Boom5426/HPA-v25-subcellular

Human Protein Atlas v25: subcellular localization and full annotation tables

A verbatim mirror of three Human Protein Atlas (HPA) release files, packaged together with an explicit provenance and schema description. Nothing here is derived, filtered, or recomputed: the bytes are exactly what the upstream endpoints served.

The point of this repository is reproducibility. HPA serves only the current release at a stable URL, so a file downloaded six months from now will silently differ from the one used in a past experiment. Each file below is pinned with a SHA-256 and a retrieval date.

What HPA is

The Human Protein Atlas is a Swedish research program that maps the human proteome using antibody-based imaging, mass spectrometry, and transcriptomics. The subcellular section (the Cell Atlas) assigns proteins to organelles and subcellular structures by immunofluorescence (IF/ICC) in human cell lines, with a four-tier reliability grading. The measured entity is protein, but records are keyed by gene (ENSG), because HPA entries are gene-centric.

Files

File	Size	Rows x Cols	HPA version	Retrieved	SHA-256
subcellular_location.tsv	2,046,017 B	13,603 x 14	v25	2026-07-05	19ee031772375318e75556e43a5103cd8d919f7757094ef3fd134fba134fc095
proteinatlas.xml.gz	729,899,935 B	20,162 entries	v25 (declared in file)	2026-04-21	a8d12feab9d1c9846787fdd3f5efbe843d4247a898227830ff870ac18287f8ca
unversioned_mirror/proteinatlas.tsv	36,611,106 B	20,162 x 107	undetermined, pre-v25	2026-04-04	5f629a170520b747d12e2602471e0bbc2ee6a5c4968353f00143ffe26cec154a

Upstream sources:

• subcellular_location.tsv from https://www.proteinatlas.org/download/tsv/subcellular_location.tsv.zip (unzipped, byte-identical to the archive member). Upstream Last-Modified was 2025-11-06, consistent with the v25.0 data drop (v25.0 released 2025-11-11).
• proteinatlas.xml.gz from https://www.proteinatlas.org/download/proteinatlas.xml.gz. The file declares its own version: schemaLocation="https://v25.proteinatlas.org/download/proteinatlas.xsd", schemaVersion="3.0", and every <entry version="25">. Identifiers are Ensembl 109 on GRCh38.p13, GENCODE 43. gzip -t passes.
• unversioned_mirror/proteinatlas.tsv was not obtained from HPA directly. It came from a third-party data-lake mirror (the Biomni data lake). See the version warning below.

subcellular_location.tsv schema

13,603 genes, 14 tab-separated columns. This is the Cell Atlas summary table: one row per gene, aggregating IF observations across the cell lines that gene was stained in.

#	Column	Non-null	Description
1	Gene	100%	Ensembl gene ID (ENSG). 13,603 unique. The reliable join key.
2	Gene name	100%	HGNC symbol. Only 13,597 unique, see caveats.
3	Reliability	100%	Gene-level summary grade: Approved 6,999 / Supported 4,582 / Enhanced 1,230 / Uncertain 792
4	Main location	100% (13,602)	Primary localization, semicolon-separated, multi-label
5	Additional location	46.5%	Secondary localization, semicolon-separated
6	Extracellular location	5.4%	Single value Predicted to be secreted for 740 genes. A prediction, not a measurement.
7	Enhanced	9.6%	Locations whose evidence reached the Enhanced tier
8	Supported	38.7%	Locations at the Supported tier
9	Approved	58.8%	Locations at the Approved tier
10	Uncertain	17.9%	Locations at the Uncertain tier
11	Single-cell variation intensity	26.8%	Compartments showing cell-to-cell intensity heterogeneity (3,650 genes)
12	Single-cell variation spatial	1.4%	Compartments showing spatial heterogeneity (186 genes)
13	Cell cycle dependency	3.5%	Cell-cycle-dependent structures (479 genes), e.g. Cytokinetic bridge, Mitotic spindle, Mitotic chromosome
14	GO id	100%	Location terms with GO accessions, e.g. Cytosol (GO:0005829)

How columns 7-10 relate to columns 4-5

These four columns are not a repeat of the reliability grade. They re-partition the same location labels by evidence strength. Verified on this file:

• For all 13,603 rows, Enhanced U Supported U Approved U Uncertain equals Main location U Additional location exactly. Zero exceptions.
• 3,099 genes carry locations spanning more than one evidence tier (2 tiers: 2,806; 3 tiers: 292; 4 tiers: 1). Reliability reports only that gene's best tier.

Practical consequence: to drop low-confidence localizations, remove the terms listed in the Uncertain column, not the rows where Reliability == "Uncertain". The Uncertain column is non-null for 2,430 genes but only 786 of those have Reliability == "Uncertain", so row-level filtering leaves uncertain labels on 1,644 genes.

Multi-label structure

Main location count per gene: 1 location for 10,035 genes, 2 for 3,065, 3 for 407, 4 for 79, 5 for 10, 6 for 5, 7 for 1, and 1 gene with none. Additional location is empty for 7,277 genes.

Location vocabulary

49 distinct terms across Main location and Additional location, heavily skewed:

Nucleoplasm 6283, Cytosol 5267, Vesicles 2488, Plasma membrane 2263, Golgi apparatus 1279,
Mitochondria 1132, Nucleoli 1116, Nuclear bodies 626, Endoplasmic reticulum 576,
Centrosome 513, Nuclear speckles 507, Primary cilium 467, Basal body 451,
Principal piece 371, Microtubules 356, Mid piece 352, Cell Junctions 331,
Nucleoli fibrillar center 325, Nuclear membrane 295, Actin filaments 253,
Cytokinetic bridge 245, Centriolar satellite 228, End piece 195, Mitotic spindle 161,
Nucleoli rim 153, Focal adhesion sites 148, Intermediate filaments 147,
Primary cilium tip 130, Flagellar centriole 116, Acrosome 112,
Primary cilium transition zone 96, Connecting piece 92, Equatorial segment 82,
Mitotic chromosome 79, Cytoplasmic bodies 74, Perinuclear theca 70, Calyx 70,
Midbody 56, Annulus 50, Lipid droplets 40, Midbody ring 25, Peroxisomes 24,
Lysosomes 20, Rods & Rings 19, Aggresome 19, Endosomes 16, Kinetochore 8,
Microtubule ends 7, Cleavage furrow 2

Two terms (Nucleoplasm, Cytosol) account for a large share of all annotations. Any analysis that treats compartment sharing as a uniform feature will be dominated by these two.

proteinatlas.xml.gz

The complete HPA v25 release in XML, 20,162 gene entries, roughly 700 MB compressed. This is the superset: everything in the TSV tables is derivable from it, plus data that has no TSV equivalent. Per-entry sections include identifier, name, synonym, proteinClasses, proteinEvidence, predictedLocation, antibody, tissueExpression, cellExpression, cellTypeExpression, cancerExpression, rnaExpression, and proteinstructure.

The antibody blocks are the reason to use the XML rather than the TSV. They carry per-antibody, per-cell-line IF results, including the individual imageUrl entries pointing at the raw immunofluorescence JPGs on images.proteinatlas.org. The summary TSV collapses all of that to one row per gene.

Stream it, do not load it. xml.etree.ElementTree.iterparse over gzip.open works fine and keeps memory flat.

unversioned_mirror/proteinatlas.tsv and its version warning

20,162 genes x 107 columns: the wide per-gene annotation table covering protein class, tissue and single-cell RNA specificity, blood and brain expression, antibody reliability, subcellular location, secretome, expression clusters, and TCGA prognostics.

This file is not from a pinned HPA release and its version could not be determined. It was mirrored from a third-party data lake, not from proteinatlas.org. It predates the v25 files above: its cancer column is RNA cancer specific FPKM, whereas v25 uses RNA cancer specific pTPM. Its subcellular columns also disagree with subcellular_location.tsv: Subcellular location and Subcellular main location are non-null for 13,534 and 13,533 genes, against 13,603 and 13,602 in the v25 table.

It is included because it is a convenient wide-format annotation table, and because leaving it out of a mirror while it circulates elsewhere unlabeled is worse than shipping it with the caveat attached. Do not cite it as a specific HPA release. If you need the wide table at a known version, download proteinatlas.tsv fresh from HPA and record the release number.

Caveats before you use this

1. Cell-line context. HPA IF localization is measured mostly in U2OS, A-431, and U-251MG. The summary table is a cross-cell-line consensus, not a context-specific label. It does not describe localization in your cell line of interest.
2. Antibody-dependent. Localization comes from antibody staining, so it inherits antibody specificity limits. The reliability tiers exist for this reason and should be used.
3. Duplicate symbols. Gene name has 6 symbols appearing twice, each mapping to two distinct ENSGs: NPIPA9, POLR2J3, MATR3, HERC3, PRODH, PINX1. Keying by symbol silently drops or merges records. Key by ENSG.
4. Extracellular location is a prediction, not an observation, and should not be pooled with the measured location columns.
5. Gene-level keying. Records are per gene, so isoform-resolved or proteoform-resolved localization is not available here. Antibody epitopes usually span regions shared by isoforms, making each label an isoform mixture.
6. Not a quantitative readout. These are categorical assignments with a reliability grade, not abundances or fractional distributions across compartments.

Not included

The raw immunofluorescence images (roughly 92,000 JPGs, about 28 GB) are not mirrored here. Their URLs are inside proteinatlas.xml.gz under the antibody blocks and can be fetched directly from images.proteinatlas.org.

License and citation

HPA data is released under CC BY 4.0, which permits redistribution with attribution. This mirror carries the same license. HPA notes that some integrated third-party datasets within the atlas may carry their own terms, so check before reusing a specific subset commercially.

Required attribution, per HPA: cite a primary publication and reference the Human Protein Atlas with a versioned URL, https://v25.proteinatlas.org.

Cell Atlas primary reference:

Thul PJ, et al. A subcellular map of the human proteome. Science. 2017;356(6340):eaal3321.

Loading

import pandas as pd

loc = pd.read_csv("subcellular_location.tsv", sep="\t", dtype=str)

# correct low-confidence filtering: drop terms, not rows
def clean_locations(row):
    keep = set()
    for col in ("Main location", "Additional location"):
        v = row[col]
        if isinstance(v, str):
            keep |= {x.strip() for x in v.split(";")}
    unc = row["Uncertain"]
    if isinstance(unc, str):
        keep -= {x.strip() for x in unc.split(";")}
    return keep

loc["locations"] = loc.apply(clean_locations, axis=1)

# stream the XML instead of loading it
import gzip, xml.etree.ElementTree as ET

with gzip.open("proteinatlas.xml.gz", "rb") as fh:
    for _, elem in ET.iterparse(fh, events=("end",)):
        if elem.tag == "entry":
            ...
            elem.clear()