T-cell epitopes, i.e., peptides capable of inducing a T-cell mediated immune response, represent suitable components for vaccines against infectious diseases and cancer. The development of accurate T-cell epitope prediction methods is thus of great interest to immunologists and the pharmaceutical industry. Whether a particular peptide is a T-cell epitope depends on the availability of (a) an MHC molecule capable of presenting the peptide on the cell surface and (b) a suitable T cell. In order to ensure self-tolerance of the immune system, T cells reactive to self-peptides are eliminated via negative selection processes. The composition of the Tcell repertoire thus depends on the host proteome. These complex dependencies along with a lack of data render Tcell epitope prediction a rather challenging problem. It is commonly reduced to the simpler MHC binding prediction problem. While state-of-the-art MHC binding prediction methods are highly accurate, the actual prediction of T-cell epitopes leaves room for improvement. Previously proposed approaches to T-cell epitope prediction do not take the dependencies on the host proteome into account but utilize peptide sequence information only. Their low prediction accuracies can be attributed to this limited view on T-cell reactivity and to a biased data basis. Moving from simple sequence-based predictors to predictors taking system-wide properties into account, we present a novel approach to T-cell epitope prediction combining sequence information and information on self-tolerance. In a thorough study on a small but unbiased data set, our method outperforms purely sequence-based predictors, indicating the validity of our approach. Copyright © 2011 ACM.