songlab/gpn-star-umap-regions

GPN-Star UMAP Regions

A set of 111,329 labeled 100 bp windows of the human genome (hg38/GRCh38), spanning seven functional region classes (100 bp ≈ the median human coding-exon length). This is the exact region set used for the GPN-Star embedding UMAP visualization in Ye et al. 2025.

It can be used as:

• a benchmark for genomic region / functional-element classification, and
• a labeled region set for interpretation of genomic models (e.g. UMAP / probing of sequence embeddings).

Dataset structure

A single test split (test.parquet). Each row is one 100 bp window.

Column	Type	Description
chrom	string	Chromosome, autosomes 1–22 (no chr prefix), hg38/GRCh38
start	int64	Window start, 0-based, inclusive (BED-style)
end	int64	Window end, 0-based, exclusive; end - start == 100 for every row
cons	float32	Window conservation: the 75th percentile, over the window, of per-base primate phastCons scores (posterior probability of conservation, range [0, 1]) from the Zoonomia 2021 Cactus track hub, with undefined bases set to 0.
label	string	Functional region class (see below)

Labels

label	Count	Meaning	Source	Display name (paper)
CDS	20,000	Coding sequence	Ensembl 113 GFF3	CDS
five_prime_UTR	9,222	5′ UTR	Ensembl 113 GFF3	5′ UTR
three_prime_UTR	18,456	3′ UTR	Ensembl 113 GFF3	3′ UTR
lnc_RNA	11,991	Exons of long non-coding RNA transcripts	Ensembl 113 GFF3	lncRNA
PLS	11,660	Promoter-like signature cCRE	ENCODE SCREEN cCRE Registry V4	Promoter
dELS	20,000	Distal enhancer-like signature cCRE	ENCODE SCREEN cCRE Registry V4	Enhancer
background	20,000	Windows ≥ 100 bp away from any exon, cCRE, or repeat	—	Background

How it was created

The regions were produced by the GPN-Star interpretation pipeline:

• Pipeline (permalink): analysis/gpn-star/interpretation
• Rules that build this file (permalink): workflow/rules/umap.smk

To reproduce test.parquet, from analysis/gpn-star/interpretation:

uv venv --python 3.13
uv pip install -r requirements.txt
uv run snakemake --cores all \
    results/interpretation/windows/v8_0_subsamplestratified_20000.parquet

The resulting results/interpretation/windows/v8_0_subsamplestratified_20000.parquet is this dataset's test.parquet verbatim. (Building only this target downloads the public annotation tracks above; it does not require the multiple-sequence alignments or model checkpoints — those are only needed for the downstream embedding/UMAP steps.)

Construction summary

1. Source annotations. CDS, 5′/3′ UTR and lnc_RNA exons from Ensembl release-113 GFF3; PLS/dELS cCREs from ENCODE SCREEN Registry V4 (GRCh38-cCREs.bed); UCSC RepeatMasker (rmsk) and assembly gaps used for exclusion.
2. Non-overlapping classes. Each category is merged and made mutually non-overlapping (genic classes also subtract one another, cCREs and repeats; cCREs subtract exons and repeats). background = defined autosomal regions ≥ 100 bp (padded) from any exon, cCRE or repeat, and ≥ 1024 bp from undefined/gap regions.
3. Tiling. Each interval is tiled into non-overlapping 100 bp windows (step 100), keeping only full windows.
4. Conservation. Per-window cons = 75th percentile of primate phastCons over the window.
5. Autosomes only. chrX/chrY are dropped.
6. Conservation-stratified subsampling (random_state=42): background is sampled uniformly to 20,000 windows; each foreground class is sampled to up to 20,000 windows, balanced between "conserved" (cons ≥ 1) and "not conserved" (up to 10,000 each).
7. Rows are sorted by (chrom, start, end).

Note: for the UMAP figure the GPN-Star embeddings are computed by adding flanks to each 100 bp window (total context window up to 256 bp; 128 bp for vertebrate models, 256 bp for mammal/primate models), taking per-position embeddings, averaging over the central 100 positions and both strands, then standardizing (Ye et al. 2025, Methods). The 100 bp window is the labeled unit, not the full model input.

Usage

from datasets import load_dataset

ds = load_dataset("songlab/gpn-star-umap-regions", split="test")
df = ds.to_pandas()
print(df.label.value_counts())

Coordinates are 0-based half-open (BED-style); convert to 1-based inclusive (e.g. for VCF/Ensembl-style coordinates) with start + 1 .. end.

Sources

• Ensembl release-113 — https://www.ensembl.org
• ENCODE SCREEN cCRE Registry V4 — https://screen.encodeproject.org
• Zoonomia 2021 Cactus track hub (primate phastCons) — https://cgl.gi.ucsc.edu/data/cactus/zoonomia-2021-track-hub/hg38/
• UCSC RepeatMasker (hg38) — https://hgdownload.soe.ucsc.edu/goldenPath/hg38/database/

Each source is subject to its own terms of use.

Citation

If you use this dataset, please cite GPN-Star:

@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},
  pages={2025--09},
  year={2025},
  publisher={Cold Spring Harbor Laboratory}
}