Fig. 5

Mismatch resolution and chromatin context. a Biases in genomic contexts. The dot plot shows the measured bias toward T with the G:T construct. Expressed genes, silent genes, and intergenic regions are the same as in Fig. 2c. b Biases in chromatin contexts. Same representation as in a for chromatin marks. The P-values of a two-tailed Student’s t test for coefficients in linear models are indicated if they are below 0.01. c Architecture of the artificial neural network. Vertical cartridges represent neuron layers with indicated dimensions. The blue arrows indicate the forward information flow. d Learning curves for the full data set. The median loss on the test set is shown in green for 100 trainings with random restart. The interval between the 1st and 99th percentile is shown by the green shaded area. The median loss for the null model without chromatin features is shown in brown, and percentiles are shown by the brown shaded area. e Learning curves for the local GC content. The repair outcome was replaced by the GC content in a 10 kb window around the reporter. f Learning curves for conflicts among UMIs. The only observations used for learning are UMI-PCR 24 h post-I-SceI induction