standardmodelbio/hg-horizon-coevolution-erc

hg-horizon: genome-wide 1 kb-window ERC

Evolutionary-rate correlation (ERC) between physical hg38 genome windows, computed offline from the GPN processed 100-vertebrate alignment with CuEvo. This dataset contains ERC only.

What is measured — read before downloading

• 1 kb source elements, not individual nucleotides: 2,875,012 nonoverlapping physical windows cover chr1–22, including terminal partial windows. The stored catalog has 5,902,620 pairs.
• The scan considered all 4,132,845,562,566 unordered window pairs, including distant cis and trans candidates. It stores only signed ERC r > 0.25, with no top-k selection.
• Each window needs effective-taxon coverage at least 80/100; the pair's joint coverage proxy must also be at least 80/100. At least 20 shared observed tree branches and nonzero variance are required. Only 7,516 windows pass the individual coverage gate (0.261% of physical windows). Broad scan scope does not mean dense genome coverage.
• Seven coarser spatial summaries: 10 kb, 50 kb, 100 kb, 250 kb, 500 kb, 1 Mb, 5 Mb. Every cell supplies both mean_r and max_r, with n_pairs and sum_r. The denominator is retained 1 kb pairs only. These are not recomputed native coarse-window ERC.
• Missing rows can reflect failed coverage, constant profiles, or r <= 0.25. They are not zeros. Negative, small positive, and unscorable pairs are absent. This is not a complete signed matrix.
• The cutoff is not a significance test. No p-values, q-values, genome-wide FDR control, biological validation, or independence of tree branches is asserted.
• Converted inputs retain all 1 kb window profiles, including windows that fail the score gates. Other native input tiers and exhaustive single-position pair scores are not included.

Machine-readable equivalents: manifest.json, PROVENANCE.json, schemas.json, and AGENTS.md.

Contents and exact sizes

Product	Data files	Bytes	Decimal size
pairs/	253	136,985,685	0.137 GB
pooled/	7	366,819,952	0.367 GB
inputs/	24	2,259,850,359	2.260 GB
All data	284	2,763,655,996	2.764 GB

pairs/chrN__chrM.parquet partitions retained window pairs by chromosome pair and sorts rows by the two genome coordinates. Only nonempty partitions are present; the manifest records them. Files contain float64 scores and support values, without score quantization.

pooled/genome_{resolution}.parquet supplies spatial mean and max over the same retained population. Each unordered source pair contributes once, including pairs whose windows fall inside the same coarse bin. Every resolution conserves all 5,902,620 retained source pairs.

inputs/profiles_1000/chrN.npz contains losslessly ZIP-compressed converted profiles for every physical 1 kb window on that chromosome. The arrays are posterior substitution counts posterior_counts[windows,198], binary observed_branches[windows,198], taxon_coverage[windows,100], usable_bases, genome_start, genome_end, and profile_index. NPZ is not memory-mappable; loading a compressed array decompresses that chromosome's array. inputs/expectation_1000.npy is the frozen 198-branch background expectation. inputs/profile-axes.json supplies ordered species and branch-node IDs. inputs/hg38.100way.nwk supplies the original checksum-pinned species tree.

Coordinates and support units

All coordinates are zero-based, half-open. Pair start_i/end_i/start_j/end_j are ordinary chromosome-local hg38 coordinates; chromosome labels are in the partition name and manifest. genome_start_i/genome_end_i/genome_start_j/genome_end_j and pooled/input coordinates use hg38-autosomes-padded-5mb-v1: chromosomes 1–22 each occupy ceil(length/5 Mb)*5 Mb. The total axis span is 2,935,000,000 bp; padding is not measured genome sequence. genome-axis.json gives exact offsets and biological lengths.

profile_index_i/j refer to the original global 1 kb profile rows; profile_index in each NPZ preserves those IDs. n_shared_branches counts tree branches used for Pearson correlation. effective_taxa_i/j are sums of per-taxon valid A/C/G/T fractions across each physical window; effective_taxa_joint = sum_s f_i,s * f_j,s is a coverage proxy, not a count of matched bases or independent statistical samples. A value of 80 means 80 effective taxa out of the 100-species panel. It is not a Fisher-z degrees-of-freedom estimate.

Provenance and calculation

The upstream alignment is songlab/multiz100way, file 99.zarr.zip, revision a790ef9ea692cd2bc9703f3b918337a6de08db9b: 42,017,663,762 bytes, SHA256 229e4443dfe31b31a1d49d11973025a8b832d78c17995d25fe24aad52723a657. It derives from UCSC hg38 multiz100way, with the human row replaced by the Ensembl 107 primary assembly. It is not Zoonomia241. Some upstream nonhuman unknown symbols were folded into gaps; those lost distinctions cannot be recovered from the processed file.

CuEvo inferred posterior substitution counts on the fixed 100-tip/198-branch JC69 species tree. Counts and usable-base totals were pooled over complete physical 1 kb reference windows before rate transformation. A base needs at least two comparable taxa to contribute to the count/usable planes; taxon coverage still uses the complete physical window as denominator. No MAF or informative-position subsampling defines these windows. Branch observability is pooled with OR.

For each window, rates use sqrt(counts/usable_bases), are centered across branches, projected against the centered frozen whole-tier expectation, and winsorized at the recorded 5% setting. Pearson r is then calculated after re-centering each pair on its shared observed branches. The profile inputs and stored scores retain their existing float64 representation. Source per-position posterior counts were stored in float32 before float64 window accumulation; the saved float64 arrays do not recover that earlier rounding. Calibration identity, source hashes, parameters, scan counts, and eligibility by chromosome are recorded in PROVENANCE.json; converted files have hashes and exact sizes in manifest.json.

The released catalog was checked for duplicate pairs, coordinate/profile agreement, recorded coverage gates, and r > 0.25. Independent formula replay of 128 pairs had maximum absolute score error 2e-15. Seven pooled levels conserved the pair count; 28 sampled pooled cells were checked independently for counts, means, and maxima. These numerical checks do not establish statistical significance.

Access and metadata-first download

Sign in on the dataset page and agree to share your Hugging Face username and profile email with Standard Model Bio. Access is approved automatically.

python -m pip install huggingface_hub pyarrow numpy
hf auth login

import json
from huggingface_hub import HfApi, hf_hub_download

repo = "standardmodelbio/hg-horizon-coevolution-erc"
revision = HfApi().dataset_info(repo, token=True).sha
print("Pin this revision:", revision)
path = hf_hub_download(repo, "manifest.json", repo_type="dataset", revision=revision, token=True)
manifest = json.load(open(path))
print(manifest["retained_pairs"], manifest["data_bytes"])

Read a retained pair partition

import pyarrow.parquet as pq

part = manifest["chromosome_pairs"][0]
path = hf_hub_download(repo, part["path"], repo_type="dataset", revision=revision, token=True)
table = pq.ParquetFile(path).read_row_group(0, columns=[
    "start_i", "end_i", "start_j", "end_j", "erc_r", "n_shared_branches",
    "effective_taxa_i", "effective_taxa_j", "effective_taxa_joint",
])
print(part["chrom_i"], part["chrom_j"])
print(table.slice(0, 5).to_pylist())

Read both mean- and max-pooled genome scores

path = hf_hub_download(repo, "pooled/genome_5000000.parquet", repo_type="dataset", revision=revision, token=True)
table = pq.read_table(path, columns=["genome_start_i", "genome_start_j", "n_pairs", "sum_r", "mean_r", "max_r"])
print(table.slice(0, 5).to_pylist())
# mean_r == sum_r / n_pairs; the population is retained 1 kb pairs, not all possible pairs.

Read converted profile inputs

import numpy as np

path = hf_hub_download(repo, "inputs/profiles_1000/chr22.npz", repo_type="dataset", revision=revision, token=True)
with np.load(path, allow_pickle=False) as profiles:
    print(profiles.files)
    print(profiles["genome_start"][:5])
    counts = profiles["posterior_counts"]
    print(counts.shape, counts.dtype, counts[0, :5])

Download only selected products from the command line

hf download standardmodelbio/hg-horizon-coevolution-erc --repo-type dataset --include 'pooled/*' 'manifest.json' 'PROVENANCE.json' 'genome-axis.json' --local-dir hg-horizon-erc

Original alignment, without duplicating it here

This is a 42.02 GB archive. Download it only when you need the original aligned bases; the converted inputs above are smaller and contain the processed branch profiles used here.

hf download songlab/multiz100way 99.zarr.zip --repo-type dataset --revision a790ef9ea692cd2bc9703f3b918337a6de08db9b --local-dir multiz100way

License

CC BY 4.0 See PROVENANCE.json for additional upstream data.