(174ai) Alignment-Free Prediction of HLA Class II Peptide Binding Using Deep Learning and Cross-Correlation of Amino Acid Properties

The precise prediction of antigen presentation by human leukocyte antigen (HLA) class II molecules is vital for the targeted development of immunotherapies and vaccines that activate CD4+ T cells. HLA class II molecules, structured as alpha and beta chain heterodimers, exhibit polymorphic variations predominantly around the peptide-binding domain, leading to diverse peptide binding specificities. Traditional neural network approaches for HLA class II antigen presentation predictions have primarily encoded HLA and peptide amino acid sequences independently, followed by simple concatenation for analysis. This study tests a new approach of pairwise sequence encoding for HLA-peptide pairs, utilizing cross-correlation to enhance the model's biological relevance regarding the interactive dynamics within the binding groove. We trained our neural network on extensive binding affinity datasets from the Immune Epitope Database (IEDB) across a wide array of HLA alleles. The pan-allele model is designed to predict bindings given the HLA-peptide pair, including alleles with scant or absent empirical binding data. Our model strives to capture the underlying binding motifs that are common to all alleles, although they can also account for allele-specific variation.