A clinically detectable tumor — a mass large enough to appear on a scan — contains billions of cells that have been accumulating mutations for years. Those mutations did not all arrive at once, and they did not all land in the same cell. Different lineages carry different mutations, run different transcriptional programs, and respond differently to drugs. Intratumoral heterogeneity is the term for this diversity. It is not a curiosity. It is the central reason that targeted therapy, which works so well in early scans, so rarely cures metastatic disease. To understand it, you need to think less like a pharmacologist and more like Darwin.

Peter Nowell published a two-page paper in Science in 1976 proposing that cancer behaves like a population under natural selection: cancer cells mutate, the mutations that help a cell survive and divide expand through the population, and over time the tumor evolves. He wrote this before the molecular biology existed to test it. Fifty years later, single-cell sequencing and multiregion tumor profiling have confirmed and extended every major prediction. The five principles of clonal evolution: mutation, where most are neutral passengers but a few are advantageous drivers; selection, where growth advantages expand; drift, where neutral mutations can spread by chance; bottlenecks, where hostile events crush the population; and branching, where multiple subclones evolve in parallel.

The organizing structure is the clonal tree. Every tumor has a common ancestor — the first cancer cell, which carried the founding mutations. All cancer cells in that patient are descended from that ancestor. The early, founding mutations are present in every cell; these are the trunk mutations. As the tumor grew, cells continued to mutate, and some of those later mutations were advantageous. Those cells expanded into subclones: lineages that share the trunk but carry their own additional mutations, the branch mutations. A large tumor contains many subclones, each occupying its own niche. A drug that targets a trunk driver hits every cancer cell in the body. A drug that targets a branch driver hits only the fraction of cells that carry it — the remaining subclones continue growing.

Tumors are not well-mixed flasks. Different regions carry different subclones, and a biopsy from one part of the tumor may give a different answer than a biopsy from a centimeter away. The TRACERx study made this concrete: multiple biopsies from different regions of the same non-small-cell lung cancer revealed extensive spatial heterogeneity. Some mutations were present throughout — clonal, truncal — while others appeared in only one or two sampled regions — subclonal, branch. The mutations driving metastases were sometimes different from the ones that looked most prominent in the primary. Gerlinger and colleagues showed this in renal cell carcinoma: different biopsies from the same patient led to different therapeutic conclusions. The tumor was, in a real sense, not one tumor at all.

Resistance follows directly from the clonal architecture and runs through several mechanisms that can coexist in different subclones of the same resisting tumor. Pre-existing resistance is the T790M story — the resistant cells were there before treatment, selected into dominance by the drug. Bypass signaling is what happens when MET amplification restores the same downstream effect through a parallel receptor. Phenotypic transformation is stranger: a subset of EGFR-mutant lung adenocarcinomas, when exposed to chronic inhibition, transform into small-cell lung cancer — the drug target disappears because the cell type that bore it no longer exists. Persister cells survive through a reversible epigenetic state — sequence them and they look sensitive, but they are not. And microenvironment-mediated resistance means CAFs and macrophages supply survival signals that partially substitute for the blocked pathway.

If resistance is evolutionary, then managing it requires watching the tumor evolve — which tissue biopsy does poorly and liquid biopsy does better. Cancer cells shed DNA into the bloodstream when they die. This circulating tumor DNA carries the mutations of the cells it came from. Because blood integrates shed DNA from across the whole tumor and all its metastases, a blood draw samples the tumor's heterogeneity more representatively than a needle in one spot. And because blood can be drawn repeatedly, serial ctDNA can track the clonal architecture as it changes under therapy — detecting the rise of a resistance mutation in real time, often weeks or months before it becomes visible on imaging. In the opening case, a blood draw during treatment could have detected T790M rising before the scan showed progression, prompting a switch to osimertinib before the resistant clone had time to consolidate.

Here is the chapter's central claim. The T790M that appeared at eleven months was never created by the EGFR inhibitor. It was there at diagnosis, in a cell or two among billions, carrying no selective advantage while the drug-sensitive majority dominated. The inhibitor's success — its month after month of shrinking tumors — was also the mechanism of its eventual failure: it was systematically eliminating every competitor the resistant cells had. The tumor is a population. The treatment is a selection. The oncologist is an evolutionary biologist who has not always known it. What remains open: we can reconstruct a tumor's evolutionary history after the fact, but predicting its future trajectory under a given therapy remains largely beyond us. Evolution is contingent, and that contingency may be irreducible.