Table 1.

Benchmark quantities gained on the subset of duplicate guide-target sequence pairs (testing scenario 3) using our $6\times 23$ CNN model as well as the CRISPR-Net [13] and DeepCRISPR [11] models for comparison. For piCRISPR, we also give a sequence-only version of the model in which nucleosome organisation related features and the empirical energy estimate have been set to a default value across all data points. Left column: mean squared error (MSE) between predicted cleavage score and ground truth cleavage activity, averaged over all groups of identical guide-target sequence pairs. Right column: How faithful a model is to the differences in biological environment for a given pair within such a group is measured by the average proportion of the true cleavage activity difference which the model predicts. This is zero for purely sequence-based models and unity for an ideal predictor. To emphasise small deviations which preserve the rank of predicted cleavage activities, we use the cubic root as a sign-preserving nonlinearity and term this quantity relative difference. Right panel: Example distributions of relative pairwise difference for the two models. All underlying distributions are shown in Figure S9.