Background: When proteins are subjected to proteolytic digestion and analyzed by mass spectrometry using a method such as 2D LC MS/MS, only a portion of the proteotypic peptides associated with each protein will be observed. The ability to predict which peptides can and cannot potentially be observed for a particular experimental dataset has several important applications in proteomics research including calculation of peptide coverage in terms of potentially detectable peptides, systems biology analysis of data sets, and protein quantification.
Results: We have developed a methodology for constructing artificial neural networks that can be used to predict which peptides are potentially observable for a given set of experimental, instrumental, and analytical conditions for 2D LC MS/MS (a.k.a Multidimensional Protein Identification Technology [MudPIT]) datasets. Neural network classifiers constructed using this procedure for two MudPIT datasets exhibit 10-fold cross validation accuracy of about 80%. We show that a classifier constructed for one dataset has poor predictive performance with the other dataset, thus demonstrating the need for dataset specific classifiers. Classification results with each dataset are used to compute informative percent amino acid coverage statistics for each protein in terms of the predicted detectable peptides in addition to the percent coverage of the complete sequence. We also demonstrate the utility of predicted peptide observability for systems analysis to help determine if proteins that were expected but not observed generate sufficient peptides for detection.
Conclusion: Classifiers that accurately predict the likelihood of detecting proteotypic peptides by mass spectrometry provide proteomics researchers with powerful new approaches for data analysis. We demonstrate that the procedure we have developed for building a classifier based on an individual experimental data set results in classifiers with accuracy comparable to those reported in the literature based on large training sets collected from multiple experiments. Our approach allows the researcher to construct a classifier that is specific for the experimental, instrument, and analytical conditions of a single experiment and amenable to local, condition-specific, implementation. The resulting classifiers have application in a number of areas such as determination of peptide coverage for protein identification, pathway analysis, and protein quantification.