Scope. Written for readers who work in or adjacent to genome engineering. Covers the kinetic and structural basis of target discrimination; off-target and structural-variant surveillance and where the assays disagree; the engineering lineages behind current deaminase and prime editing architectures; RNA-guided recombination and generative-model-designed effectors; delivery constraints; and the clinical record through September 2026, including the programme that killed a patient.
Stance on citations. Every quantitative claim is attributed. Where a result is preprint, single-study, or from a company release rather than peer review, that is stated inline rather than buried. Interactive figures are qualitative teaching models and are labelled as such.
Target location is not a homology search over the genome. Non-target binding affinity scales with PAM density, and sequences fully complementary to the guide but lacking a proximal PAM are ignored.1 The PI domain reads the 5'-NGG-3' motif through major-groove contacts from conserved arginines in a base-paired duplex, with minor-groove and phosphate contacts at the +1 position driving local strand separation immediately upstream.2
The unwinding step itself is ATP-independent. Cryo-EM and solution probing show that PAM binding sharply bends and undertwists the duplex, flipping nucleotides out toward the guide for interrogation; the resulting unstacked DNA hinge is accompanied by global protein rearrangement.3 This provides the physical mechanism for complementarity-guided search and predicts that local DNA geometry modulates editing efficiency at otherwise equivalent sites.
Competition assays established that strand separation and heteroduplex formation initiate at the PAM and propagate toward the distal end.1 Structures capturing intermediate states show REC2 and REC3 forming a positively charged cleft that accommodates the distal duplex early in R-loop formation; hybridisation past the seed drives REC2/REC3 rearrangement and relocates HNH into a catalytically incompetent checkpoint conformation.4
Single-molecule FRET resolves HNH sampling multiple conformations before docking. Docking requires divalent cations but not scission, and the docked state persists after cleavage.5 With mismatched targets an inactive intermediate accumulates: HNH must cross a conformational checkpoint to reach the active conformation, and docking is blocked when complementarity falls below a threshold.5 Completion of the heteroduplex triggers activation, enabled by distortion of the guide–target duplex and complementary REC2/REC3 motions.4
The practical corollaries are worth stating explicitly, because they are routinely conflated in guide-design discussions:
Recent work extends the checkpoint concept upstream of the heteroduplex itself, identifying a guide RNA repeat clasp whose integrity is allosterically coupled to catalytic efficiency, alongside the established REC-domain and bipartite seed checkpoints.6 That result is preprint as of this writing.
Guide-dependent off-target editing at homologous sites is the well-characterised mode and the one most assays are built for. Guide-independent editing is a separate problem specific to fusion editors, arising from intrinsic deaminase affinity for single-stranded substrate, and is invisible to any assay that nominates sites by sequence homology (§3).
In a head-to-head comparison in edited CD34+ HSPCs, DISCOVER-seq and GUIDE-seq showed the highest sensitivity and positive predictive value, with true off-target events appearing at the top-ranked call for each; CHANGE-seq, CIRCLE-seq and SITE-seq also ranked true sites at or near the top with lower PPV. Across the NGS validation panels the study found more than two orders of magnitude more false positives than true off-target sites.7 That ratio, not the raw nomination count, is the number to carry into a regulatory package.
| Method | Context | Detects | Blind to |
|---|---|---|---|
| Cas-OFFinder, COSMID, CRISPRme | in silico | Homology-nominated candidate sites; CRISPRme incorporates variant haplotypes | Anything not predicted by sequence similarity |
| CIRCLE-seq, CHANGE-seq, SITE-seq | in vitro | Cleavage on purified genomic DNA, high sensitivity | Chromatin context; overcalls relative to cells |
| GUIDE-seq | cell-based | dsODN capture at breaks in living cells | Requires oligo uptake; low-frequency events |
| DISCOVER-seq | cell-based, in vivo capable | MRE11 recruitment at genuine breaks; no exogenous tag | Transient or repaired-before-capture events |
| CAST-seq, LAM-HTGTS, PEM-seq | translocation-focused | Large deletions, inversions, intra- and inter-chromosomal translocations | Small indels; requires substantial input |
| Amplicon NGS across the cut site | validation | Small indels at nominated sites | Any lesion that removes a primer site: large deletions, truncations, chromothripsis |
The more consequential safety question is increasingly on-target rather than off-target. Documented outcomes at the intended site include kilobase-scale deletions, inversions, translocations, megabase chromosomal truncations, copy-neutral loss of heterozygosity, loss of imprinting, whole-chromosome loss, and chromothripsis.89 Frequent aneuploidy has been reported in primary human T cells after Cas9 cleavage.9 None of these are captured by standard genotyping.8
Cell-type dependence is real: karyotypically unstable cancer lines skew toward large aberrations and translocations, while genetically stable cells show mostly kilobase-scale deletions, insertions and rearrangements at the on-target site.8
Two process variables deserve attention because they are under the operator's control. First, the p53 trade-off: transient p53 inhibition has been reported to increase HDR efficiency by up to 17-fold in human pluripotent stem cells, but inducing double-strand breaks in a p53-compromised background plausibly raises mutational burden and structural-variant incidence.8 DNA-PKcs inhibition carries the same structural concern.10 Second, manufacturing: extended culture and proliferation-promoting reagents correlate with higher rates of long deletions and micronuclei, and G-CSF administration during transplant has been reported to amplify the p53-mediated damage response to Cas9-induced breaks.9
The counterweight is clinical: published long-term follow-up is encouraging, with CRISPR-induced aneuploidies in CAR-T products appearing to be mitigated by extended culture and after engraftment in patients.10 The honest summary is that the in vitro lesion catalogue is alarming, the clinical signal so far is not, and the surveillance methods that would reconcile the two are not yet standard in trial packages.
ABEs were built by directed evolution of E. coli TadA, a tRNA adenosine deaminase, into a variant with single-stranded DNA activity, fused to nCas9(D10A). ABE8e emerged from further evolution of the TadA monomer with a 3- to 11-fold activity improvement over ABE7.10 and a correspondingly widened editing window.11
Increased deamination rate came with costs that are mechanistically inseparable from the gain:
Screening TadA orthologs rather than further evolving one has produced ABEs, CBEs and dual ACBEs from single or double mutations across 25 orthologs, with minimised Cas9-independent DNA off-target activity and genotoxicity.14 For a target that tolerates no bystander edit, ortholog choice and window width now matter more than raw on-target efficiency.
Prime editing's version numbers are not marketing. Each increment attacks a specific, identified rate-limiting step, and knowing which step a variant addresses tells you whether two improvements will stack.
PE3 adds a nicking sgRNA on the non-edited strand to bias resolution toward the edit, at the cost of indels. MMR actively reverts replacement-type edits; a dominant-negative MLH1 (MLH1Δ754–756, MLH1dn) relieves this, either by forming a catalytically impaired MutLα with PMS2 or by saturating MSH2 binding, dose-dependently. PE2 plus MLH1dn is PE4; with the additional nick it is PE5.1516
The 3' extension is exposed to exonucleolytic degradation. Two independent solutions exist: structured motifs appended to the 3' end (evopreQ1, mpknot, giving epegRNAs), and recruitment of the La protein, which binds and stabilises pegRNA 3' tails. Fusing La to PEmax gives PE7.15 Because both address the same bottleneck, combining PE7 with epegRNAs yields similar or reduced efficiency rather than additive gain16 — the clearest illustration of why the mechanism-first framing matters. In 2026 the Liu lab reported additional epegRNA motifs outperforming the original design.16
PEmax is a codon- and NLS-optimised PE2. PE6 introduced compact evolved reverse transcriptases and improved nCas9(H840A) variants.17 Repositioning the nick in the PAM-distal direction suppresses unintended indel formation, giving the xPE and vPE architectures; vPE additionally carries the La fusion.16
A de novo MLH1 small binder (MLH1-SB) was designed with RFdiffusion and AlphaFold 3 to bind the MLH1–PMS2 dimer interface, including an AlphaFold-3-based competition test to filter candidates predicted to bind weakly. At 82 amino acids it integrates into existing architectures via 2A peptides. PE7-SB2, carrying two copies linked to PE7, reported an 18.8-fold increase over PEmax and 2.5-fold over PE7 in HeLa cells, and 3.4-fold over PE7 in mice.18 A separate 2026 report adds a PSBP module that synergises with La-mediated stabilisation and can be consolidated into a single-component editor.19
Note what these numbers are and are not: fold-changes over a stated baseline in specific cell lines and loci. Absolute efficiency remains strongly locus- and pegRNA-dependent, and cross-study comparison of fold-changes against different baselines is not meaningful.
IS110 insertion sequences encode both a recombinase and a structured non-coding RNA that binds it. This bridge RNA carries two internal loops that base-pair independently with target DNA and with donor DNA (the IS110 element itself), making it the first described bispecific guide molecule. Both loops are independently reprogrammable, enabling sequence-specific insertion, excision and inversion — the three fundamental rearrangements of genome design — through a single effector.20
The structural mechanism was resolved in parallel. The synaptic complex comprises two recombinase dimers, one binding the target-binding loop with target DNA and the other the donor-binding loop with donor DNA, forming a composite RuvC–Tnp active site spanning the dimers that positions catalytic serines adjacent to the recombination sites. Cryo-EM captured the reaction sequence: top-strand cleavage to covalent 5'-phosphoserine intermediates, strand exchange and religation into a Holliday junction, then resolution by bottom-strand cleavage.21
Demonstrations to date are in E. coli. Eukaryotic efficiency, chromatin accessibility and specificity remain the open questions, and nothing here should be read as clinic-adjacent yet. The significance is architectural: an insertion mechanism that does not route through a double-strand break or through HDR.
OpenCRISPR-1 is a Cas9-like nuclease and guide RNA generated by protein language models trained on a large CRISPR-Cas sequence corpus, retaining prototypical Type II architecture while sitting hundreds of mutations from SpCas9 or any known natural Cas protein, with NGG PAM preference. It was published in Nature in 2025 and released under an open licence.22 Reported activity, specificity and immunogenicity outperformed structure-based and baseline design methods, and the underlying CRISPR-Cas Atlas was subsequently opened.23
The immunogenicity angle is the one with therapeutic teeth. Pre-existing adaptive immunity to SpCas9 and SaCas9 in humans is a genuine constraint on redosing and on in vivo delivery of protein editors; a synthetic effector far from any natural ortholog is a plausible route around it, though this needs independent confirmation.
Hepatotropism of systemically administered LNPs is a formulation-level fact, not a therapeutic strategy. Two directions are being pursued to escape it.
Physicochemical tuning. Particle size, lipid ratio and apparent pKa shift organ tropism, with high-pKa formulations favouring lung and low-pKa formulations accumulating in spleen. Transient macrophage depletion enhances liver-directed delivery while reducing extrahepatic targeting, implicating immune clearance in biodistribution.24
Ligand targeting. DARPin-conjugated LNP formulations have achieved up to 98% binding and roughly 90% expression in human CD8+ T cells, demonstrating extrahepatic delivery to a defined population.25 Antibody-fragment approaches against CD3, CD5 and CD7 have generated anti-CD19 CAR-T in vivo in mice with resulting B-cell aplasia.26 A clinical-stage targeted-LNP CAR-T candidate entered Phase 1 in June 2025.26
Local administration bypasses systemic barriers entirely: intramuscular for dystrophin restoration, intratumoral for glioblastoma models, nebulised for pulmonary editing, intracameral for ocular tissue.24
Particle-delivered RNP. eVLPs and lentivirus-derived nanoparticles carry protein rather than nucleic acid, collapsing the exposure window from weeks to hours. AAV delivery of a base editor produced detectable off-target editing in mouse liver where eVLP delivery of the same editor did not.27 The general principle — that editor dwell time, not peak activity, drives cumulative off-target burden — argues against overloading cells with editor complexes.27
Where the pathology is expression rather than sequence, a dCas9 fused to repressive effectors avoids the break entirely. CRISPRoff-type architectures deposit heritable methylation; silencing of Pcsk9 in mouse liver has been maintained for over a year.28
The first human evidence arrived in 2026. TUNE-401 delivers mRNA encoding an epigenetic silencer plus a guide RNA against a conserved HBV control sequence, by LNP to hepatocytes, targeting cccDNA and integrated HBV DNA without altering sequence. Phase 1b/2a data presented at EASL in May 2026 reported durable repression of four HBV biomarkers in all treated patients, described by the sponsor as 95–100% silencing, with Phase 2 expected to initiate as early as late 2026.2930 These are sponsor-presented conference data, not a peer-reviewed publication.
Epicrispr's FSHD programme applies the same logic to D4Z4 hypomethylation, delivering a methyltransferase fusion by muscle-tropic viral vector; three of four treated patients evaluated at three months showed strength and function improvements exceeding matched natural-history comparators.31
The unresolved liability is heterochromatin spreading. Repressive marks deposited at integrated viral DNA, which inserts at effectively random genomic positions, could propagate into neighbouring host genes; the severity of that risk is currently unknown.28
Casgevy (exagamglogene autotemcel) remains the only approved CRISPR medicine: UK November 2023, US December 2023 for sickle cell disease with recurrent vaso-occlusive crises, extended to transfusion-dependent β-thalassaemia in January 2024, and to patients aged 2 and older on 1 July 2026 — the latter granted 53 days after filing under the Commissioner's National Priority Voucher pilot.3233 Mechanism is BCL11A erythroid enhancer disruption in autologous HSPCs to derepress fetal haemoglobin, with full myeloablative conditioning before infusion.33
Durability is strong: 16 of 17 SCD patients free of vaso-occlusive crises, 25 of 27 TDT patients transfusion-independent, some beyond three years.34 Access is the constraint. Priced at 2.2 million US dollars, 64 patients received the treatment in 2025, though initiation and cell collection rose roughly threefold year over year and the sponsor reported 90% of US patients with reimbursed access at the start of 2026.34
This programme deserves detail, because it is the field's most serious safety event and the temptation to skip past it should be resisted.
Nexiguran ziclumeran (nex-z, NTLA-2001) is LNP-delivered Cas9 targeting TTR in hepatocytes for ATTR amyloidosis, and was the first systemically administered in vivo CRISPR therapy. Phase 1 showed roughly 90% TTR reduction sustained across all 27 participants reaching two years, with stability or improvement on functional measures.34 Three low-dose participants received a second infusion at the higher dose, the first reported redosing of an in vivo CRISPR therapy — feasible precisely because LNPs do not elicit anti-capsid immunity.34
A patient dosed on 30 September 2025 in the Phase 3 MAGNITUDE trial developed grade 4 transaminase elevation with raised total bilirubin, was hospitalised on 24 October, and died on the night of 5 November 2025. The sponsor paused dosing and screening on 27 October; the FDA placed clinical holds on MAGNITUDE and MAGNITUDE-2 on 29 October 2025. Of more than 650 enrolled in MAGNITUDE with over 450 dosed, fewer than 1% experienced grade 4 transaminase elevations.3536 Patient screening in both trials resumed by the first quarter of 2026.37
The mechanism of the hepatotoxicity has not been publicly resolved. The plausible contributors — LNP-associated hepatocellular stress, innate sensing of the mRNA cargo, on-target consequences in a diseased liver — are not distinguishable from the published record, and the field should be careful not to assume it was the editing chemistry.
Lonvoguran ziclumeran (lonvo-z, NTLA-2002) for hereditary angioedema knocks out KLKB1. Phase 1/2 in 32 patients reported 97% attack- and prophylaxis-free with effects durable to 32 months; Phase 3 HAELO enrolment completed, the sponsor reported the first Phase 3 data for an in vivo gene editing candidate and initiated a rolling BLA, with launch targeted for the first half of 2027.3738
VERVE-102 delivers an adenine base editor as mRNA with guide RNA by LNP to inactivate hepatocyte PCSK9. Phase 1b Heart-2 interim results, presented at EAS and published in NEJM on 25 May 2026, covered 35 participants across six dose cohorts from 0.3 to 1.0 mg/kg, with a 27 February 2026 cutoff and median follow-up around nine months. Mean PCSK9 reduction ranged 51% to 88%; mean LDL-C reduction was 9%, 44%, 45%, 33%, 51% and 62% by ascending cohort, sustained to 18 months. Related adverse events were low-grade infusion reactions and fatigue.39 The non-monotonic LDL-C series across cohorts is worth noting rather than smoothing over; cohort sizes are small.
Beam's AATD programme is among the first attempts at direct in vivo correction of a pathogenic point mutation. Over 25 participants dosed; at the highest dose approximately 90% of circulating AAT was the corrected form by day 14, sustained at day 28.34 Prime Medicine reported the first clinical efficacy and safety data for prime editing in humans in CGD, published in NEJM in December 2025, with two participants showing restoration of the protein required for normal neutrophil function and remaining infection-free at four and six months — after which the company deprioritised the programme in a strategic reorganisation.34
KJ Muldoon, an infant with CPS1 deficiency, received a patient-specific in vivo base editing therapy developed and dosed in about six months, reported in NEJM in May 2025; he received three doses, each further reducing symptoms, and as of early 2026 was walking with mild symptoms.4034
The regulatory consequence was rapid. Prasad and Makary outlined a "plausible mechanism" pathway in NEJM in November 2025; the FDA published draft guidance in February 2026 permitting a single trial to test a platform customised per patient, conditional on the cause being defined at genetic or cellular level, the natural history being known, the therapy targeting that cause, and demonstration of both target modulation and clinical improvement. The UK MHRA announced parallel plans.4142 The CHOP group intends a urea cycle disorder platform trial under this framework.42
Separately, in vivo CAR-T generation using CRISPR was demonstrated in Nature in 2026, with clinical trials anticipated.42