MHCfovea integrates a supervised prediction module and an unsupervised summarization module to connect important residues to binding motifs.
First, the MHCfovea's predictor was trained on 150 observed alleles; 42 important positions were highlighted from MHC-I sequence (182 a.a.) using ScoreCAM. Next, we made predictions on 150 observed and 12,858 unobserved alleles against a peptide dataset (number: 254,742), and extracted positive predictions (score > 0.9) to generate the binding motif of an allele. Finally, after clustering the N- and C-terminal sub-motifs, we built hyper-motifs and the corresponding allele signatures based on 42 important positions to reveal the relation between binding motifs and MHC-I sequences.
MHCfovea takes MHC-I alleles (all alleles in the IPD-IMGT/HLA database (version 3.41.0) are available) and peptide sequences as inputs to predict the binding probability. For each queried allele, MHCfovea provides the cluster information and allele information of N- and C-terminal clusters respectively.