A Gradient Boosting Machine Learning Model for Neo-epitopes Classification

Neo-epitopes resulting from cancer specific variants are attractive targets for personalized cancer immunotherapy. The main challenge in using neo-epitopes for cancer immunotherapy is determining whether a neo-epitope will result in a tumor rejection response. Using the binding affinity of the neo-epitope to the MHC molecule or the difference between its binding affinity and that of its wildtype counterpart used to be the most common criteria for selecting neo-epitopes used in therapy. However, more recent research efforts use more features of the neo-epitope to classify it as immunogenic or not. In this work we present and test four different machine learning models for neo-epitope classification. Two models take as input on a set of numeric features, one model takes the sequences of the epitopes and the HLA allele as input, and the last one is a hybrid model that uses both numeric and sequence features. We compare the four models we present and conclude that the numeric features based Gradient Boosting model, ImmuneBoost, is the best model among the four. We then compare ImmuneBoost with two existing models, one of which uses numeric features and the other uses sequence features. We show that ImmuneBoost has higher precision and specificity but lower recall, compared to the other two models.