BeitAI-pHLA: Multiallele Peptide-HLA Class I Binding Prediction Using Protein Language Model and Multi-Instance Learning.

Human leukocyte antigen (HLA) molecules participate in cellular immune responses by binding to peptide fragments derived from antigens. Exploring this process is crucial to understanding the mechanisms and underlying factors that regulate the cellular immune system. Due to the extreme polymorphism of HLA, the peptides obtained from mass spectrometry of eluted ligand experiments typically correspond to multiple HLA alleles. They are thus poly-specific, providing various options for HLA assignments. This introduces notable challenges for accurate peptide-HLA assignment and interpretation. To address this limitation, we present BeitAI-pHLA, a deep learning framework for predicting peptide-HLA class I binding by integrating protein language model embeddings with an attention-based multi-instance learning mechanism. Our evaluations on benchmark and external datasets indicate that BeitAI-pHLA markedly outperforms other advanced prediction tools across multiple external validation datasets, achieving an area under the precision-recall curve of 0.967 on the multiallele dataset and maintaining robust performance under varying negative peptide proportions. Furthermore, BeitAI-pHLA demonstrates superior motif deconvolution capability, with higher motif similarity (13.6% increase in position-specific scoring matrix correlation coefficient), and achieves the best area under the precision-recall curve in immunogenic neoepitope prediction, offering improved prioritization of candidate peptides for immunogenicity evaluation. BeitAI-pHLA provides a highly accurate tool for binding prediction and deconvolution in multiallele immunopeptidomics data, offering considerable potential for biomedical research and clinical translation. The BeitAI-pHLA prediction tool is accessible at https://phla.kindstarbiotech.com/, and the source code is available at https://github.com/KindstarGlobalInstitute/BeitAI-pHLA.