IncleMHC v2: Predicting MHC Binding Affinity and Eluted Ligand Presentation of Longer Peptides Using a Novel Transformer-Based Framework

The adaptive immune system relies on intracellular peptide presentation to T cells to detect infected or abnormal cells [1]. This process is mediated by the Major Histocompatibility Complex (MHC), molecules that display peptide fragments for immune surveillance [1]. In humans, MHC molecules are referred to as Human Leukocyte Antigens (HLA) [1]. Peptide-MHC prediction supports epitope identification and immunotherapy development [2], [3]. This study presents IncleMHC v.2, a novel multi-allele transformer-based tool for predicting binding affinity (BA) and eluted-ligand (EL) presentation of longer peptides with MHC Class I alleles. IncleMHC v.2 learns sequence length-dependent patterns across multiple MHC alleles and across peptide lengths. Five-fold cross-validation produced a mean PCC of 0.806 ± 0.005 and SCC of 0.746 ± 0.004 for BA predictions. Length-stratified analysis showed BA PCC values of 0.853, 0.801, 0.783, 0.770, and 0.713 for 8-, 9-, 10-, 11- and 12-mers. Performance noticeably decreased to PCCs of 0.443 and 0.540 for 13- and 14-mers. In contrast, EL prediction achieved an AUC of 0.994 across all peptide lengths. Per-allele EL analysis across 29 alleles produced a mean AUC of 0.995. Evaluation included 175 test alleles, including 19 not observed during training, supporting assessment across allele diversity. The dataset and five-fold cross-validation framework were adapted from NetMHCpan 4.2. These results suggest, but do not establish, stronger prediction performance for longer peptides (11–12-mers) than published NetMHCpan 4.2 results, while performance remained comparable across shorter lengths.