MHC Attention: Identifying HLA-E presented cancer antigens through deep learning and high-throughput screening.

HLA-E presented cancer peptides can be promising cancer therapy targets, as HLA-E is minimally polymorphic and widely expressed across human populations and cancer types. However, systematic discovery of cancer associated HLA-E peptides has been constrained by sparse training data and the technical difficulty of HLA-E immunopeptidomics. Here we develop an integrated HLA-E antigen discovery platform combining a deep learning prediction model, pooled mammalian cell screening, peptide-HLA-E stability validation, and mass spectrometry. We introduce MHC Attention, a neural network that learns allele-level attention over candidate MHC alleles in multi-allele immunopeptidomics datasets, enabling direct training on patient-derived MHC peptide data. Screening an approximately 6,000-peptide HLA-E library identified stable HLA-E-presented peptides and generated HLA-E-specific training data that improved prediction performance of MHC Attention. Combining our screening assays and improved prediction algorithm, we discovered novel HLA-E-presented cancer peptides, including candidates derived from ETV4, WT1, RNF43 and BMP8A, with orthogonal support from stability assays or immunopeptidomics. These results establish a scalable framework for HLA-E peptide target discovery and provide candidate targets for broadly applicable peptide-HLA-directed cancer immunotherapies. MHC Attention 2.0 can be accessed online via https://vcreate.io/mhcattention.