General Statistical Modeling of Data from Protein Relative Expression

---

ARTICLE pubs.acs.org/jpr

General Statistical Modeling of Data from Protein Relative Expression Isobaric Tags Florian P. Breitwieser,† Andr
e M€uller,† Loïc Dayon,‡ Thomas K€ocher,z Alexandre Hainard,‡ Peter Pichler,§ Ursula Schmidt-Erfurth,|| Giulio Superti-Furga,† Jean-Charles Sanchez,‡ Karl Mechtler,z Keiryn L. Bennett,† and Jacques Colinge*,† †

CeMM, Center for Molecular Medicine of the Austrian Academy of Sciences, Vienna, Austria Biomedical Proteomics Group, Department of Structural Biology and Bioinformatics, Faculty of Medicine, University of Geneva, Geneva, Switzerland z Institute of Molecular Pathology, Vienna, Austria § CD Laboratory for Proteome Analysis, University of Vienna, 1030 Vienna, Austria Department of Ophtalmology, Medical University of Vienna, Vienna, Austria

)

‡

bS Supporting Information ABSTRACT: Quantitative comparison of the protein content of biological samples is a fundamental tool of research. The TMT and iTRAQ isobaric labeling technologies allow the comparison of 2, 4, 6, or 8 samples in one mass spectrometric analysis. Sound statistical models that scale with the most advanced mass spectrometry (MS) instruments are essential for their efficient use. Through the application of robust statistical methods, we developed models that capture variability from individual spectra to biological samples. Classical experimental designs with a distinct sample in each channel as well as the use of replicates in multiple channels are integrated into a single statistical framework. We have prepared complex test samples including controlled ratios ranging from 100:1 to 1:100 to characterize the performance of our method. We demonstrate its application to actual biological data sets originating from three different laboratories and MS platforms. Finally, test data and an R package, named isobar, which can read Mascot, Phenyx, and mzIdentML files, are made available. The isobar package can also be used as an independent software that requires very little or no R programming skills. KEYWORDS: bioinformatics, statistics, iTRAQ, TMT, quantitative proteomics

’ INTRODUCTION Proteomic technologies provide access to the protein content of biological samples1,2 and are important tools for current medical, biological, and systems biology research. Several highly efficient approaches also using MS exist to measure quantitative information related to proteins3
5 that can be combined with PTM analysis. In this work, we consider methods allowing the measurement of proteome-wide protein relative expression.5 In general, protein digestion by an enzyme, e.g., trypsin, and tandem mass spectrometry (MS/MS) are required to identify the resultant peptides.6 The samples for comparison are prepared such that the peptides from each of them are labeled in order to distinguish them after sample pooling and shared MS analysis. Several methods have been designed along this principle, e.g., ICPL,7 ICAT,8 SILAC,9 COFRADIC,10 16O/18O,11 iTRAQ,12 and TMT13 to cite the most common ones. iTRAQ is especially convenient as (1) it can be multiplexed (up to 4 samples can be analyzed simultaneously), and (2) quantitative information resides in each single MS/MS spectrum (not necessary to combine r 2011 American Chemical Society

spectra). Multiplexing is achieved through the use of isobaric tags (equal mass) to label the peptides. These tags fragment during MS/MS, thus yielding reporter peaks with distinct m/z ratios,12 e.g., 114, 115, 116, and 117 Da. Direct comparison of the reporter peak intensities, or channel intensities, provides an estimate of relative expression. TMT (2- or 6-plex) works according to the same principle, and there exists an 8-plex version of iTRAQ; the theory we develop here applies to all of them. In this work, we are interested in the prevalent experimental settings where biological samples are compared in a single experiment (with or without replicates). Experimental design that is composed of multiple iTRAQ/TMT experiments is out of the scope of this work and has been studied by others.14
16 Regarding statistical analysis, iTRAQ/TMT data have similarities with gene microarray data, though they also have clear specificities. One notable difference is the variability of available information due to the variable number of measured spectra. Received: December 23, 2010 Published: April 28, 2011 2758

dx.doi.org/10.1021/pr1012784 | J. Proteome Res. 2011, 10, 2758–2766

Journal of Proteome Research Consequently, the estimation of protein ratios comes with variable accuracy. One major breakthrough in iTRAQ data analysis was the introduction of signal intensity noise models.17
19 We present a coherent approach that extends and improves the applicability of this concept by also modeling biological sample variability and we validate our results extensively on complex and realistic test samples comprised of albumin- and IgG-depleted human plasma background and spiked ceruloplasmins (CERU). The influence of the number of available spectra to estimate protein ratios is reported as well. Many reported methods for iTRAQ and TMT data analysis do not provide statistical guidance to select regulated proteins, and ad hoc thresholds must be applied to expression fold changes. Recent developments that employed statistics14,15,17
19 are compared with our approach. We further demonstrate the application of the proposed method to several biological data sets obtained from different MS platforms (ThermoFisher Scientific ESI-LTQ Orbitrap TMT 6-plex20 and iTRAQ 4-plex,21 Applied Biosystems MALDI-TOF/TOF TMT 6-plex20). We finally show how the presented statistical framework can provide, to our best knowledge, the first practical tool for assessing the expression of proteins with no specific peptides such as isoforms. Only theoretical work22 addressed this question so far. Our layered modeling enables straightforward exploitation of quantitative proteomic data and is implemented in a R package named isobar.

’ METHODS Test Samples

Mouse and rat lyophilized CERU were dissolved and digested (trypsin) to prepare 10 fmol/μL stock solutions. These were mixed in a reciprocal fashion for each 4-plex iTRAQ channel (114:115:116:117); Set 1 = 1:2:5:10 (rat) and 10:5:2:1 (mouse), and Set 2 = 1:10:50:100 (rat) and 100:50:10:1 (mouse). A complex background peptide mixture was generated by depleting albumin and IgG (ProteoPrep) from 160 μL of human plasma. Reduced and alkylated depleted plasma was separated by 1D-SDSPAGE, and following visualization of the proteins by colloidal coomassie staining, several regions were excised and the proteins in the gel digested in situ with trypsin.23 The resultant background peptide mixture was extracted from the gel slices and purified, and four fractions were combined with 6 pmol of digested human CERU and mixed with Set 1 and Set 2 to obtain the final test samples (TS1) and (TS2). In TS1, 1:2:5:10 relates to 6.1:12.2:30.5:61 fmol CERU peptides, whereas in TS2, 1:10:50:100 relates to 0.7:6.8:34.2:68.3 fmol CERU peptides. Peptides were separated at pH 10 on a Gemini-NX column (Phenomenex, Torrance, CA). Forty fractions were collected and subsequently analyzed with a hybrid LTQ-Orbitrap XL mass spectrometer (ThermoFisher Scientific, Waltham, MA) coupled to an Agilent 1200 HPLC nanoflow system via a nanoelectrospray ion source (Proxeon, Odense, Denmark). Analyses were performed in a data-dependent acquisition mode using a top 3 higher-energy collision-induced dissociation (HCD) method. MS data were searched against human Swiss-Prot24 2010.09, with mouse and rat ceruloplasmins appended, using Mascot25 2.3 and Phenyx26 2.6.1, imposing a protein group false discovery rate (FDR)