Fast and Accurate Protein False Discovery Rates on Large-Scale Proteomics Data Sets with Percolator 3.0.

The M, MacCoss MJ, Noble WS, Käll L. · 2016

Read Fast and Accurate Protein False Discovery Rates on Large-Scale Proteomics Data Sets with Percolator 3.0. by The M, MacCoss MJ, Noble WS, Käll L. (2016) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc5059416.txt.

Introduction

Matthew The, Michael J. MacCoss, William S. Noble, and Lukas Käll present Percolator 3.0 in the Journal of the American Society for Mass Spectrometry in 2016. Percolator is a widely used program that rescues more real matches from shotgun proteomics data and gives each a sound confidence score. Proteome datasets had outgrown its speed, and users also wanted reliable scores at the whole-protein level, not just for fragments.

The new version trains on a small sample of the data, a trick stable even at a tiny fraction of the whole. It scored 73 million fragment matches in minutes on an ordinary computer, in a draft human proteome study of 25 million spectra. The team compared protein inference methods and built in the best: group proteins by their predicted fragment sets and judge each by its single best fragment. The package is free under an open license for all major systems.

Catalog record

CREATORThe M, MacCoss MJ, Noble WS, Käll L.
YEAR2016
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC5059416

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Fast and Accurate Protein False Discovery Rates on Large-Scale Proteomics Data Sets with Percolator 3.0 By Matthew The, Michael J. MacCoss, William S. Noble, Lukas Käll Published in: Journal of the American Society for Mass Spectrometry DOI: 10.1007/s13361-016-1460-7 ABSTRACT Percolator is a widely used software tool that increases yield in shotgun proteomics experiments and assigns reliable statistical confidence measures, such as q values and posterior error probabilities, to peptides and peptide-spectrum matches (PSMs) from such experiments. Percolator’s processing speed has been sufficient for typical data sets consisting of hundreds of thousands of PSMs. With our new scalable approach, we can now also analyze millions of PSMs in a matter of minutes on a commodity computer. Furthermore, with the increasing awareness for the need for reliable statistics on the protein level, we compared several easy-to-understand protein inference methods and implemented the best-performing method—grouping proteins by their corresponding sets of theoretical peptides and then considering only the best-scoring peptide for each protein—in the Percolator package. We used Percolator 3.0 to analyze the data from a recent study of the draft human proteome containing 25 million spectra (PM:24870542). The source code and Ubuntu, Windows, MacOS, and Fedora binary packages are available from http://percolator.ms/ under an Apache 2.0 license. Graphical Abstractᅟ INTRODUCTION Percolator [1] has played a prominent part in the analysis pipelines of shotgun proteomics experiments for the last decade, as a post-processor of the results from database search engines such as SEQUEST [2], MASCOT [3], X! Tandem [4], and MS-GF+ [5]. Not only does Percolator provide a significant boost in the number of statistically significant peptide-spectrum matches (PSMs) or peptides, it also provides a consistent statistical framework in which to interpret the search results. Because Percolator’s running time is usually much lower than that of the search engine, applying it as a post-processing step should be the default choice when processing shotgun proteomics data. As part of the continuous development and support of the Percolator package, we present two major additions aimed at supporting analysis of large scale proteomics studies. First, as advances in technology continue to reduce the cost and effort needed to carry out shotgun proteomics experiments, the amount of data per study will keep rising steadily. Although previous versions of Percolator are able to process the data from the vast majority of current studies in a reasonable time frame, the algorithm has some limitations for laboratories without access to a high-performance computing facility. When processing millions of PSMs, the majority of Percolator’s processing time is spent on training support vector machine (SVM) classifiers. In some settings, however, the performance of the SVM as a function of the size of its training set plateaus at a relatively low number of input PSMs [6]. Here, we use Percolator’s semi-supervised learning algorithm to train SVMs on a randomly selected subset of the PSMs and use the resulting score vectors to evaluate the rest of the PSMs. This random downsampling approach yields much shorter analysis times without any loss in statistical power. Second, we have investigated efficient ways to obtain protein-level accuracy estimates. One of the major obstacles was the question of how to deal with shared peptides and protein grouping. An implementation of Fido [7], which has been part of the Percolator package since 2012, addresses these two issues but is too computationally expensive to apply to large-scale data sets. We therefore compared four straightforward and scalable protein inference methods: using the best-scoring peptide, the two-peptide rule [8,9], the product of peptide-level posterior error probabilities (PEPs), and Fisher’s method for combining independent p values. Although each of these methods is efficient to compute, they each offer specific pros and cons. Savitski et al. [10] showed that on large-scale data sets, taking the best-scoring peptide as the representative of a protein was superior to incorporating information from lower-scoring peptides. However, this approach is unsatisfactory because the method discards all information but the best-scoring PSM for each protein. A simple way to combine evidence at the peptide level is the widely used two-peptide rule. This approach requires evidence for a second peptide to support a protein inference, thereby preventing so-called “one-hit wonders” (i.e., cases where a single, potentially spurious PSM yields a spurious protein detection). An alternative that takes into account even more evidence is to compute the product of peptide-level PEPs. This procedure takes into account all peptides within a protein and provides some protection against one-hit wonders [11]. The method also has the benefit that correctly inferred proteins are not strongly affected by incorrectly inferred peptides because these typically contribute a multiplicative term that is close to 1.0. However, a concomitant drawback to using the product of PEPs is that it is not clear how to scale the resulting product to take into account protein length. Also, it is not obvious a priori that the independence assumption implicit in taking the product applies in this case. The latter concern also applies to Fisher’s method, which is a classic technique for combining independent p values [12]. Like the product of PEPs, Fisher’s method takes into account all peptides of a protein, penalizing one-hit wonders on the basis of their many accompanying incorrect peptide inferences [13–15]. This last characteristic can, however, also be a disadvantage, as many incorrect peptide inferences can overrule a minority of correct peptide inferences. Unlike the product of PEPs, Fisher’s method explicitly accounts for the number of p values being combined and hence normalizes for protein length. METHODS We downloaded three sets of spectra, two from large-scale studies on human samples with several millions of spectra and one smaller-scale study on yeast samples. The first large-scale set comprises 2212 runs on 17 adult tissues, seven fetal tissues, and six hematopoietic cell types with a total of ∼25 million spectra [16]. The samples were analyzed on an LTQ Orbitrap Velos and Elite (Thermo Scientific) equipped with an Easy-nLC II nanoflow LC systems (Waters). We will refer to this set as the Kim data set. The second large-scale set was taken from a study aimed at studying variation of protein abundance in humans [17]. It consists of 561 runs on 51 samples from lysates of lymphoblastoid cell lines, resulting in ∼9 million spectra. The peptides were labeled with TMT 6-plex to enable quantification. The analysis took place on an LTQ Orbitrap Velos (Thermo Scientific) equipped with an online 2D nanoACQUITY UPLC system (Waters). We will refer to this set as the Wu data set. For verification of the accuracy of protein-level false discovery rate (FDR) estimates we additionally downloaded 92, 974 spectra from three injections of yeast cells grown to mid-log phase, collected on an LTQ Orbitrap Velos (Thermo Scientific), as described in Moruz et al. [18]. We will refer to this set as the “hm_yeast” set. Converting the RAW files to two separate files in the MS1 and MS2 formats [19], respectively, was done with ProteoWizard [20] with vendor peak picking for the MS2 spectra and all other options left at their default values. Next, we assigned high-resolution precursor masses and charges using information from the precursor scans with Hardklör [21] followed by Bullseye [22], both with the default parameters, through the Crux 2.0 package interface [23]. For the Kim data

Read the complete text