T cell immunogenicity prediction of peptides presented in MHC class I molecules

CARRI, IBEL; GARCIA ALVAREZ, HELI MAGALÍ; NIELSEN, MORTEN

Corrientes

Congreso; 12vo Congreso Argentino de Bioinformática y Biología Computacional; 2022

Background: Epitopes are defined as the region of an antigenic protein recognized by the immune system. Epitope identification is relevant to the study of immune responses and is crucial for vaccine development. In the immune response mediated by CD8+ T-cells, one of the most critical steps to define T-cell recognition or immunogenicity is peptide-MHC binding. Approximately, 1 in 200 peptides are recognized by the MHC complex and presented on the cellular surface. For this reason, multiple predictive tools have been developed to predict MHC binding with relative success. However, to elicit an immune response, the peptide not only needs to be presented on the cell surface but also be identified by the corresponding T-cell. For this reason, relying on peptide binding is not enough to efficiently predict immune response, especially when exploring large and complex proteomes such as parasites or cancer. The scientific community has identified that the aminoacidic composition of peptides influences immunogenicity, but predictive tools based on this information have limited performance and epitope prediction remain an unsolved challenge. This is mainly due to the combination of data scarcity and the complexity of the mechanism of T-cell recognition. In this work, we propose a novel machine learning-based tool that integrates the peptide sequences with several calculated features from the peptide and its source protein to improve immunogenicity prediction. Results: We obtained experimentally validated epitopes, non-epitopes, and their source proteins from the Immune Epitope Database. Only peptides with good binding to MHC were included to prevent the model to learn this known related feature alone. We included epitopes without the cognate MHC allele information and reconstructed the MHC specificity to enlarge the dataset. Proteic and immunological features were calculated based on peptide and protein sequences of this dataset. Proteic features, especially the ones related to protein structure, were demonstrated to correlate with immunogenicity. Different combinations of said features will be combined in a convolutional neural network. Conclusions: The analysis presented indicates that peptide and protein sequences contain complementary information related to determinants of immunogenicity, which combined may improve immunogenicity prediction and reveal rules defining T cell recognition and activation.