Peptide Retention Time Prediction using Artificial Neural Networks

Abstract: This thesis describes the development and evaluation of an artificial neural network, trained to predict the chromatographic retention times of peptides, based on their amino acid sequence. The purpose of accurately predicting retention times is to increase the number of protein identifications in shotgun proteomics and to improve targeted mass spectrometry experiment. The model presented in this thesis is a branched convolutional neural network (CNN) consisting of two convolutional layers, followed by three fully connected layers, all with leaky rectifier as the activation function. Each amino acid sequence is represented by a 20-by-20 matrix X, with each row corresponding to a certain amino acid and the columns representing the position of the amino acid in the peptide. This model achieves a RMSE corresponding to 3.8% of the total running time of the liquid chromatography and a 95 % confidence interval proportional to 14% of the running time, when trained on 20 000 unique peptides from a yeast sample. The CNN predicts retention times slightly more accurately than the software ELUDE when trained on a larger dataset, yet ELUDE performs better on smaller datasets. The CNN does however have a considerable shorter training time.