Comparison of various validated predictors of peptide-HLA binding.

Experimental peptide-HLA binding data was used to develop artificial neural networks. The numbers given in the table are the Pearson correlation coefficients between the logarithmically transformed predicted binding affinities (KD values) and logarithm transformed observed binding affinities (KD values). In bold are highlighted the maximum values in each column. (A) illustrates how poorly populated HLA molecules are more accurately predicted by the pan-specific leave-one molecule-out (Pan) predictor than by any of the conventional single allele predictors, even those generated using the data for the molecule in question. (B) illustrates that the pan-specific Pan predictor is only accurate when it has been trained on well-populated and relevant data. (C) illustrates that the pan-specific Pan predictor is inaccurate when no relevant data was included in the training sets. (D) illustrates the average performance for the HLA-A and �CB locus molecules including random negative data. Note, only non-supertype representative alleles are included in the average. The predictors are Pan: the pan-specific ANN trained on data emanate from all members of the locus in question (i.e. HLA-A or �CB) except for the member in question; Self: The most stringent comparison would be to use cross-validated ANN generated using data from the member in question, Neighbor: In the absence of self data, the next best alternatives would be to use cross-validated ANN generated using data from the most closely related member by BLOSUM comparison of the HLA-A (-or-B) pseudo-sequences, or Supertype: use cross-validated ANN generated using data from the member representing the supertype.