Off-target effects in genome editing

Off-target effects in genome editing are DNA changes that an editor makes where it was not intended to act. In practice the term has widened to cover a second and arguably more consequential category: unintended outcomes at the correct site, where the editor cut exactly where instructed and the cell's repair produced something other than the desired edit. Off-target cutting was reported within a year of CRISPR–Cas9 being adapted to human cells.¹ The on-target failure modes took most of the following decade to characterise, and both classes have absorbed a substantial share of the field's effort since.

§What counts as an off-target effect

Guide RNAs tolerate mismatches, particularly at positions distant from the PAM, so a twenty-nucleotide spacer typically has dozens to hundreds of partially matching sites in a mammalian genome. Whether any of them is actually cut depends on chromatin accessibility, the concentration and persistence of the editor, and the specific mismatch pattern. Cutting at such a site produces the same indels as an on-target cut, in a gene nobody chose.

Editors that do not cut have their own version of the problem. Base editors carry a deaminase that can act on transiently exposed single-stranded DNA anywhere in the genome, entirely independently of where the guide RNA directs the protein — a mechanism that leaves scattered single-base changes with no cut site to search near. Prime editors show lower guide-dependent off-target activity than nucleases, since three separate base-pairing events are required, but can insert fragments of their own guide RNA at the target. Editors that only bind DNA, as in Epigenome editing, perturb transcription at sites they would never cut, because binding tolerates more mismatches than cleavage does.

§Detection

No single assay is adequate, and regulators now expect at least two orthogonal ones. The families differ in what they trade away.

The standard workflow nominates candidate sites broadly, then sequences those sites deeply in the actual edited cell product. Cell-based tagging methods capture double-strand breaks as they occur; one approach instead pulls down a DNA-repair protein that gathers at breaks, which allows the measurement to be made in a living animal rather than a dish.² Whole-genome sequencing is conceptually cleaner but too insensitive for events present in a small fraction of cells.

The reference genome is not the patient

Off-target site nomination usually starts from a reference sequence. A patient carrying a common variant inside a candidate site may have a near-perfect match that the reference does not show. This became a specific question for sickle cell editing therapies, where the affected population is genetically diverse and under-represented in reference datasets.

§On-target damage

The larger surprise of the past decade was that cutting the right place can still go wrong. Repair of a Cas9-induced break produces, at measurable frequency, deletions of many kilobases, inversions and complex rearrangements that standard short-amplicon sequencing does not detect because the primer sites themselves are lost.³ More severe outcomes have been documented: whole-chromosome-arm loss, loss of heterozygosity extending far from the cut, and chromothripsis, in which a chromosome shatters and is stitched back together in scrambled order.⁴

Breaks also activate a p53-mediated damage response, which reduces editing efficiency in cells with intact p53 and therefore selects, weakly but systematically, for cells in which that pathway is impaired.⁵ The practical worry is not that editing causes cancer in any demonstrated case, but that the standard editing protocol applies a selection pressure whose direction is unfavourable.

Multiplexed editing compounds the problem. Making several cuts at once — routine in engineered cell therapies and in the heavily edited donor animals used for Xenotransplantation — creates opportunities for translocation between the cut sites, and the frequency rises with the number of simultaneous breaks.

§Editing in embryos

The embryo is the worst case for every failure mode at once. Editing after the first cell division produces mosaicism, so the resulting individual carries a mixture of genotypes and a biopsy of a few cells does not represent the rest. Studies of human embryos have reported large on-target segmental losses and loss of heterozygosity around the cut site at appreciable frequency.⁶ An influential 2017 report interpreted apparent correction of a paternal allele as repair templated by the maternal chromosome; critics argued the more likely explanation was allele dropout, in which the paternal allele was deleted and simply failed to amplify.

This is why editing in embryos is not a scaled-down version of editing in a dish. The verification problem is structural: certifying that an embryo carries no unintended change requires sequencing cells that will not become the person, and the mosaicism that makes editing risky is the same property that makes the biopsy uninformative. Where the goal is avoiding a known monogenic disease, Embryo selection achieves it without editing in almost every case, which is the strongest practical argument against Human germline editing independent of any ethical position.

§Mitigation

Several strategies reduce guide-dependent off-target activity, none to zero.

• Engineered high-fidelity nucleases with weakened non-specific DNA contacts, which lose activity faster on mismatched sites than on matched ones.⁷
• Short exposure, by delivering preassembled protein–RNA complexes that degrade within hours rather than plasmids or viral vectors that express for weeks. This is one reason ex vivo editing is safer than in vivo editing with AAV vectors, and why messenger RNA in Lipid nanoparticles is preferred to persistent expression.
• Paired nickases, requiring two adjacent binding events to produce a break.
• Guide selection that avoids sites with close genomic relatives and avoids known polymorphic positions.
• Switching editor class, using Base editing or Prime editing where a double-strand break is not needed — which trades one error spectrum for another rather than eliminating error.

§Regulatory practice

The United States Food and Drug Administration issued final guidance on genome editing products in 2024 setting out expectations: orthogonal off-target nomination assays, confirmation in the clinical cell type, justification of any residual risk, and long-term follow-up of treated patients. The approval of Casgevy in late 2023 was the first regulatory test of that framework, and the advisory discussion around it turned less on whether off-target edits had been found than on whether the assays used could have found them. That distinction — absence of evidence versus evidence of absence — remains the crux of every editing safety package.

For Somatic gene therapy the residual uncertainty is managed the way oncology risk is managed: quantified where possible, monitored over years, and weighed against the disease. The frameworks surveyed in Governance of human genome editing apply a different logic to heritable edits, and the He Jiankui affair is the standing example of what happens when the safety analysis is performed by the person with an interest in the answer.

§Open problems

The field lacks a validated way to say how much off-target editing is too much, because no assay's sensitivity floor is well characterised in the tissue that matters. There is no established method for detecting rare structural rearrangements across a whole treated organ, no consensus on how to handle patient-specific variants at candidate sites, and little long-term human data — the treated cohorts are small and young. Editors deployed outside medicine, notably in gene drives released into wild populations, face the same detection limits with none of the follow-up infrastructure, which is one reason the Precautionary principle is invoked more forcefully there than in the clinic.

§See also

• CRISPR–Cas9
• Base editing
• Prime editing
• Human germline editing
• Casgevy
• Governance of human genome editing
• Embryo selection
• He Jiankui affair

§References