Contents
56 — Cancer genomics Part 11 — Human and statistical genomics

56 — Cancer genomics

Before this: Ch 16 — Mutation · Ch 17 — DNA repair · Ch 27 — The four forces · Ch 46 — Variant calling · Time: ~80 min

Statistics needed: S2 Distributions · S4 Hypothesis testing · S7 High-dimensional data

What you'll be able to do

The core idea

Cancer is the clearest worked example of everything in Parts 3, 5 and 7, running inside one body over a few decades.

Take the definition of evolution from Ch 27 — change in frequencies in a population over time — and change what the population is. It is no longer a species. It is the cells of a tissue. They divide, they inherit their parent cell's genome with occasional error, they compete for space and nutrients, and lineages that divide faster or die less become a larger share of the tissue. Every ingredient is present: heritable variation, differential reproduction, finite population size, and time.

Three parameters differ from the organismal case, and all three consequences matter.

Generation time collapses. Cell divisions take hours to days, not decades. A lineage inside you has more generations available than the entire history of your species.

Recombination is absent. Somatic lineages are strictly clonal. There is no sex, no , no way to separate a beneficial allele from the deleterious ones sitting next to it. is total, and hitchhiking is complete: every present in the founding cell of the winning clone rides to with it, whether it contributed anything or not. This single fact is why a cancer genome is mostly noise, and why separating signal from noise is the central computational problem of the field.

The function is local and orthogonal to yours. A cell that proliferates faster than its neighbours wins the tissue, regardless of what that does to the organism. Selection has no foresight and no objective (Ch 00); it does not know the host will die. Cancer is what happens when the level at which selection acts drops from the organism to the cell.

A tumour is not a broken cell — it is a population that outcompeted its neighbours, and its genome is the fossil record of that competition. Everything you can read off a cancer genome is one of three things: a cause of the expansion, a consequence of it, or a passenger that happened to be aboard the winning cell. Almost all of it is the third.


1. Somatic and germline: the distinction that organises the clinic

Ch 16 established that you are a mosaic. Cancer genomics turns that into an operational rule: every variant found in a tumour belongs to one of three categories, and the category determines who else is affected and what happens next.

Where it isExpected in tumourExpected VAF in bloodHeritable?Found by
every cell, including gametes~0.5 or ~1.0~0.5 or ~1.0yesany tissue
Somatic (tumour)one clonal lineage in the tumourcontinuous, 0 to ~1absentnotumour vs matched normal
Clonal haematopoiesisone expanded blood cloneabsent (unless blood-derived)2–20%, drifting upward with agenoblood, and it contaminates the "normal"

The third row is the trap. Blood is the standard matched normal, and blood in an older adult is not clonal-free: expanded haematopoietic clones carrying DNMT3A, TET2, ASXL1 or PPM1D are common past 60. Such a variant appears in the "normal" at 5% and gets subtracted as germline, or appears in plasma and gets called as tumour. Good assay design sequences white cells deeply and separately for exactly this reason.

The clinical stakes of the germline/somatic split are large. A somatic BRCA1 mutation in an ovarian tumour is a property of that tumour. A germline one is a property of the patient and, with probability ½ per first-degree relative, of the family — which triggers a completely different conversation about cascade testing and surveillance. Tumour-only sequencing cannot tell them apart reliably, and returns incidental germline findings at a rate on the order of 10–15% in unselected advanced-cancer patients, most of them unsuspected from family history. That is a consent problem before it is a technical one (Ch 58).

2. The tissue as a population

Normal tissue is not a passive background. A colonic crypt is maintained by a handful of stem cells competing neutrally for niche positions; alone converts a crypt to monoclonality within years. Normal oesophagus and skin in middle-aged people are patchworks of clones carrying NOTCH1 and TP53 , some covering square centimetres. Positively selected in cancer genes are the normal state of normal tissue.

That reframes the question. Cancer is not "a cell acquired a driver". Drivers are acquired constantly and almost never produce a tumour. Cancer is the rare trajectory in which enough capabilities accumulate in one lineage, in an order that works, for the clone to escape the that hold ordinary clones in place.

The hallmarks framework (Hanahan and Weinberg 2000, extended 2011, revisited as Hallmarks of Cancer: New Dimensions, 2022) enumerates those capabilities: sustained proliferative signalling, evasion of growth suppressors, resistance to cell death, replicative immortality, angiogenesis, invasion and metastasis, metabolic reprogramming and immune evasion, with genome instability and tumour-promoting inflammation as enabling characteristics, and phenotypic plasticity, non-mutational epigenetic reprogramming, polymorphic microbiomes and senescent cells added in 2022.

Read it correctly: the hallmarks are phenotypic solutions selection can reach, not a list of causes. Many reach the same hallmark — which is why one tumour type recurs with different driver combinations across patients, and why some hallmarks are reached with no mutation at all, by chromatin state (Ch 23) rather than sequence.

3. Drivers, passengers, and a statistical trap

A driver confers a selective advantage on the clone carrying it. A passenger does not. Note carefully what this is a statement about: a effect, not a function. A mutation that destroys an important protein is a passenger if it does not change the clone's growth rate.

The ratio is brutal. Across 2,658 whole cancer genomes and 38 tumour types in the PCAWG analysis, tumours carried on average 4–5 , with no driver identified at all in about 5% of cases — against a total burden ranging from a few hundred to several hundred thousand . You are looking for four needles in a haystack whose size varies a thousandfold between patients.

The naive test, and why it fails

The obvious approach: for gene g of coding length L, count non- across N tumours, compare to the expectation under a background rate μ, and reject.

\text{observed} \sim \text{Poisson}(\mu L N), \qquad z = \frac{O - \mu L N}{\sqrt{\mu L N}}

Statistics: the Poisson count model, and why its standard deviation is the square root of its mean — which is where that denominator comes from — is covered in S2 §2.

This is correct, and it produces garbage, because μ is not a constant. From Ch 16 §7, the local somatic mutation rate varies severalfold with replication timing (late-replicating regions higher), chromatin state (closed higher), (transcription-coupled repair lowers the rate on the transcribed strand of expressed genes), and sequence context. It also varies more than a thousandfold in overall burden between tumours of the same type.

Now watch what a modest, uniform misspecification does. Suppose the tumour type has a median burden of 2 mutations per Mb, so the naive μ₀ = 2 × 10⁻⁶ per bp per tumour. Gene G has 100 kb of coding sequence, is late-replicating, heterochromatic and unexpressed in this tissue, and its true local rate is 3μ₀.

expected per tumour, naive     = 2e-6 × 1e5           = 0.2
expected over N = 1,000        = 200
observed  (true rate = 3 μ0)   = 600
excess                         = 400
z = 400 / sqrt(200)            = 28.3

Statistics: the Bonferroni correction behind the threshold below — α/m for m ≈ 20,000 genes, which is exactly 2.5 × 10⁻⁶ — is covered in S7 §2.

over ~20,000 genes at α = 0.05 needs p < 2.5 × 10⁻⁶, i.e. |z| ≳ 4.6. This gene clears it by a factor of six. And the general form is the problem:

z \;=\; \frac{(r-1)\,\mu_0 L N}{\sqrt{\mu_0 L N}} \;=\; (r-1)\sqrt{\mu_0 L N} \;\propto\; \sqrt{N}

where r is the ratio of true to assumed rate. The false positive becomes more significant as the study grows. This is not a power problem that more data solves — it is a bias problem, and bias multiplied by sample size beats noise. Running the numbers backwards, this gene falls below the significance threshold only for N < 26, which is smaller than any study that could find a real driver.

This is not hypothetical. Early exome-scale driver scans confidently reported olfactory receptor genes and TTN — a 100 kb-coding-sequence muscle protein — as recurrently mutated across cancers. The excess was entirely local background rate. Lawrence et al. (Nature, 2013) diagnosed it and gave the fix.

Two fixes

Condition on the covariates. Model the background rate per gene per patient as a function of replication timing, expression and chromatin state, using the patient's own 96-channel mutation spectrum to compute the rate expected for that gene's specific sequence composition, and borrow strength across genes with similar covariates. This is a covariate model plus empirical-Bayes shrinkage, covered in S7 §6; the biology is in knowing which covariates are load-bearing. MutSigCV is the canonical implementation.

Use an internal control instead: somatic . Synonymous mutations in gene g experience the same replication timing, the same chromatin, the same trinucleotide contexts and the same patient-specific spectrum as the non-synonymous ones. So the synonymous count within the gene is a per-gene control for every covariate you failed to model. This is exactly the Ch 33 statistic, applied within a body rather than between species, with the context correction made explicit (dNdScv). Its virtue is an interpretable effect size: dN/dS = 2 for in a gene says half the observed missense mutations there are attributable to selection, which converts directly into an estimate of how many driver mutations a tumour actually carries. Its limitation is that synonymous sites are not perfectly neutral — splice and usage put a floor under the denominator.

Three families of driver-discovery method

FamilySignal it exploitsDetectsBlind to
Recurrence above backgroundmore mutations than the local rate predictsgenes hit often in a cohortgenes with few but potent mutations; anything the background model gets wrong
Functional impactmutations skewed toward damaging predictionsgenes whose mutations are non-randomly severeinherits every bias of the pathogenicity predictor (Ch 55)
Positional clusteringmutations concentrated in 1D sequence or 3D protein structure far more than uniform-style hotspots, including ones too rare to reach recurrence significancetumour suppressors, whose mutations are deliberately dispersed

The three are complementary because they key on different consequences of selection, and consensus calls across them are more robust than any one. Non-coding driver discovery is much harder — there is no synonymous control, no codon structure to exploit, and the background model problem is worse. The great success is the TERT : two recurrent substitutions, c.-124C>T at chr5:1,295,113 and c.-146C>T at chr5:1,295,135 (GRCh38; 1,295,228 and 1,295,250 in GRCh37), each creating a new ETS-family binding site and reactivating . PCAWG's systematic non-coding scan found remarkably few others.

4. Oncogenes and tumour suppressors — readable off a picture

Two classes, distinguished by the direction of the functional change, with a mechanical consequence that is visible in a diagram at a glance.

Tumour suppressor
Change
Cellular genetics — one suffices — both copies must go
Analogyaccelerator stuck downbrake cut
How many sequence changes achieve itvery fewalmost any
Point-mutation patternclustered at hotspot residues, dispersed, enriched for nonsense, frameshift, splice
Other routesamplification, fusion, hijackingdeletion, LOH, hypermethylation
Therapeutic handleinhibit the product — druggablenothing to inhibit; only

The pattern falls straight out of the mechanism. To create a new activity you must make a specific change — lock a GTPase in its GTP-bound state, or a kinase in its active conformation. Only a handful of amino acid substitutions do that, so selection can only reward those, and the mutations pile up on a few . To destroy an activity, anything works. Every truncating mutation in the first two-thirds of the coding sequence is equivalent, so selection rewards a broad, flat distribution.

KRAS — oncogene, 189 aa
                 G12 G13                Q61
                  ▼▼▼                    ▼
      ├───────────●●●────────────────────●────────────────────┤
      1       ▓▓▓▓▓▓▓ P-loop        ▓▓▓▓▓ switch II         189
      >95% of KRAS mutations in cancer fall on three codons


APC — tumour suppressor, 2,843 aa
       ✕  ✕✕   ✕  ✕ ✕✕✕✕✕✕✕✕✕✕✕✕✕✕✕✕ ✕✕  ✕   ✕    ✕      ✕
      ├──────────────────────────────────────────────────────┤
      1   ▓▓▓ oligomerisation    ▓▓▓▓▓▓ β-catenin binding   2843
      ✕ = nonsense or frameshift, anywhere in the 5' two-thirds

APC carries a caveat worth stating, because it cuts against the diagram. The mutations are dispersed in the sense that any truncation destroys the function — but somatic APC mutations nonetheless concentrate in the mutation cluster region, codons 1286–1513, under 10% of the coding sequence carrying over 60% of the somatic hits. Truncating there removes the AXIN-binding SAMP repeats while retaining some β-catenin-binding repeats, leaving Wnt output high but not maximal — the "just-right" level for tumour growth. That is dosage tuning within the truncating class, not hotspot gain of function, and it is why real suppressor lollipops are rarely as flat as the selective logic alone predicts.

The same distinction has a quantitative form. For a gene under no selection, roughly 5–8% of coding point mutations are nonsense, set by the and the local base composition. A tumour suppressor shows 30–60% truncating; an oncogene shows near zero, often below neutral, because a truncated oncoprotein is useless to the clone. In terms: an oncogene has wmissense ≫ 1 with wnonsense ≈ 1; a tumour suppressor has wnonsense ≫ 1.

TP53 is the instructive exception. It is a tumour suppressor whose missense mutations cluster hard at a handful of DNA-binding-domain residues — p.(Arg175His), p.(Arg248Gln), p.(Arg273His), p.(Arg282Trp). The reason is that p53 works as a tetramer, so a mutant subunit poisons complexes containing wild-type subunits: a effect, plus neomorphic gain-of-function activity for some alleles. The lollipop looks like an oncogene's because the selective logic is an oncogene's, even though the gene is a suppressor. When the picture and the classification disagree, the picture is telling you about the mechanism.

Two routes bypass point mutation entirely. Amplification raises oncogene dosage (ERBB2 in breast and gastric cancer, MYCN in neuroblastoma, EGFR in glioblastoma). Structural rearrangement creates gene fusions, of two kinds: a chimeric protein with constitutive activity (BCR::ABL1 in chronic myeloid leukaemia, EML4::ALK in lung adenocarcinoma), or an intact oncogene relocated under someone else's strong regulatory element (IGH::MYC in Burkitt lymphoma, TMPRSS2::ERG in prostate cancer). The second kind is a purely regulatory driver with no change to any protein sequence.

5. Knudson's two hits, derived rather than asserted

In 1971 Knudson had no sequence data. He had ages at diagnosis for retinoblastoma, and two clinical forms:

Hereditary formSporadic form
Lateralitybilateral, multifocalunilateral, one tumour
Mean age at diagnosis~15 months~30 months
Family historyoften presentabsent

The derivation is a hazard argument. Suppose a tumour requires k independent rate-limiting events in one cell, each occurring at constant rate per cell per unit time. The probability a given cell has accumulated all k by age t goes as tk, so the incidence — the hazard, the derivative — goes as tk−1.

That single-parameter difference explains both clinical observations at once. If a carrier needs one hit in any of millions of retinoblasts, the number of tumours per carrier is Poisson with mean λ ≈ 3, which is what Knudson observed. Poisson also predicts a small fraction of carriers with zero tumours — e−3 ≈ 5% — which is why an autosomal predisposition shows incomplete without anything else going on. And bilaterality is automatic: independent hits in two eyes are two independent Poisson draws.

The gene is RB1 at 13q14, and the two hits are its two . But the second hit is usually not an independent point , because there is a much faster way to lose one allele: loss of heterozygosity (LOH) — losing the entire chromosomal region containing the wild-type copy.

LOH mechanismCopy number afterDetectable as
Mitotic recombination between homologues2 (copy-neutral)allelic imbalance with no change
loss1allelic imbalance plus coverage loss
Chromosome loss then reduplication2 (copy-neutral)as mitotic recombination
Interstitial deletion1 or 0focal coverage loss
2 (copy-neutral)short tract of allelic imbalance
hypermethylation ("epi-hit")2no genomic change at all — needs methylation data

The asymmetry is the point: the first hit is a base, the second is an arm. Mitotic recombination and chromosome missegregation happen per cell division at rates vastly higher than the per-base , so once the first allele is gone the second follows quickly. That is why two-hit inactivation is achievable at all within a human lifetime.

LOH is read straight off heterozygous SNPs. At a purity of ρ = 0.65 with copy-neutral LOH, a het site where the tumour retained the B allele sits at B-allele fraction (0.65 × 2 + 0.35 × 1)/2 = 0.825, and one where it retained A sits at 0.175:

                          B-allele fraction across 17q
                     0.0    0.25    0.5    0.75    1.0
normal tissue         ·      ·     ▓▓▓▓▓    ·      ·
tumour, het retained  ·      ·     ▓▓▓▓▓    ·      ·
tumour, LOH, ρ=0.65   ·   ▓▓▓▓      ·      ▓▓▓▓    ·
                          0.175           0.825

Coverage separates the mechanisms: copy-neutral LOH leaves the log₂ coverage ratio at 0, deletion drops it (to log₂(1.35/2) = −0.57 at this purity).

Three modern refinements to "two hits". : some suppressors are dosage-sensitive and one hit is enough. and gain-of-function: TP53 again. And the hazard exponent for common adult carcinomas is not 1 — the Armitage–Doll multistage analysis of age-incidence curves gives k ≈ 6–7 rate-limiting steps, consistent with the 4–5 drivers plus additional epigenetic and microenvironmental events that whole-genome sequencing now counts directly.

6. Hereditary cancer syndromes: germline plus somatic

SyndromeGene(s)PathwaySpectrumGenomic consequence in the tumour
Hereditary breast and ovarian cancerBRCA1, BRCA2, PALB2breast, ovarian, prostate, pancreatic signatures; PARP-inhibitor sensitivity
Lynch syndromeMLH1, MSH2, MSH6, PMS2, EPCAMcolorectal, endometrial, urothelial, othersMSI, high burden, immunotherapy sensitivity
Li–FraumeniTP53DNA damage responsesarcoma, breast, brain, adrenocortical, very early onsetwidespread instability; no targeted therapy
Familial adenomatous polyposisAPCWnt signallinghundreds of colorectal adenomasearly APC biallelic loss
RetinoblastomaRB1cell-cycle restriction pointretinoblastoma, later sarcomasLOH at 13q14
MEN2RETreceptor tyrosine kinasemedullary thyroid, phaeochromocytoma activating — the rare germline

Note the last row. Germline predisposition genes are overwhelmingly tumour suppressors, because a constitutively activated oncogene in every cell of an embryo is usually lethal or produces a developmental syndrome rather than a cancer syndrome. RET in MEN2 is the exception that shows the rule.

The unifying model: the germline variant is not the cause of the cancer — it is a head start. A carrier's cells begin one step along a path everyone's cells are on. Three predictions follow, and all three hold: earlier onset, multiple primary tumours, and — crucially — the tumour still requires somatic events, so sequencing a carrier's tumour shows biallelic inactivation, with the wild-type lost somatically. That observation is itself used as evidence in variant classification (Ch 55).

is high but not complete and not a constant. Prospective cohort estimates put cumulative breast cancer risk to age 80 at about 72% for BRCA1 and 69% for BRCA2 carriers, and ovarian cancer risk at about 44% and 17% respectively — but these are cohort averages, and individual risk is modified by variant position, family history and common polygenic background (Ch 53). Professionals reason about a distribution of risk, not a number, and translate it into surveillance intensity and risk-reducing options; this chapter describes that reasoning and does not substitute for it.

7. Mutational signatures: reading causes off the genome

Ch 16 §10 set up the encoding: each somatic single-base substitution is assigned to one of 96 channels — 6 substitution classes referenced to the of the pair (C>A, C>G, C>T, T>A, T>C, T>G) × 4 possible 5′ neighbours × 4 possible 3′ neighbours. Pyrimidine referencing is not cosmetic: in an unstranded assay you cannot tell which strand carried the lesion, so C>T and G>A are the same observation.

Stack G tumours into a 96 × G count matrix M and factorise:

\mathbf{M} \;\approx\; \mathbf{W}\mathbf{H}, \qquad \mathbf{W} \in \mathbb{R}^{96 \times K}_{\ge 0},\;\; \mathbf{H} \in \mathbb{R}^{K \times G}_{\ge 0}

Columns of W are signatures — the characteristic 96-channel profile of one mutational process. Rows of H are exposures — how much of each process ran in each tumour.

Take NMF itself as given — , the decomposition it gets contrasted with, is in S7 §5. Spend the attention on why this decomposition and what breaks:

The reference catalogue is COSMIC, currently version 3.6 (released in COSMIC v104, May 2026), with separate catalogues for single-base substitutions (SBS), doublet substitutions (DBS), small insertions and deletions (ID), and copy-number and rearrangement classes. Signature numbers are identifiers, not ranks.

SignatureProfileUnderlying processWhat it tells you
SBS1C>T at CpGspontaneous deamination of 5-methylcytosinea clock — accumulates with divisions and age
SBS5flat, all channelsubiquitous, mechanism uncleara second, slower clock
SBS4C>A, broad contextbulky guanine adducts from tobacco smoketobacco exposure
SBS7a/bC>T at dipyrimidines, CC>TT doubletsUV photoproductssun exposure
SBS2 + SBS13C>T and C>G at TCWAPOBEC3 cytidine deaminasesoften episodic and clustered
SBS6/15/20/21/26 + ID1/ID2C>T plus at homopolymersmismatch-repair deficiencyMSI, Lynch or somatic MLH1 silencing
SBS10a/bhuge excess at TCT and TCGPOLE exonuclease-domain lossultra-hypermutation
SBS3 (+ ID6, rearrangement signatures)flat SBS profile; deletions with microhomologyhomologous-recombination deficiencyBRCA1/2, PALB2, "BRCAness"
SBS31, SBS35platinum adduct spectraprior chemotherapyiatrogenic — the treatment left a record

Two disciplines of interpretation. First, a signature is evidence about a process, not proof of an exposure: SBS4 in a never-smoker means something produced that chemistry, and the inference to a cause runs through everything else you know about the patient. Second, the flattest signatures are the least specific. SBS3 alone is a weak call, which is why clinical assessment combines the substitution signature with ID6 (deletions bearing microhomology at the junction — the fingerprint of microhomology-mediated end joining substituting for HR, Ch 17) and with genomic-scar scores built from LOH extent, telomeric allelic imbalance and large-scale state transitions.

The data also carries its own internal check. Bulky adducts are removed preferentially from the transcribed strand by -coupled repair, so a genuine adduct signature shows transcriptional strand bias; replication-associated processes show replicative strand bias flipping at replication origins. A signature without the expected asymmetry deserves suspicion.

8. Instability phenotypes: mutations in the mutation rate

Most drivers change a growth rate. A smaller class changes the , and the evolutionary consequence is different — it does not increase directly, it increases the supply of variation on which selection acts. Because is total in a clonal population, a mutator hitchhikes with the beneficial it generates.

MechanismGenomic signatureClinical hook
MSI / MMR deficiencyloss of MLH1, MSH2, MSH6, PMS2 — in Lynch, somatic MLH1 hypermethylation in sporadic cases at mono- and dinucleotide repeats; high burden dominated by indels; SBS6/15/20/21/26checkpoint inhibitors (established)
POLE / POLD1 exonuclease domain lostultra-hypermutation, often >100 mut/Mb, SBS10checkpoint inhibitors
Chromosomal instability (CIN)mitotic checkpoint and cohesion defects, crisis, centrosome amplification, arm-level copy-number change, subclonal copy numberpoor prognosis; drug tolerance
BRCA1/2, PALB2, RAD51C/D loss or BRCA1 methylationSBS3, ID6, large deletions with microhomology, high scar score (established)
pulverised in a micronucleus, religated at random (Ch 20)copy number oscillating between two states, tightly clustered breakpoints, uniform join orientationscan create an amplification in one event
KataegisAPOBEC acting on single-stranded DNA at resected break endstight clusters of C>T and C>G at TCW within a few kb, colocalised with rearrangement breakpointsmechanistic marker linking APOBEC to repair
Whole-genome doublingcytokinesis failure or mitotic slippagegenome-wide > 2; ~30% of primary tumours (818 of 2,778 tumour samples in PCAWG), higher in metastasespoor prognosis

Whole-genome doubling deserves its own sentence, because why it is selected is not obvious. Doubling buffers: with four copies of everything, a subsequent deleterious mutation or chromosome loss is less likely to be lethal. It therefore raises the tolerable rate of LOH and aneuploidy, licensing the instability that follows. It also confounds every downstream inference — a mutation that occurred before doubling is duplicated with the genome and appears at higher multiplicity than one that occurred after, which is exactly how you date the doubling event against the tumour's own mutational clock.

Extrachromosomal DNA (ecDNA) is the sharpest violation of everything in Part 2. Circular megabase-scale elements carrying amplified oncogenes, with no , they segregate randomly rather than equally at mitosis. Copy number therefore changes fast and non-Mendelianly within a tumour, generating extreme heterogeneity in oncogene dosage and a rapid, reversible route to drug resistance. Inheritance inside a body does not have to obey the rules inheritance between bodies obeys.

9. Heterogeneity, clonal architecture, and resistance as selection

A biopsy is a mixture: normal cells, and one or more tumour subclones, at unknown proportions, on unknown local copy number. Ch 46 gave the simple case; here is the general relation between the observable and the quantity you want.

\text{} \;=\; \frac{\rho\,\varphi\,m}{\rho\,\text{CN}_t + (1-\rho)\,\text{CN}_n} \qquad\Longleftrightarrow\qquad \varphi \;=\; \text{VAF}\cdot\frac{\rho\,\text{CN}_t + (1-\rho)\,\text{CN}_n}{\rho\,m}

with ρ the tumour purity, φ the (CCF) — the fraction of tumour cells carrying the — m the number of mutant copies per carrying cell, CNt the local tumour copy number and CNn the normal copy number (2 on autosomes).

VAF is not interpretable. CCF is. A mutation at φ ≈ 1 is clonal — present in the founding cell, therefore in every tumour cell, therefore a valid target. A mutation at φ = 0.3 is subclonal — present in one branch.

Cluster the CCFs across thousands of mutations and the clusters are the tumour's population structure. Reconstructing the tree from them is a phylogenetics problem (Ch 34) with one extra that does most of the work:

The pigeonhole rule. Cancer cell fractions of sibling clones must sum to at most 1, because they occupy disjoint sets of cells. If cluster B is at 0.7 and cluster C at 0.5, they cannot be siblings — 1.2 > 1 — so one must be nested inside the other, and the larger is ancestral. Frequencies alone force the topology.

The reconstruction rests on the infinite sites assumption — each genomic position mutates at most once, so shared mutations imply shared ancestry. It fails at CpG hotspots, under very high , and whenever LOH deletes a mutated and erases the evidence.

graph TD N["normal cell"] -->|"driver 1<br/>TP53"| A["founding clone<br/>CCF = 1.0"] A -->|"driver 2"| B["subclone B<br/>CCF = 0.55"] A -->|"driver 3"| C["subclone C<br/>CCF = 0.30"] B -->|"resistance allele<br/>pre-existing, rare"| R["subclone R<br/>CCF = 0.01"] R -.->|"treatment removes<br/>the competition"| R2["subclone R<br/>CCF ≈ 1.0<br/><i>relapse</i>"]

Bulk sequencing has a hard floor and a spatial blind spot. Subclones below roughly 5–10% CCF are indistinguishable from noise at standard depth, and one biopsy samples one region: multi-region studies (TRACERx and its successors) showed that mutations called clonal from a single biopsy are frequently subclonal in the whole tumour. Single-cell methods trade one limitation for another — scDNA-seq gives per-cell but suffers severe allelic dropout; scRNA-seq (Ch 48) resolves expression state and large-scale copy number inferred from expression averaged along , but calls individual SNVs poorly.

Statistics: a detection floor is a statement about the power of the test, not about what is present in the tumour; computing the power of the test you just ran is S4 §5.

Resistance is selection on standing variation

This is the cleanest evolutionary argument in clinical medicine, and it is the Ch 16 fluctuation-test logic transplanted into a patient.

Two hypotheses for how a resistant tumour arises. Induction: the drug causes the resistance mutation. Selection: the resistance mutation was already present in a rare subclone, and the drug removed its competitors.

They make different predictions. Under induction, resistance should arise as a single lineage, and its timing should depend on the post-treatment mutation supply. Under selection, resistance should be present before treatment at low frequency, should appear as a sweep, and — the decisive prediction — should frequently be polyclonal: several independent resistant lineages, each with a different mutation achieving the same functional escape, expanding in parallel in the same patient.

Polyclonal resistance is what is observed. Colorectal cancers treated with anti-EGFR antibodies relapse with multiple distinct KRAS and NRAS mutant clones detectable simultaneously in plasma. EGFR-mutant lung cancers on osimertinib relapse through several independent routes at once. Chronic lymphocytic leukaemia on ibrutinib relapses with independent BTK p.(Cys481Ser) clones. Convergent evolution on the same functional target, from independent mutational origins, is the fingerprint of selection acting on pre-existing variation.

One honest complication: some cytotoxic drugs are themselves mutagens and do raise the supply — platinum agents leave SBS31 and SBS35 behind. The drug can be both a mutagen and a selective agent. But the resistance allele that actually sweeps is usually one that was already there.

The therapeutic implication is uncomfortable. Maximum-tolerated-dose treatment maximises the selection differential and therefore the speed of the sweep. Adaptive therapy — dosing to contain the tumour and deliberately preserve drug-sensitive cells so they competitively suppress resistant ones — follows directly from the population genetics and is in clinical trials. It is investigational, not standard care.

10. From genome to treatment

graph TD T["tumour + matched normal<br/>sequenced"] --> D{"actionable<br/>driver?"} T --> S{"instability<br/>phenotype?"} T --> G{"germline<br/>finding?"} D -->|"fusion, activating<br/>mutation, amplification"| TT["matched targeted therapy"] S -->|"MSI-H / dMMR<br/>TMB ≥ 10 mut/Mb"| IO["checkpoint inhibitor"] S -->|"HRD / BRCA"| PARP["PARP inhibitor<br/><i>synthetic lethality</i>"] G -->|"pathogenic, confirmed<br/>in normal tissue"| FAM["genetic counselling,<br/>cascade testing"] TT --> M["ctDNA monitoring:<br/>response, resistance, MRD"] IO --> M PARP --> M

Targeted therapy matched to a driver is the oldest and most solid part: BCR::ABL1 in chronic myeloid leukaemia; EGFR exon 19 deletions and p.(Leu858Arg) and ALK fusions in lung adenocarcinoma; ERBB2 amplification in breast and gastric cancer. Tissue-agnostic approvals now exist for NTRK and RET fusions — the alteration, not the organ, defines the indication. Even KRAS, undruggable for forty years, has covalent inhibitors against the p.(Gly12Cys) specifically, with responses that are real, modest and short-lived.

The instructive failure is BRAF p.(Val600Glu). In melanoma, BRAF plus MEK inhibition works well. In colorectal cancer the identical variant responds poorly to the identical drug, because colorectal cells answer BRAF inhibition with rapid EGFR-mediated reactivation of the pathway and melanoma cells do not. Same variant, same protein, different tissue, different result. to therapy is not a lookup table; the driver is necessary and the cellular context is decisive.

is the only handle on a lost function — you cannot pharmacologically restore a deleted tumour suppressor. The full derivation is in Ch 17 §7; the summary is that PARP inhibition is tolerable in HR-proficient normal cells and lethal in HR-deficient tumour cells, and the therapeutic window is the difference between the patient's genotype and the tumour's somatic one. Olaparib was approved in 2014 for BRCA-mutated ovarian cancer; the class now covers breast, pancreatic and prostate indications, and -positive ovarian maintenance beyond BRCA with a companion diagnostic. Resistance comes largely through BRCA reversion that restore the reading frame — and reversions are detectable in plasma, which links this section to the next.

Immunotherapy biomarkers work through neoantigens. A non-synonymous can produce a peptide absent from the self-repertoire against which T cells were tolerised; presented on MHC class I, it is a target. Checkpoint blockade does not create those T cells — it removes the inhibitory brake on ones that already exist. More mutations means more chances that some peptide is both presented and recognised, which is the mechanistic reason tumour mutational burden predicts response on average.

Two established, tissue-agnostic approvals encode this. Pembrolizumab for MSI-H / mismatch-repair-deficient solid tumours (May 2017, the first tissue-agnostic oncology approval), and for -high solid tumours at ≥10 mutations/Mb (June 2020, on a 29% objective response rate in KEYNOTE-158).

Be precise about what those approvals do and do not license as understanding:

Liquid biopsy. Dying cells shed fragmented DNA into plasma. Cell-free DNA is mostly haematopoietic, peaks at ~166 bp (the -protected length), and clears with a half-life of a couple of hours — so it is a near-real-time sample of the whole tumour burden, including metastases a needle would never reach. Circulating tumour DNA () fraction ranges from >10% in advanced disease to below 0.01% in residual disease. Four uses, at four levels of maturity:

UseMaturity as of 2026
Genotyping in advanced disease — find the actionable driver without a biopsyEstablished. FDA-approved plasma companion diagnostics exist. A positive result is actionable; a negative one requires tissue, because low shedding produces false negatives
Monitoring and resistance detection — KRAS clones on anti-EGFR therapy, ESR1 p.(Asp538Gly) on aromatase inhibitors, BRCA reversions on Established for specific indications, with matched drugs and approved plasma assays
Minimal residual disease after curative-intent surgery — tumour-informed personalised panels tracking the resected tumour's own mutations at extreme depthStrongly prognostic and widely used. Most tumour-informed MRD assays are still laboratory-developed tests, but the first randomised evidence that acting on an MRD-positive result improves survival arrived with IMvigor011 (NEJM, October 2025): ctDNA-guided adjuvant atezolizumab in muscle-invasive bladder cancer improved disease-free and overall survival, while persistently ctDNA-negative patients did well on no adjuvant therapy. The FDA approved Signatera CDx as a companion diagnostic on that basis in May 2026 — the first for blood-based MRD. Whether it generalises to other tumour types is still open; further interventional trials are running
Early detection / screeningMost contested. A blood-based colorectal screening test was FDA-approved in July 2024 with 83% sensitivity for cancer, 90% specificity — and only 13% sensitivity for advanced precancerous lesions, which is where screening's benefit mostly comes from. Multi-cancer early detection tests remain in large trials

Two on the last row that no assay improvement removes. Against a low prior — cancer prevalence in a screened population is well under 1% — specificity, not sensitivity, sets the and thus the downstream burden of investigating false positives. And a test that detects cancers earlier necessarily detects some that would never have caused harm; overdiagnosis is a property of the disease's natural history, not of the assay. Clonal haematopoiesis reappears here too: a "tumour" mutation in plasma may be from an expanded blood clone, which is why serious MRD assays sequence white cells in parallel.


Common misconceptions

What people thinkWhat's actually true
Cancer is a genetic disease, so it runs in familiesNearly all cancer are somatic and unheritable. Strong hereditary predisposition accounts for roughly 5–10% of cases
A driver mutation causes cancerA driver confers a selective advantage. Drivers are acquired constantly in normal tissue and almost never produce a tumour; on average 4–5 are needed together, in a workable order
The most frequently mutated gene is the most important driverFrequency confounds selection with background . TTN and olfactory receptors were reported as pan-cancer drivers for exactly this reason
More tumours sequenced will fix the false-driver problemThe false-positive z-score grows as √N. Bias times sample size beats noise; only a correct background model helps
A identifies an exposureIt identifies a process. Different causes can generate similar chemistry, and the flatter the signature the weaker the inference. SBS3 alone is a weak call
Tumour suppressors always need two hits, like mutant p53, and epigenetic silencing all break the strict model. "Two hits" is a statement about rate-limiting steps, and the number for adult carcinomas is closer to 6
Resistance are caused by the drugThe drug selects; the allele is usually pre-existing. Polyclonal resistance — several independent clones converging on the same escape — is the evidence. (Some chemotherapies are mutagenic and raise the supply as well, leaving their own signatures)
A of 0.5 means half the cells carry itVAF confounds purity, local copy number, mutant multiplicity and . Only CCF is interpretable, and computing it requires all four
High means immunotherapy will workThe tissue-agnostic approval rests on a 29% response rate. TMB is a weak, assay-dependent prior; neoantigen quality, clonality and immune infiltration all matter more than the count
-negative means cancer-freeIt means below the assay's limit of detection, which depends on shedding, tumour site and depth. Absence of evidence
Tumour sequencing says nothing about the familyIncidental pathogenic findings turn up in the order of 10–15% of unselected advanced-cancer patients, most without suggestive family history

Worked example: one ovarian tumour, read end to end

Whole-genome sequencing of a high-grade serous ovarian carcinoma with matched blood. Purity ρ = 0.65, ≈ 2.2, 4,320 somatic SNVs over 2.86 Gb of callable sequence.

1 — Burden. 4,320 / 2,860 Mb = 1.51 per Mb, genome-wide. A 1.1 Mb targeted panel would be expected to capture ~2 qualifying mutations, giving a panel of 1.8/Mb — with an exact Poisson 95% interval on a count of 2 running from 0.24 to 7.22 events, i.e. 0.2 to 6.6 mutations/Mb. Far below the ≥10 threshold, but note how little a panel count constrains the estimate near the boundary. Microsatellite status: stable. No immunotherapy indication.

2 — Signature refit. Fixing W to COSMIC SBS and solving for a sparse exposure vector:

SBS1  0.18  →   778 mutations    clock, CpG deamination
SBS5  0.44  → 1,901              clock, flat
SBS3  0.33  → 1,426              homologous-recombination deficiency
SBS8  0.05  →   216
                -----
                4,321  (rounding)

Corroborating evidence, which SBS3 alone would not justify: of 88 deletions longer than 10 bp, 61 carry ≥2 bp of microhomology at the junction — the ID6 pattern, i.e. microhomology-mediated end joining substituting for HR. Genomic scar score elevated. Call: .

3 — The mechanism, and the two hits. Blood carries a BRCA1 variant, NM_007294.4:c.5266dup, p.(Gln1756Profs*74) — pathogenic. Its in blood is 0.49, as expected for a heterozygote. In the tumour it reads 0.83. Which is diagnostic, because the three hypotheses predict different numbers at ρ = 0.65:

germline het VAF  =  [ ρ·m_t + (1-ρ)·1 ] / [ ρ·CN_t + (1-ρ)·2 ]      ρ = 0.65

het retained           CN_t = 2, m_t = 1 : (0.65 + 0.35) / (1.30 + 0.70) = 1.00/2.00 = 0.500
copy-neutral LOH       CN_t = 2, m_t = 2 : (1.30 + 0.35) / (1.30 + 0.70) = 1.65/2.00 = 0.825
wild-type allele lost  CN_t = 1, m_t = 1 : (0.65 + 0.35) / (0.65 + 0.70) = 1.00/1.35 = 0.741

Observed 0.83 → copy-neutral LOH. The log₂ ratio across 17q is 0.0, confirming two copies retained; a deletion would have given log₂(1.35/2) = −0.57. Knudson's second hit, made visible: germline frameshift on one , mitotic recombination removing the wild-type allele on the other.

4 — Somatic drivers, and their timing. TP53 p.(Arg273His) at VAF 0.61, local CNt = 2. Assume one mutant copy:

\varphi = 0.61 \times \frac{0.65 \times 2 + 0.35 \times 2}{0.65 \times 1} = 0.61 \times \frac{2.00}{0.65} = 1.88

Impossible — a CCF above 1 has no meaning. The multiplicity assumption was wrong. With m = 2 (mutation followed by LOH, so both retained copies are mutant):

\varphi = 0.61 \times \frac{2.00}{0.65 \times 2} = \frac{1.22}{1.30} = \mathbf{0.94}

Clonal. A CCF exceeding 1 is not an error to clamp — it is the estimator telling you the copy state.

RB1 p.(Arg320Ter), VAF 0.11, CNt = 2, m = 1:

\varphi = 0.11 \times \frac{2.00}{0.65} = \mathbf{0.34}

Subclonal — about a third of tumour cells. So the history is ordered: BRCA1 germline (present from conception) → TP53 and BRCA1 LOH in the founding clone → RB1 loss in one branch.

5 — The treatment decision. Germline BRCA1 pathogenic variant with biallelic somatic inactivation, HRD-positive by signature and scar score: PARP-inhibitor maintenance is established practice in this setting. The germline result also triggers an offer of genetic counselling and cascade testing to first-degree relatives — an entirely separate decision, about people who are not the patient.

6 — Eighteen months later. Progression on maintenance. Plasma , sequenced deeply, finds the germline BRCA1 c.5266dup plus two different somatic deletions in the same region: a 4 bp deletion removing the duplicated cytosine and three neighbours (net −3 relative to wild type, in frame) at VAF 0.9%, and a separate 7 bp deletion (net −6, in frame) at VAF 0.4%. Both restore the reading frame; both restore enough BRCA1 function to rescue HR and abolish the synthetic-lethal relationship.

Two independent reversions, in one patient, converging on the same functional escape. That is not the drug creating mutations. That is a on standing variation — Luria and Delbrück's argument, run in a human being, and readable from a tube of blood.

Connections