Single feature class performance.

Comparison of predictive performances of single feature classes. All values are taken from SVM predictions with feature files that only contain features belonging to the given class. Each prediction is an average of a ten-fold cross-validation with ten repetitions. The table shows the accuracy (ACC) and Matthews correlation coefficient (MCC) for each cell type and each feature class and is sorted by average MCC. Please note that the underlying data is imbalanced (because CpG islands tend to be unmethylated) and the average accuracy when assigning all CpG islands the unmethylated state is 0.71.