Abstract

   Over the past decade, liquid chromatography coupled with tandem mass
   spectrometry (LC–MS/MS) has evolved into the main proteome discovery
   technology. Up to several thousand proteins can now be reliably
   identified from a sample and the relative abundance of the identified
   proteins can be determined across samples. However, the remeasurement
   of substantially similar proteomes, for example those generated by
   perturbation experiments in systems biology, at high reproducibility
   and throughput remains challenging. Here, we apply a directed MS
   strategy to detect and quantify sets of pre-determined peptides in
   tryptic digests of cells of the human pathogen Leptospira interrogans
   at 25 different states. We show that in a single LC–MS/MS experiment
   around 5000 peptides, covering 1680 L. interrogans proteins, can be
   consistently detected and their absolute expression levels estimated,
   revealing new insights about the proteome changes involved in
   pathogenic progression and antibiotic defense of L. interrogans. This
   is the first study that describes the absolute quantitative behavior of
   any proteome over multiple states, and represents the most
   comprehensive proteome abundance pattern comparison for any organism to
   date.

Introduction

   System-wide investigation of gene expression at the mRNA transcript
   level has become routine and is widely used in systems biology and
   clinical studies to identify sets of genes that show distinct
   transcript profiles for a specific cellular state and to classify
   samples according to their respective molecular patterns ([35]van ’t
   Veer et al, 2002; [36]Gilchrist et al, 2006; [37]Ishii et al, 2007). It
   has also been shown that neither the concentration of transcripts
   ([38]Gygi et al, 1999; [39]Griffin et al, 2002) nor their quantitative
   change in response to perturbations ([40]MacKay et al, 2004;
   [41]Kislinger et al, 2006; [42]de Godoy et al, 2008) strongly correlate
   with the quantitative change of their corresponding proteins, the main
   functional products of gene expression. Therefore, quantitative
   proteomics holds great promise to enhance or complement the picture of
   gene expression in cells, and thus to contribute to the understanding
   of most molecular mechanisms in a cell. However, owing to the large
   heterogeneity in the amount and the physico-chemical properties of
   proteins, along with the lack of protein amplification methods,
   system-wide quantitative proteome analysis has been more technically
   challenging than transcriptome analysis.

   Recent advances in liquid chromatography–tandem mass spectrometry
   (LC–MS/MS), currently the method of choice for large-scale protein
   studies, have made the reliable identification and quantification of
   thousands of proteins in a single study a reality ([43]Brunner et al,
   2007; [44]de Godoy et al, 2008; [45]Ahrens et al, 2010). However,
   particularly due to the selection of precursor ions using a simple
   intensity driven heuristics (data-dependent analysis, DDA), results
   from such studies still show a bias against the detection of low
   abundant protein species and a decreasing level of reproducibility of
   identified peptides with decreasing abundance. Comprehensive and more
   highly reproducible proteome coverage can be achieved by extensive
   sample pre-fractionation and the mass spectrometric analysis of each
   fraction, albeit at a cost that multiplies analysis time and limits
   throughput. Additionally, the detection of different proteome subsets
   in repetitive LC–MS analyses of similar samples impairs the generation
   of consistent, reproducible quantitative data sets across multiple
   samples, a crucial prerequisite in systems biology studies ([46]Ideker
   et al, 2001; [47]Rifai et al, 2006; [48]Schiess et al, 2009).

   Therefore, several alternative or complementary MS strategies have been
   developed to overcome some of the limitations of current LC–MS/MS
   workflows ([49]Schmidt et al, 2008; [50]Picotti et al, 2009; [51]Domon
   and Aebersold, 2010). They make use of a priori information gathered
   from previous MS studies to increase the reliability, reproducibility
   and/or throughput of subsequent measurements. Specifically, in each of
   these strategies, MS analysis is focused on a few proteotypic peptides
   (PTPs) per protein, thereby minimizing instrument time without
   compromising analytical sensitivity. Two specific implementations of
   such strategies have been proposed ([52]Pan et al, 2009; [53]Schmidt et
   al, 2009; [54]Domon and Aebersold, 2010), which we have termed targeted
   and directed MS, respectively. Targeted MS is based on selected
   reaction monitoring (SRM also known as multiple reaction monitoring)
   and is typically carried out on triple quadrupole mass spectrometers.
   Because of very high selectivity and sensitivity, it is capable of
   covering the full dynamic range of proteomes in moderately complex
   organisms such as yeast ([55]Picotti et al, 2009). However, since each
   LC–MS/MS run is limited to a few hundred targeted peptides
   ([56]Stahl-Zeng et al, 2007), the throughput required for proteome-wide
   measurements is currently difficult to achieve. Directed MS makes use
   of inclusion mass lists in order to guide the MS sequencing to a
   desired, pre-determined subset of peptides ([57]Jaffe et al, 2008;
   [58]Schmidt et al, 2008, [59]2009). Directed sequencing is carried out
   on the same types of instruments as discovery measurements by DDA. In
   contrast to the SRM methodology, directed MS monitors far larger sets
   of peptides per analysis. However, because the precursor ion signal of
   the peptide of interest has to be explicitly detected to trigger its
   identification, the overall dynamic range and sensitivity of directed
   sequencing is lower than that of SRM and more dependent on the sample
   matrix ([60]Domon and Aebersold, 2010).

   Here, we have studied global and time-resolved changes in the proteome
   of cells of the human pathogen Leptospira interrogans that were
   perturbed by antibiotic stress and serum stimulation. Overall, in 31
   samples, representing 25 cellular states, 1669 proteins, representing
   75% of the Leptospira proteome discovered by saturation sequencing
   using DDA MS, were consistently detected and their cellular
   concentrations determined ([61]Supplementary Table SV). This unique
   data set was generated via an integrated inclusion list driven MS
   strategy that maximizes protein coverage in individual samples by
   focusing precious MS-sequencing time on the best flying, PTPs of each
   protein ([62]Mallick and Kuster, 2010). The cellular concentrations of
   the detected proteins were estimated in each sample by correlating the
   average of the signal intensities of the three most highly responding
   peptides per protein with a calibration curve generated with a set of
   isotopically labeled reference ([63]Malmström et al, 2009). We show
   that the protein components of entire pathways can be quantified across
   several time points and, for the first time, large-scale, consistent
   proteome data sets can be subjected to cluster analysis, a tool that
   was previously limited to the transcript level due to incomplete
   sampling on protein level. We show that the proteomic changes measured
   differ from the available transcriptomics data. We demonstrate that
   Leptospira cells adjust the cellular abundance of a certain subset of
   proteins as a general response to stress while other parts of the
   proteome respond highly specific. They furthermore react to individual
   treatments by ‘fine tuning’ the abundance of certain proteins and
   pathways in order to cope with the specific cause of stress. Using
   serum treatment we simulated the host environment and elucidate which
   proteomic adjustments underlie virulence. The method can be implemented
   with standard high-resolution mass spectrometers and software tools
   that are readily available in the majority of proteomics laboratories.
   It is scalable to any proteome of low-to-medium complexity and can be
   extended to post-translational modifications or peptide-labeling
   strategies for quantification. We therefore expect the approach
   outlined here to become a cornerstone for microbial systems biology.

Results

   To consistently detect and absolutely quantify the same, extensive
   subset of the L. interrogans proteome in multiple samples, we developed
   and deployed the general workflow displayed in [64]Figure 1. It
   consists of two main phases, proteome discovery and scoring. During the
   initial discovery phase, a comprehensive atlas of peptides and proteins
   identified by LC–MS/MS was generated by saturation sequencing of the L.
   interrogans proteome. To maximize proteome coverage, a pooled sample
   was generated and analyzed that consisted of aliquots from cells at
   different states. Subsequently, during the scoring phase, selected PTPs
   were detected in individual samples via inclusion list driven
   sequencing and quantified based on the ion current of the selected
   peptides, to generate quantitative proteome maps for each cellular
   state. Using this technique, comprehensive LC–MS/MS maps could be
   generated without the need for sample and time-consuming
   pre-fractionation steps, which significantly increases sample
   throughput.

Figure 1.

   [65]Figure 1
   [66]Open in a new tab

   Global protein profiling workflow. In the first phase of the study
   (discovery phase), the peptide samples representing different cell
   states were mixed and analyzed by data-dependent acquisition (DDA)
   followed by directed 1D-LC–MS/MS. To achieve comprehensive proteome
   coverage, all detectable precursor ions, referred to as features, were
   extracted, sequenced in sequential directed LC–MS/MS analyses and
   identified by database searching. All identified peptide sequences were
   stored in a 1D-PeptideAtlas together with their precursor ion signal
   intensity, elution times and mass-to-charge ratio. For each protein,
   mass and time coordinates from the five most suitable peptides (PTPs)
   for quantification were extracted from the PeptideAtlas and stored in
   an inclusion list. Additionally, a spectral library was generated from
   the identified spectra to improve both, the sensitivity and speed of
   spectral matching in the quantification phase. In this phase (scoring
   phase), LC–MS/MS analysis was focused on the pre-selected PTPs as well
   as a set of heavy labeled reference peptides that were added to each
   sample. This determined the concentrations of the corresponding
   proteins in the sample, which could be used as anchor points to
   translate the MS response of each identified protein into its
   concentration ([67]Malmström et al, 2009). After spectral matching,
   label-free quantification was employed to extract and align identified
   features and monitor their corresponding protein abundances redundantly
   over all samples.

Generation of a L. interrogans PeptideAtlas

   To build a PeptideAtlas ([68]Desiere et al, 2006; [69]Deutsch et al,
   2008) with maximal coverage of the L. interrogans proteome, we
   generated a pooled sample in which aliquots of extracts from different
   cell states were combined. Specifically, one aliquot of an untreated
   control sample and four aliquots of the individual perturbated cells
   (24 h treatments only, see [70]Figure 3) were pooled. We used a single
   dimension high-performance LC–MS/MS platform in combination with the
   recently introduced directed MS technique ([71]Schmidt et al, 2008) to
   maximize proteome coverage. In such measurements, precursor ion
   chromatograms are first extracted from two initial data-dependent (DDA)
   LC–MS/MS runs and the precursor ion maps (retention time versus mass
   over charge) that are also generated by these measurements are
   subjected to a peak extraction algorithm ([72]Mueller et al, 2007) to
   detect precursor ions not identified by DDA MS. In subsequent
   injections of the same sample, the mass spectrometer was then directed
   to acquire product ion spectra of previously non-selected precursor
   ions, to incrementally increase proteome coverage to saturation. We
   have shown earlier that this procedure maximizes the coverage of
   moderately complex proteomes at the peptide level while minimizing
   measurement and computational time ([73]Schmidt et al, 2008).

Figure 3.

   [74]Figure 3
   [75]Open in a new tab

   Hierarchical clustering of protein concentration changes. Hierarchical
   clustering of absolute protein abundance changes to the corresponding
   untreated control samples in copies/cell (log[10]) for all 24
   treatments. The column dendrogram representing the clustering of the
   differentially perturbed samples is displayed and the clusters (1–6)
   obtained are indicated. Significantly enriched (P<0.05) biological
   processes (BP) based on GO are indicated for all eight protein clusters
   obtained (a–h).

   Specifically, the following sequence of analyses was carried out to
   collect the data for the L. interrogans PeptideAtlas. LC–MS/MS runs #1
   and #2 were conventional DDA runs where precursor ions of different
   charge states (2 and >2, respectively) were selected. In subsequent
   LC–MS/MS runs #3–#20, precursor ions selected by the following criteria
   were added to inclusion lists and identified by directed precursor ion
   selection: (i) all features detected by a feature detection algorithm
   ([76]Mueller et al, 2007) in the initial DDA runs; (ii) precursor ions
   corresponding to all PTPs extracted from a recently published
   large-scale proteome analysis on the same species ([77]Beck et al,
   2009); and (iii) predicted precursor ion signals for all PTPs that were
   computed but not observed from the L. interrogans genomic sequence. PTP
   predictions were carried out by the algorithm PeptideSieve ([78]Mallick
   et al, 2007). The L. interrogans proteome is highly accessible for the
   LC–MS analysis employed here since for the majority of gene products
   (3402/3658) five or more PTPs could be predicted ([79]Supplementary
   Figure S1). The fragment ion spectra generated from each of these
   analyses were database searched and the resulting data were filtered to
   a peptide and protein level false discovery rate (FDR) of 1%
   ([80]Reiter et al, 2009). At each stage, already identified features as
   well as proteins identified with more than five PTPs were excluded from
   further analysis in the subsequent stages.

   In the two initial DDA LC–MS/MS runs, we detected 37 833 unique
   features of which 7776 could be assigned to a peptide sequence,
   resulting in 6861 peptide identifications corresponding to 1223
   proteins ([81]Table I). The remaining features (27 968) for which no
   MS/MS spectra were acquired were split into four inclusion lists, each
   comprising around 7000 features. These were then specifically sequenced
   by directed LC–MS/MS analyses. Thereby, the PeptideAtlas could be
   extended by 2356 (228) additional peptides (proteins). Finally, 12 and
   10 additional directed LC–MS/MS-sequencing runs for the identification
   of missing proteins using PTPs from a recently published PeptideAtlas
   or predicted PTPs, respectively, increased the overall number of
   identifications to a total of 13 113 features, corresponding to 11 611
   peptides and 1680 proteins. To reach this coverage, 28 LC–MS/MS runs
   were required ([82]Table I). As is evident from [83]Figure 2A, the
   number of protein identifications reaches saturation toward completion
   of each experimental phase, after rising at the beginning of the phase,
   indicating that different peptide subsets are identified at each of the
   analytical stages. The final feature map generated in this discovery
   phase contains the exact mass and time coordinates of each identified
   feature and represents a rich resource for the directed sequencing of
   all detected proteins in the scoring phase. Importantly, the identified
   features are well distributed by time and mass ([84]Figure 2C), which
   allowed their specific sequencing in a high number of samples by
   directed LC–MS/MS.

Table 1. Number of unique features and peptides identified in the discovery
phase.

   Data filtering Number of entries for all 1680 proteins identified in
   the discovery phase Number of entries for GroEL
   ^aDetected using the SuperHirn algorithm ([85]Mueller et al, 2007).
   ^bIdentified by database searching. FDR was set to 1%.
   ^cPTPs are defined as features that assigned peptide sequences show
   full tryptic cleavage, contain no modification and only match to one
   protein sequence in the database used.
   ^dUp to the five most intense PTPs per protein were selected for
   screening phase.
   Detected unique features^a 37 833 Not available
   Identified unique features^b 13 113 86
   Identified unique peptides^b 11 611 76
   Identified unique PTPs^b,c 6889 23
   Selected unique PTPs^d 4953 5
   [86]Open in a new tab

Figure 2.

   [87]Figure 2
   [88]Open in a new tab

   Directed LC–MS/MS analysis of the L. interrogans proteome. A pool of
   peptide samples generated from different perturbations was LC–MS
   analyzed to generate a comprehensive protein/peptide atlas of L.
   interrogans. (A) This was achieved by accumulating the MS data obtained
   from (i) two non-directed (DDA) LC–MS runs followed by (ii) directed
   LC–MS analysis of all detected features, (iii) previously detected PTPs
   ([89]Beck et al, 2009) and (iv) predicted PTPs ([90]Mallick et al,
   2007). Proteins detected with five or more PTPs were excluded in the
   following analysis. The numbers of identified proteins (y axis) and
   peptides (inset) versus the number of identified tandem mass spectra
   are displayed. For comparison, the protein discoveries obtained from a
   non-directed LC–MS analysis of 24 OGE fractions (OGE/LC) are shown in
   red. A recently developed algorithm was deployed to estimate the
   increase in protein/peptide discoveries with additional LC–MS
   experiments (dashed lines) ([91]Claassen et al, 2009). (B) Venn diagram
   showing the overlap of proteins identified with the LC-only and the
   OGE/LC–MS approach. (C) LC–MS map of all identified features (green).
   Precursor ions identified as tryptic peptides of the GroEL protein are
   shown in red. The sequences as well as the coordinates of the five PTPs
   selected for GroEL monitoring in the scoring phase are indicated in
   blue.

   We next evaluated the extent of proteome coverage achieved by this
   iterative directed sequencing strategy with that achieved by more
   conventional proteome analyses via extensive sample fractionation and
   DDA analysis of each fraction. For the latter strategy, the same
   peptide sample used for inclusion list sequencing was fractionated by
   isoelectric focusing using off-gel electrophoresis (OGE) ([92]Heller et
   al, 2005) and each of the 24 fractions was analyzed once by DDA
   LC–MS/MS analysis. Intriguingly, this data set contained 60% more
   peptide identifications, but only 19% additional protein hits (number
   versus number, [93]Figure 2A), indicating a higher peptide per protein
   ratio of 12 (OGE) over 7 (LC only). We thus conclude that 81% of the
   proteins detected by the OGE–LC–MS/MS approach were also detected by
   the directed LC–MS/MS method, most of them with a sufficient number of
   peptides for accurate quantification in the scoring phase. Notably,
   only a slight increase in protein identifications is expected by
   additional LC–MS/MS analyses ([94]Claassen et al, 2009), demonstrating
   that we have detected most of the proteins identifiable by the two
   LC–MS/MS strategies employed ([95]Figure 2A, dashed lines). As
   expected, the majority of proteins (67.9%) were identified with both
   approaches. However, 23.3/8.9% of identified peptides were exclusively
   detected by the OGE–LC/LC-only approach, respectively ([96]Figure 2B).
   Functional annotation revealed that many of the 194 protein hits
   exclusively identified by the directed (LC only) LC–MS/MS approach and
   missed by the OGE–LC–MS/MS approach are membrane proteins
   ([97]Supplementary Figure S2), suggesting a decreased recovery of
   hydrophobic peptides after OGE. Conversely, the OGE–LC–MS/MS strategy
   showed an increased coverage, particularly of low abundant proteins,
   like transcription factors and regulators, confirming the higher
   protein concentration range accessible after extensive sample
   fractionation. In general, extensive proteome coverage was achieved
   with both strategies, which is supported by the lack of biases against
   any functional groups ([98]Supplementary Figure S2).

   Overall, of the 13 113 different features identified by directed
   LC–MS/MS ([99]Supplementary Table SII), 6889 represented suitable PTPs
   for protein quantification ([100]Supplementary Table SIII). For each
   protein, the five most suitable PTPs for protein quantification,
   referred to as top five PTPs, were extracted from the feature list
   considering the following attributes; (i) specificity to a single
   database entry, (ii) true tryptic cleavage termini, (iii) lack of
   modifications and (iv) high MS-signal response determined by the
   SuperHirn algorithm ([101]Mueller et al, 2007). The selected 4953 PTPs
   ([102]Table I) covered the whole feature intensity range
   ([103]Supplementary Figure S3) and all 1680 identified proteins
   ([104]Table I). The feature intensity range for the PTP precursor ions
   on the inclusion list spanned more than three orders of magnitude, a
   dynamic range that is expected to capture most of the L. interrogans
   proteome ([105]Malmström et al, 2009). The benefits of focusing on the
   most suitable PTPs for monitoring each protein can be demonstrated in
   the case of the chaperone GroEL. For this abundant protein, 86
   different features could be identified ([106]Table I) of which the five
   most intense fulfill all PTP selection criteria ([107]Figure 2C, blue),
   supporting the observation that unspecifically proteolyzed or modified
   peptides constitute a minor but detectable fraction of the total ion
   current generated by the peptides from a protein ([108]Picotti et al,
   2007). By focusing on these PTPs, >90% of the MS-sequencing cycles
   required to detect and monitor GroEL levels in the following scoring
   phase could be saved and thus used for measuring different proteins of
   interest. It is important to note that this effect is more pronounced
   for highly abundant and larger proteins for which high numbers of
   peptides are identified.

   Finally, 38 heavy labeled reference peptides from 19 proteins were
   added to estimate absolute protein concentration on a system-wide scale
   in each sample following a recently described protocol ([109]Malmström
   et al, 2009) ([110]Figure 1; [111]Supplementary Table SI). Thus, the
   final inclusion mass list was distributed over two LC–MS/MS runs and
   the coordinates of the heavy reference peptides and their endogenous
   counterparts were included in both runs. Therefore, the data generated
   in the discovery phase of the project allowed us to establish a method
   in which 1680 proteins per sample could be detected and absolutely
   quantified in two inclusion list LC–MS/MS runs with a total analysis
   time per sample of 4 h.

   To increase the speed and identification yield of the selected PTPs in
   the scoring phase, we computed a spectral library from the acquired
   MS-sequencing data in the discovery phase using SpectraST ([112]Lam et
   al, 2009). We included additional MS data from a recent large-scale
   LC–MS/MS study on the same species ([113]Beck et al, 2009) to further
   enhance the quality of the consensus spectra in the spectral library
   and applied very stringent filtering criteria to keep the overall FDR
   <0.2%. Overall, 321 498 identified MS2 spectra were merged to 33 766
   distinct consensus spectra covering >2300 proteins. The library was
   added to the current L. interrogans PeptideAtlas and can be downloaded
   from [114]http://www.peptideatlas.org.

   Next, we assessed the performance of the described approach by
   analyzing a single control sample and comparing the number of
   identified peptides/proteins to the conventional shotgun LC–MS/MS
   methodology using the same number of runs. While the non-directed DDA
   LC–MS/MS analysis ([115]Supplementary Figure S4A, blue) identified a
   larger number of peptides, 404 (40%) additional proteins could be
   detected by the directed strategy (1593) ([116]Supplementary Figure
   S4A, red). The coverage was particularly enhanced for proteins of
   mid-to-low abundance, indicating an increased identification efficiency
   for these proteins by the directed MS approach compared with DDA
   LC–MS/MS-based strategies ([117]Supplementary Figure S4B).

   Finally, we assessed the utility of the generated inclusion
   list/spectral library on a different LC–MS platform in a different
   proteomics laboratory. After adjusting the retention times of the PTPs
   to the new LC system, the identified proteins could be detected with
   the same high consistency ([118]Supplementary Figure S5A and B) and
   coverage ([119]Supplementary Figure S5C) as on the LC–MS platform that
   was used to build the inclusion list and spectral library. This
   demonstrates the value of the generated data for the application in
   other laboratories and the usefulness of the generated, global
   PeptideAtlas and inclusion mass list for the proteomics community.

Quantitative time course measurements of perturbed L. interrogans cells

   We next used the method established above to acquire quantitative
   proteome profiles of Leptospira cells grown under different conditions.
   Specifically, cells were cultured in EMJH supplement (control samples)
   and in the presence of fetal bovine serum (FBS; 10% v/v) and
   antibiotics (5 μg/ml ciprofloxacin, 10 μg/ml penicillin G, 15 μg/ml
   doxycycline, respectively) in EMJH supplement. The underlying molecular
   mechanisms of the individual treatments are displayed in [120]Figure 3.
   Samples were taken after 3, 6, 12, 24, 48 and 168 h of treatment. Thus,
   overall 31 protein samples were generated, including 7 controls. We
   used label-free quantification to generate proteome maps of all
   detected PTPs and employed them for absolute protein quantification
   within each sample as well as relative protein quantification across
   all samples. Two technical replicates were acquired and averaged for
   all samples, to improve quantification accuracy.

   We first evaluated the combined technical and biological
   reproducibility of the relative protein quantification by comparing the
   proteome maps of three different control samples ([121]Supplementary
   Figure S6). The high squared Pearson correlation R^2 (0.945–0.965) and
   the near straight lines indicated the nearly optimal linear
   relationship between the replicates. Specifically, minimal abundance
   variations between the replicate samples were observed by the inclusion
   list driven LC–MS/label-free quantification approach even for proteins
   of low abundance ([122]Supplementary Figure S6A–C). Consequently, with
   the measured coefficient of variances of the protein ratios being <26%
   between all controls, 1.5-fold changes (2 × σ) with a P-value <0.05
   (ANOVA) can be confidently detected for most proteins by the described
   approach ([123]Supplementary Figure S6D–F).

   We next used the proteome maps to estimate the absolute quantities of
   the proteins in each perturbed sample and thus, in conjunction with the
   number of cells used to generate the samples, the cellular
   concentrations of the detected proteins. This was accomplished by
   translating the signal intensities of the high responder peptides from
   each detected protein into absolute protein quantities, using a
   recently published approach with some modifications ([124]Malmström et
   al, 2009). First, the absolute protein quantity of a consistent set of
   proteins was accurately determined in each sample by comparing the
   signal intensities of the sample intrinsic peptides with the
   corresponding signals generated from known amounts by isotopically
   labeled reference peptides of identical sequence that were added to
   each sample. Since these peptides were included in the directed LC–MS
   analysis, no additional SRM LC–MS analyses were required for their
   quantification. In this way, the precise concentrations of 29 peptides
   corresponding to 19 proteins could be calculated ([125]Supplementary
   Table SI). The concentrations of these proteins spanned almost three
   orders of magnitude, from 68 copies/cell for the flagellar M-ring
   protein ([126]YP_001355.1) to 13 649 copies/cell for the GroEL protein
   ([127]YP_001299.1, [128]Supplementary Table SI), confirming the high
   dynamic abundance range covered by the method ([129]Supplementary
   Figure S3). In general, the protein abundances determined by multiple
   heavy reference peptides per protein showed good agreement, even for
   low abundance proteins ([130]Supplementary Table SI). Moreover, the
   values determined here matched very well with those published in a
   recent study and the structural benchmarks employed therein
   ([131]Malmström et al, 2009) ([132]Supplementary Figure S7). In a
   second step, these abundance values were aligned with the average
   intensities of the three PTPs of each protein with the highest MS
   response, the same peptides that were in the focus of the directed
   LC–MS analysis for peptide identification. In the same operation, we
   therefore consistently estimated the absolute abundances of all
   identified proteins in each of the samples. On average, a high squared
   Pearson correlation (R^2=0.805) of the absolute abundances accurately
   determined by heavy peptide references and their average feature