---
license: cc0-1.0
task_categories:
- tabular-classification
- feature-extraction
tags:
- tcga
- bulk-rna-seq
- methylation
- mutation
- cancer
- genomics
- multi-omics
pretty_name: TCGA PanCancer Atlas (8-subtype base set + per-architecture extensions)
---

# TCGA

A multi-omics subset of the TCGA PanCancer Atlas (RNA, methylation, mutation, clinical),
for use with [AUTOENCODIX](https://github.com/jan-forest/autoencodix_package) tutorials.

## Base set (for `Vanillix` and other Tutorials)

Covers **8 cancer types**: BRCA, OV, LUAD, UCEC, LUSC, COAD, READ, UCS.

- `rna.parquet` — 3552 samples × 17448 genes (RNA expression)
- `methylation.parquet` — 3875 samples × 9829 genes (per-gene methylation)
- `mutation.parquet` — 3461 samples × 20304 genes (combined mutation/CNA score per gene;
  restricted to samples that belong to the 8 cancer types above — the source file is a
  pan-cancer TCGA mutation matrix, trimmed here to match this cohort)
- `clinical.parquet` — 3902 samples × 56 columns (sample-level clinical/annotation
  metadata, incl. `CANCER_TYPE_ACRONYM`, `SUBTYPE`)

Gene identifiers are NCBI/Entrez gene IDs throughout, consistent with the ontology files
below.

## `ontology/` — extension for `Ontix` Tutorial

Two ontology hierarchies, each as a 2-level gene-grouping file (gene ID → level-1 group,
level-1 group → level-2 group):

- `chromosome_ont_lvl1_ncbi.txt`, `chromosome_ont_lvl2.txt` — chromosome/cytoband-based
  grouping
- `reactome_ont_lvl1.txt`, `reactome_ont_lvl2.txt` — Reactome pathway-based grouping
  (level 2 uses human-readable Reactome pathway names)

## `xmodalix/` — extension for `XModalix` Tutorial

Synthetic MNIST-style images paired to TCGA samples, one per sample, for the
image↔omics translation tutorial:

- `xmodalix/tcga_fake_images.zip` — 3907 grayscale 28×28 PNGs
- `xmodalix/tcga_image_mappings.txt` — tab-separated `sample_ids`, `img_paths`,
  `extra_class_labels`, `CANCER_TYPE_ACRONYM`; 3230 of the 3902 base-set clinical
  samples have a paired image (one image label per cancer type)

## Loading

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

rna_path = hf_hub_download(repo_id="autoencodix/tcga", repo_type="dataset", filename="rna.parquet")
meth_path = hf_hub_download(repo_id="autoencodix/tcga", repo_type="dataset", filename="methylation.parquet")
mut_path = hf_hub_download(repo_id="autoencodix/tcga", repo_type="dataset", filename="mutation.parquet")
clin_path = hf_hub_download(repo_id="autoencodix/tcga", repo_type="dataset", filename="clinical.parquet")

rna = pd.read_parquet(rna_path)
```

Ontix additionally needs, e.g.:

```python
chr1 = hf_hub_download(repo_id="autoencodix/tcga", repo_type="dataset", filename="ontology/chromosome_ont_lvl1_ncbi.txt")
rea1 = hf_hub_download(repo_id="autoencodix/tcga", repo_type="dataset", filename="ontology/reactome_ont_lvl1.txt")
```

XModalix additionally needs:

```python
images_zip = hf_hub_download(repo_id="autoencodix/tcga", repo_type="dataset", filename="xmodalix/tcga_fake_images.zip")
mapping = hf_hub_download(repo_id="autoencodix/tcga", repo_type="dataset", filename="xmodalix/tcga_image_mappings.txt")
```

## License

**CC0 1.0 (Public Domain Dedication).** The TCGA PanCancer Atlas open-access data
(RNA, methylation, mutation, clinical) is designated public domain by the NIH Genomic
Data Commons, with no restrictions on reuse. The Reactome pathway groupings in
`ontology/` are separately licensed under CC0 by Reactome. The synthetic `xmodalix/`
images are generated by this project and released under the same CC0 terms.