Figure 2. Performance comparison of the tested classifiers for IVIG binding prediction on the test set peptides.

ROC analysis for our machine learning approach with an ensemble classifier (ML-advanced), the machine learning method of El-Manzalawy et al. (2008) [22], [23] and a PWM approach using an PWM derived from the training set. Both machine learning approaches were trained on the original training set (“original”: three times more non-binding than binding peptides) and on the balanced training set (“balanced”: equal number of binding and non-binding peptides) and finally applied on the test set. AUC values are indicated as well. Note that the curves based on the original and balanced training set of our ML-advanced method show almost complete overlap.