baSeq docs › CRISPR amplicon-seq

CRISPR amplicon-seq

One short piece of the genome, sequenced deeply, so every molecule in the library is one editing outcome: how much of the DNA was changed, into what, by which repair pathway, and whether the change you wanted was made.

What to upload

One or two FASTQ files per sample: the reads of the amplicon themselves. This is the one analysis that reads FASTQ directly: there is no count table to make from an amplicon library, because aligning its reads to the amplicon is the analysis. Paired reads are merged back into the molecules they came from, and a library whose pairs do not overlap is told so rather than quietly losing them.

Beside them, a sheet naming each amplicon's sequence and the guide aimed at it. The sheet is not a formality: the reads say what the DNA is now, not what it was meant to be, so without the amplicon sequence there is nothing to compare against. Columns are matched by what they mean, so a sheet written for CRISPResso or ampliCan is read as it is: sample, amplicon, amplicon sequence, guide, nuclease, the primers, a coding sequence, an HDR donor, a pegRNA's extension and scaffold, or a base editor's conversion. Anything else in the sheet is kept as a sample column.

Name the unedited sample mock (or untreated, unedited, no-guide, nuclease-only). Without one, a low editing rate cannot be told from the indels a PCR and a sequencer make on their own, and every small number has to be read as an upper bound. Note that wt is not read as untreated: in a specificity experiment it usually means wild-type Cas9 against a high-fidelity variant, which is an edited sample.

Already run another tool? Upload its allele frequency table, or a whole CRISPResso_on_<name> folder, and the rest of this analysis is available on top of it. The alignment it made is kept rather than redone, so its numbers come back as they were.

The three choices that move every number

The same library can read 20% or 26% edited depending on choices that often go unrecorded, so each is a setting here, shown beside the number it produced.

Where an edit counts. The default is the guide and 10 bases either side, with the primer ends excluded, because an "edit" inside a primer is the primer. The published tools' default of one base either side of the cut is offered, and so is the whole amplicon; Editing Efficiency draws the same reads under all three. A deletion counts if any of its possible placements falls in the window, because inside a microhomology the placement is not knowable.
Substitutions are not editing, by default. A nuclease makes insertions and deletions; a sequencing run makes substitutions at about the rate of a real low-level edit. They are counted and reported separately, and counted as editing only when you ask, or in base editing, where the substitution is the edit and is measured against the untreated sample.
What is not known is not an edit. A base the sequencer was unsure of is neither a match nor a mismatch. A molecule that does not cover the whole window is counted apart, never as unedited. A read that fits no amplicon is counted and named, never forced onto the nearest one.

How much can you see?

Every rate comes with its confidence interval and with the smallest rate the depth could have detected: against a typical 0.1% background that is about 0.67% at 300 reads, 0.24% at 2,500 and 0.13% at 50,000. It is what makes "no editing detected" mean something: a site with 300 reads simply cannot show a 0.5% edit, and reporting it as zero would be a statement about the sequencing rather than about the DNA.

For an off-target panel this is the whole game. A site is reported as edited only when an untreated sample cannot explain its edits, and a clean site is reported as below its own floor rather than as zero.

Comparing conditions

With replicates, editing rates are compared with a beta-binomial test, which asks whether the conditions differ by more than the replicates of one condition do. This matters more than it sounds: reads are PCR copies of a few hundred thousand molecules from one dish of cells, not independent samples of the biology, so a test that pools them calls a difference between 30.0% and 30.5% overwhelming when replicates routinely differ by more. With one sample per condition nothing can be said about spread, and the step reports a pooled test and labels it a screen rather than evidence.

Sample datasets

Browse Sample Datasets in a CRISPR amplicon-seq session lists four published experiments, each built from the authors' own deposited reads and read by this upload exactly as your own would be.

EMX1 and nine off-target sites in primary human stem cells (Amit et al. 2021): CD34+ cells from mobilised peripheral blood, the guide delivered in four formats with and without the electroporation enhancer, and a mock. Every library covers all ten sites, so this is the one for Background Noise and the Off-target Panel: one near-match site is edited at 20% while the mock shows 1% apparent editing at some of the others, which is the difference the untreated sample buys you. The enhancer is worth four to eight times the editing, and the cleanest guide format is also the least efficient.
Prime editing at five sites, with every nicking-guide position (Anzalone et al. 2019): the pegRNA alone (PE2) against the same pegRNA with a second nick from 116 bases before to 90 bases after it, three replicates each. The second nick roughly doubles the intended edit and brings deletions between the two nicks with it, so the best position is a trade rather than a maximum.
Four base editors at the same target bases: BE2 and BE4max (cytosine) and ABEmax and ABEdmax (adenine) at HEK3, EMX1 and FANCF. For efficiency, product purity and the bystander bases, which is where the editors differ. The sheet names the cytosine conversion, so choose A to G in the Base Editing step for the two adenine arms.
Correcting the sickle-cell mutation, with the clonal lines it starts from: thirteen pegRNA designs for the same single base change, three replicates each, and the six clonal lines that carry the mutation. For Clonal Genotyping, which has to tell a clone (one allele) from a pool (dozens of them).

Step reference

Step	Stage	What it does	You get	Key inputs
Editing Overview	Setup	Reads one or two FASTQ files per sample: merges the paired reads back into the molecules they came from, collapses them to the distinct molecules, aligns every one to each amplicon of the experiment, and turns the alignments into alleles, edits and outcomes. An allele table from another tool (CRISPResso, ampliCan) is read instead, keeping the alignment it made.	Where every read went, how much of each sample was edited with its confidence interval, and what the editing was.	A sheet naming each amplicon's sequence and guide. Without one the amplicon is read off the data and said to be a guess.
Read Quality	QC	Where every read went (merged, assigned to an amplicon, or lost to quality, length, or fitting no amplicon), with each sample's depth and base quality, and the smallest editing rate that depth could have detected.	The fate of every read per sample, depth against what the depth can see, and base quality.	Samples shown.
Panel Coverage	QC	How the reads were shared between the amplicons of a multiplexed panel, and which sites were sequenced deeply enough to be called clean. A site with 200 reads cannot answer whether editing is below 0.1%.	Reads per amplicon per sample, how evenly they are spread, and each amplicon's detection floor.	Samples shown.
Background Noise	QC	What an untreated sample says the PCR and the sequencer do on their own, position by position, and which edits in the treated samples are therefore real. Each combination of edit kind, length and position is weighed as noise against editing, so a homopolymer slip is not read as an edit and a real edit at a noisy position is not thrown away.	The noise along the amplicon, every edit type as treated against untreated, and what the correction changes.	The amplicon. Needs an untreated sample, and says so when there is none.
Editing Efficiency	Analysis	The headline rate: the share of molecules carrying an edit, with its confidence interval, the background of an untreated sample taken off, and the depth it rests on. The same reads are also shown under each window, because a rate is only comparable with another computed the same way.	The rate per sample with its interval, what the editing was, and the same data under all three windows.	Samples, amplicon, where an edit counts, whether substitutions count, and the background.
Allele Table	Analysis	Every allele the editing produced, commonest first, drawn as the field draws it: the reference on top and each allele beneath with matching bases as dots, so a deletion reads as dashes and an insertion as letters in a gap.	The alleles around the cut with their shares, and how concentrated the editing is: one allele or forty.	Samples, amplicon, how many alleles, how many bases either side of the cut, and the unedited allele.
Indel Profile	Analysis	Where along the amplicon the editing sits and how big it is. Real nuclease editing is a sharp peak at the cut; a flat or multi-peaked profile means artefacts or a guide that is not where it was thought to be. A deletion whose placement is ambiguous is spread over the placements it could have.	Insertions, deletions and substitutions per position per sample, and the distribution of length changes with the frame marked.	Samples, amplicon, and whether to show reads or shares.
Repair Outcomes	Analysis	Which repair pathway made the editing: microhomology at the deletion junctions (measurable because the deletion's placement is ambiguous by exactly the homology's length), single-base insertions that duplicate the base the polymerase had to copy, and how repeatable the outcome is. Beside them, what the sequence itself predicted before any sequencing.	The microhomology share per sample, how long the homology is, the templated share, and precision.	Samples and amplicon.
Knockout Prediction	Analysis	What share of the DNA was changed in a way that usually breaks the protein (out of frame, or an indel of 21 bases or more), and, with a coding sequence, where the frameshift puts the next stop codon. Reported as an upper bound: about a third of frameshift knockouts verified by sequencing still made protein.	What the editing does to the protein per sample, every outcome by its length change, and where the stop codon lands.	Samples and amplicon.
Base Editing	Analysis	What a base editor converted, position by position across the protospacer (not at the cut, where almost no base edit sits), with its efficiency over all reads and over indel-free reads, its product purity, the bystander bases that changed in the same molecule, and the indels it is not supposed to make.	The conversion along the protospacer, efficiency and purity, bystanders per molecule, and indel byproducts.	Samples, amplicon, the conversion, and the intended position.
Prime Editing	Analysis	How much of the DNA carries the edit the pegRNA was designed to make, how much carries it with other changes, how much copied the guide's scaffold into the genome, and how much was simply broken. The intended product is built from the pegRNA's extension by following the chemistry.	Every outcome per sample, efficiency and purity, and the unintended products.	Samples. Needs the pegRNA's extension in the sheet, and its scaffold to find scaffold incorporation.
Off-target Panel	Analysis	Which sites of the panel were really edited and which were only sequenced badly: each site's rate with its interval, a test against the untreated sample, and the smallest rate its own depth could have detected, so a clean site is reported as below a number rather than as zero. The specificity ratio per sample is what a nuclease variant or a guide design is judged on.	Editing at every site, treated against untreated site by site, and specificity per sample.	Samples and the background. Needs more than one amplicon.
Clonal Genotyping	Analysis	Each sample read as a picked clone: which alleles it carries, whether both copies were edited, and whether the well holds one clone at all: three or more alleles means it does not, and that is the result. A single allele is a homozygote or a heterozygote that lost its other copy, and an amplicon cannot tell those apart.	The alleles of each sample with the genotype written out, and how the call depends on the threshold.	Samples, amplicon, the threshold an allele must pass, and the copies of the locus.
Differential Editing	Analysis	Whether editing differs between two conditions (a nuclease against a high-fidelity variant, one guide design against another), tested so that replicates count: a beta-binomial model asks whether the conditions differ by more than the replicates do. Reads are PCR copies, not independent samples of the biology, so pooling them makes a fraction of a per cent look overwhelming.	The rates side by side per amplicon, and every sample behind them.	The parameter and its two values, and the amplicon.