Mutations and outcome
An oncoplot shows how often a gene is mutated; survival analysis asks whether patients carrying the mutation fare differently. The recipe looks simple: split into mutant and wild-type, draw Kaplan–Meier curves. The step that goes wrong most often is deciding who counts as wild-type.
Who is mutant, who is wild-type
Mutant: a patient with at least one non-synonymous mutation in the gene. Wild-type: a patient who was sequenced, with the gene adequately covered, and has no non-synonymous mutation in it.
The catch is that a MAF lists only samples with variants. A patient who was sequenced but had no mutation detected has not a single row in it. Pick patients from the MAF alone and that patient vanishes from the analysis; the correct move is to count them as wild-type. Conversely, patients in the clinical table who were never sequenced are neither mutant nor wild-type and must be excluded.
The TCGA-LAML MAF holds 193 patients and the clinical table 200. The 7 patients without a MAF record were sequenced too, so they count as wild-type in survival analysis. A further 12 have no follow-up time and cannot enter the survival analysis.
Three more rules:
- One row per patient. When a patient has several samples (diagnosis and relapse, say), collapse to the patient first, mutant if any sample is mutant, before any survival model.
- Fix the analysis set before grouping. Exclusions (no follow-up, not sequenced, duplicate samples) are decided before looking at results.
- Report patients and events. How many patients and deaths each group has matters as much as the p-value.
KM curves and the log-rank test
Pick a gene and the widget draws Kaplan–Meier curves for mutant and wild-type patients. Each step down is a death and each tick a censored patient (alive at last follow-up); the shaded band is the 95% confidence interval, with the numbers at risk below. The log-rank test asks whether the two curves differ overall; the HR comes from a Cox model with wild-type as reference and gives the size of the effect.
Try a few things:
- Compare TP53, DNMT3A and FLT3: which separate clearly and which do not? Take the p-values from the widget.
- Untick “count patients without any MAF record as wild-type” and see how the numbers and the result change.
- Pick a gene with few mutant patients and watch for the small-sample warning. With fewer than 10 mutant patients, or few deaths in a group, the tail of a curve rests on a handful of people, and the result is exploratory at best.
Common pitfalls
- Using only the samples in the MAF as the denominator. Sequenced patients without mutation records are lost, the wild-type group shrinks, and the result is biased.
- Counting unsequenced patients as wild-type. Their mutation status is unknown; exclude them.
- Lumping different mutation types together. FLT3-ITD and FLT3-TKD differ in biology and prognostic meaning; mixed together they can cancel out.
- Analysing samples instead of patients. Several samples from one patient must be collapsed first.