Statistical comparison of established T-cell epitope predictors against a large database of human and murine antigens

Identification of T-cell epitopes within a protein antigen is an important tool in vaccine design. The T-cell epitope prediction schemes often are exploited by workers but have proved unreliable in comparison with experimental techniques. We compared published T-cell epitope predictors against two databases of human and murine T-cell epitopes. Each predictor was assessed against random cyclic permutations of epitopes in order to determine significance. Predictor performance was expressed in terms of two parameters, specificity and sensitivity. Specificity is an expression of the quality of predictions, whereas sensitivity is an expression of the quantity of epitopes predicted. Against the human data set, the strip-of-hydrophobic helix algorithm [Stille et al., Molec. Immun. 24, 1021-1027 (1987)] was the only significant predictor (p < 0.05), whereas against murine data only, the Roth2 pattern [Rothbard and Taylor, EMBO J. 7, 93-100 (1988)] was significant (p < 0.05). Not only were the majority of algorithms no better than random against both data sets, against the murine data two schemes were significant (p < 0.05) anti-predictors. This report indicates which predictors are relevant statistically and is the first to describe anti-predictors which can themselves be useful in the identification of T-cell epitopes.