Accurate prediction of HLA class II antigen presentation across all loci using tailored data acquisition and refined machine learning

Abstract

Accurate prediction of antigen presentation by human leukocyte antigen (HLA) class II molecules is crucial for rational development of immunotherapies and vaccines targeting CD4 + T cell activation. So far, most prediction methods for HLA class II antigen presentation have focused on HLA-DR because of limited availability of immunopeptidomics data for HLA-DQ and HLA-DP while not taking into account alternative peptide binding modes. We present an update to the NetMHCIIpan prediction method, which closes the performance gap between all three HLA class II loci. We accomplish this by first integrating large immunopeptidomics datasets describing the HLA class II specificity space across all loci using a refined machine learning framework that accommodates inverted peptide binders. Next, we apply targeted immunopeptidomics assays to generate data that covers additional HLA-DP specificities. The final method, NetMHCIIpan-4.3, achieves high accuracy and molecular coverage across all HLA class II allotypes .

Proteogenomics guided identification of functional neoantigens in non-small cell lung cancer

“…The motif most closely matched the patients HLA-DRB1*03:01 allotype with peptide AYKMDMSLDDIIKLN predicted as a weakly binding peptide [30]. We identified 1135 missense mutations for patient A147 (Table S1) potentially yielding 6 neoantigens, representing 0.5% missense derived neoantigens, of which we observed one.…”

How to predict binding specificity and ligands for new MHC-II alleles with MixMHC2pred

“…An alternative approach to infer motifs is to use the pan-allele predictors, perform predictions for a large set of random peptides (at least 100 000 peptides), select the best scoring peptides (typically the top 1% best scoring peptides), and build the motif based on the predicted binding cores of these peptides. Such an approach is feasible with MixMHC2pred and NetMHCIIpan ( 21 ) in any species. Other MHC-II ligand predictors accommodating alleles without experimental ligands do not output the predicted binding cores and can moreover only be used for HLA-DR ( 25 ) or human/mouse ( 26 ) alleles.…”

“…Recent developments in MHC-II peptidomics have made it possible to obtain thousands of MHC-II ligands per sample in a single experiment in a high-throughput manner ( 6–18 ). Together with the development of powerful motif deconvolution algorithms to assign peptides to their cognate MHC-II alleles and identify peptide binding cores ( 12, 19 ), these data have enabled us and others to characterize the binding specificity of close to 100 alleles ( 6, 18, 20, 21 ). As of today, most of the common MHC-II alleles in human have well defined experimental motifs, but the coverage of experimental binding specificities falls short of the full diversity of human alleles (>10’000 alleles, as of November 2023 in IPD-IMGT/HLA database ( 1 )).…”

“…This limited coverage of the MHC-II alleles in terms of experimentally determined binding motifs both in human and other species has motivated the development of tools that can make predictions of MHC-II ligands even for alleles without any known ligand (so called pan-allele predictions) ( 6, 21 ). These tools capitalize on existing data to learn patterns in MHC-II sequences that can predict the preference for specific ligands.…”

Integrating Machine Learning-Enhanced Immunopeptidomics and SARS-CoV-2 Population-Scale Analyses Unveils Novel Antigenic Features for Next-Generation COVID-19 Vaccines

“…The rules of antigen processing and presentation (APP) have been studied, assessed, and incorporated into prediction algorithms to predict T-cell epitopes 53,54 . For instance, MHCflurry 55,56 and NetMHCpan [57][58][59] are two widely used algorithms that merge scores derived from APP properties. MHCflurry encompasses two essential predictors: an "antigen processing" predictor, which models MHC allele-independent effects like proteosomal cleavage, and a "presentation" predictor that combines processing predictions with HLA-peptide binding affinity (BA) predictions to yield a composite "presentation score (PS)".…”