Despite the availability of major histocompatibility complex (MHC)-binding peptide prediction algorithms, the development of T-cell vaccines against pathogen and tumor antigens remains challenged by inefficient identification of immunogenic epitopes. CD8(+) T cells must distinguish immunogenic epitopes from nonimmunogenic self peptides to respond effectively against an antigen without endangering the viability of the host. Because this discrimination is fundamental to our understanding of immune recognition and critical for rational vaccine design, we interrogated the biochemical properties of 9,888 MHC class I peptides. We identified a strong bias toward hydrophobic amino acids at T-cell receptor contact residues within immunogenic epitopes of MHC allomorphs, which permitted us to develop and train a hydrophobicity-based artificial neural network (ANN-Hydro) to predict immunogenic epitopes. The immunogenicity model was validated in a blinded in vivo overlapping epitope discovery study of 364 peptides from three HIV-1 Gag protein variants. Applying the ANN-Hydro model on existing peptide-MHC algorithms consistently reduced the number of candidate peptides across multiple antigens and may provide a correlate with immunodominance. Hydrophobicity of TCR contact residues is a hallmark of immunogenic epitopes and marks a step toward eliminating the need for empirical epitope testing for vaccine development.
Keywords: MHC class I; T cell; epitope prediction; immunogenicity; nonself; vaccine.