---
license: other
task_categories:
  - tabular-classification
tags:
  - immunogenomics
  - AIRR
  - TCR
  - BCR
  - repertoire-sequencing
  - TCGA
  - cancer
  - tumor-infiltrating-lymphocytes
  - RNA-seq
pretty_name: TCGA — tumour AIRR (TCR/BCR) repertoires with clinical annotation
size_categories:
  - 10M<n<100M
---

# TCGA — tumour AIRR (TCR/BCR) repertoires

**Adaptive immune receptor (TCR + BCR) repertoires extracted from TCGA tumour bulk RNA-seq**,
for **9,591 samples** spanning **33 cancer types**, with matched **clinical / demographic
metadata** and — critically — the **total raw sequencing depth** of each sample. Built as a
benchmark for **sample-level (repertoire) analysis in solid tumours**, where the recovered
receptor content conflates sequencing depth with immune infiltration (TIL fraction) and must be
handled as such.

> The results shown here are in part based upon data generated by the **TCGA Research Network**
> (<https://www.cancer.gov/tcga>). AIRR rearrangements were extracted from tumour RNA-seq; for the
> RNA-seq receptor-profiling approach see **Bolotin DA, Poslavsky S, Davydov AN, et al.
> "Antigen receptor repertoire profiling from RNA-seq data." *Nature Biotechnology* 2017;35(10):908–911.**
> [DOI: 10.1038/nbt.3979](https://doi.org/10.1038/nbt.3979) · PMID
> [29020005](https://pubmed.ncbi.nlm.nih.gov/29020005/).

## Contents

| File | What |
|---|---|
| `samples.tar.gz` | 9,591 per-sample AIRR tables, `samples/<sample_id>.tsv` (git-LFS) |
| `metadata.tsv` | one row per sample, keyed by `sample_id` — clinical, demographic, read counts |
| `metadata.hla.tsv` | one row per **donor**, keyed by `subject_id` — HLA class-I typing (OptiType) |
| `load.py` | minimal `huggingface_hub` + `pandas` bootstrap loader |
| `SOURCES.md` | provenance of every field (source file, fetch, experimental vs derived) |

### Per-sample AIRR tables (`samples/<sample_id>.tsv`)

Tab-separated, AIRR Rearrangement field names, one clonotype per row:

| column | meaning |
|---|---|
| `locus` | `IGH/IGK/IGL/TRA/TRB/TRG/TRD` |
| `v_call`, `j_call`, `c_call` | IMGT V / J / constant (isotype) gene calls |
| `junction`, `junction_aa` | junction nucleotide / amino-acid sequence (conserved anchors included) |
| `duplicate_count` | number of reads supporting the clonotype (RNA-seq abundance) |
| `sample_id` | back-reference to the metadata key |

The dataset is **immunoglobulin-dominant** (as expected for tumour RNA-seq, where the plasma-cell /
B-cell infiltrate contributes most receptor reads):

| locus | clonotypes | share |
|---|---:|---:|
| IGK | 9,036,576 | 44.4% |
| IGL | 5,580,569 | 27.4% |
| IGH | 5,045,108 | 24.8% |
| TRB | 352,641 | 1.7% |
| TRA | 306,824 | 1.5% |
| TRG | 23,243 | 0.1% |
| TRD | 5,133 | 0.03% |

Per sample: **median 634**, mean 2,121, max 52,388 clonotypes (all loci pooled) — the
low-coverage regime typical of bulk RNA-seq, 20,350,094 clonotypes total.

### `metadata.tsv` (one row per sample)

Key `sample_id` (`TCGA-XX-XXXX.N`) matches the per-sample table names exactly. Columns:

- **identity** — `sample_id`, `subject_id` (case), `study_id`/`cohort` (e.g. `TCGA-GBM`),
  `center_id`, `center_name`, `metacohort_id`.
- **phenotype** — `cancer_type`, `disease`, `tumor_type`, `source_type`, `source_tumor`,
  `source_icd10`, `sex`, `race`, `age`, `cancer_stage`, `tumor_stage`,
  `therapy`, `therapy_info`, `time_point`, `response`, `response_type`,
  `OS`, `OS_event`, `PFS`, `PFS_event`, `AE`, `AE_type`, `AE_severity`.
- **sequencing depth** — `total_reads`, `aligned_reads` (see below).

#### Read counts — total (raw) vs aligned

`total_reads` is the **total number of reads in the raw sequencing library** (all reads, mapped
*and* unmapped) — **not** the number of reads mapped to a receptor. It is the read total of the
GDC `rna_seq.genomic.gdc_realn.bam` (`Total_Reads`), verified to equal `tcga_meta_slim.total_reads`
on matched samples (median ratio 1.000). `aligned_reads` is the genome-aligned subset
(`Aligned_Reads`) — provided as a bonus.

This distinction matters in tumours: the recovered receptor content (`Σ duplicate_count`, or clone
count) mixes **sequencing depth** with **immune infiltration**. `total_reads` is the technical
denominator that lets you separate them — e.g. the receptor read fraction
`Σ duplicate_count / total_reads` is a depth-normalised infiltration proxy, whereas
`duplicate_count` alone is not.

### `metadata.hla.tsv` (one row per donor)

Germline **HLA class-I** typing (4-digit A/B/C) from TCGA RNA-seq, keyed by `subject_id` (the
donor — HLA is germline, so it is per-donor not per-sample). Join to `metadata.tsv` on
`subject_id`.

| column | meaning |
|---|---|
| `subject_id` | donor barcode `TCGA-XX-XXXX` — matches `metadata.tsv` `subject_id` |
| `HLA-A_1`, `HLA-A_2`, `HLA-B_1`, `HLA-B_2`, `HLA-C_1`, `HLA-C_2` | the two alleles per locus, e.g. `HLA-A*02:01` |
| `hla_source` | which call set the row came from — `panimmune` or `optitype` |

Two OptiType-based call sets are unioned to maximise coverage: the curated **TCGA PanImmune** HLA
table (primary; Thorsson et al. 2018) plus the raw **OptiType** 2017 call set (Szolek et al. 2014)
filling donors PanImmune lacks. Both derive from OptiType, so they are methodologically consistent.
Provenance:

> - **PanImmune** (`hla_source = panimmune`, 7,649 donors) — Thorsson V, Gibbs DL, Brown SD, et al.
>   "The Immune Landscape of Cancer." *Immunity* 2018;48(4):812–830.e14.
>   [DOI 10.1016/j.immuni.2018.03.023](https://doi.org/10.1016/j.immuni.2018.03.023) · PMID
>   [29628290](https://pubmed.ncbi.nlm.nih.gov/29628290/); via the TCGA PanImmune resource.
> - **OptiType** (`hla_source = optitype`, 1,474 donors) — Szolek A, Schubert B, Mohr C, et al.
>   "OptiType: precision HLA typing from next-generation sequencing data." *Bioinformatics*
>   2014;30(23):3310–3316. [DOI 10.1093/bioinformatics/btu548](https://doi.org/10.1093/bioinformatics/btu548)
>   · PMID [25143287](https://pubmed.ncbi.nlm.nih.gov/25143287/).

Coverage: **9,123 / 9,450 donors** carry HLA (**9,263 / 9,591 samples**); 187 donors are in neither
call set.

## Coverage / caveats

- **All 9,591** AIRR samples are keyed in `metadata.tsv`.
- **`total_reads` present for 9,549 / 9,591** (99.6%); 42 samples lack a matchable GDC BAM.
  Where `tcga_meta_slim` left it blank (556 samples), it was recovered from the GDC read-count
  table for the 514 single-BAM cases (unambiguous).
- **Clinical annotation present for 9,104 / 9,591** (95%); 487 samples (mostly `.2` second
  aliquots) have `NA` clinical fields but retain demographics + read counts.
- `aligned_reads` present for 9,358 / 9,591 (97.6%).
- **HLA class-I present for 9,123 / 9,450 donors** (96.5%, `metadata.hla.tsv`) → 9,263 / 9,591
  samples (PanImmune ∪ OptiType); 187 donors are in neither call set.

## Usage

```python
from load import load_metadata, load_hla, load_sample, iter_samples

md = load_metadata()                       # DataFrame, one row per sample
hla = load_hla()                           # one row per donor (subject_id)
md = md.merge(hla, on="subject_id", how="left")   # attach HLA to each sample
df = load_sample(md.sample_id.iloc[0])     # one sample's AIRR clonotypes
for sample_id, clonotypes in iter_samples():   # stream all 9,591
    ...
```

## License

TCGA-derived. Use is subject to the **NIH Genomic Data Sharing Policy** and the TCGA data-use
terms; the sequences here are de-identified RNA-derived receptor clonotypes and standardised
clinical annotation. Cite the TCGA Research Network as above.