Cleavage score estimation.

To understand amino acid preference for HLA-II peptide cleavage, we compared amino acid frequency around HLA-II peptide cleavage sties with a background distribution. Cleavage sites included six amino acids upstream of the HLA-II peptide N terminus (−6 to −1) and six amino acids downstream of the HLA-II peptide C terminus (+1 to +6) assuming a N terminus to C terminus direction. We randomly generated an equal number of gene-matched and length-matched peptides on the basis of our HLA-II peptides, and the amino acid frequency from the same cleavage sites was used as the background distribution. Enrichment and depletion level were calculated as HLA-II peptide amino acid i frequency at j position divided by background distribution amino acid i frequency at j position (Fig. 2b and Supplementary Fig. 4). The analysis was done for MHC-I as well (Supplementary Fig. 4a).

We built a neural network to quantitatively estimate cleavage scores given the flanking regions of a query gene-peptide pair (Supplementary Fig. 2b). The model determines six amino acids upstream of the query peptide N terminus and six amino acids downstream of the query peptide C terminus with a human proteome dictionary, encodes these 12 amino acid sequences, processes them with hidden layers and outputs a probability score between 0 and 1 (score of the positive output neuron T). The neural network encodes amino acids using one-hot encoding as described before and contains two fully connected hidden layers of 32 neurons. To avoid overfitting when applying this model to lymphoma data, we trained the cleavage model on an independent dendritic cell line (MUTZ3) ligandome⁵⁴ (Supplementary Fig. 4d).