TCGA — The Cancer Genome Atlas (via NCI Genomic Data Commons)

At a glance

Labels and annotations

Per-case clinical data (stage, grade, vital status, survival), somatic mutations (MAF), copy number, RNA-seq expression, methylation, miRNA, and slide images; slide-level labels are the pathology diagnosis, not region annotations.

Details

What is in it

• Harmonised genomic data re-processed by the GDC against GRCh38.
• Diagnostic (FFPE) and frozen-tissue slides in Aperio SVS format.
• Clinical and biospecimen supplements; cBioPortal offers a browsing front end.

How to get it

• Open-access files download without an account through the portal or the gdc-client.
• Controlled data require dbGaP authorisation via an institutional signing official.

Known pitfalls

• Slide scanners, sites and batch effects vary; several pathology foundation models pretrained on TCGA, so it is not an external test set for them.
• Survival follow-up is short for some projects.

Models trained or evaluated on it

• evaluation UNI Mahmood Lab, Brigham and Women's Hospital / Harvard Medical School — subtyping and biomarker tasks drawn from TCGA cohorts
• evaluation CONCH Mahmood Lab, Brigham and Women's Hospital / Harvard Medical School
• evaluation Prov-GigaPath Microsoft Research / Providence Health / University of Washington
• pretraining Phikon-v2 Owkin — part of PANCAN-XL — not a valid external test set
• training CHIEF Yu Lab, Harvard Medical School
• evaluation H-optimus-0 Bioptimus
• training Phikon-v2 Owkin

Sources

1. GDC Data Portal
2. NCI — The Cancer Genome Atlas Program