A Novel Machine Learning Model for Predicting Neoantigen Immunogenicity Using T Cell Assays and Binding Affinities

Cancer immunotherapy, particularly personalized treatments targeting neoantigens, has revolutionized oncology. Neoantigens—new proteins that form on cancer cells during certain mutations of tumor DNA—are crucial in these therapies, but predicting their immunogenicity remains challenging. Several machine-learning models have been made in the past to predict neoantigen immunogenicity using graphical neural networks (GNNs) and other algorithms. This study aimed to develop a new model to make predictions of the immunogenic response of novel neoantigens more accurate by improving on the framework of existing GNNs. The model architecture was altered for robustness, used early stopping to prevent overfitting, and Optuna for hyperparameter optimization. By modeling the correlation between binding affinities and immunogenicity within the GNN framework, the model achieved superior accuracy in predicting T cell responses, offering a significant advancement in personalized cancer immunotherapy.