR is self-contained; the orchestrator depends on the three primitives, but each primitive is independently useful for non-LocusCompare asks.

Sources & licensing

• eQTL Catalogue v7+ (Kerimov 2021 Nat Genet 53:1290), CC-BY-4.0
• Per-study attribution: original publication of each constituent dataset
• Skill code: MIT

---

🤖 Generated with Claude Code