Conservation Genomics Reanalysis

Assessing genetic status across the modern conservation-genomics toolkit — from the new global biodiversity-treaty (GBF) indicators on an endangered SE Asian bird, to realized-vs-masked genetic load in an inbred carnivore — and showing how fragile each verdict is to analyst choices.

Two chapters:

1. Green peafowl (Pavo muticus) — GBF indicators, contemporary N_e, robustness, population structure, and a damage-aware museum–modern comparison that remains geographically confounded. Dryad 10.5061/dryad.ffbg79ctc.
2. Scandinavian wolf (Canis lupus) — GERP-weighted load proxies (realized vs masked) and map-limited ROH autozygosity proxies. Dryad 10.5061/dryad.7m0cfxpzj.

Open the complete project website — the project summary, evidence gallery, searchable deliverable library, reproducibility record, and links to every public artifact. It is a dependency-free static site and works offline.

Sendable website bundle: dist/conservation-genomics-reanalysis.zip. Phase 5B is the current, cleanly verified release; its two isolated builds, bounded-memory evidence, science-hash comparison, and no-op proof are recorded in results/PHASE5B_REPRODUCIBILITY.md.

Personalize the website without editing HTML: double-click EDIT_WEBSITE.command, click an outlined heading or paragraph in the live preview, and choose Save website text. Keep the Terminal window open while editing and press Ctrl+C when finished. Undo, restore-original, and JSON import/export controls are included; scientific results, decision gates, links, figures, and generated Phase-5B evidence remain locked. See EDIT_WEBSITE.md.

Understand the project in 60 seconds

• Question: do conservation-genomics conclusions survive defensible analyst choices, or does the reading move when filtering, population structure, GERP site classes, or ROH definitions move?
• Green peafowl: no sampled population confidently clears the GBF $N_e>500$ target under any filtering preset. The headline is robust, but population estimates and even estimator convergence are fragile at n = 7/10/5. In Phase 5B, 0/15 simulated sample-size × marker-density designs resolved both predeclared resolution truths ($N_e=250$ and $N_e=1{,}000$), so the study does not support an empirical sample prescription; only 1/4 museum–modern diversity metrics survived every damage and influence check, with geography still confounded and no upgrade to temporal-erosion language.
• Scandinavian wolf: all 15/15 predeclared GERP × ROH cells support a positive partial realized-load association after the generic-homozygosity control. This is association between genomic proxies, not fitness, individual harm, or causal inbreeding depression. The authorized §13.4 localization found 24/24 predeclared population × GERP cells at the 1.5-Mb cutoff beyond the circular-shift null; that is genomic colocalization, not a phenotypic or causal result.
• Optional annotation study: §13.5 is DEFER_DUPLICATES_SOURCE_STUDY because 0/1 independent annotation definitions were available and the local classes overlap the source study for 211/211 samples.
• Contribution: a config-driven Snakemake reanalysis, fail-closed scientific decision gates, and a read-only explorer that makes uncertainty visible instead of forcing uncertain estimates into pass/fail.
• Boundary: this demonstrates genomic-proxy association, conservation relevance, and analytical sensitivity, not measured fitness, inbreeding depression, individual harm, or a new field estimate.

Start here: project website · four-page project-wide illustrated brief · research journey · questions and answers · 2–3 minute walkthrough · offline technical report

Deliverables

• Reproducible repo — one-command Phase-5B reproduction (snakemake --snakefile workflow/Snakefile phase5b --cores 4 --resources mem_mb=7000), pinned environment.lock.yml (R packages included — R is conda-owned), the Snakefile, a config of all thresholds, the results cube, all figures, and a paper-grade report/report.qmd.
• Fragility explorer (app/streamlit_app.py) — a live tool whose two knobs (filtering stringency × structure modeled/ignored) move the peafowl $N_e$ estimates and three-state per-population reading in real time; the pooled result is shown only as a cautionary structure contrast. The explorer also presents both GBF Goal-A indicators, a robustness matrix, and a GONE2 recent-trajectory context panel. (The N_e-method knob was dropped in Phase 2 — GONE2 measures a different time horizon, so it is context, not an axis.) Phase-5B panels add the peafowl damage falsification and scenario-based $N_e$ design landscape, plus the wolf individual-association and ROH-localization surfaces.
• Static project website (index.html) — an offline-readable portfolio with the scientific summary, visual evidence, searchable deliverable explanations, complete artifact index, and reproduction instructions. Its separate localhost editor changes only allowlisted narrative text and stores personalization outside the HTML. Validate both with python scripts/python/validate_website.py.
• Sendable website bundle (dist/conservation-genomics-reanalysis.zip) — a deterministic, self-contained ZIP of the site and every local dependency/deliverable. Rebuild and verify it with snakemake --snakefile workflow/Snakefile package_website --cores 4 --resources mem_mb=7000, so Snakemake passes the complete fail-closed artifact inventory to the packager.

Reproduction

# Build the pinned osx-64 toolchain. Classic conda OOM-kills on 8 GB RAM — use micromamba.
CONDA_SUBDIR=osx-64 micromamba create -n congen -f environment.lock.yml
micromamba activate congen
scripts/build_tools.sh                   # compile currentNe2 + GONE2 (pinned commits)
scripts/fetch_data.sh && scripts/extract_data.sh   # Dryad data (needs DRYAD_CLIENT_ID/SECRET)
snakemake --snakefile workflow/Snakefile phase5b --cores 4 --resources mem_mb=7000
streamlit run app/streamlit_app.py       # explore the verdict

# Optional explicit post-reproduction checks:
python -m unittest discover -s tests -v
report/render_report.sh

R packages (SNPRelate / dartR / hierfstat / adegenet) ship with the conda env — R is conda-owned (renv not used); for exact pins use environment.lock.yml.

For release maintainers creating a reproducibility record from an otherwise record-free checkout, phase5b_science is the non-release aggregate used in two isolated workspaces to collect real timing, return-code, RSS, validator, and science-hash evidence. After writing that evidence to results/PHASE5B_REPRODUCIBILITY.md, the canonical phase5b target above performs the synchronized rendering, record validation, offline packaging, and delivery gate. The first-release bridge is explicit: SCIENCE_PASS produces one release-content WARN until the canonical no-op is observed; promotion to PASS requires that distinct no-op proof and returns the final gate to zero WARN. Ordinary reproduction uses only the canonical phase5b command.

Layout

config/      ALL filtering thresholds, N_e settings, paths (config.yaml)
workflow/    Snakefile + rules/ (qc, structure, ne, gbf, load) — builds the results cube
scripts/     python/ + R/ analysis scripts called by the rules
results/      cube/ (results_cube.parquet) + tables/ + figures/
app/          streamlit_app.py — the fragility explorer (reads the cube)
report/       report.qmd — paper-grade methods/results/limitations
website/      static-site CSS, JavaScript, favicon, and preview notes (entry point: index.html)
dist/         sendable, validated project-website ZIP
data/         raw VCFs + intermediates (symlinked to internal SSD; see data/README.md)
tools/        currentNe2 / GONE2 (built from source) — see tools/README.md

See PROJECT_MAP.md for a one-page orientation (pipeline diagram, directory map, invariants), PROJECT_STATUS.md for the current phase table, CGplan.md for the full execution plan, and SETUP.md for environment setup status. Phase 5B status: COMPLETE. Sections 13.1, 13.2, 13.3, 13.4, and 13.6 are complete; §13.5 records the evidence-backed DEFER_DUPLICATES_SOURCE_STUDY decision, with reopening requiring at least one independent definition, at least 0.95 common-sample coverage, and at least 0.90 auditable callable-denominator coverage. The synchronized report, explorer, static site, manifest, offline ZIP, and release gates all pass; see the clean verification record.

Credits

• Data and original methods: green peafowl genomes from Dong et al. 2021 (Dryad 10.5061/dryad.ffbg79ctc); Scandinavian wolf genomes, VEP/GERP annotations, and the GERP $>4$ load anchor from Smeds & Ellegren 2022 (Dryad 10.5061/dryad.7m0cfxpzj). The wolf load calculation re-implements their gerp_load_polarized.py (linneas/wolf-deleterious). Full citations are in report/references.bib.
• Scientific software: currentNe2, GONE2, PLINK, bcftools, VCFtools, ADMIXTURE, SNPRelate, hierfstat, adegenet, dartR, Snakemake, Streamlit, Quarto, and pandoc.
• AI coding assistance: Claude (Anthropic), used through Claude Code for workflow implementation, debugging, code review, and prose drafting. The research question, design decisions, and final scientific claims are the author's; see the ownership and assistance record.