Fig. 1: Model architecture and performance of DeepPhospho.
From: DeepPhospho accelerates DIA phosphoproteome profiling through in silico library generation
a The DeepPhospho deep learning architecture for indexed retention time (iRT) and fragment ion intensity prediction for any given phosphopeptide. Given a peptide sequence and the precursor charge as input, our model first uses a bi-LSTM network to compute an initial representation of all the amino acids, which are then refined by a Transformer module. The resulting global features are fed into a linear regressor network to generate predictions for fragment ion intensity and iRT. b Evaluation of DeepPhospho and three other models based on the distribution of Pearson correlation coefficient (PCC) and spectral contrast angle (SA) calculated between predicted and experimental MSMS spectra from two datasets. Median PCC and SA are indicated; n is the number of phosphopeptides in the test set. Boxplot center line, median; box limits, upper and lower quartiles; whiskers, 1.5× interquartile range. c Evaluation of DeepPhospho based on the correlation of predicted and experimental iRT values. Correlation coefficient of linear regression (R2) and median absolute error (MAE) are indicated. Source data for this figure are provided as a Source data file.
