Run a GWAS on case-control genotype data

Hand Claude a PLINK/VCF genotype set and a phenotype column; get back a QC’d, PCA-adjusted genome-wide association scan with Manhattan-ready summary statistics and a documented QC trail.

   
Problem class Data analysis
Subject areas Translational Medicine, Molecular and Cellular Biology
Evidence level Proposed
Complexity One skill or MCP
Availability Fully open
Compute Workstation (laptop fine for <10k samples)

Problem

A statistical geneticist who has genotyped a case-control cohort (custom array, imputed dosages, or sequencing-derived calls) faces the same checklist every time: filter low-call-rate samples and variants, drop SNPs failing Hardy-Weinberg equilibrium in controls, prune for linkage disequilibrium, derive principal components to correct for population stratification, then run the per-variant logistic regression with those PCs as covariates. The steps are individually standard and individually easy to get subtly wrong — the wrong HWE threshold, forgetting to compute PCs on an LD-pruned set, mixing case/control HWE filtering. “Solved” looks like: point Claude at the genotype files, get back a reproducible QC log, a covariate file of genotype PCs, and an association table ready to plot, with every threshold stated.

  1. Install the PLINK2 skill from the SciAgent-Skills collection (clone the repo, /plugin install sciagent-skills). The skill reads PLINK .bed/.bim/.fam, VCF, and BGEN, and exposes QC, IBD, PCA, and regression GWAS through local Python/Bash.

  2. Prompt with the cohort files, phenotype, and QC thresholds. A worked version:

    Run a case-control GWAS on the PLINK fileset cohort.{bed,bim,fam}
using the PLINK2 skill. Phenotype is in the .fam (1=control, 2=case).
Assembly is GRCh38.

Stage 1 — sample QC:
  - drop samples with call rate < 0.98 (--mind 0.02)
  - report and flag sex-check mismatches (--check-sex); do not auto-remove
  - flag heterozygosity outliers (>3 SD from the mean F)
Stage 2 — variant QC:
  - --geno 0.02, --maf 0.01
  - --hwe 1e-6 applied in CONTROLS only
Stage 3 — stratification:
  - LD-prune (--indep-pairwise 50 5 0.2) to an independent SNP set
  - compute 10 principal components on the pruned set (--pca 10)
Stage 4 — association:
  - logistic regression (--glm) on the QC'd genotypes,
    covariates = first 10 PCs
  - output Manhattan-ready summary stats (CHR, POS, SNP, A1,
    OR, SE, P)

Emit a QC log table (samples in/out, variants in/out per step) and
the genomic inflation factor lambda_GC. Save the association table
as gwas_results.tsv.

  3. Read off lambda_GC before trusting any hit. A genomic inflation factor far above ~1.05 signals residual stratification or cryptic relatedness — re-prompt to add more PCs, tighten relatedness filtering (--king-cutoff 0.0884 to remove up to second-degree relatives), or note that a mixed model is needed (see Alternatives).

  4. Plot and annotate the top loci. Ask Claude to emit a Manhattan + QQ plot from gwas_results.tsv, then hand genome-wide-significant SNPs (P < 5e-8) to the GWAS Catalog skill to check whether each locus is already a known association for the trait.

Why this assembly

Rung 2. A single skill (PLINK2) covers the entire QC → PCA → association arc; PLINK is the reference implementation these steps were defined against. Claude’s value is orchestrating the multi-stage protocol correctly and keeping an auditable QC log, not replacing the engine. Claude Code alone (rung 1) cannot run PLINK and would hand-wave the genotype math. A rung-3 toolbelt adds nothing here — the GWAS Catalog lookup in step 4 is annotation, not analysis, and is optional. Escalate to rung 3/4 only for biobank-scale cohorts where a linear mixed model (regenie, SAIGE) is required to control inflation — and those tools are not yet catalogued (see Alternatives).

Availability

Fully open. PLINK2 is GPL-3.0; the skill ships in the CC-BY-4.0 SciAgent-Skills collection. The GWAS Catalog REST API used for annotation is public and needs no auth. Genotype data itself is yours — no external upload; the skill runs locally, which matters for consented human-subjects data under your IRB/dbGaP data-use agreement.

Compute requirements

Laptop-sufficient for cohorts up to ~10k samples × ~1M variants: PLINK2’s bit-level parallelism runs QC and a --glm scan in minutes with a few GB of RAM. For larger imputed sets (10M+ variants, tens of thousands of samples) move to a workstation with 32–64 GB RAM; the --pca and --glm steps are the heavy ones. No GPU is used by PLINK itself — the front-matter tier reflects the upper end of in-scope cohorts. Biobank scale (100k+ samples) exceeds this recipe; see Alternatives.

Evidence

Proposed. No published benchmark of an LLM-driven PLINK2 GWAS assembly is known. The closest grounding is the protocol literature this recipe encodes: PLINK’s second-generation engine and its --glm/--pca/QC operators are documented in Chang et al., GigaScience 4:7 (2015), and the exact QC-then-association staging (call rate, HWE-in-controls, LD pruning, PCA covariates, logistic-regression GWAS, lambda_GC check) follows the widely used tutorial of Marees et al., Int. J. Methods Psychiatr. Res. 27:e1608 (2018). Both are standard references with thousands of citations; the assembly’s individual steps are validated, the LLM-orchestrated composition is not independently benchmarked.

Alternatives considered

  • Biobank-scale cohorts (rung 3/4). With 100k+ samples or substantial relatedness/structure, fixed-effect logistic regression inflates; a linear mixed model (regenie, SAIGE, BOLT-LMM) is the standard. None is catalogued as a Claude component today, so this recipe stays at the array/small-cohort scale. Surfaced as a missing-component note for the catalog curator.
  • Claude Code alone (rung 1). Insufficient — no live genotype-processing engine; the model would confabulate per-variant statistics.
  • The GWAS Catalog skill alone. That queries published SNP-trait associations; it interprets results, it does not compute them. Use it for step 4 annotation or when you only want prior-art lookups, not a fresh scan.

See also

Sources

Tried this recipe?

Share feedback — what worked, what didn’t, what you’d change. The form opens with this recipe pre-selected and a link back to this page.