Contents
16 — Mutation Part 3 — Genome instability

16 — Mutation

Before this: Ch 04 DNA replication · Ch 07 The genetic code · Ch 09 Mitosis and meiosis · Time: ~40 min

What you'll be able to do

The core idea

Two facts sit in tension, and is what falls out between them.

DNA is chemically fragile. It is a large organic molecule sitting in warm water. Bonds hydrolyse, bases oxidise, amino groups fall off. Tens of thousands of chemical lesions occur in every one of your cells every day, with no mutagen present at all.

The observed error rate is about 10⁻⁸ per base per generation. That is a fidelity of roughly one part in a hundred million, achieved by a molecule that is falling apart continuously.

The gap between those numbers is repair (Ch 17). Mutation is not what happens to DNA; mutation is the residue — the small fraction of chemical insults that repair missed or mis-fixed, then handed to the next round of replication, which copied the error faithfully and made it permanent.

That reframe has a consequence worth stating immediately:

Damage is not mutation. A lesion is a chemical alteration to one strand; it is reversible, and the cell can see it because the complementary strand disagrees. A mutation is a change to the sequence that is now correctly base-paired on both strands. Once a lesion has been replicated past, there is no longer any evidence that anything is wrong. The information is gone, not corrupted.

Everything in this chapter is either a mechanism for producing lesions, a reason a particular lesion evades repair, or an accounting of the result.


1. Germline and somatic: the distinction that comes first

Before any molecular classification, ask which cell lineage the is in.

Somatic
WhereCells that give rise to gametesEvery other cell
Inherited?Yes — passed to offspringNo
Present inEvery cell of the offspringOnly the descendants of the cell it arose in
Detected bySequencing blood/saliva; triosSequencing the affected tissue and subtracting a normal sample
Matters forInherited disease, evolution, population geneticsCancer, ageing, mosaic disorders

You are not a single . You are a clonal population descended from one zygote, and every division since has been an opportunity for error. Normal adult tissues accumulate on the order of tens of per cell per year, so a cell in your colon at 60 differs from the zygote at thousands of positions. That is , and it is the normal state.

Two consequences that are easy to get wrong. A somatic mutation in a "disease gene" is not a diagnosis of an inherited disease — it sits in one clone, and the only question that matters is whether that clone expands (Ch 56). And "de novo" does not mean recurrence risk is zero: if the mutation arose early in a parent's development, that parent is a gonadal mosaic — unaffected, negative on a blood test, carrying the variant in a substantial fraction of their gametes. Empirical recurrence risks for apparently de novo conditions are typically ~1%, not 0.

2. Classification by scale

ScaleNameTypical sizeNotes
1 bpPoint / SNV1Substitution of one base for another
1–50 bpinsertion or deletionIn coding sequence, a length not divisible by 3 shifts the frame
>50 bp50 bp – megabasesDeletions, duplications, , , mobile-element insertions
Chromosomalwhole arms or chromosomes, large-scale rearrangement (Ch 20)

The 50 bp boundary is a technological artefact, not a biological one — it is roughly where short- stops being able to see the event inside a single read. Different mechanisms dominate at different scales, but the classes grade into each other, and the "SV" category exists partly because those events are hard to call (Ch 46).

3. Classification by molecular type: the transition puzzle

Substitutions split into two classes by chemistry. (A, G) have two fused rings; (C, T) have one.

Count them. Each of four bases can change to three others: 12 possible substitutions. Four are transitions, eight are transversions. If substitutions were chosen uniformly at random, the transition/transversion ratio (Ti/Tv) would be 0.5.

Observed Ti/Tv in human whole-genome variant calls is about 2.0–2.1. In exomes it is about 3.0–3.3. So transitions are enriched roughly fourfold over chance genome-wide, and more in coding sequence. Both figures are standard variant-calling QC metrics precisely because sequencing errors are closer to the null 0.5, so a callset drifting toward 0.5 is a callset filling up with artefacts.

Where does the fourfold enrichment come from? The largest single identifiable contributor is one reaction on one base, which the next section derives — though it carries only about a third of the excess, and §5 says what carries the rest.

4. Classification by consequence

For a variant inside a protein-coding gene, read the consequence off the code (Ch 07):

reference   ATG CTG AAA GGT CAT TGG ...     Met Leu Lys Gly His Trp
synonymous  ATG CTC AAA GGT CAT TGG ...     Met Leu Lys Gly His Trp   (codon 2 CTG->CTC, still Leu)
missense    ATG CCG AAA GGT CAT TGG ...     Met Pro Lys Gly His Trp   (codon 2 CTG->CCG, Leu -> Pro)
nonsense    ATG CTG TAA GGT CAT TGG ...     Met Leu STOP              (codon 3 AAA->TAA, premature stop)
frameshift  ATG CTG AAG GTC ATT GG. ...     Met Leu Lys Val Ile ...   (codon 3, 1 bp deleted; everything downstream is garbage)
ClassWhat changesTypical severity
Synonymous changes, amino acid does notUsually mild — but see below
One amino acid substitutedAnywhere from nothing to complete
NonsensePremature stop codonUsually severe; often triggers , destroying the transcript
FrameshiftReading frame shiftsUsually severe; scrambles everything downstream
Splice-siteDisrupts an boundaryOften severe — exon skipped or intron retained (Ch 06)
RegulatoryNon-coding; changes how much protein is madeUsually subtle individually; collectively the class in complex disease

Why synonymous is not silent

The code is redundant, so a synonymous change leaves the protein sequence intact. It does not leave everything else intact.

  1. . Exons contain exonic splicing — short motifs marking the exon for inclusion. A synonymous change can destroy one and cause the whole exon to be skipped, or create a new cryptic splice site. This is the commonest route to a pathogenic synonymous variant.
  2. speed. Codons are read by tRNAs of very different abundance, so a rare codon stalls the . Proteins fold while still being synthesised, so changing the pause pattern can change the folded outcome from an identical amino acid sequence.
  3. mRNA structure and stability. The change alters RNA secondary structure, and hence half-life and to ribosomes and microRNAs.
  4. It may not be synonymous in every isoform. Genes produce multiple transcripts with different boundaries and frames.
  5. Selection can see it. Codon usage bias is maintained by selection in organisms with large (Ch 33).

Treating synonymous variants as a null class is a modelling convenience — usually reasonable, occasionally a clinically consequential error (Ch 55).

5. Where spontaneous mutations come from

No mutagen required. Five processes account for most of it.

graph TD D["Chemical lesion<br/><i>~10⁴–10⁵ per cell per day</i>"] R["Repaired correctly<br/><i>the overwhelming majority</i>"] M["Mis-repaired, or replicated<br/>past before repair"] F["<b>Mutation</b><br/>correctly base-paired,<br/>now invisible"] D --> R D --> M --> F P["Replication error<br/><i>polymerase mis-incorporation</i>"] --> M

Depurination. The bond joining a to the sugar hydrolyses spontaneously — roughly 10,000 purines lost per mammalian cell per day, leaving an abasic site, a backbone position with no base at all. ( loss runs at about a twentieth of that rate.) Base-excision repair handles nearly all of them; a polymerase reaching an unrepaired abasic site tends to insert A opposite it, so a lost G becomes a G→T transversion.

Deamination — the big one. Cytosine's exocyclic amino group hydrolyses off, converting C to uracil, at roughly 100–500 events per cell per day. Uracil pairs with A, so unfixed before replication the C:G pair becomes T:A — a transition.

Now the crucial asymmetry. Uracil does not belong in DNA. Uracil-DNA glycosylase excises it on sight and base-excision repair restores the C. The lesion is trivially recognisable because it is chemically foreign.

But cytosines in the dinucleotide CpG are frequently methylated as an epigenetic mark (Ch 23). Deaminating 5-methylcytosine does not produce uracil. It produces thymine — a perfectly legitimate DNA base. Repair now faces a G:T mismatch with no way to tell which base is the intruder. It must guess, and half the time it guesses wrong.

That single fact explains a great deal:

Oxidative damage. Reactive oxygen species from ordinary metabolism oxidise guanine to 8-oxoguanine, which can rotate into a conformation that pairs with A, yielding G:C → T:A transversions. This is one of the few common transversion routes, and the source of the notorious "OxoG" artefact in sequencing libraries oxidised during preparation.

Replication error. The classical explanation for polymerase mis-incorporation is a tautomeric shift: bases sit overwhelmingly in one tautomeric form but flip transiently to a rare form with different hydrogen-bonding geometry, letting G pair with T. Watson and Crick proposed it in 1953; NMR has since detected transient mispairing states populated at roughly 10⁻⁵ — which is about where base selection alone leaves the raw error rate. Two further filters act in series on that number: 3'→5' (a factor of 10²–10³) and (another 10²–10³), taking replication fidelity to roughly 10⁻⁹–10⁻¹⁰ per base per replication. Watch the units. That is per replication; the ~10⁻⁸ per base per generation figure of §7 is a different quantity, summing hundreds of divisions plus lesions that were never replication errors at all.

Strand slippage. At a tandem repeat the nascent strand can detach and re-anneal out of register, because the template offers many equally good re-annealing positions. Result: a gain or loss of one repeat unit.

template   ...CA CA CA CA CA CA CA...
nascent    ...GT GT GT GT GT GT...        slips back one unit -> insertion of CA
                          ^ re-anneals here instead of here

This is why microsatellites (tandem repeats of 1–6 bp) mutate at roughly 10⁻³–10⁻⁴ per locus per generation, four to five orders of magnitude above the point- — which made them the workhorse markers of classical mapping and forensic profiling. When mismatch repair fails, through germline loss of MLH1 or MSH2 in Lynch syndrome or somatic silencing in a tumour, slippage goes uncorrected and the genome acquires microsatellite instability (MSI): detectable at a handful of mononucleotide marker loci, and now a predictive biomarker for immune checkpoint therapy.

6. Induced mutagenesis, and how you test for it

Mutagens do not create new classes of . They raise the rate of lesion types the cell already deals with, sometimes with a characteristic bias.

AgentLesionMutational outcome
UV (UVB/UVC)Covalent of adjacent — cyclobutane dimers and 6-4 photoproductsC→T at dipyrimidines; the diagnostic CC→TT tandem substitution
Ionising radiationDouble-strand breaks, mostly indirect via hydroxyl radicals from water radiolysisDeletions, ,
Alkylating agents (EMS, nitrogen mustards, temozolomide)Alkyl groups on bases; O⁶-alkylguanine pairs with TG:C → A:T transitions, heavily biased
Base analogues (5-bromouracil, 2-aminopurine)Incorporated in place of a normal base, then mispairTransitions
Intercalators (proflavine, acridines, ethidium)Slot between stacked base pairs, distorting the helix±1 bp — frameshifts

Two of these earned a place in the history of the subject. Intercalator-induced frameshifts were how Crick and Brenner established that the code is read in non-overlapping triplets: single insertions and single deletions each destroyed function, but three insertions together restored it. And EMS is still the standard mutagen for forward genetic , precisely because its bias toward point mutations yields an allelic series rather than a clean (Ch 37).

The Ames test turns mutagenesis into an assay. Take a Salmonella typhimurium strain carrying a mutation in a histidine-biosynthesis gene, so it cannot grow without histidine supplied, and plate it on medium lacking histidine. Only cells acquiring a reverting mutation that restores the gene form colonies. Colonies against compound dose gives a slope — mutagenic potency. Add rat-liver microsomal extract ("S9"), because many compounds are not mutagenic until mammalian liver metabolism converts them. Different tester strains carry frameshift versus base-substitution starting mutations, so which strain responds tells you the mechanism.

It is a mutagenicity assay used as a carcinogenicity screen. The mapping is good but not perfect: non-genotoxic carcinogens pass it.

7. The rate, derived

From reference/verified-facts.md:

QuantityValue
Genome-wide SNV rate~1.1–1.3 × 10⁻⁸ per bp per generation
Recent long-read estimate1.30 × 10⁻⁸ per bp per generation
Fraction of of paternal origin~80%
Additional de novo mutations per year of paternal age~1.3–1.5

A human genome is about 6.2 × 10⁹ bp (2 × 3.1 Gb). Multiply:

low     1.1e-8 per bp x 6.2e9 bp = 68.2
high    1.3e-8 per bp x 6.2e9 bp = 80.6

So the arithmetic predicts roughly 68–81 new point per person that neither parent carried. Published trio studies typically report ~60–70. That gap is not an error — it is informative. Short-read trio studies can only call variants in the callable fraction of the genome: they exclude , and long repeats, which is on the order of 10–15% of the sequence. Apply that:

0.85 x 80.6 = 68.5

which lands squarely on the reported range. The naive product is the true number; the reported number is the true number minus what the technology cannot see. This is exactly why the long-read pedigree estimate sits at the top of the range — long reads recover the repetitive fraction, which is also the fraction that mutates fastest.

The parental split. About 80% are paternal: of ~70 mutations, roughly 56 from the father and 14 from the mother. The conventional explanation is cell divisions — the oocyte lineage undergoes a couple of dozen in total, while spermatogenesis continues through adult life, accumulating several hundred by age 30 and more thereafter, each one a replication opportunity. (Damage in non-dividing cells contributes too, so the picture is not purely division-driven, but the direction is unambiguous.)

Paternal age adds ~1.3–1.5 mutations per year. A child of a 45-year-old father carries roughly 20 × 1.4 ≈ 28 more de novo mutations than a child of a 25-year-old — a ~40% increase on a baseline of ~70.

Note what this does not say. Maternal age is the risk factor for (Ch 20) — a different mechanism entirely, failure in oocytes arrested since fetal life. Paternal age drives point mutations; maternal age drives errors.

Rate heterogeneity

The per-base rate is an average over a genome where the local rate varies severalfold:

This is not a footnote. Any test assuming a uniform background rate will generate false positives in high-rate regions. The canonical failure was in cancer genomics: early driver-gene scans flagged large, late-replicating, lowly-expressed genes as recurrently mutated when the excess was entirely local background rate. The fix — modelling the covariates of the background rate explicitly — is now standard, and it is the same lesson as covariate control in .

8. Mutation is undirected

Here is the misconception that survives everything else in this chapter.

is undirected: the probability that a particular mutation occurs is unrelated to whether it would be useful. Bacteria exposed to an antibiotic do not mutate toward resistance. A population under heat stress does not preferentially generate heat-tolerance . There is no mechanism by which the usefulness of a sequence change could feed back on the chemistry that produces it, and no such mechanism has ever been found.

This is the load-bearing distinction between mutation and selection. Mutation proposes, blind to consequence; selection disposes, on the basis of consequence. Collapse the two and you get Lamarckism, which is not merely historically wrong but mechanistically incoherent given how DNA replication works.

Two refinements, because the naive version of "random mutation" is also wrong:

Undirected is not uniform. Mutation is heavily biased in kind and in place — toward transitions, toward CpG, toward late-replicating regions, toward repeat tracts. "Random with respect to " is the precise claim; "uniform over the genome" is a different claim and it is false.

Stress can raise the rate without aiming it. Bacteria under stress upregulate error-prone polymerases, raising the genome-wide. That is a rate change, not a direction change: it increases the supply of variation of all kinds, most of it harmful. Reports that mutation rates are systematically lower in functionally important regions of some genomes describe another rate bias — a bias in where, not a bias toward what would help now.

The experiment that settled this is in the worked example below.

9. Repeat expansion and anticipation

Slippage at a short tandem repeat usually changes the tract by one unit. At a small number of loci, once the tract exceeds a threshold length, it becomes unstable: the expanded can form hairpin and slipped-strand structures that escape correction, and the tract grows further with each transmission. The becomes a function of the current allele length — a positive feedback.

DiseaseGeneRepeatLocationNormalPathogenic
Huntington diseaseHTTCAGcoding (polyglutamine)≤2627–35 intermediate; 36–39 reduced ; ≥40 fully penetrant
Fragile X syndromeFMR1CGG5' UTR<4545–54 grey zone; 55–200 ; >200 full
Myotonic dystrophy type 1DMPKCTG3' UTR5–3435–49 mutable normal (premutation); ≥50 fully penetrant
Friedreich FXNGAA 1<33~66–1300 ()

The mechanisms differ by location. In HTT the repeat is translated into a long polyglutamine tract that makes the protein aggregate — a gain of toxic function. In FMR1 the expanded CGG tract becomes methylated and the gene is silenced — a . Same class of mutation, opposite molecular logic.

follows directly. Onset age correlates inversely with repeat length, and repeat length tends to grow through transmission, so the disease appears earlier and more severely in successive generations. There is a parent-of-origin asymmetry: large HTT expansions overwhelmingly come through the father, whereas an FMR1 premutation expands to a full mutation only through the mother.

Worth noting for a statistically minded reader: anticipation was for decades dismissed as , and that scepticism was methodologically correct. If you ascertain families through an affected child, you will systematically find early-onset cases in the younger generation and late-onset cases in the older one, manufacturing apparent anticipation from nothing. The hypothesis was right to raise. It happened to be wrong here, and only molecular measurement of repeat length could show that.

If this section reads like the opening of a longer story, it is: D3 — Repeat-expansion disorders, in the optional track, expands the four-row table into the full disorder class — the mismatch-repair engine behind the instability, thresholds and premutations, and anticipation made quantitative.

10. Mutational signatures

Every mutational process leaves a characteristic fingerprint in the joint distribution of substitution type and sequence context. Encode a genome's as counts over 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. Each tumour becomes a 96-vector.

Factorise a matrix of many such vectors — non-negative matrix factorisation, which the reader already owns — and the recovered components are : recurring processes, each with a characteristic 96-channel profile, present in each tumour at some exposure level.

The catalogue is real and interpretable. C>T at CpG accumulating linearly with age (a clock, and exactly the deamination reaction of §5). C>T at dipyrimidines with CC>TT tandems in melanoma (UV). C>A in lung tumours of smokers (tobacco adducts). Distinctive profiles for APOBEC enzyme activity and for homologous-recombination deficiency in BRCA1/BRCA2-mutant tumours — the last of which is clinically actionable.

This is the payoff of everything above: because each chemical mechanism produces a biased, context-dependent spectrum, the spectrum is invertible. You can read the causes off the genome. Ch 56 develops it fully.


Common misconceptions

What people thinkWhat's actually true
are randomRandom with respect to — yes, and that is the whole point. Uniform over the genome — emphatically not. The rate varies severalfold with context, chromatin, replication timing and , and by 10–50× at CpG
Organisms mutate in response to a challengeLuria–Delbrück ruled this out in 1943. Stress can raise the , which increases variation of all kinds; nothing aims it. There is no channel by which "useful" could reach the chemistry
Damage and mutation are the same thingA lesion is chemical and repairable because the other strand still holds the answer. A mutation is a correctly paired sequence change with no evidence left that anything happened
are silentThey alter , speed and co-translational folding, mRNA stability, and may be non-synonymous in another isoform. Pathogenic synonymous variants are well documented
Older mothers are the main source of new mutations in a childPoint mutations are ~80% paternal and rise ~1.3–1.5 per year of paternal age. Maternal age drives — a different mechanism entirely
CpG mutates fast because methylation damages DNAMethylation does no damage. The point is the deamination product: 5-methyl-C deaminates to thymine, a normal base that repair cannot identify as the intruder. Unmethylated C deaminates to uracil, which is foreign and excised on sight
Mutagens cause distinctive kinds of mutation you would never otherwise seeThey raise the rate of lesion classes the cell already handles, with a bias. The bias is what makes readable — but the lesions are the same chemistry
A in a cancer gene means the family is at riskOnly variants are transmitted. Distinguishing the two requires sequencing a normal tissue alongside the tumour — which is why paired sequencing is the standard design

Worked example: the fluctuation test

The question, in 1943. Bacteria exposed to phage T1 mostly die; a few resistant colonies grow. Two hypotheses:

Both predict the same mean number of resistant colonies. Luria's insight was that they predict wildly different variances.

Prediction under H₁. Each of N plated cells independently converts with probability p at the moment of plating. Count ~ Binomial(N, p) ≈ Poisson(Np). Therefore Var = Mean, and the variance/mean ratio is 1.

Prediction under H₂. Grow a culture from a tiny inoculum to N cells. A mutation occurring at the k-th division, when the population is about k cells, leaves about N/k descendants by plating time. Mutational events are approximately Poisson over the ~N divisions, so the total count M is a compound Poisson with jump size N/k:

E[M]   = μN · (1/N) Σ_{k=1..N} (N/k)     = μN · H_N        ≈ μ N ln N
Var(M) = μN · (1/N) Σ_{k=1..N} (N/k)²    = μ N² Σ 1/k²      ≈ μ N² · π²/6

Var/Mean ≈ (π²/6) · N / ln N

The variance/mean ratio does not equal 1. It grows almost linearly with N. The reason is visible in the algebra: a mutation early in growth contributes an enormous clone, and those "jackpot" cultures dominate the second moment while barely affecting the first.

The design that separates them. Two sets of platings, same expected count:

H₁ predicts Var ≈ Mean in both. H₂ predicts Var ≈ Mean in Set B and Var ≫ Mean in Set A.

Illustrative numbers (patterned on the original result):

Set A  (independent cultures):  1, 0, 0, 7, 0, 303, 0, 1, 0, 48
Set B  (aliquots of one culture): 29, 41, 35, 38, 30, 44, 36, 33, 40, 34

Both sum to 360, so both have mean 36.0.

Set A:  Σ(x - x̄)² = 81,204      s² = 81,204 / 9 = 9,022.7      s²/x̄ = 250.6
Set B:  Σ(x - x̄)² =    208      s² =    208 / 9 =    23.1      s²/x̄ =   0.64

Test each against Poisson with the index-of- statistic, (n−1)s²/x̄ ~ χ² on 9 df:

Set B:   9 × 0.64  =    5.8   on 9 df   — entirely consistent with Poisson
Set A:   9 × 250.6 = 2,255    on 9 df   — p astronomically small

Conclusion. Set B behaves exactly as Poisson sampling should, confirming the assay is well behaved and the aliquots differ only by chance. Set A does not: a few cultures are jackpots, and jackpots are the signature of mutations that occurred generations before the phage was ever added. Resistance was pre-existing.

What makes the argument airtight is what it did without. Luria and Delbrück never observed a mutation, never sequenced anything, and had no molecular model of a gene. They inferred the timing of an unobservable event purely from the shape of the noise.

One caveat worth carrying: the Luria–Delbrück distribution is heavy-tailed enough that its theoretical variance is dominated by events too rare to ever observe, so estimating μ from a sample variance is unstable. Modern practice uses the proportion of cultures with zero mutants, or maximum likelihood on the full distribution.

Connections

Back to:

Forward to: