Abstract Mutations in cancer cells may lead to the formation of neo-epitopes potentially presented by both major histocompatibility complex (MHC) class I or II. These neo-epitopes may be recognized by CD8+ or CD4+ T-cells, and trigger an immune response. Only a small fraction of the neoepitopes will be displayed by the MHC class I or II. One of the challenges of cancer immunotherapy is therefore to predict which neoepitopes are susceptible to elicit a T-cell response. Software tools such as netMHC, MHC Flury and many others are over-predictive as the bulk part of the data used to train these methods are based on affinity assays. Several publications have indicated that stability assays may provide data that better correlate with epitope presentation by MHC. Prediction tools trained on stability assays may therefore be better at selecting neoepitopes resulting in more effective cancer vaccine design. We performed stability assay measurements for 10 MHC class I and 10 MHC class II alleles using a peptide scan library approach. In brief, random 9-mers where one position is known were used to measure stability of the peptide MHC complex. Next, the data were used to train a prediction tool, PrDx, that relies on a combination of different machine learning methods (random forest, feed forward neural networks and recurrent neural networks), of which the outputs are gathered in an assemble model. The model was then further trained with peptides predicted to bind with high stability, to the MHC alleles, until satisfactory performances were attained. To our surprise, PrDx showed new binding patterns for the alleles we trained. Although mostly similar to the binding patterns seen with affinity data trained method, the stability trained method is able to show new important positions in the binding patterns of the peptide-MHC complexes. Through retrospective analysis, our method seems able to select more accurately peptides susceptible to elicit a T-cell response, compared to state-of-the-art epitope prediction methods. Our results suggest that PrDx may be an attractive prediction tool for neo-epitopes discovery. Citation Format: Stephan Thorgrimsen, Sune Justesen, Nicolas Rapin. Development of prediction software PrDx, trained on peptide-MHC stability assays, shows new important positions in the binding patterns of the peptides-MHC I and II complexes abstract. In: Proceedings of the Fourth CRI-CIMT-EATI-AACR International Cancer Immunotherapy Conference: Translating Science into Survival; Sept 30-Oct 3, 2018; New York, NY. Philadelphia (PA): AACR; Cancer Immunol Res 2019;7(2 Suppl):Abstract nr B093.