Evaluation of classifiers and feature types for superclass prediction.

<p>The classification performance of representative and widely used machine learning methods incorporating different features for superclass prediction was assessed my means of threshold-averaged ROC curves obtained from stratified 4×4-fold nested cross-validation. The differently colored curves correspond to distinct classification methods (see legend). For each classifier the area under the curve (AUC) is denoted. ROC curves were obtained from classifiers incorporating (<b>A</b>) our novel bit score percentile features, (<b>B</b>) <i>k</i>-mer features (<b>C</b>) PSSM profile features (<b>D</b>) functional domain features and (<b>E</b>) pseudo amino acid features.