Peptide Identification from Tandem Mass Spectrometry Data Using Convolutional Neural Networks

Mass spectrometry analysis of peptides obtained by proteolysis is one of the most common approaches for identifying proteins in biological samples. An essential part of the analysis is database searching where each mass spectrum is compared against a database of either theoretically or experimentally computed peptide spectra. Due to physiochemical factors, mass spectra contain a lot of noise which affects database searching. Auxiliary peaks are usually considered as noise and excluded from the input spectra. However, many of these peaks contain useful information which could be exploited. In this paper, we introduce an approach to automatically learn the importance of each peak in a spectrum that utilizes convolutional neural networks. The model is trained to reconstruct the peak positions in the input spectra from their auxiliary peaks. By forcing center weight of the convolution kernel to zero we managed to avoid overfitting and learn the weights that correspond to the importance of auxiliary peaks. As a result, we can decrease the intensities for noise peaks which significantly improves performance. Experiments on publicly available human proteome data show that our method yields 20-50% more identified peptides than standard approaches at a given false discovery rate. Keywords: tandem mass spectrometry; proteomics; database searching; neural networks;