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.

Corrected MIp coevolution: explicit position selection, signed scores, mean and max

This is a new offline CuEvo computation from Zoonomia241 alignment inputs. It uses A/C/G/T-only pairwise-complete observations, exact identity-neighbor species weights, weighted column coverage >=0.70, pseudocount 0.5, float64 calculations, and observed APC. Signed MIp residuals are preserved. It does not use historical five-state/gap-as-a-state scoring or positive clipping. MIp is MI minus average-product correction, in natural-log units (nats); it is not Pearson r. No score threshold, p-value, q-value, or significance claim is applied.

This dataset contains the corrected MIp release at the original repository URL. Access is automatically approved after sharing your Hugging Face username and profile email.

Resolution and coverage — read first

Product	Source population	Delivered resolution	Pooling
`cis/zoonomia241/chr*`	Spatially balanced informative positions, nominal cap 150,000 per chromosome; then weighted-coverage filtering	50 kb, 100 kb, 250 kb, 500 kb, 1 Mb, 5 Mb, chr1–22,X	All finite unordered selected-position pairs; signed mean and numerical max
`genome/zoonomia241/`	One qualifying position near each physical 250 kb bin center, chr1–22; then weighted-coverage filtering	250 kb, 1 Mb, 5 Mb, includes cis and trans	All finite unordered pairs in this sampled panel; signed mean and numerical max
`inputs/selected/zoonomia241/`	Exact scored/candidate column axes, species codes, weights and frozen APC backgrounds	Selected single-position alignment columns, not a pair catalogue	No spatial pooling
`inputs/msa-informative/zoonomia241/`	Larger informative-column conversions before the spatial cap	0.5–2 Mb physical shards containing retained variable positions	No pair scores

This is not an exhaustive nucleotide-resolution genome pair catalogue. The finest delivered score map is 50 kb for cis data. Large-bin maps are pooled selected-position data. The genome overview is sampled-position data, not a calculation between all bases of physical windows. Full coordinate lists and observed counts are supplied; never infer dense coverage from a chromosome-spanning filename. X is available only in the chromosome product; Y is absent.

Each chromosome has its own selected-column species weights and whole-chromosome APC background. The sampled autosomal overview has one joint sampled-genome background. These products have different populations and backgrounds: do not merge them as interchangeable scores or treat resolution levels as independent observations. Within one product, coarser maps summarize the same measured pair population.

Pooling and schema

Every Parquet row contains pos_i, pos_j (int64 bp), n (int64 pair count), and float64 sum_mip, mean_mip, min_mip, max_mip. Mean=sum_mip/n. Max=max_mip is the largest numerical signed value, not largest absolute magnitude. min_mip also permits reconstruction of a signed largest-magnitude value when required. The denominator includes negative and positive finite scores, with no positive-hit filter. Missing/unscorable pairs are not zeros.

Only unordered distinct source-position pairs are counted. Different source positions falling inside one coarse bin contribute to its diagonal cell; self-pairs do not. All coarser levels conserve the same measured-pair count. Files are sorted by (pos_i,pos_j).

Cis coordinates are zero-based chromosome-local bin starts. Genome coordinates are zero-based on hg38-autosomes-padded-5mb-v1: each chromosome is padded to a multiple of 5 Mb, total span 2,935,000,000 bp. Padding is not measured sequence. Exact offsets are in genome-axis.json. Biological chromosome lengths, including X, are in chromosome-lengths.json; terminal bin ends are clipped to the chromosome length.

Method and provenance

CuEvo revision: a638bf5388c9078dd74c4a95cfa47e9a3318e3bb. The GPU implementation was checked against the independent NumPy reference on synthetic special-symbol cases and real Zoonomia input. The two-pass streamed algorithm was checked against the public dense CuEvo calculation. Every completed population also passed independent NumPy pair audits; pooled counts are checked at every resolution. Scientific validation results are in verification.json. These are numerical correctness checks, not statistical significance or biological validation.

The larger converted source retained positions with variation >=0.1 and roughly 50% non-gap species coverage. For the chromosome cap, positions are ranked by 1-major-base-frequency among observed A/C/G/T and balanced across 400 physical-span windows, up to 375 positions per window, with earlier coordinates breaking ties. Terminal-window behavior can put the nominal cap slightly above 150,000; actual counts are recorded for each chromosome. Source gaps and upstream unknown/missing symbols were already irreversibly conflated into code 4. The corrected estimator excludes that code; it cannot recover the original distinctions or filtered bases.

Species redundancy weights use identity >=0.8 over all selected columns, with regular-symbol comparability. Coverage is evaluated after reweighting. MI uses jointly observed species for each pair and a uniform total pseudocount 0.5 across 16 A/C/G/T joint states. APC uses finite observed matrix entries, with estimable diagonal entries zero. Row and global backgrounds are fitted once over the full declared population before signed scores are exported.

The original UCSC/Zoonomia 241-way alignment is linked rather than duplicated: cactus241way.bigMaf is approximately 1.23 TB. The substantially smaller converted inputs are included here. PROVENANCE.json, manifest.json, and per-chromosome counts identify the exact source files, sizes, checksums, selected positions, masking, scoring policy, and pooling populations.

Contents and measured sizes

Product	Data files	Bytes
`cis/`	138	4,246,304,884
`genome/`	3	538,982,130
`inputs/selected/`	24	303,760,782
`inputs/msa-informative/`	1,617	55,122,091,826
All data	1,782	60,211,139,622

inputs/selected/zoonomia241/chrN.npz contains chromosome-local positions, codes[241,L] (uint8, A/C/G/T=0..3 and upstream gap/unknown/missing=4), ordered species, weights[241], weighted_coverage[L], apc_row_mean[L] and scalar apc_grand_mean. genome_sampled.npz uses padded-axis positions and its own joint sampled-genome weights/background. The files retain the complete selected column axis, including subsequently masked columns; unestimable APC row means are NaN. Compressed NPZ arrays are not memory-mappable.

The larger informative-column shards contain positions, codes[241,L], and species; some also preserve reference_start, reference_codes, and reference_observed for the human reference interval. These are preprocessed informative-column inputs, not complete raw cross-species sequence. They preserve the original conversions without additional rewriting.

Automatic email access and copy/paste examples

Open the dataset page, sign in, and agree to share your username and profile email with Standard Model Bio. Access is approved automatically. Authenticate locally with hf auth login; never paste a token into chat, examples, or issues. Each hg-horizon dataset has its own access request.

python -m pip install huggingface_hub==1.25.1 pyarrow numpy
hf auth login

import json
from huggingface_hub import HfApi, hf_hub_download

repo = "standardmodelbio/hg-horizon-coevolution-mip"
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["data_bytes"], len(manifest["files"]))

import pyarrow.parquet as pq

path = hf_hub_download(repo, "cis/zoonomia241/chr22_50000.parquet", repo_type="dataset", revision=revision, token=True)
table = pq.ParquetFile(path).read_row_group(0, columns=["pos_i", "pos_j", "n", "mean_mip", "max_mip", "min_mip"])
print(table.slice(0, 5).to_pylist())

path = hf_hub_download(repo, "genome/zoonomia241/genome_5000000.parquet", repo_type="dataset", revision=revision, token=True)
table = pq.read_table(path)
print(table.slice(0, 5).to_pylist())
# This overview uses a sampled-position panel; see the coverage table above.

import numpy as np

path = hf_hub_download(repo, "inputs/selected/zoonomia241/chr22.npz", repo_type="dataset", revision=revision, token=True)
with np.load(path, allow_pickle=False) as data:
    print(data["positions"][:5], data["codes"].shape)
    print(data["weights"].shape, data["weighted_coverage"][:5])

To download only the score maps rather than all input alignments:

from huggingface_hub import snapshot_download
snapshot_download(repo, repo_type="dataset", revision=revision, token=True,
                  allow_patterns=["cis/*", "genome/*", "manifest.json", "PROVENANCE.json", "genome-axis.json"],
                  local_dir="hg-horizon-corrected-mip")

License and citation

Standard Model Bio's derived scores, conversions and documentation are CC BY 4.0. Cite this dataset's resolved Hub revision, CuEvo, and the original alignment producers/Zoonomia project. Upstream assembly terms and credits apply: https://genome.ucsc.edu/goldenPath/credits.html .

Downloads last month: 124

Total file size:

60.2 GB

Collection including standardmodelbio/hg-horizon-coevolution-mip

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