Figure 2

SVM performance dependence on testing and nearest training neighbour sequence similarity. Using leave one domain out cross validation (top), domain specific ROC and Precision/Recall AUC scores for SVM (blue) and nearest neighbour predictor (black) were grouped according to a given testing domain's similarity to its nearest training neighbour. The same was done for peptides using leave one peptide out cross validation (bottom). The similarity between two domains was calculated as the percentage of matched residues between their binding site sequences. The similarity between two peptides was calculated as the percentage of matched residues. Numbers in parentheses indicate the number of domains or peptides in each boxplot.