Training in bad data only produces bad models. I have seen a lot of deep-learning models in proteomics and almost no research/talks about the data that was used for training.