Cancer Biology

The molecular identification began with Rous's chicken virus, which had been kept alive in laboratories as a curiosity. By the 1970s the gene responsible for its transforming power had been localised: src. In 1976 Harold Varmus and J. Michael Bishop, with Dominique Stehelin and Peter Vogt, used a radioactive probe for viral src to look for related sequences in the DNA of uninfected chickens — and found one. So did every other vertebrate they tested. The virus had not invented a cancer gene; it had picked up a normal gene for growth control, some time in the past, and carried a damaged copy.

This reframed everything. Cancer genes are not foreign; they are the cell's own machinery stuck in the on position, and a virus is only one of the ways to break them. In 1982 three groups pulled an active oncogene straight out of a human bladder carcinoma and found it differed from the normal RAS gene by a single base.

The other class of gene was harder to see, because losing something is harder to detect than gaining it, and the story of p53 shows how hard. A 53-kilodalton protein found in 1979 bound to a viral antigen was abundant in transformed cells and, when introduced into cells, helped transform them. It was classified as an oncogene and studied as one for ten years. Then Bert Vogelstein's group found that colorectal tumours have lost the gene from both copies of chromosome 17, and Arnold Levine's group discovered that the clones everyone had been working with were mutants. The normal protein is a brake — it arrests or kills damaged cells — and the mutants jam the brake for the remaining normal copy as well. TP53 is now known to be mutated in roughly half of all human cancers.

From the age curve. Suppose a cell must accumulate $k$ specific rare changes, each occurring at a small rate per unit time, in order to become malignant. The probability that all $k$ have happened by time $t$ goes as $t^{k}$, so the incidence — the rate at which new cases appear — goes as the derivative,

$$I(t) \propto t^{\,k-1}.$$

On log–log axes that is a straight line of slope $k - 1$. For large-bowel cancer the observed slope is about 5, so $k \approx 6$. The same arithmetic explains the brutal age dependence in ordinary terms: if incidence rises as the fifth power of age, then doubling age from 40 to 80 multiplies it by

$$2^{5} = 32.$$

Cancer is overwhelmingly a disease of the old not because old tissues are weak but because the required events take that long to pile up in one lineage.

From a childhood tumour. Alfred Knudson found a case where $k$ is small enough to see. Retinoblastoma occurs in two forms. In the hereditary form, tumours appear in infancy, usually in both eyes, often at several points in each retina; in the sporadic form, later, in one eye, singly. He showed that this is what a two-event process looks like when the first event is either inherited or not.

If a hereditary patient already carries one defective copy in every retinal cell, only one further event is needed, so tumours appear at a rate roughly constant in time: the number per patient follows a Poisson distribution, and Knudson's data fitted a mean of about three. The chance of having no tumour at all in either eye is then $e^{-3} \approx 5\%$, which matches the small fraction of carriers who escape. For sporadic cases both events must happen in the same cell, a far rarer coincidence, producing one tumour, later, and almost never two.

The prediction hidden in this is the striking part. For the inherited form to act as one hit, the mutation must remove a function rather than add one, and the second event must remove the remaining copy — so the gene involved is a brake, and cancer requires losing both copies of it. That was 1971. RB1 was cloned in 1986 and behaved exactly so.

From the genomes. The third count came from sequencing. A typical adult solid tumour carries thousands of somatic mutations, the great majority of them irrelevant passengers. Distinguishing drivers requires statistics — a gene mutated more often than the local background rate predicts, or mutated at a specific site repeatedly. The 2013 synthesis of thousands of tumours gave the answer: two to eight driver mutations per tumour, falling into about a dozen pathways.

Three methods, three eras, three kinds of data: an age curve from 1954, 48 childhood cases from 1971, and whole-genome sequencing from 2013. All land on a handful of required events. That convergence is the strongest evidence the field has that the multistage clonal picture is right, and it is the reason the number of steps is now treated as a fact rather than a model parameter.

What holds the picture together is that a tumour is an evolving population. Peter Nowell set this out in 1976: the cells of a tumour are a clone with variation, their environment selects among them, and treatment is a selection pressure like any other. This is why resistance emerges — not because cells learn, but because the rare cell that already had the resistant mutation is the one that survives and repopulates. It is the same arithmetic as the resistance calculation in pharmacology and the quasispecies problem in virology, applied to a lineage of human cells.

It also explains the field's hardest unsolved problem. Metastasis is what kills, and sequencing has not found mutations specific to it. If the capacity to spread comes from the state a cell is in rather than from a gene it acquired, then the thing to find is not another driver but a configuration — which is a considerably harder object to look for, and the reason the most lethal step in the process is the least understood.