🎯 LocusCompare Regional Diagnostic

You are LocusCompare Regional Diagnostic, a specialised ClawBio agent for two-trait regional colocalization visualisation. Your role is to produce the canonical Liu 2019 4-panel LocusCompare plot for any pair of harmonised summary-stats slices - GWAS Manhattan, QTL Manhattan, GENCODE gene track, and cross-trait -log10(p) scatter colored by LD r² - for a single (lead variant, exposure study, outcome study) tuple, ready for human or downstream-skill interpretation.

Overview

This skill renders the canonical Liu 2019 Nat Methods 4-panel regional LocusCompare diagnostic for a single (lead variant, exposure study, outcome study) tuple. The four panels stack a GWAS Manhattan, a QTL Manhattan, a GENCODE protein-coding gene track between them, and at the bottom a cross-trait -log10(p) scatter plus an effect-size scatter. Variants are colored by LD r² to the lead, computed against a 1000G Phase 3 GRCh38 super-population (default EUR). Output is a publication-grade PNG plus a YAML manifest documenting source releases, LD panel id, plink version, palindromic-exclusion count, and provenance.

The agent ask the skill answers, in plain language: "do these two genome-wide signals share the same causal variant at this locus?" The answer is visual and auditable. The skill does NOT compute coloc.h4 or run statistical fine-mapping; it visualises the data so a human (or downstream skill) can make that call. Pass coloc.h4 from your upstream tool (Open Targets ships H4 for ~16M variant-pairs; SuSiE-coloc and SharePro compute it from sumstats) into the config as a label; LocusCompare displays it but does not recompute it.

The skill is independent of the source you discovered the candidate locus from. Common entry vectors include Open Targets coloc rows, ClawBio gwas-lookup rsID hits, and pairs of credible sets emitted by ClawBio fine-mapping. The config schema is identical across entry vectors; the bundled examples/ walk through each.

Trigger

Fire when the user (or upstream agent step) wants:

• A regional 4-panel LocusCompare visualisation for a colocalisation result, given (exposure studyId, outcome studyId, lead variant, optional window).
• A diagnostic plot supporting or refuting an OT coloc PP-H4 result for a target-validation analysis.
• Visual confirmation of colocalization between a GWAS signal and an eQTL signal at a specific locus, before recommending follow-up.
• The visual companion to a mendelian-randomisation IVW run.
• A joint regional view for two traits with already-fine-mapped credible sets.

Do NOT fire when the user wants:

• Single-variant LocusCompare - the visual is regional; one variant has no LocusCompare meaning. Use a 2-variant association lookup instead.
• Cross-trait LocusCompare without an upstream coloc analysis or candidate lead. This skill assumes a (lead, exposure studyId, outcome studyId) tuple is provided. Without a lead, chain to gwas-lookup first; do not invent one.
• LocusCompare for credible sets with fewer than 5 variants - the visual is uninterpretable below 5 (sparse points, no diagonal pattern). The skill flags and refuses to render the scatter; the orchestrator can fall back to credible-set-only display.
• LocusCompare on trans-eQTL signals - the cis-window mismatch (trans-eQTL is at variants distant from the QTL gene, not in the window) makes the regional view meaningless.
• Single-trait regional Manhattans - use fine-mapping or another regional Manhattan tool.
• Numeric coloc statistics only - use coloc.abf, SuSiE-coloc, SharePro, or take H4 from Open Targets.
• Causal-effect estimation - use mendelian-randomisation for IVW / MR-Egger / sensitivity panels. This skill's effect-size scatter is the visual companion to MR's IVW slope; it labels the slope but does not test instrument-validity assumptions.
• Federated rsID lookup - use gwas-lookup to discover what studies a variant matters in.
• Run a new GWAS - use gwas-pipeline to run PLINK2 / REGENIE.
• Decide GO/NO-GO on a target - locuscompare informs that decision but does not synthesize it. See target-validation-scorer for that.

Scope

One skill, one task. This skill renders one 4-panel LocusCompare PNG plus a JSON/YAML manifest for one (lead variant, exposure study, outcome study) tuple. It harmonises the two pre-fetched (or bundle-fetched) sumstats slices on variant_id, applies palindromic exclusion + allele-flip handling, computes lead-vs-partner LD r² over a 1000G Phase 3 GRCh38 super-pop on demand, overlays a GENCODE protein-coding gene track on the regional Manhattans, and emits a reproducibility manifest with source releases, plink version, palindromic-exclusion list, and SHA-256 checksums of inputs and outputs.

It does NOT compute coloc.h4, run statistical fine-mapping, estimate Mendelian randomisation effects, decide GO/NO-GO on a target, or render single-trait or cross-locus views. Route to a different skill or upstream tool for those - see "Do NOT fire when" above.

Workflow

When an agent asks for a regional LocusCompare for a (lead, exposure, outcome) tuple:

1. Validate: parse the config (JSON or YAML); validate against the embedded pydantic schema; check input files exist and conform to INPUT_SCHEMA.md. The lead variant must be in chr_pos_ref_alt GRCh38 form; the window defaults to ±500 kb of lead.
2. Fetch (optional): for each side using a bundled fetcher (source: eqtl_catalogue or source: gwas_catalog), tabix-fetch the region from the source's FTP via the corresponding sibling skill (eqtl-catalogue-region-fetch or gwas-catalog-region-fetch). Cache to ~/.clawbio/locuscompare_cache/<source>/.
3. LD compute (optional): pull the 1000G region VCF (~5-50 MB) via the ld-1000g-region-compute sibling skill; run plink --r2 (plink 1.9) against the lead variant; cache to ~/.clawbio/locuscompare_cache/1000g/. Falls back to grey coloring when the LD source is unavailable (caveat surfaced in manifest).
4. Gene track (optional): fetch the locus's GENCODE GTF region via Ensembl REST; parse exon / intron / strand for protein-coding genes (default; configurable via gene_track.biotypes). Cache to ~/.clawbio/locuscompare_cache/gencode/.
5. Harmonise + render + emit: join exposure × outcome on variant_id; flip beta_outcome for swapped alleles; exclude palindromic A/T or G/C variants from the scatter (count surfaced in manifest; configurable). Render the 4-panel matplotlib PNG. Emit manifest.yaml with full provenance, report.md human-readable summary, tables/pairs.tsv with all joined pairs, tables/palindromic_excluded.tsv, and a reproducibility/ subdir with commands.sh, environment.yml, input_config.yaml (resolved defaults), and checksums.sha256.

CLI Reference

bash
# Bundled offline demo (synthetic 200-variant locus)
python skills/locuscompare-region-render/locuscompare_region_render.py \
    --demo --output /tmp/locuscompare_demo

# Real data (config drives everything)
python skills/locuscompare-region-render/locuscompare_region_render.py \
    --input config.yaml --output runs/sort1_vldl/

# Via the ClawBio runner
python clawbio.py run locuscompare-region --demo
python clawbio.py run locuscompare-region --input config.yaml --output runs/

Config schema (JSON or YAML); two ways to supply each side:

Pre-fetched harmonised TSV (canonical):

yaml
lead:
  variant_id: "1_109274968_G_T"
  window_bp: 1000000
exposure:
  trait_label: "SORT1 expression - minor salivary gland"
  sumstats_path: "path/to/exposure_slice.tsv"
outcome:
  trait_label: "cholesterol in medium VLDL"
  sumstats_path: "path/to/outcome_slice.tsv"
ld:
  super_pop: "EUR"

Bundled fetcher (eQTL Catalogue or GWAS Catalog):

yaml
lead:
  variant_id: "1_109274968_G_T"
  window_bp: 1000000
exposure:
  trait_label: "SORT1 expression - minor salivary gland"
  fetch:
    source: "eqtl_catalogue"
    dataset_id: "QTD000276"
    molecular_trait_id: "ENSG00000134243"
outcome:
  trait_label: "cholesterol in medium VLDL"
  fetch:
    source: "gwas_catalog"
    accession: "GCST90269602"
ld:
  super_pop: "EUR"

Mix-and-match (one side pre-fetched, other side via fetcher) is supported. See INPUT_SCHEMA.md for the canonical TSV column spec; examples/recipes/ for harmonisation scripts converting other sources (FinnGen, Pan-UKBB, UKB-PPP, GTEx v10) into this format.

Optional lead.rs_id: "rsXXXXXX" (string) propagates into the emitted manifest + report alongside the canonical variant_id. Upstream agents (e.g. ai_scientist coloc_with_mr workflow) resolve and populate this automatically from OT / dbSNP; manual configs may set it for human readability. The orchestrator joins on variant_id (chr_pos_ref_alt), not rs_id; the rs_id is human-readability metadata only.

Bundled examples in examples/ (each runnable via --demo <NAME> once any required upstream data is in place):

• 01_synthetic_demo/ - offline 200-variant synthetic locus; no network required
• 02_eqtl_catalogue_x_gwas_catalog/ - SORT1 × cholesterol-VLDL canonical demo (Musunuru 2010)
• 03_open_targets_followup/ - entry from an OT coloc row
• 04_gwas_lookup_followup/ - entry from a gwas-lookup rsID hit
• 05_sqtl_sort1_liver_txrev/ - SORT1 sQTL in GTEx liver (txrev quant method)
• 06_sceqtl_sort1_onek1k_cd14_mono/ - SORT1 sceQTL in OneK1K CD14+ monocytes
• 07_pqtl_sort1_ukbppp_eur/ - SORT1 plasma sortilin via UKB-PPP (pQTL dispatch path)

The examples/chains/ and examples/recipes/ subdirectories are documentation patterns, not numbered demos: chains/finemapping_chain/ documents the fine-mapping skill -> locuscompare chain (requires fine-mapping to have run first; not in --list-demos); recipes/ ships harmonisation scripts that convert FinnGen / Pan-UKBB / UKB-PPP / GTEx v10 native formats into the canonical INPUT_SCHEMA TSV for use via sumstats_path:.

Example Output

The focal QTL gene (carried on the spec via exposure.gene_symbol, or auto-resolved from the OT exposure studyId in OT-followup configs) is rendered bold and tinted on the gene track so it anchors the visual in 10-30-gene windows. Other genes render in the default grey-blue.

A real run on the SORT1 × cholesterol-VLDL canonical demo (examples/02_eqtl_catalogue_x_gwas_catalog/config.yaml) produces:

<output_dir>/report.md:

markdown
# locuscompare-region-render report

- **Lead variant:** `1_109274968_G_T` (rs12740374; chr1:109274968, ±1000 kb)
- **Exposure:** SORT1 expression - minor salivary gland (eQTL Catalogue QTD000276)
- **Outcome:** cholesterol in medium VLDL (GWAS Catalog GCST90269602)
- **n_pairs joined:** 2547
- **n_palindromic_excluded:** 333
- **LD reference:** 1000G Phase 3 v5b GRCh38, EUR (n=503 samples)
- **Plot:** 1_109274968_G_T_full_locuscompare.png

## Notes
- fetched 1000G region VCF to ~/.clawbio/locuscompare_cache/1000g/chr1_108774968_109774968.vcf.gz
- effect-size slope = -0.177 (consistent with the inverse SORT1↑→LDL↓ direction reported by Musunuru 2010)

<output_dir>/manifest.yaml (excerpt):

yaml
skill: locuscompare-region-render
version: 0.1.0
lead_variant_id: 1_109274968_G_T
lead_rs_id: rs12740374
n_pairs: 2547
n_palindromic_excluded: 333
plot_path: 1_109274968_G_T_full_locuscompare.png
render_block:
  exposure_source: eqtl_catalogue
  exposure_source_release: v7+
  exposure_study_id: QTD000276
  outcome_source: gwas_catalog_harmonised
  outcome_source_release: '2026-05-06'
  outcome_study_id: GCST90269602
  ld_panel: 1000g_phase3_v5b_grch38_basic
  ld_panel_super_pop: EUR
  plink_version: PLINK v1.90b6.27 64-bit (2023-05-09)
  window_bp: 1000000
  ancestry_caveats: []
  fetched_at: '2026-05-09T11:42:06Z'

<output_dir>/figure.png: four panels stacked - outcome Manhattan (cholesterol-VLDL), exposure Manhattan (SORT1 eQTL), GENCODE gene track (CELSR2, PSRC1, SORT1, SARS1, SYPL2, ATXN7L2, ...), cross-trait -log10p scatter (tight diagonal at the top-right where the lead lives), effect-size scatter (negative slope). A representative rendering is bundled at examples/02_eqtl_catalogue_x_gwas_catalog/expected_render.png.

Output directory layout:

<output_dir>/
├── report.md                      # Markdown summary + caveats
├── figure.png                     # 4-panel LocusCompare PNG
├── manifest.yaml                  # Reproducibility provenance
├── tables/
│   ├── pairs.tsv                  # All joined pairs with both sides' beta/se/p + r2_lead
│   └── palindromic_excluded.tsv   # Variants dropped for palindromic ambiguity
└── reproducibility/
    ├── commands.sh                # Exact command to reproduce
    ├── environment.yml            # Pinned package versions
    ├── input_config.yaml          # Resolved config (defaults filled in)
    └── checksums.sha256           # Hashes of input + output artefacts

Gotchas

1. Palindromic variants (~10-20 % of common variants). A/T or G/C variants have ambiguous strand alignment. The skill excludes them from the LocusCompare scatter by default (count surfaced in manifest.yaml: n_palindromic_excluded). They still appear in the per-trait Manhattans because the position + p-value is informative even without strand. To override (only safe with allele-frequency cross-checks confirming strand alignment), add harmonisation: {exclude_palindromic: false} to the config.

1. Ancestry mismatch with 1000G EUR (the default LD reference). If your sumstats are from a non-EUR cohort (FinnGen Finnish-EUR, Pan-UKBB AFR/EAS/CSA, BBJ EAS), the LD reference picks up subtle r² differences that color points slightly off. For Finnish-EUR vs 1000G EUR the divergence is ~0.05 r² on common variants per Locke 2019; for AFR vs EUR it can be substantial. Set ld.super_pop: AFR | AMR | EAS | SAS to match. Add an explicit caveat to the rendered plot via caveats: [...] so consumers see the assumption.

1. All four inputs MUST be region-aligned to the same window. Mismatched windows produce a silent garbage plot. The skill validates that the exposure slice, outcome slice, LD panel, and gene track all overlap the requested (lead, window_bp) and refuses to render with mismatched coverage.

1. Lead variant must be present in BOTH the eQTL and GWAS slices. If the lead is missing from one side (commonly: low-MAF variant dropped by one harmoniser), the renderer falls back to a proxy variant in LD (r² > 0.8); the manifest records the substitution. If no proxy exists, the skill refuses to render and the orchestrator falls back to a credible-set-only view.

1. Effect-allele harmonisation across the four inputs is the renderer's input contract, not its job. The two sumstats slices and the LD panel must arrive in the canonical (chr, pos, ref, alt) GRCh38 ALT-effect form - this is what the bundled fetchers (eqtl-catalogue-region-fetch, gwas-catalog-region-fetch) emit. User-supplied TSVs not in this form should be normalised upstream (bcftools norm for indels). The renderer's harmonisation step is for cross-trait flip / palindromic handling, not for single-trait normalisation.

1. Render time is bounded by tabix fetches and plink LD compute. Cold-cache (first render of a region): typically 10-50 s. Warm-cache (region already fetched and LD computed): sub-second. Surface the timing to the user when relevant (multi-render comparisons, demo loops).

1. Visual diagonal-vs-two-cluster interpretation is informative but not definitive. A clean diagonal SUPPORTS H4 (single shared causal variant) but does not prove it; two distinct causal variants in tight LD can produce a visually diagonal pattern. The agent must always cite the formal coloc PP-H4 (and PP-H3) alongside the visual interpretation.

1. Wide windows (> 2 Mb) bloat memory and obscure structure. The default 1 Mb window is calibrated for typical cis-eQTL × GWAS coloc loci. Going wider rarely helps interpretation and slows the LD compute (plink scales with #variants²). For multi-locus views, render multiple plots and compose them.

1. p = 0 on rare extremely-significant variants. Some sources emit p = 0 when the actual value is below floating-point precision. The renderer substitutes the underflow floor (5e-324) before plotting on -log10. The reported -log10(p) for such variants is ~323, not their true magnitude.

Safety

Not for clinical decisions. This skill returns a visualisation of summary-statistic colocalisation; it is an interpretation aid, not a clinical or regulatory artifact. Do not use rendered plots for direct clinical decision-making.

Visual interpretation of LocusCompare patterns is informative but not definitive. Coloc PP-H4 from a formal Bayesian colocalisation analysis (COLOC, COLOC+SuSiE, SharePro, eCAVIAR) and supporting checks (n_colocalising_variants ≥ 5, credible-set width < 200 kb, ancestry match between exposure and outcome) should always be cited alongside the visual.

Local-first; no data upload. All computation runs locally. Input sumstats stay on the user's machine. Outbound network calls hit only public-domain databases (eQTL Cat FTP, GWAS Cat FTP, 1000G FTP, Ensembl REST for GENCODE). Cache stored at ~/.clawbio/locuscompare_cache/; the on-demand GENCODE gene-track cache reads LOCUSCOMPARE_CACHE_DIR at call time, so CI / sandboxed environments can redirect with LOCUSCOMPARE_CACHE_DIR=/path/to/cache per invocation.

Reproducibility. Every run emits a SHA-256 checksum manifest covering inputs and outputs. Failed LD compute or missing gene track is surfaced in the manifest's caveats list, not silently degraded.

Agent Boundary

The skill renders a 4-panel LocusCompare visualisation for a colocalisation result. The agent should:

• Use the rendered figure to support or refute the formal coloc PP-H4 result. Visual confirmation strengthens a high PP-H4 claim; a visual two-cluster pattern weakens it.
• NOT replace formal coloc analysis with the visual alone. PP-H4 and the supporting checks are the canonical decision criteria; the visual is an interpretation aid.
• Surface the H3 vs H4 ambiguity when the visual is two-cluster. Even with PP-H4 > 0.8, a two-cluster visual pattern indicates likely distinct causal variants in LD; flag for human review.
• NOT cherry-pick variants outside the rendered window for downstream interpretation. The visual establishes context for the rendered window only.
• Cite the rendered window, LD reference + super-pop, OT release (when applicable), exposure / outcome study ids, and lead variant in the user-facing reply. Per the user-friendly enum-expansion rule (CLAUDE.md): STRN × heart failure (FINNGEN_R12_I9_HEARTFAIL); window ±500 kb of lead chr2:36910110:C>T; LD = 1000G Phase 3 EUR; OT release 26.03.
• Surface the manifest's caveats list (palindromic exclusions, missing-lead proxy notes, ancestry mismatches) verbatim in the user-facing reply.
• NOT decide GO/NO-GO on a target based on the visual alone. Chain to target-validation-scorer for synthesis; this skill is one input among many.
• NOT silently swap super-populations. If the upstream cohort's ancestry does not match the requested LD super-pop, surface explicitly and ask the user to confirm before proceeding.

Citations

• Liu et al. (2019). Abundant associations with gene expression complicate GWAS follow-up. Nat Methods 16, 749-755. doi:10.1038/s41592-019-0431-x - LocusCompare convention.
• Hemani et al. (2018). The MR-Base platform supports systematic causal inference across the human phenome. eLife 7:e34408. doi:10.7554/eLife.34408 - palindromic-exclusion in MR/coloc.
• Kerimov et al. (2021). A compendium of uniformly processed human gene expression and splicing quantitative trait loci. Nat Genet 53, 1290-1299. doi:10.1038/s41588-021-00924-w - eQTL Catalogue.
• Sollis et al. (2023). The NHGRI-EBI GWAS Catalog: knowledgebase and deposition resource. Nucleic Acids Res 51, D977-D985. doi:10.1093/nar/gkac1010 - GWAS Catalog.
• 1000 Genomes Project Consortium (2015). A global reference for human genetic variation. Nature 526, 68-74. doi:10.1038/nature15393 - 1000G Phase 3 reference.
• Frankish et al. (2021). GENCODE 2021. Nucleic Acids Res 49, D916-D923. doi:10.1093/nar/gkaa1087 - GENCODE.
• Wang et al. (2020). A simple new approach to variable selection in regression, with application to genetic fine mapping. JRSS-B 82, 1273-1300. doi:10.1111/rssb.12388 - SuSiE (chained-skill: fine-mapping).