Structure modeling and specificity of peptide-MHC class I interactions using geometric deep learning

Major Histocompatibility Complex (MHC) plays a major role in the adaptive immune response by recognizing foreign proteins through binding to their peptides. In humans alone there are several hundred different MHC alleles, where each allele binds a specific subset of peptides. The peptide-MHC complex on a cell surface is identified by a T-cell receptor (TCR) and this binding invokes an immune response. Therefore, predicting the binding specificity of peptide-MHC pairs is necessary for understanding the immune recognition mechanism. Here, we develop an end-to-end novel deep learning model, MHCfold, that consists of structure and specificity prediction modules for simultaneous modeling of peptide-MHC class I (pMHCI) complexes and prediction of their specificity based on their modeled structure. MHCfold produces highly accurate structures of pMHCI complexes with mean Cα RMSD of 0.98Å and 1.50Å for the MHC α chain and the peptide, respectively. The binding specificity is also predicted with high accuracy (mean AUC of 0.94). Furthermore, the structure modeling component is orders of magnitudes faster than state-of-the-art methods (modeling of 100,000 pMHCI pairs in four hours on a standard computer), enabling high-throughput applications for large immunopeptidomics datasets. While peptide-MHC specificity can be accurately predicted from the sequence alone, TCR specificity prediction likely requires modeling of the 3D structures. We anticipate our model can be further used in structure-based prediction of TCR specificity. MHCfold is available @ https://github.com/dina-lab3D/MHCfold