Fig. 3: The framework and performance of model-based approach for multi-allelic immunopeptidomics deconvolution.

a The framework of the deconvolution method to convert multi-allelic immunopeptidomics data into pseudo-mono-allelic. Using a mono-allelic model, predictions were made for all HLA alleles (up to six) of each sample. To make the predicted scores for different alleles comparable, raw scores were calibrated using percent rank values from a background set of 500,000 random peptides. Peptide was assigned to the allele with the lowest rank (best binder). b Benchmark comparison of our approach, NetMHCpan-4.1 and MixMHCpred 2.2, to retrieve HLA-bound peptides observed in patient-derived tumor datasets. AUROC, AUPRC, and PPV stratified by samples (n = 47) were calculated. c Mean AUROC, AUPRC, and PPV values with 95% confidence interval (CI) stratified by both samples and epitope length (n = 317) were calculated, followed by a two-tailed Wilcoxon signed-rank test for adjusted P-values. d Average Pearson’s correlation coefficient (PCC) among alleles (n = 24) for HLA binding motifs identified by single-allelic ligands and revealed by our approach, NetMHCpan-4.1, and MixMHCpred 2.2. Two-tailed Wilcoxon signed-rank test was used for the calculation of P-values. Bars represent means and error bars are 95% CIs. e Sequence logos of binding motif for HLA-1 alleles revealed by our approach from 47 multi-allelic immunopeptidomics samples compared to that from external mono-allelic ligands. Among the two logos of the same HLA allele, the left one was obtained by deconvolution, and the right one originated from mono-allelic ligands. Source data are provided as a Source Data file.