Reliability of Computer Algorithm-Based Binding Predictions for the Identification of Leukemia Antigens.

Abstract

Major Histocompatibility Complex class I (MHC-I) molecules present antigenic peptides to T cells on the cell surface as a prerequisite for stimulating cytotoxic T cell response. Thus, the ability to reliably identify the peptides that can bind to MHC molecules is of practical importance for rapid vaccine development. Several computer-based prediction methods have been applied to study the interaction of MHC class I/peptide binding. Here we have compared three of the most commonly used predictive algorithms BIMAS, SYFPEITHI and Rankpep with actual binding of HLA-A*0201 peptides in vitro. Forty six HLA-A*0201 peptides were selected from several target oncoproteins: Wilms' tumor (WT1), native and imatinib- mutated bcr-abl p210 and JAK2 protein. Experimental peptide binding to HLA-A*0201 was assessed using a MHC stabilization assay on T2, TAP deficient cells. Peptides were considered to show positive in vitro binding if the mean fluorescence was at least 50 % of the binding of a high affinity reference peptide. Peptides qualified as positive in vitro if the BIMAS score was ≥ 100, the SYFPEITHI score ranked ≥ 24 or the Rankpep was ≥ 50. Results are summarized below:BIMASSYFPEITHIRANKPEPSensitivity84 %72 %60 %Specificity76 %71 %81 %Positive Predictive Value84 %72 %60 %Negative Predictive Value80 %68 %63 %Combining two or more computer methods did not appear to improve the predictive value. In conclusion, of the three predictive algorithms, the best correspondence with the actual MHC binding was demonstrated with the BIMAS algorithm. Predictive computer algorithms are important for preselection of potential T-cell epitope candidates for the application in vaccine design.