---
pretty_name: pMHC Data Compendium
license: other
license_name: mixed-upstream
license_link: https://huggingface.co/datasets/isalgo/pmhc_data#licensing
language:
  - en
tags:
  - immunology
  - immunopeptidomics
  - neoantigen
  - mhc
  - hla
  - t-cell
  - cancer-vaccine
  - proteomics
size_categories:
  - 10M<n<100M
---

# pMHC Data Compendium

Harmonized peptide–MHC data for neoantigen scoring: presentation, immunogenicity labels,
self/foreign reference peptidomes, expression, and clinical-outcome cohorts. Human and mouse.

Every table here is derived from public sources. Per-file provenance, filtering rules, verified
PubMed citations and row-level audits live in the `SOURCES.md` of each directory — that is the
authoritative record, not this page.

## Contents

| Directory | What is in it |
|---|---|
| [`pmhc/`](pmhc) | Epitope–MHC binding records — the presentation panel (`pmhc_full`, and `pmhc_shortlist` at ≥2 references) |
| [`immunogenicity/`](immunogenicity) | T-cell immunogenicity corpora: immunogenic vs **self** ligands and vs **non-self**, each with an HLA-matched variant; labelled set; HLA population frequencies |
| [`neoantigens/`](neoantigens) | Tested and candidate neoantigens — minigene, epitope-level and per-mutation screens, human and mouse |
| [`ligandome/`](ligandome) | Foreign and tissue reference peptidomes — viral ligands, pan-viral ORFs, cancer-testis targets, per-tissue corpora |
| [`thymus/`](thymus) | Thymic self-peptidome and thymic expression, human and mouse — the central-tolerance reference |
| [`expression/`](expression) | Reference expression, human and mouse, plus a single-pipeline table and mouse protein abundance |
| [`proteome/`](proteome) | Human and mouse reference proteomes plus 16 bacterial/viral proteomes for mimicry scans |
| [`homology/`](homology) | Cross-species homology mappings |
| [`tcga/`](tcga) | Expressed mutant-peptide–HLA binders across 8,505 donors |
| [`hla_loh/`](hla_loh) | HLA-I loss of heterozygosity and genetic immune escape over 6,319 tumours, with three-caller typing concordance |
| [`icb_cohorts/`](icb_cohorts) | Harmonized checkpoint-blockade cohorts with outcome |
| [`vaccines/`](vaccines) | Shipped personalized-vaccine cassettes and their immune monitoring |

Formats are mixed by design: gzipped TSV for the large harmonized tables, parquet where the schema
is wide or typed, `.npz` for dense matrices, gzipped FASTA for proteomes.

## Shared schema

The immunogenicity tables share one schema:

```
peptide, mhc_a, mhc_b, mhc_class, mhc_species, source_species, host_species,
assay_type, immunogenicity{1,0,null}, expression, affinity, affinity_wt,
agretopicity, foreignness, validated, dataset_origin, reference_id, n_references
```

plus `patient_id, cancer_type` on the neoantigen tables. **Canonical columns hold as-reported
values only**; anything predicted lives in a `pred_*` column, so `affinity` is always experimental.
`affinity` units differ by source and MHC alleles are as-reported.

## Two things to know before filtering

- **`neoantigens/neoag_tested.tsv.gz` contains mouse rows labelled as human.** 6,186 CEDAR rows
  carry mouse MHC alleles (`H2-Db` 1,020, `H2-Kb` 932, `H2-b class I` 719, `H2-Kd` 525,
  `H2-d class II` 505, …) while their `mhc_species` reads `HomoSapiens` and `host_species` reads
  `human`. `SIINFEKL`, the canonical H-2Kb ovalbumin epitope, is among them. Filtering that table
  on `mhc_species` therefore silently admits mouse pMHC. The defect is upstream and is left
  unfixed; `neoantigens/neoag_tested_mmu.tsv.gz` is the correctly-labelled murine set.
- **`neoag_tested_{mmu,hsa}` overlap `neoag_tested.tsv.gz` by design** — all 2,273 peptides appear
  in both, since both are IEDB-derived. Dedupe on peptide × MHC when joining.

## Loading

```python
from huggingface_hub import hf_hub_download
import pandas as pd

path = hf_hub_download("isalgo/pmhc_data", "pmhc/pmhc_shortlist.tsv.gz", repo_type="dataset")
df = pd.read_csv(path, sep="\t")
```

Grab one directory rather than the whole repo:

```bash
hf download isalgo/pmhc_data --repo-type dataset --include "neoantigens/*" --local-dir ./pmhc_data
```

[`mhcmatch`](https://pypi.org/project/mhcmatch/) consumes this compendium directly — `mhcmatch
bootstrap` stages the panel and reference tables from here.

## Upstream sources

Which table came from which, and under what filtering, is in the `SOURCES.md` of each directory.

| | |
|---|---|
| IEDB | https://www.iedb.org |
| CEDAR | https://cedar.iedb.org |
| HLA Ligand Atlas | https://hla-ligand-atlas.org |
| UniProt | https://www.uniprot.org |
| GTEx | https://gtexportal.org |
| Human Protein Atlas | https://www.proteinatlas.org |
| FANTOM5 | https://fantom.gsc.riken.jp/5/ |
| UCSC Xena | https://xenabrowser.net |
| GDC / TCGA | https://portal.gdc.cancer.gov |
| ICGC / PCAWG | https://dcc.icgc.org |
| Hartwig Medical Foundation | https://www.hartwigmedicalfoundation.nl |
| Allele Frequency Net Database | https://www.allelefrequencies.net |
| PRIDE | https://www.ebi.ac.uk/pride/ |
| MassIVE | https://massive.ucsd.edu |
| GEO | https://www.ncbi.nlm.nih.gov/geo/ |

## Licensing

Each upstream source carries its own terms; this compendium redistributes derived, harmonized
tables and does not relicense them. Check the `SOURCES.md` of the directory you are using before
redistributing, and cite the primary publications recorded there rather than this repository.