---
license: other
license_name: nih-genomic-data-sharing
license_link: https://gdc.cancer.gov/analyze-data/data-analysis-policies
pretty_name: TCGA Patients (prototype)
tags:
  - cancer
  - tcga
  - clinical
  - genomics
configs:
  - config_name: patients
    data_files: patients/*.parquet
---

# TCGA Patients (prototype)

One row per patient. Open-access TCGA clinical data from the NCI Genomic Data
Commons (GDC), pre-joined into a single nested record per case so multi-modal
oncology models can stream rich patient context in batches without doing joins.

- **Projects included:** `TCGA-CHOL`, `TCGA-DLBC`
- **Generated:** 2026-05-03 23:41:10 UTC
- **Source:** NCI GDC `/cases` endpoint with expansions on `demographic`,
  `diagnoses`, `diagnoses.treatments`, `follow_ups`, `exposures`,
  `family_histories`, `project`.
- **Access tier:** open. Redistributable; see [License & redistribution](#license--redistribution) below.
- **Partitioning:** one Parquet file per TCGA project, at
  `patients/<TCGA-XXXX>.parquet`. The `patients` HF config globs all of them.

## Loading

```python
from datasets import load_dataset

# Stream all projects (HF concatenates the per-project parquets):
ds = load_dataset("<repo_id>", "patients", split="train", streaming=True)
for patient in ds:
    case_id = patient["case_id"]
    diagnoses = patient["diagnoses"]                # list of dicts
    treatments = [tx for dx in diagnoses for tx in dx["treatments"]]
    ...

# Restrict to one project: load only its parquet
chol = load_dataset("<repo_id>", data_files="patients/TCGA-CHOL.parquet")
```

Or load into pyarrow / polars / pandas directly:

```python
import pyarrow.parquet as pq

table = pq.read_table(
    "patients/TCGA-CHOL.parquet",
    columns=["case_id", "project_id", "demographic", "diagnoses"],
)
```

## Verification

This dataset's source is the NCI GDC. The same TCGA cases are also browsable on
[cBioPortal](https://www.cbioportal.org/). cBioPortal hosts **multiple versions**
of TCGA per disease (Firehose Legacy, PanCancer Atlas, GDC-sourced); the
`*_tcga_gdc` family shares our upstream source and is the like-for-like
comparison. Disease prefixes use cBioPortal's own taxonomy, not always the
TCGA project suffix (e.g. `TCGA-DLBC` → `dlbclnos_tcga_gdc`).

```
https://www.cbioportal.org/patient/summary?studyId=<study>&caseId=<case_submitter_id>
```

For example, patient `TCGA-W5-AA39` (project `TCGA-CHOL`) →
`https://www.cbioportal.org/patient/summary?studyId=chol_tcga_gdc&caseId=TCGA-W5-AA39`.

## Schema

```
case_id: string
case_submitter_id: string
project_id: string
primary_site: string
disease_type: string
demographic: struct<
  demographic_id: string,
  submitter_id: string,
  gender: string,
  sex_at_birth: string,
  race: string,
  ethnicity: string,
  vital_status: string,
  age_at_index: int64,
  days_to_birth: int64,
  days_to_death: int64,
  year_of_birth: int64,
  year_of_death: int64,
  country_of_residence_at_enrollment: string,
>
diagnoses: list<struct<
  diagnosis_id: string,
  submitter_id: string,
  primary_diagnosis: string,
  morphology: string,
  tissue_or_organ_of_origin: string,
  site_of_resection_or_biopsy: string,
  icd_10_code: string,
  ajcc_pathologic_stage: string,
  ajcc_pathologic_t: string,
  ajcc_pathologic_n: string,
  ajcc_pathologic_m: string,
  ajcc_staging_system_edition: string,
  age_at_diagnosis: int64,
  days_to_diagnosis: double,
  year_of_diagnosis: int64,
  prior_malignancy: string,
  prior_treatment: string,
  synchronous_malignancy: string,
  classification_of_tumor: string,
  last_known_disease_status: string,
  days_to_last_follow_up: double,
  days_to_last_known_disease_status: double,
  days_to_recurrence: double,
  residual_disease: string,
  diagnosis_is_primary_disease: bool,
  treatments: list<struct<
    treatment_id: string,
    submitter_id: string,
    treatment_type: string,
    treatment_or_therapy: string,
    treatment_intent_type: string,
    treatment_outcome: string,
    therapeutic_agents: string,
    days_to_treatment_start: double,
    days_to_treatment_end: double,
    initial_disease_status: string,
  >>,
>>
follow_ups: list<struct<
  follow_up_id: string,
  submitter_id: string,
  timepoint_category: string,
  disease_response: string,
  progression_or_recurrence: string,
  days_to_follow_up: double,
  days_to_progression: double,
  days_to_recurrence: double,
  ecog_performance_status: string,
>>
exposures: list<struct<
  exposure_id: string,
  submitter_id: string,
  tobacco_smoking_status: string,
  cigarettes_per_day: double,
  years_smoked: double,
  alcohol_history: string,
  alcohol_intensity: string,
  bmi: double,
  weight: double,
  height: double,
>>
family_histories: list<struct<
  family_history_id: string,
  submitter_id: string,
  relationship_type: string,
  relative_with_cancer_history: string,
  relationship_primary_diagnosis: string,
>>
```

Notes on the nested layout:

- `demographic` is a single `struct` (1:1 with the patient).
- `diagnoses`, `follow_ups`, `exposures`, `family_histories` are `list<struct<...>>`
  (0..N per patient).
- `treatments` is nested *inside* each diagnosis (`diagnoses[i].treatments`), so
  the source 1:N relationship between diagnosis and treatment is preserved.
- All leaf fields are nullable.

## License & redistribution

This dataset only contains TCGA data fetched from the GDC's **open-access** tier
(`/cases` endpoint with `access=open`). Per the
[NCI GDC Data Analysis Policy](https://gdc.cancer.gov/analyze-data/data-analysis-policies):

> The GDC itself places no restrictions (other than attempts at reidentification)
> on analysis or publication of open access data provided through the GDC Data Portal.

Per the [NCI TCGA citation page](https://www.cancer.gov/ccg/research/genome-sequencing/tcga/using-tcga-data/citing):

> Moratoria on all cancer types are now lifted and all TCGA data are available
> without restrictions on their use in publications or presentations.

Per the [GDC Data Access Processes and Tools page](https://gdc.cancer.gov/access-data/data-access-processes-and-tools):

> Open access data generally includes high level genomic data that is not
> individually identifiable, as well as most clinical and all biospecimen data
> elements.

The clinical data here is squarely within that scope.

## Restrictions on use

The single hard restriction inherited from the GDC policy applies to all
downstream users:

> Users of any data provided by GDC, whether open or controlled access, agree
> not to attempt to reidentify any individual participant in any study
> represented by GDC data, for any purpose whatever.
> ([source](https://gdc.cancer.gov/analyze-data/data-analysis-policies))

## Required acknowledgement

If you publish or present results derived from this dataset, please include the
[NCI-required TCGA acknowledgement](https://www.cancer.gov/ccg/research/genome-sequencing/tcga/using-tcga-data/citing):

> The results <published or shown> here are in whole or part based upon data
> generated by the TCGA Research Network: https://www.cancer.gov/tcga.

Suggested citations for the GDC and TCGA:

- Grossman, R. L., et al. (2016). Toward a Shared Vision for Cancer Genomic Data.
  *NEJM*, 375(12), 1109-1112.
- The Cancer Genome Atlas Research Network. https://www.cancer.gov/tcga
- NCI Genomic Data Commons. https://gdc.cancer.gov

Additional policy references:

- [GDC Policies (umbrella)](https://gdc.cancer.gov/about-gdc/gdc-policies)
- [GDC Encyclopedia — Controlled Access](https://docs.gdc.cancer.gov/Encyclopedia/pages/Controlled_Access/) (defines what is *not* in this dataset)
- [NIH Genomic Data Sharing Policy](https://sharing.nih.gov/genomic-data-sharing)

## Disclaimer

Prototype dataset for engineering validation. Schemas, projects, and column
coverage will change as additional modalities (mutations, expression, copy number)
are added — likely as additional top-level nested columns on the same patient row.
Re-derive from the GDC for any analysis where freshness matters.