How CRISPR Actually Edits DNA
The real mechanism behind CRISPR gene editing — guide RNA, PAM sequences, double-strand breaks, and the repair pathways that decide the outcome.
By Velloreth
CRISPR is usually described as “molecular scissors” or a “word processor for DNA.” Both metaphors hide the thing that actually matters: CRISPR-Cas9 does not rewrite DNA. It makes a cut at a location you choose, and then the cell’s own repair machinery — imperfect, probabilistic, and different in different cell types — decides what the sequence looks like afterward. Understanding that handoff is the difference between a working edit and a failed experiment.
This is the mechanism as it actually works, from bacterial immunity to clinical therapy.
Where the system comes from
CRISPR (Clustered Regularly Interspaced Short Palindromic Repeats) is a bacterial immune system. When a virus infects a bacterium, the bacterium sometimes captures a snippet of the viral DNA and stores it in its own genome, between CRISPR repeat sequences. If the same virus attacks again, the bacterium transcribes those snippets into RNA, loads them into a Cas protein (CRISPR-associated), and the complex searches for and destroys matching viral DNA.
The 2012 work by Jinek, Doudna, Charpentier and colleagues showed that the system could be reprogrammed: the bacterial CRISPR RNA and its helper (tracrRNA) could be fused into a single guide RNA (sgRNA) of about 20 nucleotides. Change those 20 letters, and you retarget the cut to almost any sequence in any organism. That programmability — one protein, one RNA, one target — is why CRISPR displaced zinc-finger nucleases and TALENs almost overnight.
The targeting mechanism
Cas9 from Streptococcus pyogenes (SpCas9) is the workhorse. The complex of Cas9 plus sgRNA scans DNA for a protospacer adjacent motif (PAM) — for SpCas9, the three-letter sequence NGG (N is any base, G is guanine). PAMs occur roughly once every 16 base pairs in a random genome, so target sites are abundant, but not every position is editable.
Once Cas9 finds a PAM, it unwinds the adjacent DNA and the sgRNA attempts to base-pair with the target strand. If pairing is good enough across the ~20-nucleotide spacer, two nuclease domains — HNH and RuvC — cut each DNA strand. HNH cuts the strand paired with the guide; RuvC cuts the other. The result is a double-strand break (DSB) about three nucleotides upstream of the PAM.
Two constraints follow directly. First, you cannot cut just anywhere: you need a PAM nearby, and the target must match the spacer. Second, mismatches near the PAM-proximal end (the “seed” region) are less tolerated than mismatches at the distal end — which matters for predicting off-target cuts.
What happens after the cut: the cell decides
Cas9 has done its job. Everything after this point is the cell repairing itself, and the repair pathway chosen determines the edit.
NHEJ: the knockout pathway
Non-homologous end joining (NHEJ) ligates broken ends back together, and it is error-prone. At the junction it frequently inserts or deletes a few nucleotides (indels). If the cut lands in a coding exon, a frameshift mutation often knocks the gene out of function. This is how most CRISPR “knockouts” work: you are not deleting the gene, you are letting the cell introduce a frameshift when it stitches the ends back together.
NHEJ works in dividing and non-dividing cells, is efficient, and needs no donor template. For gene disruption — knocking out a viral receptor, disabling a disease allele, creating a loss-of-function cell line — it is the standard approach.
HDR: the precise pathway
Homology-directed repair (HDR) uses a template with homology to the sequences flanking the break. Supply an exogenous donor — typically a single-stranded oligodeoxynucleotide (ssODN), which is more efficient than double-stranded donors — and the cell can copy your sequence across the break site. That gives you a specific nucleotide change, a small insertion, or a correction of a pathogenic variant.
The limitations are real. HDR is much less efficient than NHEJ: the two pathways compete for the same break, and NHEJ usually wins. HDR occurs primarily in the S and G2 phases of the cell cycle, so it largely fails in post-mitotic cells like neurons and cardiomyocytes — exactly the cells many therapies target. Cells also have evolved to suppress HDR in somatic tissue; embryonic stem cells use it far more readily.
Researchers tilt the balance with small molecules that inhibit NHEJ factors (the “CRISPY mix” of DNA-PK inhibitors and related compounds), cell-cycle synchronization, and careful donor design. But HDR remains the bottleneck for precise correction.
Other pathways
Two more repair routes matter in practice. Microhomology-mediated end joining (MMEJ) uses short microhomologies flanking the break and produces predictable deletions — useful for certain engineering applications, a nuisance for precision editing. Single-strand annealing (SSA) deletes everything between direct repeats, which is destructive and usually unwanted.
From cuts to precision: base editors and prime editing
Double-strand breaks are powerful but blunt. They cause indels, chromosomal rearrangements in rare cases, and p53-mediated DNA damage responses. Two generations of tools avoid them entirely.
Base editors
Developed in David Liu’s lab in 2016, base editors fuse a catalytically impaired Cas9 (a nickase that cuts only one strand) to a deaminase enzyme. The guide positions the editor over a target window — typically positions 4–8 of the protospacer — and the deaminase chemically converts one base to another: cytidine base editors do C→T (and G→A on the opposite strand), adenine base editors do A→G.
No donor template, no double-strand break, and the edits work in non-dividing cells. Point mutations account for more than half of known pathogenic genetic variants in humans, and C→T and A→G conversions cover a large fraction of those. The trade-offs: only certain conversions are possible, and bystander edits within the editing window can occur if nearby bases are also substrate for the deaminase.
Prime editing
Prime editing (Liu lab, 2019) is more general. A Cas9 nickase is fused to a reverse transcriptase, and the guide — a prime editing guide RNA (pegRNA) — carries both a targeting sequence and a template encoding the desired edit. The nickase cuts one strand; the exposed 3′ end hybridizes to the pegRNA’s primer-binding site; the reverse transcriptase copies the edit template directly into the genome.
The result is a flap that the cell resolves — if the edit is kept, the change is permanent, with no double-strand break and no donor template required. Prime editing can install any of the twelve possible base-to-base conversions, plus small insertions and deletions. PE3 improves efficiency further by nicking the non-edited strand to bias repair toward the new sequence. The trade-off: prime editing is currently less efficient than base editing and requires more complex guide design.
CRISPR in the clinic: what has actually shipped
The first approved CRISPR therapy is Casgevy (exagamglogene autotemcel), authorized in the UK in November 2023, by the FDA in December 2023, and by the EMA in February 2024. It treats sickle cell disease and transfusion-dependent β-thalassemia.
The mechanism is not correcting the patient’s mutation in their body. Instead: hematopoietic stem cells are harvested, edited ex vivo with CRISPR-Cas9 to disrupt the BCL11A gene — specifically its erythroid enhancer — which de-represses fetal hemoglobin (HbF) production. The edited cells are reinfused after myeloablative conditioning. Fetal hemoglobin compensates for the defective adult hemoglobin.
The trial results were strong for a single-arm study: among sickle cell patients followed long enough, 28 of 29 were free of vaso-occlusive crises for at least a year; in β-thalassemia, 39 of 42 patients became transfusion-free. The side effects are those of the conditioning regimen — myeloablation, febrile neutropenia, mucositis — not the edit itself, though long-term follow-up continues.
Other CRISPR therapies are in trials for transthyretin amyloidosis (in vivo editing in the liver), hereditary angioedema, and various cancers (editing immune cells rather than the patient’s genome directly). Each faces the same delivery problem: getting the editing machinery into the right cells, at sufficient efficiency, without editing the wrong ones.
Off-target effects and the safety picture
Any programmable nuclease can cut at sites that resemble the target. Off-target DSBs can create indels or, in combination with on-target breaks, translocations. The field’s response has been threefold:
- High-fidelity Cas9 variants (eSpCas9, HypaCas9, evoCas9) that reject mismatched targets, trading some on-target efficiency for specificity.
- Unbiased detection methods — GUIDE-seq, CIRCLE-seq, DISCOVER-seq — that map off-target cuts genome-wide rather than testing a predicted list.
- Break-free editing — base and prime editing — which reduces but does not eliminate off-target issues (deaminases can act at off-target sites too; prime editing’s nickase approach cuts far less aggressively).
Whole-genome sequencing of prime-edited cells has generally shown minimal genome-wide changes beyond intended edits and a small number of indels near the additional nick site in PE3. That is reassuring, but the assays have detection limits, and long-term safety data in humans is still accumulating for everything beyond Casgevy.
Designing an edit: the practical sequence
If you were planning an experiment today, the decision tree looks like this:
- Knockout? Use Cas9 + sgRNA, let NHEJ do the work. Choose a guide in an early coding exon with a good PAM and a clean off-target profile. Verify by amplicon sequencing, not just phenotype.
- Point mutation correction? If the change is a C→T or A→G class, base editing first — it is more efficient and works in non-dividing cells. Otherwise, prime editing.
- Small insertion or precise sequence? Prime editing for anything up to tens of base pairs; HDR with ssODN when efficiency is acceptable and cells are dividing.
- Large knock-in? HDR is still the main tool, with viral or plasmid donors; newer approaches (twinPE, PASTE) are extending the size range but remain mostly experimental.
At every step: check the PAM, check the seed region matches, run off-target prediction, and measure the edit directly. The guide RNA is easy to design; the biology of repair is what determines whether it works.
The through-line
CRISPR’s power is that it moved gene editing from a specialist craft (custom zinc-finger proteins per target) to something a graduate student can do in a week with a catalog-synthesized RNA. Its limits are equally clear: the cell, not the enzyme, performs the edit, and the cell’s priorities are its own. Every advance in the field — base editors, prime editing, better delivery, better off-target assays — is a response to that fact. The scissors were the easy part.