Reader Comments

Post a new comment on this article

Referee Comments: Referee 3

Posted by PLOS_ONE_Group on 03 Sep 2007 at 21:09 GMT

Reviewer 3's Review

“This manuscript shows a novel ANN-based predictive algorithm for the prediction of HLA class I binding peptides, a very important topic of immunology as applied to vaccines. It claims it can outperform similar predictive algorithms given their taking into consideration both the peptide binding data as well as the variant sequences in each HLA class I molecule. Data do seem to show a significant improvement in prediction accuracy, and are in line with previously observed However, this reviewer felt its predictive properties should be more extensively explored,in two ways:

i. Regarding retrospective comparison with peptides with known binding prperties, authors should calculate sensitivity (as related to false negatives) and specificity (as related to false positives)

ii. Regarding prospective validation (depicted in Fig 1) authors should test some more uncharacterized HLAs; they should also perform and show empirical binding assays from peptide sequences predicted not to bind that particular HLA class I molecule by the algorithm.”

N.B. These are the general comments made by the reviewer when reviewing the originally submitted version of this paper. The manuscript was revised before publication. Specific minor points addressed during revision of the paper are not shown.

RE: Referee Comments: Referee 3

mniel replied to PLOS_ONE_Group on 19 Sep 2007 at 11:27 GMT

ad i) Since our method allows for quantitative predictions of IC50 binding affinities for peptide MHC binding, the most informative validation measure of prediction accuracy is a correlation coefficient. If we were to calculate specificity/sensitivity values data would have to be classfied into binders versus non-binders, and the power of quantitative predictions would be lost. However, we did conduct a sensitivity/specificity validation in the paper using both endogenously presented peptides and known HIV immunogens. For details on this experiment I refer to the paper.

ad ii) We have conducted a large scale leave-one-molecule out benchmark, where peptide binding data for one HLA allele was excluded from the training of the pan-specific method. This type of benchmarking reflects the performance of the pan-specific method on uncharacterized HLA. This LOO validation was performed for all 42 HLA molecules included in the benchmark data set. In the LOO benchmark both peptide binders and non-binders were included.