Statistical-genetics DAG in the scrna_qc shape: config-driven
thresholds, a Snakemake pipeline, a deterministic synthetic cohort plus
a real public-data mode, committed result artifacts, and the same
claim-verification layer that audits every number in the markdown
report against recorded results.
data -> qc -> strat -> assoc -> report
python -m pip install -e .[test]
snakemake -s workflow/Snakefile --cores 1 # demo cohort
PYTHONPATH=src python -m pytest tests/ -q
snakemake -n -s workflow/Snakefile # DAG dry rundataset.mode: demo generates 800 samples x 4,000 variants: two
ancestry groups differentiated by a Balding-Nichols model (Fst = 0.08),
2% missing calls, 8 planted causal variants at recorded effect sizes,
and a quantitative trait with an ancestry-shift confounder. The truth
lands in results/planted_truth.json, so recovery is auditable.
What the committed demo artifacts show:
- QC waterfall reconciles. 4,000 -> 3,840 variants (23 fail MAF, 137 fail pooled HWE) and 1 of 800 samples fails the heterozygosity screen. Two of the planted causals die at HWE: mixing ancestry groups produces the Wahlund effect, and the filter removes exactly the stratified variants it should. Het outliers are z-scored within ancestry group, so a real population's het level is not mistaken for contamination.
- PC1 recovers the planted structure at |r| = 1.00 with the ancestry label on 6.0% of genotype variance.
- The relatedness scan (GRM off-diagonal on PC-projected genotypes, so ancestry sharing does not count as relatedness) flags 0 pairs above the 0.125 third-degree mark; a planted-duplicates test shows the same scan catching true replicates.
- PC covariates control inflation, dose-resolved: lambda_GC 2.63 with no PCs, 0.97 already at 1 PC, 0.97 at 2 PCs. One component absorbs the entire two-group confound.
- Signals come back without false positives. All 4 Bonferroni hits are planted causals (4 of 6 post-QC recovered. Beta correlation 0.99). The two misses are sub-threshold effects, reported as such.
Demo numbers exercise the pipeline. They are not genetic findings.
config/demo_logistic.yaml reruns the same cohort with the liability
thresholded to a binary trait and the Rao score test: committed under
results/demo_logistic/, lambda 1.72 -> 0.99 with PCs, 3 Bonferroni
hits, all causal (a thresholded trait trades power for the case/control
shape).
config/kg_chr22.yaml reads a 14 MB range-fetch of the phase3 chr22
call set (the first 20,000 biallelic variants across 2,504 real
samples) plus the integrated-call panel for superpopulation labels and
a real sex covariate. Fetch commands are in the config header. Nothing
public is committed beyond the derived artifacts under
results/kg_chr22/.
The call set ships no phenotype, so the config simulates a trait on the
real genotypes with a recorded seed. Causal variants are planted only
on common variants with limited continental frequency spread
(max_anc_spread), keeping the trait unconfounded. Setting
ancestry_shift > 0 deliberately reintroduces stratification.
Committed artifacts show what real data does to a GWAS pipeline:
- 20,000 -> 1,866 variants: 86% fail MAF >= 0.01 (the call set is dense with rare variation) and 971 fail pooled HWE across continental groups. 4 of 2,504 samples fail the within-superpopulation het screen.
- PC1 alone carries 4.5% of ancestry variance across five superpopulations. Real structure needs multiple PCs.
- lambda_GC by covariate PCs: 11.0 / 5.1 / 7.6 / 2.0 / 2.5 / 1.9 for 0..5 PCs. Non-monotone because different PCs capture different structure than the planted trait loads on. It stays above 1 because a chr22-head slice is one LD-dense region and genome-wide lambda needs genome-wide sampling.
- The GRM relatedness scan flags 79,583 pairs at kinship > 0.125 with off-diagonal sd 0.058 (4x the demo's dispersion). 1,866 colocalized variants cannot estimate kinship reliably; the report prints the count with that caveat rather than letting it read as a cohort of cryptic relatives.
- 19 Bonferroni hits decompose as 2 recovered causals + 17 LD proxies within 250 kb + 0 unexplained. Two planted causals are perfectly collinear (identical statistics 619 bp apart). Their opposing true effects cancel marginally, which is the failure mode fine-mapping exists to solve.
src/statgen/claims.py is the same verifier scrna_qc uses (ported
from oncology_coscientist), vendored byte-identical and sha256-pinned
by the repo-root parity test. This package is its second consumer.
Every number in report.md must re-derive from
results/{qc_waterfall,strat_summary,assoc_summary,summary,planted_truth}.json
at display tolerance. Unbound tokens fail the check and land in
claims_check.json.
config/config.yaml: demo cohort, thresholds, association settingsconfig/demo_logistic.yaml: binary-trait score-test runconfig/kg_chr22.yaml: real-data mode with fetch commandssrc/statgen/data.py: cohort generation + VCF/panel loadingsrc/statgen/qc.py: missingness/MAF/HWE filters, reconciling waterfallsrc/statgen/strat.py: genotype PCA, PC1-ancestry diagnosticssrc/statgen/assoc.py: linear (FWL) / logistic (score) scans, lambda_GCsrc/statgen/report.py: figures, summary, markdown, claim checksrc/statgen/claims.py: vendored claim verifier (second consumer)workflow/Snakefile: the DAGresults/holds committed demo artifacts,results/demo_logistic/andresults/kg_chr22/hold the binary-trait and real-data runs
Single-marker marginal tests only, no LD-aware fine-mapping, imputation, or mixed models. The logistic path is a score test, not a fitted logistic regression. Simulated traits are pipeline exercises with recorded seeds and planted effects, never findings.