Prime Editing Efficiency Predictor — Rank pegRNAs with PRIDICT2.0
Score every PBS/RTT combination for one prime edit with PRIDICT2.0 and rank the pegRNAs — HEK293 and K562 scores, library percentiles, the full pegRNA, and Golden Gate cloning oligos.
🌐 Nothing you paste is logged or stored — every tool is also callable via REST & MCP, and in bulk from the batch tools
Predicted, not measuredSpearman ρ = 0.85 on held-out data from the libraries it was trained on
- How good is it?
- Spearman rho of about 0.85 for intended edits on held-out library data — the best-validated figure of any model in this registry, and roughly double OSTIR's 0.39 on independent data. That figure is still within the library and cell lines it was trained on.
- Only valid for:
- human sequence, and efficiency ranking within one locus. It is parameterised on HEK293 and K562; your cell type, delivery method, and chromatin context will all move the absolute efficiency, chromatin alone by severalfold. Nothing here is predicted for a non-human host or for editors outside the PE2/PE3 architecture the training libraries used.
- Fitted on:
- PRIDICT2.0, an attention-based deep model trained on measured prime-editing outcomes from large pegRNA libraries in two human cell lines, HEK293 and K562 (Mathis, Allam, Kissling et al., Nat Biotechnol 2024).
Write your target with the edit in brackets — context, then (original/edited), then context — and PRIDICT2.0 enumerates the primer-binding-site and reverse-transcriptase-template length combinations that could install it, scores every one, and returns the highest-ranked designs with the full pegRNA sequence and its Golden Gate cloning oligos. PRIDICT2.0 (Mathis, Allam, Kissling et al., Nature Biotechnology 2024) is an attention-based deep model fitted on measured prime-editing outcomes from large pegRNA libraries in HEK293 and K562 cells, and reports a Spearman rho of about 0.85 for intended edits on held-out library data — the best-validated model on SeqBench. Use it to choose between designs at one locus: it ranks pegRNAs, and a score is not a promised editing percentage in your cells.
Context, then the edit in brackets, then context — …ACGT(A/G)ACGT… — with roughly 100 bp or more on each side, which the model reads as features. Unchanged flanking bases stay outside the brackets: T(a/g)C, not (TAC/TGC). An insertion or deletion leaves one side empty: (/AGG) inserts AGG, (AGG/) deletes it.
Both scores always come back; this only picks which one the ordering uses. There is no generic-mammalian option because the model has no such training data.
How to use the Prime Editing Efficiency tool
- 1Paste the target region with the edit in brackets: context, then (original/edited), then context — for example ...CTACGGCCGA(G/C)GTGCGAGGCC... — with roughly 100 bp or more each side. Unchanged flanking bases stay outside the brackets (T(a/g)C, not (TAC/TGC)); an insertion or deletion leaves one side empty, as in (/AGG) or (AGG/).
- 2Check the read-back under the box: it reports how many bracketed edits it found, what kind of edit each one is, and how much context sits on each side.
- 3Pick which cell line the ranking should use (HEK293 or K562 — both scores always come back), set how many designs to return, and click Score pegRNAs. A run takes roughly 8-10 seconds once the service is warm.
- 4Compare the ranked designs, then select one by its number to copy the full pegRNA and its two Golden Gate cloning oligos — and read the "not checked" list before you order anything.
Frequently asked questions
Which model produces these scores?
PRIDICT2.0 (Mathis, Allam, Kissling et al., Nature Biotechnology 2024), an attention-based deep model trained on measured prime-editing outcomes from large pegRNA libraries in HEK293 and K562 cells. It is MIT-licensed and runs on SeqBench's own service; the scores, percentiles, pegRNA sequences and cloning oligos are all emitted by the model's own pipeline rather than recomputed here.
How accurate is it?
PRIDICT2.0 reports a Spearman rho of about 0.85 for intended edits on held-out library data — the best-validated figure of any predictive tool on SeqBench, and roughly double the 0.39 behind the RBS Designer. That is still held-out data from the same libraries and the same two cell lines, so it measures how well the model orders pegRNAs in the setting it was fitted on.
Does a score of 60 mean 60% of alleles get edited?
No. The number is on the model's own 0-100 scale, and the percentile places a design against the library it was fitted on. Cell type, delivery method and chromatin context all move absolute editing efficiency — chromatin alone by severalfold — so treat the output as an ordering of these designs against each other at one locus, not as a yield you can plan an experiment's success around.
Why only HEK293 and K562?
Those are the two contexts the training libraries were measured in, so those are the two the model can speak to. There is deliberately no generic-mammalian option, because there is no such training data. Both scores always come back on every candidate; choosing a cell line only decides which one the ranking is sorted by, and comparing the two columns is a cheap way to see how context-sensitive a design is.
Does it score PE3 nicking guides?
No. Scoring a nicking guide requires DeepSpCas9, whose weights PRIDICT2's MIT licence does not cover, so they are removed from this deployment's image and no nicking-guide number is ever produced here. Design PE3 nicks in the Prime Editing Studio and validate their specificity separately — off-target activity is not evaluated by either tool.
How do I write the edit, and how much flanking sequence do I need?
As context, then the edit in brackets, then context: ...ACGT(A/G)ACGT... . Keep unchanged bases outside the brackets — T(a/g)C rather than (TAC/TGC) — and leave one side empty for an insertion or a deletion, e.g. (/AGG) or (AGG/). Roughly 100 bp each side is the guidance, but the exact requirement is PRIDICT2's own and it measures the context itself after normalising unchanged bases out of the brackets, so the page reports the counts it read and lets the model decide. If it refuses, add more flanking sequence rather than moving bases into the brackets.
What are the Golden Gate oligos for?
Each design comes with the two oligos PRIDICT2 emits for its standard Golden Gate pegRNA cloning route: one for the spacer, one for the 3' extension. They are passed through verbatim, so check the overhangs against your own vector before ordering, and check the pegRNA for restriction sites your assembly cares about — nothing here inspects your backbone.
Is my sequence stored, and can I run this from code?
No. Your sequence goes to SeqBench's own PRIDICT2 service over a private network — not to any third party — and nothing is persisted to disk. prime_editing_efficiency is also callable from the REST API and the MCP server. Each run pushes hundreds of candidates through a deep model on a shared service, so it is rate limited, and the five-fold averaging option costs five times as much.
More
Related tools
Design knockdown reagents against an mRNA — Reynolds/Ui-Tei-scored siRNAs with ready shRNA cassettes, or 5-10-5 ASO gapmers screened for known liabilities.
Design KASP/ARMS SNP-genotyping primers — two allele-specific forward primers with FAM/HEX tails and an engineered ARMS secondary mismatch, plus a common reverse primer.
Diagnose gel problems (no bands, smearing, wrong size, primer-dimer, degradation) with a fully deterministic symptom guide that links to the right SeqBench tools.