thymus/SOURCES.md

# `thymus/` — sources

The central-tolerance reference: what a developing T-cell repertoire was negatively selected
against. All three files are **experimental** (mass spectrometry / antibody-based expression); the
harmonisation into this schema, and every FDR cutoff, is computed.

Human and mouse are separate files rather than one two-species table, matching the `_mmu` / `_hsa`
convention in `neoantigens/`. Every consumer already filters on `mhc_species`, and the human deposit
is an LFS object all of them read.

## `thymus_immunopeptidome.tsv.gz` — 53,878 rows

Thymus-eluted self peptides, **HLA Ligand Atlas** (`dataset_origin = hla_ligand_atlas_thymus`).

| `mhc_class` | rows |
|---|--:|
| MHCII | 27,987 |
| MHCI | 25,891 |

Schema: `peptide`, `mhc_a`, `mhc_class`, `mhc_species`, `source_protein`, `species`,
`dataset_origin`. Human only. `mhc_a` is empty throughout — the atlas gives donor-level candidate
alleles, not a per-peptide deconvolution.

**A hit here is a tolerance argument and an autoimmunity flag, and those are the same fact read two
ways**: a neoantigen resembling a thymically presented peptide had its reactive clones plausibly
deleted (lower expected immunogenicity), and a vaccine built on it risks cross-reacting with self.
`mhcmatch.mimics` keeps it as its own category for exactly that reason and never sums it with the
foreign ones — see `mhcmatch.mimics.KINDS`.

## `thymus_immunopeptidome_mmu.tsv.gz` — the mouse counterpart

Same schema, column for column. `mhc_species = species = MusMusculus` throughout. Added 2026-08-21;
this file is what the `## Not here` block of this document used to say was impossible.

**8,151 rows / 6,791 distinct peptides** as deposited here.

| `dataset_origin` | `mhc_class` | `mhc_a` | peptides |
|---|---|---|--:|
| `massive_msv000087031_thymus` | MHCII | `I-Ab` | 1,488 |
| `pride_pxd008733_thymus` | MHCI | `H-2Db` | 1,574 |
| `pride_pxd008733_thymus` | MHCI | `H-2Kb` | 1,089 |
| `pride_pxd008733_thymus` | MHCII | `I-Ab` | 2 |
| `pride_pxd031966_thymus` | MHCI | *(empty)* | 1,694 |
| `pride_pxd042241_tec` | MHCI | *(empty)* | 2,304 |

1,490 class-II and 5,301 class-I distinct peptides. 148 rows carry an empty `source_protein`.

MHC-I lengths run 8–11 with a median of 9. The two MHC-II rows are all that survive the 11–25 window
from PXD008733's 16 thymic anti-I-A/I-E PSMs — that immunoprecipitation returned mostly 9–10mers,
which is the class-I-length contamination the Nanaware paper is about. 121 rows carry an empty
`source_protein`; they are PXD042241 peptides from non-canonical ORFs that are not in UP000000589.

Unlike the human file, `mhc_a` is populated where the deposit records an immunoprecipitating
antibody, in the `pmhc/`+`ligandome/` spelling `H-2Kb` / `H-2Db` / `I-Ab` — **not** the `H2-Kb` form
found in the contaminated CEDAR rows of `neoantigens/neoag_tested.tsv.gz`. It is as-reported
attribution; nothing is predicted, and it is empty where the immunoprecipitation was pan-H-2.

### `pride_pxd008733_thymus` — whole thymus, H-2Kb and H-2Db

The thymus slice of the murine MHC-I tissue atlas; the other 18 tissues are in
`ligandome/tissue_self_mmu.tsv.gz`, which carries the full provenance and the FDR derivation.

> Schuster H, Shao W, Weiss T, Pedrioli PGA, Roth P, Weller M, Campbell DS, Deutsch EW, Moritz RL,
> Planz O, Rammensee HG, Aebersold R, Caron E.
> **A tissue-based draft map of the murine MHC class I immunopeptidome.**
> *Sci Data* 2018;5:180157.
> PMID [30084848](https://pubmed.ncbi.nlm.nih.gov/30084848/) ·
> doi:[10.1038/sdata.2018.157](https://doi.org/10.1038/sdata.2018.157)

Taken from the deposit's own Trans-Proteomic Pipeline PSM tables, not its raw spectra. Peptide-level
FDR recomputed from the deposit's `DECOY_` entries at 1%: `probability_ip` ≥ 0.974142, 35,470 target
against 354 decoy peptides.

### `pride_pxd042241_tec` — sorted cortical and medullary thymic epithelial cells

The cells that actually induce central tolerance, which whole thymus dilutes.

> Larouche JD, Laumont CM, Trofimov A, Vincent K, Hesnard L, Brochu S, Côté C, Humeau JF,
> Bonneil É, Lanoix J, Durette C, Gendron P, Laverdure JP, Richie ER, Lemieux S, Thibault P,
> Perreault C.
> **Transposable elements regulate thymus development and function.**
> *eLife* 2024;12:e91037.
> PMID [38635416](https://pubmed.ncbi.nlm.nih.gov/38635416/) ·
> doi:[10.7554/eLife.91037](https://doi.org/10.7554/eLife.91037)

`mhc_a` is empty: the deposit does not name the immunoprecipitating antibody.

**Its FDR is not recomputable, and the cutoff is a stated convention.** The mzid declares
`no threshold` at protocol level and carries no decoy entries, and its `passThreshold` attribute is
not a quality filter — it passes PSMs down to −10lgP = 5.00 while rejecting others up to 35.62.
The retained set is rank-1 PSMs at **PEAKS −10lgP ≥ 20**. That cutoff is doing real work: 8–11mers
rise from 67.2% to 82.0% of distinct peptides and 9-mers from 28.0% to 38.1%. It is **not** a
measured 1% FDR and must not be described as one.

Source proteins in this deposit are Ensembl and transposable-element identifiers from a personalized
mTEC/cTEC database, so peptides derived from non-canonical ORFs carry an empty `source_protein`
after re-mapping.

### `pride_pxd031966_thymus` — NOD mouse, whole thymus

The NOD-mouse MHC-I deposit covers thymus *and* pancreas, and **its run labels name neither**:
`NOD2_3`, `NOD4`, `NOD5`. PMID 36210013 is closed-access with no PMC copy, and the deposited tables
give only 30–50 peptides per run, far too few to tell the two apart. This was recorded here as an
open labelling problem, and its peptides were deliberately withheld rather than guessed.

The re-search settled it. Searching the nine raw files against the whole mouse proteome (rather than
the deposit's own predicted-binder FASTA) gives 1,700–5,800 peptides per run — enough for the source
proteins to speak:

| run | thymus-exclusive proteins | acinar zymogen peptides | call |
|---|---|--:|---|
| `NOD2_3` | **`Prss16` ×3, `Psmb11` ×1** | 4 | **thymus** |
| `NOD4` | none | 160 | pancreas |
| `NOD5` | none | 145 | pancreas |

**The call is made on presence, not on a ratio.** `Prss16` is the thymus-specific serine protease and
`Psmb11` the thymoproteasome subunit β5t, confined to cortical thymic epithelium; neither has a
tissue of expression outside thymus in the mouse, so a peptide from either cannot have been eluted
from pancreas. The converse holds for the exocrine zymogen program. A ratio is the wrong instrument
and would have got this wrong — on a 3× rule `NOD2_3` reads "unresolved" at 11 thymic against 4
pancreatic markers, while carrying three `Prss16` peptides and a `Psmb11` one.

An earlier, looser marker set inflated the thymic signal in *both* pancreas runs, because it counted
`Krt5`/`Krt8`/`Krt14` — thymic-epithelium markers in a thymus context but generic epithelial keratins
elsewhere — and `Lck`/`Zap70`/`Tcf7`, which mark the T-cell infiltrate that is the defining lesion of
the NOD pancreas. Those genes are now explicitly excluded; the rule is in
`src/assign_pxd031966_tissue.py`.

Only `NOD2_3` is deposited here. `NOD4` and `NOD5` are in `ligandome/tissue_self_mmu.tsv.gz` as
pancreas.

> Wang L, Li X, Yang S, Chen X, Li J, Wang S, Zhang M, Zheng Z, Zhou J, Wang L, Wu Y.
> **Proteomic identification of MHC class I-associated peptidome derived from non-obese diabetic
> mouse thymus and pancreas.**
> *J Proteomics* 2022;270:104746.
> PMID [36210013](https://pubmed.ncbi.nlm.nih.gov/36210013/) ·
> doi:[10.1016/j.jprot.2022.104746](https://doi.org/10.1016/j.jprot.2022.104746)

`mhc_a` is empty: the immunoprecipitation used the pan-H-2 antibody M1/42.3.9.8, which resolves no
allele. NOD is H-2^g7 (K^d, D^b), so these peptides are **not** C57BL/6 K^b/D^b ligands — read
`dataset_origin` before pooling them with the PXD008733 rows.

**Re-searched, not taken as deposited.** The authors searched a predicted-MHC-binder FASTA at 5% PSM
FDR; a database of peptides someone already predicted to bind can only return peptides someone
already predicted to bind. This is nf-core/mhcquant 3.2.0 against UP000000589 at 1% peptide-level
FDR.

### `massive_msv000087031_thymus` — I-Ab, the MHC class II arm

**1,488 distinct peptides — the only mouse class-II thymic ligands in the compendium.** No
identifications were deposited, so this is a re-search of the raw spectra with nf-core/mhcquant
3.2.0: unspecific cleavage, Comet, MS²Rescore (DeepLC + MS²PIP `Immuno-HCD`), Percolator, 1%
peptide-level FDR, against UP000000589 plus contaminants.

**Checked as class-II ligands rather than assumed.** Median length **16 aa** (mean 16.0, range
11–25), matching the ~16 aa thymic peak the source paper reports. **49% of the peptides (733 of
1,488) share a nested core with another peptide in the set** — ragged N- and C-termini around a
common binding core is the defining signature of MHC class-II processing and does not arise for
class I, where the termini are trimmed to the closed groove.

**The deposit's 11–25 window removes a peak the paper itself flags.** That paper is titled for this
problem: the thymic I-Ab length distribution is bimodal, with a second mode near 10 aa that is
class-I-length material co-purifying with a limiting-abundance class-II preparation. The window
applied here (`mhcmatch.mimics._LEN["mhc2"]`) drops it by construction — 1,303 rows fell outside the
per-class windows across all arms. That is a filter, not a measurement, and it is why the
distribution here is unimodal where the paper's is not.

> Nanaware PP, Jurewicz MM, Clement CC, Lu L, Santambrogio L, Stern LJ.
> **Distinguishing Signal From Noise in Immunopeptidome Studies of Limiting-Abundance Biological
> Samples: Peptides Presented by I-A^b in C57BL/6 Mouse Thymus.**
> *Front Immunol* 2021;12:658601.
> PMID [33995376](https://pubmed.ncbi.nlm.nih.gov/33995376/) ·
> doi:[10.3389/fimmu.2021.658601](https://doi.org/10.3389/fimmu.2021.658601)

Twelve of the deposit's 37 runs — the four thymus samples (`120116WTII`, `020617WT_MHCII`,
`061417MHCII_WT1`, `061417MHCII_WT2`) in three technical replicates each — of which **eleven were
usable**. `061417MHCII_WT1_1.raw` is unreadable: it carries the size MassIVE declares
(757,890,890 bytes) and a valid Thermo `01a1` "Finnigan" header, but ThermoRawFileParser fails on it
with `RAW file cannot be processed`, so the damage is mid-file. The other eleven convert cleanly.
It costs one of three *technical* replicates of Thymus 3 — not a sample and not a condition. The splenic B-cell and
dendritic-cell arms and the human LCL runs are deliberately **not** included: this is a thymic
reference, and pooling peripheral APCs into it would destroy the distinction the file exists to make.

## `thymus_expression_mmu.tsv.gz` — 15,306 rows

The mouse side of the table below, added 2026-08-21. Schema `gene`, `species`, `evidence`, `value`;
`species = MusMusculus` throughout, and unlike the human file `value` is populated — it carries the
measurement, so a caller can threshold rather than take mere presence as evidence.

| `evidence` | rows | what it is |
|---|--:|---|
| `fantom5_mouse_thymus_tpm` | 11,862 | adult-thymus CAGE median TPM > 0 (E-MTAB-3579) |
| `geiger_silac_thymus_ppb` | 3,444 | thymus SILAC protein abundance > 0 (E-PROT-11) |

12,646 distinct genes. The two lines of evidence are kept in a column rather than pooled, exactly as
in the human file: one says the gene is transcribed in thymus, the other that its protein was
detected there. Full provenance, citations and caveats — including why `Aire` reads zero — are in
`expression/SOURCES.md`.

## `thymus_expression.tsv.gz` — 10,493 rows

Thymically expressed self, by gene. Schema: `gene`, `species`, `evidence`, `value`.

| `evidence` | rows | what it is |
|---|--:|---|
| `hpa_thymus_ntpm` | 16,279 | **HPA thymus nTPM > 0 — the measured value**, added 2026-08-21 |
| `hla_ligand_atlas_thymus_source` | 10,367 | source proteins of the peptides above |
| `hpa_thymus_cluster` | 126 | Human Protein Atlas thymus-cluster genes |

26,772 rows. The evidence lines are kept in a column rather than pooled, because they are three
different statements: a gene is *transcribed* in thymus, a peptide from its protein was *eluted*
from thymus, and it is thymus-*enriched* by cluster assignment. They support each other; they are
not interchangeable.

**Why the `hpa_thymus_ntpm` line was added.** Before it, this table was not quantitative: 10,367 of
its 10,493 rows carried an **empty `value`** — they record which proteins peptides came from, which
is not an expression measurement — and the only rows with content held a text cluster label. The
mouse table had numbers and the human one did not. Note also that **GTEx has no thymus at all**, so
the compendium's main safety read cannot answer thymic questions; `expression/SOURCES.md` records
what was searched and why HPA is the source that works.

## Caveats on the mouse file

- **It is not the human file's equal in size, and cannot be.** The human deposit is one large
  multi-donor atlas; the mouse side is assembled from three smaller studies with different
  antibodies, instruments and search pipelines. Read the `dataset_origin` column before pooling.
- **Three different FDR regimes are present**, one per `dataset_origin` — a recomputed 1%
  peptide-level FDR (PXD008733), a stated PEAKS score convention (PXD042241), and mhcquant's 1%
  peptide-level FDR (MSV000087031). They are not interchangeable; the column is the handle.
- `source_protein` is **re-derived** by exact lookup against `proteome/mouse.fasta.gz`, not taken
  from the deposits, because the three name their proteins in three incompatible ways. I and L are
  not collapsed — MS cannot distinguish them, so an I/L-collapsed match would assert an
  identification the data does not carry.
- Lengths are clipped to `mhcmatch.mimics._LEN` (MHC-I 8–11, MHC-II 11–25) because every consumer
  drops the rest silently.

## Not here

**The mouse thymus expression gap is closed** — see `thymus_expression_mmu.tsv.gz` above. What
is still absent is a *promiscuous-expression* (AIRE-driven TRA) table: the Sansom mTEC-TRA gene list
(GSE53111) is journal-supplement-only and the mouse thymus proteome (PXD007288) is a 12 GB MaxQuant
archive. Bulk data cannot substitute — `Aire` itself reads 0 TPM in whole thymus because it is
confined to a rare mTEC subset, so the promiscuous-expression question needs sorted mTEC, not this
file. **Absent, not overlooked.**

**PXD031966 is now here, and the tissue labels were settled by measurement.** See the
`pride_pxd031966_thymus` section above.

## Re-fetch / regenerate

    cd ~/vcs/projects/2026-mouse-thymus
    # results-only deposits (671 MB); verify every file against PRIDE's SHA-1 `checksum` field
    aria2c -i src/urls_results.txt -d ~/hf/pmhc_data/raw/thymus_mouse -x4 -j3 --continue
    ./src/ingest_pxd008733.py
    ./src/ingest_pxd042241.py
    # MSV000087031 raw spectra + re-search, on aldan3
    aldan3 slurm submit src/fetch_raw.sbatch -- <srcdir> <destdir>
    aldan3 slurm submit src/mhcquant.sbatch -- mhc2 /projects/hla_epitope/mouse_thymus
    ./src/build_thymus_mmu.py