Dataset: Human Protein Atlas - the protein-level and single-cell check that GTEx cannot give
also called HPA access plan · protein atlas pull · immune cell specificity
GTEx tells us how much of a gene's message a tissue makes. The Human Protein Atlas goes further: it stains actual tissue to show where the protein really is, and it breaks blood down into individual cell types. That matters when the tissue we must protect is bone marrow.
The same thing, for a scientist
HPA complements GTEx on three axes phase 2 needs: antibody-based protein localisation in tissue, single-cell RNA consensus across tissues, and - most importantly here - blood-cell-type specificity, which resolves a whole-blood TPM value into the immune and progenitor populations that would actually be exposed. The whole gene record is served as JSON from a predictable URL with no authentication, which makes it the easiest resource in this beat to automate. The limitation that must be stated with every use is that HPA's tissue immunohistochemistry rests on antibody specificity, which varies by antibody and is not uniformly validated.
What question it answers
For each candidate: is the protein (not just the transcript) present in the tissues at risk? Is expression enriched in any blood cell type, and specifically in haematopoietic progenitors? What subcellular compartment is it in?
The project's existing gene entries already carry HPA fields - for example KAT7 and PSIP1 are both recorded as 'Low tissue specificity / Detected in all / Low immune cell specificity'. Phase 2 needs the finer cut: the single-cell blood breakdown, not the summary string.
How to reach it
Append .json to the Ensembl gene page: https://www.proteinatlas.org/ENSG00000133895.json returns the full record, anonymously, no key. Verified this cycle for MEN1, which returns gene name, synonyms, UniProt accession, chromosome and position, protein class, biological process and molecular function fields 1.
For bulk work, HPA publishes downloadable TSV files of the normal-tissue, blood-atlas and single-cell datasets under its download page. HPA data is released under a CC BY-SA licence, so redistribution inside this graph requires attribution - which is a licensing note, not an access barrier.
The access barrier, and the real limitation
Access barrier: none. No registration, no agreement, no key.
The real limitation is antibody validation. An HPA immunohistochemistry panel is only as good as the antibody, and HPA grades its own reliability from 'enhanced' to 'uncertain'. Phase-2 use must read the reliability score, not just the staining. For a target where the reliability is 'uncertain', HPA protein data should not be used to support a safety argument.
A second limitation: like GTEx, HPA's normal tissue is overwhelmingly adult.
| resource | Human Protein Atlas (tissue, blood atlas, single cell, pathology) |
|---|---|
| question answered | Is the protein present where the transcript is, and which blood cell types express it - the resolution GTEx whole-blood TPM averages away? |
| access barrier | None. The limitation is antibody validation quality, graded by HPA itself; read the reliability score before using protein-level data in a safety argument. |
| priority use | Resolve whole-blood expression into progenitor versus mature populations for ENL (40.3 TPM whole blood) and menin, since the G3 window gate is about haematopoietic progenitors specifically. |
| status | planned |
| compute | laptop; no GPU |
access
{
"auth": "none",
"per_gene_json": "https://www.proteinatlas.org/<ENSG>.json",
"verified_example": "https://www.proteinatlas.org/ENSG00000133895.json",
"licence": "CC BY-SA, attribution required for redistribution"
}fields to pull per candidate
[ "RNA tissue specificity", "RNA tissue distribution", "RNA blood cell specificity", "single cell type specificity", "subcellular location", "antibody reliability score" ]
- An antibody-based absence is not an absence. Treating a negative HPA stain as evidence of safety is a known way to be wrong about on-target toxicity.
- HPA normal tissue is adult; in a disease of infants, the most relevant tissue is the least represented.
- Does HPA's blood atlas resolve CD34+ progenitors specifically, or only mature immune populations - and if the latter, what resource covers progenitors?
- What is the antibody reliability grade for each of the nine candidates, and does any safety argument in this beat rest on an 'uncertain' antibody?
- Is ENL's high whole-blood expression driven by a mature population we can afford to lose, or by progenitors we cannot?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database
The Human Protein Atlas serves a complete per-gene JSON record at a predictable URL with no authentication, verified this cycle for MEN1 (ENSG00000133895), returning UniProt O00255, chromosome 11, position 64803510-64811294, and its protein-class and molecular-function annotations.
Human Protein Atlas gene record, MEN1 (ENSG00000133895.json) · Human Protein Atlas 2026What this evidence cannot show
Fetched directly this cycle. The JSON record is a summary; tissue immunohistochemistry images and their reliability grades must be read from the web record, and antibody-based protein data is only as reliable as the antibody. TIGHTENED 2026-09-19: a complete JSON record is not a validated measurement. HPA's protein layer is antibody-based with documented field-wide specificity problems, and its RNA layer predicts protein abundance only after a gene-specific RNA-to-protein factor that spans several orders of magnitude.
Evidence that argues against this (2)
- The HPA's protein layer is antibody-based, and antibody specificity is a documented, field-wide problem: Baker's Nature analysis of the reproducibility crisis in antibody reagents reports widespread cross-reactivity and failure of commercial antibodies to recognise their stated target. An HPA immunohistochemistry call for a target in normal marrow is a reagent result before it is a biological one, and should not be used as a toxicity filter without checking the antibody's validation tier and the orthogonal RNA evidence. Baker M 2015
- The RNA and protein layers of the atlas disagree in a gene-specific way. Edfors et al. measured absolute protein copy numbers across human tissues and cell lines by targeted proteomics and found that transcript and protein levels do not correlate well unless a gene-specific RNA-to-protein conversion factor is introduced, with that factor ranging from a few hundred to several hundred thousand protein copies per mRNA. Reading an HPA RNA value as a proxy for protein abundance is therefore wrong by an unknown, gene-specific factor - which matters most for exactly the low-expressed targets a safety filter is trying to clear. Edfors F et al. 2016
executes the access route that dataset audit described, and records the substring-matching trap it did not
Both normal-tissue filters fail for the same reason in different layers: HPA's protein calls are antibody-limited and its RNA layer does not predict protein gene-by-gene (PMID 27951527), while GTEx has no marrow and no paediatric donors.
needed to resolve ENL's 40 TPM whole-blood signal into progenitor versus mature populations