Off-target checks and their limits

An off-target result is only as wide as the method that looked. Most methods look at a slice and are then described as if they had looked at the genome. This page sorts the slices: computational lists, cell-based and cell-free nomination, amplicon sequencing of a named set, and whole-genome sequencing with a real floor. How a guide becomes a hypothesis in the first place is in guide RNA design is a hypothesis. How the on-target allele is called is in checking whether a genome edit worked. The nuclease context is how CRISPR-Cas9 editing works in research.

Primers, library reagents and protein preparations are classes in the molecular biology catalogue. The assay you mean should be named on the quote request.

A list in a genome browser is a hypothesis about where to look

In silico tools search a reference for sites that resemble the spacer and sit next to a PAM the enzyme might accept. CRISPOR and the UCSC Genome Browser are public places to start. The search is only as good as the assembly, the mismatch allowance, the treatment of DNA or RNA bulges, and the PAM set. SpCas9 is not only an NGG enzyme in this context. Weaker PAMs such as NAG belong in the search if you intend to claim the list was thorough, or you must say you ignored them.

Scores then rank the list. They do not measure occupancy or cutting. Chromatin, DNA methylation and the dose and duration of the nuclease change which of those sites are cut in a given nucleus. A site absent from the list can still be cut if the cells carry a haplotype the reference lacks, or if the tool could not see a bulge. Use the list to nominate amplicons. Quote it as a nomination.

Running the search on the wrong build is the avoidable failure. A guide designed on one assembly and audited on another will confess a different off-target set. Write both identifiers down. If the parent line has been sequenced, search that sequence where it matters, not only the species reference.

Nomination methods that do not start from a mismatch table

GUIDE-seq is a cell-based class. A short double-stranded DNA tag is introduced with the nuclease and can be captured at breaks during end joining. Sequencing from the tag nominates sites that took the tag. The on-target site should appear if the experiment worked. If it does not, the map is not evidence of specificity. It is evidence that the tag was not integrated where you know a break occurred.

The limits follow from that chemistry. Cells that will not take up the oligo cannot be assayed this way. Breaks that are resealed without capturing the tag are invisible. There is a detection floor. Rare clones below it are not cleared by a blank column. The original method literature discusses sensitivity on the order of a tenth of one percent of alleles in favourable cases. That figure is a historical performance of a method, not a promise that your run reached it. Report what you did.

CIRCLE-seq is a cell-free class. Genomic DNA is circularised and cut in vitro with the ribonucleoprotein. Cleavage sites are sequenced. Sensitivity for biochemical cutting is high, and chromatin is absent, so the list over-calls sites the cell may never open. Related cell-free maps exist under other names. They are nomination tools. A CIRCLE-seq peak is a reason to sequence that amplicon in the edited cells, not a reason to write that the site was edited in the experiment.

Neither method replaces the other. A site found only in naked DNA may be irrelevant in the cells. A site found only by a tag may be missed in vitro if the circularisation or the coverage failed. When a biological claim depends on one site, confirm it in the same cells, with the same delivery, that you intend to publish.

Counting molecules at sites you already chose

Amplicon sequencing reads a PCR product from each nominated site deeply enough to estimate allele fractions. The nomination can mix the computational list, the GUIDE-seq hits and the CIRCLE-seq hits, trimmed to what you can afford. Everything off the list is unmeasured. Say that in the same paragraph as the result.

The error floor is set by the polymerase, the number of cycles, chimeric reads and the sequencer. Rare indels near that floor appear in unedited DNA. Sequence the unedited parent with the same primers and the same pipeline, and set the threshold before you look at the edited sample. Unique molecular identifiers are one way to push the floor down. They are a method choice, not a default. The background on short-read sequencing as a class is in next-generation sequencing from library to reads, and the practical run checks are in what to check before a sequencing run. A public overview of the instrument class is maintained by Illumina.

A site that matches the parent, within your threshold, supports "no edit detected here". It does not support "this guide does not cut". A site with a one-percent indel that is absent from the parent supports an off-target edit at that frequency in that sample. It does not tell you the frequency in a different delivery format. Plasmid expression and a ribonucleoprotein can rank off-target sites differently because the exposure differs.

A genome sequence, and the edits it is too shallow to see

Whole-genome sequencing of a clone can find unexpected indels, substitutions and some structural changes without a nomination list. It is the broadest screen in this set, and it is still a screen with filters. Coverage that is comfortable for finding a heterozygous variant in a pure clone will miss an edit present in one percent of cells in a pool. Mosaic tissue and bulk cultures are the wrong object for a casual genome sequence if the question was rare off-target events.

Small-indel callers and structural-variant callers are different analyses. A pipeline tuned to one-base insertions will not hand you a 20-kilobase deletion as a tidy row. Large deletions at the on-target site itself are a known way for PCR genotyping to look wild type, and the same class of event off-target is easy to under-call. If you did not run a structural analysis, do not imply one.

Reference bias remains. Reads that do not match the assembly are discarded by some settings and hide the very haplotypes where an off-target site might exist. Depth, the caller, the clone identity and the unedited parent genome, sequenced similarly, are the minimum report. Depositing the reads in a public archive such as the NCBI Sequence Read Archive is a data decision for the project. It is not a quality certificate.

Over-calling, under-calling, and the control that keeps them apart

Over-calling is what cell-free maps and error-prone amplicons do when the unedited control is missing. Under-calling is what a strict computational filter, a failed tag delivery, or a shallow genome does when you describe the blank as safety. The discipline is to show the positive control for the method: the on-target site in GUIDE-seq, a known cut in the CIRCLE-seq library, the parental amplicon beside the edited one, and a depth metric for the genome.

Phenotypes still need the on-target allele. An off-target list that is empty does not prove a phenotype is on-target. An off-target list that is full does not prove the phenotype is off-target. Two guides and a rescue experiment do that work. This assay tells you about cuts.

If a nominated site shows an indel, decide whether it hits a gene that confounds the phenotype before you expand the clone. You may still use the clone if you report the passenger. Pretending the site is a non-coding inconvenience because the phenotype is convenient is how panels fail to replicate.

Data, human sequences and institutional rules

Off-target sequencing of human cell lines can be human genomic data even when the scientific question was a reagent. Storage, sharing and deposition follow the governance your institution already applies. This guide does not waive it, does not approve a clinical genome analysis, and does not describe editing of embryos or the germline. Finding an unintended edit in a hazardous gene is a reason to revisit the risk review, not a reason to relabel the guide as unacceptable by a blog standard. The committee that approved the work is the one that can revise it.

A method with a floor must not be written as a guarantee of absence. Reviewers will ask for the floor. Write it first.

Files, power and a specification that survives

Sequencing files are the record. A power cut or a dropped network link during a copy is how a run becomes a screenshot. Keep the sample sheet, the primer sequences and the parental control under names that map to the clone. Heat and humidity matter less to the bits than to the DNA tubes waiting to be libraries. Those tubes follow the same contamination rules as any amplicon. A shared bench of off-target primers will eventually amplify the wrong site and look like a hit.

What an enquiry should specify

Name the method class, the organism, the assembly, the guide sequence, whether the sample is a clone or a pool, and the allele fraction you need to see. For a nominated panel, include the coordinates. For a genome, include the depth you are asking about. The CRISPR validation sequencing enquiry reference is a place to frame that discussion. It is an enquiry reference, not an operated off-target service. Library reagents sit in the molecular biology pathway. Ask whether a quotation is possible. Do not ask for a clean guide. Ask for a measurement with a stated edge.