drug-discovery-prompts/upstream/mims-harvard-ToolUniverse/skills/tooluniverse-cancer-genomics-tcga/SKILL.md

---
lineage_type: import
upstream_source: https://github.com/mims-harvard/ToolUniverse/blob/e2520a96/skills/tooluniverse-cancer-genomics-tcga/SKILL.md
upstream_sha: e2520a96
imported_at: 2026-06-26
prompt_class: prompt
upstream_changes: accepted
name: tooluniverse-cancer-genomics-tcga
description: TCGA/GDC cancer genomics analysis — cohort construction, clinical metadata retrieval, somatic mutation frequencies, survival analysis, and multi-omics integration. Use for TCGA-BRCA-style cohort studies, mutation prevalence by cancer type, survival-by-mutation analysis, and pan-cancer driver discovery. Always cancer-type-specific (don't use pan-cancer counts without cohort context).
triggers:
  - keywords: [TCGA, GDC, cancer cohort, somatic mutation, Kaplan-Meier, survival analysis, CNV, copy number variation, Progenetix, OncoKB, tumor, cancer genomics, mutation frequency]
  - patterns: ["TCGA-", "survival analysis for", "mutation frequency in", "copy number", "GDC project", "cancer cases", "overall survival"]
disable-model-invocation: true
---

# Cancer Genomics / TCGA Analysis

**TCGA analysis starts with: what cancer type? what data type?** Build your cohort FIRST (GDC filters), then analyze. Don't query mutations without defining the cohort — pan-cancer counts from `GDC_get_mutation_frequency` are uninformative without cancer-type context. A mutation frequency of 10% in one cancer type may be 0.5% in another; always specify `project_id`. Survival analysis (Kaplan-Meier) is hypothesis-generating in retrospective TCGA data — always report sample size and p-value, and note that TCGA cohorts are not treatment-stratified.

**LOOK UP DON'T GUESS**: never assume TCGA project IDs, NCIt codes, or gene coordinates — use `GDC_list_projects` to confirm project IDs and `Progenetix_list_filtering_terms` for NCIt codes.

Systematic TCGA/GDC analysis: define cohorts, retrieve clinical data, profile somatic
mutations, query copy number variations, run survival analysis, and interpret variants
with OncoKB.

## When to Use

- "What is the mutation frequency of TP53 in TCGA-BRCA?"
- "Get survival data for TCGA-LUAD patients"
- "Find clinical data for breast cancer cases in GDC"
- "Which TCGA projects have KRAS G12C mutations?"
- "Show CNV amplifications of EGFR in glioblastoma"
- "Annotate BRAF V600E for clinical significance in melanoma"

## NOT for (use other skills instead)

- Precision oncology treatment recommendations -> Use `tooluniverse-precision-oncology`
- Rare disease gene discovery -> Use `tooluniverse-rare-disease-genomics`
- GWAS variant interpretation -> Use `tooluniverse-gwas-snp-interpretation`

---

## Workflow Overview

```
Input (cancer type / gene / TCGA project ID)
  |
  v
Phase 1: Study Selection  -- GDC_list_projects, GDC_search_cases
  |
  v
Phase 2: Clinical Data    -- GDC_get_clinical_data
  |
  v
Phase 3: Somatic Mutations -- GDC_get_ssm_by_gene, GDC_get_mutation_frequency
  |
  v
Phase 4: CNV Analysis     -- Progenetix_cnv_search, Progenetix_search_biosamples
  |
  v
Phase 5: Survival Analysis -- GDC_get_survival
  |
  v
Phase 6: Variant Interpretation -- OncoKB_annotate_variant
```

---

## Key Identifiers

| Data Type | Format | Example |
|-----------|--------|---------|
| GDC project | TCGA-{ABBREV} | TCGA-BRCA, TCGA-LUAD, TCGA-SKCM |
| GDC case | UUID | 3c6ef4c1-... |
| NCIt cancer code | NCIT:C###### | NCIT:C4017 (breast), NCIT:C3058 (GBM) |
| RefSeq chromosome | refseq:NC_###### | refseq:NC_000007.14 (chr7) |

### Common TCGA Project IDs

| Cancer | Project ID | NCIt Code |
|--------|-----------|-----------|
| Breast | TCGA-BRCA | NCIT:C4017 |
| Lung adenocarcinoma | TCGA-LUAD | NCIT:C3512 |
| Glioblastoma | TCGA-GBM | NCIT:C3058 |
| Melanoma | TCGA-SKCM | NCIT:C3510 |
| Colorectal | TCGA-COAD | NCIT:C4349 |
| Ovarian | TCGA-OV | NCIT:C4908 |
| Prostate | TCGA-PRAD | NCIT:C7378 |

---

## Phase 1: Study Selection

**GDC_list_projects**: No params required. Returns all GDC/TCGA projects with case counts.
- Use to browse available projects and map cancer types to project IDs.

**GDC_search_cases**: `project_id` (string, e.g., "TCGA-BRCA"), `size` (int, default 10), `offset` (int).
Returns case UUIDs and basic metadata.
- Use to confirm a project exists and retrieve case counts before deeper queries.

---

## Phase 2: Clinical Data

**GDC_get_clinical_data**: `project_id` (string), `primary_site` (string, e.g., "Breast"), `disease_type` (string), `vital_status` ("Alive" or "Dead"), `gender` ("female"/"male"), `size` (int, 1-100), `offset` (int).
Returns `{status, data: [{case_id, demographics: {gender, race, ethnicity, vital_status, age_at_index}, diagnoses: [{primary_diagnosis, tumor_stage, age_at_diagnosis, days_to_last_follow_up}], treatments: [{therapeutic_agents, treatment_type}]}]}`.
- Use `project_id` + optional filters to retrieve patient-level clinical attributes.
- `age_at_diagnosis` is in days; divide by 365.25 for years.
- Multiple diagnoses or treatments per case are possible.

```python
# Get clinical data for deceased BRCA patients
result = tu.tools.GDC_get_clinical_data(
    project_id="TCGA-BRCA", vital_status="Dead", size=50
)
```

---

## Phase 3: Somatic Mutations

**GDC_get_mutation_frequency**: `gene_symbol` (string REQUIRED, alias: `gene`). Returns pan-cancer SSM occurrence count.
- Returns TOTAL count across all TCGA; no per-project breakdown.
- For cancer-specific data, use `GDC_get_ssm_by_gene` with `project_id`.

**GDC_get_ssm_by_gene**: `gene_symbol` (string REQUIRED), `project_id` (string, optional), `size` (int, 1-100).
Returns `{status, data: [{ssm_id, mutation_type, genomic_dna_change, aa_change, consequence_type}]}`.
- `mutation_type`: "Single base substitution", "Insertion", "Deletion".
- `aa_change`: amino acid change notation (e.g., "Val600Glu").

```python
# TP53 mutations in lung adenocarcinoma
mutations = tu.tools.GDC_get_ssm_by_gene(
    gene_symbol="TP53", project_id="TCGA-LUAD", size=50
)
```

---

## Phase 4: CNV Analysis (Progenetix)

**Progenetix_search_biosamples**: `filters` (string REQUIRED, NCIt code e.g., "NCIT:C4017"), `limit` (int), `skip` (int).
Returns `{status, data: {biosamples: [{biosample_id, histological_diagnosis, pathological_stage, external_references}]}}`.
- Use to find samples with CNV profiles for a given cancer type.

**Progenetix_cnv_search**: `reference_name` (string REQUIRED, RefSeq accession), `start` (int REQUIRED, GRCh38 1-based), `end` (int REQUIRED), `variant_type` ("DUP"/"DEL"), `filters` (string, NCIt code), `limit` (int).
Returns biosamples with CNV in the specified genomic region.
- `variant_type="DUP"` for amplification, `"DEL"` for deletion.
- Use `filters` to restrict to a cancer type.

```python
# EGFR amplifications (chr7:55019017-55211628) in breast cancer
result = tu.tools.Progenetix_cnv_search(
    reference_name="refseq:NC_000007.14",
    start=55019017, end=55211628,
    variant_type="DUP", filters="NCIT:C4017", limit=10
)
```

**Progenetix_list_filtering_terms**: No params. Returns all available NCIt codes and labels.
- Use when you need to find the NCIt code for a cancer type.

**Progenetix_list_cohorts**: No params. Returns named cohorts available in Progenetix.

---

## Phase 5: Survival Analysis

**GDC_get_survival**: `project_id` (string REQUIRED, e.g., "TCGA-BRCA"), `gene_symbol` (string, optional -- filters to mutated cases).
Returns `{status, data: {donors: [{id, time, censored, survivalEstimate}], overallStats: {pValue}}}`.
- Each donor has `time` (days), `censored` (bool: False=death event, True=censored), and `survivalEstimate`.
- `overallStats.pValue`: log-rank p-value (present when `gene_symbol` splits cohort).
- Without `gene_symbol`: returns full-cohort survival curve.
- With `gene_symbol`: returns survival split by mutation status (mutated vs. wild-type).

```python
# Survival for TCGA-BRCA split by TP53 mutation
surv = tu.tools.GDC_get_survival(project_id="TCGA-BRCA", gene_symbol="TP53")
pval = surv["data"]["overallStats"]["pValue"]
```

---

## Phase 6: Variant Interpretation (OncoKB)

**OncoKB_annotate_variant**: `gene` (string, alias `gene_symbol`), `variant` (string, alias `alteration`, e.g., "V600E"), `tumor_type` (string, OncoTree code e.g., "MEL").
Returns `{status, data: {oncogenic, mutationEffect, highestSensitiveLevel, treatments: [{drugs, level, indication}]}}`.
- `oncogenic`: "Oncogenic", "Likely Oncogenic", "Neutral", "Inconclusive", "Unknown".
- `highestSensitiveLevel`: FDA approval level ("LEVEL_1"=FDA-approved, "LEVEL_2"=standard of care, etc.).
- Demo mode available for BRAF, TP53, ROS1 without API key.
- Set ONCOKB_API_TOKEN for full access.

```python
# Annotate KRAS G12C in lung adenocarcinoma
result = tu.tools.OncoKB_annotate_variant(
    gene="KRAS", variant="G12C", tumor_type="LUAD"
)
```

---

## Tool Quick Reference

| Tool | Key Params | Returns |
|------|-----------|---------|
| GDC_list_projects | (none) | All TCGA/GDC projects with counts |
| GDC_search_cases | `project_id`, `size`, `offset` | Case UUIDs + metadata |
| GDC_get_clinical_data | `project_id`, `vital_status`, `gender`, `size` | Demographics + diagnoses + treatments |
| GDC_get_mutation_frequency | `gene_symbol` (alias: `gene`) | Pan-cancer SSM count |
| GDC_get_ssm_by_gene | `gene_symbol`, `project_id`, `size` | Per-mutation records with aa_change |
| GDC_get_survival | `project_id`, `gene_symbol` (optional) | Kaplan-Meier donor array + pValue |
| Progenetix_search_biosamples | `filters` (NCIt code), `limit` | Biosample records |
| Progenetix_cnv_search | `reference_name`, `start`, `end`, `variant_type`, `filters` | Biosamples with CNV in region |
| Progenetix_list_filtering_terms | (none) | All NCIt codes in Progenetix |
| OncoKB_annotate_variant | `gene`, `variant`, `tumor_type` | Oncogenicity + treatments |

---

## Example Workflows

### Workflow 1: Gene-Centric Mutation + Survival Analysis

```
1. GDC_get_mutation_frequency(gene_symbol="KRAS")
   -> Pan-cancer mutation count

2. GDC_get_ssm_by_gene(gene_symbol="KRAS", project_id="TCGA-LUAD", size=50)
   -> Specific amino acid changes in lung adenocarcinoma

3. GDC_get_survival(project_id="TCGA-LUAD", gene_symbol="KRAS")
   -> Survival split by KRAS mutation status + p-value

4. OncoKB_annotate_variant(gene="KRAS", variant="G12C", tumor_type="LUAD")
   -> Clinical significance + approved therapies (sotorasib)
```

### Workflow 2: Cohort Clinical Summary

```
1. GDC_list_projects()  -> confirm TCGA-OV exists

2. GDC_get_clinical_data(project_id="TCGA-OV", size=100)
   -> Demographics, tumor stage, treatment history

3. GDC_get_survival(project_id="TCGA-OV")
   -> Baseline overall survival curve for the cohort
```

### Workflow 3: CNV Analysis for a Gene

```
1. Progenetix_search_biosamples(filters="NCIT:C3058", limit=10)
   -> GBM biosamples with CNV data

2. Progenetix_cnv_search(
       reference_name="refseq:NC_000007.14",
       start=55019017, end=55211628,
       variant_type="DUP", filters="NCIT:C3058"
   )
   -> GBM samples with EGFR amplification
```

---

## Reasoning Framework

### Evidence Grading

| Tier | Description | Example |
|------|-------------|---------|
| **T1** | FDA-recognized biomarker with approved therapy | BRAF V600E in melanoma (vemurafenib) |
| **T2** | Well-powered clinical study, standard-of-care relevance | KRAS G12C in NSCLC (sotorasib), OncoKB Level 2 |
| **T3** | Preclinical/small cohort evidence, biological plausibility | Recurrent hotspot in TCGA but no approved therapy |
| **T4** | Computational prediction or variant of unknown significance | Low-frequency mutation, no functional data |

### Interpretation Guidance

**Mutation frequency**: A gene mutated in >10% of a TCGA cohort is likely a driver candidate (e.g., TP53 in 36% of all TCGA). Mutations at <1% frequency are typically passengers unless they occur at known hotspots. Always cross-reference with OncoKB oncogenicity annotation.

**Survival analysis (Kaplan-Meier)**: A log-rank p-value < 0.05 suggests the gene mutation is associated with differential survival. Hazard ratio (HR) > 1 indicates worse prognosis for the mutated group. Interpret cautiously: TCGA cohorts are retrospective and not treatment-stratified. Small subgroups (n < 20) produce unreliable survival estimates.

**Copy number variation**: Focal amplifications (narrow peaks) of oncogenes (EGFR, MYC, ERBB2) are more likely functionally relevant than broad arm-level events. Homozygous deletions of tumor suppressors (CDKN2A, PTEN, RB1) are strong loss-of-function signals. DUP count from Progenetix reflects sample frequency, not copy number magnitude.

### Synthesis Questions

A complete cancer genomics report should answer:
1. What are the most frequently mutated genes in this cancer type, and which are known drivers?
2. Does mutation status of the queried gene associate with survival (p < 0.05)?
3. Are recurrent CNV events (amplifications or deletions) present at known oncogene/tumor suppressor loci?
4. What is the OncoKB clinical actionability level for identified variants?
5. How does the mutation landscape compare across TCGA cancer types (pan-cancer context)?

---

## Programmatic Access (Beyond Tools)

When ToolUniverse tools return truncated results or you need bulk data, use the GDC API directly:

```python
import requests, pandas as pd

# Bulk clinical data for a TCGA project
filters = {"op":"and","content":[
    {"op":"=","content":{"field":"project.project_id","value":"TCGA-BRCA"}}
]}
all_cases = []
offset = 0
while True:
    resp = requests.post("https://api.gdc.cancer.gov/cases", json={
        "filters": filters, "size": 500, "from": offset,
        "fields": "submitter_id,demographic.vital_status,demographic.days_to_death,diagnoses.tumor_stage"
    }).json()
    hits = resp["data"]["hits"]
    if not hits: break
    all_cases.extend(hits)
    offset += len(hits)
df = pd.json_normalize(all_cases)

# Download MAF mutation file by UUID
file_uuid = "abc123-..."  # from GDC_list_files result
url = f"https://api.gdc.cancer.gov/data/{file_uuid}"
content = requests.get(url, headers={"Content-Type": "application/json"}).content

# Gene expression: query files endpoint for HTSeq counts
expr_filters = {"op":"and","content":[
    {"op":"=","content":{"field":"cases.project.project_id","value":"TCGA-BRCA"}},
    {"op":"=","content":{"field":"data_type","value":"Gene Expression Quantification"}}
]}
```

See `tooluniverse-data-wrangling` skill for pagination, error handling, and format parsing patterns.

---

## Limitations

- `GDC_get_survival` with `gene_symbol` splits on mutation presence only; no multi-gene or stage-based stratification.
- `GDC_get_mutation_frequency` returns pan-cancer total only; per-cancer frequencies require `GDC_get_ssm_by_gene` per project.
- `GDC_get_clinical_data` returns up to 100 cases per call; use `offset` for pagination.
- Progenetix uses GRCh38 coordinates; provide GRCh38 positions for `Progenetix_cnv_search`.
- `OncoKB_annotate_variant` without ONCOKB_API_TOKEN operates in demo mode (limited to BRAF, TP53, ROS1).
- Progenetix `filters` param requires NCIt CURIE format (e.g., "NCIT:C4017"), not free text.