Peto's Paradox

Is it really a paradox? Several biologists have argued that it is not, because nobody who thinks about it expects cancer defences to be constant across species. The counting argument is a null model, and the null model is wrong in an obvious way. On this reading the correct name would be "Peto's observation", and its value is in what it forces you to explain rather than in any contradiction.

The counter-argument, and the reason the name has stuck, is that the null model was not obviously wrong when Peto stated it, and the size of the discrepancy is still not fully accounted for. Knowing that elephants have extra TP53 copies tells you that they have a defence. It does not yet tell you that this defence is quantitatively sufficient to cancel a millionfold increase in expected risk. The bookkeeping is not closed.

How much of the answer is a slower clock? Leonard Nunney and others have pointed out that many large animals also grow slowly, divide their cells less often per unit time, and have lower metabolic rates per gram. If mutation accumulates per cell division rather than per year, much of the apparent discrepancy shrinks before you invoke any special gene. The 2022 somatic mutation data supports a version of this. The open question is what fraction of the gap it closes: a large fraction, in which case the exotic mechanisms are a supplement, or a modest one, in which case they are the main event.

Are the data good enough? Cancer prevalence in wild animals is hard to measure. Most figures come from zoo necropsies, which is a biased sample: captive animals are protected from predation and often live longer than wild ones, veterinary records vary in quality, and species are unevenly represented. Large-scale zoo pathology studies published since 2020 have found substantial cancer mortality in some species previously thought resistant, and the field has become correspondingly more careful. Any confident number for whale cancer incidence should be treated with suspicion, because nobody has performed a systematic necropsy programme on whales.

Hypertumours: elegant but unconfirmed. Nagy's proposal is theoretically attractive and follows straightforwardly from treating tumours as evolving populations. Direct evidence in a large animal is thin. It sits in the literature as a live hypothesis rather than an established mechanism, and it is worth labelling as such.

Why the disagreement matters. These are not academic distinctions. If the answer is mostly "slower cell division", there is little to borrow for human medicine. If it is mostly "better molecular defences", there is a great deal.

1975. Richard Peto, Frank Roe, Peter Lee and colleagues published "Cancer and Ageing in Mice and Men" in the British Journal of Cancer. It is principally a study of how the timing of carcinogenesis relates to a species' lifespan. In discussing the results, Peto observes that a human being has thousands of times more cells than a mouse and lives thirty times longer, and yet is not thousands of times more likely to develop cancer — and remarks that this ought to be puzzling.

1977. Peto set the argument out more fully in a conference paper on the epidemiology and biology of carcinogenesis, and it is this statement that is usually cited as the origin of the paradox.

Late 1970s to 2000s. The problem was known but generated relatively little research. It sat mostly in comparative pathology and in occasional theoretical papers.

2007. John Nagy, Erin Victor and Jenese Cropper published the hypertumour hypothesis, asking directly why whales do not all have cancer.

2011. Aleah Caulin and Carlo Maley published a review in Trends in Ecology & Evolution framing large animals as a natural experiment in cancer prevention. This is the paper usually credited with reviving the field, and the phrase "Peto's paradox" became common from about this point.

2013. Xiao Tian, Vera Gorbunova, Andrei Seluanov and colleagues identified high-molecular-mass hyaluronan as the basis of the naked mole rat's resistance, publishing in Nature.

2015. Two independent groups reported the elephant TP53 expansion. The Abegglen and Schiffman paper in JAMA is the most cited, and reported both the gene copy number and the elevated apoptotic response of elephant cells. The same year, the bowhead whale genome was sequenced, identifying changes in DNA repair and cell cycle genes as candidates for its exceptional longevity.

2022. Alex Cagan, Iñigo Martincorena and colleagues published somatic mutation rates across sixteen mammal species, showing the inverse scaling with lifespan.

It reframed cancer as an evolutionary problem, not just a molecular one. Cancer is somatic evolution: mutation and selection among the cells of a body. If you take that seriously, then multicellular life has been running a several-hundred-million-year arms race against its own cells, and the defences that exist are the ones that won. Peto's paradox is the clearest evidence that those defences vary enormously and are tunable.

It created comparative oncology as a research programme. The logic is straightforward and was not obvious before: rather than studying cancer only in humans and mice, study the species that solved it. Elephants, whales, naked mole rats, bats and long-lived bird species are now studied specifically for their resistance mechanisms, and the resulting list of candidate genes is a resource human oncology did not previously have.

It has produced concrete therapeutic leads. Elephant TP53 and its unusual apoptotic behaviour is under active investigation as a template for drugs that push damaged human cells towards self-destruction rather than repair. This is early-stage work and should not be oversold, but it is a direct line from a puzzle about body size to a laboratory programme.

It clarifies what cancer risk actually depends on. The naive model says risk tracks cell count and time. The comparative evidence says risk tracks the number of defences that have to fail, and that this number is a species-level trait. That reframing explains why a small number of inherited defects — a missing TP53 allele, a broken mismatch repair pathway — raise risk so dramatically: they remove a layer from a system built to have several.

Drug development from other species. The mechanisms identified in resistant species are being pursued directly: elephant p53 variants as a way to lower the threshold for apoptosis in damaged cells, and the naked mole rat's hyaluronan pathway as a target for restoring contact inhibition. Neither has produced an approved therapy, and both are worth watching rather than believing.

Understanding the limits of prevention. The evolutionary framing predicts that any single-mechanism intervention will be evaded, because a tumour is a population under selection and single-target pressure is exactly what selection handles best. Species that solved cancer did it with layered, redundant defences. That argues for combination approaches in humans, and it is one of the reasons resistance to single-agent therapy is expected rather than surprising.

A caution about scaling arguments generally. Peto's paradox is a clean example of a failure mode that recurs across science: extrapolating a relationship measured within a system to comparisons between systems, where the parameters you held fixed are the very things that differ. The same error appears in comparing crime rates between countries with different reporting practices, or model performance across benchmarks with different construction.

Species conservation and ageing research. Long-lived, large-bodied species are now studied jointly for cancer resistance and for ageing, since the mechanisms overlap heavily — DNA repair fidelity, senescence control, telomere management. The bowhead whale genome and the naked mole rat have become reference points for both fields.

The multistage model, and why it predicts a catastrophe

The Armitage–Doll multistage model treats a cell as requiring $k$ independent rate-limiting events before it becomes malignant. For a single cell with mutation rate $u$ per cell division and $d$ divisions over a lifetime, the probability of accumulating $k$ specific hits is approximately

$$p_{\text{cell}} \approx \frac{(u,d)^k}{k!}$$

For an organism with $N$ cells at risk, and treating cells as independent, the lifetime probability of at least one malignancy is

$$P_{\text{cancer}} = 1 - (1 - p_{\text{cell}})^{N} \approx 1 - e^{-N p_{\text{cell}}}$$

for small $p_{\text{cell}}$. The scaling is what matters: expected number of malignant clones goes as

$$\mathbb{E}[\text{clones}] \propto N , (u,d)^k$$

so it is linear in cell number and polynomial of degree $k$ in the number of cell divisions, and therefore in lifespan.

Now put in numbers. A blue whale has roughly $10^3$ times the cells of a human and a comparable lifespan, so on cell count alone it should have around a thousandfold higher expected clone count. A bowhead lives roughly $2.5$ times as long as a human; with $k \approx 6$, the lifespan term alone contributes a factor of $2.5^6 \approx 244$. Multiply and the naive prediction for a bowhead whale is five orders of magnitude more cancer than a human, before accounting for anything else.

The observation is that the true ratio is around one, or below it. Something must be supplying a compensating factor of $10^5$ or more.

Where the compensating factor can come from

Rearranging, the expected clone count for a species is

$$\mathbb{E} \propto N \cdot \frac{(u,d)^k}{k!}$$

There are only three places to intervene.

Reduce $u$, the mutation rate per division. Because the term is raised to the power $k$, this is extremely efficient. With $k = 6$, halving $u$ cuts expected clones by a factor of $2^6 = 64$. A tenfold reduction gives $10^6$. This is why the 2022 finding that somatic mutation rate scales inversely with lifespan is so important: a modest-looking change in $u$ does enormous work.

Increase $k$, the number of required hits. Extra tumour suppressor copies do exactly this. If an elephant needs to lose twenty TP53 copies rather than two, the relevant exponent rises, and since $u,d \ll 1$ each additional required event multiplies the difficulty by a small number. Note that increasing $k$ also flattens the age-incidence curve's dependence on cell number, because the polynomial term dominates.

Reduce $d$, divisions per cell over a lifetime. Large animals generally have lower mass-specific metabolic rates and slower tissue turnover, contributing another power-$k$ reduction.

Since these multiply, three individually modest adaptations — a threefold lower mutation rate, two extra required hits, and half the cell divisions — combine to several orders of magnitude. No single dramatic mechanism is required, which is consistent with what comparative genomics actually finds: many small changes across DNA repair, cell cycle and apoptosis pathways rather than one master switch.

Within versus between species, formally

Within a species, $u$, $k$ and the defence architecture are approximately fixed. Risk is then monotonically increasing in $N$, and the model predicts

$$\frac{\partial \log P}{\partial \log N} \approx 1$$

which is roughly what the human height–cancer cohort data shows: a positive, modest elasticity, on the order of a ten percent increase in risk per ten centimetres of height.

Between species, $N$, $u$ and $k$ co-evolve. Write $u = u(N)$ and $k = k(N)$. Observed flatness requires

$$\frac{d}{d\log N}\left[\log N + k(N)\log\big(u(N)d(N)\big) + \log\frac{1}{k(N)!}\right] \approx 0$$

This is a constraint, not a coincidence: any lineage in which the bracket grew substantially positive as body size increased would have been removed by selection before reaching large size. The flatness is a survivorship result. The species we can observe are, by construction, the ones for which the derivative was held near zero.

The hypertumour correction

Nagy's model adds a term the standard framework lacks. A tumour of size $S$ requires angiogenic investment scaling with $S$, and within it, sub-clones arise at a rate proportional to the tumour's own cell count and division rate. A sub-clone that defects on angiogenesis has a local growth advantage and a global cost. The time for such a defector to appear scales down with tumour size, while the time for a tumour to become lethal scales up with host size — a lethal tumour in a whale must be far larger in absolute terms than one in a mouse.

For sufficiently large hosts these curves cross, and the prediction is that tumours become self-limiting before they become lethal. That is the mechanism's appeal: it produces size-dependent protection with no species-specific genetics at all. It also makes a testable prediction — that large animals should carry many more benign, stalled neoplasms than small ones — which has not yet been tested at the scale required.