marin-dna/human-genome

GRCh38 — Ensembl release 115 soft-masked primary assembly

The Ensembl release 115 GRCh38 soft-masked primary assembly, provided in uncompressed and BGZF-compressed FASTA formats. Both variants include indexes for remote random-access sequence queries without downloading the entire genome.

Files

File	Purpose
Homo_sapiens.GRCh38.dna_sm.primary_assembly.fa	Uncompressed FASTA, optimized for remote byte-range queries
Homo_sapiens.GRCh38.dna_sm.primary_assembly.fa.fai	Uncompressed FASTA index
Homo_sapiens.GRCh38.dna_sm.primary_assembly.fa.gz	BGZF-compressed FASTA, optimized for full downloads while retaining random access
Homo_sapiens.GRCh38.dna_sm.primary_assembly.fa.gz.fai	BGZF FASTA index
Homo_sapiens.GRCh38.dna_sm.primary_assembly.fa.gz.gzi	BGZF block index
SHA256SUMS	SHA-256 checksums
README.md	Dataset documentation

Provenance

Source: Ensembl release 115

Source SHA-256:

651561b3065aa6083b62290ec96344a90d07388883874070aa368b745fde68fd

The source is pinned in the MarinDNA genome pipeline.

The Ensembl gzip file was decompressed to produce the uncompressed FASTA and recompressed with BGZF for indexed compressed access. Indexes were generated with samtools faidx.

Assembly details:

• Assembly: GRCh38
• Ensembl release: 115
• Sequence selection: primary assembly
• Repeat masking: soft-masked
• Ensembl-style sequence names such as 1, 2, X, Y, and MT

Remote access

For a 0-based, half-open interval [100000, 101000), convert at the samtools boundary to the 1-based, closed interval 1:100001-101000:

BASE=https://huggingface.co/datasets/marin-dna/human-genome/resolve/main

samtools faidx \
  "$BASE/Homo_sapiens.GRCh38.dna_sm.primary_assembly.fa" \
  "1:100001-101000"

The BGZF version also supports remote indexed access:

samtools faidx \
  "$BASE/Homo_sapiens.GRCh38.dna_sm.primary_assembly.fa.gz" \
  "1:100001-101000"

For reproducible applications, replace main with a pinned Hugging Face commit revision so the FASTA and indexes cannot drift apart.

Python remote access

Install pyfaidx and the HTTP support for fsspec:

uv add pyfaidx "fsspec[http]"

Query the uncompressed FASTA directly. The slice uses 0-based, half-open coordinates and is retrieved with HTTP byte-range requests rather than by downloading the complete FASTA:

import fsspec
from pyfaidx import Fasta

url = (
    "https://huggingface.co/datasets/marin-dna/human-genome/resolve/main/"
    "Homo_sapiens.GRCh38.dna_sm.primary_assembly.fa"
)

with Fasta(fsspec.open(url), as_raw=True) as genome:
    sequence = genome["1"][100_000:101_000]

assert len(sequence) == 1_000

For reproducible applications, replace main in the Python URL with a pinned Hugging Face commit revision.

Full download

hf download marin-dna/human-genome --repo-type dataset