SKILL.md
🦖 Ancestry Decomposition PCA
Place your study cohort in global genetic context by computing a joint PCA against the Simons Genome Diversity Project (SGDP) — 345 samples from 164 populations spanning every inhabited continent.
What it does
- Takes your VCF + population map as input
- Finds common variants between your cohort and the SGDP reference panel (bundled)
- Runs PLINK PCA on the merged dataset
- Separates your cohort from SGDP reference samples
- Matches SGDP samples to their population labels (164 populations)
- Generates a publication-quality multi-panel figure:
- Panel A: PC1 vs PC2 — main population structure of your cohort
- Panel B: PC3 vs PC2 with regional groupings and confidence ellipses
- Panel C: PC3 vs PC1 with language/cultural groupings
- Panel D: Global context — your samples (circles) vs SGDP (triangles)
- Produces a markdown report with variance explained, population assignments, and reproducibility bundle
Why this exists
If you ask ChatGPT to "run a PCA against a global reference panel," it will:
- Not know which reference panel to use
- Hallucinate PLINK flags for merging datasets with different variant sets
- Skip IBD removal (related individuals distort PCA)
- Not normalise contig names between your VCF and the reference
- Produce a single scatter plot with no population labels
This skill encodes the correct methodological decisions:
- Uses SGDP (the gold-standard reference for global diversity)
- Handles contig normalisation (chr1 vs 1)
- Filters to common biallelic SNPs shared between datasets
- Removes related individuals via IBD checks
- Produces publication-quality multi-panel figures with confidence ellipses
- Differentiates your samples (circles) from reference (triangles)