Hugging Face's logo

Datasets:
songlab
/
gpn-star-scores

Modalities:
Tabular
Text
Formats:
parquet
Size:
10B - 100B
Tags:
biology
genomics
variant-effect-prediction
DOI:
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
Dataset Viewer
Auto-converted to Parquet Duplicate
chrom
stringclasses
1 value
pos
int64
10.1k
249M
ref
stringclasses
4 values
alt
stringclasses
4 values
llr_calibrated
float32
-13.5
14.5
abs_llr_calibrated
float32
-4.15
13.5
1
10,129
C
A
1.991
-1.991
1
10,129
C
G
1.119
-1.119
1
10,129
C
T
1.188
-1.188
1
10,130
C
A
1.466
-1.466
1
10,130
C
G
-0.39
0.39
1
10,130
C
T
1.722
-1.722
1
10,131
C
A
1.679
-1.679
1
10,131
C
G
-0.264
0.264
1
10,131
C
T
1.865
-1.865
1
10,132
T
A
3.446
-1.292
1
10,132
T
C
3.707
-0.659
1
10,132
T
G
0.997
-0.997
1
10,133
A
C
2.884
-1.736
1
10,133
A
G
1.601
-1.601
1
10,133
A
T
2.341
-2.341
1
10,134
A
C
2.57
-2.144
1
10,134
A
G
1.422
-1.422
1
10,134
A
T
1.881
-1.881
1
10,135
C
A
1.997
-1.997
1
10,135
C
G
1.13
-1.13
1
10,135
C
T
1.184
-1.184
1
10,136
C
A
1.447
-1.447
1
10,136
C
G
-0.386
0.386
1
10,136
C
T
1.738
-1.738
1
10,137
C
A
1.67
-1.67
1
10,137
C
G
-0.273
0.273
1
10,137
C
T
1.865
-1.865
1
10,138
T
A
3.451
-1.287
1
10,138
T
C
3.712
-0.654
1
10,138
T
G
1.028
-1.028
1
10,139
A
C
2.886
-1.734
1
10,139
A
G
1.592
-1.592
1
10,139
A
T
2.337
-2.337
1
10,140
A
C
2.506
-2.209
1
10,140
A
G
1.385
-1.385
1
10,140
A
T
1.801
-1.801
1
10,141
C
A
2.044
-2.044
1
10,141
C
G
1.139
-1.139
1
10,141
C
T
1.16
-1.16
1
10,142
C
A
1.384
-1.384
1
10,142
C
G
-0.432
0.432
1
10,142
C
T
1.669
-1.669
1
10,143
C
A
1.607
-1.607
1
10,143
C
G
-0.297
0.297
1
10,143
C
T
1.803
-1.803
1
10,144
T
A
3.481
-1.258
1
10,144
T
C
3.759
-0.607
1
10,144
T
G
1.098
-1.098
1
10,145
A
C
2.96
-1.659
1
10,145
A
G
1.635
-1.635
1
10,145
A
T
2.304
-2.304
1
10,146
A
C
2.334
-2.334
1
10,146
A
G
1.364
-1.364
1
10,146
A
T
1.592
-1.592
1
10,147
C
A
2.189
-2.189
1
10,147
C
G
1.356
-1.356
1
10,147
C
T
1.24
-1.24
1
10,148
C
A
1.028
-1.028
1
10,148
C
G
-0.688
0.688
1
10,148
C
T
1.58
-1.58
1
10,149
C
A
0.861
-0.861
1
10,149
C
G
-0.875
0.875
1
10,149
C
T
1.167
-1.167
1
10,150
C
A
0.949
-0.949
1
10,150
C
G
-0.669
0.669
1
10,150
C
T
1.541
-1.541
1
10,151
T
A
3.891
-0.848
1
10,151
T
C
3.665
-0.701
1
10,151
T
G
1.392
-1.392
1
10,152
A
C
2.546
-2.073
1
10,152
A
G
1.388
-1.388
1
10,152
A
T
2.053
-2.053
1
10,153
A
C
2.701
-2.013
1
10,153
A
G
1.438
-1.438
1
10,153
A
T
1.994
-1.994
1
10,154
C
A
2.118
-2.118
1
10,154
C
G
1.193
-1.193
1
10,154
C
T
1.122
-1.122
1
10,155
C
A
1.402
-1.402
1
10,155
C
G
-0.44
0.44
1
10,155
C
T
1.624
-1.624
1
10,156
C
A
1.367
-1.367
1
10,156
C
G
-0.417
0.417
1
10,156
C
T
1.785
-1.785
1
10,157
T
A
3.653
-1.085
1
10,157
T
C
3.733
-0.633
1
10,157
T
G
1.152
-1.152
1
10,158
A
C
2.645
-1.975
1
10,158
A
G
1.513
-1.513
1
10,158
A
T
2.151
-2.151
1
10,159
A
C
2.686
-2.028
1
10,159
A
G
1.478
-1.478
1
10,159
A
T
2.068
-2.068
1
10,160
C
A
2.116
-2.116
1
10,160
C
G
1.212
-1.212
1
10,160
C
T
1.132
-1.132
1
10,161
C
A
1.488
-1.488
1
10,161
C
G
-0.387
0.387
1
10,161
C
T
1.704
-1.704
1
10,162
C
A
1.462
-1.462
End of preview. Expand in Data Studio

GPN-Star genome-wide scores

GPN-Star tracks in the UCSC Genome Browser

Genome-wide, mutation-rate-calibrated GPN-Star constraint and variant scores for eight score sets covering human, mouse, chicken, D. melanogaster, C. elegans, and A. thaliana. Canonical scores are available as chromosome-sharded Parquet. Hugging Face hosts 72 BigWigs; a multi-assembly UCSC track hub references the 64 logo/LLR views.

Overview and quick links

Resource Link
Files Browse all dataset files
Default Dataset Viewer config Human mammalian-model LLR
UCSC hub Load the multi-assembly GPN-Star hub
Human browser quick start Open UCSC's current default hg38 context plus GPN-Star-M tracks
Source code gonzalobenegas/gpn-star-scores
Immutable source snapshot d8e7e415d17865a3579885bdb9cb2728c9113970
Paper Predicting functional constraints across evolutionary timescales with phylogeny-informed genomic language models
Immutable public metadata base cdaa131f2a105fae804ffdba9c67a7f128e69f70
Release manifest manifest/release.json

Parquet is the canonical analysis product. It retains the supplied chromosome names, one-based positions, and Float32 score values. BigWigs are display products with explicit coordinate conversion and three-decimal values where documented.

What is included

  • Eight score sets: three GPN-Star models on hg38 plus mm39, gg6, dm6, ce11, and tair10.
  • 16 explicit Parquet configurations: one LLR and one entropy configuration for each score set.
  • 290 canonical chromosome shards containing 51,402,120,888 rows.
  • 72 browser BigWigs: 40 v1 entropy and sequence-logo artifacts plus 32 signed A/C/G/T LLR artifacts.
  • One multi-assembly UCSC hub spanning six working UCSC databases and referencing 64 logo/LLR BigWigs. Entropy BigWigs remain downloadable but are not displayed in the hub.
  • Machine-readable manifests containing file identities, sizes, SHA-256 checksums, row counts, and immutable revision provenance.

Score sets, assemblies, and browser databases

A score set is one model's released directory. A Parquet assembly name identifies the reference assembly and chromosome naming used by its canonical tables. A UCSC database name selects the compatible browser database; it can differ from the Parquet assembly name.

Score set Organism/model Parquet assembly Working UCSC database Model
gpn-star-hg38-v100-200m Human, 100-vertebrate alignment hg38 hg38 GPN-Star-V
gpn-star-hg38-m447-200m Human, 447-mammal alignment hg38 hg38 GPN-Star-M
gpn-star-hg38-p243-200m Human, 243-primate alignment hg38 hg38 GPN-Star-P
mm39 Mouse, 35-way mammal alignment mm39 mm39 GPN-Star mm39
gg6 Chicken, 77-way vertebrate alignment gg6 galGal6 GPN-Star galGal6
dm6 D. melanogaster, 124-way insect alignment dm6 dm6 GPN-Star dm6
ce11 C. elegans, 135-way nematode alignment ce11 ce11 GPN-Star ce11
tair10 A. thaliana, 18-way plant alignment tair10 GCF_000001735.4 GPN-Star TAIR10

Coordinate conventions

  • Parquet: supplied assembly chromosome names without an added chr prefix and one-based pos: Int64 coordinates.
  • BigWig and UCSC: UCSC chromosome names and zero-based, half-open intervals.

For example, Parquet position chrom="22", pos=20_000_001 is represented in a BigWig as the one-base interval chr22:20_000_000-20_000_001. Keep this conversion explicit when comparing table and browser values.

What the scores mean

The interpretations in this section come from the source Box README.md, file 2154252568578, version 2495703203934. The interpretation and example blocks are reproduced verbatim.

LLR and absLLR

llr_calibrated is a variant-level score. Each LLR table has one row for every genomic position × non-reference A/C/G/T allele, normally three alternate rows per covered position. The score compares the alternate allele with the reference allele.

Column Type Meaning
chrom String Supplied assembly chromosome name
pos Int64 One-based genomic position
ref String Reference nucleotide
alt String Non-reference alternate nucleotide
llr_calibrated Float32 Mutation-rate-calibrated log-likelihood ratio, alternate versus reference
abs_llr_calibrated Float32 Mutation-rate-calibrated absolute LLR

llr_calibrated: More negative = more constrained or larger effect.

abs_llr_calibrated: Magnitude of the variant's effect relative to a neutral substitution. Positive = larger effect than neutral; negative = smaller. Useful when the direction of effect is not relevant.

Example:

chrom     pos ref alt  llr_calibrated  abs_llr_calibrated
   21 5010065   T   A          -1.774               1.774
   21 5010065   T   C          -1.550               1.550
   21 5010065   T   G          -1.670               1.670

Release data note: abs_llr_calibrated is an independently supplied calibrated score. It must not be calculated as abs(llr_calibrated), and matching magnitudes in example rows do not imply a derivation.

The release exposes three distinct LLR-related representations:

  1. Canonical Parquet llr_calibrated: the supplied full-precision variant-level Float32 value.
  2. Signed A/C/G/T LLR BigWigs: browser values that retain each alternate allele's llr_calibrated; the reference allele is assigned an explicit zero. UCSC displays these tracks as -LLR without modifying the files. These tracks are rounded to three decimal places.
  3. A/C/G/T sequence-logo BigWigs: visualization heights derived from calibrated LLRs. They are not LLRs, probabilities, or canonical score products.

Entropy

entropy_calibrated is a position-level score with one row per covered genomic position.

Column Type Meaning
chrom String Supplied assembly chromosome name
pos Int64 One-based genomic position
ref String Reference nucleotide
entropy_calibrated Float32 Mutation-rate-calibrated positional entropy

Interpretation: ~1.0 = neutral; <1.0 = constrained; the lower the more constrained.

Example:

chrom     pos ref  entropy_calibrated
   21 5010065   T               0.486
   21 5010066   A               0.644
   21 5010067   A               0.591

Query the scores in code

This example joins UKB fine-mapped variants with GPN-Star (M) LLR scores. It filters each chromosome shard to the exact requested positions before joining.

The example contains 11,400 variants across the 22 autosomes. On an AMD EPYC 7543 CPU node with 16 allocated CPUs and Polars 1.42.1, the remote version took 10 minutes 41 seconds and the local version took 5.94 seconds.

The local version requires 22 Parquet shards totaling about 52 GiB.

Runtime depends not only on the number of variants, but also on their genomic locality. Variants clustered within one gene generally touch fewer Parquet row groups and run faster than sparse variants spread across a chromosome or the whole genome.

import polars as pl

keys = ["chrom", "pos", "ref", "alt"]

variants = pl.read_parquet(
    "hf://datasets/songlab/ukb_finemapped_nc_traitgym/test.parquet",
    columns=keys,
)

# Remote scores (no full score download required):
score_root = (
    "hf://datasets/songlab/gpn-star-scores/"
    "data/gpn-star-hg38-m447-200m/llr"
)

# Local scores (use this instead after downloading the LLR shards):
# score_root = "/path/to/gpn-star-scores/data/gpn-star-hg38-m447-200m/llr"

results = []

for chrom in variants.get_column("chrom").unique(maintain_order=True):
    chrom_variants = variants.filter(pl.col("chrom") == chrom)
    positions = chrom_variants.get_column("pos").unique().to_list()

    scores = (
        pl.scan_parquet(f"{score_root}/llr_chr{chrom}.parquet")
        .filter(pl.col("pos").is_in(positions))
    )

    results.append(
        chrom_variants.lazy()
        .join(scores, on=keys, how="left")
        .collect(engine="streaming")
    )

annotated = pl.concat(results)

Use the UCSC Genome Browser

The stable multi-assembly hub URL loads eight GPN-Star model groups across six working UCSC databases. Each model group includes:

  • four derived A/C/G/T sequence-logo tracks; and
  • four signed A/C/G/T tracks displayed as -LLR, with an explicit zero for the reference allele.

Entropy BigWigs remain available in the dataset but are not part of the UCSC hub. The logo and -LLR default to 16 pixels high. The -LLR rows default to a dense grayscale 0–10 view and use UCSC's display-time negation: higher displayed values correspond to more-negative source LLR and therefore greater constraint or a larger predicted functional effect. When expanded, negative source LLR appears as positive -LLR in muted blue (60,60,140), while positive source LLR appears as negative -LLR in muted red (140,60,60).

Model-specific launch links

The links below start from UCSC's clean default settings with ignoreCookie=1, retain UCSC's native default context tracks, and add one selected GPN-Star model group with the sequence logo at full and -LLR at dense. For assemblies with multiple GPN-Star models, the other GPN-Star groups are explicitly hidden so that only the selected model is displayed. The links intentionally do not use hideTracks=1, because that parameter would remove the native context shown in the default Genome Browser view.

The links do not set position. Each one therefore inherits whatever locus and native context tracks UCSC currently defines as the default for that database. If UCSC changes a default locus or adds default tracks in the future, the link inherits those changes while continuing to add the selected GPN-Star model group.

Score set UCSC database Browser Position policy
gpn-star-hg38-v100-200m hg38 Open default context + GPN-Star-V tracks Current UCSC default (dynamic)
gpn-star-hg38-m447-200m (default) hg38 Open default context + GPN-Star-M tracks Current UCSC default (dynamic)
gpn-star-hg38-p243-200m hg38 Open default context + GPN-Star-P tracks Current UCSC default (dynamic)
mm39 mm39 Open default context + GPN-Star tracks Current UCSC default (dynamic)
gg6 galGal6 Open default context + GPN-Star tracks Current UCSC default (dynamic)
dm6 dm6 Open default context + GPN-Star tracks Current UCSC default (dynamic)
ce11 ce11 Open default context + GPN-Star tracks Current UCSC default (dynamic)
tair10 GCF_000001735.4 Open default context + GPN-Star tracks Current UCSC default (dynamic)

This dynamic-default behavior is intentional. It prioritizes a familiar, up-to-date UCSC context over reproducing one historical locus forever.

Genomic span is controlled by the position interval. Image width in pixels can be controlled with UCSC's pix parameter, but the primary interactive links leave pixel width responsive to the user's browser.

File layout, implementation, and provenance

Complete public file layout 
data/
  <score-set>/
    entropy/
      entropy_chr<chrom>.parquet
    llr/
      llr_chr<chrom>.parquet
bigwig/
  <score-set>/
    entropy.bw
    A.bw
    C.bw
    G.bw
    T.bw
    llr_A.bw
    llr_C.bw
    llr_G.bw
    llr_T.bw
manifest/
  release.json
  raw-llr-validation.json
  ucsc-hub.json
ucsc/
  hub.txt
  genomes.txt
  <ucsc-database>/
    trackDb.txt
README.md

Browser transformation details

  • Entropy BigWigs contain entropy_calibrated.
  • Signed LLR BigWigs retain alternate-allele llr_calibrated, assign the reference allele an explicit zero for display, and do not read or derive abs_llr_calibrated.
  • For the derived sequence-logo tracks, the reference nucleotide receives logit zero and the three alternate nucleotides receive their independently supplied llr_calibrated values. A stable Float64 softmax produces base weights, and each height is p(base) * (2 - H) for base-2 entropy H.
  • Final browser values are stored as Float32 and rounded to three decimals. They are visualization values rather than canonical scores or raw model probabilities.

Immutable provenance

Component Immutable identity
Original interpretation source Box README.md, file 2154252568578, version 2495703203934, SHA-1 4b553074826e6a711d5308409ac8b1a0129d9f66
Canonical Parquet and v1 entropy/logo artifact revision 5c799b2ec6aa089f0caa8294ae72adb4510f81ae
Signed LLR BigWig artifact revision 47e7f051113abab49f04f43f9107cae2cbbfd34d
README publication base cdaa131f2a105fae804ffdba9c67a7f128e69f70
Dataset-card source implementation commit d8e7e415d17865a3579885bdb9cb2728c9113970

The release manifest records Parquet and BigWig identities, sizes, checksums, row counts, and artifact revisions. The UCSC manifest records browser database mappings, track URLs, and validation scope. Metadata-only card publication does not upload, rewrite, or delete any Parquet or BigWig.

License and citation

The dataset is released under the Apache License 2.0. Please cite:

Ye C, Benegas G, Albors C, Li JC, Prillo S, Fields PD, Clarke B, Song YS. Predicting functional constraints across evolutionary timescales with phylogeny-informed genomic language models. bioRxiv (2025). doi: 10.1101/2025.09.21.677619.

@article{ye2025predicting,
  title={Predicting functional constraints across evolutionary timescales with
    phylogeny-informed genomic language models},
  author={Ye, Chengzhong and Benegas, Gonzalo and Albors, Carlos and Li,
    Jianan Canal and Prillo, Sebastian and Fields, Peter D and Clarke, Brian
    and Song, Yun S},
  journal={bioRxiv},
  year={2025},
  doi={10.1101/2025.09.21.677619}
}
Downloads last month
862
Number of rows:
51,402,120,888
Total file size:
722 GB

Collection including songlab/gpn-star-scores