- What Off-Target Effects Actually Are
- How Off-Target Cuts Happen: The Molecular Mechanism
- Detection Method 1: GUIDE-seq
- Detection Method 2: CIRCLE-seq & Digenome-seq
- Detection Method 3: Whole-Genome Sequencing
- Beyond Point Mutations: Large Deletions & Translocations
- Mosaicism: The Incomplete Edit Problem
- The p53 Problem: When DNA Damage Selects for Cancer
- High-Fidelity Cas9 Variants
- The Story: The Surgeon with Trembling Hands
Section 1 — What Off-Target Effects Actually Are
An off-target effect in CRISPR gene editing is any unintended modification to the genome — a cut, insertion, or deletion at a location other than the intended target site. The term covers a spectrum of events ranging from clinically insignificant nibbles at non-functional genomic regions to potentially serious cuts within protein-coding genes or cancer-relevant sequences.
The existence of off-target effects does not mean CRISPR is unsafe. It means CRISPR is not perfect, which is different. Every medicine has side effects. Every surgical procedure carries risks. The question is not whether CRISPR produces off-target cuts — it sometimes does — but whether those cuts occur at frequencies and locations that produce clinically meaningful harm. For most well-designed guide RNAs in research applications, off-target effects are rare and inconsequential. For therapeutic applications, the bar is higher, and characterising off-target activity is a central part of the regulatory process.
Section 2 — How Off-Target Cuts Happen: The Molecular Mechanism
From Cluster 3, you know that Cas9 searches the genome for NGG PAM sequences, then checks whether the adjacent 20 nucleotides match the guide RNA spacer by forming an R-loop. The specificity of this process is high but not absolute. Off-target cuts occur when Cas9 tolerates mismatches between the guide RNA and the genomic DNA — binding the R-loop despite imperfect complementarity and proceeding to cut.
The key insight is that Cas9 has a kinetic proofreading problem: it has to decide “match or no match?” based on how stable the R-loop is at any given instant. A high-quality match produces a very stable R-loop, which quickly triggers the conformational change that activates HNH and RuvC. A poor match produces a less stable R-loop — but if the R-loop is stable enough, and if Cas9 remains bound long enough, cutting can still occur.
Which Mismatches Are Tolerated?
Not all mismatches are equal. The position, number, and type of mismatch all determine whether Cas9 will cut an off-target site:
Section 3 — Detection Method 1: GUIDE-seq
GUIDE-seq (Genome-wide Unbiased Identification of DSBs Evaluated by Sequencing), developed by the Joung lab at Massachusetts General Hospital in 2015, was the first method to comprehensively detect CRISPR off-target cuts across the entire genome in living cells. It remains one of the most widely used and trusted methods.
The principle is elegant. GUIDE-seq exploits the fact that after any DSB, the cell’s repair machinery briefly exposes single-stranded DNA at the break site. The researchers insert a short double-stranded DNA oligonucleotide (the “dsODN tag”) into cells alongside the CRISPR components. This tag has blunt ends with phosphorothioate modifications that make it resistant to cellular degradation. When Cas9 creates a DSB anywhere in the genome, the NHEJ repair machinery frequently captures and integrates one of these tags at the break site.
After editing, the researchers extract genomic DNA and use PCR with one primer matching the dsODN tag and one primer from flanking sequences, followed by high-throughput sequencing. Every site in the genome that captured a tag — every DSB site, on-target and off-target — is sequenced and its chromosomal coordinate identified. The number of sequencing reads at each site is roughly proportional to the cutting frequency.
Key advantage: Unbiased — detects DSBs regardless of their DNA sequence, including sites with bulges and unusual mismatches that computational prediction would miss. Key limitation: Requires cells to be transfected with the tag alongside CRISPR, so it cannot be used retrospectively on patient samples.
Section 4 — Detection Methods 2 & 3: CIRCLE-seq, Digenome-seq, and WGS
CIRCLE-seq: Cell-Free Ultrasensitive Detection
CIRCLE-seq is an in vitro method that avoids the need for cell transfection altogether. Purified genomic DNA is circularised using a ligase, then treated with Cas9 and guide RNA in a test tube. Because only linearised (cut) DNA can be efficiently sequenced from these circular templates, sequencing is highly enriched for cut sites. CIRCLE-seq can detect cutting events occurring in fewer than 1 in 10,000 molecules — orders of magnitude more sensitive than GUIDE-seq.
The trade-off: because CIRCLE-seq is performed on naked purified DNA, it does not account for chromatin structure. Many sites detected by CIRCLE-seq are in closed chromatin regions that would be inaccessible to Cas9 in actual cells. This makes CIRCLE-seq highly sensitive but somewhat over-predictive of true in-cell off-target activity. In practice, researchers use CIRCLE-seq to generate a comprehensive list of potential sites, then validate the most relevant ones by targeted sequencing in cells.
Digenome-seq: Whole Genome-Scale In Vitro Detection
Digenome-seq (Digested Genome sequencing) works similarly: Cas9 is used to cut purified genomic DNA in vitro, and the resulting fragments are subjected to whole-genome sequencing. Cut sites generate fragments with uniform 5’ ends at the cut position, which appear as a sharp peak in sequencing read depth at that position. Digenome-seq can detect sites cut at frequencies as low as 0.1% and covers the entire genome without preselection.
Whole-Genome Sequencing (WGS): The Clinical Standard
For clinical gene therapy programmes, whole-genome sequencing of edited cells is the definitive off-target assessment method. Deep WGS (30–100× coverage) of edited vs unedited cells from the same patient allows detection of any new variant — SNPs, indels, structural variants — that arose during the editing process. This is the method used in clinical trial safety assessments and regulatory submissions.
Section 5 — Beyond Point Mutations: Large Deletions and Chromosomal Translocations
For many years, the CRISPR safety conversation focused primarily on small indels at off-target sites — single nucleotide changes or short insertions/deletions caused by erroneous NHEJ repair. These are serious but relatively easy to detect. A more alarming set of findings emerged around 2018–2022: CRISPR-Cas9 can also produce large genomic deletions and chromosomal translocations that are harder to detect and potentially more consequential.
Large Deletions: Kilobases to Megabases
A study by Kosicki et al. (2018) reported that CRISPR-Cas9 can delete kilobases to megabases of sequence around the cut site, in a subset of edited cells. These large deletions result from extensive DNA end-processing after the DSB — instead of clean NHEJ, the broken ends are extensively resected before rejoining, removing large stretches of flanking sequence. Standard PCR-based editing efficiency assays completely miss these events because the PCR primers are within the deleted region and simply fail to amplify from affected cells.
The frequency of large deletions varies by cell type and guide RNA, but they have been detected in mouse embryos, human embryonic stem cells, haematopoietic stem cells, and primary T cells. Their clinical significance depends entirely on which genes are deleted — a 500 kb deletion in a gene-poor intergenic region may have no effect, while the same deletion spanning a cluster of regulatory elements or multiple genes could have severe consequences.
Chromosomal Translocations: When Two Breaks Meet
When CRISPR cuts at two sites simultaneously — either intentionally (in multiplexed editing) or inadvertently (an on-target cut and an off-target cut happening in the same cell) — the two broken chromosomal ends can be joined to each other instead of to their own original ends. The result is a chromosomal translocation: a rearrangement where part of one chromosome is attached to another.
Chromosomal translocations are the hallmark of many haematological cancers. The BCR-ABL translocation that causes chronic myeloid leukaemia, the PML-RARA translocation in acute promyelocytic leukaemia — both are formed by exactly this mechanism: two chromosomal breaks repaired incorrectly. CRISPR-induced translocations have been detected in edited cells at frequencies of 0.1–1% in experimental settings. At these frequencies, in a batch of tens of millions of edited cells (as in Casgevy production), some cells will carry translocations. The clinical significance requires long-term monitoring.
Section 6 — Mosaicism: The Incomplete Edit Problem
Even when CRISPR editing is highly efficient, it rarely edits every cell in a treated tissue or organism simultaneously. Mosaicism is the condition where different cells in the same organism carry different genetic compositions — some with the intended edit, some with different indels at the target site, and some unedited.
Mosaicism arises because CRISPR editing is a stochastic (random) process. Each time Cas9 creates a DSB at the target site, NHEJ repair produces a random indel. Two cells in the same editing experiment may both be edited at the same site but carry completely different indels — one cell may have a +1 insertion, another a −5 deletion. If you are trying to make a knockout, most indels will be frameshifts that disrupt the gene, so mosaicism may not matter functionally. But if you are trying to correct a specific mutation by HDR, some cells may have the correction while others have a disruptive indel at the same position.
Germline Mosaicism: A Special Concern
Mosaicism becomes especially important in the context of embryo editing. If CRISPR is delivered to a zygote (fertilised egg), editing begins immediately — but the first few cell divisions happen before the Cas9 has had a chance to edit every cell. The result can be a mosaic embryo where some cells carry the intended edit and others do not. In the worst case, the edited allele is present in only a minority of cells, rendering the edit ineffective. In the He Jiankui case (the unauthorised germline editing of human embryos), mosaicism was documented in the edited twins — suggesting the editing was incomplete and the intended outcome (CCR5 disruption in all cells) was not achieved uniformly.
Section 7 — The p53 Problem: When DNA Damage Selects for Cancer-Prone Cells
In 2018, two papers published simultaneously in Nature Medicine from the Bhatt lab (Novartis) and the Weissman lab (Karolinska) raised a concern that sent ripples through the CRISPR therapeutics community. They showed that in human pluripotent stem cells (hPSCs) and primary T cells, CRISPR-Cas9 editing triggers a p53-dependent DNA damage response that selects against successfully edited cells.
Here is the mechanism. p53 is the cell’s master guardian against DNA damage. When Cas9 creates a DSB, p53 is activated. In cells with functional p53, p53 activation causes cell cycle arrest or apoptosis (programmed cell death) — the cell is eliminated to prevent potentially damaged DNA from being replicated. This is p53 doing exactly what it is supposed to do. But the consequence for CRISPR editing is counterintuitive: the cells that are most efficiently edited (those that received more Cas9 and experienced the cut) are preferentially eliminated by p53, while unedited cells that received insufficient Cas9 survive and proliferate.
More alarmingly, the cells that survive the p53 response and expand after editing are enriched for cells with pre-existing p53 mutations — cells in which p53 is non-functional and therefore cannot mount an apoptotic response to the DSB. These cells escape CRISPR-induced death not because they were not edited, but because their p53 is already broken. Since p53 mutations are a hallmark of cancer and are found at low frequency in normal tissues (as background somatic mutations), the CRISPR editing process could potentially select for pre-cancerous cells.
Section 8 — High-Fidelity Cas9 Variants: Engineering Better Scissors
The field has not passively accepted off-target effects as an unchangeable property of Cas9. Starting around 2016, multiple groups engineered SpCas9 variants with dramatically reduced off-target activity while maintaining near-wild-type on-target efficiency. These high-fidelity variants are now standard in therapeutic applications.
The engineering strategy in most cases is similar: reduce the non-specific DNA contacts that help stabilise the R-loop, forcing Cas9 to rely more completely on guide RNA base-pairing for stable binding. This makes Cas9 less tolerant of mismatches, because the non-specific contacts that could compensate for a mismatch have been removed. The price: slightly lower on-target efficiency (typically 10–30% reduction) in exchange for dramatically reduced off-target activity.
Beyond High-Fidelity Cas9: Avoiding DSBs Altogether
The most radical solution to off-target DSB risk is to use editing tools that do not create double-strand breaks. Base editors and prime editors (covered in detail in Cluster 7) use a catalytically impaired Cas9 (nCas9 or dCas9) that creates at most a nick (single-strand break) rather than a full DSB. Without a DSB, the p53 response is not strongly triggered, large deletions and translocations are virtually absent, and the NHEJ-mediated indel risk at off-target sites is eliminated. Off-target base editing can still occur (deamination at non-target cytosines near the editing window), but this is a different and generally lower risk than full DSB-based off-targeting.
The Surgeon with Trembling Hands — and the Audit Trail
Our molecular surgeon Cas9 is extraordinarily skilled. In the vast city of 3.2 billion doors, she finds the right one almost every time. But “almost” is not “always.” On rare occasions, she pauses at a door where the address almost matches her map — eighteen of the twenty characters are correct, and she opens it anyway. Inside is not the broken machine she came to fix. It is someone else’s room, and she has just accidentally tampered with it.
This is off-target cutting. And the first question anyone asks is: does it matter? The answer depends entirely on whose room she entered. If she walked into an empty storeroom with no furniture — an intergenic region between genes — the accidental opening is inconsequential. If she walked into a room housing a critical safety mechanism — a tumour suppressor gene — and accidentally broke something, the consequences could unfold over years. Not immediately. But potentially.
How do you know which rooms she accidentally entered? That is the job of off-target detection. GUIDE-seq is like installing a GPS tag on the surgeon that leaves a footprint at every door she opens. After the procedure, you collect all the footprints and map them. Every door — intended or not — has her signature at it. You can now see: she opened 847 doors today. One was the target. Three were wrong doors. Two of the three wrong doors were in empty storerooms. One was near a gene that matters. That one needs investigation.
CIRCLE-seq is more powerful but different in character. Instead of tracking the surgeon in the real city, you build a perfect replica of the city layout in a laboratory, and send the surgeon through it there. In the replica, you can detect even the slightest touch — a door she considered opening but ultimately did not. This finds more potential wrong doors than the real city method. But some of those doors are sealed in the real city by locked chromatin — she could never have opened them there even if she wanted to. The replica gives you the list of every risk. The real-city method tells you the actual risks she acted on.
The large deletion problem is more alarming than a misidentified door. Occasionally, after opening the right door and completing the repair, the surgeon does not cleanly close the door behind her. Instead, the closing mechanism malfunctions, and the entire doorframe collapses, taking out the doors on either side and the wall between them — a large deletion, sometimes spanning tens of thousands of base pairs. You would not notice this with a quick look at the front of the building. Only a comprehensive structural survey of the entire block reveals the damage.
And then there is the p53 problem — the most counterintuitive safety issue. After the surgeon finishes her work, the building’s safety inspector (p53) comes by and evaluates the rooms where repairs were made. Rooms with good structural integrity pass inspection and continue operating. Rooms with signs of too much disruption are flagged for demolition (apoptosis). This sounds like a good system. But the rooms that pass inspection most easily are those where the safety inspector’s own office has already been vandalized — where p53 itself is non-functional. Those rooms, already pre-cancerous, escape elimination and may proliferate. The editing process has inadvertently selected for the most dangerous tenants.
The high-fidelity Cas9 variants are like training the surgeon with more demanding requirements: she can only open a door if all twenty characters of the address match, with no tolerance for close-enough approximations. Her error rate drops dramatically. She is slower, slightly more cautious, but far more trustworthy in a clinical setting where a wrong door could cost a life.
The deepest lesson of this cluster: off-target effects are not a reason to abandon CRISPR. They are a reason to measure everything, design carefully, use the right tools for the right application, and build an exhaustive audit trail before you put the surgeon to work in a patient. The approved CRISPR therapy Casgevy passed every safety assessment. Future CRISPR medicines will too — not because they are perfect, but because we have learned how to prove they are safe enough.
References & Further Reading
- Tsai et al. (2015) — GUIDE-seq enables genome-wide profiling of off-target cleavage by CRISPR-Cas nucleases. Nature Biotechnology 33:187. — The original GUIDE-seq paper.
- Yarrington et al. (2018) — Nucleosomes inhibit target cleavage by CRISPR-Cas9 in vivo. PNAS. — Chromatin effects on off-target cutting rates.
- Kosicki et al. (2018) — Repair of double-strand breaks induced by CRISPR-Cas9 leads to large deletions and complex rearrangements. Nature Biotechnology 36:765. — The large deletion finding that shocked the field.
- Haapaniemi et al. (2018) — CRISPR-Cas9 genome editing induces a p53-mediated DNA damage response. Nature Medicine 24:927. — One of the two simultaneous p53 papers.
- Kleinstiver et al. (2016) — High-fidelity CRISPR-Cas9 nucleases with no detectable genome-wide off-target effects. Nature 529:490. — SpCas9-HF1 engineering paper.
- Slaymaker et al. (2016) — Rationally engineered Cas9 nucleases with improved specificity. Science 351:84. — eSpCas9 paper.
- Chen et al. (2017) — Enhanced proofreading governs CRISPR-Cas9 targeting accuracy. Nature 550:407. — HypaCas9 paper and the conformational checkpoint mechanism.
- Off-target effects exist but are manageable. They range from clinically inconsequential to potentially serious, depending on where they occur. The question is not whether they happen, but whether they happen at clinically relevant frequencies in clinically relevant locations.
- Position and number of mismatches determine off-target risk. Seed region mismatches (nt 1–12) almost always prevent cutting. PAM-distal mismatches (nt 13–20) are well tolerated. G:T wobble pairs are disproportionately common at off-target sites.
- Use multiple complementary detection methods. GUIDE-seq for comprehensive in-cell profiling. CIRCLE-seq for ultra-sensitive in vitro screening. Whole-genome sequencing for clinical-grade safety assessment.
- Standard assays miss large deletions and translocations. Amplicon sequencing only detects small indels. Structural variant analysis by long-read sequencing or optical genome mapping is required to detect kilobase-scale deletions.
- The p53 response selects for pre-cancerous cells. DSB-induced p53 activation eliminates efficiently edited cells in some contexts, enriching for p53-mutant survivors. Base and prime editing avoid this because they do not create DSBs.
- High-fidelity Cas9 variants dramatically reduce off-target activity. SpCas9-HF1, eSpCas9, HypaCas9 achieve near-wild-type on-target efficiency with 10–100-fold reduction in off-target cutting. These are standard in therapeutic programmes.
- The safest approach avoids DSBs entirely. Base editing and prime editing create nicks rather than DSBs, eliminating the primary sources of large deletions, translocations, and p53 selection. For therapeutic applications, these are increasingly preferred where applicable.