crispAIPE

Prime editing installs precise substitutions, insertions, and deletions into the genome without double-strand breaks or donor templates, but the efficiency of any individual pegRNA design remains hard to anticipate. Successive predictors — DeepPE, PRIDICT, DeepPrime, PRIDICT2.0 — have improved point accuracy, yet each returns a single number per design, leaving experimentalists no principled way to separate a confident prediction from a guess. crispAIPE returns instead a full posterior over prime-editing outcomes for each pegRNA–target pair, plus a region of plausible outcomes whose coverage is guaranteed.

The framing is compositional. A prime-editing experiment partitions reads into three competing outcomes — intended edit, unedited, and indel-containing — whose fractions sum to one, so the prediction target is a point on the 2-simplex rather than three independent numbers. crispAIPE places a Dirichlet likelihood over that simplex, which respects the sum constraint by construction and lets the concentration parameters express how tightly the outcome is determined. The Dirichlet posterior is then wrapped in split-conformal highest-density regions calibrated on a held-out fold, giving finite-sample coverage guarantees that hold even when the Dirichlet fit is itself miscalibrated.

crispAIPE was developed by Furkan Özden, Peiheng Lu, and Peter Minary in the Department of Computer Science at the University of Oxford, with funding from Google DeepMind, and posted to bioRxiv in December 2025 with a revised version in August 2026. In the catalog it sits alongside crisprSFM, which scores guide–DNA off-target specificity: both predict what a CRISPR reagent will do at a given site from sequence, rather than designing the reagent itself.

#Key Features

• Compositional outcome modelling: The Dirichlet likelihood over edited, unedited, and indel fractions captures the competition between the three outcomes and the varying precision of the underlying screens, which independent per-outcome regressors cannot.
• Distribution-free coverage: Split-conformal highest-density regions match nominal coverage on held-out data, where region-construction baselines such as quantile regression and random forests systematically miss it.
• Leakage-controlled evaluation: Splits are made at the level of the target mutation rather than the individual pegRNA — 64,751 training, 13,828 validation, and 13,844 test pegRNAs across 9,344 / 2,002 / 2,003 mutations with zero mutation overlap.
• Calibration that transfers: The Library-1 conformal quantile is reused unchanged on PRIDICT Library-2, where region area still works as a filter for which designs carry over across cell types.
• Interpretable uncertainty: Prediction uncertainty tracks pegRNA architecture and edit context, in particular the GC content of the mutated reverse transcription template (RTT) and edit size for deletions.

#Technical Details

crispAIPE is a hybrid transformer–convolutional network. Each pegRNA–target pair is encoded as an 11-channel, 99-position matrix combining 5-bit one-hot nucleotides, 2-bit edit directionality, and 4-bit annotations marking the protospacer, primer binding site, and the initial and mutated RTT; sequences are tokenized as 3-mers. Four transformer encoder layers with four attention heads and 64-dimensional embeddings produce contextual representations that are concatenated back onto the raw channel encoding, passed through convolutional layers and an MLP, and mapped by a softplus to the three Dirichlet concentration parameters, trained under a Dirichlet negative log-likelihood.

Training uses the 92,423 PRIDICT Library-1 pegRNAs, which target 13,349 human pathogenic mutations screened with PE2 in HEK293T and K562 cells, at batch size 128 and a learning rate of 6e-4 for up to 100 epochs with early stopping. On the target-disjoint test split, crispAIPE reaches Spearman ρ = 0.835, 0.843, and 0.693 on the edited, unedited, and indel fractions (Pearson r = 0.845, 0.855, 0.659). The repository publishes the model code, the canonical target-disjoint training script, evaluation and ablation scripts, and checkpoint-driven inference; it does not distribute a pretrained checkpoint or declare a license, and the paper itself is CC BY.

#Applications

The practical output is a triage signal for pegRNA design. Given a set of candidate pegRNAs for a target mutation, a designer can rank not only by expected edited fraction but by the area of the conformal region, prioritizing designs whose predicted outcome is tightly determined and flagging those where the model's evidence is thin. Because the calibration quantile transfers to a second library, the same fixed model can screen designs for a cell type it was not trained on, with optional small-sample fine-tuning of the output head when local data is available.

#Impact

crispAIPE is the first prime-editing efficiency predictor to supply calibrated uncertainty rather than point estimates, and its conformal construction is unusual in genome-editing tooling: coverage holds without assuming the neural posterior is correct, which is the property a designer needs when acting on a prediction. The work remains a bioRxiv preprint awaiting peer review, and its scope is narrow by design — a single-task outcome predictor, not a general sequence representation model. Training on pathogenic mutations screened with PE2 in two immortalized cell lines limits transfer to newer editor variants and primary cells, and the model does not represent chromatin accessibility or DNA methylation, both of which account for real variance in editing efficiency.