# `expression/` — sources

Two human tables. `reference_expression.tsv.gz` holds three references that are **never merged** — different measurements, different units, and
a gene-keyed table and a peptide-keyed one are not the same object.

| | |
|---|---|
| file | `reference_expression.tsv.gz` |
| rows | 6,681,814 |
| schema | `key`, `key_type`, `source`, `context`, `median_tpm`, `q25_tpm`, `q75_tpm`, `n` |
| provenance | **experimental** (all three sources are measured RNA-seq); the per-context median/IQR/`n` are **computed** summaries of it |

| `source` | `key_type` | rows | contexts | what a row is |
|---|---|--:|--:|---|
| `gtex` | `gene` | 3,955,284 | 53 | a gene's expression across normal tissue — the **safety** read |
| `hpa_consensus` | `gene` | 1,025,751 | 51 | the same read over HPA's tissue panel — **the only one that contains thymus** |
| `tcga` | `peptide` | 1,700,779 | 19 | a mutant peptide's expression by tumour type — the **target** read |

Total 6,681,814 rows.

`key_type` is the join key and it differs by source on purpose: GTEx and HPA answer "is this gene on
in normal tissue", TCGA answers "is this neoantigen transcribed in this tumour". A caller that treats
them as one column is asking a question none of them answers. GTEx and HPA are both gene-keyed but
are still not interchangeable — see the units caveat in the HPA section below.

Read through `mhcmatch.expression` (`lookup`, `tissues`, `tumor_types`, `safety_profile`), which
keeps them apart. Fetched by `mhcmatch bootstrap --reference` (~115 MB).

---

## `reference_expression_toil.tsv.gz` — GTEx and TCGA on one pipeline, in one unit

Added 2026-08-26. **This is the table to use when a normal tissue and a tumour type have to be
compared, or when one supplies a scale for the other.**

| | |
|---|---|
| rows | **5,037,966** |
| schema | `key`, `key_type`, `source`, `context`, `median_tpm`, `q25_tpm`, `q75_tpm`, `n`, `unit` |
| keys | 58,581 gene symbols (`key_type = gene` on every row) |
| `source` / contexts | `toil_gtex` — 53 GTEx tissues, 3,104,793 rows · `toil_tcga` — 33 TCGA studies, 1,933,173 rows |
| `unit` | `TPM` on every row |
| size | 38.6 MB |
| provenance | **experimental** — RSEM TPM from RNA-seq; the per-context median/IQR/`n` are **computed** summaries |
| built by | `bench/neoag/expression_toil.py` in `2026-mhcmatch-benchmark` |

**Why it exists.** `reference_expression.tsv.gz` carries GTEx in **TPM** and TCGA in **RSEM
upper-quartile normalised counts**, both written into a column called `median_tpm`. They are each
correct for their own question, but they cannot be put on one axis: a scale taken from one and
divided into an abundance from the other is a units error no caller can detect, because the column
name says TPM in both cases. This table removes the problem at the source rather than by converting
one into the other — a conversion that is not available from the deposited matrix, since RSEM
normalised counts are proportional to abundance x effective length and the effective lengths are not
published with it.

**Origin.** The UCSC Xena / Toil recompute, which put GTEx, TCGA and TARGET through a **single RSEM
pipeline against one annotation (GENCODE v23)** and publishes gene-level TPM for all of them, so a
GTEx tissue median and a TCGA study median differ only in the samples that went in.

| input | file | bytes |
|---|---|--:|
| expression matrix | `TcgaTargetGtex_rsem_gene_tpm.gz` | 1,323,254,426 |
| sample phenotype | `TcgaTargetGTEX_phenotype.txt.gz` | 135,753 |
| gene annotation | `probeMap/gencode.v23.annotation.gene.probemap` | 3,244,244 |

```
base=https://toil-xena-hub.s3.us-east-1.amazonaws.com/download
curl -L -o xena/TcgaTargetGtex_rsem_gene_tpm.gz  $base/TcgaTargetGtex_rsem_gene_tpm.gz
curl -L -o xena/TcgaTargetGTEX_phenotype.txt.gz  $base/TcgaTargetGTEX_phenotype.txt.gz
curl -L -o xena/gencode.v23.annotation.gene.probemap \\
     "$base/probeMap%2Fgencode.v23.annotation.gene.probemap"
python bench/neoag/expression_toil.py --out expression/reference_expression_toil.parquet
```

> Vivian J, Rao AA, Nothaft FA, Ketchum C, Armstrong J, Novak A, Pfeil J, Narkizian J, Deran AD,
> Musselman-Brown A, Schmidt H, Amstutz P, Craft B, Goldman M, Rosenbloom K, Cline M, O'Connor B,
> Hanna M, Birger C, Kent WJ, Patterson DA, Joseph AD, Zhu J, Zaranek S, Getz G, Haussler D,
> Paten B. **Toil enables reproducible, open source, big biomedical data analyses.**
> *Nat Biotechnol* 2017;35(4):314-316.
> PMID [28398314](https://pubmed.ncbi.nlm.nih.gov/28398314/) ·
> PMC5546205 · doi:[10.1038/nbt.3772](https://doi.org/10.1038/nbt.3772)

> Goldman MJ, Craft B, Hastie M, Repecka K, McDade F, Kamath A, Banerjee A, Luo Y, Rogers D,
> Brooks AN, Zhu J, Haussler D. **Visualizing and interpreting cancer genomics data via the Xena
> platform.** *Nat Biotechnol* 2020;38(6):675-678.
> PMID [32444850](https://pubmed.ncbi.nlm.nih.gov/32444850/) ·
> PMC7386072 · doi:[10.1038/s41587-020-0546-8](https://doi.org/10.1038/s41587-020-0546-8)

**The TCGA study code is joined through the barcode, never through the disease name.** Xena labels a
TCGA sample with a disease name (`Uveal Melanoma`); the rest of this compendium keys on the study
abbreviation (`UVM`). A hand-written name->code crosswalk can be silently wrong, so the code is taken
from PanCanAtlas `merged_sample_quality_annotations.tsv` on the 15-character sample barcode, which is
exact. **17,474 of 19,131 matrix columns are used**: 9,616 TCGA tumour aliquots and 7,862 GTEx
samples. TARGET (734 columns) is dropped -- paediatric, and none of its diseases has a matched normal
here -- as are TCGA adjacent normals and aliquots the source flags `Do_not_use`.

**Upstream values are `log2(TPM + 0.001)` and are inverted on build**, so what is stored is TPM.

**Caveats.**
- **The GTEx contexts are Toil-era, not v11.** `reference_expression.tsv.gz` is GTEx v11; this is the
  earlier release Toil reprocessed, so the two GTEx halves are neither the same samples nor the same
  depth. They are also not spelled the same: Toil writes `Skin - Sun Exposed (Lower Leg)` and GTEx
  v11 writes `Skin - Sun Exposed (Lower leg)`. **A context string cannot be carried from one table
  to the other without checking.**
- **Three tissues are thin and are kept with their `n` rather than dropped**: Bladder (9),
  Cervix - Ectocervix (6), Fallopian Tube (5). They are the matched normals of BLCA, OV and CESC, so
  dropping them would remove the safety read for three tumour types. `Cervix - Endocervix` has fewer
  than five samples and is absent; CESC still resolves through Ectocervix. Every TCGA study clears 36.
- **353 of 58,581 symbols carry more than one Ensembl gene id** (0.6 %). The row kept per
  (symbol, context) is the one with the **highest median**, not a sum across loci: a lookup asks
  whether a gene is on in this context, and the dominant locus answers it without inventing a total
  across paralogues. 164,862 duplicate rows were collapsed this way.
- `n` is a **sample** count here, as in the GTEx half, and not the transcript count the mouse table's
  `n` is.

**What a floor built from it looks like**, as the 25th percentile of non-zero gene medians:
TCGA runs 0.1200 (KICH) to 0.3700 (LAML) TPM, a **3.08x** spread over 33 studies; GTEx runs 0.1000
(Whole Blood) to 0.4000 (Testis), **4.00x** over 53 tissues. Pooled, TCGA is 0.1800 TPM over 882,959
non-zero gene rows and GTEx 0.2700 over 1,402,705. **A tumour's own floor is consistently below its
matched normal's** -- SKCM 0.1600 against skin 0.3050 (0.52x), BLCA 0.1700 against bladder 0.3600
(0.47x), LUAD 0.2000 against lung 0.3500 (0.57x), BRCA 0.1900 against breast 0.3000 (0.63x).

**Format: parquet, and the record is not in the bootstrap set.** As gzipped TSV this table was
70.5 MB and 0.5 s to read with polars (5.2 s with the standard library); dictionary-encoded and
zstd-compressed it is **38.6 MB and 0.04 s**. On a link measured here at 154 kB/s that is three and
a half minutes of download saved, which is the difference between a bootstrap that finishes and one
that looks like it has hung. `mhcmatch bootstrap --reference` stages only the three companions
below — 6.6 MB between them, carrying the same numbers — and never this file.

**Label arrays in the `.npz` are fixed-width unicode, not `dtype=object`.** An object array is
pickled, and `np.load` needs `allow_pickle=True` to read it back; a pickled array in a file fetched
over the network executes on load. The deposit is written so that flag is never needed.

### Its three companions, and why scoring never reads the table itself

Added 2026-08-26 beside it. `reference_expression_toil.tsv.gz` is the record; these are what a
program actually opens.

| file | size | what it is |
|---|--:|---|
| `toil_matrix.npz` | 6,632,665 B | the whole table as a dense **58,581 gene x 86 context float32 matrix**, contexts keyed `"<source>|<context>"` |
| `toil_floors.tsv` | 8,365 B | 88 rows -- each context's abundance floor at q05, q10, q25, plus a pooled row per cohort |
| `context_synonyms.tsv` | 29,334 B | 448 rows, 184 aliases -- how a free-text origin becomes a context |

**The matrix exists so a lookup does not cost a whole table.** Answering *what is this gene in this
tumour type* from the TSV means building a five-million-entry dictionary of dictionaries: measured,
**5.20 s and 3,168 MB resident**. The matrix is **0.05 s and 29 MB** -- 100x the speed on 1/109 of
the memory -- and every lookup a resolution chain makes is one array index. Both return the same
floor to four decimals, which is how the matrix is checked against the table it came from. A gene absent from a context is `0.0` and not `nan`: every gene here was
measured in every context that cleared the sample threshold, so absence is silence, and silence is a
measurement. Whether a *gene* is in the reference at all is a separate question and is reported
separately -- silence and ignorance must not collapse into one number.

**`context_synonyms.tsv` is derived, not hand-written.** Long format, one target per row, so it
sorts and joins like data rather than holding lists in cells:

| alias_kind | where the aliases come from | rows |
|---|---|--:|
| `code` | the 33 TCGA studies in the matrix | 83 |
| `site` | Xena `_primary_site`, joined to the PanCanAtlas study code on the 15-character barcode | 218 |
| `disease` | Xena `detailed_category`, joined the same way | 83 |
| `gtex` | the 53 GTEx contexts in the matrix | 64 |

**One alias resolves to two things and each is its own row**, which is why the alias count and the
row count differ: 150 rows carry `target_kind = tcga_code` and 298 carry `gtex_context`, so
`hepatocellular` names both `LIHC` and `Liver`. 152 of the 448 rows are `derived:*` and 296 are
`curated:*`, and the `provenance` column says which on every row. The curated majority is the
matched-normal half and the adjectival forms a free-text origin actually arrives in
(`cervical`, `pancreatic`, `uterine corpus`); the derived half is what the matrix and the barcode
join supply on their own. Hand-writing this would have been wrong in the places nobody checks: **an organ is
more than one study more often than not** -- `Lung` is `LUAD` *and* `LUSC`, `Kidney` is three -- and a
crosswalk that picks one is silently wrong for the rest. The tumour-to-matched-normal half is carried
from `mhcmatch.expression.TUMOR_TISSUE` against the spellings this matrix uses, because Toil writes
`Brain - Frontal Cortex (Ba9)` where GTEx v11 writes `(BA9)` and an exact match drops it.

An alias that does not resolve **raises**. Falling back to the pooled reference would return a
plausible number computed from the wrong distribution with nothing downstream able to tell.

**A note on what these floors are for.** `c` is a property of a transcriptome, and the numerator is
one candidate's abundance. Where the numerator is not in reference TPM the two stop cancelling, and
the obvious repair -- a median of ratios against the reference -- **must not be estimated from a
candidate list**. A mutation reaches such a list only if it was called, and only where the gene was
seen in RNA, so the ratio measures that conditioning instead of the library: on three screens whose
columns are all deposited as TPM it returns 1.78, 2.18 and 3.15, never below 1. `mhcmatch` gates it
on coverage of the reference transcriptome for that reason, and a candidate list cannot clear it.

---

## Mouse — `reference_expression_mmu.tsv.gz` and `protein_abundance_mmu.tsv.gz`

Added 2026-08-21. The compendium's normal-tissue safety read was GTEx, and GTEx is human, so a mouse
epitope had no reference to be scored against. These two files are the mouse counterpart.

**They are two files, not one, and that is deliberate.** The human table's value column is literally
`median_tpm`. RNA and protein are different measurements in different units, and writing a SILAC
abundance into a `_tpm` column would be a units lie no caller could detect.

### `reference_expression_mmu.tsv.gz` — the GTEx analogue

| | |
|---|---|
| rows | **659,050** |
| schema | `key`, `key_type`, `source`, `context`, `median_tpm`, `q25_tpm`, `q75_tpm`, `n` — **column for column identical to the human file** |
| keys | 18,830 gene symbols (`key_type = gene`) |
| contexts | **35 adult tissues** |
| `source` | `fantom5_mouse` |
| provenance | **experimental** — CAGE tag counts; the per-gene median across transcripts and the quantile summary are **computed** |

**Origin.** EBI Expression Atlas [E-MTAB-3579](https://www.ebi.ac.uk/gxa/experiments/E-MTAB-3579),
the RIKEN FANTOM5 mouse CAGE atlas. `q25_tpm` / `median_tpm` / `q75_tpm` are taken from the Atlas'
own five-number summary per gene × group; they are not re-derived here.

> Lizio M, Harshbarger J, Shimoji H, Severin J, Kasukawa T, Sahin S, Abugessaisa I, Fukuda S,
> Hori F, Ishikawa-Kato S, Mungall CJ, Arner E, Baillie JK, Bertin N, Bono H, de Hoon M, Diehl AD,
> Dimont E, Freeman TC, Fujieda K, Hide W, Kaliyaperumal R, Katayama T, Lassmann T, Meehan TF,
> Nishikata K, Ono H, Rehli M, Sandelin A, Schultes EA, 't Hoen PAC, Tatum Z, Thompson M,
> Toyoda T, Wright DW, Daub CO, Itoh M, Carninci P, Hayashizaki Y, Forrest ARR, Kawaji H.
> **Gateways to the FANTOM5 promoter level mammalian expression atlas.**
> *Genome Biol* 2015;16:22.
> PMID [25723102](https://pubmed.ncbi.nlm.nih.gov/25723102/) ·
> doi:[10.1186/s13059-014-0560-6](https://doi.org/10.1186/s13059-014-0560-6)

**Only plain `adult` groups are kept** — 35 of the deposit's 170. GTEx is adult normal tissue;
folding an E14 embryo, a pregnancy timepoint or a lactating mammary gland into a "is this gene on in
normal tissue" safety read answers a different question. The other stages are in the source and are
recoverable by re-running without the filter.

**A parsing trap worth recording.** Atlas group labels are two `;`-separated parts and **the order
is not fixed**: 136 of the 170 are stage-first (`adult; thymus`) and 32 are tissue-first
(`adrenal gland; adult`). Taking either position blindly mislabels a third of the table — the first
build of this file produced contexts named `adult` and `whole body` that way. The stage is matched by
vocabulary wherever it sits, and the tissue is whatever remains.

**Checked against known biology, not assumed.** Of 35 tissues, thymus ranks **first** for `Prss16`
(31 TPM), `Psmb11` (88), `Dntt` (961), `Rag1`, `Cd8a` (131) and `Lck` (12) — thymic epithelium and
thymocyte genes. As a negative control `Cela1` peaks in pancreas at 225,010 TPM against 0.3 in
thymus.

**Caveats.**
- **`n` is 1 for every row.** FANTOM5 provides one library per adult tissue, so `q25`/`q75` describe
  the spread across transcripts of a gene, **not across animals**. GTEx `n` runs to several hundred
  donors. Do not read the two `n` columns as the same quantity.
- **`Aire` is 0 TPM in thymus here, and that is a detection floor, not biology.** Aire is expressed
  in a rare mTEC subset; bulk CAGE at this depth cannot see it. Any promiscuous-expression question
  needs sorted mTEC data, not this file.
- CAGE measures 5′-capped tag density. It is reported as TPM and behaves like expression, but it is
  not RNA-seq TPM and the two are not numerically interchangeable across sources.

### `tumor_expression_mmu.tsv.gz` — the mouse TUMOUR read, and why it is a separate file

Added 2026-09-04. `reference_expression_mmu.tsv.gz` answers the *safety* half of the expression
question -- is this gene on in normal tissue -- and there was no counterpart for the *ranking* half.
Mouse has no TCGA, so a mouse neoantigen's own abundance fell back to the gene's normal-tissue
median, which is the number `expr_norm` already carries. Measured on the 923 class-I rows of the
IEDB mouse neoantigen deposit, `expr_lvl` and `expr_norm` came out **identical on 100 %** of the
rows lacking a deposited abundance: two fitted terms on one column.

| | |
|---|---|
| rows | **149,640** |
| schema | `key`, `key_type`, `source`, `context`, `median_tpm`, `q25_tpm`, `q75_tpm`, `n` — **column for column identical to the two files above** |
| keys | 24,940 gene symbols (`key_type = gene`) |
| contexts | **6 syngeneic tumour models**: B16F10, CT26, E0771, LLC, MC38, Panc02 |
| `source` | `gse245293_syngeneic` |
| `n` | **3 on every row** — three biological replicates per model, so `q25`/`q75` are a spread across *animals*, unlike the FANTOM5 file's spread across transcripts |
| provenance | **experimental** — RNA-seq counts; the length normalisation, the per-model quantiles and the gene-level aggregation are **computed** |

**Origin.** NCBI GEO [GSE245293](https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE245293),
supplementary file `GSE245293_Normalized_count_table_tumor.txt.gz` (2.9 MB), staged at
`raw/expression_mouse/`. 23 samples in the series; the 18 `*_tumor_rep*` columns are used and the
8 CAF (cancer-associated fibroblast) columns are **not** -- those are stroma, not the tumour
transcriptome, and folding them in would answer a different question.

    curl -O https://ftp.ncbi.nlm.nih.gov/geo/series/GSE245nnn/GSE245293/suppl/GSE245293_Normalized_count_table_tumor.txt.gz
    python src/build_tumor_expression_mmu.py     # in ~/vcs/projects/public_expression_peptidomes

> Comparison of syngeneic tumor models reveals different tumor microenvironments with distinct
> subsets of cancer-associated fibroblasts polarized by infiltrating T cells. GEO GSE245293,
> deposited 2023-10-13, released 2024-01-29. *Mus musculus*, expression profiling by high
> throughput sequencing.

**Units are computed and the file says so.** The deposit is **normalised counts** with a `Length`
column, not TPM. Length-normalising each sample and rescaling to a million gives a TPM-like
quantity comparable *within* this table; it is **not** interchangeable with FANTOM5 CAGE TPM or
with GTEx RSEM TPM. `mhcmatch.rank.expr_level` is unit-free when its floor is a quantile of the
same column, which is exactly why a floor for these contexts must be taken from this file.

**Checked against known biology, not assumed.** Three independent melanocytic markers separate the
melanoma line by three orders of magnitude: `Tyr` **398.9**, `Dct` **4,611.8** and `Pmel`
**4,636.7** TPM in B16F10 against ≤ 4.3 in all five other models. They would all move together if
the sample-to-model mapping were wrong.

**Colon markers were tried and do not work here; do not re-try them.** `Muc2` reads ~0 in every
model -- CT26 is an undifferentiated colon carcinoma and makes no goblet-cell mucin -- and
`Cdx1`/`Cdx2` do not separate CT26 and MC38 from the rest (Cdx2 is highest in Panc02, 8.8 against
CT26's 1.4). The expectations were wrong, not the data.

**Coverage of the IEDB mouse neoantigen deposit**, joining on model × gene symbol: **347 of 968
class-I rows (36 %)** and **304 of 522 class-II rows (58 %)**. The class-II figure is the reason
this file exists -- that arm previously had a deposited abundance on 13 % of its rows.

**Two models it does not carry** that the neoantigen deposit names often: `P815` (89 rows) and the
`T3`/`CMS5`/`Meth A` fibrosarcomas (150 rows between them). Those still take the pooled floor.

### `PRJNA1271699` — located, not fetched

[PRJNA1271699](https://www.ncbi.nlm.nih.gov/bioproject/PRJNA1271699), *"Mouse syngeneic model
RNAseq sequencing"* (CrownBio), submitted 2025-06-04: **120 runs over 12 models × 10 replicates**
-- A20, B16BL6, B16F10, CT26, EMT6, H22, Hepa16, LL2, MC38, Pan02, RM1, Renca -- paired-end
Illumina NovaSeq 6000, **418 GB** of raw reads. It would cover Renca and EMT6, which nothing here
does.

**NCBI carries raw reads only; there is no processed matrix.** The BioProject abstract advertises
an interactive resource at `https://muscreen.crownbio.com/syngeneic_omics`, and **as of 2026-09-04
that URL is dead** -- `/syngeneic_omics` returns 404 and the host root returns a bare Apache
default page (403). So the processed tables the abstract promises are not currently retrievable,
and using this project means downloading 418 GB and running an alignment pipeline. Recorded here
so the cost is known rather than rediscovered.

**This is the source to re-process if these models are wanted here** -- it is the one that covers
P815, Renca and EMT6, which `tumor_expression_mmu.tsv.gz` does not.

    # metadata only, no reads:
    curl -sS "https://trace.ncbi.nlm.nih.gov/Traces/sra-db-be/runinfo?acc=PRJNA1271699"

### `toil_matrix_mmu.npz` + `toil_floors_mmu.tsv` — the harmonised mouse compendium

Added 2026-09-04. The mouse counterpart of `toil_matrix.npz` / `toil_floors.tsv`, same array names
and the same `source|context` key shape, so code written against the human pair reads this one.

| | |
|---|---|
| shape | **26,737 genes x 68 contexts**, dense float32 |
| `genes` | official NCBI symbol, synonyms resolved via Ensembl release-100 (= GENCODE M25) and NCBI Gene |
| `contexts` | `fantom5_mouse\|<tissue>` x 35 normal, `gse245293\|<model>` x 6 tumour, `prjna1271699+\|<model>` x 27 tumour |
| `values` | linear TPM-like, **harmonised across sources**; `NaN` = the gene is absent from that source, which is not the same as measured at zero |
| `n_samples` | 3 for GSE245293 (replicate **mean**), 1 for FANTOM5, 0 where the source reports none |
| `toil_floors_mmu.tsv` | 71 rows -- each context's floor at q05/q10/q25, plus a pooled row per source, same seven columns as the human file |

**Sources.** FANTOM5 mouse CAGE (EBI E-MTAB-3579); GEO **GSE245293** (6 syngeneic models x 3
replicates); and the CrownBio syngeneic panel, labelled `prjna1271699+` for NCBI BioProject
**PRJNA1271699** plus the further models that panel carries beyond the 12 the accession lists.
GSE245293 wins the five models both cover. The 27 filled models include **P815, Renca and EMT6**,
which no other deposit here reaches.

**Harmonised, and here is what that means.** The three sources sat on three scales -- pooled mean
`log2(1 + TPM)` of 1.721, 2.513 and 2.059. Each source's non-zero values are quantile-mapped onto
the mean of the three sources' quantile functions; **per source, not per context**, so
between-context structure survives. Zero maps to zero. The map is monotone and its grid spans each
source's support, so **every within-context rank is preserved exactly** -- which is what
`mhcmatch.rank.expr_level` consumes. Non-zero spread falls 0.577 -> 0.087 log2 units, and the
pooled q25 floor agrees to 0.714--0.717 across all three.

Two things it does not do. It cannot correct a **gene-specific** platform effect: `Actb` reads 3,032
in B16F10 and 10.4 in FANTOM5 thymus after harmonisation, because CAGE at that promoter genuinely
ranks it mid-distribution. And it cannot correct the **detection-rate** difference (65.2 / 78.1 /
65.0 % non-zero), because a monotone map that respects zero cannot move it.

**Absolute values are no longer any one source's own number.** That is the trade: comparability
across all 68 contexts, at the cost of each source's internal scale. The unharmonised per-source
tables are not deposited; `tumor_expression_mmu.tsv.gz` and `reference_expression_mmu.tsv.gz` above
remain the unharmonised reads of two of the three sources.

Built by `src/build_deposit_mmu.py` in `~/vcs/projects/public_expression_peptidomes`; the protocol
and its numbers are `results/expression_harmonization.md` there.

**For comparison, the human pair needed none of this.** Same measurement on `toil_matrix.npz`:
`toil_gtex` and `toil_tcga` pooled means differ by **0.011** log2 units, against 0.565 for the
mouse CAGE-vs-RNA-seq seam. Both human halves went through one Toil RSEM recompute, and it worked.

### `protein_abundance_mmu.tsv.gz` — the protein-level read

| | |
|---|---|
| rows | **133,848** |
| schema | `key`, `key_type`, `source`, `context`, `abundance`, `unit`, `n` |
| keys | 5,148 gene symbols |
| contexts | **26 tissues**, including thymus |
| `source` / `unit` | `geiger_silac_mouse` / `ppb_ibaq` |
| provenance | **experimental** — SILAC-normalised MS abundance |

**Origin.** EBI Expression Atlas [E-PROT-11](https://www.ebi.ac.uk/gxa/experiments/E-PROT-11).

> Geiger T, Velic A, Macek B, Lundberg E, Kampf C, Nagaraj N, Uhlen M, Cox J, Mann M.
> **Initial quantitative proteomic map of 28 mouse tissues using the SILAC mouse.**
> *Mol Cell Proteomics* 2013;12(6):1709–1722.
> PMID [23436904](https://pubmed.ncbi.nlm.nih.gov/23436904/) ·
> doi:[10.1074/mcp.M112.024919](https://doi.org/10.1074/mcp.M112.024919)

**Why keep a protein table at all.** A peptide can only be presented if the protein was translated,
so for immunopeptidome work protein abundance is the more direct evidence and RNA is a proxy. It is
much shallower — 5,148 genes against 18,830 — so it complements the RNA table rather than replacing
it.

### Why not E-PROT-13

[E-PROT-13](https://www.ebi.ac.uk/gxa/experiments/E-PROT-13) (Huttlin's mouse phosphorylation and
expression atlas) was the first candidate considered. It covers **9 tissues and does not include
thymus**, so it cannot serve either purpose here. Of the **65 mouse baseline experiments in
Expression Atlas, only six contain thymus**; E-MTAB-3579 (35 adult tissues) and E-PROT-11 (26) are
the only two of those with broad tissue coverage.

### Regenerate

    cd ~/vcs/projects/public_expression_peptidomes && ./src/build_expression_mmu.py

---

## Human — the `hpa_consensus` arm, and why GTEx alone was not enough

Added 2026-08-21, in the same pass that built the mouse references.

**GTEx has no thymus.** Its 53 contexts contain no thymus, no lymph node and no bone marrow — adult
thymus involutes and is not routinely sampled, so the normal-tissue safety read was blind to the one
tissue central tolerance actually happens in. Expression Atlas is no better: of its **123 human
baseline experiments only three carry any thymus group**, and all three are compromised for this
purpose — E-MTAB-3871 is fetal tissue from fetuses with congenital defects, E-MTAB-3827 is sorted
thymocyte subsets rather than tissue, and E-MTAB-9641 is thymic epithelium cultured on a 3T3-J2
feeder layer.

HPA samples thymus directly, so its consensus tissue set is what fills the hole.

| | |
|---|---|
| rows | **1,025,751** |
| keys | 20,151 gene symbols (`key_type = gene`) |
| contexts | **51 tissues, including thymus** |
| `source` | `hpa_consensus` |
| provenance | **experimental** — RNA-seq; the HPA consensus normalisation is **computed** |

**Origin.** [`rna_tissue_consensus.tsv.zip`](https://www.proteinatlas.org/download/tsv/rna_tissue_consensus.tsv.zip)
from the Human Protein Atlas.

> Uhlén M, Fagerberg L, Hallström BM, Lindskog C, Oksvold P, Mardinoglu A, Sivertsson Å, Kampf C,
> Sjöstedt E, Asplund A, Olsson I, Edlund K, Lundberg E, Navani S, Szigyarto CA, Odeberg J,
> Djureinovic D, Takanen JO, Hober S, Alm T, Edqvist PH, Berling H, Tegel H, Mulder J, Rockberg J,
> Nilsson P, Schwenk JM, Hamsten M, von Feilitzen K, Forsberg M, Persson L, Johansson F, Zwahlen M,
> von Heijne G, Nielsen J, Pontén F.
> **Tissue-based map of the human proteome.**
> *Science* 2015;347(6220):1260419.
> PMID [25613900](https://pubmed.ncbi.nlm.nih.gov/25613900/) ·
> doi:[10.1126/science.1260419](https://doi.org/10.1126/science.1260419)

**It is added as a third `source`, not merged into GTEx and not replacing it.** The schema was
designed with a `source` column for exactly this; every existing `gtex` and `tcga` row is untouched.

**Checked against known biology.** Of 51 tissues, thymus ranks **first** for `PRSS16` (770 nTPM),
`PSMB11` (28.5), `DNTT` (306), `CD1A` (296), `RAG1` (368), `CCL25` (718), `CD3E` (596) and `LCK`
(326). Negative controls behave: `ALB` peaks in liver at 198,524 nTPM against 0.3 in thymus, `INS` in
pancreas, `MYH7` in skeletal muscle.

**Caveats.**
- **nTPM is not GTEx TPM.** Both are TPM-family units but the normalisation differs, so compare
  *within* a `source`, not across. This is what the `source` column is for.
- `n` is 1 — HPA consensus publishes one value per gene × tissue, so `q25_tpm` and `q75_tpm` equal
  `median_tpm` and carry no spread. GTEx `n` runs to several hundred donors; the two `n` columns are
  not the same quantity.
- **`AIRE` is detectable here (1.2 nTPM, rank 2 of 51) but reads 0 in the mouse CAGE atlas.** That is
  a difference in assay depth, not in biology — do not read it as a species difference.