Prediction-based reduction of the search space in metaproteomics

Metaproteomics search databases typically take on enormous sizes since the a priori unknown composition of metaproteomics samples requires the inclusion of proteomes of hundreds to thousands of species that could potentially be found in the samples. A major consequence is that the identification rate in metaproteomics experiments remains drastically below the identification rate in single-species proteomics. Therefore, reducing the database size will not only decrease computation time, but can simultaneously increase identification rate. To reduce database size, I used predictions from the machine learning algorithm CP-DT. This algorithm, originally intended to predict likely tryptic cleavage sites based on an ensemble of decision trees, has been shown to also be a useful predictor of the likelihood of observing a given peptide in a proteomics experiment. Indeed, if a large database (1.85 million protein sequences) is in silico digested using CP-DT, most peptides are marked as highly unlikely to be observed by the mass spectrometer. Moreover, if the peptide search space is reduced to only the top-35% scoring peptides according to CP-DT, more than 95% of the peptides that were actually observed by the mass spectrometer, are recovered. From these results I can conclude that the search space can be drastically reduced using CP-DT. Ongoing work will show if this reduction in search space will lead to an increased identification rate, while keeping the false discovery rate (FDR) under control