Cross validation classification results on the H1-hESC training data set.

We show the averaged results of a 10-fold cross validation experiment on . For each , we plot the sensitivity (for a specificity of 99%) against the number of leaves. Error bars show double standard error. We observe that the sensitivity increases with model complexity and reaches a maximum at approximately 120 leaves, which corresponds to . With further increasing complexity, the sensitivity remains comparatively stable until it starts to drop when the model has more than 1,000 leaves. We observe that taking into account intra-motif dependencies improves the classification accuracy up to the point where the model is too complex, resulting in overfitting.