Selection isn't causation: using patient age to find real cancer drivers

This paper quietly undermines an assumption I’d been building on, which is exactly why I wanted to present it. The standard way we call cancer drivers, finding mutations more frequent than expected and inferring positive selection (dN/dS, MutSigCV, OncodriveFML), conflates two different things. Cheek and colleagues show that selection and causation are not the same, and then give a surprisingly cheap way to tell them apart: patient age.

The NOTCH1 problem

Positive selection is everywhere in normal ageing tissue, skin, oesophagus, blood, colon all carry clonally expanding mutations without becoming cancer. Because tumours arise from cells already sitting in expanded clones, those normal-tissue mutations show up enriched in cancer genomes too.

The cleanest cautionary tale is NOTCH1 in oesophageal cancer. It’s strongly selected in oesophageal squamous cell carcinoma, and classified as a driver, yet it’s even more frequently mutated in normal oesophagus, and mouse experiments show NOTCH1 mutations actively inhibit tumour growth. A “driver” that isn’t one. The obvious worry: how many others are wrong?

The carcinogenic effect

Their fix is to define the quantity we actually care about, a carcinogenic effect w = r₁/r₀: the rate at which a cell with the mutation initiates cancer over the rate without it. Estimated as an odds ratio comparing mutation frequency among cancer founder cells (y) versus normal tissue cells (x). w ≫ 1 is a strong driver; w ≈ 1 is no carcinogenic effect; w < 1 is cancer-inhibiting (hello, NOTCH1).

Applied with normal-tissue data, the spread is enormous: in AML, carcinogenic effects span five orders of magnitude. DNMT3A/TET2/ASXL1 (the classic clonal-haematopoiesis genes) have weak carcinogenic effects (~10) despite strong selection, while NPM1, FLT3 and IDH2 confer 800–10,000×. Selection signal and carcinogenic potency are genuinely different axes.

The clever bit: age, when you have no normal tissue

You rarely have matched normal-tissue mutation frequencies. So they ask whether patient age can substitute, and the logic is elegant:

• A mutation that’s positively selected in normal tissue but not carcinogenic just accumulates with age, so it shows up in older patients.
• A potent cancer-causing mutation accelerates onset, so it shows up in younger patients.

Crucially, enrichment alone can’t separate the two (both raise frequency in cancers), but the age distribution can. They validate it: across young vs old AML, carcinogenic effects correlate (ρ = 0.72) while raw frequencies don’t (ρ = 0.12), and across tissues, stronger carcinogenic effects sit at younger mean ages (colorectal ρ = −0.65). Different mechanisms hitting the same tumour suppressor (deletions vs point mutations) agree on age bias too, a nice internal check.

What it says about breast cancer

In breast invasive carcinoma (1,562 samples), GATA3 and TP53 show significant young-age bias, strong carcinogenic effect. PIK3CA does not, despite its high prevalence, hinting its carcinogenic effect may be weaker than its frequency implies, which uncomfortably aligns with the disappointing track record of anti-PI3K therapy. (The honest caveat: failing the age test doesn’t prove PIK3CA isn’t a driver, only that this test can’t detect the signal.)

My take

The framework is a scalar simplification, it collapses multi-stage carcinogenesis, epistasis and temporal dynamics, and the branching-process model is, in the authors’ own words, “a cartoon” for tissues like breast. It’s also built entirely on protein-coding mutations; extending it to the non-coding genome is not done here. But that’s precisely the opening. The conceptual correction, selection ≠ causation, applies just as forcefully to non-coding drivers, where our null models are even shakier. Borrowing the age-bias idea for non-coding hotspot calling is one of the most concrete, low-cost things I took from any paper this year. It also adds justification for the CRISPRi validation in my Objective 4: sometimes intervention is the only way to break the selection-vs-causation tie.