Accurate peptide fragmentation predictions allow data driven approaches to replace and improve upon proteomics search engine scoring functions

Abstract

Motivation The use of post-processing tools to maximize the information gained from a proteomics search engine is widely accepted and used by the community, with the most notable example being Percolator—a semi-supervised machine learning model which learns a new scoring function for a given dataset. The usage of such tools is however bound to the search engine’s scoring scheme, which doesn’t always make full use of the intensity information present in a spectrum. We aim to show how this tool can be applied in such a way that maximizes the use of spectrum intensity information by leveraging another machine learning-based tool, MS2PIP. MS2PIP predicts fragment ion peak intensities. Results We show how comparing predicted intensities to annotated experimental spectra by calculating direct similarity metrics provides enough information for a tool such as Percolator to accurately separate two classes of peptide-to-spectrum matches. This approach allows using more information out of the data (compared with simpler intensity based metrics, like peak counting or explained intensities summing) while maintaining control of statistics such as the false discovery rate. Availability and implementation All of the code is available online at https://github.com/compomics/ms2rescore. Supplementary information Supplementary data are available at Bioinformatics online.

Lost and found: re-searching and re-scoring proteomics data aids the discovery of bacterial proteins and improves proteome coverage

“…Comparing the fragmentation of synthetic peptides can additionally be used to validate such one-hit wonders (Friedman et al 2017). Due to the recent development and accuracy of peak intensity prediction tools such as MS²PIP (Degroeve and Martens 2013) or Prosit ) -the latter trained on synthetic peptides -the comparison of predicted and empirical peptide fragmentation spectra is automated and can deliver additional PSM scoring metrics to be used by semi-automated machine learning tools such as Percolator Silva et al 2019). Notably, we used correlation to MS²PIP-predicted spectra before to discriminate high-confidence unannotated peptides in Arabidopsis thaliana (Willems et al 2017).…”

“…A concatenated target-decoy peptide database was searched using MS-GF+Percolator (Kall et al 2007;Granholm et al 2014). Next to MS-GF+ features, we relied on an additional set of 11 features, referred to as the 'auxiliary' feature set and including the experimental RT deviation (ΔRT) to predictions (Moruz et al 2010), the number of missed cleavages, features related to the number of matched b/y-ions, and features describing b/y-ion intensity correlation to those of MS²PIPpredicted spectra (Degroeve and Martens 2013;Silva et al 2019). A full description of the in total 34 features is provided in supplemental Table S1.…”

“…This model was used to predict the RT for all matched peptides and the absolute deviation of the predicted and experimental RT (in mins) was used as additional scoring feature in the auxiliary feature set (supplemental Table S1). In addition, the reScore algorithm (Silva et al 2019) was used to compare the fragment peak intensities of empirical spectra with those predicted MS²PIP (algorithm v20190312, HCD trained model v20190107) (Degroeve and Martens 2013). Of the resulting features, 'spec_pearson_norm', 'dot_prod_norm' and 'spec_mse' were used in the auxiliary feature set given their relatively high weights in Percolator learning.…”

“…Comparing these predicted fragment ion intensities with those matched fragment ions aids in discriminating correct PSMs by machine learning Silva et al 2019;Tiwary et al 2019). As such, these fragment intensity correlation features are especially useful to attain a higher confidence for novel (i.e.…”

Scop3P: a comprehensive resource of human phosphosites within their full context

“…In order to obtain uniform data, we collected all '.RAW' files from PRIDE and converted these to Mascot Generic Format (MGF) peak files using ThermoRawFileParser 48 (Hulstaert et al, 2019). The resulting peak files were then searched against the human complement of UniProtKB/Swiss-Prot (release-2018_02, containing 20259 protein sequences) with the target/decoy approach using ionbot (https://ionbot.cloud/; ionbot is based on MS 2 PIP 49-51 and ReScore 52,53 ). Results were filtered at 1% FDR.…”