UNCOVERseq
Which places in the genome a CRISPR nuclease really cut, each one tested against a no-guide control, ranked by how strong it is, and tiered into the panel worth confirming.
What the assay does
A blunt double-stranded tag goes into living cells with the nuclease, and non-homologous end joining captures it at double-strand breaks. A tag-specific amplification then reads outwards from every captured tag, so the molecules pile up where the nuclease cut: the intended site, and the off-target sites no prediction found. That much is GUIDE-seq. UNCOVERseq (Kinney et al., Nature Communications 2026) rebuilds everything around it.
The genomic DNA is fragmented enzymatically and adapter-ligated, then amplified in a single reaction with a blocking oligo against each strand of the tag, which takes the tag-to-adapter artefact from between 37% and 67% of reads down to about half a percent. Reads are collapsed to molecules by their molecular identifier. The guide's spacer is then placed in each site's window by an alignment that allows a bulge, and every site in every treated library is tested against the pooled no-guide controls. Measured against deep amplicon confirmation of two benchmark guides, that reached 97.6% analytical sensitivity at 78% precision, where GUIDE-seq reached 35.9% at 94%.
Why it is a kind of its own, and not GUIDE-seq with a wider net
Three things differ, and each changes what comes out.
It tests rather than subtracts. GUIDE-seq marks a site its mock library also holds and leaves the subtraction to the reader. Here each site's share of a treated library is compared with its share of the pooled control by a one-sided hypergeometric test, corrected across that library's own sites. A site can then be background at 500 control molecules and real at three, which no fixed window or floor can express: a control library holds tag integrations too, at the breaks a cell makes on its own and wherever the library preparation leaves an artefact, and those are not evenly spread.
It matches on an edit distance, with a gap allowed. The spacer is placed by a glocal Needleman-Wunsch alignment at the published scoring, and a site is kept if its Levenshtein distance to the spacer is under seven, counting a substitution and a one-base bulge alike. The PAM is read off the bases beside the match and reported, never scored and never a gate. On 48 guides this qualified a median of 30% more sites than the regular expression the current GUIDE-seq pipeline uses and 150% more than its historical alignment, which forbade a gap outright; sites found only this way were later confirmed as edited at frequencies up to 68%.
It counts replicates. A site nominated in every replicate is, on this method's own confirmation data, almost certainly real. Run as a single replicate, 30% to 40% of the sites worth confirming are missed or mis-ranked, and those are not the strong ones, which any replicate finds.
What a nomination rests on
A site is nominated when all three of the published criteria hold: a significant adjusted P value against the pooled control in at least one library, sequencing evidence from both sides of the cut in at least one library, and an edit distance to the spacer under seven. A site nominated in one replicate of three is still nominated; how reproducible it was decides its tier, not its nomination.
One number baSeq adds is the detection floor, and it explains more disappointing runs than anything else on the quality page. The test draws its power from the control's depth, so a control sequenced much more shallowly than its treatment makes every test powerless: the chance of all of a site's molecules landing in the treated library is high by construction, and nothing reaches significance. Library Quality turns the two depths into the fewest molecules a site must hold to be able to be nominated at all. At matched depth that is a handful of molecules. At a hundredth of the depth it is over a thousand, and a run like that nominates the on-target site and nothing else while looking, in every other respect, like a clean experiment.
Severity bins, and what they mean
A site's severity bin is its molecules as a share of its own guide's on-target site, in its own group of replicates: over a half, a tenth to a half, a hundredth to a tenth, a two-hundredth to a hundredth, and under a two-hundredth. The bins are worth knowing because the method's own confirmation says what each one turned out to mean. Of a 60-site panel taken to deep amplicon sequencing, every site in bins 1 to 3 came back edited; 30% of bin 4 and 72.3% of bin 5 could not be confirmed down to 0.01% indels, which is at least as likely to be the confirmation assay running out of sensitivity as the nomination being wrong, since the nomination frequencies track the confirmed ones closely.
A site with no bin is not a weak site. It is a site with nothing to compare against, because its guide's own on-target site was never found. Those are counted and named rather than swept into bin 5: "under half a percent of the on-target" and "we never found the on-target" are different statements, and only one of them is a measurement.
The tiers are baSeq's own
Tiering answers a practical question: which sites are worth a confirmation panel's depth? It weighs the three things the method names, which are frequency, reproducibility and genomic consequence. Tier 1 is a site nominated in every replicate and over a hundredth of the on-target. Tier 2 is one nominated in every replicate but weaker, or reproduced and strong, or in a coding exon, where even a rare cut changes a protein. Tier 3 is nominated once and strong, or reproduced and weak. Tier 4 is nominated once, weak, and not in coding sequence. A site that failed the test is not tiered at all, and is on no panel.
Each tier is a sentence rather than a score, and every site carries its own reason, so a laboratory can disagree with one rule and know exactly which. The method's own tiering lives in a supplementary data file whose terms restrict commercial use, so it is not copied; what baSeq uses is what the paper states in its text and figures.
Nomination is half the method
The other half is confirmation. A nominated site is a candidate until the site itself is amplified and sequenced deeply enough to measure the editing there, and a promiscuous proxy cell line nominates far more sites than the cells a therapy uses will edit. The Confirmation Panel step produces the panel: the sites in order, the amplicon window either side of each cut, and the depth each site needs, which is a floor of 1,000-fold on every site and past that five times the inverse of the frequency being claimed, so a 0.01% edit needs 50,000-fold. The panel's own depth is its deepest site's. The reads of such a panel can then be analysed in baSeq itself, in a CRISPR Amplicon-seq analysis.
Every nominated site is also a candidate for base-editing activity, whether or not it shows indels: a base editor is fused to a nickase, so it binds where the nuclease binds, and in this method's own data the double-strand-break frequencies of a high-fidelity nuclease rank-correlate with base-editing frequencies at the same sites. Whether anything can happen at a site is a property of its sequence, so the panel says where an adenine or a cytosine editor's window holds an editable base, on the strand the editor actually sees.
The mitochondrion
A guide cannot be aimed at mitochondrial DNA. The guide RNA is not imported into the mitochondrion, which is why editing mtDNA needed protein-only base editors rather than Cas9 at all. So a cut site on chrM is not the nuclease's work: it is a tag captured at one of the breaks mtDNA carries anyway, or a read from a nuclear mitochondrial segment that mapped to chrM instead of to its nuclear copy. Being high-copy, chrM collects those whatever guide was used, and a chrM site can match a guide by chance, which is exactly how one ends up read as a real off-target. Every step that lists sites leaves them out by default and says how many it left.
What it does not tell you
It says where the nuclease cut, not how much editing followed. A site with many molecules was cut often; turning that into an editing frequency is what the confirmation panel is for. And a nomination list from a promiscuous cell line is deliberately long: its purpose is to be sensitive, with the paper's own measurement being that between 2.4% and 34.5% of sites nominated in a promiscuous line converted to confirmed editing in primary cells, while the frequency ranking was largely conserved.
Step reference
| Step | Stage | What it does | You get | Key inputs |
|---|---|---|---|---|
| Data Upload | Setup | Reads one library per sample: a BAM of the tag-specific reads, or a table of sites a pipeline already called. Calls the integration sites, places each guide's spacer in every site's window by edit distance, and tests every site in every treated library against the pooled no-guide controls. | What each library caught, every site against the control, how strong the nominated sites are, and the fewest molecules each library could have nominated at all. | Name the no-guide libraries control. A sheet naming each library's guide, group and DNA input, a GTF and a genome file are taken if uploaded. |
| Library Quality | QC | Whether each library could have answered the question: reads against the depth the method's own sensitivity rests on, the genome copies amplified, the share of molecules read from both sides of the cut, the on-target site's depth, the share on the mitochondrion, and how deep its pooled control is. A measure the upload never took is left out rather than drawn from a default. | Each measure library by library, the ones to look at starred, and each library's detection floor. | None. |
| Nominated Sites | Analysis | Every nominated site with its sequence set against the guide that explains it: mismatches, and bulges where a one-base loop lines the site up better. One panel per guide, each site scored against its own guide. The sites that failed the test are one click away rather than hidden. | One panel per guide: that guide, then each of its sites beneath it with matching bases as dots and mismatches in colour, the molecules beside each; how far the sites are from their guide; and how much was cut against how far. | Nominated sites, every site, or every site matched to a guide. Sites drawn per guide, their order, the libraries, and whether to keep the mitochondrial sites. |
| Reproducibility | Analysis | How many of a treatment's replicates nominated each site, how quickly further replicates stop finding new ones, and which sites each library alone would have missed, counted separately for the sites worth confirming. | Sites by how many replicates hold them, where the molecules sit, what one library alone would miss, and how strong the sites at each level are. | A replicate group, or all of them. The mitochondrial sites. |
| Severity and Tiers | Analysis | Sorts the nominated sites into the five published severity bins by their strength against their own guide's on-target site, and into tiers by strength, reproducibility and where they land. A site with no on-target site to compare against has no bin, which is not the same as a weak one. | How many sites in each bin, what each bin meant when the method's own panel was confirmed, the bins per guide, every site against its on-target site, and what to confirm in order. | Split by guide, by replicate group, or all together. The mitochondrial sites. |
| Guide Specificity | Analysis | How specific each guide was: the share of molecules on target, the specificity ratio as this method defines it, the number of nominated off-target sites, and the cumulative off-target ratio. The CFD and MIT scores are shown only for a nuclease they were trained on. | Each guide's specificity side by side, where its molecules went, and its sites by severity. | Group by guide, nuclease or library. The sites and the libraries. |
| Mismatch Patterns | Analysis | Where along the spacer a mismatch is tolerated, what the sites look like as a logo, which PAMs the sites carry, and how many sites only a gapped alignment finds at all. | Mismatches per position of the protospacer, the sites' sequence logo, the PAMs used, the spread of edit distances, and the sites a mismatch-only matcher would have dropped. | Count each site once or by the molecules captured there. The sites and the libraries. |
| Site Annotation | Analysis | Where in the genome the nominated sites landed: the nearest gene, how far its start is, and which part of the gene the site sits in. A site in a coding exon is what raises a tier whatever its strength. | Every site across the genome as deep as its molecules, what the sites sit in, and the sites in coding sequence named with their genes. | The sites in coding sequence alone or every one. The libraries and the mitochondrial sites. |
| Confirmation Panel | Analysis | The panel to order: which sites to amplify, in what order, the coverage each one needs to be measurable at the frequency it is likely to show, the amplicon window either side of each cut, and where a base editor could act on the same site. | What is on the panel, how deep each site has to be sequenced, how many sites at each depth, where a base editor could act, and a CSV a primer design tool can take. | Which tiers to include, and how far either side of the cut an amplicon must reach. |
| Predicted vs Detected | Analysis | Searches the whole genome for every sequence within a few mismatches of each guide with a PAM the nuclease accepts, and lays the predicted list beside the nominated sites. A nominated site the search could not have reached is marked rather than counted against the prediction. | What the prediction found and missed, how many places the genome holds at each mismatch count, how many of them were cut, and the sites no prediction found. | The genome build, the mismatches searched for, and which PAMs. |
| Across Conditions | Analysis | Whether what was nominated in one group of libraries holds in another: a promiscuous proxy cell line against the cells a therapy uses, wild-type against a high-fidelity nuclease, one editor against another. Says when the two groups carry different guides, which confounds the comparison by design. | The sites each group holds and those they share, the rank correlation of their frequencies, what share of one group's editing the other accounts for, and a test per site where there are replicates. | The column to compare and its two values. The sites and the mitochondrial sites. |
Every kind of analysis baSeq runs: baseq.bio/analyses