gabrielaltay/tcga-patients-open

TCGA Patients (Open Access)

Open-access TCGA data from the NCI Genomic Data Commons (GDC). Covers all 33 TCGA projects.

This view presents one HuggingFace subset per TCGA project, with one row per patient. See the tcga-tabular-open companion for a per-table view of the same underlying data.

• Generated: 2026-08-14 01:39:43 UTC
• Schema: derived from the GDC Data Dictionary.
• GDC data release: Data Release 45.0 - December 04, 2025

Data model

Where the data comes from

Three sources feed each project's data, all open-access:

• Case-level clinical structure — fetched from the GDC /cases endpoint, returning the full nested case JSON (demographic + diagnoses → treatments + follow_ups + exposures + family_histories + samples → portions → analytes → aliquots). The biospecimen subtree on each case:

  case          one patient (TCGA-XX-1234)
  └── sample    physical specimen taken from the patient at one timepoint
                (Primary Tumor, Solid Tissue Normal, Blood Derived Normal, ...)
      └── portion    a piece of that sample for a specific lab process
          └── analyte    extracted material of one type (DNA or RNA)
              └── aliquot    a vial of that analyte handed off for sequencing
  

• Per-modality files — discovered via /files (filtered by the clauses in the table below) and downloaded via /data. Each combination locks one data_type to a specific GDC pipeline so a future GDC addition can't quietly substitute a different pipeline under the same data_type. This covers both the molecular modalities and the scanned Pathology Report PDFs, which are carried verbatim — no text extraction is applied, so consumers can run whichever parser they trust against the original document.

• BCR Clinical Supplement biotabs — original Biospecimen Core Resource (BCR) clinical forms shipped as per-project TSVs (one per form: patient, follow_up, nte, drug, radiation, etc.). The harmonized /cases endpoint drops or under-populates a number of clinical fields the BCR-original biotabs preserve. The schema varies by cancer type (e.g. BLCA's BCG-response columns don't exist in CHOL's hepatic-marker forms), so each project's biotabs ship only the columns they actually carry. Discovered the same way (/files then /data) — see the filter table below.

Source data filters (canonical)

Same in both views of the dataset; each row locks the /files query for one source:

data_type	data_format	data_category	experimental_strategy	analysis.workflow_type
Masked Somatic Mutation	MAF	Simple Nucleotide Variation	WXS	Aliquot Ensemble Somatic Variant Merging and Masking
Gene Expression Quantification	TSV	Transcriptome Profiling	RNA-Seq	STAR - Counts
Pathology Report	PDF	Clinical	``	``
miRNA Expression Quantification	``	Transcriptome Profiling	miRNA-Seq	BCGSC miRNA Profiling
Protein Expression Quantification	TSV	Proteome Profiling	Reverse Phase Protein Array	``
Allele-specific Copy Number Segment	TXT	Copy Number Variation	``	``
Masked Copy Number Segment	TXT	Copy Number Variation	``	``
Clinical Supplement	bcr biotab	Clinical

How each source appears in this view

Source	Where it lands
GDC /cases	nested fields on each patient row (demographic, diagnoses, follow_ups, exposures, family_histories, samples); gdc_portal_url link added
Masked Somatic Mutation MAFs	samples_masked_somatic_mutation array on each patient row (sample FKs resolved alongside GDC's aliquot UUIDs)
Gene Expression Quantification	samples_gene_expression_quantification array on each patient row (stranded_first / stranded_second dropped — GDC harmonizes as unstranded)
BCR Clinical Supplements	clinical_supplement struct on each patient row, with sub-fields patient (1 dict) and follow_ups / ntes / drugs / radiations / ablations / omfs (lists of dicts). Sub-fields with no data for the project are omitted.
Pathology Reports	samples_pathology_report array on each patient row; pdf_bytes holds the scanned PDF verbatim, joined to its sample via pathology_report_uuid
Allele-specific Copy Number Segment	samples_allele_specific_copy_number_segment array — one record per (aliquot, caller) with segments as index-aligned arrays; filter on workflow_type
Masked Copy Number Segment	samples_masked_copy_number_segment array — one record per aliquot, log2 ratios in segment_mean
miRNA Expression Quantification	samples_mirna_expression_quantification array — one record per aliquot, ~1,881 miRNAs as index-aligned arrays
Protein Expression Quantification	samples_protein_expression_quantification array — one record per portion (not aliquot), ~487 antibodies
BCR Biospecimen Supplements	biospecimen_supplement struct on each patient row, with list-valued sub-fields (sample, portion, analyte, aliquot, slide, protocol, ssf_*, ...). Sub-fields with no data for the project are omitted.
MSigDB gene sets + RNA-Seq	samples_ssgsea_<collection> array columns — pathway activity per aliquot; see the ssGSEA section below

Specific to this view

• Convenience: each row carries samples_<modality> array columns so you can column-project just the molecular data you need without walking the nested GDC entities.
• Loading: the tcga2hf package ships a typed TcgaHfPatient pydantic model that mirrors this schema and adds convenience joins (tumor/normal pairs, mutations-by-gene, expression-by-gene, longitudinal timeline).

Provenance pinned per build

• GET /status → data_release / tag / commit saved in each project's gdc_status.json.
• GET /v0/submission/_dictionary/_all → schema dictionary snapshot saved alongside the raw data; its SHA-256 is recorded in gdc_status.json.

See the repository for full request payloads, filter clauses, and the build pipeline source.

Survival endpoints (survival_derived)

We have provided a supplement to the GDC source data: re-derived survival endpoints — Overall Survival (OS), Disease-Specific Survival (DSS), Progression-Free Interval (PFI), Disease-Free Interval (DFI) — following the algorithm published by Liu et al. 2018 (DOI 10.1016/j.cell.2018.02.052).

Each patient row carries a top-level survival_derived struct with eight sub-fields: os_event / os_time, dss_event / dss_time, pfi_event / pfi_time, dfi_event / dfi_time. *_event is 0/1 (event observed vs censored); *_time is days from index_date (TCGA: diagnosis date). DFI is null for SKCM / THYM / UVM / LAML — Liu specifies no DFI for those tumor types.

We've reimplemented Liu's method against the current TCGA data and find broad agreement with the original curated CDR. Differences exist and are expected: this is a newer release of the underlying GDC data, so re-curated clinical values, post-2018 patient additions, and schema migrations all contribute to the gap. This work is evolving; see the repository for the full reproduction report and per-endpoint methodology.

Why we don't ship Liu's curated 2018 values directly: the CDR is a frozen 2018 snapshot derived from a since-modified GDC release. Including those values would lock in irreproducible source-data drift. We re-derive on every build, so the values reflect the current GDC and are reproducible from this dataset's other tables alone.

Pathology reports

Scanned surgical pathology reports as GDC serves them — 11,208 reports covering 11,121 cases across 32 projects. Each patient row carries a samples_pathology_report array; pdf_bytes holds the document.

TCGA-LAML has none, which is expected rather than missing: acute myeloid leukaemia has no surgical resection specimen to report on.

The bytes, not a text extraction

The PDFs are carried verbatim, with no text extraction applied. Any parse is specific to the tool and version that produced it, so extracting at publication time would freeze one tool's output into the dataset and lose the original. Shipping the source document means a better parser can be run later without re-downloading from GDC, and a canonical parse — if one is added — becomes an additional clearly-labelled column rather than a replacement.

Practical notes for anyone parsing them:

• These are page scans. Most carry an OCR text layer added upstream of GDC, so a pure-Python extractor returns several hundred to a few thousand characters for nearly every report — but that layer transcribes the barcode strip and handwritten margin notes as noise, and its fidelity varies by submitting institution.
• Patient identifiers are redacted out of the page image by GDC before distribution.

Joining to a sample

Every report links to the sample it describes. The GDC file name is <case_submitter_id>.<REPORT_UUID>.PDF, and that UUID is the same value GDC reports on sample.pathology_report_uuid — a key this dataset has always carried, so reports join to samples without anything new being invented. Where GDC names the sample directly in the file's associated_entities, that is preferred, with the file-name UUID as fallback.

Copy number

Copy number ships at segment level, exactly as GDC serves it, in two array columns on each patient row that answer different questions. Each record covers one assay run, with the per-segment values as index-aligned arrays inside it.

Column	Caller	Measurement	Files
samples_allele_specific_copy_number_segment	ASCAT2, ASCAT3, AscatNGS	Integer total copy number plus its split into major_copy_number / minor_copy_number	23,225
samples_masked_copy_number_segment	DNAcopy	Relative log2(sample / reference) in segment_mean, germline CNVs masked out	22,629

copy_number = major_copy_number + minor_copy_number holds everywhere. minor_copy_number = 0 with major_copy_number > 0 is loss of heterozygosity.

Filter on workflow_type

All three allele-specific callers ship for overlapping aliquots, and each fits tumour purity and ploidy independently, so they can disagree. On TCGA-CHOL, ASCAT2 and ASCAT3 give the same length-weighted modal copy number for 33 of 36 shared aliquots — but where they differ they differ substantially (one aliquot is modal 2 under ASCAT2 and modal 4 under ASCAT3), and ASCAT3 segments far more coarsely (2,469 segments against ASCAT2's 6,580 over the same aliquots). AscatNGS is the WGS-based caller; the other two run on genotyping arrays, recorded in experimental_strategy.

A query that does not filter on workflow_type is pooling three different answers to the same question. ASCAT3 is GDC's current standard.

The two views are not interchangeable

Nesting each masked segment inside its containing ASCAT3 segment on TCGA-CHOL (2,590 pairs) gives Spearman +0.56, with median segment_mean rising monotonically across integer copy number:

copy_number	0	1	2	4	8
median segment_mean	−1.93	−0.44	+0.07	+0.22	+1.25

The correlation is only moderate, and that is a property of the data rather than a defect: ASCAT corrects for purity and ploidy while DNAcopy's ratio is against a diploid reference, so in a hyperdiploid tumour integer copy number 3 is copy-neutral relative to its own baseline yet still reads near log2 0. Use the allele-specific calls for absolute copy number, the masked segments for reference-relative ratio.

One formatting difference is carried through from the source rather than normalized: the allele-specific segments write chr1, and the masked segments write bare 1.

A small tail of over-fragmented masked segments

Most masked files hold 60-100 segments (median 77 in TCGA-LAML, 91 in TCGA-BRCA, 67 in TCGA-CHOL). A handful hold tens of thousands: 32 of 22,629 files (0.14%) exceed 200 KB, the largest carrying 50,780 segments against TCGA-BRCA's per-file maximum of 1,029. They cluster in TCGA-LAML (12), TCGA-BLCA (9) and TCGA-BRCA (7), and the most extreme are all -11A- matched normals.

This is the signature of a noisy genotyping array, where circular binary segmentation fails to merge and emits many tiny spurious calls. It is genuine GDC content and is shipped unmodified, but it is a real trap: an unfiltered query over this table gets a few samples contributing tens of thousands of junk rows each, enough to skew any per-segment aggregate. num_probes is the filter — the spurious segments are supported by very few probes.

Gene-level copy number is deliberately absent

GDC also serves Gene Level Copy Number — the same calls projected onto GENCODE v36 — at roughly 34 GB per workflow. It is not shipped here because it is exactly reproducible from the allele-specific segments rather than being independent evidence. (Verified against GDC's own files on TCGA-CHOL: projecting segments onto the gene model reproduced every gene call with zero mismatches across three aliquots, and for genes straddling a segment boundary GDC's min_copy_number / max_copy_number are the min and max over the overlapping segments.) It may be added later as a clearly-labelled derived table.

miRNA-Seq and protein expression (RPPA)

samples_mirna_expression_quantification

One record per aliquot, holding ~1,881 miRBase v21 mature miRNAs as index-aligned arrays, from 11,441 files across TCGA. read_count is raw; reads_per_million_mirna_mapped is normalized within the aliquot and sums to exactly 1,000,000 per aliquot.

cross_mapped is Y when reads for that miRNA also aligned elsewhere in the genome, so its count is not uniquely attributable. GDC ships the flag rather than dropping the row and so do we; filter it out if you need clean attribution. The source column is spelled cross-mapped — renamed here only because the hyphen is not a legal bare SQL identifier.

Isoform-level quantification (Isoform Expression Quantification, ~4 GB) is not shipped.

samples_protein_expression_quantification

Reverse Phase Protein Array. One record per portion, holding ~487 antibodies as index-aligned arrays. 7,906 files covering 7,827 of 11,428 TCGA cases — the narrowest coverage of any modality here, because RPPA was only run on a subset.

Three things to know before using it:

• It is the only modality that attaches to a portion, not an aliquot, so it carries portion_id and resolves sample_id through the portion.
• The antibody panel grew over the project's life, and set_id records which version a measurement came from. A peptide_target absent for a sample may mean "not on that panel" rather than "measured as zero".
• protein_expression is null where the source says NA — a failed or missing measurement, not a zero. On TCGA-CHOL that is 930 of 14,370 cells (6.5%).

Values are replicate-based normalized log2 signal, centred near 0, and the sign is meaningful. Agreement with matched RNA is modest and positive, as expected for protein-vs-transcript: median Spearman +0.26 across shared targets on TCGA-CHOL.

Biospecimen supplements

The counterpart to the clinical_supplement struct: where that describes the patient, the biospecimen_supplement struct describes the specimen chain — how a tumour got from the operating room to a sequencer, and the pathologist's read on each slide along the way. 340 BCR biotab files across TCGA (~76 MB) covering 11,315 cases, one list-valued sub-field per form. Every sub-field is a list — unlike clinical_supplement, which has a single-dict patient slot — because each of these forms is one-row-per-entity.

Some of it restates what the case structure already nests (sample / portion / analyte / aliquot ids and types). The forms worth reaching for are the ones with no /cases equivalent:

Sub-field	What is in it
slide	Per-slide percent_tumor_nuclei, percent_necrosis, percent_stromal_cells, percent_lymphocyte_infiltration, section_location — the QC layer behind "is this sample actually tumour?", and the standard covariate for purity and deconvolution work
analyte	a260_a280_ratio, concentration, extraction method — nucleic-acid quality, which drives batch effects
protocol, shipment_portion	Plate, shipment and centre each specimen moved through — the raw material for batch-effect analysis
ssf_tumor_samples, ssf_normal_controls	Site-specific factors: the disease-specific pathology fields the pan-cancer clinical schema has no column for
cqcf	The submitting centre's clinical quality control form (TCGA-LUAD only)

Like the clinical supplements these are flex-schema: the column set differs by project and by submitting centre, so the shape is inferred per project rather than padded into a pan-cancer union. Forms with no data for a project are omitted entirely (only TCGA-LUAD has cqcf; only 9 projects have auxiliary). All fields are typed as strings with BCR sentinels like [Not Available] preserved verbatim.

Records are keyed to the patient by BCR barcode. The specimen-level forms are keyed on their own entity and several omit the patient barcode column entirely, in which case it is recovered as the first three groups of the entity barcode (TCGA-3X-AAV9-01A-11D-A42S-01 → TCGA-3X-AAV9) — a property of the TCGA barcode grammar, not a heuristic.

Two submitters ship these files: nationwidechildrens.org for 334 of the 340, and genome.wustl.edu for 6 (all TCGA-LUAD). Where both ship the same form for one project their records are concatenated, and the parquet schema is the union of their columns.

Pathway activity (ssGSEA)

Single-sample gene set enrichment scores for every RNA-Seq aliquot, in one samples_ssgsea_<collection> column per MSigDB collection. Each entry is one scored aliquot, with pathway, pathway_url, the gene counts and score_raw as index-aligned arrays — the same struct-of-arrays shape samples_gene_expression_quantification uses.

pathway_url links to the authoritative MSigDB definition of each gene set — so what a score means is one click away from the score itself.

Collections

Pinned to MSigDB 2026.1.Hs and verified by md5, because gene-set membership changes between MSigDB releases and feeds directly into every score.

collection	contents	file	md5
hallmark	MSigDB Hallmark — 50 coherent, deliberately non-redundant signatures	h.all.v2026.1.Hs.symbols.gmt	367eec875967c2cfbf664a1a065b7b8d
reactome	MSigDB C2:CP:REACTOME — 1,839 canonical pathways	c2.cp.reactome.v2026.1.Hs.symbols.gmt	1516b5d15611415d1996c92b7cb6d1cc
pid	MSigDB C2:CP:PID — 196 NCI-Nature cancer signalling pathways	c2.cp.pid.v2026.1.Hs.symbols.gmt	291508046f73d82d13e5efb47492fa47
oncogenic	MSigDB C6 — 189 oncogenic signatures (oncogene / tumour-suppressor perturbation)	c6.all.v2026.1.Hs.symbols.gmt	aba0e2214ff63327ae3fb0ce4bcd11c2
cancer_cell_atlas	MSigDB C4:3CA — 148 Curated Cancer Cell Atlas meta-programs (single-cell derived)	c4.3ca.v2026.1.Hs.symbols.gmt	ff9902288655ff2ab88fcb5cbc4a95dd

MSigDB is released under CC BY 4.0; some constituent collections carry extra restrictions, so we ship only collections we can redistribute scores from. See the MSigDB licence terms.

Method

Barbie et al. (2009) ssGSEA as implemented by Bioconductor GSVA, transcribed to Python and validated against GSVA 2.6.6 to floating-point noise (Pearson/Spearman 1.0000000000, max relative difference 4.8e-13). alpha=0.25, gene sets filtered to a minimum of 10 genes after mapping onto the expression matrix; no maximum size.

Scored on tpm_unstranded over a gene universe of protein-coding genes plus the functional immunoglobulin / T-cell-receptor segments. That last inclusion matters for tumour-immune biology: GENCODE gives Ig/TCR segments their own biotypes, and without them MSigDB's B-cell-receptor and complement pathways match as little as 8% of their genes. Note that V/D/J segments are somatically rearranged, so their expression reports lymphocyte infiltration rather than regulation of a fixed locus.

Because ssGSEA weights ranks rather than expression values, any strictly monotonic transform of the input leaves scores unchanged — there is no reason to log-transform before scoring.

Why score_raw, and how to normalize

score_raw is the only score column, and it is a property of its own sample: it does not depend on which other samples or gene sets were scored alongside it. GSVA's optional normalization divides by the range of the entire score matrix, which would make every value depend on cohort and collection composition — adding Reactome to a Hallmark run widens that divisor by ~49% on this data, silently restating previously published scores.

The reference distributions needed to normalize live in the tabular view's ssgsea_stats_<collection> tables, which have no counterpart here: they are cohort-level aggregates, and this view is one row per patient. Load them from there to z-score against a population, or to recover GSVA's divisor as MAX(max) - MIN(min) over their pan_cancer rows. The scores themselves are identical in both views.

Loading

from datasets import load_dataset

# One config per TCGA project.
luad = load_dataset("gabrielaltay/tcga-patients-open", "TCGA-LUAD")

Each row is one patient with the full GDC case structure nested in-place plus the survival_derived struct.

GDC references

• Data dictionary (every entity + field definition)
• Biospecimen Encyclopedia
• MAF format spec
• Gene Expression Quantification spec
• Sample Type codes
• TCGA Barcode reference

License & redistribution

Per the NCI GDC Data Analysis Policy:

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:

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:

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

Restrictions on use

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)

Required acknowledgement

If you publish or present results derived from this dataset, include the NCI-required TCGA acknowledgement:

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

Suggested citations:

• 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

Policy references: GDC Policies, GDC Encyclopedia — Controlled Access (defines what is not in this dataset), NIH Genomic Data Sharing Policy.

Disclaimer

This project is not affiliated with the NCI, GDC, or the TCGA Research Network. It is an experimental open-source pipeline that may change significantly between versions. Pipeline source: galtay/tcga2hf.