Contents
38 — Genome editing Part 8 — Methods

38 — Genome editing

Before this: Ch 17 · Ch 25A §§6–8 — gene targeting, which is the thing this chapter replaced; §2's argument is hard to feel without it · Ch 36 · Ch 37 · Time: ~45 min

What you'll be able to do

The core idea

Every genome editor is two things bolted together:

  1. A programmable DNA-binding module that finds one address in 3.1 Gb.
  2. A payload that does something there — usually break the DNA, sometimes chemically modify a base, sometimes write new sequence.

And then there is a third component you do not control: the cell's own repair machinery, which decides what the final sequence actually is. You aim; the cell edits. Nearly every hard problem in this field is a consequence of that split.

The thirty-year history collapses into two sentences. First, targeting moved from protein engineering to base pairing — from recompiling a binding module for every new target to passing a 20-nucleotide parameter. Second, the field has been steadily taking the decision away from the repair machinery, from "cut and hope" through to , where the new sequence is written by a polymerase you brought with you.


1. The repair pathways are the editing machinery

A double-strand break is repaired by one of two competing routes (Ch 17):

graph TD DSB["double-strand break<br/>at your target"] --> NHEJ["non-homologous end joining<br/><i>available all cell-cycle phases</i>"] DSB --> HDR["homology-directed repair<br/><i>needs a template + S/G2</i>"] NHEJ --> PERF["religated perfectly<br/>→ substrate for re-cutting"] NHEJ --> INDEL["small indel<br/>→ frameshift → <b>knockout</b>"] HDR --> PRECISE["template copied in<br/>→ <b>precise edit</b>"]

Perfect religation regenerates the target, so cuts again; the reaction runs until an destroys the site. Indels are an absorbing state. That is why efficiencies are high without you doing anything clever.

-directed repair needs a donor template with homology arms, a sister chromatid available as the normal substrate, and the S/G2 of the cell cycle. Typical HDR efficiency in dividing cells is single-digit to low-double-digit percent; in post-mitotic or quiescent cells — neurons, cardiomyocytes, the largely non-dividing hepatocytes of an adult liver, the tissues you most want to fix — it collapses. Neurons and cardiomyocytes are terminally post-mitotic and HDR there is effectively zero; adult hepatocytes are merely G0-quiescent — they retain full proliferative capacity, which is why the liver regenerates after partial hepatectomy — so HDR is very inefficient rather than absent, and works appreciably only in neonatal or regenerating liver.

This is the central practical of the field. Breaking a gene is easy everywhere. Correcting a point is hard in dividing cells and impossible in most of the cells that matter. and exist because of this one fact.

Repair outcomes are not noise. The indel spectrum at a given cut site is reproducible across replicates and largely predictable from the ±20 bp of local sequence, because microhomology-mediated end joining favours particular deletions. Outcome prediction is a supervised learning problem over sequence context, and the models work well enough that guide choice is partly a choice of outcome distribution.

2. From recompiling to passing a parameter

The row above this table is gene targeting, and it is worth having in view because it is the baseline every entry below is an improvement on: no nuclease at all, at its natural frequency, and the whole difficulty pushed into selecting the rare correct event — positive/negative selection, ES cells, chimeras, transmission, and roughly a year to a founder animal. Ch 25A §§6–8 builds it, and §8 there marks the boundary precisely: editing replaced the addressing step and left the repair biology, the design and the breeding untouched.

PlatformHow targeting is specifiedRecognitionCost per new target
MeganucleasesRedesign the enzyme's own DNA-contacting residues~18–24 bpMonths of protein engineering; few targets reachable
Zinc-finger nucleasesAssemble fingers, ~3 bp each, fused to the FokI nuclease domain, which must dimerise — so targets come in pairs~18 bp per pairWeeks; context-dependent finger interactions make it unreliable
TALENsOne TALE repeat per base, base identity set by two residues (the RVD) — a genuine one-to-one code~30–40 bp per pairDays, but each construct is a long, highly repetitive gene to synthesise and clone
–20-nt guide RNA, Watson–Crick pairing to the target20 bp + PAMHours; an oligonucleotide order

TALENs already had a code — the field knew how to specify any sequence. The revolution was not specificity, it was binding energy supplied by base pairing instead of by a folded protein surface. A protein interface has to be designed, expressed and validated. A guide RNA is a string literal.

The asymptotics matter more than the constant factor. Editing N targets costs N protein-engineering campaigns with ZFNs or TALENs, and one plasmid plus N oligos with CRISPR. Pooled (§9) are simply not constructible on the older platforms.

3. Cas9 mechanism

is a bacterial adaptive immune system: fragments of past invaders are stored as spacers in a genomic array, transcribed, and used to guide a nuclease against matching sequence. Two RNAs are involved naturally — a crRNA carrying the spacer and a tracrRNA that scaffolds the protein — and the engineering step was fusing them into one single-guide RNA (sgRNA).

The PAM is the self/non-self test. Streptococcus pyogenes (SpCas9) will not cut unless the target is immediately followed by NGG. The bacterium's own CRISPR array contains the spacer sequence but not the adjacent PAM, so Cas9 cannot cut the array that encodes it. Without the PAM, a CRISPR system is an autoimmune disease. For us, the PAM is a on the addressable space: NGG occurs about once every 8 bp on average across both strands, so most positions are reachable but not all.

Recognition is ordered and directional. Cas9 collides with DNA constantly and interrogates only PAMs; on finding one it locally melts the duplex and tries to zipper the guide onto the target strand starting from the PAM-proximal end. Mismatches near the PAM (the seed, roughly the PAM-proximal 8–12 nt) abort the process; mismatches at the far end often do not. A complete R-loop — guide RNA paired to the target strand, displaced non-target strand looped out — licenses the two nuclease domains, HNH and RuvC, to cut one strand each, producing a blunt cut 3 bp upstream of the PAM.

                               protospacer (20 nt)            PAM
                     |-------------------------------------| |---|
   5'- C C T G G T A G A C T T C A G C A G T A C C G G A C A T G G C C T A A G -3'
   3'- G G A C C A T C T G A A G T C G T C A T G G C C T G T A C C G G A T T C -5'
                                                      ^
                                                  blunt cut
                                       (between protospacer 17 and 18)

   sgRNA spacer:  5'- G A C U U C A G C A G U A C C G G A C A -3'

One kinetic detail with practical consequences: Cas9 is effectively single-turnover and remains clamped to the cut product for a long time. Cutting is fast; release is slow; total off-target damage therefore scales with how long the enzyme is present, not with how fast it works. That argument reappears in §10.

4. Guides, off-targets, and assays that disagree

On-target efficiency varies over an order of magnitude between guides at the same locus, driven by guide sequence composition, secondary structure, occupancy and chromatin (Ch 23). Scoring models are trained on measured cleavage across thousands of guides — and carry the usual transfer problem, since the training data come from particular cell types and delivery formats.

Off-target activity is where naive intuition fails. Hamming distance to the genome is the wrong model: a 3-mismatch site in the seed is usually dead, a 3-mismatch site at the PAM-distal end can be cut efficiently, and RNA or DNA bulges create sites that a fixed-length string search never enumerates. Exhaustive in-silico enumeration has excellent recall and terrible precision.

Empirical assays split by where the cut is measured:

Assay ideaImplementationSeesMisses
Capture breaks in living cells with a tagGUIDE-seq — a blunt double-stranded oligo is integrated at breaks by , then its junctions are sequencedReal cellular chromatin contextSites in cells that take up the tag poorly; low-frequency events
Census of cuttable sites in vitroCIRCLE-seq, CHANGE-seq — circularised genomic DNA is cut by the ribonucleoprotein and linearised fragments sequencedVery high sensitivity, no cell neededNo chromatin — nominates sites never cut in vivo
Pull down a repair protein at breaksDISCOVER-seq — ChIP for MRE11, which loads at breaksWorks in tissues and animalsRequires the repair protein to accumulate detectably

These methods do not agree. Nominated-site lists overlap only partially, and the union is usually much larger than the intersection. There is no gold standard, so the defensible posture is: run orthogonal assays, take the union as candidates, then deep-sequence those candidates in the actual therapeutic cell type, and report the detection floor (typically ~0.1% ) rather than claiming zero.

A second class of risk is invisible to standard assays. A single cut can produce kilobase-scale deletions, loss of heterozygosity, or ; two simultaneous cuts can produce (Ch 20). Amplicon sequencing across the cut site cannot see these, because the events destroy the primer binding sites — the assay conditions on the damage not having happened. That is an , not a measurement error, and it is why long-read or -level assessment belongs in any therapeutic package.

5. The variant zoo, organised by what problem it solves

VariantIdea it implements
High-fidelity SpCas9 (engineered contact mutants)Remove excess non-specific binding energy so that a mismatched R-loop no longer clears the activation threshold. Specificity by making the enzyme worse at binding
Compact orthologues (e.g. SaCas9)Fit inside a viral vector's cargo limit (§10)
Relaxed-PAM variantsWiden the addressable space when the edit must land at a specific base, e.g. for
Cas12a (Cpf1)T-rich PAM (TTTV) opens AT-rich sequence; cuts ~18–23 bp distal to the PAM leaving staggered ends, which favours directional insertion; needs only a crRNA, and processes its own crRNA array — so one transcript can deliver many guides, making multiplexing cheap
Cas13Targets RNA, not DNA. Knockdown without touching the genome, and the basis of amplification-free nucleic-acid diagnostics via its collateral cleavage of reporter RNAs

6. dCas9: targeting as an API

Inactivate both nuclease domains (D10A + H840A) and becomes dCas9 — a programmable DNA-binding protein with a payload slot. This decouples which locus from what to do there.

FusionEffectWhy it is used
Transcriptional domain (CRISPRi)Blocks initiation/elongationKnockdown without a cut; dose-tunable; works on non-coding elements where frameshifts are meaningless
Activator domains (CRISPRa)Recruits the machinery — tests sufficiency, not necessity
Histone acetyltransferase / DNA methyltransferase / demethylaseWrites or erases a chromatin markTests whether a mark is causal, and whether it self-sustains once the editor leaves (Ch 23)
Fluorescent proteinLabels a locusLive imaging of dynamics (Ch 50)

7. Base editing: chemistry instead of breakage

A base editor is a deaminase fused to a nickase. The R-loop exposes a short single-stranded stretch of the non-target strand; the deaminase chemically converts a base within a window of roughly 4–8 nt (counted from the PAM-distal end of the protospacer). No double-strand break, no donor template, no dependence on HDR — so it works in non-dividing cells.

Cytosine base editor (CBE)Adenine base editor (ABE)
EnzymeCytidine deaminase + uracil glycosylase inhibitorA tRNA adenosine deaminase evolved in the laboratory to accept DNA — no natural DNA adenine deaminase exists
ChemistryC → U, read as TA → inosine, read as G
Net changeC•G → T•AA•T → G•C

The nickase nicks the unedited strand, which biases repair to use the edited strand as template — the same "nick the strand you want overwritten" trick reuses.

The limitation is structural: the conventional deaminase editors above install transitions only. ↔purine and ↔pyrimidine. Transversions need prime editing, a nuclease, or the newer glycosylase-coupled editors — CGBEs (a CBE with the uracil glycosylase added rather than inhibited, so the U is excised to an abasic site and installs C•G→G•C) and the A-to-Y family such as AYBE (an ABE fused to a methylpurine glycosylase). These fit the same "deaminase on a nickase" definition and are still cut-free, but they are lower-efficiency and narrower-window than CBEs and ABEs, so the practical statement is transitions are routine, transversions are not. This matters less than it sounds, because spontaneous deamination of methylated cytosine makes G•C → A•T by far the most common class of pathogenic point (Ch 16) — so ABEs address a large share of correctable variants. Other real limitations: bystander editing of every susceptible base in the window; sequence-context preference; and guide-independent deamination of any transiently single-stranded DNA or RNA the editor encounters, which is invisible to guide-based off-target assays entirely.

8. Prime editing: bring your own polymerase

fuses a nickase to an engineered reverse transcriptase and replaces the sgRNA with a pegRNA carrying two extra 3′ elements: a primer binding site and a reverse-transcriptase template containing the desired edit.

1. nick the PAM-containing strand
2. the freed 3' end anneals to the PBS on the pegRNA
3. RT extends that end, copying the RTT  →  a 3' flap carrying the edit
4. edited 3' flap competes with the original 5' flap; 5' flap excised, ends ligated
5. heteroduplex →  mismatch repair decides which strand wins

Step 5 is the efficiency bottleneck, and it is attacked directly: nicking the non-edited strand biases toward the edit (the PE3 design), and transiently suppressing mismatch repair with a MLH1 raises efficiency further (PE4/PE5).

Nuclease + HDRPrime editing
Double-strand breakyesnono
Donor template neededyesnono (template is in the guide)
Substitutions possibleall 124 routinely (transitions); transversions only via glycosylase editorsall 12
Insertions / deletionsyes, template-drivennoyes, up to ~tens of bp
Works in non-dividing cellspoorlyyesyes
Main costHDR efficiencywindow/bystander large construct, lower efficiency, more design parameters

Prime editing's design space (PBS length, RT template length, nick placement) is empirical and larger than a base editor's, and its cargo is bulky enough to make delivery harder. In exchange it installs any substitution — transversions included — or a small insertion without a cut, at a generality the glycosylase editors do not reach.

9. Pooled screens: genetics as a counting experiment

This is where a programmable, cheap-to-multiplex nuclease changes what is askable (Ch 37).

graph LR LIB["guide library<br/>4–10 guides × 20k genes"] --> LV["lentiviral transduction<br/>MOI &lt; 1 → one guide per cell"] LV --> T0["T0 sample"] LV --> SEL["selection<br/>proliferation · drug · FACS sort"] T0 --> SEQ["PCR the integrated guide cassette<br/>and sequence"] SEL --> SEQ SEQ --> STAT["count-based differential abundance<br/>→ guide effects → gene effects"]

The readout is a count matrix, and the statistics are the ones you already know from (Ch 47): negative-binomial models with shared estimation, then aggregation of the several guides targeting a gene into a gene-level statistic (rank aggregation or a mixed model), with the guide-level replication doing the work that biological replicates do elsewhere. Power scales with cells per guide — of 100–1000× is the analogue of read depth.

Two artefacts to know:

Modalities map onto different questions: asks is this gene necessary; CRISPRi asks the same for essential genes and non-coding elements without a cut; CRISPRa asks is it sufficient. Tiling place a guide at every available PAM across hundreds of kilobases and read out a reporter or marker — turning discovery into a positional signal detection problem (Ch 52). Single-cell readouts replace the scalar with a full transcriptome per perturbed cell (Ch 48).

10. Delivery is the bottleneck

The editing chemistry is largely solved. Getting it into the right cells, at the right dose, for the right duration, is not.

ModalityCargo limitDuration of expressionNotes
Ribonucleoprotein, electroporatednonehoursBest specificity — exposure time is minimal. Ex vivo only
Lipid nanoparticle, mRNA cargolargedaysRedosable, no genomic integration; tropism strongly hepatic after IV dosing
AAV~4.7 kb totalmonths to years (episomal)SpCas9 coding sequence alone is ~4.1 kb, so SpCas9 + guide + does not comfortably fit — hence compact orthologues and split-vector designs. Long expression means long off-target exposure; pre-existing antibodies exclude many patients

Because off-target damage integrates over exposure time, the ranking RNP > mRNA > integrating vector is a specificity ranking, not just a convenience one.

Ex vivo editing removes cells, edits them as RNP, expands and quality-controls them, and reinfuses — you get to inspect your work before dosing. It is limited to cell types you can harvest and transplant, and it requires conditioning to make marrow space, which is where most of the toxicity actually lives. In vivo editing has no such limits and no such checkpoint; the field's early in vivo successes are all hepatic because that is where lipid nanoparticles go.

11. Where therapy actually stands

The approved case. Exagamglogene autotemcel (Casgevy) treats sickle cell disease and transfusion-dependent β-thalassaemia by editing the patient's own haematopoietic stem cells ex vivo. Approvals: UK November 2023, US December 2023 (sickle cell) and January 2024 (β-thalassaemia), EU February 2024, Canada September 2024; the US label was extended to age 2 and above in July 2026. List price is roughly $2.2M in the US.

Its design is the best single illustration of §1. Repairing the causal HBB would require HDR in stem cells — the hard problem. Instead the therapy makes a : it disrupts the erythroid-specific in 2 of BCL11A, the that silences fetal haemoglobin after birth. Losing that enhancer de-represses HBG1/HBG2 in red-cell precursors only, and fetal haemoglobin substitutes for the defective adult protein. A precise-repair problem was re-specified as a break-something problem, because breaking things is what the repair machinery does well.

In vivo and beyond nucleases. Lipid-nanoparticle-delivered editors targeting liver-expressed genes — TTR in transthyretin amyloidosis, PCSK9 for cholesterol — are in clinical trials, as are base-edited cell therapies. The prime editor PM359 entered trials in 2024 for chronic granulomatous disease, with two treated patients reported to be effectively disease-free by December 2025.

The bespoke case. An infant with CPS1 deficiency received a personalised adenine base editor designed for his private variant, delivered as mRNA in lipid nanoparticles across three intravenous doses beginning February 2025; the therapy went from diagnosis to dosing in about six months and was reported in the New England Journal of Medicine in June 2025. The scientific content is unremarkable by 2026 standards. The structural content is not: it demonstrates an n-of-1 regulatory and manufacturing pathway for variants too rare to support a trial (Ch 54).

12. Germline editing

Everything above is somatic: the edit affects the treated person and dies with them. Editing an embryo or gamete makes the change heritable. The distinction is legal and ethical, not technical — the same reagents work; only the target cell differs.

In 2018 He Jiankui edited CCR5 in human embryos to attempt HIV resistance, resulting in the birth of twin girls and later a third child. At the sequence level the experiment failed on its own terms: neither twin carried the intended naturally occurring Δ32 ; one was heterozygous for a novel edit, the other carried novel ; both showed . He was convicted of illegal medical practice in December 2019, sentenced to three years and fined ¥3M, and released in April 2022. China subsequently amended its Civil Code and Criminal Law to prohibit non-compliant human genome editing.

The scientific objections stand independently of the ethics:

ObjectionWhy it is not fixable by better technique
No unmet needSperm washing already prevents paternal HIV transmission. The risk/benefit ratio had no numerator
MosaicismThe editor acts over a window during which the zygote divides. You biopsy cells you discard and infer about cells you keep — the verification problem is intrinsic
The edit was not the tested alleleΔ32 has a century of human population data behind it. A novel indel in the same gene has none
Known costsCCR5 loss carries elevated susceptibility to some flaviviruses. There is no such thing as a free
Irreversibility and consentThe subjects cannot consent, their descendants cannot consent, and there is no mechanism for withdrawal

Governance as of August 2026. Heritable human genome editing is prohibited by law in more than 70 countries and by the Oviedo Convention, a binding Council of Europe treaty. The 2020 International Commission convened by the US National Academies and the UK Royal Society concluded that no clinical use should proceed and specified narrow preconditions and a translational pathway if it ever did; the WHO issued a governance framework in 2021 recommending against heritable editing while building international registry and oversight mechanisms; the Third International Summit on Human Genome Editing (London, March 2023) concluded that heritable editing remains unacceptable at this time. Chapter 58 takes the argument apart properly (Ch 58).

13. Gene drives

A homing is a cassette encoding plus a guide targeting the wild-type at the drive's own locus. In a heterozygote, the drive cuts the wild-type ; HDR copies the drive across using the drive chromosome as template. The heterozygote becomes a homozygote, and transmission rises from 50% toward 100%.

This breaks the arithmetic that all of Part 5 rests on. A costly allele cannot normally invade from low frequency (Ch 27); a drive can, because super-Mendelian transmission supplies a frequency-independent boost that outruns selection against it. That is precisely the appeal — suppressing a malaria vector — and precisely the danger.

Its failure mode comes from §1. When the cut is repaired by instead of HDR, the resulting destroys the guide's target site: a drive-resistant allele, immune to conversion and positively selected. The same pathway competition that limits HDR in a patient's cells limits drive spread in a wild population. Mitigations: multiplex guides so several targets must be destroyed simultaneously, and target ultraconserved sequence such as the female-determining exon of doublesex, where indels are themselves lethal.

Containment designs trade reach for reversibility: split drives (Cas9 unlinked from the guide, so the drive cannot spread alone), daisy-chain drives (a chain of elements that exhausts after a bounded number of generations), and reversal drives intended to overwrite a released drive. None has been demonstrated at ecological scale.

No gene-drive organism has been released into the wild as of August 2026. Field programmes have proceeded through staged releases of non-drive modified mosquitoes, and the Convention on Biological Diversity requires case-by-case risk assessment with the engagement and consent of affected communities. The governance problem is genuinely novel: a drive has no geographic containment, can move between hybridising species, and is not recallable.


Common misconceptions

What people thinkWhat's actually true
"edits" DNACRISPR cuts DNA. The cell writes the edit. Base and prime editors are the exceptions that write it themselves
A precise correction is the standard usePrecise correction requires HDR, which is inefficient in dividing cells and effectively absent in non-dividing ones. Most successful therapies are engineered as
Off-target effects are measured, and the number is knownAssays disagree substantially and none is a gold standard. Reports state a detection floor, not zero
Base editors can make any substitutionThe standard deaminase editors do transitions only — C•G→T•A and A•T→G•C. Transversions require , a nuclease, or the newer glycosylase-coupled editors (CGBE, AYBE), which are real but lower-efficiency and narrower-window
Amplicon sequencing at the cut site shows what happenedIt is blind to large deletions, and , because those destroy the primer sites
and somatic editing differ technicallyIdentical reagents. The difference is which cell you put them in, and whether the change is inherited
spread because they are beneficialThey spread because they copy themselves into the homologous . cost slows them; it does not stop invasion
A guide with no perfect off-target match is safeMismatch tolerance is position-dependent, not Hamming-distance-dependent. A 3-mismatch PAM-distal site can be cut efficiently

Worked example

Design a , then compute what you should expect.

Step 1 — find a cut site. Using the sequence in §3 (a synthetic example), scan the top strand for NGG. The trinucleotide TGG at positions 28–30 qualifies. The 20 nt immediately 5′ of it, positions 8–27, are the protospacer:

protospacer  5'- GACTTCAGCAGTACCGGACA -3'      PAM = TGG
position       1                   20

Step 2 — locate the cut. Blunt, 3 bp upstream of the PAM: between protospacer positions 17 and 18, i.e. ...GTACCGG ▲ ACA TGG. The sgRNA spacer is the protospacer as RNA: GACUUCAGCAGUACCGGACA.

Step 3 — expected knockout fraction. Suppose amplicon sequencing of the edited population gives a 62% rate across reads, and that 68% of indel have length not divisible by 3 (a typical spectrum, dominated by 1 bp insertions and short deletions).

So a population with a headline "62% editing efficiency" contains fewer than one in five true knockout cells. That is why single-cell cloning or a selectable marker is standard.

The independence assumption is also wrong, and wrong in a known direction: cells differ in how much editor they received, so allele outcomes within a cell are positively correlated. The true biallelic fraction exceeds the binomial square — measured allele frequencies are marginal over an overdispersed cell-level distribution.

Step 4 — how many off-target candidates to expect. The number of 20-mers at Hamming distance exactly d from the guide is C(20, d)·3^d:

d = 0 :      1
d = 1 :     20 ×  3 =     60
d = 2 :    190 ×  9 =  1,710
d = 3 :  1,140 × 27 = 30,780
                       ------
total ≤ 3            = 32,551

The probability that a random 20-mer falls in that set is 32,551 / 4²⁰ = 32,551 / 1.0995 × 10¹² = 2.96 × 10⁻⁸.

Sites carrying an NGG PAM: a 3.1 Gb genome read on both strands gives 6.2 × 10⁹ positions, of which a fraction (1/4)(1/4) = 1/16 are followed by NGG — about 3.88 × 10⁸ candidate targets.

Expected sites within 3 mismatches = 3.88 × 10⁸ × 2.96 × 10⁻⁸ ≈ 11.5.

So roughly a dozen close homologues by chance alone, in a genome that is not uniform random — repeats and paralogues make the real count worse. Almost none of these will actually be cut, because most of the mismatches will fall in the seed. The calculation bounds how many candidates you must test; it does not estimate how many are real. That gap is exactly why the empirical assays in §4 exist, and why they are run in the target cell type.

Connections

Back to:

Forward to: