Which genes matter in lung cancer? Three ways of answering that with maftools, on 616
tumours from TCGA — and what each one gets wrong.
Learning project. Public TCGA data, meant to be read and re-run.
The obvious first move is to count how many tumours carry a mutation in each gene and take the top 20.
TP53 leads at 50% and is a real driver. Most of the rest are not. TTN (43%), MUC16 (41%), CSMD3 (39%) and RYR2 (34%) are among the largest genes in the genome. TTN runs to roughly 100,000 coding bases; KRAS is about 570. A tumour accumulating random damage hits a big target more often than a small one, so these genes reach a third of samples without being selected for at all.
EGFR ranks 45th. It is mutated in 14% of this cohort and is the target of an entire drug class — and frequency ranking buries it beneath ten passenger genes.
Frequency measures how big a gene is, not how much it matters.
oncodrive asks something different: do a gene's mutations pile up at particular
residues? Random damage spreads out along a gene. Selection concentrates it where the
protein's function changes. Size cancels out.
Two genes pass FDR < 0.1: KRAS (0.00025) and EGFR (0.039). EGFR moves from 45th to 2nd.
But TP53 vanishes. Mutated in half the cohort, unquestionably a driver, and completely absent from these results. That is structural, not a bug: TP53 is a tumour suppressor, switched off by damage almost anywhere in the gene. Clustering detects the opposite — oncogenes, which need one specific change to switch on. KRAS needs codon 12. EGFR needs L858R. TP53 just needs breaking.
EGFR: 104 variants, 85 samples. L858R appears 23 times and the exon 19 deletion (E746_A750del) 18 times — together 41 of 104. Nearly all sit inside the kinase domain (residues 704–1019), holding the receptor in its active state. That is why these tumours respond to TKIs.
KRAS: 163 variants, one cluster, 94.5% of mutations inside it — codon 12, in the P-loop of a 189-residue protein, where substitution blocks GTP hydrolysis and locks the protein on.
Both genes cluster where function changes. That is what oncodrive detected.
One tumour out of 241 carries both (p = 9.6e-11). They activate the same RAS/MAPK pathway, so the second mutation adds nothing once the first has happened.
This is not a statistical curiosity — it defines two diseases. EGFR-mutant tumours respond to TKIs; KRAS-mutant tumours do not.
Across all 616 tumours, C>A dominates — the tobacco signature, where benzo[a]pyrene damage to guanine produces G>T.
But EGFR-mutant lung cancer is common in never-smokers, and KRAS-mutant disease is strongly smoking-associated. Splitting on that:
| C>A | C>T | |
|---|---|---|
| EGFR-mutant (n=90) | 24.5% | 40.4% |
| KRAS-mutant (n=157) | 39.5% | 28.1% |
Wilcoxon p = 6.5e-13. The spectra are inverted. KRAS tumours carry tobacco damage; EGFR tumours carry the age-related signature of spontaneous cytosine deamination. Mutational burden follows — 4.66 against 1.43 per Mb (p = 5.2e-09).
The pooled figure describes neither group.
- Frequency ranks gene size. Clustering ranks oncogenes. Neither finds tumour suppressors.
- A cohort average can describe none of the groups inside it.
- A MAF contains SNVs and small indels only — no fusions (ALK, RET, ROS1) and no copy number, whatever question you ask of it.
BiocManager::install(c("TCGAbiolinks", "maftools"))Then open maftools_TCGA_LUAD.Rmd in RStudio and knit it, from the repo root. The GDC download caches to
GDCdata/ on the first run (a few hundred MB, not tracked here). R 4.6.1,
TCGAbiolinks 2.40.0, maftools 2.28.0 — full versions in results/sessionInfo.txt.
All figures are in figures/, all tables in results/, including the cohort summary,
driver-panel oncoplot, Ti/Tv breakdown and EGFR survival curve not shown above.
- Mayakonda A et al. (2018). Maftools. Genome Research 28:1747–1756.
- Colaprico A et al. (2016). TCGAbiolinks. Nucleic Acids Research 44:e71.
- TCGA Research Network (2014). Comprehensive molecular profiling of lung adenocarcinoma. Nature 511:543–550.
- Lawrence MS et al. (2013). Mutational heterogeneity in cancer. Nature 499:214–218.
- Related published work on a different NSCLC subtype: Tripathi V, Khare A, Shukla D, Bharadwaj S, Kirtipal N & Ranjan V. Genomic and computational- aided integrative drug repositioning strategy for EGFR and ROS1 mutated NSCLC. International Immunopharmacology, 2024 — squamous carcinoma rather than adenocarcinoma, and a drug-repositioning question rather than a methodological one.
MIT