NCI Genomic Data Commons

The NCI Genomic Data Commons is the National Cancer Institute's unified repository and knowledge base for cancer genomic data - one harmonized home for the molecular and clinical records of 50,571 cancer cases spread across more than 1.3 million files as of Data Release 46.0 (August 10, 2026). Named contributing programs pour into it: TCGA with 11 projects, TARGET with 6, the CPTAC-2 and CPTAC-3 proteogenomics efforts, plus CGCI, CMI, BEATAML, Exceptional Responders, MATCH, ALCHEMIST, ORGANOID, REBC, CTSP and FM.

Harmonization is the part worth dwelling on. Raw sequencing alignments land as BAM, variant calls as VCF and MAF, expression quantification as standardized matrices, methylation arrays as IDAT, whole-slide pathology images as SVS, and clinical and biospecimen records as TSV, JSON or XML - all reprocessed against a common pipeline rather than left in each program's native shape. A single versioned data dictionary describes every entity, published both as browsable documentation and as machine-readable schemas.

The practical consequence: joining a lung adenocarcinoma case's stage, survival status, diagnostic slide and somatic mutation calls stops being a fourteen-pipeline reconciliation project and becomes a lookup on one case identifier. Get a sample of this dataset and inspect real records before committing pipeline time.

Two records straight from the corpus: one case row, and the release stamp every delivery ships under.

case_id               : 6e3b6b72-142d-4b8d-a462-28a205796e41
submitter_id          : TCGA-44-3918
project_id            : TCGA-LUAD
primary_site          : Bronchus and lung
disease_type          : Adenomas and Adenocarcinomas
primary_diagnosis     : Adenocarcinoma, NOS
ajcc_pathologic_stage : Stage IA
demographic.race      : white
vital_status          : Alive

data_release          : Data Release 46.0 - August 10, 2026
status                : OK

Read the first block top to bottom and you have a complete patient anchor: the case carries a permanent GDC universally unique identifier and the project-native donor barcode (TCGA-44-3918) side by side, so records join inward on either key. Disease is pinned three ways - topography (Bronchus and lung), morphology class (Adenomas and Adenocarcinomas) and narrative-text diagnosis (Adenocarcinoma, NOS) - with AJCC pathologic stage Stage IA, self-reported race and vital status on the same flat row.

The second block is the release banner. Every delivery arrives pinned to a numbered release, which means an analysis run against Release 46.0 stays reproducible even as later releases re-harmonize underlying files. Beneath each case row hang its child records - alignments, variant calls, expression matrices, slides - reachable through the same case identifier shown above.

Thirteen core fields form the case-level spine, every one carrying a verified definition. Identifiers case_id and submitter_id anchor the record on both the repository-native and program-native sides; project.project_id, primary_site, disease_type, diagnoses.primary_diagnosis and diagnoses.ajcc_pathologic_stage classify the disease; demographic.race, demographic.vital_status and days_to_lost_to_followup carry the patient context survival analysis needs; batch_id and consent_type document submission provenance; updated_datetime timestamps the last revision.

Beyond the spine sit the remaining entity modules - file provenance, biospecimen chain of custody, treatment and therapy detail, read-group metadata - delivered as additional fields on request, so a pipeline only pulls the modules its analysis actually touches.

Geography - United States-centric cancer cohorts with international contributions mixed in. Because topography and morphology ride ICDO classifications on every case, cohorts pool cleanly across programs instead of fragmenting along institutional coding habits.

Temporal - the specimen era opens with TCGA samples collected 2005-2013, and later programs continue collecting today. The commons itself launched in 2016 and currently stands at Data Release 46.0 (August 10, 2026); analyses pin to a release number and stay reproducible as new releases land.

Granularity - one row per case, with per-file children across genomic, transcriptomic, epigenomic, proteomic, imaging and clinical data types hanging off that case key. Depth varies by program and case, so counting child files per case - rather than assuming completeness - is the honest way to size a cohort.

Scale check: 50,571 cases and 1,337,360 files, with the format mix led by VCF (about 364,000 files), BAM (about 200,000), TSV (about 199,000) and MAF (about 172,000), rounded out by IDAT methylation arrays and SVS pathology slides. Rebuilding that breadth independently would mean re-running harmonized pipelines over raw data from fourteen programs - years of work the commons already absorbed.

API, files, or your warehouse. Daily, weekly, or hourly.

• Biomarker discovery and validation - join stage, vital status and follow-up duration against per-case molecular profiles so a candidate marker is checked for prognostic signal inside one query rather than one literature review.
• Cohort construction and study feasibility - filter by project code, primary site, disease type and pathologic stage to size a recruitable cohort before a protocol commits to it.
• Model training - 50,571 outcome-labeled cases give classifiers and survival models supervision that arrives pre-harmonized instead of assembled from fourteen inconsistent pipelines.
• Pathology-imaging AI - SVS whole-slide images paired with structured diagnoses and outcomes make slide-level models trainable against ground truth that is already coded, not hand-annotated after the fact.
• Competitive diligence - program-by-program depth shows who studied which diseases, at what granularity, before anyone claims a gap in the evidence.

The data scientists working in biotechnology briefing maps more of these workflows end to end.

Bioinformaticians and computational biologists get clinically anchored molecular data on one identifier per case - the variant call and the survival status stop living in different documents. Data scientists and ML engineers get a large, outcome-labeled, schema-stable corpus suited to supervised problems; see the data scientists use cases page. Competitive-intel product teams read program coverage as a map of where the evidence already runs deep. Developers building data products get stable universally unique identifiers that join outward without fuzzy matching - see developers builders use cases. Investors and diligence teams test whether a claimed cohort effect survives contact with the actual cases. Start from the biotechnology data hub, then read NCI Genomic Data Commons vs NCBI GenBank for where a harmonized clinical-molecular commons beats a sequence archive.

Notes that pair well with this page:

• Biotechnology data hub - the pooled industry view this record sits inside, alongside sequence, structure and interaction corpora.
• NCI Genomic Data Commons vs NCBI GenBank - harmonized clinical-molecular case data versus a flat sequence archive; the pair brackets patient-anchored versus molecule-first design.
• The NCI GDC Data Portal source profile - program scope, release behavior and dictionary structure of the underlying portal.
• GWAS Catalog data - variant-trait associations to layer onto per-case genotype evidence.
• ClinicalTrials.gov data - trial-level context for the same cancers the commons covers at the patient level.
• European Nucleotide Archive data - the European counterpart archive for raw sequencing reads.
• Best biotechnology datasets - the ranked shortlist this record competes on.