Share your Hugging Face username and profile email with Standard Model Bio for automatic access. We use this information to understand who accesses the data.

hg-horizon: genome-wide LD computed with CuEvo

LD pair scores are calculated offline with CuEvo. Spatial summaries provide both mean and max pooling. Inputs are the 1000 Genomes NYGC high-coverage GRCh38 panel's 2,504 unrelated phase-3 individuals. AFR, AMR, EAS, EUR and SAS are calculated separately. This is unphased diploid dosage Pearson correlation, not phased haplotype-frequency LD, ancestry-adjusted correlation, MI, or a statistical significance catalogue.

CuEvo revision: bde2ef7eeb452dbf51fe3331e499828c3952222e. The native engine uses integer genotype counts and the exact rational test 10 C² ≥ Vx Vy for r² ≥ 0.1. Scores are delivered as float64 signed r, without int16/float16 quantization. The export orients r to each population's minor alleles; explicit allele dictionaries permit other orientations. Numerical validation is recorded in verification.json. Squaring stored r can differ from the exact cutoff by floating-point roundoff; eligibility was decided by the integer test, rather than by re-filtering rounded scores.

Coverage, resolution, and pooling

Product	Coverage	Delivered detail
`pairs/{population}/chr*.parquet`	Cis, chr1–22 and X	Retained individual variant pairs, zero-based chromosome positions
`tiles/{population}/chr*_{resolution}.parquet`	Same cis pair population	10 kb, 50 kb, 100 kb, 500 kb, 1 Mb spatial bins; mean and max
`full-genome/cis-trans/v2/{population}/pairs/chr*.parquet`	Every retained cis and trans pair among chr1–22	Individual pairs, indexed by population-specific variant row IDs
`full-genome/cis-trans/v2/{population}/variants/chr*.parquet`	Matching autosomal variant dictionaries	Row IDs, chromosome, one-based position, alleles and frequencies
`full-genome/cis-trans/v2/{population}/tiles/genome_*.parquet`	Same full autosomal cis/trans population	250 kb, 1 Mb, 5 Mb bins; mean and max on a declared padded genome axis
`inputs/plink/`	Reusable autosomal genotype conversion	2,504 people, 20,276,768 biallelic A/C/G/T SNVs; no phase
`inputs/vcf/{population}/chrX.vcf.gz`	Exact chrX scoring input per population	Filtered GT calls with the declared 0/2 male coding

Smaller bp bins mean finer resolution. Tiles are spatially pooled summaries, not single-variant pair tables. Their mean r² = sum_r2/n and max r² = max_r2 describe retained r² ≥ 0.1 pairs only, not every possible variant pair. Missing pairs are not measured zeros. Counts and sums allow correctly weighted re-pooling; resolution levels repeat the same observations and must not be added as independent measurements. max_r preserves the sign of the strongest-|r| source pair; it is not a maximum of signed r.

Each scan uses population MAF ≥1%, inclusive. Eligible autosomal variants are drawn from the reusable panel selected at MAF ≥1% in at least one superpopulation. CuEvo tests the complete unordered upper triangle with no genomic distance or top-K limit. The scan's internal memory windows do not restrict genomic distance. chrX is a separate cis product: X–autosome trans, Y, and MT are not supplied.

Population	People	Eligible autosomal variants	Retained autosomal cis pairs	Retained autosomal trans pairs	Retained chrX cis pairs
AFR	661	15,391,796	1,503,015,246	74,548,741	83,649,831
AMR	347	10,154,246	2,558,209,078	4,472,972,035	351,576,562
EAS	504	7,880,833	1,087,705,883	40,187,517	37,135,228
EUR	503	8,890,126	1,293,544,087	47,955,947	45,657,547
SAS	489	9,353,742	1,243,359,677	97,671,664	47,999,261

Exact compressed download sizes

Product	Data files	Bytes	Decimal GB
`inputs/plink/`	66	13,368,980,720	13.369
`inputs/vcf/`	5	646,625,652	0.647
`pairs/`	115	43,814,427,403	43.814
`tiles/`	575	6,618,360,308	6.618
`full-genome/cis-trans/v2/`	235	70,357,289,081	70.357
All data	996	134,805,683,164	134.806

Small documentation/index files are additional. Use files.tsv or manifest.json to plan a subset download; avoid downloading the entire repository by default. Fine cis rows also occur in the full cis/trans product.

Automatic email access and copy/paste downloads

Open the dataset page, sign in and agree to share your profile email and username. Access is approved automatically. Use a token from that approved account for downloads; no private decoder or application code is required. Each hg-horizon dataset has a separate access request.

python -m pip install huggingface_hub pyarrow numpy
hf auth login

import json
import pyarrow.parquet as pq
from huggingface_hub import HfApi, hf_hub_download

repo = "standardmodelbio/hg-horizon-ld"
revision = HfApi().dataset_info(repo, token=True).sha
print("Pin this revision:", revision)
def download(name):
    return hf_hub_download(repo, name, repo_type="dataset", revision=revision, token=True)

manifest = json.load(open(download("manifest.json")))
print("Data bytes:", manifest["data_bytes"])

# A small genome-wide overview, with both pooling strategies in the same file.
path = download("full-genome/cis-trans/v2/EUR/tiles/genome_5000000.parquet")
tile = pq.read_table(path, columns=["pos_i", "pos_j", "n", "mean_r2", "max_r2"])
print(tile.slice(0, 5).to_pylist())

# Finer cis data; batch iteration does not materialize the whole pair table.
fine = pq.ParquetFile(download("pairs/EUR/chr22.parquet"))
batch = next(fine.iter_batches(batch_size=5))
print(batch.to_pylist())

Full cis/trans files contain gidx_i, gidx_j, padded genome_pos_i/j, r_signed and joint_calls. Join gidx to that same population's variant dictionary for chromosome, one-based pos_1based, A1/A2 alleles and frequencies. gidx is a row ID, not a base-pair coordinate. chrX is a separate index namespace: its fine-pair gidx values refer to zero-based record order in inputs/vcf/{population}/chrX.vcf.gz and must not be joined to the autosomal dictionaries. See the format guide.

from huggingface_hub import snapshot_download
# Consult the manifest sizes before choosing a population-wide download.
snapshot_download(repo, repo_type="dataset", revision=revision, token=True,
                  allow_patterns=["pairs/EUR/*", "tiles/EUR/*"], local_dir="hg-horizon-ld")
# Or retrieve the useful genotype conversion for a single chromosome.
snapshot_download(repo, repo_type="dataset", revision=revision, token=True,
                  allow_patterns=["inputs/plink/chr22.*"], local_dir="hg-horizon-ld")

Inputs, provenance and limitations

The original VCFs and sample metadata are linked to their providers rather than duplicated. The source release is NYGC 20220422_3202_phased_SNV_INDEL_SV, with chrX .v2. The autosomal PLINK inputs provide filtered, reusable genotype calls. LD is calculated from those genotype calls. BED calls are unphased and do not reconstruct the original haplotypes. Sample rosters, input hashes, eligibility rules, software pin, score conventions and numerical checks are recorded in PROVENANCE.json.

The recomputation does not impute missing calls or silently replace them with zeros. The native CUDA path is admitted only after verifying complete calls in the scoring inputs. For these inputs, joint_calls is the full population sample size. Undefined invariant pairs are excluded. No p-value, biological validation, or genome-wide FDR guarantee is implied by the r² cutoff. When A1 frequency equals 0.5, A1 defines the sign orientation.

chrX male haploid calls are duplicated into the existing 0/2 dosage convention; female calls remain diploid. This mixed-ploidy statistic can reflect sex composition. It must not be interpreted as sex-specific or phased haplotype LD.

The padded autosomal axis (hg38-autosomes-padded-5mb-v1) pads each chromosome length to a multiple of 5 Mb, span 2,935,000,000 bp. Padding is not measured sequence. Exact offsets are supplied in genome-axis.json. Chromosome-local fine/tile positions are zero-based; BIM and dictionary pos_1based positions are one-based.

Derived results are CC BY 4.0; genotype inputs remain subject to the original 1000 Genomes/IGSR terms. Cite Byrska-Bishop et al., Cell 2022 and the original resource. Curated by Standard Model Bio, in the hg-horizon collection.

Downloads last month: 475

Total file size:

135 GB

Collection including standardmodelbio/hg-horizon-ld

hg-horizon

Collection

Curated maps of linkage disequilibrium, coevolution, chromatin contacts, and GWAS correlations for exploring the genome at every scale. • 5 items • Updated 1 day ago