Quantitative proteomics approach for identifying protein–drug interactions in complex mixtures using protein stability measurements

Abstract

Knowledge about the protein targets of therapeutic agents is critical for understanding drug mode of action. Described here is a mass spectrometry-based proteomics method for identifying the protein target(s) of drug molecules that is potentially applicable to any drug compound. The method, which involves making thermodynamic measurements of protein-folding reactions in complex biological mixtures to detect protein–drug interactions, is demonstrated in an experiment to identify yeast protein targets of the immunosuppressive drug, cyclosporin A (CsA). Two of the ten protein targets identified in this proof of principle work were cyclophilin A and UDP-glucose-4-epimerase, both of which are known to interact with CsA, the former through a direct binding event (K_d ∼ 70 nM) and the latter through an indirect binding event. These two previously known protein targets validate the methodology and its ability to detect both the on- and off-target effects of protein–drug interactions. The other eight protein targets discovered here, which include several proteins involved in glucose metabolism, create a new framework in which to investigate the molecular basis of CsA side effects in humans.

Knowledge about protein–drug interactions is critical for understanding the therapeutic effects and pleiotropic activities of drugs. Whereas genome- and proteome-based expression profiling studies of organisms and cell lines grown in the presence of drugs can provide valuable information about the functional analysis of drugs (e.g., the physiological processes and biological pathways impacted by both the on- and off-target effects of drug action), such expression profiling studies do not provide a readout of the specific protein interactions involved in a drug’s mode of action. The comprehensive analysis of such drug-induced protein interactions requires the ability to assay all the proteins in a proteome for drug binding and to quantitatively assess their affinity for the drug target.

Several protein ligand-binding assays, such as the yeast two-hybrid assay (1, 2) and affinity chromatography techniques coupled with mass spectrometry (3–6), exist for the large-scale analysis of protein–protein interactions. However, generic protein ligand-binding assays are lacking for the detection of protein–drug interactions on the proteomic scale. Although applications of the yeast two-hybrid assay and affinity chromatography techniques to nonprotein ligands have been reported (7–9), these approaches are frequently problematic with nonprotein ligands (e.g., small drug molecules) due to the inherent difficulties associated with conjugating such ligands to the necessary bait protein in the yeast-based assay and to the necessary solid support in the affinity chromatography and mass spectrometry-based assay. Conventional protein ligand-binding assays (e.g., those based on calorimetric and spectroscopic techniques) are also not amenable to analyses on the proteomic scale because they require relatively large amounts of purified protein, which are not easy to generate on a large scale.

Currently, there is also a need for new analytical methodologies capable of detecting both the on- and off-target effects of protein ligand-binding. Most protein ligand-binding assays, including the yeast two-hybrid assay and affinity chromatography techniques, detect on-target effects, or direct binding events. Off-target effects, or indirect binding events, are more difficult to detect and are often the cause of deleterious side effects. Such off-target effects can arise, for example, when the direct binding interaction between one protein and a target ligand precludes and/or induces the interaction of one protein to a second protein. Off-target effects go undetected in the yeast two-hybrid assay, and the affinity chromatography technique is only sensitive to off-target effects in which a direct binding event recruits additional proteins to the protein complex. Off-target effects in which a direct ligand-binding event to a target protein precludes interaction(s) between the target protein and other protein(s) currently escape detection by all existing protein ligand-binding assays.

Here we describe a general mass spectrometry-based proteomics method for making thermodynamic measurements of protein-folding reactions in complex biological mixtures to detect protein–drug interactions. We demonstrate the ability to simultaneously assay the protein-folding and ligand-binding properties of 327 proteins in a yeast cell lysate. The new platform’s ability to simultaneously analyze multiple proteins for binding to a target ligand is unmatched by existing methodologies for the detection and quantitation of protein ligand-binding affinities, and it is this unique multiplex and large-scale capability that enables the detection of both on- and off-target effects of drug binding. The proof of principle study described here was designed to investigate the protein targets of an already well-studied immunosuppressive drug, cyclosporin A (CsA) (10). However, the described methodology can be readily adapted to the study of a wide variety of other ligands and their protein targets in other complex biological mixtures.

Results and Discussion

General Strategy.

The experimental strategy developed in this work (Fig. 1 and Fig. S1) combines a chemical modification- and mass spectrometry (MS)-based method, termed Stability of Proteins from Rates of Oxidation (SPROX) (11), with quantitative MS-based proteomics. The proteomics platform used here relied on a multidimensional protein identification technology (MudPIT) (12) and the use of isobaric mass tags (13) for peptide quantitation. However, in theory any quantitative proteomics platform could be utilized, provided the oxidation reaction products generated in the SPROX experiment can be effectively quantified.

Fig. 1.

Schematic representation of the experimental strategy developed in this work. A complex biological mixture of proteins (e.g., a cell lysate) is subjected to two analyses, one in the absence of drug and one in the presence of drug. In each analysis, aliquots of the protein mixture (or the protein mixture combined with drug) are diluted into a series of buffers containing increasing concentrations of a chemical denaturant [e.g., guanidinium hydrochloride (GdmCl)]. The protein samples in each GdmCl-containing buffer are reacted with the same amount of hydrogen peroxide for the same amount of time. The reaction time and concentration of hydrogen peroxide are tuned such that the thioether groups in the side chain of methionine residues are selectively oxidized. The protein oxidation reaction is quenched and the oxidized protein samples are submitted to a quantitative proteomic analysis and the nonoxidized and oxidized methionine-containing peptides are quantified as a function of the SPROX buffer denaturant concentration.

The SPROX methodology used in this work utilizes the chemical denaturant-dependent oxidation rates of methionine residues in a protein to report on the thermodynamic properties of its global and/or subglobal unfolding/refolding reactions. Our initial development of SPROX included a matrix-assisted laser desorption/ionization (MALDI)- and/or electrospray ionization (ESI)-based readout of the intact protein (11). The integration of SPROX into the LC-MS-based proteomics platform described here requires a peptide readout in which the disappearance of nonoxidized methionine-containing peptides and/or the appearance of oxidized methionine-containing peptides are quantified as a function of the SPROX buffer denaturant concentration. Because the oxidation reaction is performed on the intact proteins in the SPROX buffers (i.e., before the protease digestion step) the oxidation state of the detected methionine-containing peptides provides biophysical information on the protein from which it was derived.

The platform’s peptide readout creates a fundamental upper limit to the number of proteins in a complex mixture that can be analyzed. This upper limit is determined by the number of proteins that are identified with methionine-containing peptides in the proteomics experiment. In our experiments approximately 20% of the peptide sequences identified in the MudPIT analyses contained at least one methionine residue (see Table S1). These methionine-containing peptides mapped to approximately 33% of the approximately 1,400 proteins identified in the MudPIT analyses (see Table S2). The yeast lysate proteins that could be assayed for CsA binding in our experiments were also restricted in our experiment to those in which the methionine-containing peptides used for their identification were consistently detected and accurately quantified (SI Text) in the MudPIT analyses performed on the samples sets generated in the presence and absence of CsA. Ultimately, 322 different nonoxidized methionine-containing peptides and 564 different oxidized methionine-containing peptides covering a total of 327 unique proteins (see Table S2) yielded useful SPROX data, which included 188 cases in which both the disappearance and appearance of the nonoxidized and oxidized forms of the same methionine-containing peptide was successfully quantified.

Many of the methionine-containing peptides identified in this work (Table S2) were derived from highly expressed proteins in our yeast system (e.g., those proteins involved in glycolysis). However, whereas methionine-containing peptides from more abundant proteins in a given mixture are more likely to appear in the assay, there is not a bias for such proteins to appear as hits. This is because the protein mixtures analyzed in the absence and presence of ligand are identical; and hits are detected by a change in protein stability and not by a change in protein expression. Indeed, the identified protein targets in this work have a range of different expression levels (see Table 1).

Table 1.

Identified protein targets of CsA

Proteins and peptide hits	Molecules/cell*	Met-Residue#	ΔTransition Midpoint ([GdmCl])†
Cyclophilin A (YDR155C) VIPDFMLQGGDFTAGNGTGGK‡	86,000	Met-59	+1.5
VIPDFM(Ox)LQGGDFTAGNGTGGK§		Met-59	+1.3
Carbamoylphosphate Synthetase (YJL130C) ILAIDVGMK‡	11,000	Met-236	+1.0
Glycogen Synthase (YLR258W) YKNKGADMFIEALARLNYR‡	14,600	Met-330	+0.5
Glutamate Dehydrogenase (YDL215C) GGVTSSSM(Ox)EVLASLALNDNDFVHK‡	7,400	Met-927	+1.5
Uncharacterized ORF (YIL137C) VMSQVLFLNK§	5,110	Met-703	> +1.5
Pyruvate Kinase (YAL038W) SNLAGKPVICATQM(Ox)LESM(Ox)TYNPR¶	291,000	Met-300/304	−0.5
AEVSDVGNAILDGADCVMLSGETAK¶		Met-330	+0.5
VTDGVMVAR¶		Met-261	+0.5
ACDDKIMYVDYK¶		Met-125	−0.5
UDP-Glucose-4-Epimerase (YBR019C) SLEMVYK¶	Not available	Met-518	+0.5
Cytoplasmic Arginyl-tRNA synthetase (YDR341C) M(Ox)LSFEGDTGPYLQYAHSR¶	20,600	Met-478	+0.7
Ubiquitin-like Protein of SUMO Family (YDR510W) IQADQTPEDLDM(Ox)EDNDIIEAHR¶	2,940	Met-83	+1.0
NAD-Specific Isocritrate Dehydrogenase (YDL066W) VKQPVVELDGDEMTR¶	Not available	Met-34	+0.7

^*Values are taken from data in ref. 26.

^†Values were estimated by visual inspection of SPROX data obtained in the presence and absence of ligand (see Fig. 3 and Fig. S5).

^‡Normalized reporter ion intensity differences were > 0.6 at 4 consecutive denaturant concentrations.

^§Normalized reporter ion intensity differences were > 0.6 at 3 consecutive denaturant concentrations.

^¶Normalized reporter ion intensity differences were > 0.6 at 2 consecutive denaturant concentrations.

Transition Midpoint Analysis.

The quantitative proteomics platform employed here relied on the reporter ion intensities generated in the MS/MS analysis of each tandem mass tagged methionine-containing peptide for quantitation. The reporter ion intensities were subject to two normalizations (see SI Text) including one that accounted for reporter ion intensity differences that were observed from peptide to peptide and one that accounted for reporter ion intensity differences that were observed for the different tandem mass tags in a given MudPIT run (Table S3 and Fig. S2). The normalized reporter ion intensities calculated for each methionine-containing peptide in Table S2 could be used to construct two sets of SPROX data (i.e., normalized reporter ion intensity versus denaturant concentration plots) for each methionine-containing peptide, including one set of data using the intensities obtained in the absence of ligand and one set of data using the intensities obtained in the presence of ligand (Figs. 2 and 3).

Fig. 2.

SPROX data obtained on a selected methionine-containing peptide “nonhit” from glycerol dehydrogenase, which was not identified as a CsA target. (A) and (B) Data obtained on the nonoxidized and oxidized (respectively) forms of the peptide. Data obtained from analyses conducted in the absence and in the presence of CsA are shown in blue and red bars, respectively. The dotted lines mark the SPROX transition midpoints, which were all within a small range, < 0.5 M, and within experimental error.

Fig. 3.

SPROX data obtained on a selected methionine-containing peptide hit from CypA, which was identified as CsA target. (A) and (B) Data obtained for the nonoxidized and oxidized (respectively) forms of the peptide. Data obtained from analyses conducted in the absence and in the presence of CsA are shown in blue and red bars, respectively. The hashed bars represent points that were missing from the data and assigned a value that was the average of the experimentally determined normalized reporter ion intensity on either side of the missing denaturant concentration. In this case the missing data was a result of poor reporter ion intensities in one of the MudPIT runs. The arrows indicate cases where there was a significant difference (> 0.6) between the normalized TMT reporter ion ratios obtained with and without CsA. The dotted lines mark the SPROX transition midpoints, which were significantly shifted (> 0.5 M) to higher denaturant concentrations in the presence of ligand.

An analysis of the SPROX data obtained on the methionine-containing peptides identified in the yeast lysate sample (see SI Text and Fig. S3) revealed a distribution of transition midpoints that ranged from < 0 M (i.e., only a posttransition baseline was detected) to > 4 M (i.e., only a pretransition baseline was detected) (Fig. S4). The methionine-containing peptides that yielded transition midpoints > 0 M (approximately 85% of the methionine-containing peptides detected) are expected to contain methionine residues that are globally or subglobally protected in the context of the intact proteins to which they map; whereas the methionine-containing peptides that yielded transition midpoints < 0 M (approximately 15% of the methionine-containing peptides detected) are expected to contain methionine residues that are either solvent exposed or only locally protected in the proteins to which they map.

In theory, transition midpoints observed for the nonoxidized and oxidized forms of the same methionine-containing peptide should be similar. This was observed for the glycerol dehydrogenase peptide highlighted in Fig. 2 and in about half of the other 187 cases in which both forms (nonoxidized and oxidized) of a given methionine-containing peptide were detected. In the remaining cases, which included methionine-containing peptides like the CypA peptide highlighted in Fig. 3, the transition midpoint of the oxidized peptide form was generally at a lower denaturant concentration than that of the nonoxidized form. Cases in which such transition midpoint differences existed can be explained by the oxidized form of the protein being less stable than the nonoxidized form. Indeed, the results in Fig. 3 are consistent with the destabilizing effects of methionine oxidation that we have previously measured in the CypA system (11). In general, SPROX data obtained on nonoxidized peptides are expected to provide the most accurate measure of a protein’s wild-type stability. However, it is noteworthy that the nonoxidized and oxidized forms of the CypA peptide highlighted in Fig. 3 showed similar transition midpoint shifts in the presence of CsA. The oxidaton reaction perturbs CypA’s folding behavior, but not its CsA binding properties.

Analysis of Protein–Drug Binding.

The goal of this work was to identify proteins in the yeast cell lysate where the transition midpoints of identifying methionine-containing peptides shifted in the presence of the CsA ligand. A transition midpoint shift to a higher or lower denaturant concentration signifies a CsA-induced stabilization or destabilization, respectively, of the protein to which the peptide maps. Such shifts could result from either direct and/or indirect interactions of a protein with CsA. Proteins with shifted transition midpoints were identified in this work by examining the differences between the normalized reporter ion intensities generated for a given methionine-containing peptide in the absence and presence of ligand. Differences > 0.6 were deemed significant based on the observed distribution of the measured differences (Fig. 4), which revealed 85% of the measured differences were within 0.6 of the average difference that was close to zero. We also reasoned that SPROX data with significant midpoint shifts (i.e., > 0.5 M) would have a minimum of two normalized reporter ion intensity differences > 0.6 and that these differences would occur at consecutive denaturant concentrations in the SPROX data.

Fig. 4.

Distribution of the normalized reporter ion intensity differences obtained by subtracting the normalized reporter ion intensities of the 886 methionine-containing peptides (nonoxidized and oxidized) in the absence of ligand from those intensities obtained in the presence of ligand. Approximately 85% of the observed differences were > -0.6 and < 0.6. The differences outside this range were deemed significant.

Peptides with normalized reporter ion intensity differences > 0.6 in at least two consecutive denaturant concentrations were identified in a series of filtering steps. A total of 196, 66, and 10 peptides were identified with differences > 0.6 in at least 2, 3, and 4 consecutive denaturant concentrations (respectively). None of the identified peptides had normalized reporter ion intensity differences > 0.6 at five or more consecutive denaturant concentrations. Visual inspection of the SPROX data obtained on the 196 peptides identified in the above filtering process revealed a group of 14 peptides with SPROX transition midpoints that were shifted > 0.5 M in the presence of CsA (see Table 1, Fig. 3, and Fig. S5). These 14 peptide “hits” mapped to the 10 different proteins summarized in Table 1. The 182 false peptide hits that were eliminated in the visual inspection step were generally those peptides with at least one poor quality SPROX curve (Fig. S6), that required additional filtering steps (see SI Text) to extract a transition midpoint. Approximately 25% of the 1772 SPROX curves recorded in the work were so-called poor quality, thus it is not surprising that the false peptide hits totaled 182 (i.e., approximately 20% of the 886 methionine-containing peptides detected in this work). It is noteworthy that the sets of 66 and 10 peptides identified with normalized reporter ion intensity differences > 0.6 in at least 3 and 4 consecutive denaturant concentrations (respectively) had decreasing numbers of false peptide hits that had to be identified in the visual inspection step (i.e., 56 and 6, respectively).

A determination of which protein hits in Table 1 originate from direct interactions with CsA (i.e., on-target effects) and which originate from indirect interactions (off-target effects) requires additional studies of CsA binding to the purified proteins. The requisite studies for two of the hits, CypA and UDP-glucose-4-epimerase, have been reported in the literature (14–17). CypA, which was one of two protein hits identified using the most stringent filtering criteria (i.e., significant normalized reporter ion intensity differences at 4 consecutive denaturant concentrations), is known to directly bind CsA (K_d 30–200 nM) (15–17). The transition midpoint shifts of approximately 1.5 and 1.3 M that were detected for the nonoxidized and oxidized CypA peptides (respectively) in Table 1 are consistent with the ligand induced stabilization of CypA expected for the direct binding of CsA to CypA. The shifts, which are within experimental error of the 1.5 M transition midpoint shift measured in our earlier SPROX experiments on a purified form of CypA binding to CsA (11), can be used to calculate K_d values of 26 and 100 nM for the CsA-CypA complex using the nonoxidized and oxidized peptide shifts, respectively (see SI Text).

A second protein hit, UDP-glucose-4-epimerase, identified in Table 1, is also known to have its protein-folding and ligand-binding behavior modulated by CsA, albeit indirectly (14). The peptidyl prolyl cis-trans isomerase activity of CypA has been shown to increase the refolding reaction of the UDP-glucose-4-epimerase homodimer (14). Significantly, CypA has no effect on the refolding behavior of UDP-glucose-4-epimerase in the presence of CsA (14), which inhibits the peptidyl prolyl isomerase activity of cyclophilin A. The CsA-induced stabilization of UDP-glucose-4-epimerase detected in our experiments is most likely due to a thermodynamic consequence of CsA inhibition of CypA’s kinetic effect on the UDP-glucose-4-epimerase folding reaction. Based on the proposed mechanism of CypA’s peptidyl prolyl isomerase activity (18) CypA in the absence of CsA presumably interacts with Xxx-Pro sequences in the unfolded state(s) of UDP-glucose-4-epimerase to produce more native-like and lower energy UDP-glucose-4-epimerase molecules. In the presence of CsA, such interactions of CypA with UDP-glucose-4-epimerase are precluded and UDP-glucose-4-epimerase molecules must fold from less native-like and higher energy denatured state(s). It is noteworthy that such an off-target effect would have escaped detection in both the yeast two-hydrid assay and the affinity chromatography assay, even if the CsA ligand were successfully conjugated to the necessary bait protein in the yeast-based assay and to the necessary solid support in the affinity chromatography assay.

To our knowledge eight of the proteins identified in Table 1 have no known interactions (either direct or indirect) with CsA. These eight newly identified protein targets of CsA create a unique molecular framework in which to better understand documented side effects of CsA, which include posttransplant diabetes mellitus (PTDM) (19). Whereas retrospective studies of different patient populations have established a link between PTDM and the use of CsA to prevent and treat organ rejection in transplant patients (20, 21), very little is known about the molecular basis of this deleterious side effect of CsA. More detailed biochemical studies are clearly needed to better characterize how CsA binding interactions with the newly identified protein targets in Table 1 might lead to PTDM. However, it is interesting to note that two of the protein hits in Table 1, pyruvate kinase and glycogen synthase, are involved in glucose metabolism. Work is in progress to screen purified constructs of these new protein targets for CsA binding to differentiate those that are direct targets of CsA from those that are indirect targets. Ultimately, the network of protein interactions responsible for the indirect targets will be studied in additional experiments on selected protein pools.

One motivation for selecting the CsA and yeast model system for this proof of principle work was to determine if the known indirect protein target of CsA, calcineurin, could be detected. The binding of CsA to CypA is known to induce a binding interaction between CypA and calcineurin, which ultimately inhibits the phosphatase activity of calcineurin (22). SPROX data were obtained for three different methionine-containing peptides from calcineurin A (see YLR433C peptides in Table S2). However, these peptides were not identified as hits. Our inability to identify calcineurin as an indirect protein target of CsA in this work underscores several important caveats of the described methodology. These caveats include (i) the need to detect methionine-containing peptides from the specific domain and/or domains of target proteins that have their thermodynamic properties modulated (either directly or indirectly) by the ligand; and (ii) the need to have sufficiently high protein and ligand concentrations to produce a measurable transition midpoint value shift in the SPROX experiment (see SI Text).

Another fundamental limitation of SPROX for the evaluation of protein ligand-binding affinities is that the oxidation reaction used to probe the protein folding and unfolding reaction can perturb the protein-folding and ligand-binding properties of proteins. Whereas we have previously shown that accurate ΔΔG_f values for some protein systems can be determined by SPROX, even if the protein oxidation products are destabilized (11), it is clear that not all protein-folding and ligand-binding interactions will be free from perturbation. However, as described above SPROX data obtained on nonoxidized forms of methionine-containing peptides are expected to be useful regardless of the perturbing effects of the oxidation reaction.

The described methodology creates a discovery platform for identifying the protein targets, both direct and indirect, of drug molecules. The results of this proof-of-principle study not only demonstrate the platform’s ability to detect both the on- and off-target effects of protein–drug interactions, but they also identify unique protein targets of CsA. Future applications of the described methodology to newly discovered drug molecules promise to aid in the prediction of the side effects of new drugs before they are widely prescribed. Early incorporation of the described platform into drug discovery efforts could also lead to the production of drugs with fewer unwanted side effects. More broadly, this platform provides a unique approach to examine protein-folding. As the protein-folding environment has been implicated in many orphan diseases and aging, this method could be used to assess, in large-scale, the folding environment in “misfolding” diseases and in aging cells (23). Such protein-folding and ligand-binding analyses add a new dimension to mass spectrometry-based proteomics platforms.

Materials and Methods

Preparation of Yeast Cell Lysate.

Yeast GAL1 overexpression strains (24) for cyclophilin A (YDR155C) and calcineurin A (YLR433C) were obtained from Open Biosystems, cultured and lysed according to standard protocols, which are described further in SI Materials and Methods. The lysates from both strains were combined into one sample that was subsequently divided into two equal portions, one containing 450 μL lysate and 50 μL of DMSO that was the minus ligand stock and one containing 450 μL lysate and 50 μL of a 10 mM CsA DMSO solution that was the plus ligand stock. The total protein concentration in each stock was approximately 18 mg/mL.

SPROX Analyses.

The denaturant-containing buffers used for SPROX analyses were comprised of 20 mM Phosphate (pH 7.4) and specific [GdmCl] that ranged from 0.2 to 8.1 M. The [GdmCl] in each buffer was determined by a refractive index measurement as described elsewhere (25). In the SPROX analyses, 20 μL aliquots of the above stock lysate solutions were combined with 25 μL of the above SPROX buffers to yield a series of 12 solutions that contained the lysate (or the lysate plus CsA) and concentrations of GdmCl that ranged from 0 to 4 M. The lysate (and lysate plus CsA) samples in the denaturant-containing SPROX buffers were equilibrated for 15 min, and the oxidation reaction was initiated with the addition of 5 μL of 9.8 M H₂O₂. The final concentrations of GdmCl in the 12 oxidation reactions were 0.1, 0.5, 1.0, 1.3, 1.5, 1.8, 2.0, 2.3, 2.6, 3.0, 3.5, and 4.0 M. The oxidation reactions were allowed to proceed for 3 min in 0.98 M H₂O₂ before they were quenched with 1 mL of 300 mM Methionine. The protein in each SPROX buffer was precipitated using trichloroacetic acid, and the resulting protein pellets were washed three times with 300 μL of ice-cold ethanol.

Proteomic Sample Preparation and Analysis.

The protein pellets from each denaturant-containing buffer were dissolved in 120 μL of buffer containing 100 mM triethyl ammonium bicarbonate and 0.1% SDS. A 10 μL aliquot of each sample was removed and used in a Bradford assay to determine the concentration of protein in each sample, which was determined to be between 0.7 and 1 mg/mL. The 24 dissolved protein samples were reduced, alkylated, digested with Trypsin, labeled with one of the six tandem mass tags (TMT) in a TMTsixplex™ Isobaric Mass Tagging Kit (Thermo Scientific Pierce Products), appropriately pooled, and analyzed in a total of four MudPIT runs using a LTQ-Orbitrap mass spectrometer (Thermo Fisher Scientific). Ultimately, the tandem mass tag reporter ion intensities generated in the MudPIT analyses were used to quantify the oxidized and nonoxidized forms of the methionine-containing peptides in each sample. Additional details about the pooling strategy, the MudPIT analyses, and the tandem mass tag reporter ion analysis are provided in SI Materials and Methods.