Penalized regression analysis identifies features of the peptide:MHC class I interaction and mRNA expression as key to prioritizing neoantigen immunogenicity

Abstract

Abstract Current approaches to prioritization of tumor-specific neoantigens have low sensitivity and specificity, directly impacting the development of personalized vaccines and the prediction of immunotherapy efficacy. Many biological features have been suggested to contribute to accurate neoantigen prioritization. We calculated predictors for a comprehensive set of biological features for presentation and recognition of neoantigens on a validated neoantigen dataset. One thousand random, evenly-split combinations of immunogenic and non-immunogenic neoantigens were analyzed using penalized regression (lasso) for variable selection. This analysis isolated the peptide:MHC class I dissociation constant, binding stability, and mRNA expression as the key predictive variables of tumor-specific neoantigens. Based on these results, we fit a logistic regression model and tested its performance on three independent datasets. We compared our results to the models by Wells et al., Łuksza et al., and Zhou et al. Our model has improved performance as measured by the area under the receiver operator curve (AUC) compared to Łuksza and Zhou et al. Our model selects the same terms as Wells et al. and performs equivalently. However, it has the advantage of providing a predicted immunogenicity score for each individual neoantigen. Individual antigen approaches are useful for prioritizing vaccine candidates and predicting response to immunotherapy. Future work will focus on the application of this predictor to patient response to immunotherapy.