Deep learning boosts sensitivity of mass spectrometry-based immunopeptidomics

Accurate prediction of relative ion intensities of tandem mass spectra is a long sought goal in mass spectrometry based proteomics. The authors have previously released a deep-learning based tool Prosit. In this study they have updated this resource by training the model with new mass spec experimental data of ~300k non-tryptic peptides. They show that the prediction of non-tryptic peptide fragmentation spectra improved. The new model was then applied toward identifying and rescoring MHC-bound immunopeptides which improved identification rates. Overall this study looks promising and should improve upon a useful resource. I have the following comments: