Predictive performance in AUC on 12 human HLA MHC class I alleles (Peters data set) and on 14 HLA-DR MHC class II alleles (Wang similarity reduced SR dataset).

For MHC class I no significant difference is found in predicted performance between the NNAlign, SMM and ANN method (p>0.5, binomial test). The values for the SMM and ANN methods were taken from Peters et al. [27]. The method was trained using a fixed motif length of 9 corresponding to the peptide length, and constructing a network ensemble with multiple architectures using respectively 2,3,4,5 and 7 hidden neurons. Performance was measured in cross-validation, training each network for a fixed number of 500 iterations per sequence.The different MHC class II prediction methods are NN-align[20], SMM-align[19], and Propred[31], [32]. NNAlign server is the method described here. Performance values for first 4 methods are taken from [25]. NNAlign was trained with a motif length of 9, flanking regions of 3 amino acids, Blosum encoding including peptide length and flanking region length, and an ensemble of 2, 3, 5, 9 and 12 hidden neurons for each of 10 initial random configurations.In bold is highlighted the best performing method for each MHC allele. The column # gives the number of the peptides in the data set for the given allele.