Interactions between peptide and MHC class II (pMHC-II) are central to T-cell recognition, yet predicting the binding core remains challenging due to the high polymorphism of MHC-II molecules and their ability to accommodate diverse peptide lengths. C...
Interactions between peptide and MHC class II (pMHC-II) are central to T-cell recognition, yet predicting the binding core remains challenging due to the high polymorphism of MHC-II molecules and their ability to accommodate diverse peptide lengths. Computational approaches have therefore become indispensable. Traditionally, sequence-based models such as NetMHCIIpan and DeepMHCII have been the mainstay, while more recently, AlphaFold-based structural prediction has emerged as a powerful alternative. In this work, we compared four representative methods— NetMHCIIpan, DeepMHCII, AlphaFold2 fine-tuned for peptide binding (AF2- FT), and AlphaFold3 (AF3)—with an emphasis on binding core prediction. AF3 and AF2-FT showed strong performance for positive binders, reaching recall values of 0.86 and 0.81, respectively, though both tended to overpredict binding for unbound peptides. NetMHCIIpan, on the other hand, was far more effective at identifying non-binders, achieving the highest negative recall (0.93), but its ability to capture positive binders was limited (0.44). DeepMHCII showed balanced yet unremarkable performance across the board. Interestingly, combining the strengths of both approaches yielded the best outcome: consensus strategies that integrated AlphaFold-based predictions for binder identification with NetMHCIIpan as a filter improved precision substantially, reaching 0.94 in known cases and 0.87 in previously unseen ones. Overall, our findings emphasize the complementarity of structural and sequence-based approaches. AlphaFold-based methods bring high sensitivity for true binders, whereas NetMHCIIpan provides robustness against false positives. Moving forward, improving the treatment of unbound peptides in structure-based models, and enhancing binding discrimination in sequence- based methods, will be key to more accurate pMHC-II prediction.