Architecture, specificity, and choosing an editor

Assumes you know what a nucleotide, a protein domain and a reading frame are. Covers how SpCas9 actually finds a target, why specificity is a kinetic problem rather than a sequence-matching one, and what distinguishes the four editor families in practice.

Domain architecture

Streptococcus pyogenes Cas9 is a bilobed protein of roughly 1,370 amino acids. The two lobes do different jobs, and almost everything interesting about specificity comes from how they talk to each other.

The recognition (REC) lobe holds the guide RNA and reads the growing RNA–DNA duplex. It is subdivided into REC1, REC2 and REC3. The nuclease (NUC) lobe contains the two catalytic domains plus the PAM-interacting domain: RuvC (an RNase H fold, split across three segments of the primary sequence) cleaves the non-target strand, HNH cleaves the target strand, and the PI domain reads the PAM.

The single-guide RNA fuses two natural RNAs. The crRNA supplies the 20-nucleotide spacer that pairs with DNA; the tracrRNA supplies the scaffold that Cas9 grips. Charpentier's discovery of tracrRNA is what made the two-component fusion possible.

Domain organisation of SpCas9. RuvC is discontinuous in sequence but assembles into one active site. HNH sits on a flexible linker, which is what allows the conformational gating described below.
REC lobe
RuvC
HNH
PAM-interacting

Target search and R-loop formation

Cas9 does not scan processively along DNA. It samples by three-dimensional collision, and what it tests at each collision is the PAM, not the guide sequence. For SpCas9 the PAM is 5'-NGG-3', read in the major groove of the duplex by two arginines in the PI domain.

PAM engagement is the licence to proceed. Only after binding a PAM does Cas9 locally distort the adjacent DNA, bending and undertwisting it so that bases flip out of the duplex and become available to the guide. A sequence perfectly complementary to the guide but lacking an adjacent PAM is ignored entirely.

Hybridisation then proceeds directionally, starting at the PAM-proximal end and extending outward, displacing the non-target strand as it goes. The resulting three-stranded structure (RNA–DNA hybrid plus displaced ssDNA) is the R-loop.

The first 8 to 12 nucleotides adjacent to the PAM are the seed. Because pairing starts there, a mismatch in the seed aborts the process almost immediately: the complex dissociates within a fraction of a second. Mismatches at the PAM-distal end are a different story, and that difference is the whole basis of off-target behaviour.

Specificity is a conformational problem

The naive model, that Cas9 cuts wherever the guide matches, is wrong in an important way. Binding and cleavage are separate events with separate requirements.

Completing the R-loop drives a rearrangement of REC2 and REC3 that repositions HNH. Crucially, HNH first adopts a catalytically incompetent intermediate, a checkpoint conformation. Only a fully or near-fully complementary heteroduplex supplies the energy to push HNH into its docked, active state. With PAM-distal mismatches, Cas9 can bind stably and sit there without ever cutting.

This is why binding assays and cleavage assays disagree, and why an off-target site identified by a binding method is not automatically a cut site. It also explains a practical trick: truncating the spacer to 17 nucleotides raises specificity, because it removes the energetic slack that would otherwise let a single PAM-distal mismatch still trip HNH activation.

Once HNH docks, both strands are cut, producing a predominantly blunt double-strand break three base pairs from the PAM. The docked conformation persists after scission, and Cas9 remains bound to the product for a long time, which matters for turnover.

Binding and cutting are gated separately. A near-match can reach stable binding and stall at the checkpoint without cleavage.

What the break becomes

Cas9 is a nuclease, not an editor. The edit is whatever the cell's repair machinery produces.

Non-homologous end joining

Dominant in most cell types and active throughout the cell cycle. Ends are ligated directly, frequently with small insertions or deletions at the junction. In coding sequence, roughly two thirds of indels shift the reading frame and cause loss of function through premature termination or nonsense-mediated decay. Indel spectra are not random: they are reproducible per target site and predictable from sequence context, which is why knockout screens work at all.

Microhomology-mediated end joining

An alternative end-joining route that uses short homologous tracts flanking the break, producing characteristic deletions. Worth knowing because it explains much of the reproducibility above.

Homology-directed repair

Requires a donor template and is largely restricted to S and G2 phase, since it depends on the recombination machinery. In non-dividing cells such as neurons, hepatocytes at rest, or quiescent stem cells, HDR is effectively unavailable. This single constraint is why base and prime editing were developed.

Everything else

A double-strand break is a genuinely dangerous lesion. Beyond small indels, Cas9 cutting can produce kilobase-scale deletions, inversions, translocations, chromosomal truncations, loss of heterozygosity and, at low frequency, chromothripsis. Standard amplicon sequencing across the cut site does not detect any of these, because the primer binding sites are gone. The PhD tier covers the detection methods and what is known about frequency.

The four editor families

All four share the same targeting module. What changes is what is fused to it and whether a double-strand break is made at all.

FamilyBreakWhat it can installMain constraint
Nuclease
Cas9, Cas12a
Double-strand Knockouts, deletions between paired guides, HDR knock-ins with a donor Repair outcome is stochastic; structural-variant risk; HDR needs dividing cells
Base editor
CBE, ABE
Nick only C→T and A→G transitions (and their complements) Transitions only; bystander bases in the activity window; deaminase-driven off-target on DNA and RNA
Prime editor
PE2 → PE7
Nick only All 12 substitutions, small insertions and deletions Lower efficiency, highly locus- and pegRNA-dependent; large cargo
Epigenome editor
dCas9 fusions
None Transcriptional repression or activation; heritable methylation Does not change sequence; durability and heterochromatin spreading are the open questions

Base editing, mechanically

A catalytically impaired Cas9 (usually the D10A nickase) is fused to a deaminase. R-loop formation exposes a short stretch of single-stranded non-target DNA; the deaminase acts on that window, typically positions 4 to 8 of the protospacer. Cytosine base editors deaminate C to uracil, read as T; adenine base editors use an evolved TadA to convert A to inosine, read as G. A nick on the opposite strand biases repair toward keeping the edit.

Two consequences follow directly from the mechanism. Any target base of the right identity inside the window gets edited, not just the one you wanted, which is the bystander problem. And because the deaminase has intrinsic affinity for single-stranded nucleic acid, it can act independently of the guide entirely, producing Cas9-independent off-target editing on DNA and, for ABEs derived from a tRNA deaminase, on RNA.

Prime editing, mechanically

A Cas9 nickase fused to a reverse transcriptase, paired with a pegRNA that carries both the spacer and a 3' extension containing a primer binding site and a reverse-transcription template. The nicked 3' flap anneals to the primer binding site, the RT writes the new sequence directly onto the genomic strand, and the resulting 3' flap competes with the original 5' flap for incorporation. No donor template, no double-strand break, no dependence on HDR.

The efficiency ceiling comes from three places: pegRNA degradation, competition between flaps, and mismatch repair actively reverting the edit. Each has been attacked separately, and that history is the version numbering, from PE2 through PE7.

Choosing an editor

Pick the edit you need and the tool narrows quickly.

Delivery, the actual bottleneck

Editing chemistry is no longer the limiting factor for most therapeutic programmes. Getting the machinery into the right cells is.

Lipid nanoparticles carry mRNA plus guide, express transiently, and can be redosed because they do not provoke the anti-capsid immunity that viral vectors do. Systemically administered LNPs go to the liver, largely because they adsorb ApoE and enter hepatocytes through the LDL receptor. That is a gift if your target is a hepatocyte gene and a wall otherwise.

Adeno-associated virus reaches several tissues and persists episomally, but its roughly 4.7 kb cargo limit forces split-intein or dual-vector designs for anything larger than SpCas9 alone, and sustained expression increases cumulative off-target exposure.

Engineered virus-like particles package the editor as ribonucleoprotein rather than nucleic acid. The payload is present for hours rather than weeks, which measurably lowers off-target editing relative to AAV delivery of the same editor.

Ex vivo sidesteps delivery altogether: remove cells, electroporate ribonucleoprotein, quality-control the product, reinfuse. This is how Casgevy works, and it is why Casgevy requires myeloablative conditioning and a transplant centre.

For where targeted LNPs, receptor-binder conjugates and extrahepatic tropism currently stand, see the PhD tier.

Sources

  1. Anders C. et al., Structural basis of PAM-dependent target DNA recognition by the Cas9 endonuclease, Nature 2014, PubMed 25079318.
  2. Sternberg S.H. et al., DNA interrogation by the CRISPR RNA-guided endonuclease Cas9, Nature 2014, PubMed 24476820.
  3. Dagdas Y.S. et al., A conformational checkpoint between DNA binding and cleavage by CRISPR-Cas9, Science Advances 2017, science.org.
  4. Pacesa M. et al., R-loop formation and conformational activation mechanisms of Cas9, Nature 2022, PMC9433323.
  5. Cofsky J.C. et al., CRISPR-Cas9 bends and twists DNA to read its sequence, Nature Structural & Molecular Biology 2022, PMC9189902.
  6. Repair-outcome overview, BioMed Research International 2018 Fig. 2, PMC6167567; and J Clin Invest 2022 Fig. 1, PMC8718140.
  7. Structural variants after editing, Nature Communications 2025, s41467-025-62606-z.
  8. Base editing review, J Cell Mol Med April 2026, PMC13109818.
  9. Prime editing review, Experimental & Molecular Medicine 2025, nature.com.
  10. Delivery platforms review, PMC12865601; in vivo delivery strategies, Int J Biol Sci July 2026, ijbs.com.