immunogenicity/SOURCES.md

# `immunogenicity/` — sources, provenance, filtering and a complete audit

General immunogenicity corpora: **is this presented peptide seen by a T cell?** Neoantigen datasets
moved to [`../neoantigens/`](../neoantigens/SOURCES.md) — this directory no longer holds NCI,
`neoag_*` or any tumour-specific screen.

Every citation resolved against **PubMed**; PMIDs and DOIs given. Audit regenerated by
`bench/neoag/audit_pmhc_data.py` in `2026-mhcmatch-benchmark`; corpora rebuilt by
`bench/neoag/corpus_chowell.py` + `bench/neoag/deposit_corpora.py`.

Written 2026-08-17.

---

## Complete audit

Measured from the files, not copied from any prior description.

| file | rows | distinct peptides | pos | neg | prevalence | HLA alleles | length |
|---|--:|--:|--:|--:|--:|--:|:--:|
| `chowell_rebuilt.tsv.gz` | 511,301 | 315,191 | 19,866 | 491,435 | 0.0389 | 210 | 8–11 |
| `chowell_rebuilt_hla_matched.tsv.gz` | 115,592 | 99,763 | 19,866 | 95,726 | 0.1719 | 183 | 8–11 |
| `kesmir_rebuilt.tsv.gz` | 65,737 | 36,155 | 22,613 | 43,124 | 0.3440 | 196 | 8–11 |
| `kesmir_rebuilt_hla_matched.tsv.gz` | 59,941 | 35,202 | 22,613 | 37,328 | 0.3773 | 188 | 8–11 |
| `iedb_labeled.tsv.gz` | 790,868 | 516,092 | 34,362 | 756,506 | 0.0435 | 250 | 2–50 |
| `hla_pop_freqs.tsv.gz` | 7,171 | — | — | — | — | 4,234 | — |

No donor/patient identifier exists in any file here — these are peptide-level corpora aggregated
over studies, not per-donor cohorts. (Donor counts are meaningful in `../neoantigens/` and are
reported there.)

---

## `chowell_rebuilt.tsv.gz` — immunogenic vs presented **self**

Chowell's construction, rebuilt from the current IEDB dump rather than read off the 2015
spreadsheet, which is copyrighted supplementary material and frozen at 2015.

> Chowell D, Krishna S, Becker PD, Cocita C, Shu J, Tan X, Greenberg PD, Klavinskis LS,
> Blattman JN, Anderson KS. **TCR contact residue hydrophobicity is a hallmark of immunogenic CD8+
> T cell epitopes.** *Proc Natl Acad Sci USA* 2015;112(14):E1754–62.
> PMID [25831525](https://pubmed.ncbi.nlm.nih.gov/25831525/) ·
> doi:[10.1073/pnas.1500973112](https://doi.org/10.1073/pnas.1500973112)

### Filtering procedure, in order

1. **Source** — `iedb_labeled.tsv.gz` (see below), itself an IEDB export.
2. **Class** — `mhc_class == "MHCI"`.
3. **Host** — `host_species ∈ {human, mouse}`; kept as a column and **never pooled**. The two hosts
   have different MHC and different thymic repertoires, so a fit across them without the split is
   fitting a mixture.
4. **Peptide** — uppercased, whitespace-stripped; length **8–11**; canonical 20 amino acids only.
5. **Exclusions** — `SLLMWITQV`, `KLGGALQAK` (NY-ESO-1 / MAGE-A3 epitopes that appear on both sides
   of the literature with contradictory annotation).
6. **Positives** — any peptide with a positive T-cell assay (`dataset_origin == "iedb_tcell"`).
7. **Negatives** — eluted ligands whose `source_species` equals the **host** (self), plus the HLA
   Ligand Atlas thymus immunopeptidome for the human arm.
8. **Label is a property of the (peptide, host), not the row.** 1,075 peptides are both eluted and
   positively assayed. Elution is evidence of *presentation*; it is never evidence against a T-cell
   response. Assigning per row makes the same sequence positive under one allele and negative under
   another — it sits on both sides of the fit. Computing the positive set once per (peptide, host)
   and applying it everywhere raised self-sourced positives by **57%**.
9. **Aggregation** — one row per `(peptide, allele, host_species)`, label = max.

### Not controlled, deliberately

**Allele and length.** Immunogenic and eluted sets genuinely differ in both: some HLAs are
protective and some are not, and immunogenic peptides have a real length distribution that differs
from bulk ligands. Matching those away removes effects, not confounds. `allele` and `length` ship
as columns so a stratified analysis is a filter, not a rebuild — and the HLA-matched variant below
exists for when allele composition *is* the thing to remove.

---

## `kesmir_rebuilt.tsv.gz` — immunogenic vs presented **non-self**

The Calis/Kesmir contrast. Same shape, same rules, one difference: negatives are eluted ligands
from a source organism that is **not** the host.

> Calis JJA, Maybeno M, Greenbaum JA, Weiskopf D, De Silva AD, Sette A, Keşmir C, Peters B.
> **Properties of MHC class I presented peptides that enhance immunogenicity.**
> *PLoS Comput Biol* 2013;9(10):e1003266.
> PMID [24204222](https://pubmed.ncbi.nlm.nih.gov/24204222/) ·
> doi:[10.1371/journal.pcbi.1003266](https://doi.org/10.1371/journal.pcbi.1003266)

**This arm requires `../ligandome/viral_foreign_iedb.tsv.gz`, and is wrong without it.** Every
eluted-ligand negative in the IEDB table itself is self-sourced — 449,449 human-self and 41,986
mouse-self against 149 + 63 cross-species, and zero from any other organism. Filtering
`source_species != "human"` on a **mouse** host therefore selects mouse *self* ligands and rebuilds
Chowell under another name. The foreign ligandome supplies 41,294 genuine human non-self negatives.

Kesmir removes the self/non-self axis that Chowell confounds with immunogenicity, and is the harder
problem: prevalence 0.344 against Chowell's 0.039, and every negative is a foreign peptide that was
presented and still not attacked.

---

## The HLA-matched variants

`*_rebuilt_hla_matched.tsv.gz`, built by `deposit_corpora.py --ratio 5 --seed 20260817`.

There are 5–25× more negatives than positives and the two sides do not share an allele
distribution, so a classifier can separate them on **allele composition alone**. For each allele,
negatives are drawn without replacement to **5 per positive of that same allele**:

* alleles with no positives contribute **no** negatives — keeping them would reintroduce exactly the
  imbalance being removed;
* alleles with too few negatives contribute all they have, and the shortfall stays visible in the
  per-allele counts rather than being topped up from another allele;
* the draw is seeded (`20260817`), so the file is reproducible rather than one sample.

Effect: Chowell 491,435 → 95,726 negatives, 210 → 183 alleles, prevalence 0.039 → 0.172. Kesmir
43,124 → 37,328, 196 → 189 alleles.

Use the matched file when the question is about peptide sequence; use the full file when allele
composition is part of the signal you want. Neither is "the" corpus.

---

## Two residue-level artefacts, measured

Both are properties of how the data was generated. A model will fit either as if it were biology.

| corpus | Cys % positives | Cys % negatives | ratio | C-term R/K % pos | C-term R/K % neg |
|---|--:|--:|--:|--:|--:|
| `chowell_rebuilt` | 10.69 | **1.55** | **6.9×** | 9.21 | 9.50 |
| `kesmir_rebuilt` | 10.94 | 14.50 | 0.8× | 9.91 | **20.33** |
| `iedb_labeled` | 14.56 | **2.52** | **5.8×** | 10.90 | 13.98 |

### Cysteine — an MS detection artefact, and it hits Chowell hardest

Cysteine is systematically **under-detected in mass-spectrometry immunopeptidomics** unless samples
are alkylated. Where positives are T-cell-assayed (synthetic peptides, no MS step) and negatives are
MS-eluted, the corpus carries a Cys gradient with no biological content. Chowell shows **6.9×**;
finer-grained, the thymus MS subset runs at 0.17% against 11.59% in the assayed positives.

Fitted freely on the Chowell corpus, Cys took the **single largest coefficient** in a position-role
naive Bayes (+1.84 anchor / +2.05 TCR-facing). Masking it costs grouped-CV AUROC 0.712 → 0.690 and
*improves* transfer to neoantigen cohorts. The shipped `mhcmatch.posbayes` tables zero it.

**`kesmir_rebuilt` does not have this problem** (0.8×) — its negatives are IEDB-curated foreign
ligands rather than a single MS pipeline. That makes it the safer corpus for fitting residue-level
terms, and is a reason to run both.

### C-terminal Arg/Lys — genuine anchor preference, or tryptic carry-over

A basic C-terminus is the real anchor preference of **HLA-A\*03, A\*11, A\*31, A\*68 and B\*27** —
and Arg/Lys are exactly the residues **trypsin cleaves after**. An elevated rate can be either, and
the two are **not separable from sequence alone**.

`kesmir_rebuilt` negatives run at **20.33%** against 9.91% in positives — a 2× gradient in the
corpus that is otherwise the cleaner of the two. Before attributing anything to a C-terminal basic
residue on this corpus, stratify by allele: if the effect survives within A\*03/A\*11/B\*27 and
within non-basic-preferring alleles separately, it is anchor preference; if it only appears pooled,
it is composition or contamination.

---

## `iedb_labeled.tsv.gz`

The IEDB export both corpora are built from. **This is also the `ipred` training set**, so any
`ipred` figure quoted on it is in-sample.

790,868 rows · 516,092 distinct peptides · 34,362 immunogenic · 250 alleles · lengths 2–50
(unfiltered — the 8–11 restriction is applied downstream).

> Vita R, Mahajan S, Overton JA, Dhanda SK, Martini S, Cantrell JR, Wheeler DK, Sette A, Peters B.
> **The Immune Epitope Database (IEDB): 2018 update.** *Nucleic Acids Res* 2019;47(D1):D339–D343.
> PMID [30357391](https://pubmed.ncbi.nlm.nih.gov/30357391/) ·
> doi:[10.1093/nar/gky1006](https://doi.org/10.1093/nar/gky1006)

Provenance: **experimental** — curated T-cell assay and MHC-ligand records. Raw export kept at
`raw/immunogenicity/iedb_tcell_assays_29032026.tsv.gz`.

Thymus negatives come from the HLA Ligand Atlas (`thymus/thymus_immunopeptidome.tsv.gz`); its
`mhc_a` is null by design — the deposit is not allele-deconvolved.

---

## `hla_pop_freqs.tsv.gz`

Class-I allele frequencies (A/B/C) for EUR/ASN/AFR populations, 7,171 rows over 4,234 distinct
alleles. Source: Allele Frequency Net Database (AFND), <http://www.allelefrequencies.net>.
Provenance: **experimental** (population genotyping), aggregated.

---

## Regenerate

    # in ~/vcs/projects/2026-mhcmatch-benchmark
    python bench/neoag/corpus_chowell.py                 # plain build (Chowell arm)
    python bench/neoag/corpus_chowell.py --with-viral    # + foreign ligandome (Kesmir arm)
    python bench/neoag/deposit_corpora.py --ratio 5 --seed 20260817
    python bench/neoag/audit_pmhc_data.py

`corpus_chowell.parquet` must keep its exact row counts (human 464,310 / 14,712 pos; mouse 47,203 /
5,154) — the shipped `mhcmatch.posbayes` tables are fitted on it, and folding extra sources into it
in place breaks that provenance silently. That is why `--with-viral` writes a separate file.

## `chowell_iedb_full.tsv.gz` / `chowell_iedb_full_matched.tsv.gz` (2026-08-18)

Rebuilt from the **full** IEDB export (build 2026-08-11) rather than `iedb_labeled.tsv.gz`. That
harmonised table kept only *positive* T-cell assays, so its `assay_type` is perfectly collinear with
its label and it cannot supply a measured negative at all; the full export carries 361,962 negative
T-cell assay rows, for mouse as well as human.

`chowell_rebuilt.tsv.gz` and `kesmir_rebuilt.tsv.gz` are **unchanged and still current** — the
shipped `mhcmatch.complement` and `posbayes` tables are fitted on them, they are named in the
library's `REFERENCE_FILES`, and every AUROC in `bench/results/complementarity.md` belongs to them.
These two files are additional, not replacements.

### The rules, in the order applied

1. MHC class I.
2. Host ∈ {human, mouse}; a column, never pooled.
3. Length 8–11, canonical 20 amino acids.
4. Exclude `SLLMWITQV`, `KLGGALQAK`.
5. **Immunogenic** — the peptide has ≥1 positive T-cell assay in that host (`Positive`,
   `Positive-Low`, `-High`, `-Intermediate`).
6. **Non-immunogenic** — an eluted **self** ligand appearing in no positive T-cell assay in any
   host. 2,086 rows are eluted but positive elsewhere and are excluded from the negatives rather
   than counted as such.
7. The label is a property of `(peptide, host)`, never of an assay row.
8. The restriction is resolved, and **no row is dropped for lacking one**: 1,677,439 of 3,278,355
   surviving class-I rows name only `HLA class I`. netMHCpan-4.2 picks the best `%Rank_EL` among
   whatever the recorded string still permits — an allele group or serotype restricts to that
   group, an H-2 haplotype to the molecules it carries, a bare `HLA class I` to a 30-allele
   population panel built per locus. A **known** non-human/non-mouse MHC on a human or mouse host
   (2,812 rows) is a curation error, not a missing value, and is excluded rather than imputed.
9. One row per `(peptide, allele group, host)`, label = max. The key is the allele **group**
   (`HLA-A*02`), which is what lets serotype-only records join.

### Counts

| file | rows | positive | negative | peptides | allele groups |
|---|--:|--:|--:|--:|--:|
| `chowell_iedb_full` (human) | 789,047 | 24,293 | 764,754 | 579,322 | 75 |
| `chowell_iedb_full` (mouse) | 65,472 | 7,187 | 58,285 | 49,253 | 43 |
| `chowell_iedb_full_matched` (human) | 20,360 | 10,180 | 10,180 | 17,181 | 18 |
| `chowell_iedb_full_matched` (mouse) | 9,132 | 4,566 | 4,566 | 7,969 | 25 |

Columns add `allele_2f`, `allele_group`, `allele_source` ∈ {`reported`, `imputed_group`,
`imputed_panel`} and `allele_rank_el`, so an analysis can exclude imputed restrictions and say how
much sample it lost. Human: 375,462 reported against 413,585 imputed; mouse 51,003 against 14,469.

The matched arm resamples both classes to one profile at 1:1 so the allele group carries no signal
about the label — human to population frequency from `hla_pop_freqs.tsv.gz`, mouse to the positives'
own profile, there being no H-2 frequency table. It is **small**: 52 of 75 human allele groups are
too scarce to fill their population weight and are dropped, which is the cost of the matching, not a
bug.

Regenerate, in `~/vcs/projects/2026-mhcmatch-benchmark`:

    python immunogenicity/fetch.py            # stage the export from ~/hf/pmhc_data/dump
    bench/neoag/corpus_iedb.py --workers 12   # ~55 min, 14.7M netMHCpan predictions
    bench/neoag/deposit_arms.py

Arm counts, the filter cascade and the selection tree are `bench/results/corpus_arms.md`.

**Caveat carried forward.** These negatives are still *inferred* — an eluted ligand nobody has
tested is assumed non-immunogenic. 38,322 rows in the rebuild have a measured negative T-cell assay
instead, held under evidence `tcell_negative` and deliberately **not** merged into the arm: swapping
the negative definition changes what every AUROC recorded on these corpora means.

### `immunogenicity_legacy_arms.tsv.gz` (2026-08-19)

The published deposits and the legacy Kešmir arms in one table, keyed `corpus` ∈ {`chowell_vanilla`,
`kesmir_S1`, `kesmir_S2`, `kesmir_S3`}, so a held-out evaluation can be reproduced without the
copyrighted spreadsheets. 15,991 rows.

**Species is keyed on the restricting allele, not on the deposit's own `Species` column.** The two
disagree on 777 of the 2,508 Calis rows (31%), which carry `Species = Mus` with an HLA restriction —
HLA-transgenic mice, murine host and human restricting molecule. Both readings describe the
experiment; they build different corpora. Keyed on the host, the Calis human arm is 1,113
immunogenic against **one** non-immunogenic; keyed on the allele it is **1,619 against 272** and the
mouse arm is H-2 throughout. The reported host is retained as `host_reported`.

Regenerate with `bench/neoag/corpus_legacy.py` then `deposit_arms.py`.

### What the recognition model is fitted on

`chowell_iedb_full` is the training arm for `mhcmatch.recognition` (0.15.0). Two choices were made
by measurement and are recorded in `bench/results/recognition_model.md` and
`feature_importance.md`:

- **The unmatched arm ships.** Resampling negatives to match population HLA usage costs 0.019
  (human) and 0.058 (mouse) AUROC on held-out published deposits.
- **ESM2 and the physicochemical features are both kept.** ESM2 adds +0.022/+0.042 over physchem
  alone; physchem adds +0.009/+0.019 over ESM2 alone. Permutation importance on held-out data ranks
  the two ESM face-pools first (0.19–0.27), then `kf_tcr`; the whole-peptide Kidera aggregate is
  worth 0.0165/0.0012 and is not shipped, being exactly the sum of the two role columns.

## `staged/` — the IEDB full export, as staged (2026-08-19)

The **root of the provenance chain**. Every rebuilt corpus in this directory
(`chowell_iedb_full*`, `immunogenicity_legacy_arms`) is derived from these two tables, and until
now they lived only in a gitignored working directory — one bad checkout from being unreproducible.
Deposited so the chain closes without re-downloading 450 MB from IEDB.

| file | provenance | rows |
|---|---|--:|
| `staged/iedb_tcell.parquet` | **experimental** — IEDB `tcell_full_v3_tsv.zip`, build 2026-08-11, downloaded 2026-08-18 | 570,434 linear-peptide rows, 233,636 distinct peptides |
| `staged/iedb_ligand.parquet` | **experimental** — IEDB `mhc_ligand_full_tsv.zip`, same build | 5,773,495 linear-peptide rows, 1,704,996 distinct peptides |
| `staged/fetch.py` | the script that produced both | — |
| `staged/FETCH.md` | its record: zip mtimes, export row counts, alignment failures, class and qualitative-measurement breakdowns | — |

Ten columns are kept from each export; nothing is filtered beyond `Linear peptide` and a non-empty
name, because length, alphabet, class and host filters are corpus-construction rules and belong
where they can be stated as rules. **0 rows failed to align** in either table — IEDB ships two
header lines and embedded newlines in comment fields, so `fetch.py` reads with `csv.reader` rather
than splitting on tabs, and reports the misalignment count rather than assuming it is zero.

The T-cell export is what makes a *measured* negative possible: 361,962 `Negative` rows across all
hosts, where the older harmonised `iedb_labeled.tsv.gz` keeps only positive T-cell assays and its
`assay_type` is therefore perfectly collinear with its label.

Regenerate (needs `~/hf/pmhc_data/dump`, which is gitignored — the raw zips are not deposited):

```zsh
python immunogenicity/staged/fetch.py
```

## `corpus_iedb_mhc2.parquet` — the MHC class-II arm (2026-08-19)

**Derived/computed**, from `staged/iedb_tcell.parquet` and `staged/iedb_ligand.parquet` — the same
IEDB build 2026-08-11 export the class-I corpus is built from, by the same rules, so the two arms are
comparable by construction rather than by assertion. This is what the shipped class-II
complementarity model (`mhcmatch.complement`, `cls="mhc2"`, v0.16.0) is fitted on.

| field | value |
|---|--:|
| rows (peptide, allele group, host) | 1,096,034 |
| distinct peptides | 643,744 |
| immunogenic peptides | 77,943 |
| human rows | 1,036,041 |
| human positives | 65,486 |
| human distinct peptides | 603,781 |
| human allele groups | 121 |
| mouse rows | 59,993 |
| mouse positives | 12,457 |
| mouse distinct peptides | 50,258 |
| mouse allele groups | 40 |
| peptide length | 11–25, median 15 |

Three rules differ from the class-I arm and nothing else does:

- **Length** is 11–25 rather than 8–11.
- **Restriction is parsed, never imputed.** Class-II restriction is written a dozen ways across IEDB
  (`HLA-DRB1*15:01`, `HLA-DPA1*01:03/DPB1*04:01`, `HLA-DR15`, serotypes, `H2-IAb`), so the corpus
  carries a parser and keeps whatever it cannot resolve as an explicit `unresolved` group rather than
  guessing an allele. It parses **1,003,554 / 1,974,413 rows (50.8 %)**; the remainder are retained,
  not dropped, because the complementarity fit reads only peptide and label. There is therefore **one
  class-II arm where class I has two** — no HLA-matched variant is built, since matching needs a
  resolved allele on both sides.
- **Rule 8** (pathogen epitopes) is applied by parsing the source organism rather than by set-match.

`evidence` records why each row is what it is: `eluted_but_positive_elsewhere=5,305`,
`eluted_foreign=103,351`, `eluted_self=1,156,574`, `eluted_unknown=440,458`, `tcell_negative=110,312`,
`tcell_positive=158,413`.

`corpus_iedb_mhc2.log` is the generator's own run record, deposited beside it; the counts in the table
above are read from it. The build is deterministic — regenerated 2026-08-19 from a clean checkout, every
count reproduced exactly.

Regenerate (needs this mirror at `$MHCMATCH_PMHC_DIR`, and the benchmark repo):

```zsh
python bench/neoag/corpus_iedb.py --cls mhc2
```