Split neoantigens from immunogenicity; add per-directory SOURCES with verified citations and audits

Directory split. immunogenicity/ now holds only general immunogenicity corpora; every neoantigen
dataset moved to neoantigens/, and nci_complete.tsv.gz was renamed nci_complete_minigenes.tsv.gz
because its unit is a minigene (25-mers, mhc_a blank) and the old name hid that.

New corpora, rebuilt from the IEDB dump so they can be shared and regenerated rather than
redistributing copyrighted supplementary files:

- chowell_rebuilt.tsv.gz 511,301 rows / 19,866 immunogenic / 210 alleles
- kesmir_rebuilt.tsv.gz 65,737 rows / 22,613 immunogenic / 196 alleles
- *_hla_matched.tsv.gz negatives down-sampled per allele to 5 per positive of that
allele, seed 20260817, so a classifier cannot separate the
classes on allele composition alone

Three SOURCES.md files with the full filtering procedure, provenance per column group
(experimental vs derived), and complete per-dataset audits: rows, distinct peptides, positives,
negatives, prevalence, HLA counts, donor counts and length ranges, all measured by
bench/neoag/audit_pmhc_data.py rather than copied from prior descriptions.

Two residue artefacts measured and written up, because a model fits either as if it were biology:
- Cysteine. MS immunopeptidomics under-detects it unless alkylated. chowell_rebuilt runs 10.69%
Cys in positives against 1.55% in negatives -- a 6.9x gradient with no biological content.
kesmir_rebuilt does NOT have it (0.8x), which makes it the safer corpus for residue-level terms.
The neoantigen screens have no gradient at all, so a Chowell-trained residue term meets an
inverted distribution there.
- C-terminal Arg/Lys. Genuine anchor preference for A*03/A*11/A*31/A*68/B*27, and also what trypsin
cleaves after -- not separable from sequence alone. kesmir_rebuilt negatives run 20.33% against
9.91% in positives; stratify by allele before attributing anything to it.

Citations resolved against PubMed with PMIDs and DOIs. One was caught wrong in the process: the
Gartner 2021 PMID was 34676374, which is Pearlman et al. "Targeting public neoantigens" -- a real,
resolvable identifier attached to the wrong paper. Corrected to 34927080. Sources that could not be
verified (CTpedia; the CEDAR/Neopep/Gfeller/ITSNdb/GBM/VACCIMEL/Bjerregaard cohort references) are
marked unresolved rather than guessed.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

DESCRIPTION.md

@@ -48,16 +48,30 @@ the mouse thymus proteome (PXD007288) is a 12 GB MaxQuant archive, and the mTEC-
 
 ## Public datasets (this repo)
 
+Per-directory provenance, filtering procedures, verified PubMed citations and complete
+per-dataset audits (rows, positives, negatives, HLA counts, donor counts, residue artefacts):
+[`immunogenicity/SOURCES.md`](immunogenicity/SOURCES.md) ·
+[`neoantigens/SOURCES.md`](neoantigens/SOURCES.md) ·
+[`ligandome/SOURCES.md`](ligandome/SOURCES.md).
+
+> The `#` column below is undocumented and does not consistently mean column count, row count or
+> dataset count. Do not rely on it; the SOURCES files carry measured counts.
+
+
 | Path | # | Dataset | One-line |
 |------|---|---------|----------|
 | `pmhc/pmhc_full.tsv.gz` | 6 | IEDB epitope–MHC | positive IEDB MHC-ligand binding records (human IMGT/HLA + mouse H-2) |
 | `pmhc/pmhc_shortlist.tsv.gz` | 6 | IEDB epitope–MHC (≥2 refs) | high-confidence pairs (≥2 references) |
+| `immunogenicity/chowell_rebuilt.tsv.gz` | — | Chowell corpus, rebuilt | 511,301 rows / 315,191 peptides, 19,866 immunogenic, 210 alleles, 8–11mers. Immunogenic vs **self** eluted ligands, rebuilt from the IEDB dump (Chowell et al., PNAS 2015, PMID 25831525). Carries a 6.9x cysteine artefact — see SOURCES.md |
+| `immunogenicity/chowell_rebuilt_hla_matched.tsv.gz` | — | …HLA-matched | 115,592 rows; negatives down-sampled per allele to 5 per positive of that allele, seed 20260817 |
+| `immunogenicity/kesmir_rebuilt.tsv.gz` | — | Kesmir/Calis corpus, rebuilt | 65,737 rows / 36,155 peptides, 22,613 immunogenic, 196 alleles. Immunogenic vs **non-self** eluted ligands (Calis et al., PLoS Comput Biol 2013, PMID 24204222). Requires `ligandome/viral_foreign_iedb.tsv.gz`; no cysteine artefact, but a 2x C-terminal Arg/Lys gradient |
+| `immunogenicity/kesmir_rebuilt_hla_matched.tsv.gz` | — | …HLA-matched | 59,941 rows, same matching rule |
 | `immunogenicity/iedb_labeled.tsv.gz` | 7 | IEDB immunogenicity — **ipred training set** | immunogenic (T-cell+, 34,362) vs non-immunogenic (MHC-ligand+ in healthy self, 756,506); `host_species` (human/mouse) is the cross-validation axis; ipred featurizes `peptide` (MJ/hydropathy/Kidera) |
-| `immunogenicity/neoag_tested.tsv.gz` | 9 | Neoantigens — tested (workhorse) | tested patient-derived neoantigens, immunogenicity 0/1 (TESLA, Neopep, CEDAR, Gfeller, ITSNdb, GBM, VACCIMEL, Bjerregaard). **Human only by label — see the mouse-contamination caveat below.** |
-| `immunogenicity/neoag_candidates.tsv.gz` | 9 | Neoantigens — candidates | untested neoantigen candidates (Neopep `not_tested`) |
-| `immunogenicity/neoag_tested_mmu.tsv.gz` | 12 | Neoantigens — tested, **mouse MHC** | 1,977 murine tested neo-epitopes (IEDB), immunogenicity 0/1 (740 immunogenic); `mhc_species=MusMusculus` |
-| `immunogenicity/neoag_tested_hsa.tsv.gz` | 12 | Neoantigens — tested, **human MHC in mouse models** | 717 rows from the same IEDB export whose restriction is HLA — HLA-transgenic mice plus human-host epitopes (KRAS G12D, CDK4 R24C, PMEL); `mhc_species=HomoSapiens` |
-| `immunogenicity/nci_complete.tsv.gz` | 9 | Neoantigens — complete NCI screen (variant) | 1,326 mutant 25-mer minigenes across 46 GI-cancer patients with CD8/CD4 T-cell screen outcomes (28 CD8-immunogenic) — a predictor-agnostic *screened* set with real negatives. Standalone variant, **minigene-level** (`peptide` = minigene, `mhc_a` blank — not epitope-deconvolved). Source: Koşaloğlu-Yalçın et al., iScience 2022, Table S6 ([doi:10.1016/j.isci.2022.103850](https://doi.org/10.1016/j.isci.2022.103850), PMID 35128348) |
+| `neoantigens/neoag_tested.tsv.gz` | 9 | Neoantigens — tested (workhorse) | tested patient-derived neoantigens, immunogenicity 0/1 (TESLA, Neopep, CEDAR, Gfeller, ITSNdb, GBM, VACCIMEL, Bjerregaard). **Human only by label — see the mouse-contamination caveat below.** |
+| `neoantigens/neoag_candidates.tsv.gz` | 9 | Neoantigens — candidates | untested neoantigen candidates (Neopep `not_tested`) |
+| `neoantigens/neoag_tested_mmu.tsv.gz` | 12 | Neoantigens — tested, **mouse MHC** | 1,977 murine tested neo-epitopes (IEDB), immunogenicity 0/1 (740 immunogenic); `mhc_species=MusMusculus` |
+| `neoantigens/neoag_tested_hsa.tsv.gz` | 12 | Neoantigens — tested, **human MHC in mouse models** | 717 rows from the same IEDB export whose restriction is HLA — HLA-transgenic mice plus human-host epitopes (KRAS G12D, CDK4 R24C, PMEL); `mhc_species=HomoSapiens` |
+| `neoantigens/nci_complete_minigenes.tsv.gz` | 9 | Neoantigens — complete NCI screen (variant) | 1,326 mutant 25-mer minigenes across 46 GI-cancer patients with CD8/CD4 T-cell screen outcomes (28 CD8-immunogenic) — a predictor-agnostic *screened* set with real negatives. Standalone variant, **minigene-level** (`peptide` = minigene, `mhc_a` blank — not epitope-deconvolved). Source: Koşaloğlu-Yalçın et al., iScience 2022, Table S6 ([doi:10.1016/j.isci.2022.103850](https://doi.org/10.1016/j.isci.2022.103850), PMID 35128348) |
 | `neoantigens/neoantigens_tested_peptides.tsv.gz` | 16 | Neoantigens — tested, **epitope-level** | 423,085 screened 8–12mers with their wild-type counterpart and restricting alleles, 178 CD8-immunogenic, 98 patients, over three datasets (`NCI` 420,786 · `HiTIDE` 1,563 · `TESLA` 736) and the authors' own `split`. This is the screen at the resolution a pMHC model is scored at — the complement to the minigene-level `nci_complete`. Harmonized from `raw/immunogenicity/NCI_dataset_only_tested.txt` by `bench/neoag/ingest_tested.py` (2026-mhcmatch-benchmark). Predicted quantities keep a `pred_` prefix; `peptide`, `peptide_wt`, `mhc_a` and `immunogenicity` are as-reported |
 | `neoantigens/nci_gartner_mmp.tsv.gz` | 45 | NCI — confirmed CD8+ minimal epitopes | 147 rows / 144 distinct minimal mutant peptides (mmp) with restricting HLA; 120 carry the wild-type mmp, so agretopicity is **measurable rather than predicted**. `split` preserves the paper's train/test. Source: Gartner et al., Nat Cancer 2021, Suppl. Tables 3 & 20 ([doi:10.1038/s43018-021-00197-6](https://doi.org/10.1038/s43018-021-00197-6)) |
 | `neoantigens/nci_gartner_nmers.tsv.gz` | 21 | NCI — full screened nmer universe | 9,541 screened mutations as WT/mutant 25-mers (144 CD8+), with gene, mutation type, RNA-seq presence and exome VAF decile — the negative universe at the unit the assay used. Same source, Suppl. Tables 11 & 12 |

immunogenicity/SOURCES.md

@@ -0,0 +1,198 @@
+# `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.

immunogenicity/chowell_rebuilt.tsv.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:585983d4c9f9e3400a2b1bb3cdc6fc08ca8145b4e7627d668bd4fbd216bde42a
+size 4432513

immunogenicity/chowell_rebuilt_hla_matched.tsv.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2411a25837bb531bcd7d76aaf81c2c3d87a5a14285947d0f04db805a670886ed
+size 928735

immunogenicity/kesmir_rebuilt.tsv.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf4dbaf44f0409109e21aa37502224c895ea29d668c31d711a7dd43b05a70c38
+size 575885

immunogenicity/kesmir_rebuilt_hla_matched.tsv.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c413096c1752caf14712700f896f49716c8d1cc8f941c0d832183c1338a47794
+size 501103

ligandome/SOURCES.md

@@ -0,0 +1,164 @@
+# `ligandome/` — sources, provenance and citations
+
+Three reference sets used as *background* rather than as labelled training data: what a foreign
+peptide looks like, what a novel viral ORF looks like, and which self genes are tumour-restricted
+enough to be targeted safely.
+
+Every citation below was resolved against **PubMed** and carries its PMID and DOI. Where a source
+could not be resolved, that is stated as unresolved rather than filled in with a plausible guess.
+
+Audited and rewritten 2026-08-17. The previous entries in `DESCRIPTION.md` were one-liners with no
+citations, and the row `#` was wrong for all three files.
+
+---
+
+## `viral_foreign_iedb.tsv.gz`
+
+Foreign (pathogen-derived) peptides observed presented on MHC — the non-self reference.
+
+| | |
+|---|---|
+| rows | **82,368** (44,993 distinct peptides) |
+| columns | `peptide`, `mhc_a`, `mhc_class`, `mhc_species`, `source_organism`, `dataset_origin` |
+| class | MHC-I 55,084 · MHC-II 19,762 · unlabelled 7,522 |
+| host | HomoSapiens 71,207 · MusMusculus 3,639 · unlabelled 7,522 |
+| source organisms | 909 distinct |
+| provenance | **experimental** — MS-eluted / assayed ligands as curated by IEDB |
+
+**Origin.** Export of the Immune Epitope Database restricted to non-human source organisms.
+
+> 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)
+
+**Why it matters here.** This is the only source of *non-self eluted ligands* we hold. Every
+eluted-ligand negative in `immunogenicity/iedb_labeled.tsv.gz` is self-sourced (449,449 human-self
+and 41,986 mouse-self against 149 + 63 cross-species), so a Calis/Kesmir-style corpus — immunogenic
+vs presented-**foreign**-but-not-immunogenic — cannot be built without this file.
+
+**Caveat.** 7,522 rows (9.1%) carry no `mhc_class` or `mhc_species`. Filter on those columns
+explicitly; do not assume MHC-I.
+
+---
+
+## `viral_orfs_gse272406.tsv.gz`
+
+Translated **unannotated** viral ORFs — proteins the canonical viral proteome does not contain.
+
+| | |
+|---|---|
+| rows | **5,673** ORFs |
+| columns | `orf_id`, `protein`, `virus`, `virus_family`, `gene`, `orf_type`, `dataset_origin` |
+| coverage | 365 viruses, 31 viral families |
+| ORF type | uORF 1,954 · mORF 1,304 · N_ext 779 · iORF 751 · ouORF 603 · N_trunc 282 |
+| provenance | **experimental** — ribosome-profiling-derived ORF calls, translated to protein |
+
+**Origin.** GEO series [GSE272406](https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE272406).
+
+> Weingarten-Gabbay S, Bauer MR, Stanton AC, Yu Y, Freije CA, Welch NL, Boehm CK, Klaeger S,
+> Verzani EK, López D, Hensley LE, Clauser KR, Carr SA, Abelin JG, Rice CM, Sabeti PC.
+> **Pan-viral ORFs discovery using massively parallel ribosome profiling.**
+> *Science* 2025;388(6752):1218–1224.
+> PMID [40504907](https://pubmed.ncbi.nlm.nih.gov/40504907/) ·
+> doi:[10.1126/science.ado6670](https://doi.org/10.1126/science.ado6670)
+
+**Correcting a prior mislabel.** An earlier spec filed this as thymus *self*. It is neither thymic
+nor self: these are viral ORFs, and the source paper reports peptides from noncanonical viral ORFs
+presented on class-I HLA in infected cells. Using it as a self reference would invert its meaning.
+
+---
+
+## `cancer_targets_tsarina.tsv.gz`
+
+Tumour-antigen genes with **tissue-restriction and safety annotation** — not merely a gene list.
+The previous description ("curated shared tumor-antigen genes", 3 columns) understated it by an
+order of magnitude.
+
+| | |
+|---|---|
+| rows | **393** genes, **51** columns |
+| restriction call | TESTIS 268 · SOMATIC 79 · REPRODUCTIVE 26 · PLACENTAL 11 · NO_DATA 9 |
+| confidence | MODERATE 167 · HIGH 152 · LOW 65 · NO_DATA 9 |
+| with immunopeptidomics evidence | 222 genes carry `ms_pmids` (178 distinct PMIDs) |
+| provenance | **mixed — see below** |
+
+**Column provenance, which is the point of this file and must not be flattened:**
+
+| group | example columns | provenance |
+|---|---|---|
+| identity | `Symbol`, `Aliases`, `Ensembl_Gene_ID`, `Canonical_Transcript_ID`, `biotype` | reference annotation |
+| membership | `source_databases` | **curated** — which CTA database(s) list this gene |
+| expression | `rna_testis_ntpm`, `rna_max_somatic_ntpm`, `protein_testis`, … | **experimental** (HPA) |
+| filters / calls | `passes_filters`, `rna_*_pct_filter`, `restriction`, `restriction_confidence`, `safety_flags` | **derived/computed** from the above |
+| MS evidence | `ms_pmids`, `ms_healthy_somatic_tissues`, `ms_restriction` | **experimental** — per-gene PMIDs of immunopeptidomics studies observing the gene as an HLA ligand |
+
+**Pipeline origin.** Built by [`pirl-unc/tsarina`](https://github.com/pirl-unc/tsarina) (Apache-2.0),
+which delegates CTA membership and reproductive-tissue sets to
+[`pirl-unc/oncoref`](https://github.com/pirl-unc/oncoref). **Neither repository declares a DOI or
+PMID for itself** — cite the tool by URL, and the underlying data by the references below.
+
+### Underlying sources, as tagged in `source_databases`
+
+`CTpedia` · `CTexploreR_CT` / `CTexploreR_CTP` · `daSilva2017` / `daSilva2017_protein` ·
+`placental_antigen` · `paralog:<gene>`
+
+> **daSilva2017** — da Silva VL, Fonseca AF, Fonseca M, da Silva TE, Coelho AC, Kroll JE,
+> de Souza JES, Stransky B, de Souza GA, de Souza SJ.
+> **Genome-wide identification of cancer/testis genes and their association with prognosis in a
+> pan-cancer analysis.** *Oncotarget* 2017;8(54):92966–92977.
+> PMID [29190970](https://pubmed.ncbi.nlm.nih.gov/29190970/) ·
+> doi:[10.18632/oncotarget.21715](https://doi.org/10.18632/oncotarget.21715)
+
+> **CTexploreR** — Loriot A, Devis J, Gatto L, De Smet C.
+> **A survey of human cancer-germline genes: Linking X chromosome localization, DNA methylation and
+> sex-biased expression in early embryos.** *PLoS Genet* 2025;21(10):e1011734.
+> PMID [41091706](https://pubmed.ncbi.nlm.nih.gov/41091706/) ·
+> doi:[10.1371/journal.pgen.1011734](https://doi.org/10.1371/journal.pgen.1011734)
+> — the paper describing the R/Bioconductor package and its reference list of 146 cancer-germline
+> genes.
+
+> **Human Protein Atlas** — source of every `protein_*` / `rna_*` column.
+> Uhlén M, Fagerberg L, Hallström BM, Lindskog C, Oksvold P, Mardinoglu A, et al.
+> **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)
+
+> **`ms_pmids`** — per-gene immunopeptidomics evidence, 178 distinct PMIDs. Spot-verified against
+> PubMed; the four most frequent are
+> [35051231](https://pubmed.ncbi.nlm.nih.gov/35051231/) (influenza-infected human tissue
+> immunopeptidome, 98 genes),
+> [32938616](https://pubmed.ncbi.nlm.nih.gov/32938616/) (melanoma spliced-peptide immunopeptidome,
+> 96),
+> [31844290](https://pubmed.ncbi.nlm.nih.gov/31844290/) (Sarkizova et al., mono-allelic HLA
+> peptidome / HLAthena, 76) and
+> [29557506](https://pubmed.ncbi.nlm.nih.gov/29557506/) (glioblastoma stem-like cell HLA ligandome,
+> 76). Resolve any individual PMID directly; they are not reproduced here.
+
+### Unresolved
+
+**`CTpedia` has no citation I could verify.** It is not resolvable from the tsarina or oncoref
+repositories, and a PubMed search did not return a record that unambiguously matches this tag.
+Recorded as unresolved rather than guessed — do not cite it until someone confirms which resource
+is meant (candidates include the CTdatabase / CT antigen database lineage, but that is a hypothesis,
+not a verification).
+
+---
+
+## Re-fetch / regenerate
+
+None of these three are regenerated by this repository — they are deposited as fetched.
+
+    # viral_foreign_iedb: IEDB export, non-human source organisms
+    #   https://www.iedb.org/  (manual export)
+    # viral_orfs_gse272406:
+    #   https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE272406
+    # cancer_targets_tsarina:
+    #   pip install tsarina  # https://github.com/pirl-unc/tsarina  (CTA table export)
+
+## Audit note on `DESCRIPTION.md`
+
+The `#` column in the main table is undocumented and does not correspond to column count, row count
+or dataset count consistently — for these three files it read 2, 1 and 3 against actual column
+counts of 6, 7 and 51. The ligandome rows now carry real counts in their one-liners; the `#` column
+itself should be defined or dropped repo-wide.

neoantigens/SOURCES.md

@@ -0,0 +1,170 @@
+# `neoantigens/` — sources, provenance, citations and a complete audit
+
+Every tumour-neoantigen dataset. Moved here from `immunogenicity/` on 2026-08-17 so that directory
+holds only general immunogenicity corpora; `nci_complete.tsv.gz` was renamed
+`nci_complete_minigenes.tsv.gz` because its unit is a minigene, not an epitope, and the old name hid
+that.
+
+Citations resolved against **PubMed**, PMIDs and DOIs given. Anything unverified is marked as such
+rather than filled in. Audit regenerated by `bench/neoag/audit_pmhc_data.py`.
+
+---
+
+## Complete audit
+
+| file | rows | distinct peptides | pos | neg | prevalence | HLA | donors | length |
+|---|--:|--:|--:|--:|--:|--:|--:|:--:|
+| `neoantigens_tested_peptides.tsv.gz` | 423,085 | 420,196 | 178 | 422,907 | 0.00042 | 82 | 98 | 8–12 |
+| `neoag_tested.tsv.gz` | 525,989 | 491,495 | 29,951 | 496,038 | 0.05694 | 601 | 127 | 3–50 |
+| `neoag_candidates.tsv.gz` | 1,355,926 | 1,344,070 | 0 | 0 | — | 82 | 99 | 8–12 |
+| `neoag_tested_hsa.tsv.gz` | 717 | 535 | 373 | 344 | 0.52022 | 34 | — | 8–35 |
+| `neoag_tested_mmu.tsv.gz` | 1,977 | 1,757 | 740 | 1,237 | 0.37430 | 17 | — | 8–35 |
+| `nci_gartner_nmers.tsv.gz` | 9,541 | 9,492 | 144 | 9,397 | 0.01509 | — | 70 | 11–168 |
+| `nci_gartner_mmp.tsv.gz` | 147 | 144 | 147 | 0 | 1.00000 | 43 | 77 | 8–12 |
+| `nci_gartner_samples.tsv.gz` | 98 | — | — | — | — | 85 | 97 | — |
+| `nci_complete_minigenes.tsv.gz` | 1,326 | 1,303 | 28 | 1,298 | 0.02112 | **0** | 46 | 13–316 |
+
+`neoag_candidates` is the **untested** candidate pool — 0 positives and 0 negatives is correct, its
+`immunogenicity` is null throughout. `nci_gartner_mmp` is positives-only by construction (it is the
+confirmed-epitope table); its negatives live in `nci_gartner_nmers`. `nci_complete_minigenes` has
+**zero** alleles because the deposit is not allele-deconvolved.
+
+### Residue composition, by label
+
+| file | Cys % pos | Cys % neg | C-term R/K % pos | C-term R/K % neg |
+|---|--:|--:|--:|--:|
+| `neoantigens_tested_peptides.tsv.gz` | 15.17 | 22.21 | 14.04 | 11.10 |
+| `neoag_tested.tsv.gz` | 19.50 | 23.03 | 11.40 | 11.38 |
+| `neoag_tested_hsa.tsv.gz` | 20.38 | 18.60 | 10.19 | 7.27 |
+| `neoag_tested_mmu.tsv.gz` | 20.81 | 26.27 | 5.27 | 6.71 |
+| `nci_gartner_nmers.tsv.gz` | 40.97 | 39.68 | 11.81 | 11.13 |
+| `nci_complete_minigenes.tsv.gz` | 25.00 | 39.21 | 3.57 | 11.86 |
+
+**No cysteine artefact here, and that is the point of contrast.** The MS-elution depletion that
+gives `immunogenicity/chowell_rebuilt.tsv.gz` a 6.9× Cys gradient (10.69% positives vs 1.55%
+negatives) is absent: these are screened synthetic peptides, so both classes go through the same
+assay. Cys rates are *high* (15–41%) simply because longer peptides are more likely to contain one —
+the 25-mer minigene tables sit near 40%.
+
+A model fitted on Chowell and applied here therefore meets an inverted cysteine distribution. That
+is a corpus-transfer artefact, not biology, and it is one measured reason a Chowell-trained residue
+term does not transfer to a neoantigen screen. See `../immunogenicity/SOURCES.md` for the full
+treatment of both the Cys and the C-terminal Arg/Lys issue.
+
+---
+
+## `neoantigens_tested_peptides.tsv.gz` — the screens at epitope resolution
+
+423,085 screened **8–12mers** with wild-type counterpart and restricting alleles; 178 CD8-positive;
+98 patients. Three datasets in one file with the authors' own split: `NCI` 420,786 · `HiTIDE` 1,563 ·
+`TESLA` 736.
+
+Harmonized from `raw/immunogenicity/NCI_dataset_only_tested.txt` by `bench/neoag/ingest_tested.py`.
+Canonical columns as-reported; every predicted quantity carries a `pred_` prefix.
+
+Provenance: **experimental** (CD8 T-cell screening) for `peptide`, `peptide_wt`, `mhc_a`,
+`immunogenicity`; **derived/computed** for every `pred_*` column.
+
+> **TESLA** — Wells DK, van Buuren MM, Dang KK, Hubbard-Lucey VM, Sheehan KCF, Campbell KM, et al.
+> **Key Parameters of Tumor Epitope Immunogenicity Revealed Through a Consortium Approach Improve
+> Neoantigen Prediction.** *Cell* 2020;183(3):818–834.e13.
+> PMID [33038342](https://pubmed.ncbi.nlm.nih.gov/33038342/) ·
+> doi:[10.1016/j.cell.2020.09.015](https://doi.org/10.1016/j.cell.2020.09.015)
+
+**Prevalence 4.2 × 10⁻⁴ — one immunogenic peptide in 2,376.** This is the operational base rate a
+ranking model faces, and it is ~75× lower than the training-corpus prevalence of
+`chowell_rebuilt` (0.039). Reading a corpus-prevalence probability as an operational one overstates
+it by that factor.
+
+---
+
+## `nci_gartner_*.tsv.gz` — the NCI screen, deconvolved
+
+Three tables, deliberately not merged; they are different units of observation.
+
+* **`nci_gartner_mmp`** — 144 distinct confirmed CD8+ **minimal mutant peptides** with restricting
+  HLA. 120 carry the wild-type mmp, so agretopicity is **measured, not predicted**. `split`
+  preserves the paper's train/test.
+* **`nci_gartner_nmers`** — 9,541 screened mutations as WT/mutant 25-mers (144 CD8+): the negative
+  universe at the unit the assay used.
+* **`nci_gartner_samples`** — 98 patients with full class-I genotype and histology. This is what
+  makes a per-patient allele-aware scan possible at all.
+
+> Gartner JJ, Parkhurst MR, Gros A, Tran E, Jafferji MS, Copeland A, Hanada KI, Zacharakis N,
+> Lalani A, Krishna S, Sachs A, Prickett TD, Li YF, Florentin M, Kivitz S, Chatmon SC,
+> Rosenberg SA, Robbins PF. **A machine learning model for ranking candidate HLA class I
+> neoantigens based on known neoepitopes from multiple human tumor types.**
+> *Nat Cancer* 2021;2(5):563–574.
+> PMID [34927080](https://pubmed.ncbi.nlm.nih.gov/34927080/) ·
+> doi:[10.1038/s43018-021-00197-6](https://doi.org/10.1038/s43018-021-00197-6)
+
+Supplementary Tables 1, 3, 11, 12, 18, 20 of `43018_2021_197_MOESM2_ESM.xlsx`, ingested by
+`bench/neoag/ingest_gartner.py`.
+
+**One join caveat, recorded.** `(ID, Variant key)` is not unique in Suppl. Tables 11 or 12; an outer
+join inflates 9,541 rows to 10,521. The ingest deduplicates the annotation side and left-joins with
+`validate="m:1"` so the row count stays the screened-nmer count.
+
+---
+
+## `nci_complete_minigenes.tsv.gz` — the same screen, minigene-level
+
+1,326 mutant 25-mer minigenes across 46 GI-cancer patients, 28 CD8-immunogenic. **`mhc_a` is blank
+by design** — not epitope-deconvolved. Formerly `immunogenicity/nci_complete.tsv.gz`.
+
+> Koşaloğlu-Yalçın Z, Lee J, Greenbaum J, Schoenberger SP, Miller A, Kim YJ, Sette A, Nielsen M,
+> Peters B. **Combined assessment of MHC binding and antigen abundance improves T cell epitope
+> predictions.** *iScience* 2022;25(2):103850.
+> PMID [35128348](https://pubmed.ncbi.nlm.nih.gov/35128348/) ·
+> doi:[10.1016/j.isci.2022.103850](https://doi.org/10.1016/j.isci.2022.103850) — Table S6.
+
+**Do not promote its k-mers to positives.** The immunogenicity readout is per minigene: a positive
+minigene contains one epitope and dozens of innocent k-mers. Use `nci_gartner_mmp` when the epitope
+identity is needed, or score this table max-over-k-mers at the minigene level.
+
+---
+
+## `neoag_tested.tsv.gz` / `neoag_candidates.tsv.gz` — the aggregated workhorse
+
+`neoag_tested`: 525,989 rows over 9 cohorts — CEDAR 91,651 · Neopep 422,292 · Gfeller 7,758 ·
+Gfeller_GBM 2,948 · TESLA 605 · Bjerregaard 291 · ITSNdb 199 · GBM 151 · VACCIMEL 94.
+`neoag_candidates`: the untested Neopep candidate pool.
+
+Deduplication rule: primary datasets take precedence over the aggregated compendia (CEDAR, Neopep) —
+a duplicate peptide–MHC already present in a primary dataset is dropped from CEDAR/Neopep.
+
+**`neoag_tested_mmu` / `_hsa` split on `mhc_species`, not host.** H-2 restrictions go to `_mmu`,
+HLA restrictions (HLA-transgenic mice *and* human hosts) to `_hsa`; `host_species` /
+`source_species` preserve the distinction within each file. `_hsa` has no populated `patient_id`.
+
+### Per-cohort citations — **partially unresolved**
+
+Verified:
+
+* **TESLA** — Wells et al. 2020, PMID [33038342](https://pubmed.ncbi.nlm.nih.gov/33038342/)
+  (above).
+
+**Not yet verified, and deliberately not written out:** CEDAR, Neopep, Gfeller, Gfeller_GBM,
+ITSNdb, GBM, VACCIMEL, Bjerregaard. Raw inputs for each are under `raw/immunogenicity/`
+(`cedar_tcell_assays_25-03-2026.tsv`, `Gfeller_peptides.tsv.gz`, `Gfeller_GBM.tsv.gz`,
+`ITSNdb.tsv`, `GBM_Peptides.tsv.gz`, `VACCIMEL.tsv`, `Bjerregaard_Peptides.csv.gz`), which is the
+provenance that is actually established. Resolve each against PubMed before citing it anywhere
+downstream — do not infer a reference from a cohort name.
+
+### Cohort overlap — these are not independent holdouts
+
+Measured: TESLA and NCI/Neopep share **71.8%** of the smaller set (30 of TESLA's 37 positives);
+Gfeller and Gfeller_GBM share **96.5%**. Four "independent" cohorts are two families. Treat them as
+two, not four, in any leave-one-cohort-out design.
+
+---
+
+## Regenerate
+
+    # in ~/vcs/projects/2026-mhcmatch-benchmark
+    python bench/neoag/ingest_tested.py       # -> neoantigens_tested_peptides.tsv.gz
+    python bench/neoag/ingest_gartner.py      # -> nci_gartner_{mmp,nmers,samples}.tsv.gz
+    python bench/neoag/audit_pmhc_data.py     # -> the audit tables above
+
+`neoag_tested*` / `neoag_candidates` / `nci_complete_minigenes` are deposited as fetched and are not
+regenerated by this repository.