refbook

One catalog for the many published versions of the human reference genome.

"GRCh38" names dozens of different FASTA files: UCSC, Ensembl, GENCODE, NCBI's analysis sets, Broad/GATK, 1000 Genomes, iGenomes. They differ in which sequences they contain, how they name and order them, and whether some bases are masked or rewritten. refbook describes every one of them as a recipe applied to the same roots, the original INSDC sequences, and counts a version as cataloged only when its recipe reproduces the published file sequence by sequence.

Release 0.1.0 covers 17 published GRCh38 files, which reduce to 13 distinct sequence sets.

Publisher	Files reproduced
Broad / 1000 Genomes	GRCh38_full_analysis_set_plus_decoy_hla, Homo_sapiens_assembly38
Ensembl 116	primary_assembly, toplevel
GENCODE 50	p14, primary_assembly
GRC (NCBI analysis sets)	full, full + hs38d1, no_alt, no_alt + hs38d1
Illumina iGenomes	GATK/GRCh38, NCBI/GRCh38, NCBI/GRCh38Decoy, UCSC/hg38
UCSC	hg38, hg38.analysisSet, hg38.p14

What is in it

All coordinates are 0-based, half-open. Sequence digests follow GA4GH refget (MD5 and SQ. sha512t24u); collection digests follow GA4GH seqcol.

File	Contents
roots.parquet	The 3,097 roots: GRCh38 p0–p14 (711 sequences), the hs38d1 decoys (2,385) and EBV, with length and refget digests
roots_external.parquet	IMGT/HLA 3.18.0 alleles, by identifier and digest only (see below)
root_blocks.parquet	The bases of every root, uppercase with IUPAC codes kept, in 64 kb blocks, laid out for HTTP range reads
assembly_members.parquet	Every row of every GRCh38 and hs38d1 assembly report
aliases.parquet	Names of each root in the GenBank, RefSeq, UCSC, GRC, Ensembl and IMGT namespaces
layer_*.parquet	Annotations in root coordinates: AGP components and gaps, alt/patch placements, named regions (centromeres, PARs), the GRC hard-masks, IUPAC positions, N runs, NCBI soft-masking
recipes.parquet	Each published file's recipe and seqcol digest
flavor_sequences.parquet	Each published file's sequences in order: name, root, length, MD5, SQ. digest
family_tree.json	How the files relate: each one's nearest relative and the exact difference
seqcol/*.json	A seqcol (level 2) document per published file
evidence/	What each published file was observed to contain, and the root verification reports
catalog.json	The release manifest: every file's size, sha256 and role

Read any table directly with DuckDB, Polars or pyarrow, for example:

import polars as pl
base = "https://huggingface.co/datasets/abdenlab/refbook/resolve/v0.1.0/"
fs = pl.read_parquet(base + "flavor_sequences.parquet")
fs.filter(pl.col("name") == "chr21").select("flavor", "md5").unique("md5")

A Python library that identifies a header against this catalog and fetches bases by region is in development.

Sources and licensing

The catalog tables are released under CC-BY-4.0. The underlying sequences are INSDC records (GenBank), distributed by NCBI without restrictions on use. The published files used as reproduction targets belong to their publishers and are not redistributed here; the evidence/ tables record only their sequence names, lengths and digests.

IMGT/HLA alleles are not redistributed. The IPD-IMGT/HLA Database is licensed under CC BY-NoDerivs, and its maintainers ask not to be mirrored. This catalog records only allele identifiers, names and digests, with links to the IMGT/HLA release. If you use the HLA entries, please cite IPD-IMGT/HLA (Barker et al., Nucleic Acids Research 2026; Robinson et al., Tissue Antigens 2000).

Versions

Each release is a tag (v0.1.0, …). Pin URLs to a tag, for example https://huggingface.co/datasets/abdenlab/refbook/resolve/v0.1.0/catalog.json.