Hierarchical progressive learning for zero-shot peptide-HLA binding prediction and automated antigenic peptide design

Predicting the binding of peptides to human leukocyte antigen (HLA) alleles is critical for identifying epitopes initiating immune responses in vaccine development and immunotherapies. Despite the progress in predicting binding peptides of 1% HLA class-I alleles with in silico prediction tools, it remains challenging to predict binding peptides for alleles lacking epitopes. To achieve the binding prediction on those zero-shot alleles, we developed a hierarchical progressive learning (HPL) framework that progressively learns sequence patterns of specific peptide-HLA complexes. HPL models improve the prediction performance for zero-shot HLA class-I alleles by 60.8% (1,414.0% for non-classical ones). We further developed an automated peptide mutation search program, which automatically identifies binding peptides for any target HLA class-I allele and mutates weak-binding or non-binding peptides to enhance binding affinity under the guidance of HPL models. Our program generated high-affinity binding candidates of any target allele with limited mutations in more than 38.1% of testing cases.