It holds three kinds of thing: source tables the pipeline ingests, trained models, and the published per-cell surfaces those models produce.

This is interim storage. The project's artifact home is GCS (#33); this dataset exists so the store lives in one addressable place until that is set up. Everything here is plain files, so the migration is a copy.

raw/
  afnd_populations.tsv        1,821 AFND populations: coordinates, settlement class, ascertainment
  afnd_frequencies.tsv        123,502 HLA/KIR/MIC/cytokine allele frequencies
  map_hbs_surveys.csv         1,287 Malaria Atlas Project HbS surveys
  map_g6pd_surveys.csv        1,749 MAP G6PD deficiency surveys
store/
  artifacts/<variant>__<model_version>__<data_version>/
    cells.parquet             per-H3-cell posterior summaries  <- the citable output
    manifest.json             what was published, under which assumptions
  fits/<variant>.fit.pkl      trained PyMC models  <- a CACHE, not an artifact (see below)
  INVENTORY.json              sha256 of every file

How it was made

raw/afnd_populations.tsv — harvested from allelefrequencies.net by scripts/fetch_afnd.py, in two public hops: pop6001b.asp lists every population as a ?pop_name= link in one request, and pop6001c.asp?pop_name=<name> prints the coordinate. 1,821 of 1,825 retained; the four refusals print no coordinate at all. Coordinates are kept in AFND's printed sexagesimal (6º 25' S) because the printed precision is what bounds them — 41º 0' N is degree precision and earns a ~78 km uncertainty floor, not an arcminute fix.

raw/afnd_frequencies.tsv — not ours. Redistributed by slowkow/allelefrequencies (MIT-licensed code), scraped from AFND's public frequency search endpoints.

raw/map_*.csv — Malaria Atlas Project WFS layers Explorer:HbS_Data and Explorer:G6PD_Data, by scripts/fetch_map_hbs.py. Open, no credentials.

store/fits/ — scripts/build_surfaces.py. A binomial/beta-binomial likelihood over a Gaussian process on the unit sphere (not lon/lat — a degree of longitude is 111 km at the equator and 47 km at 65°N), with inducing points placed on an H3 geodesic grid, sampled with numpyro NUTS. A fit that has not mixed is refused rather than published.

store/artifacts/ — scripts/publish_artifacts.py, predicting each fit onto H3 res-3 land cells.

Using it

from huggingface_hub import snapshot_download
import pandas as pd

path = snapshot_download("bschilder/genomeos-data", repo_type="dataset")

The surfaces (start here)

cells = pd.read_parquet(f"{path}/store/artifacts/chr11-5227002-T-A__v1__map-2026-08/cells.parquet")
print(cells.columns.tolist())
# ['h3_index', 'variant_id', 'post_median', 'post_mean', 'post_sd',
#  'q025', 'q975', 'q25', 'q75', 'support', 'posterior_contraction',
#  'dist_nearest_obs_km', 'model_version', 'data_version']

Read the support column before the numbers. It is the point of the whole design:

`support`	meaning
`observed`	a survey sits in this cell
`interpolated`	inferred, with data within twice the correlation range
`unknown`	no data close enough — the model is not making a claim here
`prior_dominated`	the posterior never moved off the prior

# Never aggregate without masking. A mean over unmasked cells is a mean over the prior.
claimed = cells[~cells["support"].isin(["unknown", "prior_dominated"])]
print(f"{len(claimed)}/{len(cells)} cells carry a claim")
print(claimed["post_median"].describe())

dist_nearest_obs_km travels with each row so you can apply a stricter threshold than ours without refitting.

The manifest

import json
m = json.load(open(f"{path}/store/artifacts/chr11-5227002-T-A__v1__map-2026-08/manifest.json"))
# correlation_range_km, likelihood, lengthscale_sigma, n_observations, support_counts, ...

A fitted correlation range is meaningless without the prior that produced it, which is why the manifest carries both. Artifacts are keyed (variant_id, model_version, data_version) and are immutable: a refit publishes alongside rather than overwriting, so an older citation keeps resolving.

The observations

pops = pd.read_csv(f"{path}/raw/afnd_populations.tsv", sep="\t")
freq = pd.read_csv(f"{path}/raw/afnd_frequencies.tsv", sep="\t")
# Join on population name — AFND's own public key, shared by both tables, so it is exact.
joined = freq.merge(pops, left_on="population", right_on="population")

Watch for two things that cost real data when missed: n carries thousand separators ("3,732"), and alleles_over_2n is a frequency, not a count — allele counts are reconstructed as round(af * 2n).

The fits — read this first

from genomeos.surfaces.fit import load_fit          # genomeOS must be installed
fit = load_fit(f"{path}/store/fits/chr11-5227002-T-A.fit.pkl")

A fit is a cache, not an artifact. It is a pickled PyMC graph, so it is coupled to the exact PyMC/pytensor versions that wrote it, and pickle executes arbitrary code on load — only load files you trust. It is here because refitting costs ~20 minutes, not because it is archival. The parquet is the durable form: 1.1 MB against a 121 MB fit, carrying everything a consumer needs.

Reproduce the environment exactly with requirements.lock.

Provenance and terms

This dataset is private, and that is deliberate. AFND publishes no licence — its "Licensing" link carries only a disclaimer, and re3data's "public domain" record is third-party catalogue metadata rather than a grant. Collection here proceeds on an assumed-open basis (#117); keeping the dataset private makes this storage rather than redistribution, which is what that decision covers. Redistribution of anything derived from indigenous-population panels is separately unsettled (#66).

Cite the sources, not this store:

AFND — Gonzalez-Galarza et al., Allele frequency net database (AFND) 2020 update, Nucleic Acids Research 48:D783. doi:10.1093/nar/gkz1029
MAP HbS — Piel et al., Global epidemiology of sickle haemoglobin in neonates, The Lancet 381:142 (2013).
MAP G6PD — Howes et al., G6PD deficiency prevalence and estimates of affected populations in malaria endemic countries, PLoS Medicine 9:e1001339 (2012).
Frequency redistribution — slowkow/allelefrequencies.

Known limitations

No golden test has passed. §8's HbS parity test is blocked on population-weighted national aggregation.
Held-out skill is marginal. Under spatially-blocked cross-validation the HbS surface scores MAE 0.0388 against a constant baseline's 0.0391 (#109).
HLA surfaces are not published here. A screen of twenty alleles fitted correlation ranges of 1,036–4,111 km — a range that long makes the field near-constant and indistinguishable from the intercept (#122).
One population can dominate. AFND's DKMS German donor entry has an = 6,912,132, about 23,000× the median; a binomial likelihood weights by an (#123).
G6PD is a phenotype, not a variant. phenotype:g6pd-deficiency aggregates ~200 alleles assayed by enzyme activity (#116).

Downloads last month: 29

Total file size:

901 MB

You need to agree to share your contact information to access this dataset

genomeOS Atlas data store

Contents