NeoTImmuML: A Machine Learning-Based Prediction Model for Human Tumor Neoantigen Immunogenicity

Yan ShaoShuguang GeRuizhe DongWei JiChaoran QinPengbo Wen^*

• School of Medicine Information and Engineering , Xuzhou Medical University, Xuzhou, China

Tumor neoantigens, due to their high specificity and immunogenicity, have become key targets for personalized cancer immunotherapies, such as mRNA vaccines and T cell therapies. However, identifying neoantigens and evaluating their immunogenicity remain challenging. These tasks often rely on time-consuming experimental validation, which greatly limits the efficiency of vaccine development. To address this problem, we introduced two main initiatives. First, we upgraded the TumorAgDB database by integrating publicly available neoantigen data published in the past two years, resulting in TumorAgDB2.0. Second, we developed NeoTImmuML, a weighted ensemble machine learning model for predicting neoantigen immunogenicity. Using data from TumorAgDB2.0, we calculated the physicochemical properties of peptides. We then systematically evaluated eight commonly used machine learning algorithms through five-fold cross-validation. Among them, LightGBM, XGBoost, and Random Forest performed best. These three models were combined into a weighted ensemble to build NeoTImmuML. The model showed strong generalization ability on both internal and external test sets. SHapley Additive Explanations (SHAP) feature importance analysis revealed that peptide hydrophobicity and length are key factors influencing immunogenicity prediction. NeoTImmuML is now integrated into the TumorAgDB2.0 platform. Overall, TumorAgDB2.0 provides a comprehensive data resource for neoantigen research. NeoTImmuML offers an efficient and interpretable tool for predicting neoantigen immunogenicity. Together, they provide strong support for the design of personalized neoantigen vaccines and the development of cancer immunotherapy strategies.