Week 1 Day 4 — TCGA-LAML independent-cohort ETL

Summary

Fetched TCGA Acute Myeloid Leukemia PanCancer Atlas 2018 cohort (n=173 patients with RNA-Seq V2 data) from cBioPortal’s public REST API, built the same 80-dim patient feature schema as BeatAML, and saved canonical tables for Week 5 independent-cohort validation.

Key numbers

Stage Count
TCGA-LAML samples with mRNA 173
Unique patients 173 (1:1 mapping)
Genes fetched via cBioPortal 30,000 → 16,596 retained after pivot
Mutation rows (25 curated genes) 357 (145 unique samples)
Clinical attributes available 6 (OS_MONTHS, OS_STATUS, SUBTYPE, SAMPLE_COUNT, CANCER_TYPE_ACRONYM, IN_PANCANPATHWAYS_FREEZE)
Deceased at last follow-up 114/173 (65.9%)
Median OS (months) 11.0
RNA 50-PC cumulative variance 79.1%

Top mutation prevalence (25-gene panel)

Gene TCGA-LAML BeatAML (comparison)
FLT3 30.1% 35.2%
NPM1 27.7% 32.0%
DNMT3A 24.9% 17.5%
IDH2 10.4% ~10%
RUNX1 9.8% 13.4%

Good match with BeatAML and with AML literature (FLT3 ~30%, NPM1 ~30%, DNMT3A ~20%). DNMT3A is slightly higher in TCGA — expected since TCGA-LAML is biased toward older, de-novo cases where DNMT3A is more common.

Important caveat: clinical fields are BeatAML-median placeholders

cBioPortal’s PanCancer Atlas API only exposes OS_MONTHS, OS_STATUS, SUBTYPE for TCGA-LAML patients. AGE, SEX, RACE, ELN risk are listed as available attributes but return empty values (privacy or data-sparsity reasons).

To preserve the 80-dim feature schema for later transfer, we fill the 5 clinical columns with BeatAML medians (age=62, ELN ordinal=1.0, blast_pct=70%, secondary_aml=0.16, fit_for_intensive=0.5). This means the clinical features carry no TCGA-specific information; all personalization for TCGA predictions comes from RNA PCA + mutation features.

Practical implication: Week 5 TCGA validation should lean on the mechanism-prior combo scoring (uses mutations + expression signatures), not on direct transfer of the BeatAML-trained Baseline A MLP (which expects real clinical features).

Outputs

data/canonical/
├── tcga_laml_patient_features.csv    # 173 × 81 (patient_id + 80 features)
├── tcga_laml_clinical.csv            # 173 × 4 (patient_id + OS + SUBTYPE)
└── tcga_laml_manifest.json           # PCA variance, mutation prevalence, config

data/raw/tcga_laml/                   # cBioPortal API response cache (1.9 GB)
├── sample_ids.json
├── clinical_data.json
├── mutations.json
├── all_genes.json
├── expr_batch_000.json  …  expr_batch_007.json    # 8 expression batches
└── expression.parquet                              # pivoted gene × sample

(Raw cache is gitignored; only canonical CSVs get committed.)

Implementation notes

cBioPortal REST API quirks discovered:

  1. POST /clinical-data/fetch expects entityId, not sampleId or patientId in the identifier body.
  2. POST /genes/fetch takes a plain array body ["FLT3","NPM1"], and the geneIdType=HUGO_GENE_SYMBOL must be in the query string (not body).
  3. Expression fetch is batched at 4000 Entrez IDs × 173 samples per request (~100 MB JSON per batch). Total wall time ~17 min (network-bound). Full cache is 1.9 GB; subsequent runs reload from cache in <1 s.
  4. TCGA sample IDs embed the patient ID: TCGA-AB-3008-03 → patient TCGA-AB-3008. No extra API call needed for mapping.

Tests

Existing 46 tests still pass. No TCGA-specific tests added yet — the data comes directly from a public API with well-known schema, and the code-under-test is mostly plumbing (JSON reshape + PCA on the same pipeline as BeatAML, which already has tests).

Next (Day 5)

End-to-end integration smoke test: load all canonical tables, verify schema compatibility between BeatAML (613 patients), TCGA (173 patients), and DrugComb (186 strict pairs), and produce a one-page data-quality dashboard confirming Week 1 is complete.