gabrielaltay/tcga-gene-expression-quantification-open

TCGA Gene Expression Quantification — Open Access

Every open-access TCGA RNA-Seq gene expression file from the NCI Genomic Data Commons (GDC), stacked into matrices: one row per sequenced tube of RNA, called an aliquot, and one column per gene. Every value is exactly as GDC published it.

• Size: 11,505 aliquots × 60,660 genes, from 10,517 patients in 33 TCGA projects
• Source: GDC Data Release 46.0 (2026-08-10)
• Built: 2026-09-23 04:56:55 UTC

The source files

GDC runs every aliquot through its mRNA pipeline: the STAR aligner counts the sequencing reads on each of 60,660 GENCODE v36 genes. Each output file has six numbers per gene, which this card calls measures:

measure	what it is
unstranded	reads counted on the gene
stranded_first, stranded_second	the same reads, split by DNA strand (see Before you model)
tpm_unstranded	transcripts per million: counts adjusted for gene length, scaled so each aliquot sums to 1,000,000
fpkm_unstranded	fragments per kilobase per million: counts adjusted for gene length and sequencing depth
fpkm_uq_unstranded	FPKM with depth taken from the 75th-percentile gene count instead of the total

Four more lines in each file count the reads that no gene received.

Layout

Each measure is its own config, a table you load by name, with one row per aliquot and a values list of 60,660 numbers:

config	dtype	size
unstranded	int32	755 MB
stranded_first	int32	734 MB
stranded_second	int32	741 MB
tpm_unstranded	float32	1,554 MB
fpkm_unstranded	float32	1,326 MB
fpkm_uq_unstranded	float32	1,349 MB

Two more configs index the axes. Row i of aliquots describes row i of every measure config, and row j of genes describes position j of every values list. Measure configs also repeat aliquot_index, aliquot_id, project_id, sample_type and the patient's case_submitter_id, so training needs no join.

Load it

from datasets import load_dataset

REPO = "gabrielaltay/tcga-gene-expression-quantification-open"
ds = load_dataset(REPO, "tpm_unstranded", split="train")

aliquots and genes are in matrix order, so they become an AnnData's obs and var:

import anndata as ad
import numpy as np

obs = load_dataset(REPO, "aliquots", split="train").to_pandas().set_index("aliquot_id")
var = load_dataset(REPO, "genes", split="train").to_pandas().set_index("gene_id")
X = np.stack(ds.with_format("numpy")["values"])   # (11,505, 60,660) float32

adata = ad.AnnData(X=X, obs=obs, var=var)

The examples below use this adata.

Where each row comes from

TCGA tracks material as a tree from patient to sequenced tube. GDC's data dictionary defines five levels:

level	what it is
case	"all data related to a specific subject in the context of a specific project": one patient
sample	material taken from the patient, such as a tumour or adjacent normal tissue
portion	"a physical sub-part of any sample"
analyte	"a liquid bulk product" extracted from a portion; here, RNA
aliquot	a measured volume of analyte; the thing that was sequenced

Every level has a UUID and a barcode. *_id columns hold GDC's UUID and *_submitter_id columns hold the human-readable TCGA barcode. aliquots has both for every level, so labels at any level join directly.

A barcode extends its parent's barcode. For example, TCGA-02-0047-01A is a primary-tumour sample from patient TCGA-02-0047:

TCGA-02-0047	-01	A	-01	R	-1849	-01
case	sample type (codes)	vial	portion	analyte type (codes)	plate	sequencing center (codes)

Join on the ID columns, not on sliced barcodes. Some barcodes in GDC's records do not extend their parent's, and slicing them finds no entity or the wrong one. To list them:

from itertools import pairwise

barcodes = adata.obs[[f"{level}_submitter_id" for level in ["case", "sample", "portion", "analyte", "aliquot"]]]
nests = barcodes.apply(lambda r: all(child.startswith(parent) for parent, child in pairwise(r)), axis=1)
adata.obs.loc[~nests, barcodes.columns]

Rows are not independent. A sample sequenced twice has two rows, and if they share an analyte_id they are one RNA extraction sequenced twice. A patient can also have several samples, such as a tumour and its matched normal.

adata.obs.sample_id.duplicated(keep=False).sum()   # rows from samples sequenced more than once
adata.obs.sample_type.value_counts()               # tumours, normals and other sample types

Before you model

Model the unstranded measures. GDC normalizes only the unstranded counts:

To facilitate harmonization across samples, all RNA-Seq reads are treated as unstranded during analyses.

— mRNA Analysis Pipeline, Introduction

The stranded counts matter only for checking the lab protocol. strand_balance is stranded_first / (stranded_first + stranded_second) over all genes: near 0.5 when the protocol did not record strand, near 0 or 1 when it did.

adata[adata.obs.strand_balance.between(0.4, 0.6)]   # aliquots whose protocol did not record strand

Condition on the share of reads assigned to genes. It varies between aliquots and between projects. The n_* columns count the reads no gene received, taken from the file's unstranded column, so together with the unstranded counts they total every read. TPM and FPKM are normalized per aliquot and cannot show this share.

counts = load_dataset(REPO, "unstranded", split="train").with_format("numpy")
assigned = np.stack(counts["values"]).sum(axis=1)
unassigned = adata.obs[["n_unmapped", "n_multimapping", "n_nofeature", "n_ambiguous"]].sum(axis=1)

fraction = assigned / (assigned + unassigned.to_numpy())

Every gene is kept. No expression or gene-type filter is applied; filter with gene_type:

adata[:, adata.var.gene_type == "protein_coding"]

Gene coordinates come from GDC's gene-level copy number files, which use the same GENCODE v36 annotation. Genes missing from those files have null coordinates.

Versions

GDC files rarely change between releases. GDC versions each file separately, and a new release leaves most files untouched. source_file_id and source_file_md5sum identify the exact file behind each row.

Pin a revision. Each upload to the Hub is a commit, and the card in each commit names its GDC release:

load_dataset(REPO, "tpm_unstranded", revision="<commit sha>")

Check any row against GDC

source_file_url downloads the row's file from GDC's data endpoint, with no login:

import hashlib, io, urllib.request
import pandas as pd

row = adata.obs.iloc[0]
raw = urllib.request.urlopen(row.source_file_url).read()
assert hashlib.md5(raw).hexdigest() == row.source_file_md5sum

tsv = pd.read_csv(io.BytesIO(raw), sep="\t", comment="#").iloc[4:]   # drop the four uncounted-read lines
assert (tsv.gene_id.to_numpy() == adata.var_names).all()
assert (tsv.tpm_unstranded.to_numpy(dtype="float32") == adata.X[0]).all()

A 404 means GDC no longer serves the file.

Data dictionary

Columns not marked computed here are exactly as GDC published them.

aliquots, one row per aliquot:

column	meaning	source
aliquot_index	Row number; row i of every measure config	computed here
aliquot_id	Aliquot UUID	GDC aliquot.id
aliquot_submitter_id	Aliquot barcode	GDC aliquot.submitter_id
analyte_id	UUID of the analyte the aliquot was drawn from	GDC analyte.id
analyte_submitter_id	Analyte barcode	GDC analyte.submitter_id
analyte_type	Kind of molecule extracted, e.g. RNA	GDC analyte.analyte_type
portion_id	UUID of the portion the analyte was extracted from	GDC portion.id
portion_submitter_id	Portion barcode	GDC portion.submitter_id
sample_id	UUID of the sample the portion was cut from	GDC sample.id
sample_submitter_id	Sample barcode	GDC sample.submitter_id
sample_type	Where the sample came from, e.g. Primary Tumor	GDC sample.sample_type
case_id	UUID of the patient	GDC case.id
case_submitter_id	Patient barcode	GDC case.submitter_id
project_id	TCGA project, e.g. TCGA-BRCA	GDC project
source_file_id	UUID of the GDC file this row was read from	GDC gene_expression.id
source_file_md5sum	GDC's MD5 checksum of that file	GDC gene_expression.md5sum
source_file_version	The file's version number	GDC /files/versions
source_file_first_release	First GDC release that served the file	GDC /files/versions
source_file_url	Download URL for the file	computed here
strand_balance	stranded_first / (stranded_first + stranded_second) over all genes	computed here
n_unmapped	Reads that did not align to the genome	GDC expression file
n_multimapping	Reads that aligned to more than one place	GDC expression file
n_nofeature	Aligned reads that overlap no gene	GDC expression file
n_ambiguous	Aligned reads that overlap more than one gene	GDC expression file

genes, one row per gene:

column	meaning	source
gene_index	Position in every values list	computed here
gene_id	Ensembl gene ID with version, e.g. ENSG00000000003.15	GDC expression file
gene_name	Gene symbol	GDC expression file
gene_type	GENCODE gene type, e.g. protein_coding	GDC expression file
chromosome	Chromosome (GRCh38); null where the source omits the gene	GDC gene-level copy number file
start	Start position (GRCh38)	GDC gene-level copy number file
end	End position (GRCh38)	GDC gene-level copy number file

Measure configs (unstranded, tpm_unstranded, ...), one row per aliquot:

column	meaning	source
aliquot_index	Row number, as in aliquots	copied from aliquots
aliquot_id	As in aliquots	copied from aliquots
case_submitter_id	As in aliquots	copied from aliquots
project_id	As in aliquots	copied from aliquots
sample_type	As in aliquots	copied from aliquots
values	One number per gene, in genes order	GDC expression file

GDC references

• mRNA analysis pipeline (STAR counts, TPM / FPKM / FPKM-UQ, GENCODE v36)
• Sample Type codes
• TCGA Barcode reference
• Data dictionary (every entity + field definition)

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.