Off-target effects are unintended edits that a gene editor makes at the wrong place in the genome, and researchers measure them by combining computer prediction, test-tube cleavage assays, and sequencing of treated cells. No single method finds everything. That is why regulators ask developers to use several, and why a clean result in one assay is a starting point rather than a verdict.

The question sounds simple: did the editor change anything it was not supposed to? Answering it turns out to be a hard measurement problem, because the genome has about three billion letters and a rare event can hide in that haystack.

Two kinds of unwanted edit

The term “off-target” usually means an edit at a different site that resembles the intended address. Cas9 can tolerate a few mismatches between its guide and the DNA, so look-alike sequences elsewhere sometimes get cut.

A second category is less obvious. Even at the intended site, a cut can go wrong. A 2018 study in Nature Biotechnology by Kosicki, Tomberg and Bradley found that repair of Cas9 double-strand breaks often produced large deletions spanning many kilobases and complex rearrangements, in mouse stem and blood progenitor cells and in a human cell line. Standard short-range checks had missed them. The authors found them with long-read sequencing and long-range PCR (Nature Biotechnology, 2018). The tools below mostly look for the first kind of error, and the second kind needs its own tests.

Layer one: predict the likely sites by computer

Software can scan a genome for sequences that resemble a guide RNA and rank them by how many mismatches they carry. It is quick and cheap, and it is a necessary first filter when designing a guide. It is also incomplete. In the original GUIDE-seq paper, most of the off-target sites the assay detected had been missed by computational prediction (Tsai et al., Nature Biotechnology, 2015).

Layer two: cut naked DNA in a tube

Biochemical assays mix purified genomic DNA with the editor and then sequence what gets cut. They are sensitive, because no cell machinery interferes, and they can be run in large batches. Three are worth knowing.

  • Digenome-seq (2015). The Jin-Soo Kim group digested genomic DNA with Cas9 and guide RNAs in vitro, sequenced the whole genome, and located cut sites from reads that shared the same ends. They reported detecting insertions or deletions at off-target sites below 0.1 percent frequency (Kim et al., Nature Methods, 2015).
  • CIRCLE-seq (2017). Genomic DNA is sheared and circularized so that only molecules the editor linearizes are sequenced. The authors reported higher sensitivity than existing cell-based or biochemical methods, no need for a reference genome, and the ability to build a personalized profile that reflects one person’s variants (Tsai et al., Nature Methods, 2017).
  • CHANGE-seq (2020). A scalable, tagmentation-based refinement. Applied to 110 guide targets across 13 therapeutically relevant loci in human T cells, it identified 201,934 off-target sites. The study also found that, in cells, off-target activity was two to four times more likely near active promoters, enhancers and transcribed regions, and that human sequence variation changed activity at roughly 15 percent of the sites analyzed (Lazzarotto et al., Nature Biotechnology, 2020).

The catch is that these tubes over-report. DNA that is tightly packed or silent in a real cell may be cut readily in a dish. Many of the sites they flag never show editing in living cells, so each hit must be checked.

Layer three: look in cells

Cell-based assays ask what actually happens inside a living cell, including the effects of chromatin packaging and DNA repair. GUIDE-seq, introduced in 2014 and published in 2015, lets double-stranded oligonucleotides get captured into breaks, which tags the cleavage sites for sequencing. Applied to 13 guide RNAs in two human cell lines, it showed that off-target activity varied widely between guides. It also revealed breakpoint hotspots independent of the editor, and found that truncated guide RNAs gave far fewer off-target breaks (Nature Biotechnology, 2015).

The final step is always targeted sequencing. After a candidate list is assembled from the layers above, researchers sequence those exact regions deeply in treated cells and count how often they changed. Software such as CRISPResso2, described in 2019, was built to analyze this kind of amplicon sequencing and added support for base editor data (Clement et al., Nature Biotechnology, 2019).

Comparison at a glance

Approach Where it runs Strength Weakness
Computer prediction Software Fast, cheap, useful for guide design Missed many real sites in early comparisons
Digenome-seq, CIRCLE-seq, CHANGE-seq Purified DNA in a tube Sensitive, scalable, can reflect individual genomes Can flag sites that never change in cells
GUIDE-seq Living cells Captures real cellular context Needs efficient tag uptake into cells
Targeted deep sequencing Treated cells or tissue Gold standard for confirming candidate sites Only checks sites you already suspect
Long-read and long-range assays Treated cells Reveals large deletions and rearrangements Less routine, harder to quantify

What regulators ask for

The FDA’s January 2024 final guidance, Human Gene Therapy Products Incorporating Human Genome Editing, recommends combining methods such as in silico, biochemical and cell-based assays, and says genome-wide analysis helps reduce bias in finding candidate sites. It asks sponsors to verify sites with methods sensitive enough to detect low-frequency events, and to use relevant human cell types, ideally from several donors. For products given inside the body, the cell types should include the major ones where editing is detected. The guidance also flags chromosomal rearrangements as a consequence of DNA cleavage, calls for monitoring of off-target editing in trials, and recommends long-term follow-up for up to 15 years after administration (FDA guidance).

Nothing in that list is a numerical threshold. How much off-target editing is acceptable depends on where it happens and what that site does, and that judgment is made case by case.

How designers reduce the risk before measuring it

Measurement is only half the work. The other half is choosing parts that make errors less likely in the first place, and the literature above contains several concrete examples.

  • Trimmed guides. The GUIDE-seq authors reported that truncated guide RNAs produced substantially fewer off-target breaks.
  • Modified guides. The Digenome-seq team found that replacing a promiscuous guide with a version carrying two extra guanines reduced off-target effects.
  • Short exposure. The FDA guidance tells sponsors to minimize how long nuclease activity lasts and to optimize guide components. Delivering the editor as short-lived mRNA instead of a permanent gene follows that logic.
  • Machine learning. The CHANGE-seq dataset was used to train a model that predicts off-target activity, which could improve the first computational filter.

Myth versus fact

Myth: if no off-target site shows up, there are none. Fact: every assay has a detection limit. The Digenome-seq authors, for instance, reported being able to see insertions or deletions below 0.1 percent frequency, and that is a good number, but it is a limit nonetheless.

Myth: a cell-line result carries over to patients. Fact: sequence variation between people can change which sites matter. CIRCLE-seq was pitched partly for its ability to build a personalized profile for that reason, and CHANGE-seq found that human variation affected activity at a meaningful fraction of sites.

Myth: only the tiny, scattered sites matter. Fact: where an unintended edit lands matters more than how many there are. A change in a stretch of DNA with no known function is a different matter from one inside a gene that regulates cell growth.

Do base editors and prime editors fix the problem?

They remove one source of damage. Base editors generate few insertions and deletions, typically at or below 1 percent in the original cell-line experiments (Broad Institute record of Komor et al.), and the prime editing paper reported much lower off-target editing than Cas9 nuclease at known problem sites (Anzalone et al., 2019). Neither removes the need to look. Different editors fail in different ways, so assays built for Cas9 cuts need adapting. Editors that make no break leave nothing for a break-capture method to tag, which is part of why CRISPResso2 added base-editor analysis.

A short guide to reading an off-target claim

  1. Which methods were used, and was there more than one layer?
  2. Were the cells relevant to the disease, and were multiple donors tested?
  3. Did the team check for large deletions and rearrangements as well as small edits?
  4. How deep was the sequencing, and what was the detection limit?
  5. Is there follow-up in treated animals or patients, rather than only in dishes?

This is general scientific background, not medical advice.

Sources