Proteomics for Biological Discovery [2 ed.] 9781119081692, 1119081696

Table of contents :
Cover
Proteomics for
Biological Discovery
© 2019
Contents
Foreword
List of Contributors
1 Quantitative Proteomics
for Differential Protein
Expression Profiling
2 Protein Microarrays
3 Protein Biomarker
Discovery: An Integrated
Concept
4 Biomarker Discovery
with Mass Spectrometry
Imaging and Profiling
5 Protein–Protein
Interactions
6 Mass Spectrometry of Intact
Protein Complexes
7 Cross‐linking Applications
in Structural Proteomics
8 Functional Proteomics:
Systematic
Characterization of the
Physical and Functional
Organization of
Cell Systems
Pierre C.
9 High‐Resolution
Interrogation of Biological
Systems via Mass
Cytometry
10 Characterization of Drug–
Protein Interactions by
Chemoproteomics
11 Phosphorylation
12 Large‐Scale
Phosphoproteomics
13 Probing Glycoforms
of Individual Proteins
Using Antibody‐Lectin
Sandwich Arrays: Methods
and Findings from Studies
of Pancreatic Cancer
Index

Citation preview

Proteomics for  Biological Discovery

Proteomics for Biological Discovery Second Edition

Edited by Timothy D. Veenstra Watertown, Wisconsin

John R. Yates III Torrey Mesa, California

This Second edition first published 2019 © 2019 John Wiley & Sons, Inc. Edition History John Wiley & Sons, Inc. (1e 2006) All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, except as permitted by law. Advice on how to obtain permission to reuse material from this title is available at http://www.wiley.com/go/permissions. The right of Timothy D. Veenstra and John R. Yates III to be identified as the Editors of the editorial material in this work has been asserted in accordance with law. Registered Office(s) John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030, USA Editorial Office Boschstr. 12, 69469 Weinheim, Germany For details of our global editorial offices, customer services, and more information about Wiley products visit us at www.wiley.com. Wiley also publishes its books in a variety of electronic formats and by print‐on‐demand. Some content that appears in standard print versions of this book may not be available in other formats. Limit of Liability/Disclaimer of Warranty While the publisher and authors have used their best efforts in preparing this work, they make no representations or warranties with respect to the accuracy or completeness of the contents of this work and specifically disclaim all warranties, including without limitation any implied warranties of merchantability or fitness for a particular purpose. No warranty may be created or extended by sales representatives, written sales materials or promotional statements for this work. The fact that an organization, website, or product is referred to in this work as a citation and/or potential source of further information does not mean that the publisher and authors endorse the information or services the organization, website, or product may provide or recommendations it may make. This work is sold with the understanding that the publisher is not engaged in rendering professional services. The advice and strategies contained herein may not be suitable for your situation. You should consult with a specialist where appropriate. Further, readers should be aware that websites listed in this work may have changed or disappeared between when this work was written and when it is read. Neither the publisher nor authors shall be liable for any loss of profit or any other commercial damages, including but not limited to special, incidental, consequential, or other damages. Library of Congress Cataloging‐in‐Publication Data Names: Veenstra, Timothy Daniel, 1966– editor. | Yates III, John R., editor. Title: Proteomics for biological discovery / edited by Timothy D. Veenstra, John R. Yates III. Description: Second edition. | Hoboken, NJ : Wiley-Blackwell, 2019. | Includes bibliographical   references and index. | Identifiers: LCCN 2019015129 (print) | LCCN 2019015652 (ebook) | ISBN 9781119081692   (Adobe PDF) | ISBN 9781119081722 (ePub) | ISBN 9781118279243 (hardback) Subjects: | MESH: Proteomics | Computational Biology–methods Classification: LCC QP551 (ebook) | LCC QP551 (print) | NLM QU 460 | DDC 612.3/98–dc23 LC record available at https://lccn.loc.gov/2019015129 Cover Design: Wiley Cover Image: © Scala /Art Resource Set in 10/12pt TimesTen by SPi Global, Pondicherry, India 10 9 8 7 6 5 4 3 2 1

Contents Foreword

vii

List of Contributors

ix

1. Quantitative Proteomics for Differential Protein Expression Profiling

1

Christian K. Frese, Henk van den Toorn, Albert J.R. Heck, and Shabaz Mohammed

2. Protein Microarrays

29

Fernanda Festa and Joshua LaBaer

3. Protein Biomarker Discovery: An Integrated Concept

63

Andrei P. Drabovich, Eduardo Martínez‐Morillo, and Eleftherios P. Diamandis

4. Biomarker Discovery with Mass Spectrometry Imaging and Profiling

89

Sage J.B. Dunham, Elizabeth K. Neumann, Eric J. Lanni, Ta-Hsuan Ong, and Jonathan V. Sweedler

5. Protein–Protein Interactions

125

Claire M. Delahunty and John R. Yates III

6. Mass Spectrometry of Intact Protein Complexes

145

Jonathan T.S. Hopper and Carol V. Robinson

7. Cross‐linking Applications in Structural Proteomics

175

Evgeniy V. Petrotchenko, Jason J. Serpa, and Christoph H. Borchers

v

vi  Contents

  8. Functional Proteomics: Systematic Characterization of the Physical and Functional Organization of Cell Systems

197

Pierre C. Havugimana, Pingzhao Hu, and Andrew Emili

  9. High‐Resolution Interrogation of Biological Systems via Mass Cytometry 215 Heather M. Grundhofer, Michelle M. Kuhns, and Edgar A. Arriaga

10. Characterization of Drug–Protein Interactions by Chemoproteomics

247

Markus Schirle, Marcus Bantscheff, and Bernhard Kuster

11. Phosphorylation

265

Timothy D. Veenstra

12. Large‐Scale Phosphoproteomics

291

John R. Yates III

13. Probing Glycoforms of Individual Proteins Using Antibody‐Lectin Sandwich Arrays: Methods and Findings from Studies of Pancreatic Cancer

311

Brian B. Haab

Index329

Foreword What Is Proteomics? A critical advance in biology was the sequencing of the human genome ­approximately 15 years ago. A dedicated effort to advance technology has made it ­feasible and cost‐ effective to sequence the entire genomes of individuals with a growing use in clinical diagnosis. The growing collection of DNA sequence data has provided a powerful resource for studies involving protein biochemistry, in particular to create a better understanding of how disease mechanisms manifest from genes to proteins. Advanced methods in large‐scale protein biochemistry or proteomics have broadened the types of experiments possible. How Is This Driving Biological Research? Understanding diseases requires discovering the mechanisms by which biological processes are disrupted. These mechanisms are often manifested through proteins and their functions. Proteomic methods are now able to measure protein expression, the composition of organelles, posttranslational modifications, and protein– protein interactions to determine how proteins are changed as a function of disease. A variety of methods make these measurements possible, including mass spectrometry and protein arrays. Protein arrays allow the study of large‐scale ­protein expression. They also allow scanning for circulating reactive antibodies that associate with disease. These advanced methods are increasingly used for studies to identify markers for disease. Increasingly, proteomic tools are being used in the development of therapeutic treatments. In this second edition of Proteomics for Biological Discovery, chapters describe research meeting these needs. Mohammed and Heck describe recent advances in quantitative proteomics using mass spectrometry. Veenstra describes proteome analysis of posttranslational modifications. Delahunty and Yates describe mass spectrometry‐based methods and vii

viii   Foreword

applications to use affinity purification mass spectrometry for characterization of protein complexes. Diamandis and Drabovich cover the process of biomarker ­discovery. Yates discusses the large‐scale analysis of phosphorylation in biological systems. Robinson discusses the characterization of intact protein complexes using native mass spectrometry. Borchers describes the use of protein cross‐linking to characterize protein structures and protein–protein interactions. Emili describes the use of proteomics to understand protein function. Haab discusses the use of antibodies for proteomic profiling. LaBaer describes the use of protein arrays in proteomics. Sweedler describes the use of mass spectrometry imaging. An important new area of proteomics is single cell mass cytometry which is described by Edgar Arriaga. Kuster describes how to characterize drug–protein interactions.

List of Contributors Timothy D. Veenstra Department of Applied Science Maranatha Baptist University Watertown, WI, USA John R. Yates III Departments of Molecular Medicine and Neurobiology The Scripps Research Institute, LaJolla, CA, USA Edgar A. Arriaga Department of Chemistry, University of Minnesota Minneapolis, MN, USA Marcus Bantscheff Cellzome, Heidelberg, Germany Christoph H. Borchers University of Victoria – Genome British Columbia Proteomics Centre Vancouver Island Technology Park, Victoria, BC, Canada Claire M. Delahunty Departments of Molecular Medicine and Neurobiology The Scripps Research Institute, LaJolla, CA, USA Eleftherios P. Diamandis Samuel Lunenfeld Research Institute, Mount Sinai Hospital, Toronto, ON, Canada Andrei P. Drabovich Department of Laboratory Medicine and Pathology, University of Alberta, Edmonton, AB, Canada ix

x  List of Contributors

Sage J.B. Dunham Department of Chemistry and the Beckman Institute of Science and Technology University of Illinois at Urbana–Champaign, Champaign, IL, USA Andrew Emili Donnelly Centre for Cellular and Biomolecular Research University of Toronto, Toronto, ON, Canada Fernanda Festa Departments of Pediatrics and Biochemistry/Molecular Biology College of Medicine, Penn State University, Hershey, PA, USA Christian K. Frese Biomolecular Mass Spectrometry and Proteomics Bijvoet Center for Biomolecular Research and Utrecht Institute for Pharmaceutical Sciences, Utrecht University, Utrecht The Netherlands Heather M. Grundhofer Department of Chemistry, University of Minnesota Minneapolis, MN, USA Brian B. Haab Van Andel Research Institute, Grand Rapids, MI, USA Pierre C. Havugimana Donnelly Centre for Cellular and Biomolecular Research, University of Toronto, Toronto, ON, Canada Albert J.R. Heck Biomolecular Mass Spectrometry and Proteomics, Bijvoet Center for Biomolecular Research and Utrecht Institute for Pharmaceutical Sciences, Utrecht University, Utrecht, The Netherlands Jonathan T.S. Hopper Department of Chemistry, Physical and Theoretical Chemistry Laboratory, University of Oxford, Oxford, UK Pingzhao Hu Department of Biochemistry and Medical Genetics, University of Manitoba, Winnipeg, MB, Canada Michelle M. Kuhns Department of Chemistry, University of Minnesota, Minneapolis, MN, USA Bernhard Kuster Cellzome, Heidelberg, Germany. Technical University Munich, F.reising, Germany

List of Contributors  xi

Joshua LaBaer Virginia G. Piper Center for Personalized Diagnostics, Biodesign Institute, Arizona State University, Tempe, AZ, USA Eric J. Lanni Department of Chemistry and the Beckman Institute of Science and Technology, University of Illinois at Urbana–Champaign, Champaign, IL, USA Eduardo Martínez‐Morillo Samuel Lunenfeld Research Institute, Mount Sinai Hospital, Toronto, ON, Canada Shabaz Mohammed Biomolecular Mass Spectrometry and Proteomics Bijvoet Center for Biomolecular Research and Utrecht Institute for Pharmaceutical Sciences, Utrecht University, Utrecht The Netherlands Elizabeth K. Neumann Department of Chemistry and the Beckman Institute of Science and Technology University of Illinois at Urbana–Champaign, Champaign, IL, USA Ta‐Hsuan Ong Department of Chemistry and the Beckman Institute of Science and Technology, University of Illinois at Urbana–Champaign, Champaign, IL, USA Evgeniy V. Petrotchenko University of Victoria – Genome British Columbia Proteomics Centre Vancouver Island Technology Park, Victoria, BC, Canada Carol V. Robinson Department of Chemistry, Physical and Theoretical Chemistry Laboratory University of Oxford, Oxford, UK Markus Schirle Novartis Institutes for BioMedical Research, Inc., Cambridge, MA, USA Jason J. Serpa University of Victoria – Genome British Columbia Proteomics Centre Vancouver Island Technology Park, Victoria, BC, Canada Jonathan V. Sweedler Department of Chemistry and the Beckman Institute of Science and Technology University of Illinois at Urbana–Champaign, Champaign, IL, USA Henk van den Toorn Biomolecular Mass Spectrometry and Proteomics, Bijvoet Center for Biomolecular Research and Utrecht Institute for Pharmaceutical Sciences Utrecht University, Utrecht, The Netherlands

1 Quantitative Proteomics for Differential Protein Expression Profiling Christian K. Frese, Henk van den Toorn, Albert J.R. Heck, and Shabaz Mohammed Biomolecular Mass Spectrometry and Proteomics, Bijvoet Center for Biomolecular Research and Utrecht Institute for Pharmaceutical Sciences, Utrecht University, Utrecht, The Netherlands *Christian K. Frese and Henk van den Toorn contributed equally.

1.1 INTRODUCTION Mass spectrometry (MS)‐based proteomics has become an integral analytical technology in life science research (Aebersold and Mann 2003; Chait 2011). ­ Coupling liquid chromatography (LC) to MS facilitated the routine analysis of several thousands of proteins in parallel within a few hours and has mostly replaced two-dimensional gel electrophoresis‐based approaches. Over the past decade, MS has matured from a technique that produces purely qualitative information into a versatile tool that provides accurate quantitative data on protein abundance. Quantitative ­information is a fundamental necessity to interrogate the highly dynamic global proteome of living organisms. Driven by recent advances in mass spectrometric instrumentation and data analysis software, today, MS‐based proteomics builds the analytical framework for various challenges in biomarker discovery, systems biology, and structural biology (Bensimon et  al. 2012; Altelaar et al. 2013b).

Proteomics for Biological Discovery, Second Edition. Edited by Timothy D. Veenstra and John R. Yates III. © 2019 John Wiley & Sons, Inc. Published 2019 by John Wiley & Sons, Inc.

1

2  Proteomics for Biological Discovery

1.2  QUANTIFICATION APPROACHES Multiple quantification strategies for MS‐based proteomics have been reported. They can be categorized into two main regimes: absolute and relative quantification. In this chapter we will focus on the latter and discuss practical aspects, ­applicability, and problems of the currently most popular quantification strategies. For those interested in absolute quantification, a detailed insight is given in a review by Brun et al. (2009). The need for reliable peptide identification underlies all MS‐ based quantification strategies. Protein identification in the commonly employed bottom‐up proteomics strategy involves enzymatic proteolysis followed by one‐ dimensional or two‐dimensional peptide chromatographic separations. Trypsin and Lys‐C are the routinely used enzymes where the former generates peptide lengths between 10 and 20 amino acids for the majority of the proteolytic peptides. These “MS‐friendly” peptides are sequenced by tandem mass spectrometry (MS/MS) and identified by database search algorithms (Eng et  al. 1994; Perkins et  al. 1999). Peptide quantification is performed either at the MS or MSn‐level, depending on the quantification strategy. The use of isotopes to label specific molecules has greatly expanded the toolbox in biochemistry. Radioisotopes such as 32P, 33P, 35S, or 125I and stable isotopes such as 2 H, 13C, and 15N are widely utilized for quantifying, tracking protein–protein interactions, determining enzyme kinetics or dissecting metabolic pathways. Besides proteins, nucleic acids, lipids, and carbohydrates have also been subjected to  isotopic labeling. The main techniques to analyze biomolecules involving the use of nonradioactive, stable isotopes are nuclear magnetic resonance (NMR) and  MS. Today, MS‐based proteomics aims to provide large‐scale quantitative information on protein abundances. The basic principle of stable isotope‐based ­labeling for peptide and protein quantification is that the physicochemical c­ haracteristics of the differentially labeled peptides are nearly identical (Gouw et al. 2010). These similarities include sample preparation procedures, LC‐separation performance, ionization efficiency and MS/MS fragmentation behavior. An additional assumption is that the different isotope variants do not influence any cellular process in the case of metabolic labeling. Most labeling strategies introduce stable (nonradioactive) isotopes with distinct masses that allow each sample to be distinguishable at the MS stage. The stable isotopes can be introduced via chemical reactions of a labeling reagent with a distinct functional group, by metabolic processes or enzymatic activity (Figure  1.1). Differentially labeled samples are combined prior to MS analysis assuming an overall near‐complete labeling efficiency. The shift in absolute mass depends on the number of heavy stable isotopes incorporated into the peptide or protein. The actual signal shift between the isoforms (in the m/z range) of a differentially labeled peptide depends on the absolute mass shift and the peptide ion charge state. For each peptide, the area under the curve of the isotopic envelope is integrated over the LC elution time. Relative quantification is performed by calculating the ratio of the peak areas of the differentially labeled peptides. Intensity‐based label‐free quantification is based on the same principle; however, peptide abundances are retrieved from consecutive LC‐MS/MS analyses since each sample is analyzed individually.

Quantitative Proteomics for Differential Protein Expression Profiling  3

metabolic labeling

chemical labeling

label free

sample

protein level

digestion

peptide level

MS analysis 500

f(x)

m/z

1000

vs

data analysis

Figure 1.1  Workflows in quantitative proteomics. Dashed lines highlight the steps where samples are treated separately, which introduces experimental variation. Horizontal lines illustrate at which step samples are combined. Starting with the step when samples are labeled, samples are highlighted in orange and green, respectively. In metabolic labeling, samples are combined at the earliest possible step which, in theory, minimizes variation introduced at each sample preparation step. In chemical labeling, samples are combined after labeling. Label‐free quantification requires every sample to be processed individually. Here, the samples are combined at the data analysis level. Source: Adapted from Ong and Mann (2005). (See color insert.)

1.2.1  Metabolic Labeling In metabolic labeling, the stable isotopes are introduced into the proteins in living cells or organisms during protein synthesis. This procedure requires that the organism or cell line must be auxotrophic for the source of the isotope used for labeling. Thus, after a few cell cycles the heavy isotopes are fully incorporated into all proteins. One of the first applications of metabolic labeling in proteomics involved the use of 15N for labeling of bacteria (Oda et al. 1999). The heavy nitrogen is introduced into the proteins mostly via 15N‐ammonium or nitrate salts in the culture medium. This leads to all N‐atoms being replaced by the heavy nitrogen isotope during cell growth and division. Complete labeling is usually achieved after 5–10 cell cycles. Consequently, all peptides can be quantified by MS independent of the amino acid composition. Thus, every enzyme can be utilized for proteolytic digestion. 15N‐­labeling is mostly applied in bacterial or Saccharomyces cerevisiae cultures (Washburn et  al. 2002) but it has also been successfully applied in  Escherichia coli (Pasa‐Tolic et  al. 1999). Multicellular organisms can be labeled by feeding them 15N‐enriched yeast, bacteria or algae. This approach has

4  Proteomics for Biological Discovery

extended the technique to model systems such as Drosophila melanogaster and Caenorhabditis elegans (Krijgsveld et al. 2003). Metabolic labeling of mammals was first reported by Yates and coworkers who introduced stable isotopes into rats (Wu et al. 2004). Moreover, metabolic labeling of chicken has also been reported (Doherty et al. 2005). However, complete incorporation in rodents is hard to achieve even after long‐term feeding with 15N‐enriched nutrients. In plant biology, 15N can be introduced via liquid nutrition solutions (Schaff et al. 2008), axenic liquid cultures (Huttlin et al. 2007), or agar (Hebeler et al. 2008). A number of different plants have been metabolically labeled for quantitative proteomic experiments using 15N, including the most studied plant model, Arabidopsis thaliana (Nelson et al. 2007). A comprehensive review on the use of 15N in plants has been published by Arsova et al. (2012). The main advantage underlying metabolic labeling strategies is that the isotopic labels are introduced during protein synthesis in living cells. After a few cell cycles, the heavy isotopes are typically fully incorporated into all proteins. This strategy allows c­ ombination of the differentially labeled samples at the earliest possible stage in proteomic workflows which in theory minimizes variation that can potentially be introduced by systematic errors at every sample preparation step (see Figure  1.1). One of the caveats of 15N‐based labeling is that the mass difference between the labeled and unlabeled version of the peptide varies with peptide length and amino acid composition, which complicates data analysis. Moreover, small 14N contaminations can cause incomplete peptide labeling and thus hamper peptide identification and correct assignment of the corresponding peptide pairs. One also has to consider that, despite the fact that all isotopes have the same chemical properties, enzymatic reactions can be compromised that may lead to a bias against the 15N isotopes in some cellular processes (Evans 2001). Besides these drawbacks, 15N‐ based labeling is too expensive or laborious for some biological systems. Labeling via stable isotopes by amino acids in cell culture (SILAC) is another strategy that facilitates complete and proteome‐wide labeling. For this purpose, amino acid‐deficient cell culture medium is supplemented with stable isotope‐ enriched amino acids that are taken up by the cells and incorporated into all proteins. Logically, this strategy requires the cells to be auxotrophic for the labeled amino acid (Ong et al. 2002). Typically, one or two specific amino acids that contain the stable isotopes are supplemented to the culture medium. Unlabeled and labeled versions of the same peptide are separated by a defined mass difference, depending on the number of labeled amino acids per peptide. Leucine (Ong et al. 2002) and methionine (Hunter et al. 2001), among other amino acids, have been utilized for introducing deuterium as the stable heavy isotope. However, due to the different reversed‐phase chromatography high‐performance liquid chromatography (RP‐ HPLC) retention behavior of deuterium‐labeled peptides compared to their unlabeled counterparts, most labeling approaches have switched to 13C and 15N as a source of stable isotopes. Besides the need for the cell line to be auxotrophic for the particular amino acid, one also has to consider the variation in the frequency of occurrence of the amino acids in the total proteome and the need for a corresponding protease. For instance, tryptophan and cysteine are relatively less abundant amino acids, which makes them less suitable for stable isotope labeling. Nowadays, arginine (Blagoev et al. 2004) and lysine (Berger et al. 2002) are the amino acids most commonly utilized to introduce stable isotopes into the proteome (Figure  1.2.a). Using these two amino acids in

Quantitative Proteomics for Differential Protein Expression Profiling  5

(a) SILAC Arg

Lys O

H N

CH C

Arg-6

OH H N

CH

Lys-4 O

O C

H

H N

OH

C

C

Arg-10

OH

H N

Lys-8 O

O H

OH

CH C

H

N

O OH H

C

C

N

H

C

CH

CH

CH

CH

CH

CH

CH

CH

CD

CH

CH

CH

CH

CH

CD

CH

CH

NH

CH

CH

NH

NH

C

NH

C

NH C

NH

NH

NH

CH

C

OH

CH NH

NH

NH

NH

(b) dimethyl O

R–NH

H

H

NaBH CN

N

R–NH

H C D

NaBH CN

R–NH

NaBD CN

D N

CH

C

C N

OH

CH

D HC

CH

C

CHD

CH

CH

CH

CHD

CH

CH

CH

CH

CH

CH

N

N

CD R–N

H C

CD

D HC

CH

(c) TMT

N D

CHD

CD

C

iTRAQ O

O

O

O

O N

N

N

N H

reporter group

R-NH

balancer group

N

iTRAQ-(O-NHS) pH 8

amine reactive group

O

TMT-(O-NHS) pH 8

N

O O

reporter group

balancer group amine reactive group

O H N

N H H N

N N

N

O O

R-NH

OH

CH

C

D

R–N

C D

D HC OH

CH

O D

C

CH

O D

CH

O

O

O

H C

CH R–N

R

R

O

Figure 1.2  Chemical structures of the stable isotope labeled peptides in quantitative proteomics. (a) In SILAC, 13C, 2H, and 15N isotopes are introduced into the proteins via lysine and arginine (Δm = +4/+6 Da  for intermediate and +8/+10 Da  for heavy label, respectively). (b) Dimethyl labeling targets free amine groups via reductive amination. Formaldehyde and cyanoborohydride with varying number of heavy isotopes (13C and 2H) are employed to methylate free amine groups which results in a mass increase of 28, 32, and 36 Da for the light, intermediate, and heavy label, respectively. (c) Isobaric labeling via TMT or iTRAQ utilizes tags composed of three parts which add up to the same molecular mass for all channels.The isobaric tags label peptides via NHS chemistry through an amine‐specific reactive group.The number of 13C and 15N atoms is constant, but the position of the heavy atoms varies between the different labels. The identity of the tags is revealed by the reporter group upon fragmentation of the peptide (the balancer group undergoes neutral loss).

6  Proteomics for Biological Discovery

combination with trypsin or Lys‐C digestion generates peptides that contain at least one C‐terminal‐labeled amino acid. Up to three samples (unlabeled, 13C‐ or 13C/15N‐ containing amino acids) are usually differentially labeled and mixed in a single experiment. A caveat of SILAC labeling is the potential metabolic conversion of arginine to proline (van Hoof et al. 2007). This conversion results in an undesired mass shift of proline‐containing peptides that hampers quantification and leads to underestimation of the original heavy peptide abundance. This issue may be addressed by careful titration of the amount of arginine (Ong et al. 2003), the addition of proline to the culture medium (Bendall et al. 2008) or by computational correction of the obtained peptide ratios (Park et al. 2009). Over recent years, the SILAC toolbox has been extended to label whole ­organisms such as D. melanogaster (Sury et al. 2010) and Mus musculus (Kruger et al. 2008). The so‐called SILAC mouse has been successfully used for in‐depth comparison of the left and right ventricle proteomes (Scholten et al. 2011) as well as for investigation of global proteome changes during aging (Walther and Mann 2011). SILAC has also been used for pulse labeling to study global protein turnover (Milner et al. 2006). In these so‐called pSILAC experiments, cells are exposed to heavy amino acids for a defined period of time prior to lysis. The incorporation of the heavy stable isotopes into the proteins, that is, the protein turnover rate, can be determined by the ratio between the heavy and light peptide versions. Selbach and coworkers extended the experimental procedures to allow investigation of changes in global protein turnover upon an external stimulus (Schwanhausser et al. 2009). Currently, SILAC labeling is not applicable to human tissue or any clinical sample. However, the Mann group recently introduced the so‐called super SILAC approach for quantification of animal and human tissue proteins (Geiger et al. 2010). For this purpose, they used a mixture of SILAC‐labeled cells as an internal standard (Geiger et al. 2011). The tissue complexity arises from its heterogeneous composition from various cell types. To mimic the tissue of interest as closely as possible, they combined different cell lines at differing levels for the super SILAC mixture. The proteins of the unlabeled samples of interest are first quantified against the internal standard. Next, the differences between the tissue samples are determined by calculating the ratios between initially determined ratios against the internal standard. The super SILAC approach theoretically expands the SILAC toolbox to all kinds of samples. However, the challenge arising is the need for a defined reference sample that ideally represents exclusively the same proteome as the tissue sample of interest without adding too many unspecific proteins that would increase complexity and thus complicate quantification. Moreover, the super SILAC approach cannot reflect the tissue microenvironment of living cells. 1.2.2  Chemical Labeling Chemical stable isotope labeling strategies make use of the presence of reactive groups of peptides and proteins. The main difference from metabolic labeling is that chemical labeling approaches are applicable to virtually all samples, including body fluids and clinical samples because the stable isotope labels are ­introduced in the sample preparation process. In 1999, Aebersold and coworkers introduced isotope‐coded affinity tags (ICAT) that target the thiol group of cysteine residues

Quantitative Proteomics for Differential Protein Expression Profiling  7

(Gygi et  al. 1999). This tag is composed of three parts: a thiol‐specific reactive group, a linker that contains either zero or eight deuterium atoms, and a biotin affinity tag. The latter enables affinity purification of the labeled ­peptides via streptavidin, which leads to a reduction in complexity of the sample. In contrast, ICAT labeling is difficult for smaller proteins that contain only a few cysteines. ICAT labeling has also been successfully employed for organelle proteomics (Dunkley et al. 2004). Since the introduction of ICAT, other chemical groups such as carboxyls (Goodlett et al. 2001; Syka et al. 2004) have been explored for chemical labeling. However, the majority of the techniques target the amine group at the peptide N‐ terminus and at the ε‐position of lysines. Primary amines can be easily derivatized using various approaches, such as acylation or sulfonation (Regnier and Julka 2006). Derivatizations via specific N‐hydroxysuccinimide (NHS) chemistry or reductive amination are among the more frequently used techniques. Labeling can be performed on either peptide or protein level. The former approach has the advantage that the samples can be combined at an earlier step in the workflow. However, the fact that the most popular proteases trypsin and Lys‐C do not cleave derivatized lysine residues limits the applicability. Moreover, incomplete labeling is more likely to occur when labeling is performed on an intact protein level. This limitation is likely the reason why chemical labeling of intact proteins is not so frequently used in proteomic experiments. In stable isotope dimethyl labeling, the amine groups are labeled by forming a Schiff base upon reaction with formaldehyde, which is then rapidly reduced by cyanoborohydride (Hsu et al. 2003; Huang et al. 2006; Boersema et al. 2008; Boersema et al. 2009). Two cycles of these reactions will convert primary amines into tertiary amines with two methyl groups – “dimethyls” (Figure 1.2.b). Isotopomers of formaldehyde and cyanoborohydride are utilized to introduce Cα‐bound deuterium and 13C atoms as stable isotope labels to each peptide. Typical tryptic peptides contain either one (C‐terminal arginine) or two labels (C‐terminal lysine) which translates to a mass increase of 4 or 8 Da, respectively. As in SILAC, commonly three samples are differentially labeled and analyzed together in one run (Figure 1.3). Several advantages make dimethyl labeling an attractive option in quantitative proteomics. First, the required reagents are cheap and the reaction rate is very fast  and efficient. Second, the labeling protocol allows great scalability from low ­microgram to milligram amounts of sample. Third, using dimethyl labeling prior to immunoprecipitation (IP) at the  peptide level reduces variation potentially being introduced at the IP‐ or LC‐MS/MS steps (Ding et al. 2011). Moreover, the labeling can be performed in solution, on C18 solid‐phase extraction columns or online coupled to HPLC and MS (Raijmakers et al. 2008, 2009). The use of deuterium leads to a small retention time shift when using reversed phase‐based peptide separation coupled to MS (Zhang et al. 2001). This shift can influence accuracy if peptide quantification is solely based on a single MS spectrum, but has little to no effect on quantification if the area under the whole chromatographic peak is integrated (Ji and Li 2005; Cox and Mann 2008). A comprehensive overview of applications using dimethyl labeling was published by Kovanich et al. (2012). Acylation of primary amines via esters such as NHS provides another elegant way to introduce stable isotopes into peptides and proteins. Münchbach et al. used H4/D4‐ nicotinoyloxysuccinimide esters to selectively label the N‐terminus of

8  Proteomics for Biological Discovery (a)

40

60

80

100

120

140

Rentention time (min)

(b) MS-Quantification

400

450

500

550

600

650

700

750

800

m/z

(c) MS2-Quantification

400

900 m/z

MS3-Quantification

(d)

126

127 m/z

128

Figure 1.3  Peptide quantification on MS‐ and MSn‐level. (a) Typical chromatogram of a complex ­peptide mixture separated by reversed‐phase nano‐HPLC. (b) Full scan MS spectrum at a given time point of the analysis. Inset shows a zoom‐in of a labeled peptide (three different labels). MS‐level quantification is performed by comparing the MS abundance of the differentially labeled peptides. (c) MS/ MS spectrum of a peptide after collision‐induced dissociation. Inset shows the low mass region where the reporter ions of iTRAQ/TMT labeling are present. Accuracy in MS2‐based reporter ion quantification is hampered by interference of coisolated and cofragmented background peptides (highlighted in red) which leads to ratio distortion. (d) The true intensities of the reporter ions are unveiled by MS3 methods which remove interfering background reporter ions.

Quantitative Proteomics for Differential Protein Expression Profiling  9

t­ryptic p ­ eptides, which induces a mass shift of 4 Da (Munchbach et  al. 2000). However, mixing differentially coded samples increases sample complexity, which can negatively affect proteome coverage and quantification accuracy. A potential solution to this is so‐called “isobaric” tandem mass tags that also employs NHS chemistry to target free amines and are currently among the most popular chemical labeling strategies (Figure 1.2c). The most common ones allow multiplexing of up to six tandem mass tags (TMT) (Thompson et al. 2003) or eight isobaric tags for relative and absolute quantitation (iTRAQ) (Ross et al. 2004) samples. Contrary to other chemical labeling strategies, isobaric labeling introduces labels that have the same total mass. The tags are composed of a reactive group, a reporter, and a ­balancer group. The number of 13C and 15N atoms remains the same between the ­different tags but the position of the heavy carbon and nitrogen atoms within the reporter and balancer group changes. Thus, the same peptide originating from samples that are differentially labeled are indistinguishable in the full MS scan. Upon fragmentation, they generate defined reporter ions in the low mass region of the MS/MS spectra that are used for quantification (Figure 1.3c). The benefit of this approach is its superior multiplexing capabilities because neither LC separation nor full MS spectra suffers from increased complexity. This feature makes isobaric tagging a favorable technique for clinical samples or time‐course experiments. Gygi and coworkers showed that a combination of metabolic and isobaric labeling further increases multiplexing and allows analysis of up to 18 samples simultaneously (Dephoure and Gygi 2012). Another advantage of TMT and iTRAQ quantification is the higher signal‐to‐noise ratio in the reporter ion m/z range of MSn spectra compared to MS‐level quantification. One limitation of isobaric tagging approaches is that not all mass spectrometers are suitable for the analysis because of the reporter ion masses being in the low mass region (3) with a relatively poor reproducibility (CV >50%), such as spectral counting, or more complex approaches for measurement of relatively small fold changes (>1.5) with good reproducibility (CV ≥10%), such as stable isotope labeling by amino acids in cell culture (SILAC). Being a part of protein identification phase, biomarker qualification provides some evidence that there is an association between the abundance of biomarker candidates and the clinical outcome. Specific filtering criteria are usually applied to select a manageable number of candidates to proceed to the verification phase (Makawita et al. 2011). The aim of biomarker verification is to measure the most promising candidates in a large set of samples and exclude false candidates. ELISA and SRM (Picotti and Aebersold 2012) are commonly used assays for biomarker verification. Advantages of ELISA include measurement of low‐abundance proteins, especially in biological fluids of high complexity, such as blood serum. SRM assays facilitate biomarker verification through multiplexing capabilities and measurement of proteins for which immunoassays are not available. Biomarker verification is followed by development of a preclinical assay and biomarker validation. Proper validation includes measurement of each biomarker in hundreds to thousands of samples from multiple centers, standardized blinded analysis, and establishment of reference values for each protein in normal, benign, and disease conditions. Some initiatives, such as the Biomarkers Definitions Working Group (Group 2001), were launched to define the individual stages of  biomarker discovery and provide clear recommendations for each stage (Ransohoff 2004). Finally, a clinical‐grade assay is developed and submitted for approval by the FDA. The list of FDA‐approved clinical assays currently includes more than 200 proteins (Anderson 2010). Most FDA‐approved protein assays utilize ELISA, and not a single MS‐based protein assay has been approved for clinical use yet (Li et al. 2011). In addition, there is no single FDA‐approved protein biomarker that has been discovered by MS and proteomics techniques. Recently approved cancer biomarkers, HE4 protein and PCA3 transcript, were discovered by cDNA microarray‐based differential transcriptomic approaches. Due to the long timeline of biomarker discovery, FDA‐approved protein biomarkers are expected to emerge in the near future from proteomics projects initiated as early as 10 years ago.

68  Proteomics for Biological Discovery

3.4  BIOLOGICAL SAMPLES FOR BIOMARKER DISCOVERY A variety of biological samples can be used to discover clinically useful biomarkers. This variety includes blood, urine, proximal fluids, tissue samples, cell lines, and laboratory animals (Figure 3.2). Blood serum or plasma and urine are the bodily fluids commonly used for routine clinical analysis due to their simple and minimally invasive collection. Since blood and urine contain numerous markers reflecting the changes in patient health status, these samples are an important source of clinical information. Human blood, in spite of being the fluid of choice for diagnostics and a rich source of biomarkers, is the most challenging sample to analyze by proteomic techniques. Blood plasma proteins have a dynamic range of concentrations of more than 10 orders of magnitude, with albumin and cytokines being the most and the least abundant proteins, respectively (Anderson and Anderson 2002). Such a wide dynamic range allows for identification of only high‐ and medium‐abundance proteins, and masks low‐­abundance proteins that may be more specific disease biomarkers. Urine, a liquid by‐product of the body, also seems to be a good source of biomarkers of renal and urological diseases, since only a few organs from the urinary system, such as kidneys and bladder, contribute to its proteome. Besides, urine is collected noninvasively and in large amounts. However, under normal physiological conditions, the total amount of urine proteins excreted daily is less than 150 mg. Since concentration of proteins depends on water excretion, adjustment of concentration by urinary creatinine is essential. Similar to blood, the wide dynamic range of protein concentrations in urine makes analysis of low‐abundance proteins a challenge (Konvalinka et al. 2012). Due to the lower complexity of their proteome and elevated levels of disease‐­ relevant proteins, proximal fluids that surround the diseased tissue or organ can be a better alternative for biomarker discovery. Low‐abundance blood and urine proteins are present at much higher concentrations in proximal fluids (Drabovich and Diamandis 2010; Makawita et al. 2011). Available examples include cerebrospinal fluid for the study of neurodegenerative diseases, amniotic fluid for fetomaternal screening, seminal plasma and prostate secretions for male infertility and prostate cancer detection, and pancreatic fluid for pancreatic cancer studies (Drabovich et al. 2014). However, the use of biological fluids for biomarker discovery has some limitations such as invasive collection procedure, low availability of samples, and small amounts of samples obtained from healthy individuals (for example, cerebrospinal fluid). Another disadvantage of proximal fluids may be concurrent contamination by blood and high‐abundance blood proteins (Kulasingam et al. 2010). Tissue samples collected by biopsy or surgery are another source for biomarker discovery. The major advantage of tissues is the higher concentration of tissue‐ specific proteins and thus easier identification of biomarker candidates. However, distinct limitations of tissue proteomics include low availability of normal tissues and their heterogeneous composition, which challenge quantitative analysis of ­tissue proteins. Laser capture microdissection was introduced to facilitate s­ ampling of homogeneous population of cells in the tissue and thus provide quantification of proteins in epithelial cells, but not in stromal cells or the extracellular matrix (Wisniewski et al. 2012).

Type of sample

Characteristics

Blood

Urine

Sampling Minimally-invasive Availability Study Detection of lowabundance proteins Suitability to develop a clinical assay

Other body fluids

Tissue

Cell line

Animal model

Non-invasive

Invasive*

Invasive

Non-invasive

Invasive

High

High

Medium

Low

High

High

Any disease

Some diseases

Some diseases

Some diseases

One disease

One disease

Difficult

Difficult

Moderate

Easy

Easy

Moderate

Suitable

Suitable

Unsuitable*

Unsuitable

Unsuitable

Unsuitable

Figure 3.2 clinical samples available for biomarker discovery. the major characteristics of clinical samples are displayed, including difficulty of sample collection, availability of samples, number of diseases that can be studied, difficulty of detection of low‐abundance proteins, and suitability of samples for the development of protein‐based clinical assays. *collection of some body fluids, such as sweat, tears, saliva, and seminal plasma can be noninvasive, and the sample can be suitable for the development of a clinical assay. †Proteins can be measured in biopsy‐obtained tissues using immunohistochemistry. (see color insert.)

70  Proteomics for Biological Discovery

A cell culture‐based approach for biomarker discovery relies on the identification of disease‐associated proteins in the cell lysate or cell secretome. The use of cell lines allows for detection of low‐abundance secreted and membrane proteins (Kulasingam et al. 2010). Secreted and membrane proteins play important regulatory and signal transduction roles and can act locally and systemically, which makes them promising biomarker candidates. The main advantages of cell lines include their availability, cost‐effective analysis, and detection of low‐abundance blood proteins in the cell line secretome. In addition, several cell lines can represent different stages of disease progression. However a single cell line will not represent the whole spectrum of disease heterogeneity. Cell line models do not account for tissue microenvironment and possible interactions with neighboring cells and connective tissue. As opposed to primary cells, indefinitely proliferating immortalized cell lines possess mutations or viral vectors that deregulate the cell cycle and thus may affect many signal transduction or metabolic pathways. As a result, an immortalized cell line may be a very simplified or even irrelevant model of a primary cell line (Pan et al. 2009). Laboratory animal models, such as mouse xenografts, facilitate mining for biomarkers expressed in the microenvironment with in vivo interaction between the host mouse tissues and the diseased human cells. Since multiple genetically identical mice can be grown and studied, animal models minimize genetic and environmental variabilities. In addition, sample collection is simple and can be performed at any stage of disease development. In some cases, such as xenograft models of human cancers, the size of the tumor in proportion to the body weight is significantly larger in mice relative to humans, resulting in increased levels of potential biomarkers in circulation and higher likelihood of their identification. The altered tumor microenvironment, however, should be taken into consideration. Proteomic analysis of tissues and biological fluids obtained from genetically modified mice has the same limitations as for human samples. Furthermore, it is still not clear whether every orthologous genomic alteration in mice can be translated into models of human diseases (Frese and Tuveson 2007). 3.5  INTEGRATION OF ‐OMICS APPROACHES TO SELECT BIOMARKER CANDIDATES Recent advances in genomics, epigenomics, transcriptomics, and proteomics has resulted in global profiling of genes, mRNA, and proteins in health and disease. Disease‐specific alterations may be translated to the proteome level not only directly ­ echanisms through transcription and translation, but also indirectly through multiple m including, but not limited to, epigenetic regulation, alternative splicing, noncoding RNA regulation, altered signal transduction, and posttranslational modification (PTM). Disease‐specific alterations at the proteome level may be qualitative, such as single nucleotide polymorphisms (SNPs) resulting in single amino acid substitutions, or quantitative, such as changes in protein abundance. For example, genomic alterations that can be used for protein diagnostics include somatic mutations and SNPs in the protein‐coding genes, protein‐coding gene fusions, copy number gains and losses, disease‐specific alternative splicing, alterations in DNA methylation, histone modification, nucleosome remodeling and miRNA levels, and disease‐specific

Protein Biomarker Discovery: An Integrated Concept  71

DNA

DNA PTMs

mRNA

Protein

- SNPs - Mutations - Copy numbers - Gene fusions

- DNA methylation - Histone modif ication - miRNA

- Alternative splicing - Dif f erential regulation

- Dif f erential regulation - Protein isof orms - Protein variants

Genomics Epigenetics Transcriptomics Proteomics Figure 3.3  Integration of genomic, epigenetic, transcriptomic, and proteomic approaches to select protein biomarker candidates. Disease‐specific alterations are translated to the proteome level not only directly through transcription and translation, but also indirectly through multiple mechanisms, including, but not limited to, epigenetic regulation, alternative splicing, noncoding RNA regulation, altered signal transduction, and posttranslational modification. Global proteome analysis identifies changes in protein abundances, while the exact mechanism may not be known. Integration of multiple ‐omics technologies provides complementary biomarker candidates and also facilitates elucidation of molecular mechanisms resulting in disease‐specific alterations of the proteome.

­ rotein  isoforms or PTMs (Figure  3.3). Qualitative genomic alterations, such as p SNPs, are  usually not amenable to measurement by protein immunoassays, while MS‐based SRM assays facilitate analysis of SNPs and other subtle genomic alterations. Integration of disease‐specific genomic, epigenomic, transcriptomic, and ­proteomic alterations provides a comprehensive approach to discovery of protein biomarkers as well as elucidation of molecular mechanisms leading to such qualitative or quantitative changes (Dimitrakopoulos et al. 2016). Completion of the Human Genome Project provided the reference genome in 2001 (Venter et al. 2001). Since then, further elements of the human genome have been investigated by large‐scale international projects. For example, the International HapMap Project was intended to identify genes and SNPs that affected both health and disease conditions (Frazer et  al. 2007). The Encyclopedia of DNA Elements (ENCODE) Project was aimed at discovering functional elements in the human genome sequence, such as regions of transcription, chromatin structure, transcription factor association, and histone modifications (Dunham et al. 2012). The 1000 Genomes Project was launched to discover genetic variations, such as SNPs, insertions, deletions, and copy number variations, in 1000 individuals from 14 different populations (Abecasis et al. 2012). Causal mutations for more than 4000 mendelian disorders were cataloged in the Human Gene Mutation Database (Stenson et  al. 2009). The International Cancer Genome Consortium will provide a comprehensive description of genomic, transcriptomic, and epigenomic changes in 50 different tumor types and subtypes (Hudson et al. 2010). Similar large‐scale projects included the Cancer Genome Project initiated in the UK (Campbell et  al. 2008) and the Cancer Genome Atlas project initiated in the US (Network 2008, 2011; Hammerman et al. 2012; Network 2012a, b). Epigenomic changes in disease are currently being investigated by the NIH Roadmap Epigenomics Mapping Consortium (Bernstein et  al. 2010). Differential transcriptomic changes and mRNA alternative splicing

72  Proteomics for Biological Discovery

under different biological conditions are compiled in the microarray‐ and RNA sequencing‐based NCBI Gene Expression Omnibus (GEO)(Barrett et al. 2010) and EBI Array Express (Kapushesky et al. 2012) databases. Finally, the Human Proteome Project launched in 2011 was presented as a global effort to catalog abundance, subcellular localization, and function of all human proteins (Legrain et al. 2011). Integration of search results obtained with several ‐ omics databases can be used to generate a comprehensive list of biomarker candidates (see Figure 3.3). Such integration would provide certain complementarity for selection of candidates. For example, it has been shown that mRNA levels explain only ~40% of variation of protein levels, so only a fraction of proteins significantly dysregulated in disease will have significantly dysregulated mRNA transcripts (Schwanhausser et al. 2011). In addition to genetic, epigenetic, transcriptomic, and proteomic changes, tissue‐ specific expression of genes and proteins can be used as a complementary factor to search for biomarkers with high specificity. Abnormal changes in concentration of tissue‐specific proteins may indicate an ongoing pathological process in the disease‐ relevant organ or tissue. For example, the success of PSA protein, is mostly due to its tissue specificity. Tissue‐specific transcripts and proteins are cataloged in a number of databases including BioGPS (www.biogps.org), Human Protein Atlas (www.pro teinatlas.org) and GTEx Portal (https://gtexportal.org). The BioGPS database is based on mRNA expression profiles of all human genes in 84 tissues and cells, while the Human Protein Atlas is the most comprehensive proteomic database; its current version 18.1 includes immunohistochemistry‐based protein expression profiles of ~87% of the human protein‐coding genes ( almost 17 000 genes based on more than 26 000 antibodies). The immense ‐omics database has not yet been utilized to its full extent for the purpose of clinical diagnostics. Little is currently known as to which genomic, epigenetic, and transcriptomic alterations in disease are translated to the protein level. Furthermore, little is known about which of these alterations can be measured in biological fluids and tissues and thus be used for clinical diagnostics. There is little doubt that extensive integration of multiple ‐omics technologies will facilitate protein biomarker discovery through stratification of patients and disease subtypes, more targeted and rational design of discovery pipelines, complementarity of ­alterations at different ‐omics levels, and elucidation of molecular mechanisms contributing to protein biomarkers. 3.6  PROTEIN IDENTIFICATION BY MASS SPECTROMETRY In recent years, shotgun proteomics has been extensively used to identify thousands of proteins in various biological samples. Until now, the vast majority of proteomic data has been generated by tandem mass spectrometry (MS/MS). Tandem mass spectrometers are instruments with more than one analyzer, which perform multiple stages of mass analysis separation. These instruments can be classified into two categories: tandem‐in‐space (e.g., QqQ and Q‐TOF) and tandem‐in‐time (e.g., ion traps and Orbitraps). Ion sources, such as electrospray ionization (ESI) (Fenn et  al. 1989) or matrix‐assisted laser desorption/ionization (MALDI) (Tanaka et al. 1988), are coupled to the mass spectrometers to ionize proteins or peptides. Protein identification

Protein Biomarker Discovery: An Integrated Concept  73

approaches include top‐down and bottom‐up methods. In top‐down proteomics, intact proteins are fractionated and ionized and then ions undergo multiple stages of gas‐ phase fragmentation inside the mass spectrometer. Distinct advantages of top‐down proteomics include analysis of protein isoforms and PTMs. Even though there were significant achievements in top‐down proteomics and increasing numbers of protein identifications (Tran et al. 2011), this approach is still too complex, less efficient than bottom‐up proteomics and is not currently used for biomarker discovery. Bottom‐up proteomics approaches provide unsurpassed capabilities in terms of the number of protein identifications. Quantification of complete proteomes of 11 mammalian cells (>8000 proteins) in just a few hours has been claimed using bottom‐up proteomics approaches combined with high‐resolution quadrupole‐Orbitrap mass spectrometer (Geiger et al. 2012; Mann et al. 2013). In a typical bottom‐up proteomics experiment, proteins in a complex mixture are denatured, and the thiol groups of cysteine residues are reduced and alkylated, to prevent reformation of disulfide bonds (Drabovich et al. 2013b). Following that, proteins are cleaved using a proteolytic enzyme, such as trypsin, into relatively short peptides, and peptides are subjected to the first‐dimension separation, typically strong‐cation exchange (SCX) chromatography. Since each tryptic peptide has a net charge of +2 or higher in the acidic pH (due to the basic N‐terminus and lysine or arginine residues at the C‐terminus), tryptic peptides are separated from neutral or singly charged molecules. During the second separation step, which typically involves reverse‐phase chromatography, peptides in each SCX fraction are separated based on their hydrophobicity. Following two‐dimensional separation, positively charged peptides are desolvated and transferred to the mass spectrometer for gas‐phase separation, isolation, and collision‐induced fragmentation by inert gas molecules. Upon detection of fragment ions, the resulting MS/MS spectra are assigned to peptide sequences using bioinformatics search algorithms such as database matching, de novo sequencing or hybrid approaches (Nesvizhskii et al. 2007). For the large‐scale proteomics studies, database matching remains the most frequently used method for peptide and protein identification. Each acquired fragment ion spectrum is matched to the theoretical spectrum predicted for each peptide in a protein sequence database. A scoring system is used to calculate the probability of the match between experimental and theoretical spectra, and candidate peptides are ranked according to the computed search score. Finally, the best scoring peptide match is selected for a subsequent statistical analysis, which estimates the false‐positive rate of peptide and protein identifications. Conventional software tools for peptide and protein identification include MASCOT (Perkins et  al. 1999), X!Tandem (Fenyo and Beavis 2003), SEQUEST (Eng et al. 1994), ProteinProspector (Chalkley et al. 2005), and Andromeda (Cox et al. 2011). Proteome composition of many biological fluids has been extensively studied (Anderson et al. 2004; Adachi et al. 2006; Kuk et al. 2009; Hudson et al. 2010; Schiza et al. 2019). Clinical applications remain the major motivation for the in‐depth proteomic analysis of biological fluids. For example, extensive analysis of seminal plasma proteome was conducted in order to facilitate diagnostics of urogenital diseases such as male infertility, prostatic inflammation, and prostate cancer ­ (Pilch and Mann 2006; Batruch et al. 2011, 2012). Several biomarkers for the noninvasive differential diagnosis of male infertility were verified and validated (Drabovich et al. 2011, 2013a; Korbakis et al. 2017; Schiza et al. 2014). Analysis of

74  Proteomics for Biological Discovery

the proteome of amniotic fluid represents another example of extensive identification of proteins in biological fluids. Amniotic fluid is a liquid that protects the fetus against mechanical and thermal shock, possesses antimicrobial activity, and contains nutritional factors (Tisi et al. 2004). Fetal and pregnancy‐related proteins in the amniotic fluid have great potential as specific biomarkers for fetal diseases or complications of pregnancy (Cho et al. 2011). Recent studies identified more than a thousand of proteins in the amniotic fluid of pregnant women carrying chromosomally normal fetuses and fetuses with Down syndrome at different gestational ages, and the subsequent verification of the most promising candidates has been completed (Cho et al. 2007, 2011; Martinez‐Morillo et al. 2012). Integrated proteomic approaches for biomarker discovery combine different biological sources and maximize the likelihood of discovering true‐positive ­markers. Since each type of biological sample has its own advantages and disadvantages, it  has been hypothesized that multiple approaches will complement each other and facilitate discovery of useful biomarkers. For example, an extensive search for pancreatic cancer biomarkers was performed through integration of proteomic analyses of secretomes of six pancreatic cancer cell lines, a near‐normal human pancreatic ductal epithelial cell line, and two pools of pancreatic fluid samples (Makawita et al. 2011). In total, 3479 nonredundant proteins were identified with high confidence, of which 40% were extracellular or cell membrane bound. To complement this study, a proteomic analysis of ascites fluid from patients with advanced pancreatic ductal adenocarcinoma and presence of metastases was completed (Kosanam et al. 2011). In total, 816 proteins were identified using four different methods of peptide fractionation. Proteomes of cell line‐conditioned media, pancreatic fluid, and ascites were compared and filtered through several criteria, including tissue specificity (Chan et al. 2013). The most promising biomarker candidates identified through this integrative approach are currently being ­subjected to further verification and validation. 3.7  PROTEIN QUANTIFICATION BY MASS SPECTROMETRY In general, methods of quantitative proteomics can be classified into four groups based on the type of protein or peptide labeling: label‐free quantification, metabolic labeling, chemical labeling, and the use of synthetic heavy isotope‐labeled peptide internal standards. Label‐free quantification methods rely on spectral counting or measurement of signal intensity and integrated area of the chromatographic peak. Spectral counting is based on the observation that more abundant peptides will be selected for fragmentation in the data‐dependent acquisition mode and will result in the number of  MS/MS spectra proportional to the amount of protein (Neilson et  al. 2011). Advanced spectral counting techniques, such as Protein Abundance Index (PAI/ emPAI) (Rappsilber et al. 2002; Ishihama et al. 2005), Absolute Protein Expression (APEX) (Lu et  al. 2007), Normalized Spectral Abundance Factor (NSAF) (Zybailov  et  al. 2006), and Normalized Spectral Index (SIN) (Griffin et  al. 2010) account for  additional parameters, such as the number of theoretical tryptic ­peptides per protein. In general, label‐free quantification is a straightforward, fast

Protein Biomarker Discovery: An Integrated Concept  75

and affordable option for protein quantification (Neilson et  al. 2011). However, more accurate and precise techniques are required for biomarker verification. Among the available label‐based quantitative techniques, SILAC and isobaric tags for relative and absolute quantification (iTRAQ) are extensively used for biomarker identification and qualification. SILAC relies on the metabolic labeling of the cell proteome using a mixture of 13C‐ and 15N‐labeled arginine and lysine (Ong et al. 2002). Since protein digestion by trypsin results in peptides with C‐terminal arginine and lysine residues, all tryptic peptides except the C‐terminal peptide will include 13C‐ and 15N‐isotope labels. Cells grown in the culture media with heavy isotope‐labeled amino acids (“heavy” media) will contain >98% heavy proteins after six cell doublings. “Heavy” cells are subjected to certain perturbations or hormonal stimulation while cells grown in the media with nonlabeled amino acids (“light” media) are used as a control. Finally, “heavy” and “light” cells are lysed, and cell lysates are mixed in an equimolar ratio prior to liquid chromatography‐mass spectrometry (LC‐MS) analysis. Following that, heavy and light forms of peptides are identified, and the intensities of peptide ions are used to  determine relative abundances of the heavy peptides and estimate relative ­abundance of the corresponding protein (Figure 3.4a). Apart from accurate quantitative capabilities, advantages of SILAC include easy implementation and its use to study PTMs or dynamic protein turnover. SILAC, however, has a relatively narrow dynamic range of around 20‐fold and can hardly be applied to the slowly proliferating primary cells. SILAC‐based metabolic labeling is also used to generate heavy isotope‐labeled proteome, which may be spiked in as a relative standard to quantify proteins in biological fluids and tissues. Such approaches are implemented in the SILAP (Yu et al. 2009), super‐SILAC (Geiger et al. 2010) and SILAC mouse (Kruger et al. 2008). In super‐SILAC, a mixture of several SILAC‐labeled cell lines is used as an internal standard for quantification of proteins in tissue extracts. The major advantages of super‐SILAC are the low cost and good compatibility with existing SILAC workflows. However, the mixture of cell lines cannot provide the complete coverage of proteins expressed in the tissue of interest. iTRAQ allows for protein identification and relative quantification using MS/MS peptide fragments and low mass reporter ions (Ross et al. 2004). Amine groups at the peptide N‐terminus and lysine side chain are labeled with isobaric tags that include a reporter group, a mass balance group, and a peptide‐reactive group. The reporter groups range from 113 to 121 m/z, the balance groups range from 32 to 24 m/z, while the combined m/z ratio is kept constant. Following fragmentation of the tag by collision‐induced dissociation (CID), the balance groups are lost as n ­ eutral fragments and reporter groups acquire different m/z ratios. All other sequence‐ informative fragment ions remain isobaric, and their signal intensities are additive. The relative abundance of the peptides is thus deduced from the relative intensities of the corresponding reporter ions (Figure 3.4b). While often used in biomarker discovery studies, iTRAQ has some limitations, such as underestimation of protein fold changes and interference by cross‐label isotopic impurities (Evans et  al. 2012). Alternative quantitative approaches based on chemical labeling of intact proteins include isotope‐coded affinity tagging (iCAT) (Gygi et al. 1999) and tandem mass tags (TMTs) (Thompson et al. 2003).

76  Proteomics for Biological Discovery Lysine (12C) Arginine (14N)

Light lable

Heavy lable

Lysine (13C) Arginine (15N)

(b)

Samples

Proteins

Peptides

Digestion

Cell with heavy isotope labeled proteins

Cell doubling times (≥6)

Intensity (counts)

Mix 1:1

Target proteins

6-10 Da

MS2

Sample 1 N -

-C

Sample 2 N -

-C

113

32

-C

121

24

-C

Neutral loss

8 Da 121 113

w

R

(13C, 15N) (13C, 15N)

Heavy labeled peptide

Protein expression

Proteins

Peptides

Digestion

QconCAT

Chromatogram

Light

R K R

Heavy R

Sample

MS1

Intensity (counts)

K K

Medium with isotope labeled amino acids K

z

(d)

Target peptides

Expression vector

y x

m/z

K

Gene

PRG

Reporter Balance

Labeling

Digestion

Cell

group PRG

Quantitation Identification Fragmentation

m/z

Gene construct

Peptide reactive

32

113

Sample 2 Sample 1

Peptide 1

MS1

(total protein)

(c)

Sample 2 Sample 1

TAGs

Isobaric 145 m/z labels 121 24

Intensity (counts)

(a)

6-10 Da

m/z

Retention time (min)

Figure 3.4  Approaches for quantitative proteomics. Commonly used approaches include (a) stable isotope labeling by amino acids in cell culture (SILAC). Cells are cultured in light (12C and 14N isotopes) or heavy media (13C and 15N isotopes). After six cell doublings, cells cultured in heavy media will contain mostly heavy isotope‐labeled proteins (>98%). Following cell culturing,“heavy” and “light” cells are lysed, and cell lysates are mixed in equimolar amounts prior to mass spectrometry analysis. SILAC labeling results in a mass shift of 6–10 Da for each tryptic peptide of the “heavy” protein. (b) Isobaric tags for relative and absolute quantification (iTRAQ). Amine groups are labeled with isobaric tags, which include a reporter group (e.g., 121 and 113 m/z), a mass balance group (e.g., 24 and 32 m/z), and a peptide‐reactive group, while keeping the combined m/z ratio constant. Following fragmentation of the tag, the balance groups are lost as neutral fragments and reporter groups acquire different m/z ratios. All other sequence‐informative fragment ions remain isobaric, and their signal intensities are additive. (c) Quantification concatemers (QconCAT). A gene for the target peptides is constructed, inserted into a plasmid vector and expressed in the medium containing heavy isotope‐labeled arginine and lysine.Then, the concatemer with the proteotypic peptides is produced and spiked into the sample prior to digestion. (d) Absolute quantification (AQUA) peptides. Heavy isotope‐labeled peptides with identical physical and chemical properties as endogenous proteotypic peptides are spiked into the sample and used to quantify proteins. (See color insert.)

Biomarker verification requires accurate relative or absolute quantification of proteins. Absolute quantification methodologies are aimed at measuring protein concentration in biological samples, enabling the comparison of data between ­laboratories and between different analytical methods. In general, stable isotope dilution‐based quantification provides excellent reproducibility, linear response, and precision regardless of the quantification standards used (Brun et  al. 2009). Protein standard absolute quantification (PSAQ) standards, quantification concatemer standards (QconCAT), and proteotypic heavy isotope‐labeled peptides with a trypsin‐cleaved quantification tag facilitate absolute quantification of proteins.

Protein Biomarker Discovery: An Integrated Concept  77

Full‐length isotope‐labeled proteins are the ideal internal standards for absolute quantification (Brun et al. 2007). Such standards retain most physical and chemical properties of the intact protein during preanalytical sample preparation, protein and peptide separation, and peptide ionization prior to quantification by MS. PSAQ standards, however, do not carry PTM and their production is laborious (Brun et al. 2009). QconCAT strategy is based on the production of a concatemer, an artificial protein that contains several proteotypic peptides (Beynon et al. 2005). A gene of such construct is subcloned into a vector and expressed in a medium containing heavy isotope‐labeled arginine and lysine. Concatemers, similar to intact protein standards, are spiked into the sample prior to trypsin digestion (Figure  3.4c). This approach is a cost‐effective approach since as many as 50 tryptic peptides can be included into a single QconCAT construct (Rivers et al. 2007). Even with the use of QconCAT standards, the completeness of tryptic digestion has to be assessed for each proteotypic peptide (Brun et al. 2009). Finally, the absolute quantification (AQUA) strategy (Gerber et al. 2003) is used to quantify proteins through spiking heavy isotope‐labeled peptides that have identical physical and chemical properties as endogenous proteotypic peptides (Figure  3.4d). Although fast and straightforward, the AQUA approach does not account for the variability of trypsin digestion (Brun et al. 2009). Selected reaction monitoring assays are often used for quantification of proteins and verification of biomarker candidates (Cho et  al. 2013;  Drabovich et  al. 2012, 2016; Karakosta et al. 2016; Korbakis et al. 2015). In general, SRM is a quantitative analytical assay performed using a triple‐quadrupole mass spectrometer. A typical SRM assay includes liquid chromatography separation of proteotypic peptides, ionization of peptides with the electrospray source, filtering of peptides in the first quadrupole, fragmentation of peptides in the second quadrupole, filtering of peptide fragments in the third quadrupole, measurement of three fragment ion intensities, and integration of the three signals (Picotti and Aebersold 2012). With state‐of‐the‐art SRM assays, up to 100 peptides representing 100 medium‐to‐high abundance proteins can be measured simultaneously in the unfractionated digest of biological fluid, while achieving coefficients of variation under 20%. Addition of stable isotope‐labeled PSAQ, QconCAT or AQUA standards can aid in increasing the reliability of the quantification as it enables absolute quantification and determination of the correct analyte in the presence of co‐eluting ­peptides. Combination of SRM assays with SISCAPA (Stable Isotope Standards and Capture by Anti‐Peptide Antibodies) ­technology promises to allow for verification of low‐abundance biomarker candidates in clinical samples, including urine and blood plasma (Whiteaker et al. 2011; Kuhn et al. 2012). 3.8  BIOMARKER VERIFICATION Protein identification and qualification phases (see Figure 3.1) of the biomarker discovery pipeline supply long lists of putative biomarkers. As our experience shows, most of these biomarker candidates are found differentially expressed due to the systematic biases occurring at the identification phase. Such biases include analytical bias resulting from variations in the sample preparation protocols and label‐free quantification, as well as biological bias, such as large inter‐ and intra-individual

78  Proteomics for Biological Discovery

­ iological variability. Verification phase should eliminate the majority of false posib tives and supply only a few candidates for subsequent validation. Most common flaws of biomarker discovery projects include data overfitting, the lack of multiple hypotheses testing in the statistical analysis and poor estimation of the sample size required for identification and verification phases. Overfitting is a result of data analysis in which a very large number of parameters are measured in a small set of independent samples, and a model is used to find a multiparametric pattern. To avoid overfitting, any pattern discovered in the training set of samples should be verified by cross‐validation or tested in an independent set of samples (Simon et al. 2003). Since multiple proteins are measured at the identification and verification phases, multiple hypotheses testing should be performed, and the false discovery rate‐adjusted p‐values should be reported instead of p‐values (Pencina et al. 2010; Jenkins et al. 2011). Regarding the number of samples, statistical estimations show that there should be at least 10 events, or independent samples, per variable, or protein, measured (Peduzzi et al. 1996). A biomarker discovery approach that includes identification of several thousand proteins in only one normal and one disease sample followed by verification of just a few of those candidates (typically those proteins for which immunoassays are available) has a very low chance of success due to a very high false‐positive rate. Indeed, simulations show that if we assume a 10‐fold intra-individual variation of a protein concentration in 95% of samples, a log‐normal distribution, and a two-fold change cut‐off, measurement of differential expression of proteins in one normal and one disease sample will result in a false‐positive rate of 21%. With 1000 proteins measured, this equals about 206 false‐positive candidates. Assuming that 1000 proteins also contain 20 true positive biomarker candidates, a list of top 20 candidates will include mostly false positives (around 18 false‐positive and two true‐positive candidates). However, if we measure 1000 proteins in 10 normal and 10 disease samples, the false‐positive rate will drop to 1.2%. As a result, a list of top 20 candidates will include mostly true positives (around 13 true‐positive and seven false‐positive candidates). Even though biomarker identification is often completed in tissues, cell lines or proximal fluids, the verification phase would require biomarker measurement in the biological sample suitable for use in the clinic (Pepe et al. 2001). Clinical parameters such as age, sex, concurrent medications, etc. should be collected and matched for control and disease groups. In addition, sample collection and preparation should be standardized to exclude possible bias due to protein precipitation or instability. 3.9  BIOMARKER VALIDATION Biomarker validation is set to assess only the few most promising candidates that were previously verified, and for which accurate and precise analytical assays were developed. Several different definitions of validation are used in the literature (Ransohoff 2004), so the concept of validation in biomarker discovery should be further developed and standardized (Ransohoff 2004; Buyse et al. 2010). Validation protocols are very well developed and strictly regulated in the drug discovery field but still not well accepted or standardized in the field of biomarker discovery (Ransohoff 2004). The validation phase must acknowledge the final clinical use of

Protein Biomarker Discovery: An Integrated Concept  79

the biomarker and include prospective and retrospective validation or validation for general population screening. Clinical parameters of samples used for validation should be well defined and include age, sex, race, concurrent medications, and lifestyle habits, such as smoking. Such considerations are set to exclude biases when some parameters of disease are associated not with the primary cause of disease, such as cancer, but with some ­secondary features of disease, such as concomitant inflammation or stress. Bias in the sample collection is a major challenge for proper validation. It is not uncommon for healthy control samples to be collected in a general clinic with a relatively low psychological stress, while disease samples are usually obtained in specialized clinics prior to surgery, which adds additional stress to patients and may result in discovery of biomarkers of stress, rather than disease. Clinical cohorts for validation should include samples with different subtypes and stages of disease since biomarkers ­typically perform better at the late stage. Even though many analytical assays are available for protein quantification, immunoassays still dominate in the field of clinical proteomics. Due to low sensitivity and low throughput, MS‐based assays cannot as yet compete with ELISA. To be competitive, MS‐based validation assays should be able to analyze thousands of clinical samples with an interday CV 25 Å) did not have any noticeable effect on the simulations (when compared to nonconstrained simulations). Thus, both short‐distance nonselective cross‐links and the DMD algorithm were essential for the success achieved by the approach. CL‐ DMD is an independent method for the determination of unknown protein structures by cross‐linking, and we hope that it will play a role in the protein structure determination field, especially for cases where standard structural biology methods cannot be applied. 7.7.4  Proteome‐Wide Interaction Networks Another exciting application that has recently received a lot of attention is the proteome‐wide determination of the protein–protein interaction networks by in-vivo cross‐linking (Tang and Bruce 2010). Incremental improvements in high‐mass accuracy high‐performance MS instrumentation, combined with developments in the cross‐linking reagents, techniques, and data analysis, have finally made it possible to tackle proteome‐wide cross‐linking experiments. The cross‐linking in vivo of whole organelles, cells, or even tissues allows one to capture all of the proteome‐wide ­protein–protein interactions that exist at the moment of reaction. The problem of the combinatorial nature of the cross‐linked peptides, as mentioned above, is exacerbated in this case, as every cross‐link can potentially consist of pairs of peptides from the entire proteome. A critical feature of cross‐linking reagents that allows one to address this problem is the MS/MS cleavage of the cross‐linker spacer. The cleavage produces individual peptides that constitute the cross‐link, which subsequently can be analyzed using standard proteomic approaches for peptide identification. Another challenge of the approach is that, for each cross‐link, the identification of each of the interacting proteins is based on a single peptide. Thus, confident identification requires cross‐link‐specific MS signatures, rich MS/MS fragmentation, and robust search and validation algorithms. In addition, cross‐linked and digested samples are inherently complex, so the cross‐link enrichment and fractionation strategies described above are required to increase the number of identifications. We have recently performed a proof‐of‐ principle cross‐linking analysis of intact mitochondria (~1200 proteins) (Rudashevskaya et al., unpublished data). In this experiment, yeast mitochondria were cross‐linked with CBDPS, lysed in hypotonic buffer, and the soluble and membrane‐bound proteins were separated by centrifugation and digested with trypsin. Digests were fractionated by SCX chromatography, and the cross‐links were affinity enriched on immobilized avidin. The samples were analyzed using LC‐MS/MS on an Orbitrap Fusion Lumos mass spectrometer. Precursors with doublets of signals due to the isotopic coding of the cross‐linker were selected on the fly for CID

Cross-linking Applications in Structural Proteomics  191

MS/MS fragmentation using the mass tags acquisition method. Data were analyzed using the Kojak search engine (Hoopmann et al. 2015), and cross‐link identifications were validated by Percolator (The et al. 2016). Over 100 000 cross‐link signals were detected, out of which about 10 000 inter-peptide cross‐links and over 600 protein interacting pairs were identified with a false discovery rate of 3% of all purifications), including ribosomal proteins, were removed from further consideration. To assign confidence scores, the MS data were combined and subjected to machine learning procedures to stringently prune spurious interactions. To this end, we performed training and cross‐ validation using a reference set, or “gold standard,” of manually curated yeast protein complexes downloaded from the popular Munich Information Center for Protein Sequences (MIPS) database. However, this gold standard was small and biased (i.e., not fully representative), with only 68 annotated complexes, which likely affected performance. It is also not easy to define clearly which pairs of proteins actually interact directly within a complex. Nevertheless, using a median interaction probability precision of ~0.7, we identified a core set of ~7000 putative high‐­confidence PPI (i.e., bait–prey interactions using the so‐called “spoke” model) among over 2700 proteins, which covered roughly half the expressed yeast proteome. 8.4  DEFINING COMPLEX MEMBERSHIP BY NETWORK PARTITIONING A multiprotein complex in a network can be viewed graphically as a subnetwork or module of proteins (vertices) that are densely connected (edges), with fewer links to the rest of the network. In general, the algorithms used to define protein complexes from interaction networks can be broadly differentiated between classic similarity or distance‐based methods and the more recently developed graph‐based approaches. Our studies focused mainly on the latter since these have a strong mathematical/ theoretical basis. For our yeast work, we applied the popular MCL algorithm (Enright et al. 2002), which simulates random “walks” within graphs (wherein nodes are selected at random and a fixed number of PPI “edges” is crossed) to estimate relatedness. Through an iterative process of many such walks, the algorithm splits the proteins into exclusive groups based on the flow across more highly traversed regions, with high connectivity being a sign of clusters (i.e., complexes). Using clustering parameters that optimize overlap with the MIPS reference complexes (Mewes et al. 2004), we predicted a set of 547 nonoverlapping soluble protein complexes (Krogan et al. 2006), of which over half had not been previously reported. Extensive bioinformatics analyses based on examining the semantic similarity (i.e., tendency of pairs of proteins to have a similar functional annotation or colocalization) and evolutionary coconservation (since proteins in a complex tend to coevolve) of the components of the clusters supported the overall reliability of the complexes we defined. Yet while MCL proved to be effective with our yeast network, it is not suitable for module detection in very sparse networks where it tends to identify ­heterogeneous megacomplexes (Hu et al. 2009). A companion genome‐wide yeast tandem affinity purification‐mass spectrometry (TAP‐MS) study by Gavin and collaborators, published in parallel to our work

202  Proteomics for Biological Discovery

(Gavin et al. 2006), reported data of a similar scope, covering ~3000 purified fusion proteins and a similar number of complexes, of which again half were new (i.e., not documented in the literature) (Gavin et al. 2006). As per our strategy, this competing group also provided strong independent supporting functional evidence based on protein coexpression, colocalization, evolutionary coconservation, known multiprotein structures and previously published binary interaction data. Remarkably, however, despite using the same standardized experimental approach, the two studies showed only limited overlap in terms of the composition of the complexes reported, providing seemingly different takes on the yeast interactome that suggested systemic artifacts or errors (Goll and Uetz 2006). Subsequent reanalyses of the raw data produced in both studies demonstrated that the major source of discrepancy arose during data processing (Collins et  al. 2007; Hart et al. 2007; Pu et al. 2007), particularly the different approaches used to score the PPI networks and the different module detection algorithms applied to the networks, suggesting that careful comparison and uniform benchmarking of different methods are still needed. Where the two studies were fully con­sistent  was in terms of their limited coverage of membrane‐associated proteins, which were inefficiently extracted using buffers optimized for soluble protein complexes. 8.5  AFFINITY PURIFICATION/MASS SPECTROMETRY ANALYSIS OF MEMBRANE PROTEIN COMPLEXES IN YEAST In a collaborative follow‐up study (Babu et al. 2012) by our group and the Greenblatt laboratory, together with both computational and membrane biologists, we reported the successful application of nonionic detergents to facilitate the solubilization and ­subsequent AP of yeast membrane‐associated proteins. After evaluating ­dozens of candidates, we settled on three different detergents (Triton X‐100, DDM (n‐dodecyl‐β‐D‐ maltopyranoside), and C12E8 – octaethylene glycol monododecyl ether), which were effective at extracting nearly 1600 putative yeast membrane proteins. After solubilization and detergent removal, MS and computational analyses of the polypeptides that stably copurified with the baits revealed a physical map of 1726 high‐confidence membrane protein–protein associations. To generate a comprehensive proteome‐wide map, we then combined these new data with our soluble yeast protein AP/MS data (Krogan et al. 2006). Our integrated network consists of over 13 000 PPI among roughly two‐ thirds of the yeast proteome detectable by MS. In total, after clustering using MCL, we defined 501 putative complexes with one or more membrane protein subunits (Figure 8.1). As with our soluble protein network, we made all of the yeast membrane protein interaction data publicly accessible via the web (http://wodaklab.org/membrane) to facilitate community exploitation of this resource. 8.6  AFFINITY PURIFICATION/MASS SPECTROMETRY ANALYSIS OF THE INTERACTOME OF THE MODEL BACTERIUM ESCHERICHIA COLI Bacteria are the dominant species on Earth, and E. coli is widely recognized as one of the most intensively studied microbes, but the composition of bacterial protein complexes had not been studied in a systematic manner. Hence, in a parallel

Functional Proteomics  203

Figure 8.1  Graphical representation of putative yeast membrane‐associated protein complexes (constituent subunits similarly colored as nodes), demarcated according to subcellular localization annotations, which were inferred from an integrated network of high‐confidence PPI (edges) mapped by large‐scale AP/MS experiments (Babu et al. 2012). (See color insert.)

c­ompanion study to our yeast work, we again collaborated with the Greenblatt group to apply an analogous AP/MS strategy to characterize the protein complexes of E. coli on a large scale (Hu et al. 2009). Using a recombineering method to integrate a selectable tagging cassette into the chromosome of a K12 laboratory isolate, we generated a genome‐scale collection of affinity‐tagged E. coli strains, this time adopting a smaller (8 kDa instead of the 20 kDa TAP tag used in yeast) tandem sequential peptide affinity (SPA) tag (Zeghouf et al. 2004) to minimize possible perturbations to protein function. Unlike the original TAP procedure, wherein roughly 18% of essential yeast proteins did not tolerate a C‐terminally TAP tag (Gavin et al. 2002), we had higher success (>95%) in SPA tagging and purifying essential E. coli proteins (Babu et al. 2009a; Butland et al. 2005), implying the smaller tag was less likely to interfere with protein folding or function. Ultimately, we were able to process 1476 different E. coli SPA‐tagged strains, of which 1241 protein baits were soluble and successfully identified (i.e., ~84% success rate), again using a combination of MALDI‐TOF and LC‐MS/MS analyses. Although biased against membrane proteins, we managed to identify low‐abundance soluble proteins present down to less than one copy on average per cell (Taniguchi et al. 2010). As with our yeast study, we subjected the MS data to machine learning to segregate nonspecific binders from genuinely associated copurifying proteins. As  positives for training the algorithms, we used interactions derived using traditional low‐throughput methods curated from the literature in public PPI databases like DIP, BIND, and IntAct, while negatives were defined as pairs of bacterial

204  Proteomics for Biological Discovery

­ roteins with different subcellular localization annotations. The final weighted netp work, with a minimum likelihood score of ≥ 0.75, consisted of ~6000 PPI among 1757 ­soluble E. coli proteins. As with yeast, all these data are accessible via the web (http:// ecoli.med.utoronto.ca). We also managed to substantiate our E. coli physical network against alternate existing evidence consistent with functional associations such as gene annotations, coexpression, coconservation, etc. Graph analysis of the high‐confidence network indicated that the previously unannotated proteins (i.e., functional “orphans”) tended to have a lower overall connectivity and betweenness centrality, which is a measure of the number of shortest paths (fewest connections) going through a given node, relative to annotated E.  coli proteins, suggesting more less influential “peripheral” positions that may reflect a later evolutionary addition to the network. Conversely, however, the orphans had on average similar overall clustering coefficients, implying that in ­general they are functionally connected to, rather than isolated from, well‐studied biological systems. These observations suggested that a careful consideration of both the individual pairwise associations and the overall placement of orphans within the network would provide insight into biological function. As with our yeast studies (Babu et al. 2012; Krogan et al. 2006), we partitioned our set of high‐confidence 5993 E. coli protein interaction networks using MCL. The clustering parameters were optimized by generating a compromise between cluster efficiency (balance between the probability weights of interactions captured within the clusters versus the average cluster size) and mass fraction (fraction of interactions connecting protein nodes within the same cluster). In this way, we projected a map of 443 putative E. coli soluble multiprotein complexes, most of which consist of 2–4 bacterial polypeptides. As in yeast, over half (244 or 55%) contained at least one  unannotated subunit, with linkages implying a biological role. Moreover, the ­complexes were significantly (p 60%) unannotated proteins with high confidence (probability >0.5), some of which we were able to validate by alternate experimental approaches. As independent validations of our predictions, we performed some targeted assays testing functional dependencies (i.e., epistasis) inferred from complex ­membership. Since such piecemeal assays are hard to scale up, we also examined more global properties of the network. For example, we showed that genes encoding putative interacting proteins had elevated coconservation (i.e., higher mutual ­information scores indicating correlated phylogenetic profiles based on coabsence/ presence between pairs of interacting proteins across bacterial genomes) and more highly correlated gene coexpression patterns. 8.11 CONCLUSIONS: SURVIVING THE TRENCHES Over the last decade, our high‐throughput studies have not only led to the discovery of thousands of previously uncharacterized interactions, and many hundreds of novel components and unexpected protein complexes for model organism interactomes, but have also shed light on the global molecular and functional organization of eukaryotic and prokaryotic cells. The generation of high‐confidence interaction networks using both tag‐based and, most recently, tagless experimental strategies, combined with integrative computational scoring strategies, has also allowed us to define the biological roles of hundreds of previously uncharacterized proteins, many

(b)

Proteomic data

Genomic context-based inferences i) Gene fusions*

mass spectra

preys (e.g. annotated) A

B

D

purify

C identify bait (e.g. orphan)

LCMS* MALDI*

Genome 1 A

B

C

Genome 2 A Genome 3 A

B

C

…

B

D

ii) Phylogenetic profiles* Genes Genomes A B C D 1 1 1 1 1 2 1 1 1 1 3 1 1 0 0

…

… … … … … … … … … …

integration Integrated baits preys PI score A B 0.90 C 0.80 A A D 0.40 C E 0.75 … … …

A

B

iv) Intergenic Distances*

B vs. or B

A

p A

B

p A

B

C E

genome mapping

BDBHs

scoring

Integrated Prot2 GC score B 0.98 0.85 C 0.50 D 0.80 C … …

Prot1 A A A B …

GC network* B

A

D C

confidence cutoff = 0.8

F

filtering H

G

filtering

(c)

PI network*

p A

Operon Rearrangements

integration scoring

confidence cutoff = 0.75

iii) Gene order*

Integration of PI and GC probabilistic networks and function prediction

functional categories

B

A

D C E

B

A F G

D C E H

F G

StepPLR Relaxation Voting Relaxation Labeling orphan function prediction

E.coli genes

(a)

0.9 0.9 - 1.0 1.0 - - - 0.8 0.9 0.9 -

- 0.9 0.9 0.8 - 1.0 1.0 1.0

gene function annotations and predictions*

Figure 8.3  Network‐based protein function prediction in E. coli. We first mapped physical interactions by AP/MS (a) and, in parallel, functional associations based on genomics methods like gene coconservation (b). Next, we predicted the biological roles of previously unannotated “orphan” proteins from the combined probabilistic networks using machine learning algorithms that evaluated individual pairwise protein interaction scores, the local network topology, and correlations among existing functional annotation terms (c). Source: Hu et al. (2009). https://journals.plos.org/plosbiology/article?id=10.1371/ journal.pbio.1000096. Licensed under CC BY 4.0.

Functional Proteomics  211

of which are seemingly linked to human disease (Havugimana et al. 2012). These discoveries have not been without a myriad of challenges, some expected given the scope and technical difficulties in terms of scale‐up, but many others catching us off guard. Foremost among the latter was data analysis. This includes two aspects: the first is to create a robust and accurate curated gold standard for scoring PPIs and predicted protein complexes, respectively. The second was the need to develop methods for integrating massive different types of genomic and proteomic data for inferring interactions reliably to better advance the understanding of functional organizations of cell systems. Another difficult issue relates to multilabel, multiclass functional prediction. We tackled these complicated classification problems by taking the overall network topology and similarity of functional categories into the classification procedure. We also applied labeling approaches to encompass the uncertainty of functional labels (e.g., a protein with a given function but not with another function). The potential trade‐off is that additional error or uncertainty may have occasionally been introduced by assuming functional similarity among more “loosely” connected proteins. Although many approaches have been developed to measure the similarity between functional terms (e.g., GO), it has been known that there is a significant function bias in the current GO database since certain biological processes have been studied in far greater depth experimentally than others. Protein complexes associated with these processes are much more likely to be correctly detected by computational prediction procedures (Barutcuoglu et al. 2006). Some of the advantages of our combined experimental and computational ­strategy have been demonstrated using an integrated E. coli functional association network. Our new algorithms were able to assign functions to a large majority of orphan proteins, many of which seemingly play a role in core biological processes. Independent experimental validation of some of the predicted functions using diverse biological assays, including our group’s recently described E. coli synthetic genetic array mapping technology (Babu et al. 2011), demonstrates the utility and power of high‐throughput methods to map unexpected functional relationships. It is hoped that the scoring framework and data generation and processing pipelines painstakingly established in our laboratory, by trial and error over the past decade, will allow others to systematically probe protein function more efficiently in many different bacterial and eukaryotic species in the years to come. REFERENCES Aebersold, R. and Mann, M. (2003). Mass spectrometry‐based proteomics. Nature 422: 198–207. Alberts, B. (1998). The cell as a collection of protein machines: preparing the next generation of molecular biologists. Cell 92: 291–294. Andersen, J.S., Wilkinson, C.J., Mayor, T. et  al. (2003). Proteomic characterization of the human centrosome by protein correlation profiling. Nature 426: 570–574. Ashburner, M., Ball, C.A., Blake, J.A. et al. (2000). Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nat. Genet. 25: 25–29. Babu, M., Butland, G., Pogoutse, O. et al. (2009a). Sequential peptide affinity purification system for the systematic isolation and identification of protein complexes from Escherichia coli. Methods Mol. Biol. 564: 373–400.

212  Proteomics for Biological Discovery Babu, M., Krogan, N.J., Awrey, D.E. et al. (2009b). Systematic characterization of the protein interaction network and protein complexes in Saccharomyces cerevisiae using tandem affinity purification and mass spectrometry. Methods Mol. Biol. 548: 187–207. Babu, M., Gagarinova, A., Greenblatt, J., and Emili, A. (2011). Array‐based synthetic genetic screens to map bacterial pathways and functional networks in Escherichia coli. Methods Mol. Biol. 765: 125–153. Babu, M., Vlasblom, J., Pu, S. et  al. (2012). Interaction landscape of membrane‐protein ­complexes in Saccharomyces cerevisiae. Nature 489: 585–589. Barutcuoglu, Z., Schapire, R.E., and Troyanskaya, O.G. (2006). Hierarchical multi‐label ­prediction of gene function. Bioinformatics 22: 830–836. Behrends, C., Sowa, M.E., Gygi, S.P., and Harper, J.W. (2010). Network organization of the human autophagy system. Nature 466: 68–76. Butland, G., Peregrin‐Alvarez, J.M., Li, J. et  al. (2005). Interaction network containing ­conserved and essential protein complexes in Escherichia coli. Nature 433: 531–537. Chae, P.S., Rasmussen, S.G., Rana, R.R. et  al. (2010). Maltose‐neopentyl glycol (MNG) ­amphiphiles for solubilization, stabilization and crystallization of membrane proteins. Nat. Methods 7: 1003–1008. Chan, J.N., Vuckovic, D., Sleno, L. et al. (2012). Target identification by chromatographic co‐ elution: monitoring of drug‐protein interactions without immobilization or chemical derivatization. Mol. Cell. Proteomics 11, M111.016642. Chua, H.N., Sung, W.K., and Wong, L. (2006). Exploiting indirect neighbours and topological weight to predict protein function from protein‐protein interactions. Bioinformatics 22: 1623–1630. Collins, S.R., Kemmeren, P., Zhao, X.C. et  al. (2007). Toward a comprehensive atlas of the physical interactome of Saccharomyces cerevisiae. Mol. Cell. Proteomics 6: 439–450. Deng, M., Zhang, K., Mehta, S. et  al. (2003). Prediction of protein function using protein‐­ protein interaction data. J. Comput. Biol. 10: 947–960. Dignam, J.D., Lebovitz, R.M., and Roeder, R.G. (1983). Accurate transcription initiation by RNA polymerase II in a soluble extract from isolated mammalian nuclei. Nucleic Acids Res. 11: 1475–1489. Enright, A.J., van Dongen, S., and Ouzounis, C.A. (2002). An efficient algorithm for large‐ scale detection of protein families. Nucleic Acids Res. 30: 1575–1584. Ewing, R.M., Chu, P., Elisma, F. et al. (2007). Large‐scale mapping of human protein‐protein interactions by mass spectrometry. Mol. Syst. Biol. 3: 89. Fenn, J.B., Mann, M., Meng, C.K. et al. (1989). Electrospray ionization for mass spectrometry of large biomolecules. Science 246: 64–71. Gavin, A.C., Bosche, M., Krause, R. et al. (2002). Functional organization of the yeast proteome by systematic analysis of protein complexes. Nature 415: 141–147. Gavin, A.C., Aloy, P., Grandi, P. et al. (2006). Proteome survey reveals modularity of the yeast cell machinery. Nature 440: 631–636. Gingras, A.C., Gstaiger, M., Raught, B., and Aebersold, R. (2007). Analysis of protein complexes using mass spectrometry. Nat. Rev. Mol. Cell Biol. 8: 645–654. Goll, J. and Uetz, P. (2006). The elusive yeast interactome. Genome Biol. 7: 223. Guruharsha, K.G., Rual, J.F., Zhai, B. et al. (2011). A protein complex network of Drosophila melanogaster. Cell 147: 690–703. Hart, G.T., Lee, I., and Marcotte, E.R. (2007). A high‐accuracy consensus map of yeast protein complexes reveals modular nature of gene essentiality. BMC Bioinf. 8: 236. Hartwell, L.H., Hopfield, J.J., Leibler, S., and Murray, A.W. (1999). From molecular to modular cell biology. Nature 402: C47–C52.

Functional Proteomics  213

Havugimana, P. C. (2012). Global proteomic detection of native, stable, soluble human protein complexes. PhD thesis, University of Toronto. Havugimana, P.C., Hart, G.T., Nepusz, T. et  al. (2012). A census of human soluble protein ­complexes. Cell 150: 1068–1081. Hu, P., Bader, G., Wigle, D.A., and Emili, A. (2007). Computational prediction of cancer‐gene function. Nat. Rev. Cancer 7: 23–34. Hu, P., Janga, S.C., Babu, M. et al. (2009). Global functional atlas of Escherichia coli encompassing previously uncharacterized proteins. PLoS Biol. 7: e96. Hu, P., Jiang, H., and Emili, A. (2010). Predicting protein functions by relaxation labelling protein interaction network. BMC Bioinf. 11 (Suppl 1): S64. Hutchins, J.R., Toyoda, Y., Hegemann, B. et al. (2010). Systematic analysis of human protein complexes identifies chromosome segregation proteins. Science 328: 593–599. Jansen, R. and Gerstein, M. (2004). Analyzing protein function on a genomic scale: the ­importance of gold‐standard positives and negatives for network prediction. Curr. Opin. Microbiol. 7: 535–545. Karas, M. and Hillenkamp, F. (1988). Laser desorption ionization of proteins with molecular masses exceeding 10,000 daltons. Anal. Chem. 60: 2299–2301. Kocher, T. and Superti‐Furga, G. (2007). Mass spectrometry‐based functional proteomics: from molecular machines to protein networks. Nat. Methods 4: 807–815. Kristensen, A.R., Gsponer, J., and Foster, L.J. (2012). A high‐throughput approach for measuring temporal changes in the interactome. Nat. Methods 9: 907–909. Krogan, N.J., Cagney, G., Yu, H. et al. (2006). Global landscape of protein complexes in the yeast Saccharomyces cerevisiae. Nature 440: 637–643. Kuhner, S., van Noort, V., Betts, M.J. et al. (2009). Proteome organization in a genome‐reduced bacterium. Science 326: 1235–1240. Lee, I., Blom, U.M., Wang, P.I. et al. (2011). Prioritizing candidate disease genes by network‐ based boosting of genome‐wide association data. Genome Res. 21: 1109–1121. Letovsky, S. and Kasif, S. (2003). Predicting protein function from protein/protein interaction data: a probabilistic approach. Bioinformatics 19 (Suppl 1): i197–i204. Li, Y. (2011). The tandem affinity purification technology: an overview. Biotechnol. Lett. 33: 1487–1499. Li, X., Wu, M., Kwoh, C.K., and Ng, S.K. (2010). Computational approaches for detecting protein complexes from protein interaction networks: a survey. BMC Genomics 11 ­ (Suppl 1): S3. Liang, S., Shen, G., Xu, X. et al. (2009). Affinity purification combined with mass spectrometry‐based proteomic strategy to study mammalian protein complex and protein‐protein interactions. Curr. Proteomics 6 (1): 25–31. Lord, P.W., Stevens, R.D., Brass, A., and Goble, C.A. (2003). Investigating semantic similarity measures across the gene ontology: the relationship between sequence and annotation. Bioinformatics 19: 1275–1283. Mak, A.B., Ni, Z., Hewel, J.A. et  al. (2010). A lentiviral functional proteomics approach ­identifies chromatin remodeling complexes important for the induction of pluripotency. Mol. Cell. Proteomics 9: 811–823. Massjouni, N., Rivera, C.G., and Murali, T.M. (2006). VIRGO: computational prediction of gene functions. Nucleic Acids Res. 34: W340–W344. Mewes, H.W., Amid, C., Arnold, R. et al. (2004). MIPS: analysis and annotation of proteins from whole genomes. Nucleic Acids Res. 32: D41–D44. Murali, T.M., Wu, C.J., and Kasif, S. (2006). The art of gene function prediction. Nat. Biotechnol. 24: 1474–1475. author reply 1475–1476.

214  Proteomics for Biological Discovery Musso, G.A., Zhang, Z., and Emili, A. (2007). Experimental and computational procedures for the assessment of protein complexes on a genome‐wide scale. Chem. Rev. 107: 3585–3600. Nabieva, E., Jim, K., Agarwal, A. et al. (2005). Whole‐proteome prediction of protein function via graph‐theoretic analysis of interaction maps. Bioinformatics 21 (Suppl 1): i302–i310. Nepusz, T., Yu, H., and Paccanaro, A. (2012). Detecting overlapping protein complexes in protein‐protein interaction networks. Nat. Methods 9: 471–472. Pu, S., Vlasblom, J., Emili, A. et al. (2007). Identifying functional modules in the physical interactome of Saccharomyces cerevisiae. Proteomics 7: 944–960. Rasmussen, S.G., Choi, H.J., Fung, J.J. et al. (2011). Structure of a nanobody‐stabilized active state of the beta(2) adrenoceptor. Nature 469: 175–180. Resnik, P. (1999). Semantic similarity in a taxonomy: an information‐based measure and its application to problems of ambiguity in natural language. J. Artif. Intell. Res. 11: 95–130. Rigaut, G., Shevchenko, A., Rutz, B. et al. (1999). A generic protein purification method for protein complex characterization and proteome exploration. Nat. Biotechnol. 17: 1030–1032. Robinson, C.V., Sali, A., and Baumeister, W. (2007). The molecular sociology of the cell. Nature 450: 973–982. Ruepp, A., Waegele, B., Lechner, M. et al. (2010). CORUM: the comprehensive resource of mammalian protein complexes – 2009. Nucleic Acids Res. 38: D497–D501. Sardiu, M.E., Cai, Y., Jin, J. et al. (2008). Probabilistic assembly of human protein interaction networks from label‐free quantitative proteomics. Proc. Natl. Acad. Sci. USA. 105: 1454–1459. Schwikowski, B., Uetz, P., and Fields, S. (2000). A network of protein‐protein interactions in yeast. Nat. Biotechnol. 18: 1257–1261. Sowa, M.E., Bennett, E.J., Gygi, S.P., and Harper, J.W. (2009). Defining the human deubiquitinating enzyme interaction landscape. Cell 138: 389–403. Steen, H. and Mann, M. (2004). The ABC’s (and XYZ’s) of peptide sequencing. Nat. Rev. Mol. Cell Biol. 5: 699–711. Taniguchi, Y., Choi, P.J., Li, G.W. et al. (2010). Quantifying E. coli proteome and transcriptome with single‐molecule sensitivity in single cells. Science 329: 533–538. Vazquez, A., Flammini, A., Maritan, A., and Vespignani, A. (2003). Global protein function prediction from protein‐protein interaction networks. Nat. Biotechnol. 21: 697–700. Xu, X., Song, Y., Li, Y. et  al. (2010). The tandem affinity purification method: an efficient ­system for protein complex purification and protein interaction identification. Protein Expr. Purif. 72: 149–156. Zeghouf, M., Li, J., Butland, G. et al. (2004). Sequential peptide affinity (SPA) system for the identification of mammalian and bacterial protein complexes. J. Proteome Res. 3: 463–468. Zhang, B. and Horvath, S. (2005). A general framework for weighted gene co‐expression ­network analysis. Stat. Appl. Genet. Mol. Biol. 4: 17. Zhao, X.M., Wang, Y., Chen, L., and Aihara, K. (2008). Gene function prediction using labeled and unlabeled data. BMC Bioinf. 9: 57.

9 High‐Resolution Interrogation of Biological Systems via Mass Cytometry Heather M. G. Brown, Michelle M. Kuhns, and Edgar A. Arriaga Department of Chemistry, University of Minnesota, Minneapolis, MN, USA

9.1 INTRODUCTION The observation and analysis of single cells in the fields of genomics, metabolomics, and transcriptomics have revealed that multiple cell types or functionally distinct subpopulations often coexist in biological systems. Although bulk genomic, proteomics, and metabolomics analyses of these systems are informative, it has become increasingly apparent that meaningful understanding of cellular processes  –  from normal development to disease progression  –  is more informative when such ­analyses achieve single‐cell resolution. Accordingly, single‐cell proteomics has also evolved over the last few decades to include many different platforms for analysis, including, but not limited to, microfluidics and cytometry. Microfluidic analysis of single‐cell proteomes has been realized through cell arrays (Carlo and Lee 2006; Carlo et al. 2006; Faley et al. 2009; Faley et al. 2011) and single‐cell barcode chips (Ma et al. 2011; 2013; Shin et al. 2011; Shi et al. 2012; Wang et al. 2012; Ahmad et al. 2011; Lu et al. 2013). Multiparameter fluorescence flow cytometry is perhaps the most well‐known technique for multiple target detection and analysis in single cells. Developments of this technology have allowed

Proteomics for Biological Discovery, Second Edition. Edited by Timothy D. Veenstra and John R. Yates III. © 2019 John Wiley & Sons, Inc. Published 2019 by John Wiley & Sons, Inc.

215

216  PROTEOMICS FOR BIOLOGICAL DISCOVERY

for simultaneous analysis of up to 20 parameters, although more common analyses include far fewer parameters to eliminate the need for complicated ­deconvolution of the spectral overlap of fluorescence signals, which is inherent due to the similarity of ­excitation and emission spectra of fluorescent labels used in such studies. A technique related to fluorescence flow cytometry is mass cytometry. This technique uses mass spectrometry to detect and analyze elemental ions that report on the abundance of cellular targets in single cells. The mass resolution of mass spectrometry makes this technique adequate to detect a high number of pure isotopic elements, surpassing the number of parameters detected by flow cytometry. Whereas in flow cytometry, fluorescently tagged antibodies specific to cellular targets are used to label cells, in mass cytometry the antibodies are tagged with isotopically pure lanthanide metals. A unique feature of mass cytometry is the use of an inductively coupled plasma (ICP) that fully ionizes all components in a sample, which then allows for detection of a wide range of lanthanide ions labeling molecular targets in single cells. Before the introduction of mass cytometry, inductively coupled plasma‐mass spectrometry (ICP‐MS) had been used to detect metal tagged antibodies specific to cellular markers in cell lysates, but these findings could not be related back to specific cell types in a heterogeneous cell population (Razumienko et  al. 2008). Other relevant prior work includes the ICP‐MS analysis of single cells from suspensions that had been labeled with lanthanides and analyzed using conventional instrumentation (Ornatsky et al. 2006; Tanner et al. 2007, 2008; Quinn et al. 2002; Baranov et al. 2002; Ornatsky et al. 2008). The work headed by Scott Tanner and colleagues led to the first report on the feasibility of elemental analysis of individual, antibody‐labeled cells, establishing mass cytometry as a new technique (Bandura et al. 2009). The report of successfully using an ICP‐MS for individual cell analysis was revolutionary in that it provided high resolution and unprecedented multiplexing capabilities necessary for thorough analysis of complex biological systems. DVS Sciences first commercialized the technology in 2009, offering the CyTOF (Cytometry by Time Of Flight) which featured a default m/z range of 103–193, maximum throughput of ~1000 events/second and a dynamic range covering four orders of magnitude. Subsequent models named CyTOF 2 (second‐generation instrumentation) and Helios (third generation) further improved these and other parameters to increase the number of m/z channels that can be monitored as well as widen the dynamic range. Using state‐of‐the‐art instrumentation and methodology, the technique has seen improvements in the number of parameters that can be measured per cell (upwards of 50), peak sample introduction rates (2000 cells per second), and improved dynamic ranges (4.5 orders of magnitude). These figures of merit are either on a par with or far surpass the abilities of currently available flow cytometers. 9.2 INSTRUMENTATION Generally speaking, a mass cytometer has three modules. The first module includes an interface to introduce cells into the instrument and a set‐up to desolvate, vaporize, and atomize/ionize them (Figure 9.1, Box 1). The second module is a set of ion optics

High-Resolution Interrogation of Biological Systems via Mass Cytometry   217

3 DETECTOR

COLLECTOR f = 76.8 kHz

ACCELERATOR

LINER

SLIT

REFLECTOR

2

SHIELD DC DOUBLET

A2ʹ

A2

A1ʹ

A1

LENS B

3-CONE INTERFACE

RF QUAD

EINZEL LENS

ICP TORCH

HEATED SPRAY CHAMBER

1

SYRINGE PUMP

DEFLECTOR

MAKE-UP GAS

NEBULIZER GAS

Figure 9.1  General overview of CyTOF hardware. After samples are introduced to the instrument, they are aerosolized via the nebulizer, desolvated in the spray chamber, and sequentially vaporized, atomized, and ionized as they exit the ICP torch (Box 1). The resulting ion cloud is filtered, focused, and guided toward the detector via the RF quadrupole shown in Box 2. Separation and detection of relevant metal ions take place in the time‐of‐flight (TOF) detector in Box 3. Source: Reprinted from Bandura et al. (2009) with permission.

to select all ions of interest (Figure 9.1, Box 2). The third module is a time‐of‐flight (TOF) mass spectrometer, used to identify and quantify the abundance of each ion of interest (Figure 9.1, Box 3). These modules and their function are described below. 9.2.1  Sample Introduction and Ionization (Figure 9.1, Box 1) The introduction of cell samples into the mass cytometer has evolved with each new version of the CyTOF instrumentation. Introduction of cell samples into the first‐ or second‐generation mass cytometer is accomplished using a syringe pump and one‐ or two‐sample loop system. Using this system, an aliquot of the cell suspension is pushed into a sample loop (500 μL volume). After one of the loops is loaded, the

218  PROTEOMICS FOR BIOLOGICAL DISCOVERY

flow injection valve rotates so that flow from a syringe pump displaces the loop contents to deliver them to the nebulizer via a sample introduction capillary. For unsupervised sample injections, there are two autosamplers compatible with the ­second generation of the instrumentation: the CyTOF Autosampler 5 uses a 96‐well plate format for cell suspension aspiration into a dual‐loop system while the Super Sampler (www.victorianairships.com) is used for unsupervised, large sample volume data collection with increased cell sampling efficiency. The newest generation of  mass  cytometer, Helios, has a pneumatic sample introduction system that uses argon gas to force cell suspensions into the sample uptake probe and onward to the nebulizer. As cells from the cell suspension are introduced into the nebulizer via the sample introduction capillary, a fine spray of droplets containing individual cells is produced by the pressure exerted on the suspension from the argon gas exiting the tip of the concentric nebulizer (Figure 9.1, Box 1). This aerosol is directly injected into a heated spray chamber (~200 °C) supplied with argon makeup gas where the cell‐containing droplets are partially desolvated (Bandura et  al. 2009). The partially desolvated droplets then travel through an ICP torch consisting of a high‐temperature plasma (5500–7000 K) where they are vaporized, atomized, and ionized (Bandura et  al. 2009; Tanner et al. 2013). For each cell being ionized, the process results in a cloud of ions with a diameter of ~ 2 mm. The expansion of the ion cloud for each cell is diffusion limited, meaning that these clouds will be roughly the same size whatever the size of the cell introduced (Tanner et al. 2013). Thus, the analysis of single cells in samples with cells of multiple sizes is not biased after ionization. Furthermore, the plasma at the ICP torch is theoretically at thermodynamic equilibrium (Griem 1963; Fujimoto and McWhirter 1990; Mostaghimi and Boulos 1990) which ensures almost 100% ionization efficiency of elements with ionization potentials below 9 eV, despite differences in cellular matrices. This allows for the technique to be truly quantitative as the number of metal ions detected is directly related to the number of ions introduced (Tanner et al. 2013). 9.2.2  Ion Optics (Figure 9.1, Box 2) Ions produced at the ICP torch are sampled through a three‐cone interface, which serves as a low‐pressure, controlled entrance into the low‐vacuum ion optics chamber. Sampler, skimmer, and reducing cones reduce the pressure while cooling and focusing the ion beam. A quadrupole ion deflector then filters out neutral atoms that will interfere with downstream analysis. Positively charged ions are directed perpendicularly out of the ion deflector toward downstream ion optics for eventual mass analysis, while neutral atoms follow an uninterrupted path toward a turbomolecular pump. Another quadrupole (RF QUAD in Figure 9.1, Box 2) serves to filter out low‐mass ions (H+, C+, O+, N+, OH+, CO+, O2+, Ar+, ArH+, ArO+) that predominate in abundance in the ion cloud after ionization. Thus, ions leave the quadrupole in packets with a specific mass range, each packet corresponding to an ion cloud, which in turn corresponds to an individual cell. The filtered ion packets leaving the RF QUAD then pass through a DC quadrupole doublet (DC DOUBLET in Figure 9.1, Box 2) which flattens the ion stream to allow for injection into the TOF mass spectrometer.

High-Resolution Interrogation of Biological Systems via Mass Cytometry   219

9.2.3  Time‐of‐Flight Mass Spectrometer (Figure 9.1, Box 3) Ion packets are introduced through a vacuum interface (SLIT in Figure 9.1, Box 3) to an accelerator where ions experience sequential voltage pulses at a frequency of 76.8 KHz. That is, a portion of ions from each packet is sampled (pushed) every 13 μs, accelerated by the pulsed electric field, which impinges a constant kinetic energy regardless of the ion mass to charge ratio (m/z). In total, the ion packet from each cell is sampled 20–35 times. The kinetic energy acquired by all ions through acceleration translates into a unique TOF (t) corresponding to each m/z from the accelerator plate, via the reflector, to the detector. Using two empirically derived constants, t0 and A, the relationship between (t) and (m/z) is: t



t0

A

m (9.1) z

As deduced from Eq. (9.1), the ions with smallest m/z value reach the detector first. Ions will continue to arrive according to increasing m/z order. Typically, isotopes with masses differing by 1 amu have ions arriving at the detector 20–25 ns apart. Because the detector response is faster than 20 ns, the detector has a resolution better than 1 amu, which is sufficient to detect the various isotopes used as labels in mass cytometry. Each new generation of mass cytometer has had an increased range of masses that can be analytically analyzed, starting with a range of 103–193 amu for CyTOF (generation 1), 89–209 amu for CyTOF 2 (generation 2), and 75–209 for Helios (generation 3). 9.2.4  Integration of Mass Spectra The ion detector of a mass cytometer is a dynode electron multiplier. Depending on the number of ions with a given m/z reaching the detector during the 13 μs duration of a packet of ions, the detector operates in a pulse‐counting mode (for low number of ions) or an integration mode (for high number of ions). When the particle density approaching the detector increases, particles will arrive simultaneously, resulting in an underestimate of counts. Thus, integration of the analog signal is a more useful quantification method in this case. The relevant concentration range at which CyTOF operates requires that dual data be collected, meaning both counting and integration modes are applied to each m/z value of interest (channel). A user‐defined threshold defines when integrated intensity and pulse counting are selected for processing the signal. When the integrated intensity is used, outputs are converted to ion counts using the equation:

Count

Dual count coefficient

Intensity (9.2)

where the dual count coefficient is a constant determined experimentally for each mass channel. To obtain meaningful data, the instrument user must set criteria to identify signals (events) corresponding to single cells. These criteria are minimal signal intensity and duration of the event, typically 20–50 spectra. Signals for each m/z value (channel)

220  PROTEOMICS FOR BIOLOGICAL DISCOVERY Mass channels (m/z) 131

140

142

151

153

165

175

176

191

193

Push number

0

100

200 Figure 9.2  Consecutive mass spectra showing integrated signal for a single detected event. Each “push” is recorded as a separate mass spectrum, which can be stacked sequentially to visualize events in real time across selected mass channels. (Unpublished data, H.M.G. Brown.)

are integrated across all mass spectra collected for individual events (20–50 spectra/ event), and written into an integrated mass data file (*.imd), which is subsequently compiled into text (*.txt) and/or flow cytometry standard (*.fcs) files which include metadata such as instrument conditions and channel annotations (Tanner et  al. 2013). Events with an abnormally long, bimodal signal correspond to overlapping events, and must be removed post acquisition (Tanner et al. 2013). The frequency of analyzed events can be visualized during data acquisition in the form of a rain plot consisting of stacked mass spectra (Figure 9.2). 9.3  SAMPLE PREPARATION Preparation of samples for mass cytometric analysis is comparable to preparation of cells for flow cytometric analysis. Both rely on the use of antibodies to label specific cellular targets (epitopes) on and within cells with a molecular reporter. In flow cytometry, antibodies are labeled with fluorophores that fluoresce at a characteristic wavelength when excited by incident light. In mass cytometry, antibodies are labeled with a polymer loaded with isotopically pure lanthanide ions which are detected at a specific m/z ratio after ionization. In addition to antibodies conjugated to polymers holding lanthanide ions, other labeling reagents are also used in mass cytometry. This section describes common labeling reagents, cell preparation procedures, and an overall summary of how a CyTOF 2 instrument is operated when cell samples are analyzed. 9.3.1 Reagents Metal Chelating Polymers  For the purpose of mass cytometry, metal chelating polymers (MCPs) are loaded with a pure lanthanide isotope and conjugated to an antibody. The same MCP can be used for labeling all antibodies, but each unique antibody clone is labeled with a unique lanthanide isotope. Great similarity in chemical properties of the lanthanide series of metals makes it possible to use a wide range of isotopes with the same MCP. Furthermore, lanthanides have low natural abundance

High-Resolution Interrogation of Biological Systems via Mass Cytometry   221

O S N O O S N

O

N

O

N

HN OC

O

O

m

CO

N

O

O O

n

OC

Ln

N N O O

O

O

Figure 9.3  Metal chelating polymers are made up of three main components: a polymer backbone, metal chelating ligands, and a linking moiety that connects the polymer backbone to an antibody. Metal chelating ligands, usually DTPA or DOTA, are covalently attached to the polymer backbone at periodic intervals. The polymer is connected to the antibody through a maleimide linkage that reacts with free sulfhydryl moieties in a partially reduced antibody. Together with the number of metal ions associated with each polymer, the number of polymers attached to each antibody determines the overall number of metal ions associated with each antibody.

and are commercially availability as purified isotopes (Xudong Lou and Herrera 2007; Majonis et al. 2010; Illy et al. 2012). Loading capacity (lanthanide ion/polymer chain) of each MCP is dictated by the molecular structure of MCPs. In general, the molecular structure of MCPs includes three main covalently linked scaffolds with separate functions (Figure 9.3), including a polymer backbone, metal chelating ligand, and a linker. The most common metal chelating ligands are multidentate polyaminocarboxylates, such as diethylenetriaminepentaacetic acid (DTPA), and 1,4,7,10‐tetraazacyclododecane‐1,4,7,10‐tetraacetic acid (DOTA), because they coordinate lanthanide ions (3+) with low dissociation constants (Parker et al. 2002). The most commonly used linker relies on the reaction of maleimide with sulfhydryl groups produced by reducing disulfide bonds within the Fc portion of an antibody. Although chemistries that would covalently bind the polymer to lysine residues in the antibody are also feasible, there exists a general concern that random lysine modifications could result in a loss of antibody specificity. Because sensitivity of mass cytometry and other ICP‐MS techniques is expected to increase linearly with lanthanide content, developments in MCPs have attempted to chelate more lanthanide ions per polymer. The first report by Lou and Herrera (2007) described 33 DOTA ligands per polymer chain, representing the maximum number of lanthanide ions loaded onto a polymer. Later MCPs described by Majonis et al. were able to attach 68 ± 7 lanthanide‐chelating ligands per polymer. They also determined that on average, a conjugated antibody displayed 2.4 ± 0.3 polymers, giving a total of 161 ± 4 lanthanide ions per antibody (Majonis et al. 2010). To increase the utility of lanthanide‐chelating polymer conjugated antibodies, dual‐labeled ­antibodies have been reported (Majonis et  al. 2013; Baumgart et  al. 2017; Buckle et al. 2017), which contain both fluorophore and lanthanides as reporters to facilitate cross‐platform validation of antibody localization or cell subpopulation enrichment prior to mass cytometry analysis.

222  PROTEOMICS FOR BIOLOGICAL DISCOVERY

N H

N

H

Rh

N N

H

N

N

N

N

Ir

N N H

L

Figure 9.4  DNA metallointercalators usually contain three ligands bound to a transition element. Insertion of a planar aromatic ligand between base pairs and the associated electrostatic interactions give a relatively low dissociation constant, which allows for a stable DNA signal to be measured in individual cells via mass cytometry. Source: Reprinted from Schäfer and Sheldrick (2007) and Pyle et al. (1989) with permission.

Metallointercalators  The ability to identify signals originating from cells is critical for accurate analysis of single cell suspensions. Ornatsky et al. (2008) describe the use of the well‐characterized polypyridyl metal complexes (metallointercalators) for this purpose. These reagents can help identify single cells among cellular debris, identify cell cycle phases, and normalize signal intensities across cell populations. Similar to the MCPs mentioned above, metallointercalators usually contain a transition metal, such as Rh, Ru, or Ir, whose abundance within individual cells is an indirect measurement of DNA content. In original reports, three ligands of the metallointercalator bind to the transition metal (Schäfer and Sheldrick 2007; Pyle et al. 1989) (Figure 9.4). The strong interactions between DNA and the metallointercalator are dominated by electrostatic forces and insertion of planar aromatic ligand between nucleotide base pairs (Long and Barton 1990). This interaction has a low dissociation constant (KD  ZrO2. In a demonstration of this complementarity of IMAC and MOAC, phosphopeptides from mouse macrophages were enriched using IMAC and ZrO2 successively (Kweon and Andrews 2013). The performance of phosphopeptide enrichment methods was evaluated by calculating the ratio of the number of identified phosphopeptides versus the total number of all identified peptides. The results obtained from two replicate LC‐MS2 analyses of each fraction were combined. From the two analyses of the IMAC eluent, 378 unique phosphorylated peptides and only

Phosphorylation  277

17 nonphosphorylated peptides were identified, resulting in an overall enrichment efficiency of 96%. Of the identified phosphopeptides identified in the IMAC fraction, 77% were singly phosphorylated, 21% were doubly phosphorylated, and 2% were at least triply phosphorylated. In the IMAC flow‐through fraction that was subsequently enriched using a ZrO2 MOAC, 187 unique phosphopeptides and 525 unique nonphosphorylated peptides were identified, resulting in an enrichment efficiency of 26%. Of the phosphopeptides identified from the ZrO2 resin, 91% were singly phosphorylated. A 12% overlap was observed between phosphopeptides identified in the IMAC and consecutive ZrO2 enrichment. Overall, 504 unique phosphopeptides containing 683 unique phosphorylation sites in 294 proteins were identified in the mouse macrophage samples after four LC‐MS2 analyses. 11.4  DATABASE SEARCHING Acquiring the MS2 or MS3 spectra is only one of the steps necessary to identify the phosphopeptides. The spectra must be searched against a database containing the protein sequences from the species of interest. While a number of different search algorithms are currently available for identifying phosphopeptides, the most commonly used are SEQUEST and MASCOT (Yates et  al. 1996; Perkins et  al. 1999). While the algorithms use different comparison and scoring metrics, they both conduct a comparison of the experimentally acquired fragmentation spectrum to that ­produced from a theoretical spectrum. Since the analysis will contain some fraction of nonphosphorylated peptide, a parameter that adds a “dynamic” modification equal to the mass of a phosphorylation group on specific residues (i.e., Ser, Thr, and Tyr) is incorporated into the search algorithm. This parameter instructs the search algorithm to consider each of these residues in both the modified and unmodified state when performing the database search. This parameter increases the size of the database that needs to be searched, increasing the requirement for high‐quality MS2 spectra due to the greater number of potential matches to any given precursor ion. To gauge the confidence in phosphopeptide identifications, it is important to ­understand the factors that go into the scoring algorithm for a given database and the potential sources of error. In comparison of the theoretical and experimental spectra, the mass of the precursor (or parent) ion is directly correlated to the length of the peptide and the number of potential fragment ions. Since the overall score for a given peptide match increases with the increase in fragment ion matches, the minimum acceptable score should be considered in the context of the precursor ion mass. Therefore, phosphopeptides with higher precursor ion masses should have a greater minimum score to be considered a confident identification than smaller phosphopeptides. Many laboratories have developed false discovery rates (FDRs) calculations to enable the determination of the percentage of the total identifications that have been incorrectly matched to a sequence in a database and are therefore false positives (Stone et  al. 2011; Yadav et  al. 2013). A common FDR calculation involves searching the experimental data against a decoy database in which no correct matches would be expected to occur. Examples of decoy databases include those in which the protein sequences are scrambled or reversed. Any matches that occur when searching a decoy database are assumed to be false positives and when compared to

278  Proteomics for Biological Discovery

the number of true identifications obtained using the correct database, are used to calculate the FDR. It is generally assumed that the FDR calculated using this method is only an estimate and unfortunately does not identify those spectra that may be incorrectly assigned. Since it is accepted that potential errors exist for the identification of any phosphopeptide, before considerable effort is made in executing further validation studies, it is worth manually inspecting the MS2 or MS3 spectrum used for phosphopeptide identification. While conducting manual validation of hundreds of spectra is impractical, it is worth the time to inspect those spectra that identify key phosphorylation events that lead to in‐depth biological insight of the system being studied. The goal in manual validation is to assign fragment ions that arise from the full peptide sequence and identify as many of the fragment ions that are present in the MS2 (or MS3) spectra. Identifying all the fragment ions within a MS2 spectrum is unlikely since peptides can fragment in unpredictable ways. The ultimate validation step to ensure that the MS2 spectrum is correctly identified requires comparing this ­spectrum to that of the chemically synthesized phosphopeptide. If the MS2 spectra match, the phosphopeptide identification is accurate. It cannot be stressed enough how important it is to ensure that the correct ­phosphopeptide identification is obtained prior to further biological studies being conducted. The cost in labor and materials in pursuing further biological studies based on an incorrect identification can result in tens of thousands of dollars in wasted resources. 11.5  DISCOVERING BIOLOGICAL INSIGHT IN PHOSPHORYLATION STUDIES The previous pages have discussed the mechanics of identifying phosphorylated proteins in complex mixtures; the following pages are intended to demonstrate how these data are used to glean meaningful biological insight. To do this, I am going to focus on a couple of examples in which global phosphorylation studies were conducted in order to learn something important about a critical issue that can then lead to testable hypotheses. 11.5.1  Glioblastoma Multiforme Glioblastoma multiforme (GBM) is the most aggressive form of adult human brain tumor (Aftab et al. 2015). The median survival rate, the time at which the number of patients that do better or worse is equal, varies depending on the tumor type and treatment. Median survival for adults with an anaplastic astrocytoma who are provided standard treatment is between two and three years. The median survival of adults with aggressive forms of GBM, who are treated with temozolomide and ­radiation therapy, is about 14.6 months, with a two‐year survival rate of only 30% (Lee et al. 2014; Rapp et al. 2015). A partial cause of the low median survival rate is the dearth of biomarkers available for reliable early detection of GBM. Unfortunately, this deficiency is not unique to GBM and is too common amongst cancers. What is well known about GBM tumors is that about 50% of them overexpress a mutant form of the epidermal growth factor receptor (EGFR) (Holland et  al. 1998).

Phosphorylation  279

This mutant form, referred to as EGFRvIII, lacks exons 2–7, rendering it devoid of amino acids 6–273 of the extracellular domain normally found in the EGFR. Without this domain, EGFRvIII is unable to bind ligand, causing the receptor to exhibit weak but constitutive signaling activity. It is this residual signaling activity that results in enhanced transformation, reduced apoptosis, and resistance to therapy (Zhan and O’Rourke 2004). EGFRvIII is usually coexpressed with wild‐type (wt) EGFR. EGFRvIII activates Met, and Met contributes to EGFRvIII‐mediated oncogenicity and resistance to treatment. The addition of epidermal growth factor (EGF) has been shown to result in a rapid loss of EGFRvIII‐driven Met phosphorylation in glioma cells (Li et al. 2015). The protein Met is well established as being associated with EGFRvIII in a physical complex. Addition of EGF results in a dissociation of the EGFRvIII–Met complex with a concomitant loss of Met phosphorylation. Consistent with the abrogation of Met activation, addition of EGF results in the inhibition of EGFRvIII‐ mediated resistance to chemotherapy. The results showed that the presence of ligand in the milieu of EGFRvIII‐expressing GBM cells is likely to influence the EGFRvIII– Met interaction and resistance to treatment, and highlights a novel antagonistic interaction between wtEGFR and EGFRvIII in glioma cells. Compounding EGFRvIII’s negative activities is the receptor’s ability to evade downregulation (Grandal et al. 2007). In noncancerous cells, ligand binding results in the activation of receptor tyrosine kinases (RTKs), which is followed by internalization, ubiquitination, and degradation of the protein in lysosomes. This tightly coupled process attenuates the receptor signaling, ensuring that the signal is maintained within the time‐frame necessary for maintaining normal cellular function. While the mechanism by which EGFRvIII evades downregulation has been intensely studied, the exact mechanism has not yet been established. One study showed that inefficient internalization complemented by efficient recycling to the plasma membrane contributes to EGFRvIII’s long half‐life (Grandal et  al. 2007). Another study has suggested that hypophosphorylation of tyrosine‐1045 (Y1045) of EGFRvIII prevents the ubiquitin ligase Cbl from interacting with the receptor, inhibiting its degradation (Han et  al. 2006). Another study, however, showed that Cbl interacts with EGFRvIII in a tyrosine phosphorylation‐dependent manner (Davies et al. 2006). A study in Forest White’s laboratory at MIT sought to determine the relationship between EGFRvIII expression level in glioblastoma cell lines and the downstream pTyr‐mediated cellular signal networks that were activated or deactivated (Huang et al. 2007). The aim was to find downstream signaling events that may serve as candidates for therapeutic targeting to inhibit EGFRvIII signaling and thereby reduce tumor cell growth and survival. This group used fluorescence activated cell sorting (FACS) on isolated glioblastoma cell lines that expressed differing numbers of EGFRvIII on their surface. The glioblastoma cell lines that were isolated (and their EGFRvIII expression levels) were U87‐M (1.5 million copies of EGFRvIII per cell), U87‐H (2.0 million copies per cell), and U87‐M (3 million copies per cell). As a control, the study used U87MG cells, which express an inactive form of EGFRvIII. A schematic of the experimental strategy used is shown in Figure 11.5. After extraction of the proteins from the cells, they were digested into peptides using proteomic scientists’ favorite protease, trypsin. A panspecific pTyr antibody followed by IMAC was used to create a highly enriched sample of peptides containing

280  Proteomics for Biological Discovery U87-DK

U87-M

U87-H

PO4

U87-SH

Cell Lysis

PO4 PO4

PO4

Digestion

PO4

PO4

Anti-pTyr IP

PO4

IMAC

114

115

116

iTRAQ Labeling

Identification and Quantitation

115

117

114

116

m/z

117

Figure 11.5  Experimental strategy used for the quantitative analysis of phosphotyrosine (pTyr) ­peptides and proteins for the discovery of the mechanism of EGFRvIII in glioblastoma cells (U87MG). Four cell lines expressing different levels of active EGFRvIII (DK; dead kinase: M; medium: H; high: SH; super high) were generated. Phosphotyrosine peptides were extracted from these cell lines using a combination of immunoprecipitation and anti‐pTyr antibody and immobilized metal ion chromatography (IMAC). Stable isotope labeling using iTRAQ was used to identify and quantitate the extracted phosphopeptides. Source: Adapted from Huang et al. (2007).

phosphorylated tyrosine residues. To quantitate the amount of phosphorylation at each site in the peptides extracted from the four different cell lines, the peptides were labeled using the four‐plex isobaric tags for relative and absolute quantification (iTRAQ) reagents just prior to LC‐MS2 analysis. The final result was the quantitation of 99 unique phosphorylation sites within 69  proteins across the four cell lines. Not surprisingly, the protein in which the

Phosphorylation  281

g­ reatest number of phosphorylation sites were quantitated was EGFRvIII itself. Eight phosphorylation sites, including Y1086, were quantitated in EGFRvIII. The increase in phosphorylation at Y1086 and Y845 was found to be proportional to the number of EGFRvIII copies within each cell type. The amount of phosphorylation at Y974, Y1068, Y1114, Y1148, and Y1173 was found to be twice as much in the U87‐H cells as in the U87‐M cells. The amount of phosphorylation at these five sites did not differ between the U87‐H and U87‐SH cells, suggesting that they reach a saturation point once the expression of EGFRvIII reaches two million copies per cell. The hypothesis is that at this expression threshold, the receptor becomes ­activated and a dramatic increase in autophosphorylation occurs. Based on the amount of information known about EGFR signaling that had been gathered by previous studies, the investigators mapped the various phosphorylation sites identified in their study to the known canonical pathways associated with EGFR signaling. Quite surprisingly, it was found that increasing the expression of the EGFRvIII receptor had very little effect on the phosphorylation states of Erk1, Erk2, and STAT3 (signal transducer and activator of transcription 3). The activity of all three of these proteins had previously been shown to quickly increase in response to activation of wild‐type EGFR. An increase in the levels of EGFRvIII did result in a greater than threefold increase in the phosphorylation state of phosphatidylinositol 3‐kinase (PI3K) and Grb2‐associated binding protein (GAB1). Since PI3K has been associated with cell proliferation, survival, and migration, its activation may partially explain the role of EGFRvIII in promoting tumorigenesis in vivo. One surprising result of this study was the observation that EGFRvIII and wild‐type EGFR utilize different downstream pathways, suggesting that therapies to inhibit EGFRvIII may not disrupt EGFR signaling that is required for cell survival. 11.5.2  Validation of the Role of c‐Met Signaling While this study demonstrated the value of proteomics for accumulating data capable of deciphering protein signaling throughout pathways and networks, its results still require careful validation of specific signaling events if it is to be useful in unlocking specific targets that can be considered as prognostic indicators of GBM outcome. In this study, the dataset was binned to identify sites whose phosphorylation state changed in a mode that was similar to that found for EGFRvIII levels. Not only did the data show that EGFRvIII was resulting in constitutive activation of the c‐Met receptor pathway through phosphorylation at Y1234, but many of the downstream components of the c‐Met pathway were also activated. Within this group were Y62 of the tyrosine phosphatase SHP‐10 and Y1234 of the c‐Met receptor. While c‐Met is a well‐known tyrosine kinase, little is known about the role of phosphorylation of Y62 of SHP‐10, except that this protein is downstream of the c‐Met receptor. To further study the relationship between EGFRvIII and c‐Met activation, U87‐H cells were treated with SU11274, a c‐Met kinase inhibitor. Direct analysis of c‐Met and EGFRvIII showed that Y1234 phosphorylation was decreased by treating the cells with SU11274, but the phosphorylation state of EGFRvIII was not affected. These data suggested that the EGFRvIII‐dependent increase in phosphorylation at Y1234 was resulting in a constitutively activated c‐Met receptor pathway. This initial observation was confirmed using circumstantial evidence by virtue of the fact that

Y1234

Y974

STAT3 Y704

Y1086 Y1148

EGFRvIII EGFRvIII

c- Met

282  Proteomics for Biological Discovery

Y845 Y1068 Y1173

PLC γ

Y1254

PKC δ Y62

SHP2

GAB1

Y542

SHC

Grb2

Y318 Y605 Y646

p85 p110

SOS

Y524

RAS PKD

Akt

RAF ERK1 MEK ERK2

Y204 Y187

Figure 11.6  Diagram showing tyrosine phosphorylation sites identified in EGFRvIII or c‐Met s­ timulated U87MG cells. The phosphorylation states of the proteins in italics and dark gray circles were found to  be highly abundant in both EGFRvIII and c‐Met activated cells. Source: Adapted from Huang et al. (2007).

other phosphorylation sites identified in the global analysis could be directly identified as known elements that were downstream of the c‐Met receptor. Many of these EGFRvIII‐dependent phosphorylation sites were upregulated more than threefold. Since these proteins were both downstream of c‐Met and known to be activated by EGRvIII, it is likely that both c‐Met and EGFRvIII stimulation was required for their activation. U87‐H cells were then treated with the c‐Met kinase inhibitor SU11274 to further evaluate the linkage between EGFRvIII and c‐Met signaling (Sattler et al. 2003). Quantitative MS showed a decrease in the phosphorylation of Y1234 of the c‐Met receptor upon SU11274 treatment. The data also showed a significant overlap in proteins, including SHC, GAB1, PLCγ, p85, and p110, which were stimulated by coactivation of EGFRvIII and c‐Met receptors (Figure  11.6). To measure synergy between the receptors, U87‐H cells were treated with either an EGFRvIII or c‐Met

Phosphorylation  283

inhibitor (AG1478 or SU11274). High doses of each inhibitor were required to decrease cell viability. When the U87‐H cells were treated with a lower dose of AG1478 and a higher dose of SU11274, cell viability and death were significantly decreased and increased, respectively. This change in cell viability and death was observed even when lower overall doses of either inhibitor were used individually. These results show that c‐Met inhibitors produce responses that are driven by on‐ target effects against c‐Met itself or, perhaps, c‐Met and another kinase that is also targeted by the inhibitor. EGFRvIII‐expressing GBM xenografts are resistant to cisplatin, and a wide variety of chemotherapeutics, unless it is administered along with AG1478 (Johns et al. 2003). Based on the significant overlap of downstream proteins activated by EGFRvIII/c‐Met stimulation that were identified in this quantitative phosphoproteomics study, the investigators hypothesized that the chemoresistance of EGFRvIII‐expressing tumors was related to EGFRvIII‐mediated activation of c‐Met. To test this hypothesis, U87‐H cells were treated with a constant dose of cisplatin and various dosages of SU11274. This combination of cisplatin and c‐Met inhibitor caused a dramatic decrease in cell viability compared to cisplatin alone, confirming that the chemoresistance of EGFRvIII‐positive GBM tumors may be partially due to c‐Met activation. In this study, global phosphorylation analysis using MS revealed that EGFRvIII phosphorylation mediates the activation of c‐Met. While an impressive number of phosphoproteins were quantitated, the conclusions of this study did not require thousands of proteins to be measured to uncover important information concerning the action of constitutively active EGFRvIII. It would be very challenging to obtain the same level of information using a hypothesis‐driven method such as ELISA or western blotting, which would not only be extremely inefficient but also would have been unlikely to uncover the wealth of information discovered using this MS approach. Ultimately, the MS approach showed that treating cells that express high levels of EGFRvIII with c‐Met kinase inhibitors and cisplatin or c‐Met kinase inhibitors and EGFR kinase inhibitors results in higher cytotoxicity. The importance of these findings is demonstrated by the large number of phase I, II, and III clinical trials that have been conducted to test the efficacy of c‐Met inhibitors on many ­different tumor types (Macek et al. 2006). 11.5.3  Targeted, Quantitative Phosphoproteomics Once a list of important phosphorylation sites is determined, the goal is to be able to monitor changes in these sites under a number of conditions. Using western blotting to monitor one site per experiment becomes limited very quickly as the number of phosphorylations increases. Fortunately, MS‐based proteomic methods have been developed that help avoid this glut. Multiplex immuno‐multiple reaction monitoring (MRM) studies in which specific phosphorylated residues are targeted for analysis are able to precisely quantify a large number of phosphoproteins within cell signaling networks. Using an MRM method allows the investigator to select which peptides to measure instead of using a data‐dependent (or shotgun) acquisition in which the mass spectrometer selects based on ion intensity. As a demonstration of this technology, Whiteaker et al, developed a multiplexed immuno‐MRM assay targeting the DNA damage response network as shown in Figure 11.7 (Whiteaker et al. 2015).

284  Proteomics for Biological Discovery Identification of targets from empirical data and available databases

Select 67 immunogen peptides Generate anti-phosphopeptide affinity purified polyclonal antibodies PO4

serum Resin coupled to phosphopeptide elute Resin coupled to unmodified peptide

Elute antibodies Evaluate assay performance

Intensity

Spiked standard

Endogenous peptide

Time

Determine LLOD, LOQ, and reproducibility

Evaluate in proof-of-concept studies Figure 11.7  Summary of experimental workflow for developing multiplexed immunoassay to measure DNA damage response in both cell lines and excised tissue samples. The first stages of the study required production and purification of the antibodies directed against targeted phosphorylated peptides and their unmodified counterparts. Once the antibodies were purified, they were used to determine the analytical characteristics of the assay as it pertained to each individual target. Once these parameters were established, the multiplexed assay was evaluated in a series of proof‐of‐concept studies. LLOD, lower limit of detection; LOQ, limit of quantification. Source: Adapted from Whiteaker et al. (2015).

To identify the species to target, existing shotgun MS databases were evaluated to find 67 phosphopeptides that mapped to well‐characterized proteins involved in the DNA damage response network (Zheng and Keifer 2009; Kweon and Andrews 2013). Synthetic phosphopeptides were made from these 67 sequences and used as antigens to generate antibodies in rabbits. The resultant antibodies were able to extract both the phosphorylated and unmodified versions of each peptide.

Phosphorylation  285

The ability to immune‐extract and detect the targeted phosphopeptides and their corresponding unmodified versions was tested using a series of endogenous peptides and proteomic extracts from cell lines exposed to DNA damaging reagents. To evaluate the analytical parameters for each peptide, response curves were generated by adding known amounts of synthetically prepared “light” (i.e., nonstable isotope‐ labeled) and “heavy” (i.e., stable isotope‐labeled) versions of each peptide to the cell line extracts. By varying the amount of heavy peptide added, while keeping the amount of light peptide constant, the researchers were able to characterize the linear range, lower limit of detection (LLOD), and limit of quantification (LOQ) for each peptide. Out of the attempted assays, a total of 48 were able to detect endogenous peptides. These assays showed a linear range of ≥3 orders of magnitude and a median LLOD and LOQ of 1.4 and 2.0 fmol/mg, respectively. In a series of reproducibility experiments, the median intra‐ and interassay variabilities were 10% and 16% coefficient of variation, respectively. A series of validation studies were conducted to determine if the multiplexed assay could provide biologically relevant results. In the first study, DNA damage response was profiled in MCF 10A and peripheral blood mononuclear cells that had been exposed to either ionizing radiation (IR) or 0.5 mM methyl methanesulfonate (MMS) to induce genotoxic stress. The absolute quantities of both phosphorylated and nonphosphorylated peptides were observed to be up‐ and downregulated with each responsive analyte showing a typical pharmacodynamic (PD) curve. One ­general trend observed was that the levels of phosphorylated peptides moved in the opposite direction from their nonphosphorylated counterparts, showing that observed changes were not a result of overall increase in the concentration of the protein. This trend would be consistent with an increase or decrease in the degree of phosphorylation via the action of specific kinases or phosphatases. Some of the specific results observed included an increase in the phosphorylation status of serine residues 343 and 315 of NBS1 and p53, respectively. The PD curves for both of these peptides showed a maximal increase in phosphorylated at six and two hours in MMS‐treated and IR‐treated cells, respectively. In the next proof‐of‐principle study, the multiplexed assay was used to quantify changes in peptides from two surgically excised breast cancer tissues that had been exposed to IR ex vivo. Both tissues showed an increase in the amount of NBS1 that was phosphorylated at serine residue 343, but only the second excised tumor showed a concomitant decrease in the amount of unmodified peptide. Similar to the above result, the level of ATM phosphorylated at serine residue 367 increased in both tissues as a result of IR exposure; however, one of the tissue samples showed no change in the amount of the unmodified peptide. Surprisingly, the other tissue showed an increase in the amount of unmodified peptide, which is opposite to the general trend expected for phosphorylated peptides and their nonphosphorylated counterparts. For novel technologies such as those presented above, it is critical to compare the results to those obtained using an accepted technology for quantifying protein levels. In this case, several of the targets quantitated using the multiplexed MRM assay were analyzed using western blotting. The overall PD responses and sensitivities as measured by western blotting and multiplexed MRM assay were found to be similar. While the technology must still be rigorously validated in other settings and facilities before it becomes broadly utilized, multiplexed MRM assays for measuring changes in phosphorylation status offer a number of advantages, beyond throughput,

286  Proteomics for Biological Discovery

c­ ompared to western blotting or ELISA. Theoretically, the data should be of higher quality owing to the ability to spike known amounts of verifiable internal standards directly into the samples being analyzed. The direct detection capabilities of the mass spectrometer should eliminate errors due to nonspecific binding that can occur in other immunoaffinity‐based techniques. Ultimately, the throughput of multiplexed MRM assays should enhance our understanding of how individuals respond to treatment, enabling decisions on the effectiveness of therapeutics to be made quickly and correctly. 11.6 CONCLUSION One of the major goals of proteomics is to characterize the entire protein complement of cells. The hope is that once we fully understand the components within the cell, we can begin to put the pieces together which will then enable us to predict how cells respond to stimuli (and, most importantly, diseases) at a molecular level. While identifying the proteins involved is critical, it does not provide sufficient information to understand how cells function. Without understanding the signals that dictate how proteins interact, the dynamic nature of the cell cannot be predicted. While there are well over 100 different types of modifications, phosphorylation dominates in sheer number and importance to protein function. The advances made in recent years in identifying phosphoproteins in complex biological mixtures has been incredible. These advances have not only been in the sensitivity and speed of the mass spectrometer but have included novel sample preparation methods that enable the enrichment of phosphopeptides from complex biological matrices. Additional advances have included methods such as stable isotope labeling to measure the absolute quantities of phosphopeptides. The studies presented in this chapter demonstrate how invaluable proteomic technologies have become for characterizing phosphoproteins. While there are many variations of these studies, they all require small amounts of biological material and deliver a wealth of information. The field is moving away from experiments that focus on a single phosphorylated residue and heading toward those that measure multiple interrelated sites. This strategy fits well with our understanding of diseases such as cancer in which multiple aberrations occur within the cell and need to be studied within the same context. While our ability to collect data today may be greater than our ability to understand it, continued improvements in bioinformatics will undoubtedly get us to the goal of complete understanding of cellular mechanics. References Aftab, Q., Sin, W.C., and Naus, C.C. (2015). Reduction in gap junction intercellular communication promotes glioma migration. Oncotarget 6: 11447–11464. Beltran, L. and Cutillas, P.R. (2012). Advances in phosphopeptide enrichment techniques for phosphoproteomics. Amino Acids 43: 1009–1024. Bu, Y., Gao, L., and Gelman, I.H. (2011). Role for transcription factor TFII‐I in the suppression of SSeCKS/Gravin/Akap12 transcription by Src. Int. J. Cancer 128: 1836–1842.

Phosphorylation  287

Chu, G., Egnaczyk, G.F., Zhao, W. et al. (2004). Phosphoproteome analysis of cardiomyocytes subjected to beta‐adrenergic stimulation: identification and characterization of a cardiac heat shock protein p20. Circ. Res. 94: 184–193. Davies, G.C., Ryan, P.E., Rahman, L. et al. (2006). EGFRvIII undergoes activation‐dependent downregulation mediated by the Cbl proteins. Oncogene 25: 6497–6509. Ducret, A., Desponts, C., Desmarais, S. et al. (2000). A general method for the rapid characterization of tyrosine‐phosphorylated proteins by mini two‐dimensional gel electrophoresis. Electrophoresis 21: 2196–2208. Edbauer, D., Cheng, D., Batterton, M.N. et al. (2009). Identification and characterization of neuronal mitogen‐activated protein kinase substrates using a specific phosphomotif antibody. Mol. Cell. Proteomics 8: 681–695. Erickson, B.K., Jedrychowski, M.P., McAlister, G.C. et  al. (2015). Evaluating multiplexed quantitative phosphopeptide analysis on a hybrid quadrupole mass filter/linear ion trap/ orbitrap mass spectrometer. Anal. Chem. 87: 1241–1249. Eyrich, B., Sickmann, A., and Zahedi, R.P. (2011). Catch me if you can: mass spectrometry‐ based phosphoproteomics and quantification strategies. Proteomics 11: 554–570. Ficarro, S.B., McCleland, M.L., Stukenberg, P.T. et al. (2002). Phosphoproteome analysis by mass spectrometry and its application to Saccharomyces cerevisiae. Nat. Biotechnol. 20: 301–305. Grandal, M.V., Zandi, R., Pedersen, M.W. et al. (2007). EGFRvIII escapes down‐regulation due to impaired internalization and sorting to lysosomes. Carcinogenesis 28: 1408–1417. Han, J.M., Kim, J.H., Lee, B.D. et al. (2002). Phosphorylation‐dependent regulation of phospholipase D2 by protein kinase Cδ in rat pheochromocytoma PC12 cells. J. Biol. Chem. 277: 8290–8297. Han, W., Zhang, T., Yu, H. et al. (2006). Hypophosphorylation of residue Y1045 leads to defective downregulation of EGFRvIII. Cancer Biol. Ther. 5: 1361–1368. Hasan, N. and Wu, H.F. (2011). Highly selective and sensitive enrichment of phosphopeptides via NiO nanoparticles using a microwave‐assisted centrifugation on‐particle ionization/ enrichment approach in MALDI‐MS. Anal. Bioanal. Chem. 400: 3451–3462. Holland, E.C., Hively, W.P., DePinho, R.A., and Varmus, H.E. (1998). A constitutively active epidermal growth factor receptor cooperates with disruption of G1 cell‐cycle arrest pathways to induce glioma‐like lesions in mice. Genes Dev. 12: 3675–3685. Hresko, R.C., Hoffman, R.D., Flores‐Riveros, J.R., and Lane, M.D. (1990). Insulin receptor tyrosine kinase‐catalyzed phosphorylation of 422(aP2) protein. Substrate activation by long‐chain fatty acid. J. Biol. Chem. 265: 21075–21085. Huang, P.H., Mukasa, A., Bonavia, R. et al. (2007). Quantitative analysis of EGFRvIII cellular signaling networks reveals a combinatorial therapeutic strategy for glioblastoma. Proc. Natl. Acad. Sci. USA. 104: 12867–12872. Hunt, D.F., Shabanowitz, J., and Bai, D.L. (2015). Peptide sequence analysis by electron ­transfer dissociation mass spectrometry: a web‐based tutorial. J. Am. Soc. Mass Spectrom. 26: 1256–1258. Johns, T.G., Luwor, R.B., Murone, C. et al. (2003). Antitumor efficacy of cytotoxic drugs and the monoclonal antibody 806 is enhanced by the EGF receptor inhibitor AG1478. Proc. Natl. Acad. Sci. USA. 100: 15871–15876. Khoury, G.A., Baliban, R.C., and Fioudas, C.A. (2001). Proteome‐wide post‐translational modification statistics: frequency analysis and curation of the Swiss‐Prot database. Sci. Rep. 1:srep00090. Kim, D., Pai, P.J., Creese, A.J. et  al. (2015). Probing the electron capture dissociation mass spectrometry of phosphopeptides with traveling wave ion mobility spectrometry and molecular dynamics simulations. J. Am. Soc. Mass Spectrom. 26: 1004–1013.

288  Proteomics for Biological Discovery Korley, R., Pouresmaeili, F., and Oko, R. (1997). Analysis of the protein composition of the mouse sperm perinuclear theca and characterization of its major protein constituent. Biol. Reprod. 57: 1426–1432. Kweon, H.K. and Andrews, P.C. (2013). Quantitative analysis of global phosphorylation changes with high‐resolution tandem mass spectrometry and stable isotopic labeling. Methods 61: 251–259. Lee, J.K., Joo, K.M., Lee, J. et al. (2014). Targeting the epithelial to mesenchymal transition in glioblastoma: the emerging role of MET signaling. Oncol. Targets Ther. 7: 1933–1944. Li, L., Puliyappadamba, V.T., Chakraborty, S. et  al. (2015). EGFR wild type antagonizes EGFRvIII‐mediated activation of met in glioblastoma. Oncogene 34: 129–134. Lind, S.B., Artemenko, K.A., and Pettersson, U. (2012). A strategy for identification of protein tyrosine phosphorylation. Methods 56: 275–283. Macek, B., Waanders, L.F., Olsen, J.V., and Mann, M. (2006). Top‐down protein sequencing and MS3 on a hybrid linear quadrupole ion trap‐orbitrap mass spectrometer. Mol. Cell. Proteomics 5: 949–958. Manning, D.R., DiSalvo, J., and Stull, J.T. (1980). Protein phosphorylation: quantitative ­analysis in vivo and in intact cell systems. Mol. Cell. Endocrinol. 19: 1–19. Marcus, K., Moebius, J., and Meyer, H.E. (2003). Differential analysis of phosphorylated ­proteins in resting and thrombin‐stimulated human platelets. Anal. Bioanal. Chem. 376: 973–993. McDonald, B.J., Chung, H.J., and Huganir, R.L. (2001). Identification of protein kinase C phosphorylation sites within the AMPA receptor GluR2 subunit. Neuropharmacology 41: 672–679. Nagaraj, N., d’Souza, R.C., Cox, J. et al. (2010). Feasibility of large‐scale phosphoproteomics with higher energy collisional dissociation fragmentation. J. Proteome Res. 9: 6786–6794. Palumbo, A.M., Smith, S.A., Kalcic, C.L. et al. (2011). Tandem mass spectrometry strategies for phosphoproteome analysis. Mass Spectrom. Rev. 30: 600–625. Perkins, D.N., Pappin, D.J., Creasy, D.M., and Cottrell, J.S. (1999). Probability‐based protein identification by searching sequence databases using mass spectrometry data. Electrophoresis 20: 3551–3567. Platt, M.D., Salicioni, A.M., Hunt, D.F., and Visconti, P.E. (2009). Use of differential isotopic labeling and mass spectrometry to analyze capacitation‐associated changes in the phosphorylation status of mouse sperm proteins. J. Proteome Res. 8: 1431–1440. Rapp, M., Sadat, H., Slotty, P.J. et al. (2015). Feasibility of the EORTC/NCIC trial protocol in a neurosurgical outpatient unit: the case for neurosurgical neuro‐oncology. J. Neurol. Surg. A Cent. Eur. Neurosurg. 76: 298–302. Reinders, J. and Sickmann, A. (2005). State‐of‐the‐art in phosphoproteomics. Proteomics 5: 4052–4061. Roux, P.P. and Thibault, P. (2013). The coming of age of phosphoproteomics – from large data sets to inference of protein functions. Mol. Cell. Proteomics 12: 3453–3464. Rush, J., Moritz, A., Lee, K.A. et al. (2005). Immunoaffinity profiling of tyrosine phosphorylation in cancer cells. Nat. Biotechnol. 23: 94–101. Sacristán, C., Tussie‐Luna, M.I., Logan, S.M., and Roy, A.L. (2004). Mechanism of Bruton’s tyrosine kinase‐mediated recruitment and regulation of TFII‐I. J. Biol. Chem. 279: ­ 7147–7158. Sacristán, C., Schattgen, S.A., Berg, L.J. et al. (2009). Characterization of a novel interaction between transcription factor TFII‐I and the inducible tyrosine kinase in T cells. Eur. J. Immunol. 39: 2584–2595.

Phosphorylation  289

Sattler, M., Pride, Y.B., Ma, P. et al. (2003). A novel small molecule met inhibitor induces apoptosis in cells transformed by the oncogenic TPR‐MET tyrosine kinase. Cancer Res. 63: 5462–5469. Sickmann, A., Mreyen, M., and Meyer, H.E. (2002). Identification of modified proteins by mass spectrometry. IUBMB 54: 51–57. Stone, M.D., Chen, X., McGowan, T. et al. (2011). Large‐scale phosphoproteomics analysis of whole saliva reveals a distinct phosphorylation pattern. J. Proteome Res. 10: 1728–1736. Syka, J.E.P., Coon, J.J., Schroeder, M.J. et al. (2004). Peptide and protein sequence analysis by electron transfer dissociation. Proc. Natl. Acad. Sci. USA. 101: 9528–9533. Thingholm, T.E., Jensen, O.N., Robinson, P.J., and Larsen, M.R. (2008). SIMAC (sequential elution from IMAC), a phosphoproteomics strategy for the rapid separation of monophosphorylated from multiply phosphorylated peptides. Mol. Cell. Proteomics 7: 661–671. Thingholm, T.E., Jensen, O.N., and Larsen, M.R. (2009). Enrichment and separation of mono‐ and multiply phosphorylated peptides using sequential elution from IMAC prior to mass spectrometric analysis. Methods Mol. Biol. 527: 67–78. Vandenbogaert, M., Hourdel, V., Jardin‐Mathé, O. et  al. (2012). Automated phosphopeptide identification using multiple MS/MS fragmentation modes. J. Proteome Res. 11: 5695–5703. Whiteaker, J.R., Zhao, L., Yan, P. et al. (2015). Peptide immunoaffinity enrichment and ­targeted mass spectrometry enables multiplex, quantitative pharmacodynamic studies of phospho‐ signaling. Mol. Cell. Proteomics 14: 2261–2273. Yadav, A.K., Kadimi, P.K., Kumar, D., and Dash, D. (2013). ProteoStats  –  a library for ­estimating false discovery rates in proteomics pipelines. Bioinformatics 29: 2799–2800. Yanagida, M., Miura, Y., Yagasaki, K. et al. (2000). Matrix assisted laser desorption/ionization‐ time of flight‐mass spectrometry analysis of proteins detected by anti‐phosphotyrosine antibody on two‐dimensional‐gels of fibroblast cell lysates after tumor necrosis factor‐alpha stimulation. Electrophoresis 21: 1890–1898. Yates, J.R., Eng, J.K., Clauser, K.R., and Burlingame, A.L. (1996). Search of sequence databases with uninterpreted high‐energy collision‐induced dissociation spectra of peptides. J. Am. Soc. Mass Spectrom. 7: 1089–1098. Yu, L.R., Zhu, Z., Chan, K.C. et  al. (2007). Improved titanium dioxide enrichment of ­phosphopeptides from HeLa cells and high confident phosphopeptide identification by cross‐validation of MS/MS and MS/MS/MS spectra. J. Proteome Res. 6: 4150–4162. Zhan, Y. and O’Rourke, D.M. (2004). SHP‐2‐dependent mitogen‐activated protein kinase activation regulates EGFRvIII but not wild‐type epidermal growth factor receptor ­phosphorylation and glioblastoma cell survival. Cancer Res. 64: 8292–8298. Zheng, Z. and Keifer, J. (2009). PKA has a critical role in synaptic delivery of GluR1‐ and GluR4‐containing AMPARs during initial stages of acquisition of in vitro classical ­conditioning. J. Neurophysiol. 101: 2539–2549. Zhou, M., Meng, Z., Jobson, A.G. et al. (2007). Detection of in vitro kinase generated protein phosphorylation sites using gamma[18O4]‐ATP and mass spectrometry. Anal. Chem. 79: 7603–7610.

12 Large‐Scale Phosphoproteomics John R. Yates III Departments of Molecular Medicine and Neurobiology, The Scripps Research Institute, La Jolla, CA, USA

12.1 INTRODUCTION The tiny 1 mm worm Caenorhabditis elegans has roughly 19 000 genes and is com­ posed of only 1000 cells. Intuitively, the vastly greater complexity of humans would seem to require a greater number of genes to create the functional diversity of human physiology, and early speculation predicted that 100–300 000 genes would be found in the human genome. Instead, the human genome has been found to comprise roughly 20 000 genes. How then is human complexity derived from so few genes? The fewer than expected genes that form the human genome are significantly supple­ mented by a substantial diversity in isoforms and modified forms of gene products. These protein variants are known as proteoforms, and they allow a single gene to generate proteins with a wide range of functions. It is known that the same gene product expressed in different cells can have different or modified functions arising from slight variations in sequence, or posttranslational modification of proteins. For instance, protein phosphorylation is known to regulate function or to convey signals. The discovery of phosphorylation by Krebs and Fischer (1956) touched off many studies to identify the role of phosphorylation in biological systems, resulting in the identification of “writers,” “readers,” and “erasers” of phosphorylation modifications to proteins. More recently, efforts to identify phosphorylation sites on proteins has been facilitated by new methods based on mass spectrometry (MS). In the early studies of phosphorylation, the analytical methods used to identify sites of phosphorylation on proteins were difficult and time‐consuming, and most used

Proteomics for Biological Discovery, Second Edition. Edited by Timothy D. Veenstra and John R. Yates III. © 2019 John Wiley & Sons, Inc. Published 2019 by John Wiley & Sons, Inc.

291

292  Proteomics for Biological Discovery

r­ adioactive phosphate for detection. The radioactively labeled phosphorylated proteins and peptides relied on a variety of methods, including ion exchange chromatography, gel electrophoresis, and paper chromatography, to separate and isolate them for further analysis (van der Geer and Hunter 1994). In the 1980s, mass spectrometers became capable of ionizing and sequencing underivatized peptides and phosphopeptides, thus enabling new strategies to sequence peptides and identify sites of modifications (Gibson and Cohen 1990). Because these methods were practiced on single proteins, there was no demand to develop methods to enrich phosphopeptides. However, Andersson and Porath had already developed methods to enrich phosphoproteins using iron  ions (Fe+3) (Andersson and Porath 1986) that would subsequently be applied to phospho­ peptides by Muszyńska et al. (1992). Michel et al. were the first to apply a combination of Porath’s Fe+3 chromatography with fast atom bombardment (FAB) tandem MS to enrich and then identify the presence of phosphorylation on three photosystem II pro­ teins (Michel et al. 1988). In addition, the Michel et al. application used a combination of affinity and reversed‐phase (RP) chromatography to enrich for the peptides, the scale of the process was still small and it required nanomoles of material to obtain amino acid sequence and pinpoint the site of modification. The development of electrospray ionization (ESI) in 1988 allowed a high‐­ performance liquid chromatography (HPLC) column to be connected to the mass spectrometer, which simplified the introduction of samples and made analysis more efficient, with great improvements in detection limits for peptides. In 1993, a MS‐ specific assay for phosphosite identification was developed by Huddleston et  al. (1993). This method used negative ion mode to generate a specific fragment ion from phosphopeptides that could then be correlated with the m/z value of the intact pep­ tide for sequencing (Carr et  al. 1996). This method allowed identification and sequencing of peptides without the need to enrich and was quite useful for analyzing phosphorylation sites on single proteins. At this time, the scale of phosphorylation in the cell was assumed to be large, but it was not until newer strategies based on MS were developed that the true scale was determined. To achieve a more comprehensive analysis of protein phosphorylation, a wide variety of larger scale strategies were developed, all of which incorporated some form of enrichment prior to MS analysis. 12.2 METHODS A major challenge for the analysis of posttranslational modifications is that they are often substoichiometric, which means not all of the protein present will be modified and that different modifications can be present simultaneously, making it likely that a protein will be present in varied states of modification. The analysis of complicated mixtures such as cell lysate requires enrichment of phosphopeptides to improve the dynamic range of the analysis and to remove nonmodified peptides. For less compli­ cated mixtures, such as a protein complex, however, it is possible to assign modifica­ tion sites using a more direct analysis without enrichment. 12.2.1  Nonenrichment Methods – High Sequence Coverage The accurate assignment of a phosphorylation site requires a peptide of molecular weight between ~1000–2500 Da to achieve good fragmentation with sequence ions that bracket the modification site. A weakness of an approach that relies on a single

Large-Scale Phosphoproteomics  293

peptide spectrum to assign a phosphorylation site is that if the peptide has not achieved good fragmentation, it may be impossible to unambiguously assign the site of phosphorylation. Gatlin et al. showed that digesting separate aliquots of a protein using trypsin, subtilisin, and elastase would produce an extremely high level of pro­ tein sequence coverage (~95%+), and that the use of nonspecific proteases would result in “ladders” of peptides of different lengths covering the protein sequence, including the site of modification (Gatlin et al. 2000). Such a high level of sequence coverage allows confident identification of sequence mutations, as demonstrated by Gatlin et al. for hemoglobin variants (Gatlin et al. 2000). MacCoss et al. extended the Gatlin method to protein complexes as a means to identify posttranslational modifications (MacCoss et al. 2002). This method is agnostic to the type of modifica­ tion and has been used to identify modifications in many protein complexes (Cheeseman et al. 2002; Cao et al. 2017; Herzberg et al. 2006). A limitation to the approach has been the dynamic range of the stoichiometry (MacCoss and coworkers showed they could pick up at least a 10 to 1 difference) but this limitation may be alleviated by the better chromatographic methods and mass spectrometers that have been developed over the last decade. Application of the method to more complex mixtures of proteins required ­enrichment of the phosphoproteins prior to analysis, and as a result, a variety of new enrichment strategies have emerged. The Porath method (Andersson and Porath 1986; Muszyńska et al. 1992) was the first strategy to enrich phosphoproteins and phospho­ peptides, but there have been a number of variations that improve on the technique (Thingholm and Larsen 2016). The Porath method exploits the strong affinity between Fe+3 and phosphate to enrich phosphopeptides, but Fe+3 also acts like an ion exchanger and thus enriches for acidic peptides as well. For this reason, efforts were made to develop methods that improve the specificity and selectivity of enrichment. 12.3  ENRICHMENT METHODS Several strategies have been developed that enrich specifically for phosphopeptides based on an affinity for phosphate. As shown in Figure 12.1, these methods entail different molecular mechanisms, and thus are susceptible to different background problems. 12.3.1  Precipitation Methods Using Ca+2, Ba+2 Phosphate is known to form insoluble salts with a variety of ionic metals. By using the appropriate salt at the right pH, phosphopeptides can be selectively precipi­ tated from solution. Zhang et al. adjusted the pH of a solution of digested proteins to 10 and then added CaCl2 to the peptide mixtures (Zhang et al. 2007). After mix­ ing, the solution was centrifuged and the supernatant collected (Figure 12.1a). The pellet was dissolved in an acidic buffer for subsequent analysis. Simple mixtures of digested phosphoproteins resulted in good enrichment of phosphopeptides using calcium ion precipitation, but for more complex mixtures of proteins Zhang and coworkers combined calcium ion precipitation with immobilized metal ion affinity chromatography (IMAC) to improve on the recovery of the phosphopeptides (Zhang et  al. 2007). This method was used on proteins from tomato with good enrichment of unique phosphopeptides (Stulemeijer et  al. 2009). No tyrosine

294  Proteomics for Biological Discovery

Phosphopeptide Enrichment Methods (a) Precipitation with Ca++, Ba++

(b) IMAC Fe+3, Ga+3

+

OPO3

+

+

3PO

Fe+3

+

–

+

–O

OH

+ +

OPO3–

+ +

OPO3–

+

–O

OH

OPO3–

–O PO 3

Fe+3

–O PO 3

OH Ca++ OPO3–

+ +

++

Ca O 3 OP

+

+

+

OH

Ca++ OPO3– + Ba++ OPO3–

3PO

OH

Fe+3

+

–O PO 3

OH

–

+

(c) Titanium Dioxide O

(d) Hydroxyapatite

–O

3PO

Ca++

Ti

–O

3PO

O O

–O

O O

–O

Ti O

3PO

Ca++

–O PO 3

Ca++

–O PO 3

Ti

–O PO 3

3PO

PO4–

–O PO 3 –O

3PO

–

PO4

PO4–

–O

3PO

Figure 12.1  Four methods to enrich phosphopeptides. (a) Precipitation with cations such as Ca+2 and Ba+2 has been used as a means to enrich for phosphopeptides. (b) Enrichment of phosphopeptides using heavy metals chelated to a solid support. Both Fe+3 and Ga+3 have been used in this process. As the heavy metal cations can also function as an ion exchange resin, care is taken to minimize background binding to acidic peptides. (c) Titanium dioxide has a high affinity for phosphopeptides and has been used to enrich for phosphopeptides. An advantage to the method is that titanium is part of the crystalline structure and thus the material does not need to be charged after each use, as with the IMAC materials. (d) Hydroxyapatite is a crystalline calcium material that can be used as a solid support in chromatography to enrich for phosphopeptides. (See color insert.)

­ hosphopeptides were observed, but this outcome was attributed to the extremely p low level of tyrosine phosphopeptides present. Ruse et  al. used barium ions for the same purpose, but incorporated both pH steps and ethanol precipitation to improve the efficiency of the precipitation and differentiate phosphate motifs (Ruse et al. 2008). By precipitating at pH 3.5, 4.6, and 8.0, different phosphate motifs were enriched (Figure  12.2). After precipitation, ­peptides were not reenriched by another method but analyzed directly by reversed‐ phase LC. The method was shown to be effective for small starting amounts of pro­ tein and to enrich phosphopeptides on a proteome‐wide scale. It was also shown to be capable of enriching multiply phosphorylated peptides, which is often difficult to do with other methods.

Large-Scale Phosphoproteomics  295

4

pH 3.5

Bits

3 2

10

9

8

7

6

5

4

3

2

1

0

11

1

Sequence Position

4

pH 4.6

Bits

3 2

11

9

8

7

6

5

4

3

2

1

0

10

1

Sequence Position

4

pH 8.0

Bits

3 2

11

9

8

7

6

5

4

3

2

1

0

10

1

Sequence Position

Figure 12.2  Sequence logos show the preference of phosphopeptides for the barium precipitation method when performed at specific pH values. At pH 3.5 there is a preference for acidic residues around the phosphorylation site. At pH 4.6 there is a preference for a proline residue right after the phosphorylation site with some preference of an acidic residue three residues away from the phosphorylation site. At pH 8, after a proline residue adjacent to the phosphorylation site, there is no hard preference for amino acid residues.The sequence surrounding a site of phosphorylation can indicate the kinase that might have a preference for the site. Source: Ruse et al. (2008). Reproduced with permission of the American Chemical Society.

12.4  ION CHROMATOGRAPHY FE+3, GA+3 IMAC Porath established a Fe+3‐based enrichment strategy for phosphoproteins and phos­ phopeptides, referred to as IMAC, which researchers have been optimizing for 30 years (Figure  12.1b). Improvements of the technique have been focused on achieving more efficient binding of phosphopeptides to the metal, more efficient elution from the column, and a reduction of nonphosphopeptide background. Many groups over  the years have optimized the pH, buffers, and other conditions to improve the recovery of peptides using Fe+3 IMAC.

296  Proteomics for Biological Discovery

In one of the first applications of Fe+3 IMAC, Nuwaysir and Stults (1993) d ­ eveloped an on‐line approach to directly enrich and analyze phosphopeptides from a digested protein using ESI. In 1999 the enrichment method was optimized by Cao and Stults and interfaced with capillary electrophoresis (Cao and Stults 1999). Since 1999, many reviews have been written describing the area and summarizing optimized condi­ tions (Thingholm et  al. 2009; Shou et  al. 2002; Carr et  al. 2005). Tempst showed improved specificity for phosphopeptide enrichment using Ga+3 in place of Fe+3 (Posewitz and Tempst 1999). Matrices charged with either ion are now commercially available. Both methods still suffer from the nonspecific enrichment of acidic pep­ tides, which complicates the enrichment of phosphopeptides from more complex protein mixtures. An advance for direct analysis of complex phosphopeptides came from the seminal study of yeast phosphorylation by Hunt’s laboratory in 2002 (Ficarro et al. 2002). To improve recovery and minimize nonspecific binding originating from acidic groups, they converted all the peptides to their methyl esters. This process blocked all acidic groups in the peptides and thus reduced ionic interac­ tions with the positively charged metal ions on the column. A disadvantage to the method is the need to convert peptides to their methyl esters, which may make the more hydrophobic peptides insoluble, so extra care has to be taken to insure peptides are fully solubilized to minimize losses. Others have found that lowering the pH of the loading buffer so the acidic amino acids in peptides are protonated helps to reduce nonspecific background (Thingholm et al. 2009). IMAC has shown a preference for multiphosphorylated peptides (Shou et al. 2002). An advantage to this method is the availability of commercial columns or 96‐well plates that allow parallelization of enrichments. As discussed below, the use of preenrich­ ment methods has made the parallelization of enrichments a very powerful approach. 12.4.1  Titanium Oxide Titanium dioxide (TiO2) was discovered by Ikeguchi and Nakamura to have an affin­ ity for phosphoamino acids (Ikeguchi and Nakamura 1997) (Figure  12.1c) and Kuroda et  al. and Heck et  al. both used it to enrich for phosphopeptides (Pinkse et al. 2004; Kuroda et al. 2004). Titanium dioxide provides a solid support that elimi­ nates problems associated with “charging” the matrix that can often plague Fe+3 IMAC. Most IMAC resins consist of agarose types of solid support that cannot han­ dle high pressures, but TiO2 can withstand HPLC scale pressures. Peptides are loaded onto a TiO2 support under acidic conditions and are eluted using a basic buffer. The ease of use has made this method quite popular (Thingholm et al. 2006). Titanium dioxide is a complementary method to IMAC as monophosphorylated peptides are more often observed. 12.4.2 Hydroxyapatite A variation on the Ca+2 precipitation method uses hydroxyapatite (HAP) as a means to enrich peptides. HAP is a naturally occurring substance (in fact, a variation of HAP is one of the main ingredients in bone and teeth) that has been used in chro­ matography despite having a complicated mechanism of separation thought to be a

Large-Scale Phosphoproteomics  297

mixed mode ion exchange (Figure 12.1d). A macroporous ceramic HAP is commer­ cially available from Bio‐Rad Laboratories, which has improved mechanical stability and surface area. Mamone et al. demonstrated the ability to enrich simple mixtures of phosphopep­ tides using HAP chromatography and showed that they could elute multiphospho­ rylated peptides (Mamone et al. 2010). Fonslow et al. coupled HAP with LC/LC/MS/ MS to create an in‐line enrichment method for the analysis of complex phosphopep­ tide mixtures (Fonslow et al. 2012). They also examined the type of phosphopeptides preferred by the HAP material compared with the standard Fe+3 IMAC method. Also shown was the suitability of the method for the analysis of small amounts of complex tissues. Phosphoprotein analysis was performed using 200–500 micrograms of the amygdala section of the brain. This HAP method was shown to enrich prefer­ entially for an acidic sequence within the phosphorylation region such as S.DDE… or S.DEE… while Fe+3 IMAC (in combination with hydrophilic interaction liquid chromatography (HILIC) prefractionation) selects for more basic regions of pep­ tides, often with proline present. 12.4.3  Hybrid Enrichment Methods As discussed above, the background of nonphosphorylated peptides with the use of Fe+3 IMAC alone proved to be problematic because of the ionic interactions of acidic groups within peptides and the Fe+3 immobilized on the column. Ficarro et al. devised a derivatization strategy to block acidic groups on peptides by converting peptides to their methyl esters to reduce these interactions (Ficarro et al. 2002). Zhang et al. com­ bined Ca+2 precipitation of phosphopeptides with Fe+3 IMAC to further enrich the peptides (Zhang et  al. 2007) and reduce the background of nonphosphopeptides. Jensen and coworkers developed a sequential strategy employing both IMAC and TiO2 as a means to improve the recovery of mono‐ and multiphosphorylated peptides (Thingholm et al. 2009). In this strategy, a complex mixture of peptides is added to the Fe+3 IMAC and monophosphorylated peptides are eluted under acidic conditions onto a TiO2 column. The multiphosphorylated peptides are then eluted from the Fe+3 IMAC column using a pH 11 buffer. By segregating phosphopeptides into two different pools, the MS analysis conditions can be better tailored for the idiosyncrasies of the different types of phosphopeptides. Multiphosphorylated peptides tend to ionize and fragment poorly, and thus MS3 strategies can be employed to better sequence the peptides. Gygi and coworkers used off‐line strong cation exchange (SCX) chromatography as a means to partially enrich phosphopeptides by selecting for peptides with reduced charge (Ballif et al. 2004; Beausoleil et al. 2004). A phosphorylated peptide should be of lower charge state than an unmodified tryptic peptide and thus elute earlier in an SCX separation (Figure 12.3a). The early fractions were subjected to Fe+3 IMAC to further enrich the phosphopeptides. The main problem with this strategy is the num­ ber of steps involved and the large amounts of material required. Typical amounts would be 10 mg or more of starting material making this method unsuitable for anal­ yses of small numbers of cells or small sections of tissue, such as biopsies. However, the strategy of combining a preenrichment with SCX with Fe+3 IMAC proved to be very successful to generate large numbers of phosphopeptides. In one of the first applications of this strategy, approximately 5000 phosphopeptides were measured in a HeLa cell nuclear fraction (Beausoleil et al. 2004). It has also been used on tissue

298  Proteomics for Biological Discovery

samples, where close to 6300 phosphoproteins in total were found over nine mouse tissues (Huttlin et al. 2010). The success of SCX as a preenrichment method led others to search for new approaches. McNulty and Annan introduced the use of HILIC, which separates pep­ tides based on differences in polarity (McNulty and Annan 2008) using a gradient of high to low organic modifier (Figure 12.3b). More polar peptides such as phospho­ peptides (which are more likely to be polar because of their negatively charged phosphate group) retain longer on the column, and thus elute later in the gradient. Fractions are collected from the column and each fraction is subsequently passed through a Fe+3 IMAC or a TiO2 column. Electrostatic repulsion‐hydrophilic interaction chromatography (ERLIC) is a novel mode chromatography designed to allow selective isolation of phosphopep­ tides from a tryptic digest. It is similar to HILIC except that the solid support has the same (or similar) charge as the molecules being separated, which causes electrostatic repulsion (Hao et al. 2010). By using a high organic buffer in the mobile phase, sol­ utes are retained on the column despite being electrostatically repulsed. In this method, highly charged solutes are preferentially retained. ERLIC has been exploited to enrich phosphopeptides, but care has to be taken to ensure the pH is low enough that only the phosphate group is charged so nonphosphorylated peptides are retained Pre-Enrichment Strategies (a)

(b)

(c)

–

OH

CH3

–

OH

CH3

SO3–

OH

CH3

SO3–

OH

CH3

SO3–

OH

SO3 SO3

Hydrophilic Interaction Chromatography (HILIC)

Decreasing Organic

Increasing Salt Relative Abundance

Phosphopeptides

Time

CH3

High pH Reversed-Phase LC

Increasing Organic

Phosphopeptides Relative Abundance

Strong Cation Exchange

Relative Abundance

CH3

OH

SO3–

Time

Time

Figure 12.3  Pre-enrichment strategies for phosphopeptide analysis. (a) Strong cation exchange (SCX) can be used to enrich phosphopeptides as the typical tryptic peptide will be doubly charged and thus the presence of one or more phosphates will create a lower charge state. Typically, phosphopeptides will elute in the very early part of the separation. (b) Hydrophilic interaction liquid chromatography (HILIC) separates peptides based on polarity using a decreasing organic gradient  –  higher organic at the start of the gradient and lower at the end. Phosphopeptides are generally polar and thus come out at the end of the gradient. (c) High pH reverse‐phase chromatography provides a highly resolving separation that can improve the recovery of phosphopeptides in subsequent enrichment steps because of decreased peptide complexity. (See color insert.)

Large-Scale Phosphoproteomics  299

on the column. Because the phosphopeptides are not firmly retained on the column, separations are generally performed isocratically. ERLIC has been compared to SCX and HILIC for preenrichment of phosphopeptides, and it was found that ERLIC is better at enriching multiply phosphorylated peptides than SCX or HILIC (Zarei et al. 2011). All three preenrichment methods were followed using a TiO2 enrichment. High pH (HpH)  reversed‐phase (RP) chromatography is an alternative form of enrichment that has different selectivity from low pH RP (Song et al. 2010). Marto and coworkers applied this method in a three‐dimensional separation of phosphopeptides that combined high pH RP, SAX, and low pH RP (Ficarro et al. 2011; Wang et al. 2016). Batth et al. demonstrated the use of HpH to prefractionate phosphopeptides prior to TiO2 enrichment (Batth et al. 2014; Batth and Olsen 2016) (Figure 12.3c). They found HpH improved recovery of singly phosphorylated peptides over an SCX prefractiona­ tion method but showed no benefit over SCX for doubly phosphorylated peptides. 12.4.4  Automated Hybrid Enrichment Methods Most phosphopeptide enrichment strategies contain several steps, and there have been efforts to automate or streamline the process (Figure  12.4). Automation can obviously make the enrichment process less labor intensive, but it can also result in Pre-fractionation of peptides

Enrichment of phosphopeptides

LC/MS/MS

Computational Analysis of Data

Figure 12.4  Hybrid enrichment strategies to improve recovery of phosphopeptides. A prefractionation step is used to provide partial enrichment of phosphopeptides or decrease the complexity of peptide mixtures. These fractions are collected and then subjected to a specific phosphopeptide enrichment step that uses affinity for phosphate as its primary enriching feature. These fractions are then subjected to LC/MS/MS to collect tandem mass spectra of phosphopeptides, which are then analyzed computationally to identify the phosphopeptides and the site of modification.

300  Proteomics for Biological Discovery

the loss of less material. As described above, Marto and coworkers created an inte­ grated system for phosphopeptide analysis based on HpH RP HPLC (Ficarro et al. 2011). Heck and coworkers also developed methods for automated enrichment and separation of phosphopeptides (Polat et al. 2012) using a polyurethane polymer chip combining TiO2 and RP media to create a 2D separation (Polat et al. 2012). As TiO2 is a solid support that can withstand high pressure, this type of separation system can be created. Many IMAC‐based media use agarose and other low‐pressure types of media that prohibit the creation of integrated separations and thus TiO2 proved to be a logical choice. The creation of an integrated separation system for enrichment and separation can help reduce peptide interaction with active surfaces. The more times small amounts of peptides are transferred between enrichment systems, the greater the losses. By working with an integrated system, losses can potentially be minimized, and the analysis of smaller quantities of samples may be possible (Masuda et al. 2011). 12.4.5  Data Analysis The analysis of phosphopeptide MS data has always been uniquely difficult (Yates et al. 1995). There are three amino acids that can be modified with phosphate (four counting His), but they are often modified in a differential manner. Thus, each poten­ tial site of phosphorylation has to be separately considered as modified or not. The stoichiometry of the modification, which is often less than one, can also complicate the analysis of phosphorylation data. Enrichment is key to minimizing the impact of substoichiometric modification, but enrichment can obscure the ratio of modified to unmodified protein. In addition, if a peptide has multiple sites of phosphorylation, and different sites are modified, it is very likely the peptides will coelute and be ­cofragmented during LC/MS/MS analysis. As these spectra will have fragment ions representing all sites of modification, all the forms may be identified by a search algorithm, but with different scores. These multiplexed spectra can also confound programs that are designed to verify phosphorylation site assignments. Search algorithms to identify sites of phosphorylation rely primarily on the shared peaks model. As the databases commonly used for searching do not specify sites of modification, searches test each site. For a peptide with five potential sites of modifica­ tion, that means there are 32 possible ways the peptide may be phosphorylated, but many of these can be eliminated by virtue of the peptide mass. SEQUEST uses a cross‐ correlation analysis to compare the shared peak models to the experimental spectrum and employs a differential search strategy to identify sites of modification (Yates et al. 1995). An interesting feature of this approach is the potential to detect multiplexed tan­ dem mass spectra. Probability search programs like MASCOT use a similar shared peaks model to assign modifications, including phosphorylation sites on peptides (Creasy and Cottrell 2004; Perkins et al. 1999). Search programs such as Inspect can identify sites of modification without specifying the type of modification expected and can potentially identify mixed modification peptides (Tanner et al. 2005). These methods of searching tandem mass spectra provide an unbiased strategy to identify sites of modification. 12.4.6  Validation and Verification After a search is completed, a validation and verification algorithm is often used to reassess the site assignments and to calculate a statistical confidence for the assign­ ments. To provide the highest confidence, assignment of the site of a modification

Large-Scale Phosphoproteomics  301

requires that a site be bracketed by fragment ions. When bracketing ions are missing, adjacent fragment ions can be used as long as there is not another adjacent site of modification. One of the early validation methods, Ambiguity Score or A‐score, cal­ culates a probability for the site assignment (Schwartz and Gygi 2005; Beausoleil et al. 2006). While A‐score is a valuable tool, it is limited for the assessment of mul­ tiplexed spectra, which are often assigned a poor score even when manual inspection reveals multiple sites of modification (Keck et al. 2011). Other algorithms have been used to recognize spectra with a high likelihood of being a phosphopeptide. Colander will filter spectra to reveal those that contain losses of 98, which indicates the likely presence of a phosphopeptide (Lu et al. 2008). Debunker is a program that assesses phosphopeptide identification using a machine learning classifier (Lu et al. 2007). There are also several approaches to assess the localization of phosphosites on peptides. While most search algorithms are fairly effective at identifying a peptide sequence and assigning phosphorylation sites, there can be some ambiguity even if a site is correctly localized within the sequence. Confident localization of a site depends on the presence of two or more fragment ions that bracket the location of the site and sometimes the fragment ions used to localize the modification may be distant from the actual site. Programs have been developed like A‐score that reassess the site assignment and calculate a probability that the assignment is correct. For instance, Taus et al. evaluated the Mascot Delta Score as a means to localize sites (Taus et al. 2011). They also compared the method with data collected by different fragmentation methods and determined that it works best with high‐energy methods such as higher collision‐energy dissociation and, to a lesser extent, ­electron transfer dissociation (ETD) (Frese et al. 2013). LuciPHOr is a method that uses mass accuracy and frag­ ment ion intensities for site localization scoring and calculation of a false localization rate. LuciPHOr is the only program that generates a false localization rate and is also compatible with the Trans‐Proteomic Pipeline (Fermin et al. 2013). PhosphoRS is a software tool to assess site localization of phosphorylations that is similar to A‐score (Taus et al. 2011). PhosphoRS is compatible with ETD, collision‐induced dissociation (CID), and HCD methods of fragmentation (Frese et al. 2013). A limitation to these computational methods is the ability to identify multiplexed spectra that may contain a peptide of the same sequence but phosphorylation at dif­ ferent sites within the peptide. Often when this happens, the peptides coelute during chromatography and then cofragment in the mass spectrometer because they have the same peptide mass. In an analysis of the phosphoproteome of the yeast centrosome, Keck et al. found a number of spectra that received low A‐scores because spectra were multiplexed (Keck et al. 2011). By using postsearch methods to assess site localization, more confidence can be provided for the analysis of large‐scale datasets. A goal of large‐scale phosphoproteomic studies is to identify changes resulting from a perturbation to cells. A handy tool for assessing identified phosphopeptides in large datasets and identifying the phosphorylation motifs present is the X‐Motif (Schwartz and Gygi 2005). This program is useful for identifying phosphorylation motifs present in large datasets and thus the types of kinases that might be involved. 12.4.7  Top‐Down Mass Spectrometry One of the challenges of using bottom‐up proteomics is determining patterns of modifications across the entire protein sequence. Different modification patterns define the proteoforms of a protein, which may reflect different functions or ­activities.

302  Proteomics for Biological Discovery

In theory, top‐down MS abrogates the limitations of bottom‐up methods and can better define the modification patterns on proteins. In practice, top‐down methods are limited by the ability to create sufficient cleavage along the amide backbone to accurately and unambiguously define the modification sites. Great strides have been made in increasing the scale of top‐down experiments in proteomics, but less pro­ gress has been made to achieve large‐scale analysis of phosphoproteins, and most modifications are still identified incidentally to the analysis. A study by Han et al. used top‐down MS to determine phosphorylation sites in the Dam1 protein complex after treatment with the kinase Mps1 (Han et al. 2014). By coupling the analysis to capillary electrophoresis and tandem top‐down MS, they were able to characterize sites of phosphorylation on some components of the com­ plex. By measuring molecular weights of the proteins, it was clear when there were multiple sites of modifications. As proteins increased in molecular weight, the level of fragmentation decreased and it became more difficult to unambiguously identify sites of phosphorylation. Recent advances in protein dissociation methods such as ETD and ultraviolet photodissociation (UVPD) have improved the efficiency of fragmentation for larger proteins and should result in more comprehensive results for larger proteins (Chi et al. 2007; Brodbelt 2014; Cannon et al. 2014a, b). An advantage to the use of top‐down MS for the characterization of posttrans­ lational modifications is the ability to directly observe the patterns of modifica­ tion sites. Nowhere is that more important than in histones, which contain a variety of different types of modifications whose patterns determine the function of the histone. Zheng et al. used top‐down MS on a Fourier transform mass spectrometer (FTMS) instrument to identify how specific phosphorylations alter the function of H1 histone variants (Zheng et al. 2010). They showed that specific sites in H1.2 and H1.4 were phosphorylated at different points in the cell cycle. In a very large‐ scale study using top‐down methods for protein analysis, Catherman et al. distin­ guished many phosphorylated proteins among the 1220 proteins identified in the study. In particular, a hyperphosphorylated form of HMGA2 was identified and characterized (Catherman et  al. 2013). While top‐down methods are not yet as high throughput or as large scale as bottom‐up approaches, they do provide data about modification sites that is lost in a bottom‐up study. But until methods for large‐scale analysis become more routine, targeted analysis of specific proteins or protein complexes remains the most effective strategy to determine patterns of modifications. 12.4.8  Applications of Large‐Scale Phosphoproteomics Large‐scale phosphoproteomics studies have been enabled by the combination of enrichment, shotgun proteomic methods and large‐scale data analysis methods. The first large‐scale study was performed in Saccharomyces cerevisiae by Ficarro et al. who detected approximately 1000 phosphopeptides that identified 283 phospho­ peptides with 383 phosphosites (Ficarro et al. 2002). This study used Fe+3 IMAC to enrich phosphopeptides but combined the approach with derivatization of phospho­ peptides to create methyl esters. By converting peptides to the methyl ester, acidic functionalities were eliminated and background from acidic peptides being captured on the media was reduced. In this study, phosphopeptides with single (60), double (215), and triple (11) phosphorylations were observed.

Large-Scale Phosphoproteomics  303

The publication by Ficarro et al. opened up the floodgates for the development of new strategies for large‐scale phosphoproteomics, but also for large‐scale applica­ tions of the methods. Beausoleil et al. introduced a prefractionation step of proteins by gel electrophoresis followed by a preenrichment step using SCX prior to Fe+3 IMAC to enrich phosphopeptides from a HeLa cell nucleus. While the method required the use of a large amount of starting material, it resulted in the identifica­ tion of 2002 phosphosites from 967 proteins (Beausoleil et al. 2004). From this data­ set, the authors were able to identify known and unknown kinase phosphorylation motifs. At the time, this analysis was a record for the number of phosphorylation sites observed and helped inspire more large‐scale studies. A study of the developing mouse brain used two different brain regions from a 16.5‐day‐old mouse to measure phosphorylation patterns. This study used the same strategy as Beausoleil and coworkers, starting with a prefractionation of 10 mg of protein from each brain region by gel electrophoresis, followed by SCX to preenrich phosphopeptides, and finally Fe+3 IMAC enrichment (Ballif et al. 2004). A total of 500 phosphorylation sites were identified in brain proteins. Phosphopeptide‐binding domains were observed in the dataset for the scaffolding protein 14‐3‐3. These two studies from the Gygi laboratory used an unprecedented three steps of fractionation prior to MS, primarily due to the use of a 3D quadrupole ion trap for the MS analy­ sis. This instrument was state of the art for scan speed and dynamic range at the time these studies were performed, but it still required a fair amount of upfront fractiona­ tion to prevent undersampling in complex mixtures of peptides. In 2003, the LTQ‐FTMS was introduced, an instrument that provides high‐­ resolution and high‐accuracy mass data for peptides which ushered in a new era in phosphorylation analysis. Villen et  al. employed the LTQ‐FTMS to identify phos­ phopeptides from mouse liver (Villen et al. 2007). Because the LTQ‐FTMS is not a fast scanning mass spectrometer, the multidimensional fractionation and enrich­ ment strategy was similar to that employed by Beausoleil et al. (2004). The use of high‐accuracy mass analysis resulted in the identification of 5635 unique phospho­ peptides from 2328 proteins, and the high‐accuracy mass measurement improved confidence in the identifications. Advances in mass spectrometers continued to improve the analysis of phospho­ rylation in biological systems. The introduction of the LTQ‐Orbitrap, which scanned faster than the LTQ‐FTMS with equal mass accuracy and similar mass resolution, was a significant advance in technology. Li et al. employed this instrument to analyze the phosphorylation changes observed when α‐mating factor arrests the budding yeast S. cerevisiae (Li et al. 2007). Again, using the Beausoleil et al. strategy, 2288 unique phosphorylation sites were identified in 985 proteins upon treatment of yeast with α‐mating factor (Beausoleil et al. 2004). This study determined phosphorylation sites in a number of the components of the sterile signaling pathway, an analog of a MAPK pathway. These large‐scale methods allowed analysis of how phosphorylation differs between tissues of the mouse. The biochemistry of tissues is complex, requiring many different cell types to act as a community. Thus there may be many pleiotropic effects of proteins as a subtle means of regulating protein activity. Using multidimensional fractionation and enrichment, Huttlin et  al. examined phosphorylation in nine mouse tissues (Huttlin et  al. 2010). A total of 12 039 proteins were identified, of which 6296 were phosphoproteins with nearly 36 000 phosphosites. This study

304  Proteomics for Biological Discovery

showed that within tissues there were specific phosphorylation networks. A com­ parison of protein expression levels between tissues suggested that many proteins were regulated by phosphorylation rather than by expression. Liao et al. applied a stable isotope‐labeling method (SILAM) in rats to quantitative differences between two different time points in rat brain development (Liao et al. 2008; McClatchy et al. 2007). A comparison was performed between rat brain cortex at postnatal day 1 and postnatal day 45 using 15N stable isotope‐labeled brain tissue as an internal standard; 705 phosphopeptides were quantitated in the P1 cortex and 1477 phosphopeptides in the p45 cortex, which translated to 380 and 585 phosphoproteins (Liao et al. 2008). It was observed that differentially modified phosphoproteins were upregulated on multiple components of chromatin remodeling complexes in the p1 cortex. In this particular study, phosphopeptides were enriched directly by Ga+3 IMAC and analyzed using a LTQ‐Orbitrap using LC/LC/MS/MS. As methods improved for large‐scale phosphoproteomics, the approach was increas­ ingly used to study specific biological processes. Dephoure et al. used these methods to study the phosphorylation events that occur during the mitotic cycle (Dephoure et al. 2008), a process driven by kinases. Cells were arrested at G(1) and every other step in the mitotic process. By using stable isotope labeling, phosphorylation levels could be compared at the various steps in the mitotic process. They identified >14 000 different phosphorylation events associated with mitosis. More than 1000 of the sites were regu­ lated sites and contained CDK phosphorylation domains. Genetic studies have implicated PINK1, a kinase involved in mitochondrial homeostasis, in Parkinson disease. Qin et al. used quantitative phosphoproteomics to identify changes when PINK1 is knocked down using siRNA (Qin et al. 2014). They observed that protein expression remained mostly unchanged, but phosphorylation changes were observed in over 100 different proteins. An analysis of phosphoryla­ tion motifs suggested a proline‐directed kinase specificity. Downstream signaling nodes involved transcription factors and other nuclear proteins involved in DNA and RNA metabolism. The data obtained by Qin et al. suggested that PINK1 kinase may regulate nuclear activities (Qin et al. 2014). Studies of schizophrenia patients have not identified a strong genetic signature for the disease. Many studies to understand molecular mechanisms associated with the disease will use an animal model that replicates specific phenotypes. One ani­ mal model uses phencyclidine and prepulse inhibition to replicate a specific pheno­ type of sensorimotor gating observed in human patients. McClatchy et al. used this model to study large‐scale signaling using phosphoproteomic methods and SILAM (McClatchy et  al. 2015). It was found that PCP downregulated phosphorylation events in the long‐term potentiation (LTP) pathway. The LTP pathway is involved in strengthening signals at the synapse as a means to create events such as memo­ ries. Downregulation of phosphorylation in the LTP pathway makes sense in the context of the loss of prepulse inhibition of startle upon treatment with PCP. Prepulse inhibition of startle requires some form of memory to remember and asso­ ciate the prepulse with the startling pulse. There are many more large‐scale studies not mentioned as well as applications where the analysis of phosphorylation was focused on its role in regulating a physio­ logical process. The methods developed to study phosphorylation’s role in physiology have had an enormous impact in understanding how processes are regulated.

Large-Scale Phosphoproteomics  305

12.5 CONCLUSION Regulation of biological processes is controlled by posttranslational modification and consequently, understanding when phosphorylation occurs, the site of phos­ phorylation within a protein, the location in the cell where it occurs, and the level of phosphorylation required to trigger a response are all key to understanding the regulatory process. Mass spectrometry and phosphopeptide and protein enrich­ ment are powerful tools to derive information about protein phosphorylation. These methods can be readily used on individual proteins or on complex systems to enable global analysis of phosphorylation and provide insight into how an entire cellular system is responding. The identification of phosphorylated resi­ dues such as Ser, Thr, and Tyr is now quite routine and robust. Phosphorylation of other residues, such as His, is more challenging because of the ease with which the phosphate group can be hydrolyzed (Fuhs et al. 2015). New strategies to charac­ terize phosphorylation on other amino acid residues will help to uncover their roles in cells.

References Andersson, L. and Porath, J. (1986). Isolation of phosphoproteins by immobilized metal (Fe3+) affinity chromatography. Anal. Biochem. 154 (1): 250–254. Ballif, B.A., Villen, J., Beausoleil, S.A. et al. (2004). Phosphoproteomic analysis of the develop­ ing mouse brain. Mol. Cell. Proteomics 3 (11): 1093–1101. Batth, T.S. and Olsen, J.V. (2016). Offline high pH reversed‐phase peptide fractionation for deep phosphoproteome coverage. Methods Mol. Biol. 1355: 179–192. Batth, T.S., Francavilla, C., and Olsen, J.V. (2014). Off‐line high‐pH reversed‐phase fractiona­ tion for in‐depth phosphoproteomics. J. Proteome Res. 13 (12): 6176–6186. Beausoleil, S.A., Jedrychowski, M., Schwartz, D. et  al. (2004). Large‐scale characteri­ zation  of HeLa cell nuclear phosphoproteins. Proc. Natl. Acad. Sci. USA. 101 (33): 12130–12135. Beausoleil, S.A., Villen, J., Gerber, S.A. et al. (2006). A probability‐based approach for high‐ throughput protein phosphorylation analysis and site localization. Nat. Biotechnol. 24 (10): 1285–1292. Brodbelt, J.S. (2014). Photodissociation mass spectrometry: new tools for characterization of biological molecules. Chem. Soc. Rev. 43 (8): 2757–2783. Cannon, J.R., Cammarata, M.B., Robotham, S.A. et al. (2014a). Ultraviolet photodissociation for characterization of whole proteins on a chromatographic time scale. Anal. Chem. 86 (4): 2185–2192. Cannon, J.R., Holden, D.D., and Brodbelt, J.S. (2014b). Hybridizing ultraviolet photodissocia­ tion with electron transfer dissociation for intact protein characterization. Anal. Chem. 86 (21): 10970–10977. Cao, P. and Stults, J.T. (1999). Phosphopeptide analysis by on‐line immobilized metal‐ion affin­ ity chromatography‐capillary electrophoresis‐electrospray ionization mass spectrometry. J. Chromatogr. A 853 (1–2): 225–235. Cao, L.W., Diedrich, J.K., Kulp, D.W. et al. (2017). Global site‐specific N‐glycosylation analysis of HIV envelope glycoprotein. Nat. Commun. 8: 14954.

306  Proteomics for Biological Discovery Carr, S.A., Huddleston, M.J., and Annan, R.S. (1996). Selective detection and sequencing of phosphopeptides at the femtomole level by mass spectrometry. Anal. Biochem. 239 (2): 180–192. Carr, S.A., Annan, R.S., and Huddleston, M.J. (2005). Mapping posttranslational modifica­ tions of proteins by MS‐based selective detection: application to phosphoproteomics. Methods Enzymol. 405: 82–115. Catherman, A.D., Durbin, K.R., Ahlf, D.R. et al. (2013). Large‐scale top‐down proteomics of the human proteome: membrane proteins, mitochondria, and senescence. Mol. Cell. Proteomics 12 (12): 3465–3473. Cheeseman, I.M., Anderson, S., Jwa, M. et  al. (2002). Phospho‐regulation of kinetochore‐ microtubule attachments by the Aurora kinase Ipl1p. Cell 111 (2): 163–172. ­ roteins Chi, A., Huttenhower, C., Geer, L.Y. et al. (2007). Analysis of phosphorylation sites on p from Saccharomyces cerevisiae by electron transfer dissociation (ETD) mass spectrometry. Proc. Natl. Acad. Sci. USA. 104 (7): 2193–2198. Creasy, D.M. and Cottrell, J.S. (2004). Unimod: protein modifications for mass spectrometry. Proteomics 4 (6): 1534–1536. Dephoure, N., Zhou, C., Villen, J. et al. (2008). A quantitative atlas of mitotic phosphorylation. Proc. Natl. Acad. Sci. U.S.A. 105 (31): 10762–10767. van der Geer, P. and Hunter, T. (1994). Phosphopeptide mapping and phosphoamino acid analysis by electrophoresis and chromatography on thin‐layer cellulose plates. Electrophoresis 15 (3–4): 544–554. Fermin, D., Walmsley, S.J., Gingras, A.C. et al. (2013). LuciPHOr: algorithm for phosphoryla­ tion site localization with false localization rate estimation using modified target‐decoy approach. Mol. Cell. Proteomics 12 (11): 3409–3419. Ficarro, S.B., McCleland, M.L., Stukenberg, P.T. et al. (2002). Phosphoproteome analysis by mass spectrometry and its application to Saccharomyces cerevisiae. Nat. Biotechnol. 20 (3): 301–305. Ficarro, S.B., Zhang, Y., Carrasco‐Alfonso, M.J. et al. (2011). Online nanoflow multidimen­ sional fractionation for high efficiency phosphopeptide analysis. Mol. Cell. Proteomics 10 (11): O111 011064. Fonslow, B.R., Niessen, S.M., Singh, M. et al. (2012). Single‐step inline hydroxyapatite enrich­ ment facilitates identification and quantitation of phosphopeptides from mass‐limited ­proteomes with MudPIT. J. Proteome Res. 11 (5): 2697–2709. Frese, C.K., Zhou, H., Taus, T. et  al. (2013). Unambiguous phosphosite localization using ­electron‐transfer/higher‐energy collision dissociation (EThcD). J. Proteome Res. 12 (3): 1520–1525. Fuhs, S.R., Meisenhelder, J., Aslanian, A. et al. (2015). Monoclonal 1‐ and 3‐phosphohistidine antibodies: new tools to study histidine phosphorylation. Cell 162 (1): 198–210. Gatlin, C.L., Eng, J.K., Cross, S.T. et  al. (2000). Automated identification of amino acid sequence variations in proteins by HPLC/microspray tandem mass spectrometry. Anal. Chem. 72 (4): 757–763. Gibson, B.W. and Cohen, P. (1990). Liquid secondary ion mass spectrometry of phosphoryl­ ated and sulfated peptides and proteins. Methods Enzymol. 193: 480–501. Han, X., Wang, Y., Aslanian, A. et al. (2014). In‐line separation by capillary electrophoresis prior to analysis by top‐down mass spectrometry enables sensitive characterization of pro­ tein complexes. J. Proteome Res. 13 (12): 6078–6086. Hao, P., Guo, T., Li, X. et al. (2010). Novel application of electrostatic repulsion‐hydrophilic interaction chromatography (ERLIC) in shotgun proteomics: comprehensive profiling of rat kidney proteome. J. Proteome Res. 9 (7): 3520–3526.

Large-Scale Phosphoproteomics  307

Herzberg, K., Bashkirov, V.I., Rolfsmeier, M. et al. (2006). Phosphorylation of Rad55 on ser­ ines 2, 8, and 14 is required for efficient homologous recombination in the recovery of stalled replication forks. Mol. Cell. Biol. 26 (22): 8396–8409. Huddleston, M.J., Annan, R.S., Bean, M.F., and Carr, S.A. (1993). Selective detection of ­phosphopeptides in complex mixtures by electrospray liquid chromatography/mass spec­ trometry. J. Am. Soc. Mass Spectrom. 4 (9): 710–717. Huttlin, E.L., Jedrychowski, M.P., Elias, J.E. et  al. (2010). A tissue‐specific atlas of mouse ­protein phosphorylation and expression. Cell 143 (7): 1174–1189. Ikeguchi, Y. and Nakamura, H. (1997). Determination of organic phosphates by column‐ switching high performance anion‐exchange chromatography using on‐line preconcentra­ tion on titania. Anal. Sci. 13 (3): 479–483. Keck, J.M., Jones, M.H., Wong, C.C. et al. (2011). A cell cycle phosphoproteome of the yeast centrosome. Science 332 (6037): 1557–1561. Krebs, E.G. and Fischer, E.H. (1956). The phosphorylase b to a converting enzyme of rabbit skeletal muscle. Biochim. Biophys. Acta 20: 150–157. Kuroda, I., Shintani, Y., Motokawa, M. et al. (2004). Phosphopeptide‐selective column‐switch­ ing RP‐HPLC with a titania precolumn. Anal. Sci. 20 (9): 1313–1319. Li, X., Gerber, S.A., Rudner, A.D. et al. (2007). Large‐scale phosphorylation analysis of alpha‐ factor‐arrested Saccharomyces cerevisiae. J. Proteome Res. 6 (3): 1190–1197. Liao, L., McClatchy, D.B., Park, S.K. et al. (2008). Quantitative analysis of brain nuclear phos­ phoproteins identifies developmentally regulated phosphorylation events. J. Proteome Res. 7 (11): 4743–4755. Lu, B., Ruse, C., Xu, T. et al. (2007). Automatic validation of phosphopeptide identifications from tandem mass spectra. Anal. Chem. 79 (4): 1301–1310. Lu, B., Ruse, C.I., and Yates, J.R. 3rd (2008). Colander: a probability‐based support vector machine algorithm for automatic screening for CID spectra of phosphopeptides prior to database search. J. Proteome Res. 7 (8): 3628–3634/. MacCoss, M.J., McDonald, W.H., Saraf, A. et al. (2002). Shotgun identification of protein mod­ ifications from protein complexes and lens tissue. Proc. Natl. Acad. Sci. USA. 99 (12): 7900–7905. Mamone, G., Picariello, G., Ferranti, P., and Addeo, F. (2010). Hydroxyapatite affinity chroma­ tography for the highly selective enrichment of mono‐ and multi‐phosphorylated peptides in phosphoproteome analysis. Proteomics 10 (3): 380–393. Masuda, T., Sugiyama, N., Tomita, M., and Ishihama, Y. (2011). Microscale phosphoproteome analysis of 10,000 cells from human cancer cell lines. Anal. Chem. 83 (20): 7698–7703. McClatchy, D.B., Dong, M.Q., Wu, C.C. et al. (2007). (15)N metabolic labeling of mammalian tissue with slow protein turnover. J. Proteome Res. 6 (5): 2005–2010. McClatchy, D.B., Savas, J.N., Martinez‐Bartolome, S. et al. (2015). Global quantitative analysis of phosphorylation underlying phencyclidine signaling and sensorimotor gating in the pre­ frontal cortex. Mol. Psychiatry 21 (2): 205–215. McNulty, D.E. and Annan, R.S. (2008). Hydrophilic interaction chromatography reduces the complexity of the phosphoproteome and improves global phosphopeptide isolation and detection. Mol. Cell. Proteomics 7 (5): 971–980. Michel, H., Hunt, D.F., Shabanowitz, J., and Bennett, J. (1988). Tandem mass spectrometry reveals that three photosystem II proteins of spinach chloroplasts contain N‐acetyl‐O‐ phosphothreonine at their NH2 termini. J. Biol. Chem. 263 (3): 1123–1130. Muszyńska, G., Dobrowolska, G., Medin, A. et al. (1992). Model studies on iron(III) ion affin­ ity chromatography. II. Interaction of immobilized iron(III) ions with phosphorylated amino acids, peptides and proteins. J. Chromatogr. 604 (1): 19–28.

308  Proteomics for Biological Discovery Nuwaysir, L.M. and Stults, J.T. (1993). Electrospray ionization mass spectrometry of phospho­ peptides isolated by on‐line immobilized metal‐ion affinity chromatography. J Am Soc Mass Spectrom 4: 662–669. Perkins, D.N., Pappin, D.J., Creasy, D.M., and Cottrell, J.S. (1999). Probability‐based ­protein  identification by searching sequence databases using mass spectrometry data. Electrophoresis 20 (18): 3551–3567. Pinkse, M.W., Uitto, P.M., Hilhorst, M.J. et al. (2004). Selective isolation at the femtomole level of phosphopeptides from proteolytic digests using 2D‐NanoLC‐ESI‐MS/MS and titanium oxide precolumns. Anal. Chem. 76 (14): 3935–3943. Polat, A.N., Kraiczek, K., Heck, A.J. et al. (2012). Fully automated isotopic dimethyl labeling and phosphopeptide enrichment using a microfluidic HPLC phosphochip. Anal. Bioanal. Chem. 404 (8): 2507–2512. Posewitz, M.C. and Tempst, P. (1999). Immobilized gallium(III) affinity chromatography of phosphopeptides. Anal. Chem. 71 (14): 2883–2892. Qin, X., Zheng, C., Yates, J.R. 3rd, and Liao, L. (2014). Quantitative phosphoproteomic ­profiling of PINK1‐deficient cells identifies phosphorylation changes in nuclear proteins. Mol. Biosyst. 10 (7): 1719–1729. Ruse, C.I., McClatchy, D.B., Lu, B. et al. (2008). Motif‐specific sampling of phosphoproteomes. J. Proteome Res. 7 (5): 2140–2150. Schwartz, D. and Gygi, S.P. (2005). An iterative statistical approach to the identification of protein phosphorylation motifs from large‐scale data sets. Nat. Biotechnol. 23 (11): 1391–1398. Shou, W., Verma, R., Annan, R.S. et al. (2002). Mapping phosphorylation sites in proteins by mass spectrometry. Methods Enzymol. 351: 279–296. Song, C., Ye, M., Han, G. et al. (2010). Reversed‐phase‐reversed‐phase liquid chromatography approach with high orthogonality for multidimensional separation of phosphopeptides. Anal. Chem. 82 (1): 53–56. Stulemeijer, I.J., Joosten, M.H., and Jensen, O.N. (2009). Quantitative phosphoproteomics of tomato mounting a hypersensitive response reveals a swift suppression of photosynthetic activity and a differential role for hsp90 isoforms. J. Proteome Res. 8 (3): 1168–1182. Tanner, S., Shu, H., Frank, A. et  al. (2005). InsPecT: identification of posttranslationally ­modified peptides from tandem mass spectra. Anal. Chem. 77 (14): 4626–4639. Taus, T., Kocher, T., Pichler, P. et  al. (2011). Universal and confident phosphorylation site ­localization using phosphoRS. J. Proteome Res. 10 (12): 5354–5362. Thingholm, T.E. and Larsen, M.R. (2016). Phosphopeptide enrichment by immobilized metal affinity chromatography. Methods Mol. Biol. 1355: 123–133. Thingholm, T.E., Jorgensen, T.J., Jensen, O.N., and Larsen, M.R. (2006). Highly selective enrichment of phosphorylated peptides using titanium dioxide. Nat. Protoc. 1 (4): 1929–1935. Thingholm, T.E., Jensen, O.N., and Larsen, M.R. (2009). Analytical strategies for phosphopro­ teomics. Proteomics 9 (6): 1451–1468. Villen, J., Beausoleil, S.A., Gerber, S.A., and Gygi, S.P. (2007). Large‐scale phosphorylation analysis of mouse liver. Proc. Natl. Acad. Sci. USA. 104 (5): 1488–1493. Wang, L.D., Ficarro, S.B., Hutchinson, J.N. et al. (2016). Phosphoproteomic profiling of mouse primary HSPCs reveals new regulators of HSPC mobilization. Blood 128 (11): 1465–1474. Yates, J.R. III, Eng, J.K., McCormack, A.L., and Schieltz, D.M. (1995). Method to correlate tandem mass spectra of modified peptides to amino acid sequences in the protein ­database. Anal. Chem. 67 (8): 1426–1436.

Large-Scale Phosphoproteomics  309

Zarei, M., Sprenger, A., Metzger, F. et al. (2011). Comparison of ERLIC‐TiO2, HILIC‐TiO2, and SCX‐TiO2 for global phosphoproteomics approaches. J. Proteome Res. 10 (8): 3474–3483. Zhang, X., Ye, J., Jensen, O.N., and Roepstorff, P. (2007). Highly efficient phosphopeptide enrichment by calcium phosphate precipitation combined with subsequent IMAC enrich­ ment. Mol. Cell. Proteomics 6 (11): 2032–2042. Zheng, Y., John, S., Pesavento, J.J. et al. (2010). Histone H1 phosphorylation is associated with transcription by RNA polymerases I and II. J. Cell Biol. 189 (3): 407–415.

13 Probing Glycoforms of Individual Proteins Using Antibody‐Lectin Sandwich Arrays: Methods and Findings from Studies of Pancreatic Cancer Brian B. Haab Van Andel Research Institute, Grand Rapids, MI, USA

13.1  INTRODUCTION: THE NEED FOR PRECISE MEASUREMENT OF PROTEIN GLYCOFORMS It is increasingly recognized that the glycosylation of proteins has a great effect on protein integrity and function. Carbohydrates (glycans) are found throughout every cell of all organisms and on most secreted and membrane‐bound proteins and lipids. Glycans have been implicated in the pathology of a wide variety of diseases, including infectious disease (Marth and Grewal 2008; Rudd et al. 2001; Stevens et al. 2006), cancer (Fuster and Esko 2005), autoimmune disease (Rudd et al. 2001), and certain congenital disorders (Freeze 2006). Understanding the structures and functions of members of this class of biomolecule therefore has significant utility. Examples of the medical applications of glycans include cancer vaccines (Astronomo

Proteomics for Biological Discovery, Second Edition. Edited by Timothy D. Veenstra and John R. Yates III. © 2019 John Wiley & Sons, Inc. Published 2019 by John Wiley & Sons, Inc.

311

312  Proteomics for Biological Discovery

and Burton 2010), targeting B cells through lectin receptors (Chen et al. 2010), and biomarkers based on the detection of altered glycans secreted by cancer cells (Adamczyk et  al. 2012). Despite these considerations, glycans have been studied much less than proteins and nucleic acids. Part of the reason for the lower research effort is the relative difficulty in studying glycans. Recombinant methods for producing large quantities of specific glycans are not available, in contrast to proteins and nucleic acids, and determining the primary sequence of a purified glycan is much more difficult than for proteins and nucleic acids. Considering the major role of glycans in biology, the development of improved tools for the study of glycans is an important goal. Glycan structures can be identified using a variety of methods based on mass spectrometry  (MS), enzymatic digestion, and chromatography. Mass spectrometry methods have particularly advanced in recent years to achieve more routine, reliable, and comprehensive composition analysis (North et al. 2009). Advances in both the technology and automated analysis of spectra (Perez and Mulloy 2005) have made glycan analyses accessible to a greater range of researchers. A fully automated platform for sample preparation, chromatographic separation, and interpretation (Marino et  al. 2010) has enabled large‐scale studies of the relationships between glycans and various patient characteristics (Lauc et al. 2010; Igl et al. 2011). These methods are fundamentally important but additional, complementary approaches are needed, particularly for application to biomarker research. Biomarker research requires precise measurements of specific structures over many different clinically obtained samples, which often are available in limited amounts. Furthermore, precise measurements of the glycoforms of specific proteins could provide enhanced value. A valuable approach for measuring specific glycans and protein glycoforms in a format compatible with clinical or biomarker research is to use affinity reagents such as lectins or glycan‐binding antibodies. 13.2  ANTIBODIES AND LECTINS FOR PROBING GLYCOSYLATION OF INDIVIDUAL PROTEINS Lectins and glycan‐binding antibodies, collectively known as glycan‐binding proteins, can be used in a multitude of analytical formats (Rudiger and Gabius 2001) such as histochemistry, the probing of electrophoretic gels, affinity chromatography, solid‐phase ELISA‐type assays, and microarray assays (Figure 13.1). The information gathered from affinity reagents is highly complementary to that obtained from MS‐based approaches. Glycan‐binding proteins can provide reproducible measurements on specific structures over many samples, whereas MS typically yields information on many structures in fewer samples, with less precision. Therefore, glycan‐binding proteins miss the detailed information that MS gives but instead provide precise information on changes across samples. An important aspect of the posttranslational modification of glycosylation is that it can vary between different proteins. The glycans in many cases are uniquely directed to specific proteins, based most likely on either recognition of polypeptide sequences (Pedersen et  al. 2011) or localization of specific glycosytransferases to defined areas of the protein‐processing pathway (Sun et  al. 2011). Therefore, the

Probing Glycoforms Using Antibody-Lectin Sandwich Arrays  313

Type of microarray

Glycan binding protein Or indirect detection by antibody

Glycan Glycoprotein Lectin Biological sample

Lectin

Antibody

Microarray

Lectin Glycoprotein Slide

Isolated glycan

Conjugated glycan

Dye

Lectin Antibody

Printed antibody Protein

Figure 13.1  Microarray formats for glycoproteomics research. The types of microarrays depicted are glycan arrays, lectin arrays, antibody‐lectin sandwich arrays, and glycoprotein arrays. A detection strategy using a fluorescent dye is depicted, although other detection methods could be used, such as surface‐plasmon resonance or chemiluminescence. Source: Yue and Haab (2009).

types of structures added to a given protein cannot be inferred from the expression levels of various glycosyltransferases; protein‐specific glycosylation must be experimentally determined for each protein. The study of glycans on particular proteins is a difficult task using MS and chromatography methods, given the need to isolate large amounts of the protein. Antibody‐lectin sandwich arrays (ALSAs) enable the probing of the glycosylation state of multiple proteins in parallel (Chen et al. 2007). Antibody arrays can be used to capture many different proteins from a biological sample, and incubation with a labeled lectin provides information on the glycosylation of the captured proteins (see Figure 13.1). We and others (Li et al. 2009, 2013; Kuno et al. 2009) have previously shown the value of this approach for biomarker research. In certain cases, the detection of protein glycoforms yields better biomarker performance than conventional proteins assays, owing to greater disease‐associated changes in glycosylation than in core protein abundance (Chen et  al. 2007; Yue et  al. 2009, 2011). We are currently applying this method to problems in the diagnosis and management of pancreatic cancer and other cancers in which glycan alterations have been observed. This approach could be valuable in many other situations, given the prevalence of alterations in glycosylation state. Important considerations and guidelines for developing customized assays are given immediately below.

314  Proteomics for Biological Discovery

13.3  DEVELOPING AND VALIDATING CUSTOM ANTIBODY‐LECTIN ASSAYS 13.3.1  Finding the Right Lectins: New Bioinformatics Tools The first step in developing assays is obtaining the lectins and antibodies. Since the choice of capture antibodies is unique to each project and methods for their testing are well known, the focus of the text below is on the lectins. Previous papers have given detailed information on running the assays (Chen and Haab 2009; Haab and Yue 2011), which is not repeated here. To properly interpret data obtained using lectins, it is necessary to have precise information on the specificities of the lectins. Some lectins are highly ­specific for a single monosaccharide in a particular linkage, while others bind tri‐ or tetrasaccharides found in a variety of presentations. The exploration of lectin specificity typically has been performed using purified or synthesized glycans. For example, in frontal affinity chromatography (Hirabayashi et al. 2000), the rate at which a glycan elutes from a column containing an immobilized lectin provides information on the ­dissociation constant of the interaction. This method has been used to profile the affinities of many different lectins for hundreds of different glycans, such as in a characterization of the specificities of galectins (Hirabayashi et al. 2002). A more recently developed tool for the study of lectin specificity is the glycan microarray (Rillahan and Paulson 2011). Glycan microarrays enable the parallel probing of lectin binding to hundreds of biologically relevant glycans in a single experiment, with low reagent consumption. The low consumption of the carbohydrate structures is particularly important because of the difficulty and time required to synthesize or isolate those structures. Several demonstrations of carbohydrate microarrays appeared in 2002 with a variety of fabrication techniques (Culf et  al. 2006; Wang 2003). A significant advance in the utility and availability of glycan microarray technology came two years later through the development of glycan microarrays by the Consortium for Function Glycomics (CFG). Researchers from the CFG synthesized over 200 biologically r­elevant glycans attached to amine‐conjugated spacers, and spotted them onto N‐hydroxysuccinimide (NHS)‐activated glass slides to form covalent linkages (Blixt et  al. 2004). The arrays were initially used to characterize specificities of plant lectins, human lectins, glycan‐binding antibodies, and bacterial and viral ­proteins (Blixt et  al. 2004). Since then, the CFG has profiled the specificities of hundreds of glycan‐binding proteins for researchers participating in the consortium, with the data made available through the CFG website. The availability of these datasets presents a tremendous opportunity for researchers to identify lectins with defined specificities. Right now, the glycan array data are largely unprocessed, leaving it to the user to figure out from the relative binding to each glycan the precise specificity of a lectin. Such analysis is time‐consuming and imprecise if done manually. To enable a more rapid and objective determination of lectin specificity from glycan array data, we developed the motif segregation analysis method (Porter et al. 2010) (Figure 13.2). We have shown that this automated method of analyzing glycan array data correctly extracts the main binding specificities. An additional method, called outlier motif analysis, builds on motif segregation to enable more detailed identifications of fine specificities (Maupin et al. 2011). The output of this program is a list of component substructure motifs of the glycans and

Probing Glycoforms Using Antibody-Lectin Sandwich Arrays  315

(b)

Fluorescence

60000

ConA Glycan Array Data

40000

20000

All glycans Contain Do not contain motif Compare motif intensities p-value

50

100 150 200 250

Glycans

Motif score

12

Motif Score (log 10)

(a)

10

Mannose containing motifs

8

Terminal glucose containing motifs Other motifs

6 4 2 0 –2

0

50

100

150

Motifs Repeat for all lectins

Database of lectin binding specificities

Figure 13.2  Motif analysis of glycan array data. (a) Motif segregation for interpreting glycan array data. The left panel shows raw data after incubation of a glycan array with the lectin ConA. The amount of binding of ConA (in fluorescence units) at each of 256 distinct glycans is presented. Next, the glycans are separated according to whether they contain a particular substructure (motif). The intensities of the two groups are compared and converted to a score based on statistical significance. That process is repeated for many different motifs. The right panel shows that the highest‐scoring motifs all contain α‐mannose, the known specificity of ConA. Terminal glucose is a known secondary specificity. (b) An interface for creating user‐defined motifs. This software tool allows users to easily create arbitrary motifs by text or graphic visualization. This capability allows exploration of novel motifs in glycan array data using motif segregation. (See color insert.)

an associated score indicating the preference of the lectin for binding that motif. This relative quantitation of motifs enables an objective determination of binding specificity. Furthermore, the compilation of analyses from many different datasets enables comparisons and groupings of specificities between lectins (see Figure 13.2), as well as searches for lectins with predefined specificities. Software for the full automation of glycan array data has been developed by researchers at the Palo Alto Research Center. This new software is a key step in assembling the quantitative information necessary in developing a comprehensive resource for lectin specificities. Currently, the software is available upon request, and a complete database of motif scores for all lectins analyzed by glycan array is in development. 13.3.2  Controls for Validating and Optimizing Lectin Binding Once the lectins are obtained, it is necessary to verify their proper performance in the assay, for which controls are necessary. Positive control molecules, which display the glycan targeted by the lectin, can be spotted together in the same arrays with negative controls, which do not have the glycan. The positive controls could be a glycoprotein or a synthesized glycan, and the negative controls could be a nonglycosylated protein or an off‐target glycan. Each array also should include antibodies that should not capture anything (negative control antibodies) and antibodies that capture proteins known to be present in certain samples (positive control antibodies). The negative control antibodies could be mouse monoclonal antibodies targeted against something never present in the type of samples to be analyzed. For example, an antibody against the HA fusion tag would not be expected to capture anything present in human serum, and the signals from this antibody should extremely low. As a

316  Proteomics for Biological Discovery

positive control, an antibody against human IgM should show predictable signal in all human serum samples. A positive control for the secondary detection reagent is a biotinylated antibody, which should show strong signal from streptavidin‐based detection. Once the proper positive and negative control elements are assembled, microarrays can be produced using a microarrayer robot. The details of this process depend on the equipment that is available to the researcher. The next step is to incubate the lectins on the arrays and detect the level of binding at each spot. If a fluorescence scanner is available, a convenient method is to use biotinylated lectins followed by incubation with dye‐labeled streptavidin (see Figure  13.1). Otherwise, visualization using enzyme‐labeled streptavidin followed development with a substrate also works well (Huang et al. 2001, 2004). The lectin concentration can be varied to maximize the signals at the positive control spots and minimize the signals at the negative control spots. Some lectins have a high tendency to bind nonspecifically to all proteins or the underlying substrate and thus require lower concentrations, while others are optimally run at higher concentrations. Given the diversity among lectins, the experimental conditions need to be individually optimized. The positive and negative control spots should be included in the arrays in all subsequent experiments to verify the proper functioning of the lectins. 13.3.3  Setting the Dilution Factor and Using Calibrator and Control Samples The next step is to determine if the verified antibodies and lectins can be used to capture proteins out of biological samples and measure glycosylation on the captured proteins. Ideally, purified proteins are available for each of the proteins targeted by the antibody array. If these are available, they can be used as controls to verify that the capture antibody is functioning properly and to calibrate the linear range of the assay, which is the range of analyte concentrations in which changes in concentration can be observed. If purified proteins are not available, as is often the case, biological samples can be used that are known to contain the protein. The linear range can be assessed by running a dilution series of the purified protein or control sample with each set of experiments and then determining the lowest and highest signals at which changes are observed. The experimental samples should be diluted to a concentration that will give signals in the linear range. Since the starting concentrations of the experimental samples are now known beforehand, before running a large experiment set it is useful to run representative samples at a few dilutions each, to see which dilution best places the samples in the linear range. Negative control arrays, which should be included in each experiment set, are incubated with buffer instead of sample. These arrays should show no signal except at the biotinylated antibody spots (the positive control spots). If measurable ­signals are observed at some of the capture antibody spots, the cause could be nonspecific binding of the detection reagents to the spots or contamination of the antibody preparations. Other negative control arrays can sort out these alternatives, namely by detecting with just the secondary reagent (the dye‐labeled streptavidin) or  ­nothing at all. The solution depends on the identified source of nonspecific ­signal. Contaminants in the antibodies could be removed by dialysis (to remove small ­molecule contaminants) or ultracentrifugation (to remove aggregates), and

Probing Glycoforms Using Antibody-Lectin Sandwich Arrays  317

­ onspecific binding of lectins to the capture antibodies can be reduced using n ­chemical derivatization of the antibodies (Chen et al. 2007). Positive control arrays are incubated with samples known to contain particular proteins that are targeted by particular antibodies on the arrays. For example, an array could be incubated with a normal serum sample in order to observe the signal expected at an antibody targeting fibronectin, a protein present in high abundance in all sera. The signals from these spots can be compared to previous results to be sure that the detection reagents are working properly. In addition, samples containing the targeted proteins can be incubated in a series of dilutions on a set of arrays. The signal at each of the antibodies should reduce along with increasing dilution, finally reaching a minimum in the array incubated with just buffer. Ideally, the shape of the dilution curve is sigmoidal, with a plateau at high concentrations and a tapering off at low concentrations. A lack of consistent data from the positive control arrays could indicate a need to replace the capture or detection reagents. These controls will help to determine the quality of the data and identify routes for solving problems. Finally, one should consider the stability of the assay between runs and over long periods of time. The use of control samples with each experiment set can help assess the consistency of the assay over time. A small set (5–10 samples) of representative samples with a range of analyte concentrations can be run with each experiment, randomly interspersed with the experimental samples and treated identically, and the measurements of those samples can be compared to previous experiment sets. The use of these techniques, given stable antibody and lectin reagents, should enable the development of consistent and robust assays. Examples of approaches for designing experiments and interpreting results are provided in the following descriptions. 13.4  GLYCANS ASSOCIATED WITH PANCREATIC CANCER: FINDINGS FROM ALSA STUDIES The ALSA method has been applied most extensively to the study of glycosylation on specific proteins associated with pancreatic cancer. These studies have provided new information about the nature of protein glycosylation in this disease and have provided leads on new biomarkers. A summary of the strategic approaches and main findings follows. 13.4.1  Carrier Proteins of the CA 19‐9 Antigen in Cancer and Other Conditions The current best serological marker for pancreatic cancer is a carbohydrate structure called the CA 19‐9 antigen (Pleskow et al. 1989; Ko et al. 2005; Ritts Jr et al. 1994; Goonetilleke and Siriwardena 2007). This glycan, which can be attached to a variety of proteins and lipids, is highly upregulated in most pancreatic cancers and released into the blood. The assay for CA 19‐9 is commonly used to confirm the diagnosis of suspected pancreatic cancer and to follow the course of the disease after the commencement of treatment. It is not used for early detection or diagnosis because a significant portion of pancreatic cancer patients – between 20% and 30% (Goonetilleke and Siriwardena 2007), depending on the cut‐off used – have weak or nonexistent elevations in the blood, and because elevations can sometimes occur in

318  Proteomics for Biological Discovery

nonmalignant conditions such as liver or pancreatic inflammation. Because of the lack of a highly accurate diagnostic method, rendering a definitive diagnosis of pancreatic cancer sometimes is extremely difficult, especially at early stages of the disease, resulting in delayed use of the most effective treatment course. The clinical assay for CA 19‐9 uses a standard sandwich format, in which an immobilized antibody captures the glycan, along with the proteins to which the glycan is attached, and a detection antibody probes the glycan on the captured proteins. The proteins that carry the CA 19‐9 antigen had previously not been well characterized but were known to include mucins (Akagi et  al. 2001) and carcinoembryonic antigen (Yue et al. 2009). In addition, it was not known whether the composition of the CA 19‐9 carrier proteins is different between disease states (Chen et  al. 2007; Hollingsworth and Swanson 2004); for example whether the ­proteins bearing the CA 19‐9 antigen in pancreatitis are different from those bearing the CA 19‐9 antigen in pancreatic cancer. We recently investigated the possibility that such differences could be exploited to provide improved ­biomarker accuracy. The key to testing this hypothesis is the ability to measure the CA 19‐9 antigen on individual protein carriers. The ALSA platform is well suited to this task (Haab 2010), since the levels of a particular glycan epitope can be measured on many different proteins in parallel (Figure 13.1 and Figure 13.3a) and compared across many different samples (Forrester et  al. 2007). Another requirement in testing this hypothesis is knowledge of what proteins carry the CA 19‐9 antigen. To obtain more information about the carriers of the CA 19‐9 antigen, we used the CA 19‐9 antibody to immunoprecipitate proteins from the sera of various types of patients and identified the captured proteins using MS (Yue et al. 2011), similar to a previous study on the carriers of the related sialyl Lewis X glycan (Cho et al. 2008). A variety of novel carrier proteins were found, including apolipoprotein A, apolipoprotein E, ARVCF, fibronectin, galectin‐3 binding protein, and others. The dominant carrier was apolipoprotein A, displaying CA 19‐9 in about 25% of healthy and diseased individuals. The next step was to use the ALSA method to determine whether the carriers were different between patient groups. Using arrays containing 58 different antibodies, we probed the CA 19‐9 levels on each of the captured proteins among a set of serum samples (Yue et al. 2011). The proteins that carried the CA 19‐9 antigen in strongest association with pancreatic disease were the mucins MUC1, MUC5AC, and MUC16, all of which are produced at elevated levels in pancreatic cancer. Other proteins carried the antigen, but not in association with disease. This finding suggests that the CA 19‐9 antigen has functions on normal serum proteins and that the elevations observed in pancreatic cancer are mainly due to proteins secreted directly from the pancreas, which explains the good specificity of the marker for pancreatic and gastrointestinal disease. The next question was whether detection of CA 19‐9 on the individual, cancer‐ associated carriers MUC1, MUC5AC, or MUC16 provided better biomarker performance than the conventional CA 19‐9 assay. Again, this question was conveniently tested using ALSA (Yue et al. 2011). By probing the CA 19‐9 levels on each of the three proteins in large sets of serum samples, we found that the dominant carriers were MUC5AC and MUC16. The detection of CA 19‐9 on these individual proteins could more sensitively detect certain patients with weak CA 19‐9 (using the

(a)

Measuring total CA19-9

Measuring CA19-9 on individual proteins

(b)

Detected by CA19-9 mAb

CA19-9 antigen

CA19-9 carrier proteins

mAbs against individual core proteins

CA19-9 mAb

CA19-9 on MUC16 (830)

Probing Glycoforms Using Antibody-Lectin Sandwich Arrays  319

5

Pancreatitis Pancreatic cancer

3

1

(c) {M1} Total CA19-9 {M2} CA19-9 on MUC5AC {M3} CA19-9 on MUC16

3 Total CA19-9

5

Cancer patient samples

Classification

TP, by CA19-9 M1 M2 M3 C

FN

TP, by Panel

Non-cancer patient samples

TN

FP

Figure 13.3  Improved patient discrimination by detecting CA 19‐9 on individual protein carriers. (a) The conventional CA 19‐9 assay works by capturing and detecting the CA 19‐9 glycan on all protein carriers (left). An alternate assay is to first capture the protein carrier using specific antibodies, and then probe for the CA 19‐9 antigen on the captured proteins (right), which may provide higher accuracy for disease if certain protein carriers predominate in disease. (b) Certain cancer patients that are low in the total CA 19‐9 assay (values in the horizontal axis) are elevated in an assay for CA 19‐9 on MUC16 (values in the vertical axis). Each point is a patient sample. The lines represent 75% specificity thresholds (25% of control samples above threshold) for each marker. The cancer samples in the upper left quadrant (red arrow) were low in CA 19‐9 but high in CA 19‐9 on MUC16. (c) Improved performance was achieved using combined CA 19‐9 and CA 19‐9 on individual protein carriers. Each column represents data from a patient sample, and each row is data from a marker. A yellow square indicates that the sample exceeded a threshold set for the marker, and a black square indicates the opposite. If a sample was elevated in any one of the three markers, it was classified as a cancer sample, and if it was low in all three, it was called a control sample. Some of the samples that were low by the conventional CA 19‐9 assay (top marker row), and thus would have been false negatives, were elevated in CA 19‐9 on MUC16 or MUC5AC. This result shows improved accuracy using the detection of CA 19‐9 on individual carriers, compared to the conventional CA 19‐9 assay. FN, false negative; FP, false positive; TN, true negative; TP, true positive. (See color insert.)

c­ onventional assay), but total performance was not better than the conventional assay. However, the assays were somewhat complementary, as some of the patients with weak CA 19‐9 could be picked up by CA 19‐9 on MUC5AC or CA 19‐9 on MUC16, and vice versa (Figure 13.3b). A three‐marker panel comprising the standard CA 19‐9 assay, MUC5AC‐CA19‐9 (indicating the detection of CA 19‐9 on MUC5AC), and MUC16‐CA19‐9 improved sensitivity from 63% (using standard CA 19‐9) to 74% at a specificity of 98% (Figure 13.3c). The result was confirmed in two additional ­sample sets. The improved sensitivity was due to the fact that MUC5AC and MUC16 were the dominant carriers of CA 19‐9 in certain cancer patients, allowing improved discrimination of those patients from the pancreatitis patients and healthy subjects, who less frequently have elevated levels of CA 19‐9 on those particular proteins. This result validates the approach and also points to future needs. For example, we need to identify additional cancer‐associated carriers of the CA 19‐9 antigen. A group of

320  Proteomics for Biological Discovery

patients was identified that were elevated in CA 19‐9 but not on the individual proteins tested, indicating that some other carrier was responsible for the elevation. Identification of that carrier might allow improved detection of additional patients, as was enabled through the detection of CA 19‐9 on MUC5AC and MUC16. Another need is to identify other cancer‐associated glycans, since some patients displayed extremely low CA 19‐9. The combination of finding additional cancer‐ associated glycoproteins and identifying additional cancer‐associated glycans promises to further improve the accuracy of cancer diagnostics. The ALSA platform, in conjunction with other glycoproteomics approaches, is a key to identifying and testing novel candidate markers. 13.4.2  Other Glycans in Pancreatic Ductal Adenocarcinoma The work described above was to characterize the protein carriers of a particular glycan, the CA 19‐9 antigen. A complementary use of ALSA is to characterize the glycans on a particular protein. This goal can be pursued by using each array multiple times, each time with a different lectin or glycan‐binding antibody, to profile the glycans attached on the captured proteins. We used this approach to gain more information about cancer‐associated glycans on mucin and CEACAM‐related proteins (Yue et  al. 2009). Antibody arrays targeting 17 different proteins were incubated repeatedly with a set of serum samples from pancreatic cancer patients (n = 23) and control subjects (n = 23), and each set was probed with a different lectin (28 in all). In addition, we determined the core protein levels of the three mucins (MUC1, MUC5AC, and MUC16) using standard sandwich assays. A comparison of the core protein levels to the glycan levels allows one to evaluate whether the glycan is altered in relation to the protein level, or whether the protein abundance is altered without a corresponding change in glycosylation. All three mucins showed cancer‐associated elevations, but with notable differences (Figure  13.4). MUC16 was most frequently elevated at the protein level (~65% of cancer patients) but rarely showed glycan alterations relative to the protein level, with the exception of slightly increased fucosylation in some patients. In contrast, MUC1 and MUC5AC showed protein elevations in fewer patients (30% and 35% of cancer patients, respectively) but displayed alterations in a variety of glycan motifs in most patients (up to 65%). Based on knowledge of the specificities of the lectins, we could infer that terminal galactose‐N‐acetylgalactosamine and fucosylation were likely common alterations on MUC5AC, whereas MUC1 was more likely to just display increased lactosamine with terminal sialyl‐Lewis A (the CA 19‐9 antigen). These results show that cancer‐associated glycans can be different between proteins; the CA 19‐9 antigen was a common motif between the proteins, but increased terminal galactose‐N‐acetylgalactosamine mainly occurred on MUC5AC. In addition, these results indicate that the detection of individual glycoforms of proteins can provide better biomarker performance than the detection of the core protein. This improvement is because some patients did not display an elevation in the core protein (at least with the sandwich assay used here) but showed significant elevations in particular glycans. Further characterization of the cancer‐associated glycans and their carrier proteins, combined with the ability to optimally detect those glycoforms, is a goal for future research.

Terminal Mannose

Terminal and poly-GlcNAc

Terminal GalNAc

Gal-GlcNAc

Gal-GalNAc

AAL UEA CA19-9 BPL Jacalin PNA ECL PHAL RCA 120 DBA GSL 1 SBA Sialyl Tn VVL SJA GSL II WGA LEL STL ConA LCA

Fucose

Probing Glycoforms Using Antibody-Lectin Sandwich Arrays  321

AUC 1.0

0 MUC1 Glycan MUC1 Glycan:Protein Ratio MUC5AC Glycan MUC5AC Glycan:Protein Ratio MUC16 Glycan MUC16 Glycan:Protein Ratio

Figure 13.4  Cancer‐associated changes in protein abundance and glycosylation. The protein abundances in multiple glycan levels were measured for MUC1, MUC5AC, and MUC16 captured out of sera from pancreatic cancer patients and control subjects. Each square gives a summary statistic of the difference between the cancer and control samples, as indicated by the color bar, where yellow indicates a greater difference. (The statistic is the area under the curve (AUC) in receiver‐operator characteristic analysis.) The difference was calculated using either the glycan measurements or the glycan:protein ratio for each protein, which was obtained by dividing the lectin measurement by the protein abundance measurement. The detection lectins and antibodies are organized according to primary specificities, although some specificities are more complex or diverse than indicated here. A yellow square in the glycan:protein ratio indicates that the glycan was elevated in the cancer samples to a greater degree than the protein abundance. Source: Yue et al. (2009). (See color insert.)

13.4.3  Differences Between Cystic Neoplasms and Solid Tumors In addition to diagnosing established or incipient ductal adenocarcinoma (the focus of the work described above), a need exists for improved diagnostics of a type of precursor to pancreatic cancer, cystic neoplasms of the pancreas. Cystic lesions in the pancreas are being identified more frequently due to the increased and improved use of abdominal imaging. Some cystic lesions have the potential to progress to invasive cancer and thus should be removed, while benign cysts should not be removed. Unfortunately, the determination of the malignant potential of pancreatic cysts is not highly accurate using current diagnostic methods. The fluid from inside the cysts, which can be removed by endoscopic ultrasound‐guided, fine‐needle aspiration (EUS‐FNA), is an attractive source for biomarkers, since it is confined and concentrated in direct contact with the abnormal cells. Using the methods described for the study of serum proteins in adenocarcinoma, we profiled mucin and CEACAM glycosylation in cyst fluid samples obtained from benign and neoplastic cysts (Haab et al. 2010). The alterations observed in this study were clearly distinct from those observed for adenocarcinoma. CA 19‐9 was elevated primarily in a subset of neoplastic cysts called mucinous cystic neoplasms (MCNs) but was not frequently elevated in the more common intraductal pancreatic mucinous neoplasms (IPMNs). MUC16 and MUC1 were not good markers for neoplasm,

322  Proteomics for Biological Discovery TABLE 13.1  Comparison between adenocarcinomas and cystic neoplasms in mucin protein abundance and glycosylation. Adenocarcinoma Protein abundance

Protein glycosylation

MUC1

Increased in 30% of cancer patients

MUC5AC

Increased in 35% of cancer patients Increased in 65% of cancer patients

Increased terminal lactosamine and fucose, increased CA 19‐9 Highly increased CA 19‐9 and terminal Gal‐GalNAc Few glycan changes

MUC16

Cystic neoplasms Protein abundance Up in certain benign cysts

Protein glycosylation

Up in most neoplastic cysts Unchanged

Increased terminal Gal, GlcNAc, and GalNAc; no increased CA 19‐9 Unchanged

Unchanged

Gal, galactose; GalNAc, N‐acetylgalactosamine; GlcNAc, N‐acetylglucosamine.

but specific glycoforms of MUC5AC were highly and frequently elevated in neoplastic cysts. The lectin binding on MUC5AC indicated that terminal lactosamine, terminal N‐acetylgalactosamine, and terminal N‐acetylglucosamine were upregulated but that the CA 19‐9 antigen was not prominent. The CEA protein was elevated in most neoplastic cysts and almost always displayed the CA 19‐9 antigen, but it did not display other glycan alterations. These results revealed differences between adenocarcinoma and cystic neoplasms (summarized in Table 13.1). In contrast to adenocarcinoma, cystic neoplasms do not show upregulation of the CA 19‐9 antigen or the MUC1 and MUC16 proteins. Elevations in MUC5AC are present in both cystic neoplasms and adenocarcinoma, although the glycan features are different. Distinctive features of MUC5AC glycosylation in neoplastic cysts are the lack of CA 19‐9 antigen and the presence of nonsialylated terminal structures involving galactose, N‐acetylgalactosamine, and terminal N‐acetylglucosamine, whereas MUC5AC glycosylation in adenocarcinoma is dominated by the CA 19‐9 antigen. Thus, both protein expression and glycosylation are different between the two types of pancreatic lesions. This result may not be surprising given the distinct genetic and cellular alterations characterizing the lesions (Wu et  al. 2011; Iacobuzio‐ Donahue et al. 2012). Based on these results, we may expect to see unique types of protein and glycan alterations based on the genetic alterations and cellular differentiations states present in tumors. In both adenocarcinoma and cystic neoplasms, the detection of specific protein glycoforms outperformed the detection of core proteins for the discrimination of benign from malignant or premalignant lesions, supporting the concept of the commonality of glycan alterations in cancers and showing the importance of this approach for cancer diagnostics. 13.5  FUTURE CHALLENGES 13.5.1  Development of New Glycan‐Binding Reagents These methods need more development to be effective and useful in future research. One challenge is the continued development of high‐affinity reagents with a broad range of specificities. Relatively few lectins are commonly used in biological research,

Probing Glycoforms Using Antibody-Lectin Sandwich Arrays  323

due in part to the poor analytical qualities of many lectins, such as weak affinities or a tendency to interact nonspecificly. The result is a greatly limited spectrum of glycan structures that may be probed. This problem could be addressed in two ways. One is by further tapping into and developing the rich resource of naturally occurring lectins, assisted by the power of the glycan microarray for characterizing and screening lectins. New lectins with unusual binding properties are discovered continually from all types of organisms. Previously, it was difficult to systematically determine binding specificities in a way that allowed detailed comparisons between lectins, but with the glycan array and new bioinformatics approaches, such tasks are practical. We now can routinely determine which lectins are most specific for particular structures and which are likely to be the strongest, most robust binders. These analyses also can incorporate other types of glycan binders, including antibodies and polypeptides derived from display technologies, such as glycan‐binding proteins developed from libraries of fish antibodies (Hong et al. 2013). It is likely that extremely stable and strong binders, with unusual and narrow specificities, are abundantly available and can be identified through searches of the regularly increasing storehouse of glycan array data. Another way to increase the repertoire of useful lectins is to improve the properties of existing lectins. Site‐directed mutagenesis followed by selection of clones with desired properties can result in improved performance, as demonstrated with the lectin Aleuria aurantia (Romano et al. 2011). Alternatively, multiple lectins could be linked together to increase avidity through multivalent interactions. Lectins commonly have weak affinities individually but in the biological setting employ multivalent interactions to effect their functions (Collins and Paulson 2004). In vitro approaches for generating lectin multimers include using a bivalent antibody against a tag (such as the 6‐His fusion tag) to link two tagged proteins (Stevens et al. 2006), using the quaternary nature of streptavidin to link up to four biotinylated proteins (Kawasaki et al. 2007, 2008), and generating a polymeric backbone to which multiple proteins may be linked (Rickert et  al. 2007; Griffith et  al. 2004; Tao et  al. 2009). Lectins known to have desirable specificities but weak affinities may become more valuable through these approaches, although the effort required for individual testing and optimization will be significant. 13.5.2  Protein–Protein Complexes The analysis of proteins captured from biological solutions is complicated by the existence of multiprotein complexes. Because multiprotein complexes could be captured along with the targeted protein, as characterized previously (Bergsma et  al. 2010), lectins could bind the glycans on the associated proteins rather than on the targeted protein. Therefore, the interpretation of observed changes is not clear. This situation could be addressed by additional experiments. For example, immunoprecipitation of the targeted protein followed by gel electrophoresis and blotting for reactivity to the lectin could reveal whether the targeted protein or some other protein was the source of lectin binding. The studies of mucins described above were supplemented with such experiments, which confirmed that the mucins were the carriers of the glycans being studied (Yue et al. 2009; Haab et al. 2010). More detailed studies of individual protein glycosylation could be pursued using MS and other methods, as long as enough purified protein could be obtained from the biological samples.

324  Proteomics for Biological Discovery

This problem also could be addressed through modifications to the protocol. For example, samples could be denatured, reduced, and alkylated, or moderately digested prior to analysis on antibody arrays, which would reduce the protein–protein complexes captured by each antibody. This approach could simplify the interpretation of lectin binding. However, the sample treatment potentially could lead to increased nonspecific binding of certain proteins and weakening of the antibody affinities. Nevertheless, individual optimization of the assays should lead to conditions that maximize the specific binding of individual protein targets. 13.5.3  Analytical Properties Required for Clinical Assays To translate findings from the basic studies of protein glycosylation to biomarkers used in patient care, assays are needed that meet the rigorous performance requirements for clinical use. Features that are especially critical are reproducibility, specificity, and freedom from interference. The reproducibility must be good at both the individual assay level (tight coefficients of variation in repeats of samples) and over repeated productions of assays. The specificity and freedom from interference, for example caused by immunoglobulins that bind the assay reagents, must be demonstrated over many patient samples. No examples of lectins used in established clinical assays  currently exist, perhaps because of difficulties in producing consistent material, poor stability, frequent interference from factors in the biological matrix, or insufficient affinity. It may be necessary to modify and optimize the sequences of some lectins, or use monoclonal antibodies instead. If the binding specificity of a lectin is well characterized, the glycan structure targeted by the lectin could be used to develop an antibody with the same binding specificity. Since the performance and use of monoclonal antibodies in clinical assays are well studied, the conversion from lectin to monoclonal antibody should make the development of clinical assays straightforward. 13.6 CONCLUSION Glycan analysis methods employing lectins and antibodies are complementary to MS and chromatographic methods and can be well suited to biomarker research. Assay formats such as the antibody‐lectin sandwich array that enable measurements of glycosylation on multiple, specific proteins can be particularly informative. Our ability to characterize the binding specificities of lectins has been greatly enhanced by glycan array technology, and the automated analysis of glycan array datasets is enabling rapid searching for lectins with defined specificities. The application of these approaches to pancreatic cancer has helped to define the carriers of the CA 19‐9 antigen, characterize the glycans on cancer‐associated mucin proteins in both adenocarcinomas and cystic lesions, and develop candidate biomarkers for improved diagnostics. Whereas a spectrum of proteins carry the CA 19‐9 antigen in the circulation of healthy and diseased individuals, only the cancer‐associated mucins MUC1, MUC5AC, and MUC16 show cancer‐associated elevations of the antigen among the proteins tested so far. The glycosylation and protein expression patterns differ greatly between these proteins, with MUC5AC showing the most extensive cancer‐ associated alterations in glycosylation and MUC16 showing the least. Cystic lesions

Probing Glycoforms Using Antibody-Lectin Sandwich Arrays  325

are largely different from adenocarcinoma. Only MUC5AC shows cancer‐associated alterations in protein expression and glycosylation in cystic lesions, and the predominant glycan alteration is exposure of terminal galactose, N‐acetylglucosamine, and N‐acetylgalactosamine, in contrast to the frequent upregulation of the CA 19‐9 antigen observed in adenocarcinoma. These results show the importance of individual glycoforms in cancer studies and demonstrate a practical method for their study, particularly for biomarker research. The linking of data from these methods to more detailed glycan characterizations based on MS and chromatography will further clarify the nature of cancer‐associated glycans and provide additional leads on glycans that should be studied (Haab 2010; Zeng et al. 2010). Such studies will be fostered by the continued discovery of lectins with precise and novel glycan‐binding specificities and the further development of approaches for using and characterizing lectins.

References Adamczyk, B., Tharmalingam, T., and Rudd, P.M. (2012). Glycans as cancer biomarkers. Biochim. Biophys. Acta 1820: 1347–1353. Akagi, J., Takai, E., Tamori, Y. et  al. (2001). Ca19‐9 epitope a possible marker for muc‐1/y protein. Int. J. Oncol. 18: 1085–1091. Astronomo, R.D. and Burton, D.R. (2010). Carbohydrate vaccines: developing sweet solutions to sticky situations? Nat. Rev. Drug Discov. 9: 308–324. Bergsma, D., Buchweitz, J., Chen, S. et  al. (2010). Antibody‐array interaction mapping (aaim): a new method to detect protein complexes applied to the discovery and study of serum amyloid p interactions with kininogen in human plasma. Mol. Cell. Proteomics 9: 446–456. Blixt, O., Head, S., Mondala, T. et al. (2004). Printed covalent glycan array for ligand profiling of diverse glycan binding proteins. Proc. Natl. Acad. Sci. USA 101: 17033–17038. Chen, S. and Haab, B.B. (2009). Analysis of glycans on serum proteins using antibody microarrays. Methods Mol. Biol. 520: 39–58. Chen, S., LaRoche, T., Hamelinck, D. et al. (2007). Multiplexed analysis of glycan variation on native proteins captured by antibody microarrays. Nat. Methods 4: 437–444. Chen, W.C., Completo, G.C., Sigal, D.S. et al. (2010). In vivo targeting of b‐cell lymphoma with glycan ligands of cd22. Blood 115: 4778–4786. Cho, W., Jung, K., and Regnier, F. (2008). Use of glycan targeting antibodies to identify ­cancer‐associated glycoproteins in plasma of breast cancer patients. Anal. Chem. 80: 5286–5292. Collins, B.E. and Paulson, J.C. (2004). Cell surface biology mediated by low affinity multivalent protein‐glycan interactions. Curr. Opin. Chem. Biol. 8: 617–625. Culf, A.S., Cuperlovic‐Culf, M., and Ouellette, R.J. (2006). Carbohydrate microarrays: survey of fabrication techniques. OMICS 10: 289–310. Forrester, S., Kuick, R., Hung, K.E. et  al. (2007). Low‐volume, high‐throughput sandwich immunoassays for profiling plasma proteins in mice: identification of early‐stage systemic inflammation in a mouse model of intestinal cancer. Mol. Oncol. 1: 216–225. Freeze, H.H. (2006). Genetic defects in the human glycome. Nat. Rev. Genet. 7: 537–551. Fuster, M.M. and Esko, J.D. (2005). The sweet and sour of cancer: glycans as novel therapeutic targets. Nat. Rev. Cancer 5: 526–542.

326  Proteomics for Biological Discovery Goonetilleke, K.S. and Siriwardena, A.K. (2007). Systematic review of carbohydrate antigen (ca 19‐9) as a biochemical marker in the diagnosis of pancreatic cancer. Eur. J. Surg. Oncol. 33: 266–270. Griffith, B.R., Allen, B.L., Rapraeger, A.C., and Kiessling, L.L. (2004). A polymer scaffold for protein oligomerization. J. Am. Chem. Soc. 126: 1608–1609. Haab, B.B. (2010). Antibody‐lectin sandwich arrays for biomarker and glycobiology studies. Expert Rev. Proteomics 7: 9–11. Haab, B.B. and Yue, T. (2011). High‐throughput studies of protein glycoforms using antibody‐ lectin sandwich arrays. Methods Mol. Biol. 785: 223–236. Haab, B.B., Porter, A., Yue, T. et al. (2010). Glycosylation variants of mucins and CEACAMS as candidate biomarkers for the diagnosis of pancreatic cystic neoplasms. Ann. Surg. 251: 937–945. Hirabayashi, J., Arata, Y., and Kasai, K. (2000). Reinforcement of frontal affinity chromatography for effective analysis of lectin‐oligosaccharide interactions. J. Chromatogr. A 890: 261–271. Hirabayashi, J., Hashidate, T., Arata, Y. et al. (2002). Oligosaccharide specificity of galectins: a search by frontal affinity chromatography. Biochim. Biophys. Acta 1572: 232–254. Hollingsworth, M.A. and Swanson, B.J. (2004). Mucins in cancer: protection and control of the cell surface. Nat. Rev. Cancer 4: 45–60. Hong, X., Ma, M.Z., Gildersleeve, J.C. et al. (2013). Sugar‐binding proteins from fish: selection of high affinity “lambodies” that recognize biomedically relevant glycans. ACS Chem. Biol. 8: 152–160. Huang, R.‐P., Huang, R., Fan, Y., and Lin, Y. (2001). Simultaneous detection of multiple cytokines from conditioned media and patient’s sera by an antibody‐based protein array system. Anal. Biochem. 294: 55–62. Huang, R., Lin, Y., Shi, Q. et al. (2004). Enhanced protein profiling arrays with ELISA‐based amplification for high‐throughput molecular changes of tumor patients’ plasma. Clin. Cancer Res. 10: 598–609. Iacobuzio‐Donahue, C.A., Velculescu, V.E., Wolfgang, C.L., and Hruban, R.H. (2012). Genetic basis of pancreas cancer development and progression: insights from whole‐exome and whole‐genome sequencing. Clin. Cancer Res. 18: 4257–4265. Igl, W., Polasek, O., Gornik, O. et  al. (2011). Glycomics meets lipidomics‐associations of n‐­glycans with classical lipids, glycerophospholipids, and sphingolipids in three european populations. Mol. Biosyst. 7: 1852–1862. Kawasaki, N., Matsuo, I., Totani, K. et al. (2007). Detection of weak sugar binding activity of vip36 using vip36‐streptavidin complex and membrane‐based sugar chains. J. Biochem. 141: 221–229. Kawasaki, N., Ichikawa, Y., Matsuo, I. et  al. (2008). The sugar‐binding ability of ergic‐53 is enhanced by its interaction with mcfd2. Blood 111: 1972–1979. Ko, A.H., Hwang, J., Venook, A.P. et  al. (2005). Serum ca19‐9 response as a surrogate for ­clinical outcome in patients receiving fixed‐dose rate gemcitabine for advanced pancreatic cancer. Br. J. Cancer 93: 195–199. Kuno, A., Kato, Y., Matsuda, A. et  al. (2009). Focused differential glycan analysis with the platform antibody‐assisted lectin profiling for glycan‐related biomarker verification. Mol. Cell. Proteomics 8: 99–108. Lauc, G., Essafi, A., Huffman, J.E. et al. (2010). Genomics meets glycomics‐the first gwas study of human n‐glycome identifies hnf1alpha as a master regulator of plasma protein fucosylation. PLos Genet. 6: e1001256. Li, C., Simeone, D., Brenner, D. et al. (2009). Pancreatic cancer serum detection using a lectin/ glyco‐antibody array method. J. Proteome Res. 8: 483–492.

Probing Glycoforms Using Antibody-Lectin Sandwich Arrays  327

Li, D., Chiu, H., Chen, J. et al. (2013). Integrated analyses of proteins and their glycans in a magnetic bead‐based multiplex assay format. Clin. Chem. 59: 315–324. Marino, K., Bones, J., Kattla, J.J., and Rudd, P.M. (2010). A systematic approach to protein glycosylation analysis: a path through the maze. Nat. Chem. Biol. 6: 713–723. Marth, J.D. and Grewal, P.K. (2008). Mammalian glycosylation in immunity. Nat. Rev. Immunol. 8: 874–887. Maupin, K.A., Liden, D., and Haab, B.B. (2011). The fine specificity of mannose‐binding and galactose‐binding lectins revealed using outlier‐motif analysis of glycan array data. Glycobiology 22: 160–169. North, S.J., Hitchen, P.G., Haslam, S.M., and Dell, A. (2009). Mass spectrometry in the analysis of n‐linked and o‐linked glycans. Curr. Opin. Struct. Biol. 19: 498–506. Pedersen, J.W., Bennett, E.P., Schjoldager, K.T. et al. (2011). Lectin domains of polypeptide galnac transferases exhibit glycopeptide binding specificity. J. Biol. Chem. 286: 32684–32696. Perez, S. and Mulloy, B. (2005). Prospects for glycoinformatics. Curr. Opin. Struct. Biol. 15: 517–524. Pleskow, D.K., Berger, H.J., Gyves, J. et al. (1989). Evaluation of a serologic marker, ca19‐9, in the diagnosis of pancreatic cancer. Ann. Intern. Med. 110: 704–709. Porter, A., Yue, T., Heeringa, L. et al. (2010). A motif‐based analysis of glycan array data to determine the specificities of glycan‐binding proteins. Glycobiology 20: 369–380. Rickert, E.L., Trebley, J.P., Peterson, A.C. et  al. (2007). Synthesis and characterization of ­bioactive tamoxifen‐conjugated polymers. Biomacromolecules 8: 3608–3612. Rillahan, C.D. and Paulson, J.C. (2011). Glycan microarrays for decoding the glycome. Annu. Rev. Biochem. 80: 797–823. Ritts, R.E. Jr., Nagorney, D.M., Jacobsen, D.J. et al. (1994). Comparison of preoperative serum ca19‐9 levels with results of diagnostic imaging modalities in patients undergoing laparotomy for suspected pancreatic or gallbladder disease. Pancreas 9: 707–716. Romano, P.R., Mackay, A., Vong, M. et al. (2011). Development of recombinant aleuria aurantia lectins with altered binding specificities to fucosylated glycans. Biochem. Biophys. Res. Commun. 414: 84–89. Rudd, P.M., Elliott, T., Cresswell, P. et  al. (2001). Glycosylation and the immune system. Science 291: 2370–2376. Rudiger, H. and Gabius, H.J. (2001). Plant lectins: occurrence, biochemistry, functions and applications. Glycoconj. J. 18: 589–613. Stevens, J., Blixt, O., Paulson, J.C., and Wilson, I.A. (2006). Glycan microarray technologies: tools to survey host specificity of influenza viruses. Nat. Rev. Microbiol. 4: 857–864. Stevens, J., Blixt, O., Glaser, L. et al. (2006). Glycan microarray analysis of the hemagglutinins from modern and pandemic influenza viruses reveals different receptor specificities. J. Mol. Biol. 355: 1143–1155. Sun, Q., Ju, T., and Cummings, R.D. (2011). The transmembrane domain of the molecular chaperone cosmc directs its localization to the endoplasmic reticulum. J. Biol. Chem. 286: 11529–11542. Tao, L., Kaddis, C.S., Loo, R.R. et al. (2009). Synthesis of maleimide‐end functionalized star polymers and multimeric protein‐polymer conjugates. Macromolecules 42: 8028–8033. Wang, D. (2003). Carbohydrate microarrays. Proteomics 3: 2167–2175. Wu, J., Matthaei, H., Maitra, A. et al. (2011). Recurrent gnas mutations define an unexpected pathway for pancreatic cyst development. Sci. Transl. Med. 3: 92ra66. Yue, T. and Haab, B.B. (2009). Microarrays in glycoproteomics research. Clin. Lab. Med. 29: 15–29.

328  Proteomics for Biological Discovery Yue, T., Goldstein, I.J., Hollingsworth, M.A. et al. (2009). The prevalence and nature of glycan alterations on specific proteins in pancreatic cancer patients revealed using antibody‐lectin sandwich arrays. Mol. Cell. Proteomics 8: 1697–1707. Yue, T., Maupin, K.A., Fallon, B. et  al. (2011). Enhanced discrimination of malignant from benign pancreatic disease by measuring the ca 19‐9 antigen on specific protein carriers. PLoS One 6: e29180. Yue, T., Partyka, K., Maupin, K.A. et al. (2011). Identification of blood‐protein carriers of the ca 19‐9 antigen and characterization of prevalence in pancreatic diseases. Proteomics 11: 3665–3674. Zeng, Z., Hincapie, M., Haab, B.B. et al. (2010). The development of an integrated platform to identify breast cancer glycoproteome changes in human serum. J. Chromatogr. A 1217: 3307–3315.

Index Absolute quantification (AQUA), 76, 77 Activity‐based protein profiling, 245 Acute myeloid leukemia (AML), 239 Affinity purification MS (AP/MS), 199, 200, 202–204, 206, 207, 209, 210 Amyloid, 64, 100, 153, 155 Antibody, 30, 32–37, 39, 41, 43, 49, 51, 95, 102, 126, 130, 131, 136–138, 154, 199, 216, 220–226, 264–266, 270, 271, 275, 276, 307, 309–314, 316, 319, 320 Antibody‐lectin sandwich array (ALSA), 309, 313, 314, 316 Antibody microarray 33, 36, 39–41

Chemical labeling, 3, 6, 7, 9, 17–20, 74, 75, 133 Chemoproteomics, 244, 254, 256 Chromatography, 1, 4, 11, 20, 45, 54, 73, 75, 77, 100, 103, 129, 130, 137, 179–180, 190, 199, 200, 205, 244–246, 248, 251, 253, 254, 263, 269–272, 276, 288–295, 297 Collision‐induced dissociation (CID), 8, 9, 75, 147, 181, 268, 297 Cross‐linking, 50, 134–137, 140, 141, 146, 166, 167, 175–182, 184–192, 308–310, 321 Crystallography, 165, 166, 175, 185, 256 Cytometry by Time of Flight (CyTOF), 217–220

Biomarker, 1, 34, 40, 42, 47, 50, 54, 63–82, 89–93, 95–97, 99–101, 103–108, 244, 274, 308, 309, 313, 314, 317, 320, 321 BirA, 137, 138 BioID, 137, 138, 140

Data dependent, 86, 139, 269, 279 Diagnostic, 64–68, 70, 72, 73, 80, 81, 97, 108, 314, 316–318, 320 Desorption electrospray ionization (DESI), 90, 92, 98, 99. 108 Deuterium 4, 7, 146, 176, 188, 191 DNA array to protein array (DAPA), 48, 50 Drug affinity responsive target stability (DARTS), 252, 253

C, 2, 4–7, 9, 75, 76, 132, 133 Calmodulin binding peptide (CBP), 199 Cancer, 19, 33, 34, 36, 39, 40, 44, 46, 47, 50, 51, 63–65, 67, 68, 70, 72–74, 79, 81, 82, 97–99, 102, 156, 239, 254, 255, 274, 275, 281, 282, 307–309, 313–318, 320, 321 cDNA microarray, 67

13

Electron capture dissociation (ECD), 269 Electron transfer dissociation (ETD), 9, 269, 297, 298,

Proteomics for Biological Discovery, Second Edition. Edited by Timothy D. Veenstra and John R. Yates III. © 2019 John Wiley & Sons, Inc. Published 2019 by John Wiley & Sons, Inc.

329

330  Index Electrospray ionization (ESI), 72, 90, 99, 146, 153, 178, 199, 266, 288 Enzyme‐linked immunosorbent assay (ELISA), 29, 47, 54, 65–67, 79, 279, 282, 308 ESI‐MS, 149–152, 157, 160, 161, 163–165 Extracted ion chromatograms (XIC), 10, 11, 13 FLAG epitope tag, 130, 131, 199, 251 Flow cytometry, 215, 216, 220, 223, 226, 227, 238 Gel electrophoresis, 1, 54, 103, 126, 178, 199, 200, 263, 288, 299 Glutathione‐S‐transferase (GST), 32, 34, 49 Glycan, 39, 307–311, 313–321 Glycan microarray, 310, 319 Heat shock proteins (HSP), 155–159, 245, High energy collisional dissociation (HCD), 9, 10, 19, 297 HPLC, 4, 7, 8, 45, 103, 106, 129, 139, 180, 185, 186, 205, 254, 263, 288, 292 Hydrogen/deuterium exchange, 7, 146, 176, 188, 191 125 I, 2 ICAT, 6, 7, 12, 133, 136, 140 Immobilized metal affinity chromatography (IMAC), 271–273, 275, 276, 289–294, 296, 298–300 Immunohistochemistry, 54, 69, 72 Immunoprecipitation, 7, 29, 54, 126, 199, 251, 270, 271 Inductively coupled plasma mass spectrometry (ICP‐MS), 216, 221 Ion‐mobility, 92, 146, 148, 154, 155, 158, 161, 165–167 iTRAQ, 5, 8, 9, 11, 14, 16–18, 20, 75, 76, 133, 140, 276

Label free detection, 30, 34, 36, 95 LC‐MS/MS, 66, 135, 249, 251, 293, 295, 296, 300 Lectin, 39, 307–321 Limit of detection, 80, 280, 281 Limit of quantification, 80, 280, 281 Liquid chromatography, 1, 4, 45, 75, 77, 100, 103, 129, 130, 137, 180, 181, 199, 200, 245, 254, 269, 288, 294

MALDI, 66, 72, 90–95, 101, 103, 105–109, 181, 186, 199, 200, 210 MALDI‐MSI, 90, 94, 95, 103, 105, 107, 108 MALDI‐TOF, 37, 203, 270 MASCOT, 73, 139, 273, 296 Mass spectrometry, 1, 11, 29, 37, 67, 74, 89, 91, 126, 129, 130, 136, 145, 165, 176, 181, 190, 216, 246, 252, 261, 287, 301, 308, 314, 319 Mass spectrometry imaging (MSI), 89–92, 95, 99, 104 Matrix assisted laser desorption/ionization (MALDI), 37, 66, 72, 90–95, 101, 103, 105–108, 178, 181, 186, 199, 200, 203, 210, 274 Mechanism of action, 39, 40, 244, 255 Metabolic labeling, 2–4, 6, 17–19, 75, 132–134 Metal chelating polymers, 220–221, 223 MS/MS, 2, 9, 17, 18, 47, 72, 74, 75, 92, 105, 107, 129, 139, 180, 185, 190 MS3, 8, 10, 14, 18, 19, 265, 268–270, 273, 274, 293 MudPIT, 138, 139 Multiple spotting technique (MIST), 49 15 N, 2–6, 9, 11, 18, 75, 76, 132, 133, 159, 300 Native mass spectrometry, 151, 154, 161 N‐hydroxysuccinimide (NHS), 5, 7, 9, 314 Nuclear magnetic resonance (NMR), 2, 158, 165, 166, 175, 182, 185 Nucleic acid programmable protein microarray (NAPPA), 48–51

Open reading frame (ORF), 45, 49, 128, 129 Orbitrap, 9, 20, 72, 73, 92, 167, 190, 191, 266, 299, 300 P, 2, 263, 264 P, 2 Peptide abundance, 2, 6 Peptide microarray, 42–45, 52 Phosphoproteomics, 279, 287, 298–304 Phosphorylation, 39, 40, 46, 47, 51, 134, 239, 245, 261–267, 269–279, 281, 282, 287–289, 291–293, 296–305 Phosphoserine, 261, 262 Phosphopeptide, 266–274, 276, 280–282, 288–305 Phosphopeptide mapping, 266 Phosphothreonine, 262, 263 Phosphotyrosine, 33, 104, 262, 263, 276 Photochemical, 136 Plasma, 38–40, 65, 68, 69, 73, 77, 82, 216, 218 Plasma membrane, 223, 226, 275 32 33

Index  331

Post‐translational modification (PTM), 29, 37–39, 42, 43, 45, 46, 48, 49, 51, 70, 71, 73, 75, 77, 80, 128, 261–263, 265, 269 Principal component analysis (PCA) 96, 99, 102, 106 Prostate‐specific antigen (PSA), 65, 72 Protein abundance, 1, 2, 6, 10, 13, 70, 71, 132, 133, 309, 316–318 Protein Abundance Index, 13, 74 Protein identification, 2, 66, 67, 72, 73, 75, 77, 106, 129, 150, 246, 265 Protein in situ array, 48, 49 Protein microarray, 29–33, 36–38, 42, 45–51, 53, 54 Protein‐protein interactions, 2, 29, 31, 43, 46, 50, 51, 126, 128, 134, 136, 141, 197, 243, 255 Protein standard absolute quantification (PSAQ), 76, 77 Proteolysis, 2, 129, 153, 175, 176, 191 Quadrupole time‐of‐flight (Q‐TOF), 9, 72, 148, 172 Quantification concatemer standards (QconCAT), 76, 77 Rate zonal gradient, 136 Reverse‐phase (RP), 4, 139, 269, 288, 295, 296 Reverse‐phase microarray (RPMA), 37, 38 Ribosome, 48, 50, 150–152, 206 S, 2 SDS‐PAGE, 126, 178, 199, 200, 248, 252 Secondary ion mass spectrometry (SIMS), 90, 92, 95, 107 Selected reaction monitoring (SRM), 66, 67, 71, 77, 80, 253 Self‐assembling protein microarray, 42, 48, 49, 51, 52 SEQUEST, 73, 139, 273, 296 Serum, 33, 36–39, 44, 46, 47, 50, 53, 54, 68, 280, 311–314, 316, 317, 321

35

Shotgun sequencing, 14, 17, 136, 199, 205, 279, 280, 298 SILAC, 4–7, 11, 12, 17–20, 67, 75, 76, 133, 140, 250, 254 SILAM, 300 Single‐cell analysis, 223, 227, 238 Single chain variable fragment (scFv), 41, 49, 51 SPOT, 42 Stable isotope, 2–7, 10, 67, 76, 77, 132, 133, 137, 140, 207, 245, 250, 263, 272, 276, 281, 282, 300 Strong cation exchange (SCX), 73, 139, 179, 190, 293–295, 299 Surface plasmon resonance (SPR), 30, 36, 256, 309 Surface plasmon resonance imaging (SPRi), 36, 37 Tandem mass spectrometry (MS/MS), 2, 72, 129, 135, 137, 198, 199, 267 Tandem mass tags (TMT), 5, 8, 9, 11, 12, 14, 16–20, 75, 133 Time‐of‐flight (TOF), 92, 147, 160, 217, 219, 226, 266 Tissue, 6, 20, 30, 33, 36–38, 44, 54, 66, 68–70, 72, 74, 75, 78, 89–97, 99, 100, 102, 105, 106, 108, 138, 140, 190, 205, 226, 245–247, 251, 252, 280, 281, 293, 294, 299, 300 Tobacco etch virus (TEV), 130, 199, 200 Total ion count (TIC), 96 Triple quadrupole, 92, 160 Trypsin, 2, 6, 7, 73, 75–77, 81, 129, 133, 135, 139, 178, 190, 199, 200, 205, 247, 249, 251–253, 275, 289, Western blot, 35, 39, 54, 103, 126, 131, 279, 281, 282 Wnt, 254, 260 Yeast two hybrid (Y2H), 127–129, 134