Contents
22 — Eukaryotic transcriptional regulation Part 4 — Gene regulation

22 — Eukaryotic transcriptional regulation

Before this: Ch 21 · Ch 05 · Ch 03 · Time: ~45 min

What you'll be able to do

The core idea

A bacterium regulates a gene the way you would guard a function: put a conditional immediately in front of it, evaluate one or two inputs, proceed. The is exposed, can find it, and regulation mostly consists of blocking something that would otherwise happen (Ch 21).

Eukaryotes inverted this. DNA is wrapped in and folded; a promoter is not exposed unless something has made it exposed. Nothing happens by default. Getting a gene transcribed therefore requires assembling, at one place and one time, a coalition: several sequence-specific proteins bound to elements that may be scattered over a megabase, plus chromatin-modifying machinery they recruit, plus a bridging complex that finally contacts the polymerase.

Bacterial regulation asks "should this gene be switched off?" Eukaryotic regulation asks "has enough independent evidence accumulated to justify switching this gene on?" Default-off is the single reframe that generates everything else in this chapter: the need for many inputs, the tolerance of long distances, the existence of factors specialised for prising chromatin open, and the fact that the same 19,442 protein-coding genes can run thousands of distinct cell-type programmes.


1. What actually changes from bacteria

BacteriaEukaryotes
SubstrateNaked-ish DNA, accessibleNucleosomal chromatin, default inaccessible
CompartmentNone — and coupledNucleus separates them; an RNA-processing step intervenes
Element positionWithin ~100 bp of the start siteCore promoter plus elements up to megabases away, in either orientation
Typical inputs per gene1–3 regulators5–20 distinct factors at multiple elements
Regulator countA few hundred TFs in E. coli1,639 sequence-specific TFs in human
Co-regulation: one transcript, many genesNo operons. Each gene has its own promoter
Bridge to polymeraseActivators often touch directlyAlmost never direct — coactivators and Mediator intervene

The operon row is worth dwelling on. Bacteria bundle co-regulated genes into one transcription unit, so a single switch controls the set. Eukaryotes gave that up and achieve co-regulation instead by giving physically unlinked genes the same binding-site sequences — a broadcast pattern rather than a shared object. More expensive per gene, far more flexible: a gene joins or leaves a regulon by gaining or losing a site, without being moved.

The exception proves the rule. Nematodes retained operons — more than 17% of C. elegans genes sit in polycistronic units — but only by inventing a whole RNA-processing mechanism to make them work: the downstream cistrons are separated post-transcriptionally by trans- an SL2 spliced leader onto each.

2. The cis-regulatory parts list

ElementSizePositionJob
Core ~50–100 bp−40 to +40 of the TSSPositions and orients the pre-initiation complex
Proximal promoter~100–500 bpImmediately upstreamConstitutive/tissue TF sites; sets baseline
~200–1,000 bpAnywhere, either orientationRaises of a target promoter
Silencer~200–1,000 bpAnywhereLowers it; recruits corepressors
Insulator / boundary~200 bp–few kbBetween domainsBlocks enhancer action across it; usually -bound
Locus control region~10–50 kbOften far upstreamA cluster of elements that opens and drives a whole locus

Core promoters are less standardised than textbooks imply. Only about 10% of human promoters contain a ; roughly 60% are instead CpG-island promoters with dispersed, imprecise start sites. The Initiator (Inr), downstream promoter element (DPE), TFIIB-recognition element (BRE) and motif ten element (MTE) appear in various combinations, and the combination is itself regulatory — see §9.

The ENCODE registry of candidate cis-regulatory elements now lists 2,348,854 human cCREs on GRCh38, up from 926,535 in the previous release. Divide by 19,442 protein-coding genes: roughly 120 candidate regulatory elements per gene — an upper bound on real regulation and a lower bound on the interpretation problem.

The canonical long-range element is the ZRS, which drives SHH in the developing limb bud from about 1 Mb away, on 7q36.3, inside 5 of a different gene (LMBR1). Point in it cause preaxial polydactyly — extra digits — leaving SHH itself untouched.

3. Enhancers, and the distance paradox

Three properties define an operationally, and all three were shocks when found:

  1. Position independence — it works upstream, downstream, or inside an of its target or of a neighbour.
  2. Orientation independence — flip it and it still works. are not like this; they are directional.
  3. Distance tolerance — tens of kb routinely, up to ~1 Mb in cases like SHH.

Taken literally this is absurd. One megabase of B-form DNA has a contour length of 10⁶ × 0.34 nm = 340 μm, in a nucleus about 6 μm across. The linear distance cannot be the operative variable, because the molecule is not linear in the nucleus.

The resolution is looping. The DNA between enhancer and promoter is extruded into a loop, bringing two sequences that are a megabase apart in coordinates into physical contact:

linear coordinates
  ──[ENH]────────────────── 1 Mb of intervening DNA ──────────────────[PROM]──[gene]──►

in the nucleus
                    ╭──────────────────────╮
                    │   intervening DNA    │      loop extruded by cohesin,
                    │  (extruded as loop)  │      constrained by CTCF boundaries
                    ╰──╮                ╭──╯
  ──[ENH]══════════════╯                ╰═══════════[PROM]──[gene]──►
        └──── TF + coactivators + Mediator bridge ────┘

So the operative variable is not genomic distance but contact frequency, which is what and Micro-C measure (Ch 50). It decays with genomic separation — hence most enhancers sit near their targets — but it is also structured: topologically associating domains (), bounded by -bound insulators and shaped by -driven , make within-domain contacts far more likely than across-boundary ones.

That makes boundaries load-bearing. Deleting the TAD boundary near EPHA4 lets limb enhancers reach genes they normally cannot — enhancer hijacking — producing limb malformations without altering a single coding base. that remove boundaries activate the same way.

4. Enhancer–promoter specificity: still open

Given that an can act over a megabase, why doesn't it activate every it passes? We do not fully know. What we do know constrains the answer:

The honest statement for 2026: given an enhancer's sequence and the promoters in its , we cannot reliably predict which one it drives.

5. Transcription factors are two separable modules

A sequence-specific has, at minimum, a DNA-binding domain (DBD) that recognises a short sequence and an activation (or repression) domain that recruits machinery. The two are on the same polypeptide and are otherwise almost independent.

Human TFs (1,639 of them) fall into 25 DBD families, dominated by a few:

DBD familyHuman TFsRecognition strategy
C2H2 zinc finger747Tandem fingers, each contacting ~3 bp in the — modular and extensible
Homeodomain196Three-helix bundle; helix 3 in the major groove. Developmental patterning
bHLH108Dimerises via helix–loop–helix; basic region reads an E-box (CANNTG)
bZIP54Leucine zipper dimerises; basic regions grip the major groove as a clamp
Forkhead49Winged helix — the fold that makes FOXA1 a (§7)
Nuclear receptor46Zinc-finger DBD plus a ligand-binding domain — a TF with a built-in sensor

Almost all of them read the major groove, for the reason derived in Ch 01: it exposes enough of each base pair's edge to identify it without opening the helix.

Activation domains are the opposite of the DBDs in every respect. They are typically intrinsically disordered, often acidic, poorly conserved in sequence, and interchangeable. An activation domain does not act on DNA or on polymerase; it is a recruitment address.

Separability is why this field became engineerable. Fuse the GAL4 DBD to the herpesvirus VP16 activation domain and you get an activator that works in yeast, flies and human cells. Split the two halves onto separate proteins and only fires when those proteins interact — that is the assay. Replace the DBD entirely with a catalytically dead targeted by a guide RNA and you get CRISPRa/CRISPRi (Ch 38). None of this would work if binding and activation were one entangled function.

6. Motifs, PWMs, and why prediction fails

A TF's preference is summarised as a position weight matrix: for each position i and base b, a log-odds score log₂(f_ib / p_b). Score a window by summing; call a match above a threshold. Standard, well-behaved, and nearly useless on its own.

The reason is information-theoretic. A 's total information content is

I = Σ_i ( 2 − H_i )    bits,     H_i = − Σ_b f_ib log₂ f_ib

A perfectly specified position contributes 2 bits; a fully degenerate one contributes 0. Under a uniform background the probability a random position matches is ≈ 2^(−I), so the expected number of genome-wide matches on both strands is

E[hits] ≈ 2L · 2^(−I)

Real motifs are 6–15 bp and degenerate, giving I of roughly 8–15 bits. With L = 3.1 × 10⁹:

I (bits)Expected matches
82.4 × 10⁷
106.1 × 10⁶
121.5 × 10⁶
151.9 × 10⁵

A typical TF actually occupies 10³–10⁵ sites by . Wasserman and Sandelin named the consequence the futility theorem: a genome-wide PWM scan yields on the order of 1,000 false positives per functional site, so essentially every predicted binding site is non-functional.

This is exactly the point made in Ch 01 about specificity being relative. Individual factors are sloppy; the genome is enormous; a 10-bit preference cannot pick one site out of 6 × 10⁹. Specificity is manufactured downstream, by two multiplicative filters — combinatorial requirement and chromatin — quantified in the worked example below.

7. What TFs recruit: coactivators, remodellers, and pioneers

A bound activator rarely touches II. It recruits, and what it recruits does the work.

graph LR TF["sequence-specific TF<br/>bound at enhancer"] PIO["<b>pioneer factor</b><br/>binds nucleosomal DNA"] REM["<b>remodellers</b><br/>SWI/SNF, ISWI, CHD, INO80<br/><i>ATP-driven: slide, eject, swap</i>"] MOD["<b>histone modifiers</b><br/>p300/CBP → H3K27ac<br/>COMPASS/MLL → H3K4me1"] MED["<b>Mediator</b><br/>26-subunit core + CDK8 module"] PIC["pre-initiation complex<br/>TFIID/A/B/E/F/H + Pol II"] TXN["transcription"] PIO --> REM TF --> REM TF --> MOD TF --> MED REM --> PIC MOD --> PIC MED --> PIC PIC --> TXN

Mediator is the physical bridge. Its ~26-subunit core is organised into head, middle and tail modules: the tail contacts activator domains, the head contacts Pol II and its C-terminal domain. A separable four-subunit CDK8 kinase module (MED12, MED13, CDK8/CDK19, cyclin C) associates reversibly and modulates the whole thing. Mediator is how a disordered acidic patch bound a megabase away ends up influencing polymerase.

Chromatin remodellers use ATP to reposition, evict or swap , converting a closed site into an accessible one. Histone modifiers write the covalent marks — acetylation by p300/CBP, methylation by COMPASS-family complexes — that both loosen chromatin and act as docking sites for reader proteins (Ch 23). Repression is the mirror image: corepressors such as NCoR/SMRT, Sin3–HDAC, NuRD and Polycomb complexes are recruited by repression domains and deacetylate, methylate or compact.

Pioneer factors solve the bootstrap problem

Everything above is circular. Most TFs cannot bind DNA wrapped on a nucleosome; opening a nucleosome requires a remodeller; recruiting a remodeller requires a bound TF. How does a new programme ever start — in an embryo, in a differentiating cell, in a reprogramming experiment?

can engage their motifs on nucleosomal DNA. FOXA1's winged-helix DBD structurally resembles linker histone H1 and binds the exposed face of a nucleosome; GATA factors, OCT4, SOX2, KLF4, ASCL1 and PU.1 have comparable ability. Having bound, they recruit remodellers, open the site, and license the ordinary factors that could not have bound first. This is why the Yamanaka cocktail can reprogramme a fibroblast: OCT4, SOX2 and KLF4 need no permission from the existing chromatin state. The fourth Yamanaka factor, MYC, is not a pioneer — on its own it engages already-open chromatin, and it reaches closed sites only in the company of the other three. Two real caveats — pioneer activity is a spectrum, not a binary, and it is locus- and context-dependent.

8. Combinatorial grammar, and signals that arrive from outside

How much does the arrangement of sites within an matter? Two limiting models, both supported by real examples:

EnhanceosomeBillboard
ExampleHuman interferon-β enhancer, ~55 bpDrosophila even-skipped stripe 2, ~500 bp
OccupantsATF-2/c-Jun, IRF3/IRF7, NF-κB, HMGA1Bicoid, Hunchback (activators); Giant, Krüppel ()
Spacing/orientationCritical — a rigid, cooperative surfaceTolerant — sites can be shuffled, gained and lost
LogicStrict AND gate; all-or-nothingRoughly additive, sub-elements semi-autonomous
EvidenceSingle-bp insertions abolish activitySequence diverges heavily between Drosophila species while output stays conserved

A third model, the TF collective, covers enhancers where co-occupancy depends on protein–protein and chromatin context rather than on any motif grammar at all. Reality is a continuum, and the practical consequence is that enhancer strength in massively parallel assays is mostly predictable from motif content, with real but modest grammar effects on top.

Signal-responsive regulation is where this meets the outside world. Three architectures cover most of it, sharing one design principle: keep the synthesised but inactive, so the response needs no new protein synthesis.

FamilyLatency mechanismRelease
Nuclear receptors (GR, ER, RAR)Cytoplasmic, held by HSP90 chaperone; or DNA-bound with a corepressorLipophilic ligand diffuses through the membrane and binds the ligand-binding domain → nuclear import, corepressor→coactivator swap
NF-κBCytoplasmic, IκB masks its nuclear localisation signalSignal → IKK phosphorylates IκB → ubiquitination → proteasomal destruction → NF-κB released. Minutes
STATCytoplasmic, unphosphorylated monomerReceptor-bound JAK phosphorylates a STAT tyrosine → SH2 domain of one STAT binds the phosphotyrosine of another → dimer → nucleus

Nuclear receptors are the shortest path from a molecule to a transcriptional response in all of biology: a steroid hormone crosses the membrane and converts a transcription factor directly into its active form. No second messenger, no kinase cascade. NF-κB adds a canonical negative feedback — it induces of its own inhibitor IκB — which turns a step input into a damped oscillation.

9. Condensates and bursting: the stochastic layer

. Mediator, Pol II, BRD4 and many activation domains form microscopically visible clusters at highly active loci, and purified activation domains — being multivalent disordered polymers — undergo liquid–liquid separation in vitro. The attractive model is that happens in condensates that concentrate machinery, buffer concentration fluctuations, and let an act at a distance by nucleating a shared droplet.

Treat this as an open question, not a result. The clustering is real and reproducible. Whether it is phase separation in the thermodynamic sense is contested: many transcriptional "condensates" contain only tens of molecules, far below the scale at which classical phase behaviour is well defined; the standard perturbations (1,6-hexanediol, optogenetic droplet induction) are blunt and have off-target effects; multivalent-binding "hub" models reproduce most observations without any phase transition; and some studies find that forcing condensate formation inhibits transcription. As of 2026 this is actively argued in the primary literature. A textbook presenting it as settled is ahead of the evidence.

is on much firmer ground. Live imaging and both show that genes do not transcribe at a steady rate; they fire in bursts separated by silence. The standard model is a two-state random telegraph:

   OFF ──k_on──►  ON ──k_syn──► mRNA ──δ──► degraded
       ◄─k_off──

Burst frequency is set by k_on, burst size by k_syn/k_off (geometric, in molecules per burst). The steady-state mRNA count is a gamma-mixed Poisson — i.e. , with Fano factor above 1. If you have ever wondered why single-cell RNA-seq count models are negative binomial rather than Poisson, this is the mechanism, not a statistical convenience.

-resolved single-cell measurements localised the two parameters to different elements: enhancers set burst frequency; core elements set burst size; and cell-type differences in expression are predominantly differences in burst frequency. So an enhancer does not "turn a gene on" — it increases how often the gene fires. Bulk sees only the mean — exactly (k_syn/δ)·k_on/(k_on + k_off), which in the bursty regime k_off ≫ k_on is the product k_on·(k_syn/k_off)/δ — and cannot separate the two parameters (Ch 47, Ch 48).

10. Measuring it — and why GWAS hits land here

AssayWhat it measuresWhat it cannot tell you
/ CUT&RUN / CUT&TagWhere a protein or histone mark isWhether that occupancy does anything
/ DNase-seqWhich chromatin is accessibleWhich factor is responsible, or which gene is affected
/ Micro-CContact frequency between lociWhether a contact is functional
Reporter assayWhether a fragment can drive expressionWhether it does so in its native context
/ STARR-seqThe same, for 10⁴–10⁶ fragments at onceSame limitation, at scale
CRISPRi/CRISPRa tiling, Perturb-seqWhether the endogenous element is requiredLittle, if the element is redundant

The distinction that matters is sufficiency versus necessity. An MPRA says a sequence can drive out of context; a CRISPRi says the element in place is needed. They disagree often, and both are right about the question they asked. Ch 49 covers the assays properly.

Now the payoff. Roughly 90% of associations are non-coding, and about 76.6% of non-coding trait-associated SNPs lie within, or in perfect with, a DNase-hypersensitive site — i.e. in regulatory DNA.

Be careful about what needs explaining here. That 90% are non-coding is mostly a base-rate fact: only ~1.5% of the genome is protein-coding exon, so the null expectation with no selection at all is ~98.5% non-coding, and at 90% coding variants are over-represented several-fold — 10% observed against a ~1.5% share of the sequence. The selection argument explains something sharper — why the non-coding hits concentrate in accessible regulatory DNA rather than in random non-coding sequence, and why their effect sizes are small. A coding change that materially alters a protein is more often strongly deleterious and gets purged, so it never reaches the common frequencies GWAS is powered to detect. A regulatory change adjusts how much of a normal protein is made, in one tissue, at one stage — a smaller, more survivable perturbation, and therefore one that can to 20% frequency and appear in a study.

Which makes the hits maximally awkward to interpret:

  1. LD means the associated SNP is usually not the causal one — dozens of variants are statistically indistinguishable (Ch 29).
  2. The target gene is often not the nearest gene, and §4 says we cannot predict it from sequence.
  3. The relevant cell type and state may not be one anyone has assayed, and a regulatory element that is inert in every profiled tissue may be decisive in the right one.
  4. The effect size is small — often a few percent change in expression — which is hard to measure and hard to prove causal.

That gap between a p-value and a mechanism is the whole subject of Ch 52.


Common misconceptions

What people thinkWhat's actually true
An switches a gene onIt raises the frequency of transcriptional bursts. Expression is a rate, set by a stochastic process, and enhancers tune one of its two parameters
A motif match is a binding siteA genome-wide scan returns on the order of 1,000 false positives per functional site. Motif presence is weak evidence; occupancy requires accessible chromatin and usually partners
A peak means the factor regulates that geneMost peaks have no measurable effect when perturbed. Occupancy is necessary, not sufficient, and "HOT regions" attract many factors non-specifically
Regulatory elements sit upstream of the gene they controlEnhancers work in either orientation, upstream, downstream, in , and inside neighbouring genes. The SHH limb enhancer is ~1 Mb away inside an intron of LMBR1
The nearest gene is the targetWrong a large fraction of the time, and it is the single most common error in interpreting non-coding association signals
Eukaryotes have like lacThey do not, with the notable exception of nematodes — >17% of C. elegans genes are in operons, resolved by SL2 trans-. Eukaryotic co-regulation works by shared binding-site sequences, not shared transcripts
Activators contact directly, as in bacteriaAlmost never. They recruit coactivators, remodellers and Mediator; Mediator contacts the polymerase
happens in -separated Clustering of Pol II and Mediator is well documented. That it is genuine phase separation, and that the phase behaviour is causal, remains actively contested
Chromatin opens because the gene is being transcribedThe usual causal order runs the other way: and remodellers create , which licenses initiation
Deleting an enhancer should abolish expressionFrequently it does nothing, because shadow/redundant enhancers drive the same pattern. Redundancy is the norm in development

Worked example: how much filtering does it take to find a real enhancer?

Take a specific, answerable question. You want the erythroid bound by GATA1 in the human genome (GRCh38, L = 3.1 × 10⁹ bp haploid). Start from the motif alone and add filters.

Step 1 — information content of the motif. GATA1 binds WGATAA. Per-position frequencies and their contributions, using I = Σ(2 − H_i):

PositionCompositionH (bits)Contribution (2 − H)
1A/T, 50:501.0001.000
2G02.000
3A02.000
4T02.000
5A02.000
6A 80%, G 20%0.7221.278

I = 1.000 + 2 + 2 + 2 + 2 + 1.278 = 10.28 bits.

Check position 6: H = −(0.8 log₂0.8 + 0.2 log₂0.2) = −(0.8 × −0.3219 + 0.2 × −2.3219) = 0.2575 + 0.4644 = 0.7219.

Step 2 — expected matches genome-wide. Both strands, so N = 2L = 6.2 × 10⁹ positions, and p ≈ 2^(−10.28) = 1/1,244.

E[hits] = 6.2 × 10⁹ / 1,244 ≈ 5.0 × 10⁶

Five million matches. GATA1 in erythroid cells returns on the order of 10⁴ peaks. Roughly 1 in 500 motif occurrences is bound. The motif alone is nearly worthless as a predictor.

Step 3 — filter one: combinatorial requirement. Erythroid enhancers characteristically carry GATA1, a TAL1 E-box and a KLF1 site within a few hundred bp. Take all three at ~10 bits and a window w = 300 bp (600 strand-positions).

Expected matches of one motif per window: λ = 600 × 2⁻¹⁰ = 0.586 P(at least one) = 1 − e^(−0.586) = 0.443 P(all three present) = 0.443³ = 0.0872 Non-overlapping windows in the genome: 3.1 × 10⁹ / 300 = 1.03 × 10⁷

E[windows with all three] = 1.03 × 10⁷ × 0.0872 ≈ 9.0 × 10⁵

Requiring three factors instead of one cut the candidate set from 5 × 10⁶ to 9 × 10⁵ — a factor of ~6. Useful, and nowhere near enough. This is the quantitative form of the point in Ch 01: combinatorics is necessary but not sufficient.

Step 4 — filter two: chromatin. In any single cell type, only about 1.5% of the genome falls in an accessible peak. Intersecting:

9.0 × 10⁵ × 0.015 ≈ 1.4 × 10⁴

Now we are at the right order of magnitude for the number of cell-type-specific regulatory elements in an erythroid programme.

Step 5 — read the result. Two filters, applied multiplicatively, took 5 × 10⁶ candidates to ~10⁴:

FilterCandidates remaining
Motif alone5 × 10⁶
+ two partner motifs within 300 bp9 × 10⁵
+ accessible in this cell type1.4 × 10⁴

Chromatin did far more work than combinatorics. That is the structural reason default-off is the eukaryotic strategy: making 98.5% of the genome physically unavailable is the single largest specificity filter available, and it is the one a bacterium cannot use.

Two honest caveats. The filters are not independent — accessible regions are motif-enriched, so the true intersection is larger than the product. And ~10⁴ accessible, motif-matched elements is still far more than the number whose deletion measurably changes expression; CRISPRi tiling typically finds a small minority to be required. Getting from 10⁴ candidates to the functional subset is precisely the frontier, which is why enhancer prediction from sequence remains unsolved.

Connections