CRISPR Off-target Effects: What They Are and What Labs Check First
CRISPR off-target effects are Cas cuts at genomic sites that resemble the guide RNA but are not the intended target. They are mismatch-tolerant cleavage events, not a mysterious decimal on a design website. A score can rank nominated lookalikes in a chosen genome. It cannot certify that no other site will be cut. This page defines the effect, names why it happens, and lists what a lab checks before an oligo or plasmid order — without a clinical-safety claim.
What CRISPR Off-Target Effects Are
On-target editing is Cas activity at the site you designed the guide against. An off-target effect is the same chemistry somewhere else: the ribonucleoprotein binds a similar sequence, the nuclease cuts, and the cell repairs a break you did not ask for. Hsu et al. (2013) treated those unintended loci as a measurable specificity problem and evaluated predicted genomic off-target sites in human cells. The definition does not depend on a brand of software. If the cut is at the wrong coordinate, it is an off-target, whether or not a website painted that coordinate yellow.
This page stays with guide-dependent, sequence-similar cuts — the case most design tools search for. Other classes of unwanted change exist in the literature; they are not required to answer the first question.
Why Cas Tolerates Mismatches
SpCas9 does not require a perfect 20-nucleotide match. Hsu and colleagues showed that it tolerates mismatches between the guide RNA and target DNA in a sequence-dependent way: the number of mismatches matters, and so do their positions and how they are clustered. A mismatch near the PAM is not interchangeable with the same mismatch at the distal end. That is why "how many mismatches" is an incomplete risk sentence. Two distant mismatches can be less informative than one in the seed.
The operational consequence is simple. When you inspect a predicted off-target, read the mismatch map, not only the count. Tools that hide position are hiding the part of Hsu's result that changes the decision.
Computational Nomination Versus Experimental Validation
In silico tools search a genome for sites that look like the guide under a mismatch and PAM model, then attach a score. CRISPOR (Concordet and Haeussler, 2018) is a clear example: it finds guides in an input sequence, ranks them by off-target and on-target scores for a user-chosen genome, and lists potential off-targets with up to four mismatches. Haeussler et al. (2016) had already shown that sequence-based predictors can recover many experimentally validated sites when the search is deep enough — and that some websites missed sites because of implementation limits, not because those sites were undetectable in principle.
Nomination is still not validation. Tsai et al. (2015) developed GUIDE-seq to find Cas-induced double-strand breaks genome-wide without starting from a predicted list. Their abstract states the result labs should keep on the wall: the majority of identified sites were not detected by existing computational methods or by ChIP-seq. An open copy is on PMC4320685. The honest workflow is therefore sequential: compute a candidate set, then decide what experimental check the project's risk justifies. A quiet prediction list is not evidence that GUIDE-seq would also have been quiet.
What Labs Check Before Ordering a Guide
Before oligos or a guide plasmid go on a purchase order, the review is a short, sourcable list. It does not replace a validation assay for high-stakes work. It does stop the common error of ordering the first green-highlighted spacer.
| Check | Why it is first | What a score cannot do |
|---|---|---|
| Genome and nuclease | CRISPOR and similar tools only search the assembly and PAM you select | A human-genome score does not speak to a mouse or a cell line with a different reference |
| Mismatch depth | Haeussler and CRISPOR treat a several-mismatch search as the working nomination window | A tool that stops at one mismatch is not a complete lookalike list |
| PAM-proximal mismatches | Hsu: position and distribution change tolerance, not only count | A site with "only two mismatches" can still be the risk if both sit in the seed |
| Prediction versus validation | Tsai: most GUIDE-seq sites were missed by then-available predictors | A clean computational report is not an experimental negative |
Where On- and Off-Target Scores Fit — and Stop
Scores are useful when they are treated as ranks. They help a lab discard obvious bad guides and keep a shortlist. They do not prove the absence of off-targets, and they are not a clinical-safety result. After the concept is clear, one documented workspace example is enough: the Zettalab product page lists CRISPR guide-RNA design with on- and off-target scoring. That is the same job CRISPOR and commercial tools advertise — nomination, not clearance. If the remaining decision is which scoring environment the group will use, the CRISPOR versus commercial CRISPR software page handles that comparison. No page on this site, and no score in a workspace, should be read as "this guide has no off-targets."
Frequently Asked Questions
What is a CRISPR off-target effect in one sentence?
It is Cas cleavage at an unintended genomic site whose sequence resembles the guide. The chemistry is the same as the on-target cut; the coordinate is not.
Do off-target scores prove a guide will not cut elsewhere?
No. A score ranks nominated similar sites in the genome you searched. GUIDE-seq found real cuts that computational methods of that period missed. Treat a favorable score as a reason to keep a candidate, not as a negative genome-wide assay.
Why do some real off-target sites miss computational prediction?
Predictors look for sequence similarity under a mismatch-and-PAM model. Cellular cleavage also depends on factors those models leave out. Tsai and colleagues' unbiased DSB assay recovered a majority of sites that the computational methods then in use did not list.
What should a lab check before ordering a guide RNA?
Confirm the genome build and nuclease, inspect mismatch depth and PAM-proximal mismatches, and decide whether the project's risk requires experimental validation beyond the computational list. Ordering is a commitment of time and genotype, not a moment to trust a color code.
Related Articles