Sanger-trace quality control for CRISPR edits in the browser: knockouts, deletions between cuts, SNP corrections, small tag knock-ins, and large tag or reporter knock-ins. The companion to the Assured CRISPR Design app: the design app writes what the correct allele should look like, this app checks whether your traces show it.
Your files stay on your computer. Traces are read and analysed inside the browser (in a web worker); nothing is uploaded and there is no server.
Status: 0.3.1 (adds offset re-anchoring at the cut, an unplanned-base-change check and the functional knockout assessment; see docs/changes-0.3.0.md; engine freeze in bench/ENGINE_FREEZE_0.3.1.json). See "Accuracy and limits" before using results to accept or reject clones.
| Workflow | Input | What you get |
|---|---|---|
| Knockout (indels at one cut) | control + edited .ab1, guide(s) | editing %, knockout score, indel profile, clone genotype call |
| Deletion between two or three cuts | control + edited .ab1, guides | the above plus the deletion between the cuts |
| SNP correction (ssODN with silent blocking changes) | control + edited .ab1, design file or guides + donor (up to 300 nt) | intended edit separated from blocking-changes-only conversion, per-position readout of every designed change, clone genotype |
| Small tag knock-in (ssODN) | as for SNP | as for SNP |
| Large tag or reporter knock-in | junction or out-out reads + design file, optional band sizes | each read compared with the expected knock-in allele: junctions spanned, designed changes present, unexpected differences, out-out band classification |
To check the non-knock-in allele of a large knock-in clone, gel-purify the wild-type-sized band of the out-out PCR and run it as a Knockout sample.
Lab convention supported throughout: two guides per edit, with the donor carrying silent blocking changes for both guides.
Compare with ICE. After analysis, expand Compare with ICE and download a batch folder ZIP containing ICE workbooks, AB1 archives and a sample map. Two ssODNs become separate donor-model rows for the same trace pair. Nothing is submitted automatically. See ICE comparison.
In the browser. Drop a folder of .ab1 files (and, optionally, the design file exported by the design app, or a sample sheet) onto the start page. Files are paired automatically (control, edited, forward or reverse read) with a stated confidence for every pairing; the pairing table can be edited. The decision for each sample is accept, hold, re-sequence or reject, with the reasons, and the report, CSV, plate view, re-sequence list and a publication figure are one click each. The stepwise form is still available below the start page.
From the command line (same engine, parallel worker threads, no browser):
node cli/assured-qc.mjs run <folder> --design design.json --out results/ # or --sheet samples.csv, or --guides G1,G2 --donor <ssODN>
node cli/assured-qc.mjs figure --control control.ab1 --edited edited.ab1 --design design.json --out figure.svg
It writes results.csv, results.json, report.html, plate.html, resequence.csv and run.json (input hashes, parameters, engine version). 399 simulated pairs take 44 s on 11 threads (about 32,000 pairs per hour).
Design file. Export the design from the design app as a QC design file (JSON, schema assured-qc-design/1; see docs/design-handoff.md; needed for the reporter workflow), or type the guide sequence(s) and donor into the app. Other nucleases (SpCas9, SpCas9-NG, SaCas9, AsCas12a, LbCas12a, Mad7, legacy Cas12a, SpRY) are supported, and the cut can be inferred from the traces when no guide is given (simulation-validated only).
The edited read is aligned to the design reference using the control read to remove the sequence-specific peak pattern. Within a window around the cut sites, the per-position base mixture of the edited trace is explained as a non-negative mixture of candidate alleles: wild type, indels of 1 to 30 bp at each cut, the deletion between cuts, the intended edit, and partial conversions (blocking changes without the full edit). The mixture weights are the allele fractions. Details in docs/method.md.
Unlike the headline editing number of other Sanger tools, the intended edit and the blocking-changes-only conversion are reported separately. A synthetic clone with half intended-edit alleles and half blocking-only alleles is reported here as about 50% intended edit and 47% blocking-only; ICE 1.2.0 reports 93% HDR for it (see docs/validation.md).
The knockout score of earlier versions counts alleles whose net length change is not a multiple of three, or that are 21 bp or longer. That is a frameshift proxy: it ignores where the edit lies in the gene. With a gene record (GenBank file with the CDS and exon structure) each allele is placed in the coding sequence, translated, and judged on the protein: premature stop and nonsense-mediated decay (exon-junction rules), truncation size, restart of translation at a downstream ATG, splice-site loss (exon skipping against intron retention), start and stop loss, and in-frame changes (against essential regions you mark). A clone gets one call (predicted knockout, with caveats, not a knockout, uncertain), a functional knockout score with an interval, and a list of what to run to confirm it (Western blot or mass spectrometry, RT-qPCR downstream of the stop, RT-PCR across the exon for splice edits).
node cli/assured-qc.mjs run reads/ --design design.json --genbank gene.gb \
--regions "TIR domain:172-314" --epitope 200-250
In the browser, add the .gb file with the traces. The call is a prediction from sequence; residual protein after frameshift edits is common (about a third of verified knockout lines in one survey), so loss of protein, not the score, confirms a knockout. Rules, thresholds, limits and references: docs/functional-knockout.md.
recommend ranks the clones of a run by how strongly the sequence predicts that no functional protein is left (the knockout support: 85 at best from sequence alone, 100 only with a protein result) and names the best clone of each genotype as the panel to confirm first. The browser shows it as a Priority column; the command line writes recommendation.csv. For each sample with a gene record the app also draws the predicted protein products (residues kept, frameshifted tail, NMD, mass) with the stretches where an antibody epitope sees every product or the full-length protein only, to choose the antibody for the Western (docs/functional-knockout.md). Protein results from a Western blot (--observations protein.csv) accept clones whose protein is absent and reject those whose protein is still made. The weights are provisional; see docs/recommendation.md.
The run is also drawn on one summary graph (decision, alleles with their shares, edited share with its interval, knockout support, plan), in the browser, at the top of the report and as summary.svg. Every clone is also drawn against the wild-type control (annotated chromatograms, and all clones aligned on the same stretch; docs/aligned-chromatograms.md): in the report, in the browser and as clones_alignment.svg, clones_aligned.svg and, with --clone-figures, chromatograms/. A figure from the design to the clones to take forward (guides, cuts and designed changes over the chromatograms of the clones carried forward, and a table of all clones) is in the report, in the browser and as clone_selection.svg (docs/clone-selection.md). Every figure is also saved as a vector PDF (a (PDF) button in the browser, a .pdf next to each .svg on the command line, --no-pdf to skip; docs/pdf-export.md). A design report with several recommended donors (the ssODNs of a point edit) uses them together.
Held-out simulated test (traces built from real controls with known alleles; protocol and success criteria fixed before the run; docs/scorecard.md, docs/results.md): on 207 traces from loci not used in development, mean absolute error of the edited share is 1.9 points (95% CI 1.5-2.3) against 6.6 for ICE 1.2.0, 5.9 for peaksplit and 5.0 for Assured_TraceEdit; the intended-edit error is 4.2 against 7.0, 6.6 and 9.3, large deletions of 150 bp are recovered in 92% of traces (ICE 25%), and partial conversions are never reported as the intended edit (ICE and peaksplit always do). Criteria not met on this test: wild-type traces are called 2 or more points edited in 2 of 21 cases (the interval rule reports none of them as detected), the practical detection limit is about 5% (target 2%), HDR dilutions at 5-25% are 3 points off on average, and intended-edit intervals cover 81% of truths (target 90-98%; edited-share and knockout intervals cover 96% and 95%). The truth in this test is simulated, so these numbers are a bound on how well a tool can do on traces of this kind, not a measurement on your clones; the wet-lab mixture protocol (docs/known-mixture-protocol.md) has not been run. Only tools that can be scripted here were compared (ICE, peaksplit, Assured_TraceEdit); DECODR, TIDE, SeqScreener, EditR and Tracy were not.
On 172 laboratory pairs there is no sequencing-confirmed truth; the engine agrees with direct chromatogram readouts to 2.9 points for knockout clones (docs/validation.md). Earlier checks: 17 synthetic pairs used while the engine was built (1.5 points for editing %), and the ICE example traces (agreement within a few points, except a three-guide multiplex at 74% against ICE 63% and a 2-bp substitution at 34% HDR against ICE 44%).
- The functional knockout score is a sequence-based prediction and has not been calibrated against protein measurements; without a gene record only the frameshift proxy is reported.
- Detection limit is about 5% for an allele; Sanger traces cannot resolve smaller fractions. Every result carries an interval and a confidence tier (high, moderate, low) from the interval width; low-tier samples are not accepted automatically.
- A trace gives an unphased mixture per position. Alleles that differ only by combinations of changes, rather than by sequence, are reported as "also consistent with" where indistinguishable.
- The clone call assumes at most two alleles. A large deletion on one allele that removes a primer site shows up as apparent homozygosity; check with out-out PCR.
- Junction reads show whether a knock-in allele is correct, not how many copies or alleles are present.
- Trace quality limits everything. The app refuses swapped or mismatched control/edited pairs, reads that stop before the cut, and misaligned amplicons; it explains a control that is itself mixed (typically after a homopolymer run) rather than calling it the wrong design; it warns when the two traces already differ before the cut (discordance above 0.25, a provisional threshold from six real and several synthetic traces) and when the fit is poor.
- Real-trace accuracy beyond the ICE examples has to be confirmed with your own control/edited pairs of known outcome before this replaces your current check.
npm ci
npm test # engine tests (node:test), no browser needed (107 tests)
npm run dev # http://localhost:5174
npm run build
node scripts/validate.mjs out.csv # own engine vs known mixtures vs ICE on the synthetic traces
python bench/simulate.py jobs_real.json out/ --seed N # simulated truth from real controls
python bench/run_all.py jobs.json out/ --tools assured_qc,ice,peaksplit,traceedit
python bench/score.py jobs.json out/ prefix --split test # scorer frozen in bench/FROZEN.json
The analysis modules in src/ are plain ES modules with no browser or Node dependencies (abif.js, align.js, decompose.js, ...); src/runJobs.js is what the worker runs, src/App.jsx is the interface. scripts/make_qc_fixtures.py regenerates the synthetic traces (deterministic).
The engine is written from scratch for this project. Synthego ICE (https://github.com/synthego-open/ice, Conant et al., CRISPR J 2022) was run only as a reference to compare results; none of its code is included, and its example traces are not redistributed. The repository licence has not been set yet.
Choose AsCas12a, LbCas12a or Mad7 when entering spacer sequences, or import the enzyme-aware Design/Bench JSON or HTML report. The same control/clone upload, indel and donor analysis, clone decisions, figures, reports and CSV exports are available. Cas12a uses a 5′ TTTV PAM; Mad7 uses 5′ YTTN. Named enzymes use nominal target/non-target cuts at spacer offsets 23/18. Both strands and reverse reads are handled. The old generic Cas12a option retains its midpoint anchor for compatibility, but now models the whole predicted cut span.
The expected-allele library includes indels within the staggered-break span, paired-guide deletion junctions spanning both strand boundaries, complete donor edits and partial conversion. Upstream alignment and readable-window checks use the earliest/latest strand cut, rather than aligning through an edit near the other boundary. Figures use the configured PAM side/length and show both cuts. Results record the enzyme, config version, analysis model and predicted boundaries. An unknown or conflicting enzyme is rejected rather than treated as Cas9.
These paths have computational regression coverage on idealised signals and generated AB1 traces: wild type, heterozygous/homozygous indels, opposite-strand paired deletions, SNP repair/partial conversion, small-tag insertion and reverse reads. No laboratory Cas12a/Mad7 calibration or reference-tool equivalence is claimed. Nominal cuts can vary with enzyme, sequence and conditions; the 5% floor, intervals and clone decisions still need independent validation for these enzymes. Sanger limitations (copy number, allele dropout, long events, off-targets) remain. The verified ICE CSV template has no enzyme field; non-Cas9 cases are explicitly excluded from that exporter and require manual setup in a compatible tool.