Figure 3

Predictor performance dependence on training-testing domain sequence similarity. Leave 12% of domains out cross validation was performed with domains retained for training in each fold if their sequence similarity to all testing domains was less than a given threshold. This was performed for structure-based (blue) and sequence-based predictors (magenta). ROC and PR AUC scores were computed for each run and displayed in box plots according to training-testing domain sequence similarity threshold (top left and right). Based on significance testing using a one-tailed t-test, the mean structure-based predictor ROC and PR AUC scores are significantly higher than the sequence-based predictors scores when training-testing domain sequence similarity is < 0.7 (p-value < 0.029). The mean AUC scores for structure-based (blue) and sequence-based (magenta) predictors are plotted against sequence similarity threshold (bottom left and right).