A Deep Learning-Driven Framework Integrating Organoid-Based Functional Validation Identifies Universal Neoantigens from Recurrent Glioma Mutations

Glioblastoma (GBM) is the most common malignant intracranial tumor in adults, with a median survival of only 16-20 months. Neoantigen therapy has shown advantages in the treatment of GBM, as it improves the immunosuppressive microenvironment within the tumor. However, the identification of truly immunogenic neoantigens remains a major challenge. Current computational prediction tools primarily focus on antigen presentation, while algorithms that incorporate T cell immunogenicity features remain limited. Furthermore, standard validation methods, such as ELISpot assays, lack physiological relevance and do not fully recapitulate the tumor microenvironment. Here, we developed a neoantigen prediction algorithm, TCRscore, based on publicly available datasets by integrating human leucocyte antigen binding and T-cell receptor (TCR) recognition features. Twenty-one patient-derived glioblastoma organoid models were established from isocitrate dehydrogenase wildtype tumors to validate performance of the algorithm. Predicted neoantigens were evaluated using ELISpot assays, flow cytometry, and in vitro killing assays based on organoid-T cell co-culture systems. TCRscore outperformed six existing tools in predicting immunogenic neoepitopes. The organoid models retained the key histological and transcriptomic features of parental tumors and provided an effective platform for functional validation. Co-culture assays confirmed that neoantigen-specific T cells could induce targeted killing in GBM organoids. In particular, the analysis identified that the recurrent PIK3R1G376R mutation contributed to a potential shared neoantigen in glioblastoma. Overall, by integrating TCRscore with organoid-based validation, this study provides a high-fidelity, high-quality GBM neoantigen database with significantly enhanced prediction accuracy.