Abstract
The key roles of Major Histocompatibility Complex (MHC) Class I and II molecules in the immune system are well established. This study aims to develop a novel machine learning framework for predicting antigen peptide presentation by MHC Class I and II molecules. By integrating large-scale mass spectrometry data and other relevant data types, we present a prediction model OnmiMHC based on deep learning. We rigorously assessed its performance using an independent test set, OnmiMHC achieves a PR-AUC score of 0.854 and a TOP20%-PPV of 0.934 in the MHC-I task, which outperforms existing methods. Likewise, in the domain of MHC-II prediction, our model OnmiMHC exhibits a PR-AUC score of 0.606 and a TOP20%-PPV of 0.690, outperforming other baseline methods. These results demonstrate the superiority of our model OnmiMHC in accurately predicting peptide-MHC binding affinities across both MHC-I and MHC-II molecules. With its superior accuracy and predictive capability, our model not only excels in general predictive tasks but also achieves significant results in the prediction of neoantigens for specific cancer types. Particularly for Uterine Corpus Endometrial Carcinoma (UCEC), our model has successfully predicted neoantigens with a high binding probability to common human alleles. This discovery is of great significance for the development of personalized tumor vaccines targeting UCEC.