Dataset generation and preprocessing

Sequences of MHC-I binding peptides were obtained from the IEDB database³⁶ for the H-2D^b, H-2D^d, H-2D^q, H-2K^b, H-2K^d, H-2K^q, H-2L^d and H-2L^q haplotypes, although here we present the procedure and results of H-2K^b as a case. Given the different binding assessment methodologies considered in IEDB, elements were binarized by their MHC class I classification as positive or negative, per IEDB standards. The datasets, by entries accession number, are available at NAP-CNB.

Firstly, peptides deemed as antigenic were processed to extract their binding sites. These correspond to positive epitopes from IEDB as classified by their qualitative labels “Positive High”, “Positive Intermediate” and “Positive Low” for each MHC class I haplotype in mice irrespective of the assay type. A further selection criteria was to include only epitopes with protein identifications to generate negatives and resize the sequence to a given length. Consequently, sequences were aligned with its protein source through the Smith-Waterman algorithm³⁷ to obtain the remaining sequence as negative samples (Suppl. Fig. ¹). Additionally, epitope regions were extended through the original sequence to have a regular size (Suppl. Fig. ¹). In contrast with previous methods, a given prevalence (i.e., the fraction of the minority class) was not imposed on the dataset. In total, for H-2K^b, 4,828 peptide entries were processed into 251,049 sequences with 6714 positive entries and 244,225 negatives. A 10% split was used for test set generation. Concerning blind test data, IEDB datasets 1034799 and 1035276 were processed through the previous procedure and by the method described by¹⁵. Additional information concerning the dataset for each haplotype is available in the download section of NAP-CNB.

Further postprocessing was implemented with a majority vote algorithm that considered mutations to the most similar amino acid, given by the BLOSUM62 matrix³⁸, for each position. In other terms, a sequence modified its classification if there was a consensus among its most akin peptides.