crispAIPE: Probabilistic Modelling of Prime Editing Variant Correction Efficiency

Prime editing installs precise substitutions, insertions, and deletions without double-strand breaks or donor templates, but the efficiency of an individual pegRNA design is hard to predict in advance, and existing tools give only point estimates, leaving designers unable to judge which predictions to trust. We present crispAIPE, a transformer-based probabilistic framework that quantifies per pegRNA-target pair uncertainty by modelling the three competing prime-editing outcomes on the 2-simplex with a Dirichlet likelihood and pairing the posterior with split-conformal highest-density regions calibrated on a held-out fold, giving finite-sample coverage guarantees on the simplex without requiring the Dirichlet model itself to be well-calibrated. Trained on 92,423 PRIDICT Library-1 pegRNAs under mutation-level target-disjoint splitting, crispAIPE attains Spearman{rho} = 0.835, 0.843, and 0.693 on the edited, unedited, and indel fractions (Pearson r = 0.845, 0.855, 0.659), and its conformal regions match nominal coverage where region-construction baselines do not. pegRNA architecture and edit context, in particular GC content of the mutated reverse transcription template (RTT) and edit size for deletions, are associated with prediction uncertainty. Reusing the Library-1 calibration quantile unchanged on PRIDICT Library-2, the conformal region area generalises as an actionable filter for cross-cell-type design, identifying the pegRNAs whose Library-1 predictions transfer best while small-sample head fine-tuning further improves accuracy. Tool and trained models: https://github.com/furkanozdenn/pe-uncert.