Leveraging attention-based deep multiple instance and multiple task learning for improved neoepitope identification.

Abstract

Accurate prediction of major histocompatibility complex class I (MHC class I) neoepitopes is crucial for personalized cancer immunotherapy. Current methods struggle with predicting ligand presentation for multiple alleles and identifying neoepitopes. We introduce NeoMHCI, a deep learning model that combines attention-based multiple instance learning (MIL) and multi-task learning for precise MHC class I neoepitope identification. NeoMHCI uses MIL to generate high-quality peptide embeddings with multiple MHC class I molecules and enhances immunogenicity prioritization through fine-tuning. Analyses on benchmark datasets show NeoMHCI outperforms existing methods, achieving an area under the receiver operating characteristic curve of 0.948 and an area under the precision-recall curve of 0.496 on unobserved multi-allele ligand presentation prediction and the highest top-5 accuracy (42.3%) for neoepitope recognition, indicating potential for personalized vaccines and therapies. A record of this paper's transparent peer review process is included in the supplemental information.