Examples you can use with Claude or ChatGPT

What data can I use?

Tag.bio holds curated datasets (data products): cancer genomics, a kidney cancer trial, hospital costs and reference data. Each is one big table with one row per patient, sample or record.

Look inside one dataset

Ask what a data product contains and which analyses it can run. The answer lists its columns (collections) and its built-in analyses, which you can then ask for by name.

• 10,967 tumor samples, 30 cancer types, 56 collections: mutations, copy number, RNA expression, MSI, subtypes, survival.
• 28 built-in analyses, including survival, oncoprint, UMAP, co-occurrence, mutation prevalence and pathway alterations.

Count patients by group

The simplest question: how many tumors of each cancer type are in TCGA. The count covers every row, not a sample.

• 10,967 samples in 30 cancer types.
• Largest: breast 1,084, non-small cell lung 1,053, esophagogastric 622, colorectal 594, glioblastoma 592.

Which genes are mutated in which cancers

A driver gene is one whose mutation helps a cancer grow. This heatmap shows, for each cancer type, the share of tumors with each driver mutated.

• TP53 in 91% of ovarian tumors; IDH1 in 77% of gliomas; APC in 73% of colorectal tumors.
• KRAS in 66% of pancreatic tumors; BRAF in 59% of thyroid and 53% of melanoma; PTEN in 61% of endometrial.

Counts protein-altering mutations only; silent and intronic calls would inflate the numbers.

A map of all tumors

UMAP squeezes the activity of thousands of genes into one picture, so tumors that behave alike sit close together. Coloring by cancer type shows tumors grouping by the tissue they came from.

• 10,071 tumors with RNA-seq. 9 of 20 clusters are at least 90% one cancer type (thyroid 99.8%, kidney clear cell 98.5%).
• Squamous tumors from head and neck and lung share a cluster, as do colorectal and stomach adenocarcinomas.

A clinical trial's own data

Tag.bio also holds full trial datasets. JAVELIN Renal 101 compared avelumab + axitinib with sunitinib in kidney cancer. This compares how long the disease stayed under control (progression-free survival) in patients whose tumors make PD-L1.

• Avelumab + axitinib vs sunitinib: hazard ratio 0.54 (95% CI 0.43 to 0.69), p = 5.5e-7.
• Median progression-free survival 13.3 vs 8.2 months (270 vs 290 patients).

This is the data product's refined dataset; the published interim analysis reported HR 0.61.

What drives hospital cost overruns

The Vizient data covers about 79,000 hospital stays. The built-in cost analysis finds which patient groups, diagnoses and hospitals run most over the expected cost.

• Severity is the dominant driver: the sickest patients average $7,811 extra per stay vs $344 for everyone else.
• Hospital B runs $2,080 over per stay across 12,068 stays (about $16.4M).

Uveal melanoma: who could get darovasertib

Almost all uveal (eye) melanomas start with a mutation in GNAQ or GNA11, and darovasertib blocks that signal. Tumors that have lost BAP1 are the most likely to spread, which is why trials of treatment before or after surgery focus on them. The oncoprint shows every tumor's changes and risk group in one picture.

• GNAQ or GNA11 is mutated in 74 of 80 tumors (92%): the darovasertib-selectable group.
• Risk groups: BAP1 loss 44, SF3B1 12, EIF1AX 9, other 15.

MTAP-deleted tumors: a new drug class

Some tumors have lost both copies of a gene called MTAP, usually together with its neighbor CDKN2A. That loss makes them vulnerable to a new kind of drug (PRMT5 inhibitors) that spares normal cells. How common is the loss in each cancer?

• 978 of 10,712 tumors (9.1%) have lost MTAP; 98% of them have lost CDKN2A too.
• Glioblastoma 42%, mesothelioma 32%, bladder 26%, pancreatic 22%, non-small cell lung 16%.

KRAS is not one mutation

"KRAS-mutant" covers several different changes: G12D, G12V, G12C and others. New drugs target one specific version, so which one a patient has decides which trial they fit.

• G12D leads in pancreatic (28% of tumors) and colorectal (11%); G12C leads in lung adenocarcinoma (12%).
• Overall KRAS-mutant: pancreatic 66%, colorectal 41%, lung adenocarcinoma 30%.

Clinical sequencing finds KRAS in about 90% of pancreatic cancers; TCGA's pancreatic samples contain few tumor cells, so some mutations are missed.

Lung cancer that resists immunotherapy

Lung tumors with STK11 or KEAP1 mutations often respond poorly to standard immunotherapy, especially alongside KRAS. Trials test stronger combinations for exactly these patients. How often do these occur together, and do those patients fare worse?

• KRAS 30%, STK11 13%, KEAP1 18% of 562 tumors; STK11 occurs with KRAS far more than by chance (odds ratio 2.9).
• STK11-mutated patients live shorter (hazard ratio 1.59, p = 0.02); KEAP1 shows no difference here (p = 0.60).

Breast cancer: PIK3CA by subtype

Breast cancer comes in subtypes (Luminal A, Luminal B, HER2-enriched, Basal-like). PIK3CA mutations, the target of drugs like inavolisib, are far more common in some subtypes than others.

• Luminal A 47.5%, HER2-enriched 32%, Luminal B 30%, Basal-like 7%.
• The hormone-receptor-positive subtypes, where PI3K inhibitors are tested, carry most PIK3CA mutations.

Brain tumors: IDH status

Gliomas with an IDH1 or IDH2 mutation grow more slowly, and patients live much longer. Vorasidenib blocks the mutant IDH enzyme. This compares survival by IDH status and shows how it separates slower-growing gliomas from glioblastoma.

• IDH-mutant in 81% of lower-grade gliomas vs 6% of glioblastomas.
• Median survival 95 months (IDH-mutant) vs 15 months (IDH-wildtype).

Colorectal cancer: BRAF and microsatellite instability

About 1 in 10 colorectal cancers has a BRAF mutation, usually aggressive and treated with encorafenib + cetuximab. Many of these tumors also have high microsatellite instability (MSI-H): their DNA repair is broken, which changes how they respond to immunotherapy.

• BRAF mutated in 12% (V600E in 9%); 81% of V600E tumors are MSI-H vs 13% of KRAS-mutant tumors.
• BRAF and KRAS almost never occur together (1% of tumors).