Which TCGA cancer types does TransXplorer include?

All 33 TCGA projects, from TCGA-ACC (adrenocortical carcinoma) to TCGA-UVM (uveal melanoma). The data are pre-loaded on the server, so there is nothing to download before you start.

How does TransXplorer split patients into high and low expression groups for survival analysis?

At the median. TransXplorer keeps one primary tumor sample per patient (normal, metastatic and recurrent samples are excluded), and patients whose TPM for the chosen gene is above the median across those primary tumors are labelled HIGH; the rest are labelled LOW. The cutoff is fixed in advance rather than searched for, which avoids the inflated false-positive rate of 'optimal cutpoint' searches.

How do I read the hazard ratio in the TransXplorer survival plot?

The Cox model uses the HIGH group as the reference, so the hazard ratio describes the LOW-expression group relative to the HIGH-expression group. A hazard ratio above 1 means LOW-expression patients die at a higher rate; below 1 means they die at a lower rate. Always check the 95% confidence interval: if it includes 1, the difference is not convincing.

Why can't I make a tumor vs normal box plot for some cancers?

Nine of the 33 TCGA RNA-seq cohorts (ACC, DLBC, LAML, LGG, MESO, OV, TGCT, UCS and UVM) contain no Solid Tissue Normal samples, and several others contain only one to five. Without normal samples there is nothing to compare against, so the box plot stops with an 'Insufficient samples' message.

Is normal adjacent tissue the same as healthy tissue?

No. TCGA normal samples are usually taken from tissue next to the tumor in cancer patients. Such tissue can carry field effects and inflammation from the nearby tumor, so it sits between truly healthy tissue and tumor. Treat tumor-vs-normal differences as tumor-vs-adjacent differences.

Does a significant survival result mean my gene drives cancer progression?

No. A survival association is a correlation in observational data. Expression can track tumor subtype, stage, purity, or treatment rather than cause outcome. Confirm the association in an independent cohort, adjust for known clinical factors, and test mechanism experimentally.

Working with TCGA Cancer Data: Survival & Biomarker Tutorial

TL;DR · The short version

1 TCGA is the largest public collection of tumor RNA-seq linked to patient outcomes: 33 cancer types, each a project such as TCGA-BRCA or TCGA-LUAD. TransXplorer has all 33 pre-loaded under the TCGA Data Analysis menu.
2 Two questions, two pages. Gene Expression Analysis answers “is my gene different in tumors, and does it track with survival?” (box plot and Kaplan–Meier, median split). RNA-seq Analysis answers “what changes genome-wide between tumor and normal?” (DEG plus pathway enrichment).
3 The plots are quick to make and easy to over-read. Many cohorts have few or no normal samples. “Normal” means tissue next to the tumor, not healthy tissue. Survival associations are correlations. Read the caveats in Section 5 before you put a TCGA figure in a paper.

Section 1

What TCGA is, and why it’s worth your time

Eleven thousand tumors, one consistent pipeline, outcomes attached.

Say you’ve found a gene that changes in your cell line or mouse model. The next question from a reviewer, or from you, is usually: does this matter in human patients? The Cancer Genome Atlas is often the fastest way to get a first answer.

TCGA was a US National Cancer Institute and National Human Genome Research Institute program that profiled primary tumors from thousands of patients across 33 cancer types. Each sample was characterised with several technologies: RNA-seq, DNA sequencing, copy number, methylation, and others. Clinical information, including survival follow-up, came with each patient.¹ The data now live at the NCI Genomic Data Commons (GDC), which reprocessed everything through a single harmonised pipeline.²

Each cancer type is a project with a short code: TCGA-BRCA is breast invasive carcinoma, TCGA-LUAD is lung adenocarcinoma, TCGA-GBM is glioblastoma. TransXplorer offers all 33 in every TCGA dropdown. The full list is below so you can find your cancer’s code.

Projects: 33TCGA-ACC … TCGA-UVM
RNA-seq samples: 11,499across all 33 cohorts
Primary tumors: 10,152sample type 01
Normal samples: 740solid tissue normal, type 11

Counts are tallied from the TCGA expression files TransXplorer uses for its gene-level analyses. The remaining samples are metastatic, recurrent, blood-derived (LAML) or additional primaries.

ACCAdrenocortical carcinoma
BLCABladder urothelial carcinoma
BRCABreast invasive carcinoma
CESCCervical squamous cell carcinoma & endocervical adenocarcinoma
CHOLCholangiocarcinoma
COADColon adenocarcinoma
DLBCDiffuse large B-cell lymphoma
ESCAEsophageal carcinoma
GBMGlioblastoma multiforme
HNSCHead & neck squamous cell carcinoma
KICHKidney chromophobe
KIRCKidney renal clear cell carcinoma
KIRPKidney renal papillary cell carcinoma
LAMLAcute myeloid leukemia
LGGBrain lower grade glioma
LIHCLiver hepatocellular carcinoma
LUADLung adenocarcinoma
LUSCLung squamous cell carcinoma
MESOMesothelioma
OVOvarian serous cystadenocarcinoma
PAADPancreatic adenocarcinoma
PCPGPheochromocytoma & paraganglioma
PRADProstate adenocarcinoma
READRectum adenocarcinoma
SARCSarcoma
SKCMSkin cutaneous melanoma
STADStomach adenocarcinoma
TGCTTesticular germ cell tumors
THCAThyroid carcinoma
THYMThymoma
UCECUterine corpus endometrial carcinoma
UCSUterine carcinosarcoma
UVMUveal melanoma

Why this matters

TCGA is a discovery and validation resource, not a clinical trial. Patients weren’t randomised, treatments varied, and follow-up is uneven between cohorts. That makes it excellent for asking “is this pattern present in human tumors?” and poor at answering “does this gene cause worse outcomes?” Keep that distinction in mind and every plot below will be easier to interpret.

Section 2

Reading a TCGA barcode

Two digits tell you whether a sample is tumor or normal.

Every TCGA sample has a barcode, and you’ll see them as column names in downloaded tables. They look cryptic, but they are a fixed sequence of fields, each separated by a hyphen. You only need to recognise two of them: the participant (which patient) and the sample type (what kind of tissue).

Anatomy of a TCGA aliquot barcode

fig 2.1

The two fields that matter most. The first 12 characters (TCGA-BH-A18J) identify the patient, and that is how an RNA-seq sample is linked to survival data. The two digits after it give the sample type: 01 is a primary solid tumor and 11 is solid tissue normal. Both example barcodes are real TCGA-BRCA RNA-seq aliquots.

TransXplorer reads these sample types for you. In the RNA-seq analysis, samples are grouped by their sample-type label (“Primary solid Tumor”, “Solid Tissue Normal”, and so on), and the primary tumor group is tested against Solid Tissue Normal as the reference. The single-gene box plot keeps only Primary solid Tumor and Solid Tissue Normal samples. So you rarely have to parse barcodes yourself, but recognising -01 vs -11 lets you sanity-check anything you download.

Why this matters

One patient can contribute more than one sample: a tumor and a matched normal, or two tumor vials (01A, 01B). The barcode is how you notice. TransXplorer’s survival analysis handles this for you: it keeps only primary tumor samples (01, or 03 for primary blood-derived cancers such as LAML) and the first aliquot per 12-character patient ID, so no patient is counted twice. If you build your own survival table from downloaded data, deduplicate on the patient ID first.

Section 3

Where TCGA lives in TransXplorer

One navbar menu, two pages, five outputs.

Open TransXplorer and look for TCGA Data Analysis in the top navigation bar. It’s a dropdown with two pages, and each one answers a different kind of question:

The TCGA workflow map in TransXplorer

fig 3.1

Rule of thumb: start on the left for a quick look at a gene you already care about. Move to the right when you want an unbiased, genome-wide view of what distinguishes tumor from normal in a cohort.

The single-gene tools read a compact, pre-computed table of TPM values per cohort, so a box plot or survival curve usually appears within seconds. The genome-wide analysis loads the full matrix of GDC read counts for the cohort and runs a complete differential expression model. For the large cohorts (BRCA has more than 1,200 RNA-seq samples) it takes noticeably longer, and a progress bar keeps you informed.

Section 4

Worked example: estrogen receptor and HER2 in breast cancer

Five steps through TCGA-BRCA, from one gene to genome-wide pathways.

For a first case study you want genes whose biology is already well established, so you can tell whether the tool is telling you something sensible. Breast cancer gives us two.

ESR1 encodes estrogen receptor alpha. Most breast tumors are ER-positive and express it strongly, while a minority, including most basal-like tumors, express very little.³ ERBB2 encodes HER2. It is amplified and heavily overexpressed in a subset of breast cancers, and that amplification was linked to relapse and survival decades ago.⁴ Both patterns are “known answers” you can check against what TransXplorer shows you.

Step 1 · ~2 minutes

Is the gene different in tumors? The box plot

TCGA Data Analysis› Gene Expression Analysis› Box Plot Analysis

In the sidebar, set Select Cancer Type to Breast invasive carcinoma (BRCA) and type ESR1 into Enter Gene Symbol. The page opens with COAD and EGFR filled in as an example, so change both. Click Analyze Expression.

TransXplorer keeps only the Primary solid Tumor and Solid Tissue Normal samples and plots expression as log2(TPM + 1). It also runs a two-sample t-test. R’s default is the Welch version, which doesn’t assume equal variances, and that matters here because tumors are far more variable than normals. The Detailed Statistics tab prints the full test output. The sidebar’s Analysis Summary shows how many samples ended up in each group.

Box plot anatomy — ESR1 in TCGA-BRCA

fig 4.1 · real quartiles

Box positions are real: medians, quartiles and whisker ends come from the same TCGA-BRCA TPM file TransXplorer reads. The individual dots are a schematic subsample, because 1,224 real points would overlap into a solid smear. Fill colours match the app (light blue normal, salmon tumor).

What the real numbers say

We ran the box plot calculation on the TCGA-BRCA data TransXplorer uses (1,111 primary tumors, 113 solid tissue normals). Here is what comes out:

Gene	Normal median	Tumor median	Tumor 5th – 95th pct	Normal 5th – 95th pct	Welch t-test p
ESR1	5.32	6.56	0.61 – 8.97	2.65 – 6.93	1.04e-05
ERBB2	6.38	6.94	4.92 – 10.72	3.16 – 7.39	2.91e-13

All values are log2(TPM + 1). Computed from the TCGA-BRCA expression file used by the app; if the underlying data release changes, numbers in the app may shift slightly.

Look past the p-values. For ERBB2 the medians differ by only about half a log2 unit, so the typical tumor is barely different from normal. The interesting part is the top of the tumor range: the 95th percentile sits at 10.7, more than three log2 units (roughly 10-fold) above anything common in normal tissue. That tail is the HER2-amplified subset. For ESR1 the important feature is the spread. Tumors run from almost zero to well above normal, which is what you’d expect from a mix of ER-positive and ER-negative cancers.

Why this matters

With more than a thousand tumors, almost any real difference gets a tiny p-value. In a cohort this large, the p-value mostly tells you the sample size was big. To judge whether a difference matters biologically, look at the size of the shift and the shape of the distribution: medians, spread, tails. Sometimes the biomarker signal is in a subgroup, not in the average.

Step 2 · ~5 minutes

Does it track with survival? The Kaplan–Meier curve

TCGA Data Analysis› Gene Expression Analysis› Survival Analysis

Switch to the Survival Analysis sub-tab. It opens on BRCA. Enter your gene in Enter Gene Symbol (the default is EGFR) and click Analyze Survival. Results appear in three tabs: Survival Plot, Expression Distribution, and Detailed Statistics.

How patients are split: the median, fixed in advance

Survival curves compare groups, so a continuous expression value has to be turned into categories. TransXplorer uses the simplest defensible rule. It takes the gene’s TPM values across the cohort’s primary tumors, one sample per patient (normal, metastatic and recurrent samples are left out first), finds the median, and labels patients above it HIGH and the rest (at or below it) LOW. In TCGA-BRCA that gives 1,033 patients with survival data (151 events). The Expression Distribution tab shows exactly this per-patient set: a histogram of TPM on a log10 axis with the median drawn as a dashed red line.

The median split, as shown in the Expression Distribution tab

fig 4.2 · illustrative

Why the median? It is chosen before looking at survival, so it can’t be tuned to produce a nice p-value, and it gives two groups of equal size, which maximises statistical power for a two-group comparison. The cost is a cruder grouping: a gene that matters only in its top 10% of expressers (like the ERBB2 tail) gets diluted.

How to read a Kaplan–Meier curve

A Kaplan–Meier curve⁵ answers one question: of the patients in this group, what fraction are still alive at each point in time? Every curve starts at 1.0 (everyone alive at diagnosis) and can only step down. The clever part is how it handles patients whose follow-up ended while they were still alive. They are censored: they count toward the denominator for as long as they were followed, then drop out quietly. They are never counted as deaths.

How to read a Kaplan–Meier survival plot

fig 4.3 · illustrative data

Illustrative curves, not real data. The layout mirrors TransXplorer’s survival plot: time in days on the x-axis, HIGH and LOW curves with 95% confidence bands and censor ticks, the hazard ratio and log-rank P in the subtitle, and a caption giving the number of patients and events.

Read the plot in this order:

Check the caption first. It reads “n = … patients ( … events)”. An event is a death. The statistics run on events, not patients, so 1,000 patients with 40 deaths is a much weaker analysis than it sounds. The sidebar table gives the size of each group.
Look at the separation and when it happens. Curves that split early and stay apart tell a different story from curves that only diverge at the far right, where few patients remain and the confidence bands balloon.
Median survival is the time at which a curve crosses 0.50. If a curve never gets there, as in many indolent cohorts, the median is “not reached”. That means most patients outlived their follow-up, not that they are immortal.
The log-rank p-value tests whether the two curves differ anywhere over the whole follow-up. It says nothing about how big the difference is.
The hazard ratio (HR) from the Cox model⁶ gives the size. It compares the instantaneous death rate in one group with the other.

Read the hazard ratio the right way round

In TransXplorer’s Cox model the HIGH group is the reference, so the reported HR describes LOW relative to HIGH. HR > 1: patients with low expression die at a higher rate, so high expression goes with better survival. HR < 1: low expressers do better, so high expression goes with worse survival. The Detailed Statistics tab prints the full Cox summary; its coefficient is labelled strataLOW, which confirms the direction.

Then check the 95% confidence interval. If it includes 1.0 (for example 0.9–1.7), the data are compatible with no difference at all, however suggestive the curves look.

For ESR1 and ERBB2 in BRCA, run the analysis yourself and apply the checklist: how many events, which direction is the HR, does the confidence interval cross 1? Breast cancer is a good teaching cohort precisely because its overall survival events are relatively few for its size (about 200 of the 1,231 BRCA RNA-seq samples come from patients recorded as deceased), so intervals are wider than the sample size suggests. Remember too that ER-positive and HER2-positive patients receive targeted therapies, which reshape their outcomes. A survival curve in TCGA reflects biology and the treatment era.

Step 3 · the longest step

What changes genome-wide? Tumor-vs-normal DEG

TCGA Data Analysis› RNA-seq Analysis

Single-gene questions are hypothesis-driven. The RNA-seq Analysis page is the hypothesis-free counterpart: it tests every gene in the cohort at once for a difference between tumor and normal. The sidebar has three numbered sections and one button.

Project IDTCGA-BRCA

Any of the 33 cohorts; BRCA is pre-selected. Experimental strategy is fixed to RNA-Seq.
Analysis MethodedgeR (QL)

Choose limma-voom, edgeR (QL) or DESeq2 (the default). All three are standard count-based models.^7–9 We use edgeR’s quasi-likelihood test in this walkthrough. Our differential expression guide compares the three.
Normalization MethodTMM

For limma-voom and edgeR: TMM (recommended), RLE or Upper Quartile. DESeq2 always uses its built-in median-of-ratios normalisation.
Adj P-Value Threshold0.05

Cut-off on the multiple-testing-adjusted p-value (false discovery rate).
Log2FC Threshold2

Minimum absolute log2 fold change, so the default of 2 means at least a 4-fold change. A gene counts as up- or downregulated only if it passes both thresholds. With hundreds of samples, the fold-change cut-off does most of the filtering.

Click Run Analysis. TransXplorer loads the cohort’s raw read counts, removes genes with too few counts to test, groups samples by sample type, and fits the model you picked to test the primary tumor group against Solid Tissue Normal as the reference (for BRCA, Primary_solid_Tumor vs Solid_Tissue_Normal), whichever method you choose. A positive log2FC therefore means “higher in primary tumor than in normal”. Results fill four tabs:

Dataset Info: summary cards (genes analysed, up, down, total samples), plus Vital Status, Sample Types and Analysis Parameters tables (the last has a Comparison row, e.g. Primary_solid_Tumor vs Solid_Tissue_Normal, above the Reference row), a summary line such as “Project: TCGA-BRCA | Primary_solid_Tumor vs Solid_Tissue_Normal | N significant DEGs identified”, an MA plot, and an interactive PCA of the 1,000 most variable genes.
DEG Results: sortable tables for All DEGs, Upregulated and Downregulated. Click a gene row to open its OpenTargets information.
Data Visualization: an interactive Volcano Plot with its own fold-change and p-value thresholds, and a Heatmap (Z-scored by row) of the top N genes by p-value, by fold change, or significant DEGs only, with a choice of palette and clustering method.
Pathway Analysis: the next step.

Check the Sample Types table before interpreting

Cohorts don’t always contain just two sample types. TCGA-BRCA also has a handful of metastatic samples, and several cohorts (GBM, LGG, OV, LIHC and others) include recurrent tumors. The DE test still compares only the primary tumor group with Solid Tissue Normal, but nine cohorts have no normal samples at all, and the analysis needs at least two groups. Always open Dataset Info → Sample Types to see which groups are present and how many normals there are, and check the Comparison and Reference rows in Analysis Parameters, before you read the DEG list.

Step 4 · ~2 minutes

What biology do the DEGs point to? Pathway enrichment

RNA-seq Analysis› Pathway Analysis

Once the DEG run has finished, the Pathway Analysis tab unlocks. On the left, tick the databases you want: KEGG Pathways, GO Biological Process and Reactome Pathways are pre-selected, and GO Molecular Function, GO Cellular Component, WikiPathways, MSigDB Hallmark and BioCarta are available. On the right, choose the Gene Set to Analyze (both up and down, up only, or down only), a P-value Cutoff (default 0.05) and Min Genes per Pathway (default 3). Then click Run Pathway Analysis.

This is over-representation analysis (ORA). It asks whether your DEG list contains more members of a pathway than chance would predict. For GO, KEGG, Reactome and WikiPathways, TransXplorer uses clusterProfiler¹⁰ with Benjamini–Hochberg correction and, importantly, uses every gene that was tested as the background, not the whole genome. MSigDB Hallmark and BioCarta are queried through the Enrichr service. Results come as a Results Table (filter by database, sort by adjusted p-value, combined score or gene count), a Dot Plot, a Bar Plot, and a pathway Network.

Split up- and downregulated genes into separate runs when you can. In a tumor-vs-normal comparison, “both” tends to mix cell-cycle programmes switched on in the tumor with tissue-specific programmes switched off, and the mixed list is harder to interpret. For a ranked, threshold-free alternative, see GSEA vs ORA.

Step 5 · ~1 minute

Take it with you: downloads

Every TCGA output has its own download button:

Where	Button	You get
DEG Results › All DEGs	Download All Results	CSV of every tested gene (log2FC, p-values, Ensembl ID, gene symbol, gene type)
DEG Results › Up / Down	Download Upregulated / Downregulated	CSV of genes passing both thresholds
Data Visualization › Volcano Plot	Download PNG	Volcano PNG with the top genes labelled (set by “Labels in Download”)
Data Visualization › Heatmap	Download	Heatmap PNG
Pathway Analysis › Results Table	Download Results	CSV of enriched terms
Pathway Analysis › Dot / Bar / Network	Download Plot / Download	Self-contained interactive HTML
Box Plot Analysis	Download Plot	Box plot PNG, 300 dpi
Survival Analysis	Download Plot	Survival plot and expression-distribution PNGs, 300 dpi

The DEG CSV is the most reusable of these. You can open it in a spreadsheet, feed it into your own tools, or compare it with results from your own experiment (see bringing your own data).

Section 5

Caveats: how TCGA results go wrong

Read this before a TCGA panel goes into Figure 1.

TCGA plots are fast to make, look authoritative, and are easy to over-interpret. Most problems come from a few predictable sources.

Caveat 1 · Many cohorts have few or no normal samples

TCGA set out to profile tumors. Normal tissue was collected when it was available, not systematically. Across the 33 RNA-seq cohorts, only 740 of 11,499 samples are solid tissue normal, and they are very unevenly spread:

Solid tissue normal samples per TCGA cohort

fig 5.1 · real counts

Counts tallied from the per-cohort expression files TransXplorer reads. Nine cohorts have zero normals, and seven more have five or fewer. For those, the box plot either stops with an “Insufficient samples” message or rests on a handful of normal samples, and a genome-wide tumor-vs-normal DEG isn’t possible or isn’t trustworthy. Note that SKCM is mostly metastatic samples (368 of 473), and LAML is a blood cancer with no solid tissue at all.

Caveat 2 · “Normal” is normal adjacent tissue, not healthy tissue

TCGA normals are usually cut from tissue next to the tumor in the same patient. A pan-cancer comparison of normal-adjacent, tumor and truly healthy tissue (from GTEx donors) found that normal-adjacent tissue forms a distinct intermediate state. It shows inflammatory and wound-response signals that healthy tissue lacks.¹¹ In practice, a tumor-vs-normal difference in TCGA is a tumor-vs-adjacent difference. Some changes you care about can be muted, and some are induced in the adjacent tissue itself.

Caveat 3 · Know which survival endpoint you’re using

TransXplorer’s survival analysis uses overall survival. A patient recorded as deceased has an event at days to death; everyone else is censored at days to last follow-up. Overall survival is unambiguous, but it counts deaths from any cause, and in cohorts with good prognosis and short follow-up (thyroid, prostate, testicular) there are very few events. A curated pan-cancer clinical resource, TCGA-CDR, assessed which endpoints are reliable cohort by cohort. It urges caution with overall survival where events are few or follow-up is short, and it recommends progression-free interval as a generally robust alternative.¹² Check the events count in the plot caption before you trust a curve.

Caveat 4 · Cutpoint fishing and multiple testing

Splitting at the median is a choice made in advance. Many web tools, and many papers, instead try every possible cutoff and report the one with the smallest p-value. That “optimal cutpoint” approach badly inflates false positives: with enough cutoffs, even a gene unrelated to survival will produce a striking curve at some threshold.¹³

Why searching for the “best” cutoff misleads

fig 5.2 · illustrative

Illustrative curve. When the cutoff is scanned, the log-rank p-value jumps around by chance. Reporting the single best-looking cutoff converts noise into a “prognostic biomarker”. A pre-specified median, as TransXplorer uses, avoids this, but only if you don’t then go looking at 20 genes and 33 cohorts and report the one that worked.

That last point is the same problem one level up. Testing one gene in 33 cohorts is 33 tests; at p < 0.05 you’d expect one or two “hits” by chance alone. If you screen many genes or cohorts, say how many you tested and correct for it (a Bonferroni or FDR adjustment across your screen), or treat the screen as hypothesis-generating and confirm in an independent dataset.

More traps, quickly

Trap 5

Correlation is not causation

Expression can track tumor subtype, grade, stage or proliferation rather than drive outcome. In breast cancer, many genes “predict survival” simply because they mark the luminal vs basal split.

Fix — ask whether the gene adds anything beyond known subtypes and stage, and test mechanism experimentally.

Trap 6

Tumor purity

A bulk tumor sample is a mix of cancer cells, stroma and immune cells. A gene that looks “down in tumor” may just be expressed by normal epithelium that the tumor has displaced.

Fix — check which cell types express the gene before giving a mechanistic story.

Trap 7

Unpaired comparison of paired data

Some normals come from the same patients as tumors. The box plot’s t-test treats every sample as independent, which is a reasonable first look but not a paired analysis.

Fix — for a publication claim, consider a matched-pairs analysis on the patients with both samples.

Trap 8

Treatment era and confounders

TCGA patients were treated over many years with different therapies. Survival differences can reflect who received which treatment.

Fix — treat univariable KM results as a starting point; adjust for stage and age in a multivariable model before calling a gene prognostic.

Before you publish a TCGA panel: a checklist

I checked the Sample Types table and know exactly which groups were compared.
The normal group has enough samples to be meaningful, and I described it as normal adjacent tissue.
I reported effect sizes (fold change, hazard ratio with 95% CI), not just p-values.
I know which direction the hazard ratio runs (LOW relative to HIGH in TransXplorer).
I reported the number of events, and the cohort has enough of them for overall survival to be sensible.
The cutoff was fixed in advance, and I disclosed how many genes and cohorts I looked at.
I treat the result as an association and have a plan to validate it independently.

TCGA is best at telling you a pattern exists in human tumors. It is weakest at telling you why.

Section 6

What’s next

Bring your own data, try another workflow, or go deeper on the methods.

Keep exploring

TCGA is usually the second half of a story: you find something in your own experiment, then ask whether it shows up in patients. The sibling tutorials cover the first half, whether you’re starting from a count table, a public GEO dataset or raw sequencing reads.

Analysing Your Own Data

Format a counts table and metadata, upload them, and get the same DEG and pathway outputs for your experiment.

Read →

From GEO to Results

Pull a published dataset straight from its GSE accession and reanalyse it end to end.

Read →

From FASTQ to Results

Raw reads from the sequencing core? Quantify with Salmon and continue straight into the analysis.

Read →

Your First RNA-Seq Analysis

New to RNA-seq altogether? The beginner walkthrough on a clean two-group dataset.

Read →

Go deeper on the concepts

What is Differential Expression?

limma-voom vs edgeR vs DESeq2, dispersion, and how adjusted p-values are calculated.

Read →

Understanding Batch Effects

TCGA samples came from many hospitals and sequencing plates; here’s why that matters.

Read →

GSEA vs ORA

Why the pathway step above is ORA, and when a ranked GSEA would tell you more.

Read →

WGCNA: A Visual Guide

Beyond single genes: co-expressed modules and how they relate to clinical traits.

Read →

References

Primary sources for the data and methods discussed on this page.

Cancer Genome Atlas Research Network, Weinstein JN, et al. The Cancer Genome Atlas Pan-Cancer analysis project. Nat Genet. 2013;45:1113–1120. doi:10.1038/ng.2764
Grossman RL, Heath AP, Ferretti V, et al. Toward a shared vision for cancer genomic data. N Engl J Med. 2016;375:1109–1112. doi:10.1056/NEJMp1607591
Cancer Genome Atlas Network. Comprehensive molecular portraits of human breast tumours. Nature. 2012;490:61–70. doi:10.1038/nature11412
Slamon DJ, Clark GM, Wong SG, et al. Human breast cancer: correlation of relapse and survival with amplification of the HER-2/neu oncogene. Science. 1987;235:177–182. doi:10.1126/science.3798106
Kaplan EL, Meier P. Nonparametric estimation from incomplete observations. J Am Stat Assoc. 1958;53:457–481. doi:10.1080/01621459.1958.10501452
Cox DR. Regression models and life-tables. J R Stat Soc Series B. 1972;34:187–220. doi:10.1111/j.2517-6161.1972.tb00899.x
Law CW, Chen Y, Shi W, Smyth GK. voom: precision weights unlock linear model analysis tools for RNA-seq read counts. Genome Biol. 2014;15:R29. doi:10.1186/gb-2014-15-2-r29
Robinson MD, McCarthy DJ, Smyth GK. edgeR: a Bioconductor package for differential expression analysis of digital gene expression data. Bioinformatics. 2010;26:139–140. doi:10.1093/bioinformatics/btp616
Love MI, Huber W, Anders S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 2014;15:550. doi:10.1186/s13059-014-0550-8
Yu G, Wang LG, Han Y, He QY. clusterProfiler: an R package for comparing biological themes among gene clusters. OMICS. 2012;16:284–287. doi:10.1089/omi.2011.0118
Aran D, Camarda R, Odegaard J, et al. Comprehensive analysis of normal adjacent to tumor transcriptomes. Nat Commun. 2017;8:1077. doi:10.1038/s41467-017-01027-z
Liu J, Lichtenberg T, Hoadley KA, et al. An integrated TCGA pan-cancer clinical data resource to drive high-quality survival outcome analytics. Cell. 2018;173:400–416. doi:10.1016/j.cell.2018.02.052
Altman DG, Lausen B, Sauerbrei W, Schumacher M. Dangers of using “optimal” cutpoints in the evaluation of prognostic factors. J Natl Cancer Inst. 1994;86:829–835. doi:10.1093/jnci/86.11.829
Colaprico A, Silva TC, Olsen C, et al. TCGAbiolinks: an R/Bioconductor package for integrative analysis of TCGA data. Nucleic Acids Res. 2016;44:e71. doi:10.1093/nar/gkv1507

TransXplorer’s TCGA module is built on the Bioconductor ecosystem, including TCGAbiolinks¹⁴ for working with GDC data.

If TransXplorer helps your research, please cite us

The preprint is open access on bioRxiv. A peer-reviewed version is in submission.

Verma VM, Oler E, Syed H, Han S, Berjanskii M, Mason AL, Wishart DS, Wong GK. TransXplorer: An automated translational discovery platform for RNA-seq data. bioRxiv. 2026. doi:10.64898/2026.05.15.724657

Read the preprint →

Frequently asked questions

Which TCGA cancer types are available?

All 33 TCGA projects, from TCGA-ACC to TCGA-UVM (full list in Section 1). They are pre-loaded on the server, so there’s nothing to download before you start.

How are the HIGH and LOW groups defined?

By the median. Using one primary tumor sample per patient (normal, metastatic and recurrent samples are excluded), patients whose TPM for your gene is above the median are HIGH; the rest are LOW. The cutoff is fixed in advance rather than optimised, which keeps the p-value honest. The Expression Distribution tab shows the histogram with the median marked.

My hazard ratio is 0.7. Is high expression good or bad?

In TransXplorer the HR compares LOW to HIGH (HIGH is the reference). An HR of 0.7 means low expressers have about 30% lower hazard of death, so high expression goes with worse survival. Check that the 95% confidence interval doesn’t include 1 before reading much into it.

Why do I get “Insufficient samples” for the box plot?

The cohort has no Solid Tissue Normal samples. That applies to ACC, DLBC, LAML, LGG, MESO, OV, TGCT, UCS and UVM. Several others have only one to five normals, which is too few for a meaningful comparison.

Is TCGA “normal” the same as healthy tissue?

No. It is mostly normal tissue adjacent to the tumor, which carries its own inflammatory and field-effect signals. Describe it as normal adjacent tissue in figures and text.

Does a significant survival curve mean my gene drives progression?

No. It is an association in observational data. Expression may simply mark a subtype, stage or treatment group. Validate in an independent cohort, adjust for clinical covariates, and test mechanism experimentally.

Can I cite TransXplorer in a paper?

Yes, please do. Preprint: doi:10.64898/2026.05.15.724657. Please also cite TCGA itself and the analysis methods you report.

Working with TCGA Cancer Data

What TCGA is, and why it’s worth your time

Reading a TCGA barcode

Where TCGA lives in TransXplorer

Worked example: estrogen receptor and HER2 in breast cancer

Is the gene different in tumors? The box plot

What the real numbers say

Does it track with survival? The Kaplan–Meier curve

How patients are split: the median, fixed in advance

How to read a Kaplan–Meier curve

What changes genome-wide? Tumor-vs-normal DEG

What biology do the DEGs point to? Pathway enrichment

Take it with you: downloads

Caveats: how TCGA results go wrong

Caveat 1 · Many cohorts have few or no normal samples

Caveat 2 · “Normal” is normal adjacent tissue, not healthy tissue

Caveat 3 · Know which survival endpoint you’re using

Caveat 4 · Cutpoint fishing and multiple testing

More traps, quickly

Correlation is not causation

Tumor purity

Unpaired comparison of paired data

Treatment era and confounders

Before you publish a TCGA panel: a checklist

What’s next

Keep exploring

Analysing Your Own Data

From GEO to Results

From FASTQ to Results

Your First RNA-Seq Analysis

Go deeper on the concepts

What is Differential Expression?

Understanding Batch Effects

GSEA vs ORA

WGCNA: A Visual Guide

Try it on your gene now

References

If TransXplorer helps your research, please cite us

Frequently asked questions