baSeq docs › GUIDE-seq

GUIDE-seq

GUIDE-seq finds where a CRISPR nuclease cut the genome in living cells, without guessing the places first. A short double-stranded tag goes in with the nuclease and is captured at the breaks; the reads pile up at the cut sites. baSeq calls those sites, sets each against the guide, scores it, says where in the genome it landed, and compares what was cut with what a prediction would have found.

What you upload

One library per sample. Either a BAM of the tag-specific reads (what the published pipeline leaves behind, and what a core facility sends back), or a table of sites a pipeline already called: identified_matched.txt from the GUIDE-seq pipeline, the Bioconductor GUIDEseq package's table, iGUIDE's, or a plain BED, CSV, TSV or spreadsheet of sites. Name the mock library (the one that had the tag but no guide) mock, control or Cas9-only and baSeq will treat it as the background.

Three things are optional and each adds something: a GTF or UCSC gene table so the sites can be annotated, a sheet of guides (sample, guide, spacer, PAM, nuclease), and a genome file (.2bit, or FASTA with its .fai) for a genome baSeq cannot name or download.

baSeq does not yet count raw FASTQ for GUIDE-seq: demultiplexing a pool by its index reads and tagging the molecular identifiers has to happen first.

How the sites are called

Each read's tag-proximal end marks where the nuclease cut. In a paired library that is normally the second read of the pair, which begins in the tag. But a library amplified the other way round has it on the first, so baSeq measures how much each end piles up at shared positions and keeps the one that piles. The other end of a sheared fragment lands wherever the shearing broke it, and never piles.

Positions no more than 10 bp apart are one site. A site is kept when it holds molecules read from both orientations: a tag can be captured either way round, so a nuclease's cut is read from both sides while a single amplified molecule or a mispriming event is read from one. Molecules are counted, not reads, from the molecular identifier in the RX tag or at the end of the read name, so a molecule amplified a thousand times counts once.

Matching the guide, mismatches and bulges

Each site's 50 bp window is searched on both strands twice: once for substitutions alone, once allowing a bulge: a contiguous loop of bases on one side with no partner on the other. The budget is seven: at most seven substitutions, and each bulge costs three of the same seven, so a site with one bulge may hold four mismatches. Both searches are reported, because a site over the budget on mismatches can fall inside it once a loop is allowed.

The PAM is part of the target, not a gate. A mismatch in it costs the same as one in the spacer, so a site with an NAG PAM is found for the price of one mismatch. Refusing a non-canonical PAM outright would throw away the NAG, NGA, NAA, NGT, NGC and NCG sites the original GUIDE-seq paper detected.

Scores, and what they do not say

Every matched site gets a CFD score (Doench 2016), which separated validated off-target sites from false ones better than any other score in the largest published comparison. It comes with two warnings baSeq makes visible rather than hiding. Nine of the sixteen possible PAM dinucleotides have no value in the published table and score exactly zero. A site with an NAA PAM scores 0 whether or not it was cut, so the PAM is shown beside every score and nothing is ever filtered on CFD alone. And CFD is undefined for a bulged site, so those are left blank rather than guessed. The MIT score is reported beside it for comparability; its formula appears in no paper and it ignores the PAM entirely.

Predicted against detected

Predicted vs Detected searches the whole genome for the guide and lays the result beside what was cut. Both directions are uncomfortable, and both are the point. Most predicted places were never cut: published recall of prediction runs about 92% at two mismatches and almost nothing at six. And some cut sites were never predicted, which is why the assay exists.

One rule keeps the comparison honest. The genome search allows mismatches up to the limit chosen and no bulges, while detection allows seven substitutions or four with a bulge. A detected site outside what the search could reach is marked outside the search rather than counted as "detected but not predicted": counting it that way would blame the prediction for a question it was never asked.

What GUIDE-seq 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 needs targeted sequencing of the site itself. Cellular assays like this one are nonetheless far more specific than naked-DNA ones, whose apparent false-positive rates run from 27% to 95% against under 20% for assays run in cells.

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 published pipeline already called (identified_matched.txt, the Bioconductor GUIDEseq package's, iGUIDE's, or a BED/CSV of sites). Calls the cut sites, matches each to the guide, and reads each site's sequence from the genome so the later steps need no genome of their own.	The molecules each library captured, the strongest sites, and how far they are from the guide.	None. A sheet naming each library's guide, a GTF and a genome file are taken if uploaded.
Library Quality	QC	Six measures per library against the protocol's numbers: reads, molecules, reads per molecule, the share of molecules at sites read from both orientations, the depth of the guide's own site and whether it is the strongest, and how evenly its strands are read.	Each measure library by library, the ones to look at starred.	None.
Off-target Sites	Analysis	Every cut site with its sequence set against the guide: mismatches, and bulges where a one-base loop explains the site better than mismatches alone. Sites the guide does not explain are kept and shown apart, since the nuclease may not have made those breaks.	The guide, then each site beneath it with matching bases as dots and mismatches in colour, the molecules beside each; the spread of mismatches; and plausibility (CFD) against distance.	Count molecules or reads. Sites shown. Whether to leave out the sites the mock library holds, and whether to show the sites the guide does not explain.
Mismatch Patterns	Analysis	Where along the guide a mismatch is tolerated, what the cut sites look like as a logo, which PAMs the nuclease used, and how close each tag went in to where the nuclease cuts.	Mismatches per position of the protospacer, the sites' sequence logo, the PAMs used, and the tag's distance from the cut.	Count each site once or by the molecules captured there. The mock library's sites.
Specificity	Analysis	Four measures of how specific each guide was: the share of molecules on target, the number of off-target sites, the summed CFD of those sites, and the MIT guide score.	Where each library's molecules went, and the four measures side by side.	Compare each library on its own or the libraries of each guide together. The mock library's sites.
Site Annotation	Analysis	Where in the genome the cuts landed: the nearest gene, how far its start is, and which part of the gene the site sits in (promoter, UTR, exon, intron, downstream or between genes).	Every site across the genome as deep as its molecules, what the sites sit in, and the sites in a coding part or a promoter named with their genes.	The off-target sites alone or every site. The mock library's sites.
Predicted vs Detected	Analysis	Searches the whole genome for every sequence within a few mismatches of the guide with a PAM the nuclease accepts, and lays that prediction beside what the tag actually marked.	Predicted places by mismatch count and how many were cut, the share cut at each count, what the prediction would have caught, and the cut sites it missed.	Genome build. Mismatches searched for (3 to 6). PAMs: NGG, NAG and NGA, NGG alone, or NRG. The mock library's sites.
Shared Sites	Analysis	Which libraries found which sites, and how well pairs of them agree : reproducibility between replicates, and the comparison itself between nucleases or guide designs.	An upset chart of the libraries holding each site, and the molecules of each pair against each other.	Pairs drawn. The mock library's sites.
Differential Editing	Analysis	Compares the molecules captured at each site between two conditions, with DESeq2 where each condition has replicates and the fold change alone where it does not.	A volcano of every site tested, and the sites that differ with the guide's own site among them.	Parameter to compare and its two values. The mock library's sites.