Comparative study of the structural properties of natural major histocompatibility ligands and T-cell epitopes

GARCIA ALVAREZ, HELI MAGALÍ; CARRI, IBEL; GLAVINA, JULIANA; LÁZÁR, TÁMAS; NIELSEN, MORTEN; TOMPA, PETER; CHEMES, LUCÍA

Bratislava

Conferencia; Non-globular proteins in the era of Machine Learning; 2023

Machine Learning for non globular proteins

Antigen processing and presentation on the cell surface by the major histocompatibility complex (MHC) are crucial events for adaptive immunity, leading to potential T-cell antigen recognition and activation of the immune response. In this work, we focus on one of the two main antigen processing pathways: the endogenous pathway, through which short peptides from viral and self-proteins are presented by MHC class I molecules.With the advent of immuno-peptidomics, new massive amounts of data were generated, allowing a more precise characterization of the natural MHC ligandome. Current state-of-the-art peptide-MHC binding predictors, such as NetMHCpan or MixMHCpred, make use of these large ligand datasets, achieving a high predictive power. Nevertheless, prediction of T-cell epitopes i.e., immunogenic peptides presented on MHC, is far from optimal. The complex rules defining MHC presentation, the vast repertoire of T-cell receptors (TCR), and our limited understanding of the interactions between TCRs and peptide-MHCs make this task an open challenge.In the present work, we aim to investigate the structural properties of natural MHC ligands and T-cell epitopes, mapping them back to their source proteins. This will allow us to gain new insights into how antigen processing impacts the likelihood of MHC presentation and T-cell activation, respectively. We hypothesize that 3D peptide properties in their source protein context, such as secondary structure, disorder, and solvent accessibility, are related to the efficiency of protein degradation. Indeed, this could contribute to explain why some predicted MHC strong binders are not natural ligands, or even further what differentiates natural ligands from epitopes.To achieve this objective, we took advantage of the large public immuno-peptidomics and epitope datasets. To analyze the structural properties we used AlphaFold2 and IUPred predictions of the human and viral proteomes. As a baseline for our study, we generated a pool of artificial negative peptides with MHC-like binding motifs using the neural network-based peptide-MHC binding predictor, NetMHCpan-4.1. In line with previous studies in the literature, our preliminary results indicate that self-ligands are enriched in alpha-helices in comparison to our baseline. In contrast, viral peptides arise predominantly from disordered protein regions in comparison to self-ligands. This result supports the idea that activation of the immunoproteasome in the context of a viral infection leads to differential antigen processing. In the future, our results would improve machine learning-based viral epitope predictors by including this novel feature.