Evaluating Cross-linking-driven integrative modeling in peptide-HLAII complexes prediction with insights for refining predictive accuracy

ABSTRACTIn silico prediction of peptide-HLAII (human leucocyte antigen class II) complexes has emerged as a crucial approach in bioinformatics for deciphering antigen presentation mechanisms. Several in silico tools have been developed to predict peptide binding to HLAII alleles, trying to deconvolute the intricate peptide-HLAII binding specificity. These approaches integrate bases from molecular modeling, machine learning, and bioinformatics to predict peptide-HLAII interactions. Initially, structure-based methods relying on molecular docking algorithms were widespread, utilizing structural data of HLAII molecules and peptides to infer plausible binding conformations. These methods often faced challenges in accuracy due to the dynamic nature of peptide-HLAII interactions. Besides, the high flexibility of peptide sidechains makes their placement into the HLA-binding site even more complex. In recent years, machine learning techniques have drawn attention to peptide-HLAII binding predictions. Supervised learning algorithms, such as support vector machines (SVMs), neural networks, and ensemble methods, have been considerably applied to discriminate patterns from large datasets of experimentally validated peptide-HLAII binding affinities (like Immune Epitope Data Base, IEDB) and more recently mass spectrometry- eluted ligands from MHC-associated peptide proteomics (MAPPs) assay. The role of experiment- assisted integrative modeling in aiding peptide-HLAII complexes prediction still needs to be clarified. In this work, we benchmarked the use of AlphaLink2 (AlphaFold2 + cross-links restraints) and compared it to AlphaFold2 Multimer in predicting correct peptide binding motifs. These results can pave the way to an integrated strategy for vaccine development and protein deimmunization or autoimmunity mitigation.