This evaluation of MHC binding prediction methods available on the web cannot be considered a true measure of their value. This is due to multiple reasons described in detail in the accompanying manuscript. Most importantly, each method had very different amounts and quality of training data available when it was generated. Only by controlling both training and testing data can one make meaningful statements on the value of different prediction methods.
Also, the first returned prediction score for each website was used, and no attempts where made to optimize parameters or look for better output. This, for example, systematically underestimates the performance of SVMHC which has a more detailed output with scores for non-binding peptides available.