RankMHC: Learning\nto Rank Class‑I Peptide-MHC\nStructural Models

The binding of peptides to class-I Major Histocompability\nComplex\n(MHC) receptors and their subsequent recognition downstream by T-cell\nreceptors are crucial processes for most multicellular organisms to\nbe able to fight various diseases. Thus, the identification of peptide\nantigens that can elicit an immune response is of immense importance\nfor developing successful therapies for bacterial and viral infections,\neven cancer. Recently, studies have demonstrated the importance of\npeptide-MHC (pMHC) structural analysis, with pMHC structural modeling\nmethods gradually becoming more popular in peptide antigen identification\nworkflows. Most of the pMHC structural modeling tools provide an ensemble\nof candidate peptide poses in the MHC-I cleft, each associated with\na score stemming from a scoring function, with the top scoring pose\nassumed to be the most representative of the ensemble. However, identifying\nthe binding mode, that is, the peptide pose from the ensemble that\nis closer to an unavailable native structure, is not trivial. Oftentimes,\nthe peptide poses characterized as best by a protein–ligand\nscoring function are not the ones that are the most representative\nof the actual structure. In this work, we frame the peptide binding\npose identification problem as a Learning-to-Rank (LTR) problem. We\npresent RankMHC, an LTR-based pMHC binding mode identification predictor,\nwhich is specifically trained to predict the most accurate ranking\nof an ensemble of pMHC conformations. RankMHC outperforms classical\npeptide-ligand scoring functions, as well as previous Machine Learning\n(ML)-based binding pose predictors. We further demonstrate that RankMHC\ncan be used with many pMHC structural modeling tools that use different\nstructural modeling protocols.