Supplementary Figure 2: The Prosit deep learning model and its training.

(a) Overview of the neural network architecture of the fragment ion intensity prediction model. The model takes precursor charge, normalized collision energy and the peptide sequence as input. First, for every input a specific encoder is trained, consisting of one dense layer for precursor charge and normalized collision energy. The encoder for the peptide sequence is split in an embedding layer connected to 2 bi-directional recurrent neural networks (BDN) with gated recurrent memory (GRU) units and an attention layer. Both encoder representations are element-wise multiplied for a fixed size latent space representation. The decoder for fragment ion intensity prediction consists of one bidirectional GRU resulting in 6 predictions for up to 29 fragmentation positions. The indexed retention time (iRT) model uses the same encoder but dense layers as decoder. (b) Model performance for 5 random splits of the ProteomeTools data into Training, Test and Holdout. The main panel shows best performing models from 5 random splits of the data. The inset details the median models error with intervals (shaded regions) ranging from the best performing model to the worst performing model over the 5 splits for Training Test and Holdout. (c) Comparison of Pearson correlation and normalized spectral contrast angle (short spectral angle) as measures for spectral similarity between predicted and measured spectra contained in the holdout set for fragment ion intensity prediction.