Contents
23 — Chromatin and epigenetics Part 4 — Gene regulation

23 — Chromatin and epigenetics

Before this: Ch 03 · Ch 22 · Ch 16 · Time: ~55 min

What you'll be able to do

The core idea

A liver cell divides and produces two liver cells. Nothing in the DNA sequence changed, so the information that says be a liver cell was carried by something else, through S , past a that strips the bare.

That is the entire problem: cellular memory that survives replication without being written in sequence.

The solution is a class of chemical marks on DNA and on the histone proteins DNA wraps around. But a mark alone is not memory. Replication halves it — each daughter chromatid gets roughly half the parental histones and one unmethylated new DNA strand. A mark is only heritable if the machinery that reads it also recruits the machinery that writes it. That read–write loop turns a passive label into a self-restoring fixed point.

So the useful mental model is not "tags on the genome". It is a per-locus two-state Markov chain with feedback: loss on replication pulls the state down, reader-recruited writers push it back up, and because the write rate depends on the current state, the chain is nonlinear and can have two stable fixed points. Marks with that loop are inherited. Marks without it are just fast, reversible of what is doing right now — and calling those "epigenetic" is where most of the confusion in this field starts.


1. The definition, and why the dispute matters

Two definitions are in circulation and they are not compatible.

Strict (Riggs, Holliday, and the tradition descending from Waddington's 1942 coinage): an epigenetic change is one that is heritable through cell division and not explained by a change in DNA sequence. is the whole content of the claim. The test is operational: remove the initiating signal, let the cell divide, and see whether the state persists.

Loose (common in the literature and universal in press releases): anything to do with chromatin. Any histone modification, any methylation change, any rearrangement.

The loose definition is a disaster because it makes "epigenetic" a synonym for "regulatory", and then every gene-expression result becomes an result. Worse, it lets a claim proved in the loose sense — stress changed histone acetylation at this — be reported in the strict sense — stress changed your grandchildren. The equivocation is the mechanism by which this field generates nonsense.

This chapter uses the strict definition. Where a mark is a heritable carrier of state, it is called epigenetic. Where it is a fast, signal-dependent , it is called chromatin state and nothing more. Under that discipline, the list of genuinely epigenetic mechanisms in mammals is short: , H3K9me3/HP1 , H3K27me3/Polycomb, CENP-A identity, and . Most of what you read about is not on that list.

2. The nucleosome as regulatory substrate

147 bp of DNA wrapped ~1.65 turns around an octamer of two copies each of histones H2A, H2B, H3 and H4, joined by 20–80 bp of linker. Repeat length in human chromatin is ~180–200 bp, so a genome of 6.2 Gb carries on the order of 30 million .

 linker      nucleosome        linker      nucleosome
────────╭──────────────╮──────────────╭──────────────╮────────
        │  147 bp DNA  │              │  147 bp DNA  │
        │   octamer    │              │   octamer    │
        ╰─┬──┬──┬──┬───╯              ╰─┬──┬──┬──┬───╯
          │  │  │  │   N-terminal tails │  │  │  │
         ac me ub ph   ← modified here  ac me ub ph

Two consequences do all the work.

Occlusion. DNA on the nucleosome surface is largely unavailable to sequence-specific and to . A 's is therefore set by whether a nucleosome sits on it. Active promoters have a nucleosome-depleted region of ~150 bp flanked by strongly positioned +1 and −1 nucleosomes; that depleted region is what detects (Ch 49).

The tails. The histone N-termini protrude from the particle, disordered and solvent-exposed. They are chemically decorated, and because they stick out, the decorations are readable by other proteins without disturbing the DNA. The nucleosome is simultaneously the packaging problem and the signalling surface.

3. Histone modifications, and the "code" that isn't

Acetylation is mechanistically direct

Lysine carries a positive charge at physiological pH, and that charge is part of what binds the histone octamer to DNA's polyanionic backbone (Ch 01). Acetylation transfers an acetyl group onto the lysine ε-amino group and neutralises the charge. Histone–DNA affinity drops, higher-order folding is destabilised, and the region opens. This is the one histone modification whose effect you can derive from electrostatics alone.

Writers are histone acetyltransferases (p300/CBP, GCN5); erasers are histone deacetylases (HDAC1–11, and the NAD⁺-dependent sirtuins). Turnover is fast — minutes. Acetylation also creates a binding surface for bromodomain proteins, including BRD4 and the TAF1 subunit of TFIID.

Fast turnover and no read–write loop means acetylation is a poor carrier of memory. It tracks the current activity of the locus. It does not remember it.

Methylation is pure reader recruitment

Lysine methylation adds no charge change and does essentially nothing to histone–DNA affinity. Its entire effect is that specific reader domains — chromodomains, PHD fingers, Tudor domains — bind methylated lysines and not unmethylated ones. Methylation can be mono-, di- or tri- (me1/me2/me3), and the states are read differently. Erasers are the LSD and JmjC-domain demethylases.

MarkWhereWriterReaderWhat it means
H3K4me3Active/poised , sharp peak at TSSCOMPASS (SET1A/B, MLL1–4)TAF3, ING PHD fingersCorrelate of initiation — largely a consequence of Pol II recruitment, not a cause
H3K27acActive and promotersp300/CBPBRD4Distinguishes active enhancers from merely poised ones
H3K27me3Polycomb-repressed developmental genesPRC2 (EZH2/EED/SUZ12)PRC1 CBX chromodomain; EEDFacultative repression. Self-propagating
H3K9me3Constitutive , repeats, ZNF genesSUV39H1/2, SETDB1HP1 chromodomainPermanent silencing. Self-propagating
H3K36me3Gene bodies, 3′-biasedSETD2, co-transcriptionalPWWP domains (DNMT3B, MRG15)Marks transcribed territory; suppresses cryptic internal initiation

Two other classes matter. Phosphorylation: H3S10ph by Aurora B during mitotic condensation, which also ejects HP1 in a "phospho/methyl switch"; H2A.X S139ph (γH2AX) spreading megabases around a double-strand break (Ch 17). Ubiquitination: H2AK119ub by PRC1, part of Polycomb repression; H2BK120ub by RNF20/40, which is required for H3K4 and H3K79 methylation — cross-talk between tails on different histones.

The histone code hypothesis, honestly

Strahl and Allis (2000) proposed that combinations of modifications constitute a code: distinct marks, written by distinct enzymes, read in combination by effectors, specifying distinct downstream outcomes. As a research programme it was enormously productive. As a description it overstates the case, in four specific ways.

  1. The symbols are not independent. H3K4me3, H3K9ac, H3K27ac and depletion co-occur at active promoters because they are all downstream of the same event. A code with correlated symbols is carrying far less information than the symbol count suggests. Empirically, genome-wide segmentation (ChromHMM and similar) recovers on the order of 15–25 recurrent chromatin states, not 2ᴺ.
  2. Many marks are consequences, not causes. H3K4me3 at promoters is deposited by complexes recruited to unmethylated and to initiating Pol II. Deleting the writers often barely perturbs expression. The mark reports more than it directs it.
  3. The few causal tests give split answers. Replacing the entire histone gene cluster in Drosophila with H3K27R mutants abolishes Polycomb silencing — K27 methylation is genuinely instructive. Equivalent experiments on other residues show much weaker effects. Some marks are load-bearing; some are passengers; the code framing does not distinguish them.
  4. Readout is context-dependent. H3K9me3 in pericentromeric repeat DNA and H3K9me3 over a ZNF gene body have different consequences. The same "letter" is read differently depending on what else is present and where it sits.

The property that actually separates the load-bearing marks from the rest is the read–write loop:

Both are positive feedback: read the mark, write more of it. That is why these two, alone among the histone modifications, propagate through replication and behave epigenetically in the strict sense.

4. Remodellers and histone variants

are not fixed. Four families of ATP-dependent remodellers move, evict and edit them:

FamilyHuman complexesWhat it does
SWI/SNFBAF, PBAF (SMARCA4, ARID1A, SMARCB1)Slides and ejects nucleosomes. Opens and . Subunits mutated in ~20% of human cancers
ISWINURF, CHRAC, ACFSpaces nucleosomes into regular arrays; assembly and maturation of chromatin behind the fork
CHDCHD1, NuRD (CHD3/4 + HDAC1/2 + MBD2/3)Spacing; NuRD couples remodelling to deacetylation and to methyl-DNA reading
INO80 / SWR1INO80, SRCAPDimer exchange — swaps H2A–H2B for variant-containing dimers

Histone variants substitute for the canonical histones and change the particle's properties:

5. DNA methylation

In mammals, a methyl group on the 5-carbon of cytosine — 5-methylcytosine — almost exclusively in the dinucleotide CpG (the "p" is the phosphate; it means C followed by G on the same strand, not a C:G pair). Neurons and embryonic stem cells also carry appreciable non-CpG methylation; plants methylate in CG, CHG and CHH contexts.

CpG symmetry is the whole trick

CpG is its own reverse complement. A methylated site is therefore methylated on both strands:

5'- ...  C  G ... -3'        5'- ... mC  G ... -3'
        |  |             →           |  |
3'- ...  G  C ... -5'        3'- ...  G mC ... -5'
    unmethylated                  symmetrically methylated

after replication:      5'- ... mC  G ... -3'   parent strand
                        3'- ...  G  C ... -5'   new strand   ← hemimethylated

Replication produces a hemimethylated site, which is an unambiguous instruction: one strand says the mark belongs here. UHRF1 recognises hemimethylated CpG and recruits DNMT1, which methylates the new strand. That is the maintenance system, and it is a state machine with a self-evident restore rule.

stateDiagram-v2 [*] --> U: unmethylated CpG U --> H: DNMT3A/3B<br/>(de novo, signal-driven) H --> M: DNMT1 + UHRF1<br/>(maintenance) M --> H: replication<br/>(dilution) H --> U: replication without DNMT1<br/>(passive demethylation) M --> U: TET1/2/3 → 5hmC → 5fC → 5caC<br/>→ TDG + base excision repair<br/>(active demethylation)

DNMT3A and DNMT3B are the de novo enzymes, assisted by the catalytically dead cofactor DNMT3L. TET1/2/3 are dioxygenases that oxidise 5mC stepwise to 5hmC, 5fC and 5caC; thymine DNA glycosylase excises the last two and restores unmodified cytosine. Demethylation is therefore either passive (replication without maintenance) or active (oxidation plus repair).

CpG islands, and how methylation silences

Roughly 28–32 million CpG dinucleotides per haploid genome, of which 60–80% are methylated in a typical somatic cell. The unmethylated minority is concentrated in : stretches of ≥200 bp with GC content ≥50% and observed/expected CpG ≥0.6. There are roughly 30,000 in the human genome, and more than half of them sit at ; conversely more than 60% of human promoters contain one.

Three distinct silencing mechanisms, often conflated:

  1. Direct occlusion. Some cannot bind a methylated site. is the canonical case and the mechanism behind imprinting (§6).
  2. Reader recruitment. Methyl-CpG-binding domain proteins — MeCP2, MBD1–4 — bind 5mC and pull in NuRD/HDAC complexes and co-.
  3. Gene-body methylation, which is not silencing at all. SETD2 deposits H3K36me3 co-transcriptionally; DNMT3B's PWWP domain reads H3K36me3 and methylates the underlying DNA. Gene-body methylation therefore correlates positively with expression and suppresses spurious internal initiation.

The rule "methylation means off" is true at promoters and false in gene bodies. Any tool or intuition that treats a genome-wide average methylation level as a proxy for repression is averaging two opposite signals.

Why the genome is CpG-poor — a fossil of germline methylation

Cytosine deaminates spontaneously to uracil, which is foreign to DNA and is excised efficiently by uracil-DNA glycosylase. But 5-methylcytosine deaminates to thymine, which is a perfectly legitimate DNA base. The repair machinery sees a T:G mismatch and has no way to know which strand is wrong. The result is a C→T transition at roughly an order of magnitude above the background transition rate (Ch 16) — methylated CpG is the single most mutable position in the genome.

Run that for a few hundred million years and it shows up in composition. Human GC content is ~41%, so under independence you would expect CpG at ≈ 0.205 × 0.205 ≈ 4.2% of dinucleotides. Observed frequency is about 1% — roughly a fifth of expectation.

CpG islands are the exception precisely because they are the regions kept unmethylated in the , and therefore never exposed to the accelerated . The genome's dinucleotide composition is a several-hundred-million-year record of which sequences were methylated in germ cells.

6. Genomic imprinting

For most genes you express both parental copies. For roughly 150 human genes, expression depends on which parent the came from — one copy is silenced, deterministically, by parental origin. This is a flat contradiction of Mendel: reciprocal crosses are not equivalent.

The mark lives at an imprinting control region (ICR), a differentially methylated region whose methylation is established in the — in oogenesis or in spermatogenesis, according to sex — and which then survives the post-fertilisation reprogramming wave (§8) because ZFP57/TRIM28 specifically protects it.

IGF2/H19 — the CTCF insulator

At 11p15.5, IGF2 (paternally expressed growth factor) and H19 (maternally expressed ) share a set of downstream of H19. Between them sits ICR1, containing seven binding sites in humans. CTCF binding is abolished by methylation of those sites, so the ICR's methylation state acts as a switch on whether an insulator forms:

MATERNAL  (ICR1 unmethylated → CTCF binds → insulator blocks enhancers)

   IGF2          ICR1          H19        enhancers
   [ OFF ]══════[CTCF]═══════[ ON ]═══════[ E E ]
                  ▲                          │
                  └────── blocked ───────────┘

PATERNAL  (ICR1 methylated → no CTCF → enhancers reach IGF2; H19 promoter also methylated)

   IGF2          ICR1          H19        enhancers
   [ ON  ]══════[ mmmm ]═════[ OFF ]══════[ E E ]
      ▲                                      │
      └──────────── enhancer contact ────────┘

The prediction falls straight out, and both directions occur clinically: gain of methylation on the maternal ICR1 makes it behave paternally, giving two active IGF2 copies and the overgrowth of Beckwith–Wiedemann syndrome; loss of methylation on the paternal ICR1 makes it behave maternally, giving no IGF2 and the growth restriction of Silver–Russell syndrome. One switch, both signs, opposite .

Prader–Willi and Angelman

Both arise from the same imprinted domain at 15q11–q13, and differ only in which parent's copy is lost.

Prader–WilliAngelman
Copy lostPaternalMaternal
Critical locusSNORD116 snoRNA cluster (paternally expressed)UBE3A (maternally expressed in neurons only)
Deletion of 15q11–q13~65–75% of cases~65–75% of cases
maternal UPD, ~20–30%paternal UPD, ~3–7%
Imprinting-centre defect~1–3%~3–10%
Point in a single gene—UBE3A variants, ~5–11%

In neurons the paternal UBE3A is silenced in cis by an antisense transcript, UBE3A-ATS, running through it from the Prader–Willi imprinting centre — which is why an that truncates UBE3A-ATS and unsilences the intact paternal allele is a live therapeutic strategy.

These two syndromes are the proof, in humans, that the maternal and paternal genomes are not functionally interchangeable.

Why imprinting evolved: the kinship (conflict) theory

Haig's argument is a straightforward asymmetry in relatedness. In a species where a mother provisions offspring after fertilisation, and where a female may have offspring by more than one male, a gene copy is related to the mother's future offspring differently depending on which parent it came from. A paternally derived allele is likely not present in the mother's later offspring by other males, so its optimal strategy is to extract maternal resources aggressively. A maternally derived allele is present in half of those future offspring, so its optimum is more restrained.

The prediction: paternally expressed genes promote growth and maternal resource extraction; maternally expressed genes restrain it. IGF2 is paternally expressed and growth-promoting; H19 and CDKN1C are maternally expressed and growth-restraining; in mouse the IGF2-sequestering receptor Igf2r is maternally expressed.

The strongest evidence is convergent: imprinting arose independently in placental mammals and in the endosperm of flowering plants — the two lineages with post-fertilisation maternal provisioning — and not in birds or in egg-laying lineages generally. Two independent origins under the same selective condition is a much better test than any single locus.

It is not a complete theory. Not every imprinted gene fits the growth prediction (human IGF2R is not imprinted at all), and co-adaptation and sexual-antagonism models explain some loci better. Treat kinship theory as the best-supported general explanation, not as a law.

7. X inactivation

Females carry two X , males one. Mammals equalise dosage by silencing one X per cell, and the mechanism is the largest-scale strictly-epigenetic event in the genome.

XIST, a ~17 kb transcribed from the X-inactivation centre at Xq13, is expressed only from the X destined for inactivation and coats that chromosome in cis. It then recruits, in layers: SPEN → HDAC3 (deacetylation), PRC1/PRC2 (H2AK119ub, H3K27me3), macroH2A, SMCHD1, and finally of the CpG-island across the chromosome. Each layer is individually dispensable; together they form a lock that is essentially irreversible in somatic cells.

The choice is random per cell in the early epiblast, and then clonally inherited. A heterozygous female is therefore a mosaic of two cell populations — visible directly in the coat of a calico cat, and clinically important because it makes disease expression in heterozygous females variable and skew-dependent (Ch 13). Roughly 15–25% of X-linked genes escape inactivation.

Same sequence, two heritable states, hundreds of divisions. This is what the strict definition is for.

8. Reprogramming: the two erasures

Mammalian development wipes twice.

graph LR G["gametes<br/>highly methylated<br/>imprints set"] --> Z["zygote → blastocyst<br/><b>Wave 1</b><br/>paternal: active, TET3<br/>maternal: passive dilution"] Z --> S["post-implantation<br/>re-methylation<br/>DNMT3A/3B"] S --> P["primordial germ cells<br/><b>Wave 2</b><br/>global erasure to ~7–15%<br/><i>imprints erased here</i>"] P --> R["gametogenesis<br/>imprints re-set by<br/>sex of the individual"] R --> G

Wave 1, after fertilisation. The paternal genome is stripped rapidly — protamines are exchanged for histones and TET3 oxidises 5mC — while the maternal genome is diluted passively over the first cleavage divisions. Imprinted DMRs are specifically protected by ZFP57/TRIM28. Methylation is re-established after implantation.

Wave 2, in the primordial germ cells. Between roughly E9.5 and E13.5 in mouse, PGC methylation falls to about 7–15% genome-wide — the most complete erasure in the life cycle. Crucially, imprints are erased here, and then re-established during gametogenesis according to the sex of the individual making the gametes, not the sex of the parent they came from.

What escapes. Some IAP and other LTR retain substantial methylation through both waves; a small set of single-copy loci and rare LTR-ERV1 elements do too. That set is small, largely repeat-associated, and is the entire mechanistic space available for transgenerational inheritance of methylation in mammals.

9. Transgenerational epigenetic inheritance

First, get the generation counting right, because most popular claims fail here alone.

If a pregnant F0 female is exposed, the F1 fetus is directly exposed, and the germ cells that will make F2 are inside that fetus and are also directly exposed. F3 is the first unexposed generation. If an F0 male is exposed, his sperm are exposed, so F2 is the first unexposed generation. Anything at F1 or F2 (maternal) is intergenerational — a direct exposure effect — not transgenerational.

The two reprogramming waves are the reason this is hard, and they are not a technicality. For an environmentally induced chromatin mark to reach an unexposed grandchild it must be (i) written in the , (ii) survive erasure to ~10% genome-wide methylation in primordial germ cells, (iii) survive the post-fertilisation wave, and (iv) still alter . The known escapees from that gauntlet are mostly . Any claim of human is a claim that all four steps happened, and essentially none of the human literature tests even one of them.

Where the evidence is strong: plants and nematodes. Plants have no sequestered germline — somatic meristem cells become gametes — and no genome-wide reprogramming wave, and they possess RNA-directed . The Linaria vulgaris peloric epiallele is a hypermethylated Lcyc gene producing radially symmetric flowers; Linnaeus described it in 1742 and it is still segregating. In C. elegans, piRNA-triggered silencing is amplified by an RNA-dependent into secondary siRNAs loaded onto the nuclear Argonaute HRDE-1, and is inherited for tens of generations. Note what the nematode mechanism actually is: an RNA amplification loop, not a chromatin mark being copied.

Where it is weak: mammals. The mouse agouti viable yellow (A^vy) is the textbook case — coat colour varies with the methylation of an IAP retrotransposon inserted upstream of Agouti, and maternal methyl-donor supplementation shifts the offspring distribution. It is real. It is also (a) intergenerational, (b) entirely dependent on a specific transposon insertion — i.e. on a DNA sequence difference — and (c) a metastable epiallele, not an environmental instruction.

The best-supported mammalian vector for paternal effects is not chromatin at all: small RNAs in sperm, particularly tRNA-derived fragments, which change with paternal diet and can transmit metabolic phenotypes when injected into zygotes.

The Dutch Hunger Winter, specifically

The Dutch famine of 1944–45 is the study everyone cites. Heijmans et al. (2008) found that individuals conceived during the famine had, six decades later, 5.2% lower methylation at the IGF2 DMR than their own unexposed same-sex siblings, and that the association was specific to periconceptional exposure.

That is a careful, well-designed study. It is not evidence of transgenerational epigenetic inheritance, for six independent reasons:

Claim it is used to supportWhy it does not support it
Effects transmitted to unexposed generationsThe subjects were exposed in utero. This is F1. Generation counting alone disqualifies it
A large biological effect5.2% mean difference at one DMR, in blood, with overlapping distributions
Methylation caused the metabolic phenotypeNo causal link was tested. The methylation difference and the health outcomes could both be downstream of the same prenatal insult
It shows environment writing the epigenomeBlood is a mixture of cell types; famine exposure that shifts leukocyte composition produces apparent methylation differences with no change in any cell's methylation. Cell-composition confounding is the artefact in blood methylation studies
The effect is environmentalMethylation at most DMRs is under substantial genetic control (mQTLs). A follow-up at this locus found famine exposure and independently and additively associated with methylation
Sibling design controls for confoundingIt controls for shared genetics and household, not for birth order, maternal age, gestational timing, or postnatal environment

The honest summary: prenatal famine has real, lifelong health consequences, and there are measurable methylation correlates. Neither fact establishes that a chromatin mark is the carrier, and neither is transgenerational.

10. Epigenetic clocks

Take methylation β-values (fraction methylated) at a few hundred thousand CpGs, regress chronological age on them with an elastic net, and you get a startlingly good predictor. Horvath's 2013 multi-tissue clock uses 353 CpGs, was trained on ~8,000 samples across 51 tissues and cell types, and achieves correlation ~0.96 with an error (median absolute difference) of about 3.6 years in test data.

Three generations of clock, doing different things:

Clock typeTrained onExample
First generationChronological ageHorvath (353 CpGs), Hannum (71 CpGs, blood)
Second generationMortality / clinical biomarkersPhenoAge, GrimAge — better mortality predictors than chronological-age clocks
Third generationLongitudinal rate of changeDunedinPACE

You already know how to read this. "Epigenetic age acceleration" is the residual from regressing predicted age on chronological age. So:

Clocks are genuinely useful — forensic age estimation, mortality risk, comparing tissue ageing rates, quantifying the near-total age reset in induced pluripotent stem cells. They are measurements, not mechanisms.

11. Epigenetics in cancer

Chromatin is one of the most frequently mutated functional categories in cancer, and the details belong to Ch 56. The shape of it:

Because these are enzymatic and reversible, they are drug targets: azacitidine and decitabine (DNMT), vorinostat and romidepsin (HDAC), tazemetostat (EZH2).


Common misconceptions

What people thinkWhat's actually true
means environment changing your genesIt means heritable change in gene expression without sequence change. Most environmental effects on chromatin are neither heritable nor stable, and calling them epigenetic is the equivocation that drives most public confusion
Your experiences are written into your grandchildren's epigenomeIn mammals this requires surviving two genome-wide reprogramming waves. The known escapees are mostly . No human study has demonstrated it, and almost none have tested the required steps
The Dutch Hunger Winter proves transgenerational inheritanceThose subjects were exposed in utero — F1, not F3. The reported effect is a 5.2% methylation difference at one DMR in blood, uncorrected causally, with cell-composition and mQTL confounding unresolved
Epigenetic marks are erased, so nothing is inheritedAlso wrong, in the other direction. Imprinted DMRs are protected through wave 1, and specific retrotransposon families resist both waves. The correct statement is "the channel is narrow and mostly repeat-associated", not "the channel is closed"
switches genes offTrue at . False in gene bodies, where methylation is deposited co-transcriptionally and correlates positively with expression
Histone modifications constitute a codeThey are heavily correlated with each other and with , resolve into ~15–25 recurrent states rather than 2ᴺ, and many are consequences of transcription rather than causes. "Code" implies orthogonal symbols and deterministic readout; neither holds
All chromatin marks are heritableOnly marks with a read–write loop propagate: H3K27me3 (EED→EZH2), H3K9me3 (HP1→SUV39H1), 5mC (UHRF1→DNMT1), CENP-A. Acetylation turns over in minutes and remembers nothing
Epigenetic changes are reversible, genetic ones aren'tMethylated CpG is the most mutable position in the genome — 5mC deaminates to a legitimate T. Epigenetic marks routinely become permanent sequence changes. That is why the genome is CpG-depleted
Identical twins are epigenetically identicalThey diverge with age through stochastic maintenance error, and were never identical to begin with — choice alone is random per cell
"Epigenetic age acceleration" measures how fast you are ageingIt is a regression residual. It inherits every confounder of its features, notably blood cell composition, and its CpGs are one arbitrary representative set from many correlated equivalents
Imprinting is a rare curiosity~150 human genes, two named syndromes from one 15q11–q13 locus, two more from one 11p15.5 locus, and it is the reason and parthenogenesis fail in mammals
position is determined by DNA sequenceNeocentromeres form on ordinary unique sequence with no alpha satellite, and are inherited stably. Centromere identity is propagated by CENP-A — genuinely epigenetic

Worked example: does a methylation mark survive 40 cell divisions?

Model one CpG as a two-state Markov chain over cell divisions. Let

Step 1 — maintenance alone. Set r = 0. Survival after n divisions is (1 − q)ⁿ. Measured DNMT1 maintenance fidelity is roughly 0.96–0.99 per site per division, so q ≈ 0.01–0.04:

Fidelity 1 − qAfter 40 divisions
0.999 (hypothetical — above the measured range, for contrast)0.999⁴⁰ = 0.96
0.990.99⁴⁰ = 0.67
0.960.96⁴⁰ = 0.20

Read the measured rows carefully, because the sloppy version of this argument overstates them. At 0.96, four-fifths of the mark is gone in 40 divisions; at 0.99, two-thirds of it survives. Forty divisions is not where maintenance fails. The failure is over the lifetime of a lineage: a stem cell lineage runs for hundreds of divisions, and 0.99³⁰⁰ ≈ 0.05 — even at the top of the measured range, copying alone leaves essentially nothing. Maintenance is not sufficient, and this is a real quantitative problem, not a rounding error.

Step 2 — add de novo restoration. With both rates, the chain has a stationary probability of being methylated:

m* = r / (r + q)          relaxation time = 1 / (r + q) divisions

Take q = 0.02 and r = 0.20 (readers of the surrounding methylated block recruiting DNMT3A/3B). Then m* = 0.20 / 0.22 = 0.91, independent of the starting state. The relaxation time is 1/(r + q) ≈ 4.5 divisions — a perturbed locus returns to its set point within about five divisions.

Step 3 — see what changed conceptually. The heritable object is not the mark on any individual CpG. It is the equilibrium of a feedback loop, and it is defined over a block of sites — a has 50–100 of them. Redundancy plus feedback is an error-correcting code: individual sites flip constantly, the block-level state does not.

Step 4 — the loop is nonlinear, and that is why there are two states. r is not a constant. De novo methyltransferases are recruited by readers of nearby methylation and of H3K9me3, so r increases with the local methylated fraction. Substituting r(m) into m* = r/(r+q) turns a single fixed point into a system that can have two stable fixed points — a methylated attractor and an unmethylated attractor — separated by an unstable threshold. Bistability, not a stored bit, is what cellular memory actually is. Precisely the same argument applies to PRC2 (EED reads, EZH2 writes) and to HP1/SUV39H1.

Step 5 — the residual error is measurable, and you have already met it. Sites where the feedback is weakest have r and q closest together, so m* sits near 0.5 and the variance across cells is maximal. Those sites with cumulative divisions and time. That drift is exactly the signal an regresses on — which is why clocks work at all, and why what they measure is accumulated stochastic maintenance error rather than a programme.

Connections