# What CRISPR Can and Can't Do Yet: Off-Target Edits, One Approved Medicine, Gene-Edited Crops, and the Embryo Line

Why "no off-targets detected" is a statement about the test, why Casgevy is the only CRISPR medicine on the FDA list, how the US and EU split gene-edited plants, and which ethical arguments survive better technology

> off-target effects · gene therapy · gene-edited crops · germline editing · regulation · About 12 min · Oct 6

## Key points

1. Off-target editing arises because a 20-base guide tolerates mismatches, so near-identical sites exist throughout a three-billion-letter genome; one large screen catalogued 201,934 off-target sites for 110 guides, though many are detected in cell-free assays and are cut rarely or never in living cells.
2. Assays that measure breaks inside living cells find fewer sites and show cutting concentrating near open chromatin (promoters, enhancers, transcribed regions); "no off-targets detected" describes the sensitivity of the test, not a property of the edit.
3. Off-target risk is contextual: an edit in a tumour suppressor is categorically worse than one in inert DNA, and reviewers argue for zero tolerance there; large deletions, rearrangements, and chromosome loss at the intended site are a separate hazard that mismatch-based screens may not report.
4. Mitigations include high-fidelity Cas9 variants, truncated guides, base and prime editing that avoid double-strand breaks, and testing the actual cell product — individual genetic variation changes activity at roughly 15% of off-target sites, so the profile depends partly on the patient.
5. Casgevy (exa-cel) is the only CRISPR-edited product on the FDA's approved cell and gene therapy list: blood stem cells are removed, the BCL11A erythroid enhancer is disrupted by end joining, and fetal hemoglobin is reactivated — a knockout of a control switch rather than a repair of the sickle mutation.
6. Its real-world limits are logistical and economic: busulfan conditioning causes infertility, eligibility is narrow (age 12+, two or more severe crises per year, fit for transplant, no matched donor), the list price is about £1.65 million in England, approval came with a five-year managed access scheme, and the pivotal trial was single-arm with 30 patients at the analysis point.
7. The binding constraint on new CRISPR medicines is delivery, not editing chemistry: cells outside the body are accessible, while most organs are not.
8. USDA's SECURE rule exempts plants whose edits could have come from conventional breeding (expanded in 2024 to up to twelve modifications); the EU's 2026 regulation splits gene-edited plants into category 1 (treated as conventional, excluded from organic farming) and category 2 (full GMO assessment), covering plants only.
9. Regulators moved toward "what changed" partly because edited and conventionally bred plants can be analytically indistinguishable; "treated as conventional" is a claim about familiarity and detectability, not a blanket safety verdict, and changes in food composition still require assessment.
10. Germline editing changes every cell of the person born and passes to their children; technical obstacles (mosaicism, inability to verify a whole embryo, HDR not occurring as hoped, large deletions and chromosome loss in edited embryos) are reasons not to do it now, and would weaken as the technology improves.
11. The arguments that survive technical improvement are consent (the edited person and their descendants cannot agree), bounded harm (somatic risk ends with one lifetime; germline risk does not), and definition (whose list of "defects" counts).
12. Policy positions divide into a moratorium (not until questions are answered, implying research and registries) and a prohibition (never, implying enforcement and the risk of pushing work to unregulated places); there is no global enforcement body, and the Oviedo Convention binds only its signatories.

---

The last two articles took you inside the cell. One followed Cas9 as it read a 20-letter guide next to a short PAM and cut both strands; the other showed that the cut itself decides nothing — the cell's own repair machinery decides whether the break becomes a destroyed gene or a precise correction. That is the mechanism. This article is about what happens when the mechanism leaves the lab.

Four questions organize it. How precise is CRISPR really, and how would we know? What has actually been approved as medicine? What is it doing in farming? And where is the line drawn, and why, when the cells being edited belong to an embryo?

## Off-target effects: a guide is a proposal, not a promise

A guide RNA carries about 20 bases of targeting information. Pairing does not have to be perfect: Cas9 tolerates mismatches, especially in the half of the guide farthest from the PAM. Against a genome of about three billion base pairs, a 20-letter pattern almost always has relatives — sites that match at 17, 18, or 19 of 20 positions. Any of those is a potential cut.

How many such sites exist, and how often they are actually cut, are two different questions. A useful illustration of the first: a large screen of 110 guides aimed at 13 therapeutically relevant loci catalogued 201,934 off-target sites in human genomic DNA [2]. That number is shocking until you notice what kind of number it is. Some detection methods work on purified DNA, outside any cell, where chromatin cannot shield a site and repair cannot influence the outcome; they are extremely sensitive and generally overestimate what happens in a living cell. Cell-based methods such as GUIDE-seq, which tag the actual breaks that occur inside living cells, typically nominate far fewer sites, and cutting concentrates where the genome is open — near active promoters, enhancers, and transcribed genes, two to four times more than chance would predict [2][3].

Two consequences follow, and both matter more than any single number.

First, "we found no off-target edits" is a statement about the sensitivity of the test, not a property of the edit. Every assay has a floor below which it is blind, different methods disagree about which sites they can see, and the FDA's 2024 guidance therefore asks developers to use more than one method, including a genome-wide one [4]. The honest version of the claim is: no off-target editing was detected by these assays at these detection limits, in these cells.

Second, off-target risk is not one thing. Whether a stray edit matters depends on where it lands and who is being treated. A change in the middle of an untranscribed region may be irrelevant; an edit that disables a tumour suppressor is a different category of event, and reviewers of this field have argued for effectively zero tolerance for off-target editing in tumour suppressors and oncogenes, regardless of how rare it is predicted to be [5]. Risk tolerance also depends on the alternative: a serious off-target risk that would be unacceptable in a cosmetic application can be acceptable in a disease that shortens life and causes repeated hospitalizations.

There is also a category of damage that is not "off-target" at all, and is easy to miss: the same break that is used correctly can end in a large deletion, a rearrangement, or the loss of part of a chromosome. These look nothing like a single mismatch-based cut, so assays designed to hunt for near-matching sequences may not report them [5].

The mitigations are concrete. Choose guides with few plausible near-matches. Use high-fidelity Cas9 variants or shortened guides — the original GUIDE-seq work found that truncated guides substantially reduced off-target breaks [1]. Prefer tools that avoid double-strand breaks entirely: base editing and prime editing, which change letters without cutting both strands. And in therapies where cells can be removed, edited, and examined before they go back into the patient, test the actual product. That last point is not a formality: single-letter differences between people's genomes create or destroy off-target sites, and in one large screen individual genetic variation altered activity at roughly 15% of the off-target sites analysed [3]. The off-target profile of a guide is partly a property of the patient.

## What is actually approved

One CRISPR medicine has cleared regulators: exagamglogene autotemcel, sold as Casgevy, for sickle cell disease and transfusion-dependent beta-thalassemia [6]. The UK licensed it in November 2023, the FDA followed for sickle cell disease, and by early 2025 it was on the same approved-product list as the older cell and gene therapies — where, notably, it is the only CRISPR-edited product [6][7]. The recommended list price in England is £1,651,000 per course, and roughly CAD 2.8 million in Canada [6][8].

Read the design closely, because it uses exactly the logic from the previous article. The therapy does not repair the sickle mutation. It takes the patient's own blood-forming stem cells out of the body, makes a cut in a regulatory switch — the erythroid enhancer of a gene called BCL11A — and lets end joining scramble it. With that switch broken, BCL11A no longer suppresses fetal hemoglobin production in the developing red cells, and the red cells the patient makes afterwards are full of fetal hemoglobin, which does not sickle. In other words: an indel-based knockout of a control element, in a cell type you can extract, edit, and return. That is the easiest possible configuration of CRISPR medicine, and it works well enough to be curative in many treated patients.

Now the part that explains why there is only one. Before the edited cells can be returned, the patient's existing bone marrow must be cleared with busulfan, a chemotherapy that causes infertility and requires weeks in a transplant-capable hospital. The cells must be collected by apheresis, shipped, manufactured, and tested. Eligibility is narrow: age 12 and over, specific sickle genotypes, at least two severe pain crises per year, fit enough for a transplant, and no matched donor available [9]. NICE initially declined to recommend it on cost-effectiveness grounds before approving it in January 2025 under a five-year managed access scheme with a limited number of treatment centres [9][6]. And the pivotal evidence came from a single-arm study — no placebo group — in which 63 patients were enrolled and 30 had reached the analysis point [8]. That is normal for a rare disease with a dramatic endpoint, and it is also why regulators want long-term follow-up; the FDA recommends 15 years of it, which is expensive and hard to sustain when a one-time treatment leaves patients with little reason to keep returning [5].

Compare Casgevy with Lyfgenia, approved for the same disease around the same time and also on the approved list: it uses a lentivirus to add a working globin gene rather than CRISPR to break a switch [7]. The relevant competition for a CRISPR therapy is often another gene therapy, not a pill.

What limits the field now is delivery. Editing cells outside the body is straightforward because the cells are accessible. Editing inside the body means getting the editor into the right organ at the right dose — feasible for the liver, which soaks up lipid nanoparticles from the blood, much harder for muscle, brain, or lung. That bottleneck, not the elegance of the editing chemistry, is what stands between the current situation and a broad set of CRISPR medicines.

## Agriculture: the same molecule, a different question

In a field, CRISPR raises a question that has little to do with molecular biology: what counts as a new organism, and who gets to decide?

The American answer is product-based. Under a rule finalized in 2020 and fully implemented in 2021, the USDA exempts plants whose modifications could have been produced by conventional breeding, and it expanded those exemptions in November 2024 to allow plants with up to twelve such modifications [10]. Crops built by edits that mimic what breeding could have done may be grown and sold without passing through the genetically-modified-organism pathway.

The European answer is being rewritten. After a 2018 Court of Justice ruling put gene-edited plants under GMO law, the Commission concluded that framework was unworkable, and in June 2026 the Parliament and Council adopted a new regulation with two categories. Category 1 covers plants whose changes could have arisen naturally or through conventional breeding: after a verification procedure they are treated like conventional plants, with seed labelling and a public database, and they are not permitted in organic farming. Category 2 covers more complex modifications and stays under GMO rules, with risk assessment, authorization, traceability, and labelling. The new rules cover plants only, and only targeted mutagenesis and cisgenesis; edited animals and microorganisms remain under the old GMO framework [11].

The reason for the split is practical, and worth remembering: in many cases, once the edit is made, no analytical method can tell whether a plant was gene-edited or conventionally bred — the DNA sequences are identical [11]. A rule keyed to *how* an organism was produced becomes partly unenforceable, so regulators shift toward asking *what* changed. Note what this does and does not mean. "Treated like conventional breeding" is a statement about familiarity and detectability, not a blanket safety verdict; the EU keeps novel-food assessment for category 1 plants whose composition changes enough to affect nutrition or introduce undesirable substances [11]. A purpose-built edit that disables a gene can still have unexpected effects, and a change in composition still has to be evaluated as food.

The applications themselves are unglamorous and mostly about lost harvests and shelf life: resistance to diseases that wipe out crops, produce that browns less quickly, oils with a healthier fatty-acid profile, and removing compounds that make food bitter. The first gene-edited food sold in the US was an edited soybean oil, launched in 2019. Regulation, however, is not the only gate: a product that is legal but that nobody wants to buy, or that no exporter will take, does not make it to a plate.

## Embryos: where the line is, and how to argue about it

Everything so far involved **somatic** cells — cells of a patient's body. Changes there end with that patient. **Germline** editing means editing an embryo, an egg, a sperm, or the cells that produce them. Because every cell of the resulting person descends from the edited cell, the change is in all of them, including their eggs or sperm, and it passes to their children.

```mermaid
flowchart LR
  E["The same CRISPR edit"] --> S["Somatic cells or a patient's own blood stem cells"]
  S --> S1["Changes stay in that one person"]
  E --> G["Embryo, egg, sperm, or their precursor cells"]
  G --> G1["Present in every cell of the child born"]
  G1 --> G2["Passed on to their children"]
```

That difference in scope creates a difference in kind, not just in degree. An error in a somatic therapy is a medical event for one consenting patient. An error in the germline is inherited by people who never agreed to anything.

The technical case against doing it now is stronger than the phrase "designer babies" suggests, and it comes from the repair biology covered earlier. Editing an embryo has to happen at the one-cell stage if every cell is to carry the change; if the edit lands after the first division, some cells carry it and others do not — **mosaicism** — and a biopsy of a few cells from a later embryo cannot reliably tell you which. You cannot sequence an entire embryo without destroying it, so the thorough verification possible for a manufactured cell product is simply unavailable here. And the repair route researchers would want, homology-directed repair with a supplied template, turns out to be the route human embryos largely do not use: in the best-known human-embryo editing study, HDR was not observed at the target locus and most of the embryos were mosaic. Follow-up work in human embryos reported large deletions and loss of chromosome material at the cut site — exactly the class of damage that sequence-based off-target screens are least suited to catch. "Just fix the mutation before implantation" is a research question, not a solved procedure.

The ethical case is usually argued badly, by piling discomfort on top of biology. It is more useful to separate the questions that any policy has to answer.

**Is the goal therapy or enhancement?** Correcting the gene behind a fatal childhood disease and choosing a child's height or eye colour are different requests, but the boundary between them is not a clean line. Traits that are medically "defects" are often ordinary human variation to the people who have them; many deaf people argue that deafness is not a defect to be eliminated. A list of conditions worth editing is a social document as much as a medical one.

**Who consents?** The edited person cannot, and neither can their descendants, who inherit both the benefit and any error. Somatic medicine has an answer for this — informed consent from the person bearing the risk. Germline editing does not.

**What happens if we are wrong?** Somatic risk is bounded by a lifetime. A germline error can spread through a family line, at a rate and in a direction no one can track.

**Who decides?** There is no global enforcement body. National laws differ: many countries prohibit heritable editing, others have no explicit rule, and research on embryos in the lab is often permitted where transferring an edited embryo to a womb is not, typically bounded by the long-standing limit of around 14 days of development. The Council of Europe's Oviedo Convention is the closest thing to a binding international agreement against heritable modification, and it binds only its signatories. Scientific societies and WHO advisory bodies have repeatedly called for a halt to reproductive heritable editing — advisory, not enforceable. The concern behind those calls is not hypothetical: unauthorized embryo editing has already led to pregnancies, and no framework prevented it in advance.

One distinction is worth carrying out of this section, because it decides what policy should look like. A **moratorium** says: not until the technical and social questions have answers. A **prohibition** says: never, whatever the technology becomes. A moratorium implies funding safety research, building registries, and defining what evidence would count; a prohibition implies enforcement, criminal liability, and accepting that it may push the work to places with no rules at all.

And notice something about the ethics arguments themselves. Off-target risk, mosaicism, and embryo repair failures are all reasons not to do this *now*. As the technology improves, those reasons weaken. The arguments that survive technical improvement are consent, harm to people who cannot object, and who gets to decide what counts as a defect. If your objection to embryo editing rests only on "it isn't safe yet," you have signed up for a debate you will eventually lose.

## The pattern

Look back at the four domains and the limits turn out to be different in kind. Specificity is a biological limit, measured with instruments that have detection floors — so the honest claim is always narrower than "it's precise." Medicine is limited by delivery, manufacturing, evidence, and price: the editing was never the hardest part. Agriculture is limited by how a society defines a new organism, which is why the same edit that is exempt in one country requires a full GMO authorization in another. Embryo editing is limited by whose consent is at stake, which no improvement in the enzyme can address.

When you read a claim about CRISPR next — a cure, a superfood, a designer baby — the first useful question is which of these four limits the claim is quietly ignoring.

## Sources

1. [Tsai et al., Nature Biotechnology — GUIDE-seq, genome-wide profiling of off-target cleavage, and truncated guides reducing off-target breaks](https://www.nature.com/articles/nbt.3117)
2. [Lazzarotto et al., Nature Biotechnology — CHANGE-seq: 201,934 off-target sites across 110 guides, chromatin effects, and ~15.2% of sites affected by individual genetic variation](https://www.nature.com/articles/s41587-020-0555-7)
3. [Nature Protocols — GUIDE-seq protocol, comparing sensitive cell-based detection with cell-free methods](https://www.nature.com/articles/s41596-021-00626-x)
4. [Off-target assay guide — in silico, biochemical, cellular and in situ methods, and the FDA 2024 recommendation to use multiple methods including genome-wide analysis](https://seqwell.com/wp-content/uploads/2025/11/seqWell-Off_Target_Assay_Guide.pdf)
5. [Stewart et al., Nature (2025), author-hosted copy — detection limits, on-target large deletions and rearrangements, zero tolerance at tumour suppressors, and the 15-year follow-up recommendation](https://waltersport.com/wp-content/uploads/2025/12/NATURE-Measurement-and-clinical-interpretation-of-CRISPR-off-targets-Stewart-et-al.-2025.pdf)
6. [The Pharmaceutical Journal — UK licensing of exa-cel, the January 2025 NICE recommendation, managed access, and the £1,651,000 list price](https://pharmaceutical-journal.com/article/news/nice-approves-gene-editing-sickle-cell-therapy-for-nhs-in-england)
7. [FDA — approved cellular and gene therapy products, including Casgevy and Lyfgenia](https://www.fda.gov/vaccines-blood-biologics/cellular-gene-therapy-products/approved-cellular-and-gene-therapy-products)
8. [Canada's Drug Agency — Casgevy review: the single-arm CLIMB-121 study (63 enrolled, 30 analysed) and the ~CAD 2.8 million cost per administration](https://www.cda-amc.ca/sites/default/files/DRR/2025/SG0830_Casgevy_FINAL_Recommendation.pdf)
9. [Sickle Cell Society — Casgevy eligibility criteria, the five-year managed access scheme, and the initial treatment centres](https://www.sicklecellsociety.org/wp-content/uploads/2025/01/Casgevy-exa-cel-Frequently-Asked-Questions-March-2025-FINAL.pdf)
10. [Congressional Research Service — gene-edited plants: the SECURE rule, its 2021 implementation, and the November 2024 expansion of exemptions](https://www.everycrsreport.com/reports/IF12618.html)
11. [European Commission — the new genomic techniques framework adopted in June 2026, with the category 1 / category 2 split and the organic-farming exclusion](https://food.ec.europa.eu/plants/new-genomic-techniques_en)
12. [Regulation (EU) 2026/1388 — full text, including the finding that edited and conventionally bred products with identical DNA cannot be distinguished analytically](https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=OJ%3AL_202601388)

---

Original article: https://eulore.ai/articles/crispr-reality-limits-therapies-crops-embryos-6bc82cc5

> **Eulore** · Learn a little. Understand a lot.
>
> Eulore is an AI learning tool that turns what you want to learn into a continuing series. Share a topic, and it gets to know your starting point before creating articles you can read in 5–10 minutes. Ask as you read, and shape what comes next.This article was created in the same way.
>
> Start your own series → https://eulore.ai