Probabilistic Modelling of Base Editing Efficiency and Bystander Outcome Composition

Abstract. Base editors install single nucleotide substitutions without a double strand break or donor template, but editing at multiple susceptible positions can generate bystander alleles. Predicting the frequency of an intended outcome therefore requires both editing efficiency and the composition of edited products. Here we present crispAIBE, a probabilistic framework that uses separately trained transformer convolutional networks to predict Dirichlet distributions for editing efficiency and conditional allele composition. Split-conformal density-level regions are calibrated on data reserved from model selection, with composition thresholds fitted by outcome-set size. On held-out data, crispAIBE achieves Pearson correlations of 0.669 for efficiency and 0.896 for composition on measured allele sets, with empirical coverages of 0.902 and 0.906 at a nominal level of 0.90. On the published CRISPRon-BE benchmark, with test protospacers excluded from training, crispAIBE ranks first in seven of eight comparisons under the two-column Pearson statistic and first or second in seven of eight under the rank statistic. Within predicted-efficiency quintiles, the largest- and smallest-region deciles differ 1.9-fold in mean absolute error on average, compared with 1.7-fold for a boundary-distance ranking. When evaluated on U2OS, which was excluded from training, larger predictive regions were associated with larger absolute prediction errors. Local calibration with 100 U2OS protospacers yields median empirical coverage of 0.918 at the nominal 0.90 level. The crispAIBE repository is https://github.com/furkanozdenn/crispAIBE.