Classification of MHC I Proteins According to Their Ligand-Type Specificity

Major histocompatibility complex class I (MHC I) molecules belong to a large and diverse protein superfamily whose families can be divided in three groups according to the type of ligands that they can accommodate (ligand-type specificity): peptides, lipids or none. Here, we assembled a dataset of MHC I proteins of known ligand-type specificity (MHCI dataset) and trained k-nearest neighbor and support vector machine algorithms. In cross-validation, the resulting classifiers predicted the ligand-type specificity of MHC I molecules with an accuracy ≥ 99%, using solely their amino acid composition. By holding out entire MHC I families prior to model building, we proved that ML-based classifiers trained on amino acid composition are capable of predicting the ligand-type specificity of MHC I molecules unrelated to those used for model building. Moreover, they are superior to BLAST at predicting the class of MHC I molecules that do not bind any ligand.