Weakly supervised peptide-TCR binding prediction facilitates neoantigen identification.

Abstract

The identification of T cell neoantigens is fundamental and computationally challenging in tumor immunotherapy study. Current prediction methods mainly focus on peptide properties, human leukocyte antigen (HLA) binding affinity, or single peptide-major histocompatibility complex-T cell receptor (pMHC-TCR) interactions, often overlooking the patient-specific TCR profile in evaluating neoantigen immunogenicity. This limited scope has constrained the performance and application of these tools in real-world settings for neoantigen identification. To address these limitations, we developed "TCRBagger," a weakly supervised learning framework that uses the bagging of sample-specific TCR profiles to enhance personalized neoantigen identification. TCRBagger integrates three learning strategies-self-supervised, denoising, and multi-instance learning (MIL)-for modeling peptide-TCR binding to identify immunogenic neoantigens. Our comprehensive tests and applications reveal that TCRBagger outperforms existing tools by modeling peptide-TCR profile interactions, accordingly enhancing the capability of immunogenic neoantigen identification. Collectively, TCRBagger provides an unprecedented perspective and methodology for modeling the interaction between a peptide and patient-specific TCR profiles, facilitating neoantigen identification for personalized tumor immunotherapy. A record of this paper's Transparent Peer Review process is included in the supplemental information.