A collection of Claude Code / agent skills for statistical-genetics workflows, built to the Agent Skills standard. Each skill is a self-contained directory with a SKILL.md + helper scripts.
| Skill | Purpose |
|---|---|
gwas-fetch/ |
Fetch harmonised GWAS Catalog sumstats by accession (GCST*). OKG-first metadata resolution with REST-API fallback. Caches downloads + writes a .fetch.json provenance manifest. Compose with liftover for build alignment. |
liftover/ |
KG-aware genomic liftover for GWAS sumstats. Resolves source/target build via the OKG when available; falls back to the GWAS Catalog REST API. Refuses unknown-provenance lifts and emits a coverage-gap proposal stub. Supports --gcst <accession> for direct GWAS-Catalog fetch+lift. |
okg-search/ |
Query the statgen-analysis OKG for datasets matching a trait, EFO term, seed-dataset, or curated category. Filters: controlled-vocab ancestry (EUR/AFR/EAS/AMR/SAS/admixed), min-N power threshold, sumstats availability. EFO sibling expansion bridges the curated trait:* and scale-ingest trait:gwas_catalog_scale:* namespaces. TSV output is ready to feed into gwas-fetch GCST<X> for batch chaining (e.g. rg panels). Connects via direct Postgres (OKG_DSN); auto-installs psycopg[binary] on first use. |
manhattan-qq/ |
Quick QC for a GWAS sumstats file: Manhattan + QQ plots + a summary JSON with genomic-inflation λ_GC, GWS + suggestive hit counts, and a list of independent lead SNPs (greedy ±window-kb pruning). Auto-detects common column-name conventions (harmonised hm_*, LDSC munged, bare CHR/BP/P/SNP). Subsamples low-significance variants for fast plotting on full HM3- or 1KG-scale sumstats. Writes a .manhattan_qq.json sidecar citing the input dataset's OKG node. |
ldsc/ |
Run LD Score Regression: munge raw sumstats, estimate SNP heritability (h2), compute genetic correlation (rg), and stratified / cell-type-specific heritability (partitioned, wrapping LDSC --h2-cts for the Finucane 2018 multi-tissue / GenoSkyline-Plus / any .ldcts panel). Auto-installs the CBIIT LDSC fork; auto-detects --signed-sumstats from the file header (hm_beta → 0, hm_odds_ratio → 1, etc.); auto-supplies --merge-alleles from the on-disk HM3 snplist so indels don't crash downstream rg; writes a .ldsc.json sidecar with OKG provenance and parsed key results. |
finemap/ |
Fine-mapping with multiple interchangeable backends: sushie (mancusolab/sushie, single- and multi-ancestry SuSiE / SuShiE) as the default; SuSiE 2.0 (StatFunGen/susieR, Wang 2025) as the default SuSiE-inf implementation; and FinucaneLab/fine-mapping-inf (Cui 2024) as an opt-in (--susie-inf-impl finucane). Five subcommands: susie / sushie (individual genotypes), region (GWAS sumstats + reference VCF/PLINK; auto-detects rsid-vs-chrpos SNP-ID format of the panel), sumstats (precomputed Z + LD with method dispatch), and recommend (OKG-driven method choice given a dataset_metadata node). Each backend auto-installs on first use; writes a .finemap.json sidecar citing the method / software / paper / ld_panel / dataset OKG nodes for whichever backend ran. |
prs/ |
Polygenic risk scoring via a pluggable --method flag. Currently implements SBayesRC (Zheng 2024) through the zhilizheng/SBayesRC R package. OKG-resolves the matching HapMap3 LD eigendecomposition by --ancestry eur|eas|afr (or explicit --okg-ld-panel-id), auto-pulls N from --okg-dataset-id, converts harmonised / LDSC / GWAS-SSF sumstats to COJO format, backfills missing effect-allele frequencies from the LD reference's snp.info (GWAS-Catalog harmonised files often ship all-NA EAF; without backfill SBayesRC's tidy() drops every row), runs SBayesRC's 3-step tidy → impute → sbayesrc pipeline, and writes per-SNP PRS weights + .prs.json sidecar citing 5 OKG nodes (method, software, paper, ld_panel, dataset). LDpred2 / PRS-CS easily wired in behind the same flag. |
twas/ |
Transcriptome-wide association study via TWAS-FUSION (Gusev et al. 2016, Nat Genet) against the GTEx v8 multi-tissue weights. Resolves the (--tissue, --ancestry) pair to an OKG dataset:fusion_gtex_v8_<eur|all>:<tissue> node, downloads + caches the panel + the FUSION 1000G EUR LDREF on first use, runs FUSION.assoc_test.R per chromosome, concatenates per-gene results, and writes a .twas.json sidecar citing the method / software / paper / ld_panel / tissue / cohort / dataset OKG nodes. Auto-installs FUSION's R deps (plink2R, glmnet, here, coloc) on first use. With --okg-trait-id, queries the OKG's tissue → trait relevant_to mapping and iterates over every tissue relevant to that trait. |
locuszoom/ |
LocusZoom-style regional association plots via the locuszoomr R package. Selects a window by gene symbol, region (chr:start-end), or lead-SNP rsid; fetches pairwise r² from the LDlink REST API (1000G, configurable population); renders a multi-track PDF + PNG with the association scatter on top and Ensembl gene tracks below. With --finemap-pip or --finemap-sidecar, adds a middle panel of per-variant PIPs coloured by credible-set membership — the canonical 3-panel fine-mapped-locus figure. Requires a free LDlink API token — the skill prompts on first use and caches it under ~/.cache/locuszoom/ldlink_token. Writes a .locuszoom.json sidecar with locus selection, LD coverage stats, and OKG provenance (propagated from the upstream .finemap.json when supplied). |
variant-annotate/ |
Per-variant functional + tissue annotation against gnomAD (allele frequencies, predicted consequence, gene context) and GTEx (top eQTL / sQTL hits). Used by finemap region to enrich the printed credible-set summary table with rsid → consequence / gene / AF / eQTL annotations. Caches REST responses under ~/.cache/variant-annotate/. |
The skills follow the Agent Skills specification, so they should work with any compatible agent (Claude Code, Cursor, Goose, OpenCode, etc.).
Symlink each skill into ~/.claude/skills/:
git clone https://github.com/svdorn/statgen-skills.git
cd statgen-skills
./install.sh # symlinks all skills into ~/.claude/skills/Or symlink a single skill:
ln -s ~/statgen-skills/liftover ~/.claude/skills/liftoverAfter symlinking, restart Claude Code (or the agent client) so it picks up the new skills at startup.
Point your client at statgen-skills/<skill> or copy individual skills into the client's skills directory. Each skill's SKILL.md documents its compatibility field for any environment requirements.
Each skill is structured per the agentskills.io spec:
<skill-name>/
├── SKILL.md # required: YAML frontmatter (name, description, ...) + body
├── scripts/ # executable helpers (Python, bash, etc.)
├── references/ # detailed docs loaded on demand
└── assets/ # templates, schemas, etc.
Validate a skill before committing:
# optional: uses agentskills' reference validator
skills-ref validate ./liftoverSeveral skills here are designed around an OKG (Operational Knowledge Graph) — the upstream reference is github.com/mitdbg/okg, but any fork or compatible deployment works. To use OKG-aware features, set the OKG_REPO environment variable to your local clone:
git clone https://github.com/mitdbg/okg ~/Lab/KG/okg # or your fork
export OKG_REPO=~/Lab/KG/okgThe skills then spawn the OKG MCP server at $OKG_REPO/deployments/<deployment>/server.py (default deployment: statgen-analysis) and query for provenance metadata before acting. When the OKG can't resolve a needed entity, skills emit a coverage-gap proposal stub instead of guessing — see each skill's references/COVERAGE_GAPS.md for the workflow.
If $OKG_REPO is not set and a skill is invoked with an --okg-* flag, the skill refuses with a clear error pointing you here. Skills can still run without the OKG by passing all required metadata explicitly (e.g. --source hg38 --target hg19 for liftover).
MIT — see LICENSE.