Prediction Of Theoretical Mass Spectra Based On Deep Belief Networks

Proteomics is an important research in human life science,and protein identification is the most basic study of proteomics.Tandem mass spectrometry is a major technology for protein identification.It uses fragmentation techniques such as collision-induced dissociation to break peptides and generate MS/MS spectra containing the information of fragment ions.Predicting these tandem mass spectra signals and understanding this fragmentation process is important for high-throughput protein identification.The database search method is the most common method for identifying peptide sequences from tandem mass spectra,theoretical mass spectra prediction is an important step in this method,accurate theoretical mass spectra prediction can improve the credibility of peptide sequence identification.The intensity of fragmentation peaks in tandem mass spectra is influenced by the physical and chemical properties associated with various peptide sequences,we can use these physical and chemical properties to model the strength of the fragment ions,then predict theoretical mass spectra.In this paper,we used deep belief network algorithm to model intensity information in tandem mass spectra.Modeling process consists of three parts,namely data processing,model training and result analysis.In data processing part,we used the software of pFind,Mascot,Comet to search the data and get the result files.The PSMs of the identified results were extracted,and then we removed redundant PSMs,computed the FDR.Finally,highly reliable and non-redundant PSMs were obtained,and then we extracted the properties and intensity of ions for model training.In model training part,cross validation was used for parameter selection.In analysis part,we used the Pearson Correlation Coefficient as evaluation standard to analysis prediction results.The analysis showed that the theoretical spectra predicted by the model had a high similarity with the experimental spectra.