Plasma proteomic analysis reveals altered protein abundances in cardiovascular disease

Blood plasma specimens were collected within the context of a prospective study on CVD (part of the FP7 project SysVASC: Systems Biology to Identify Molecular Targets for Vascular Disease Treatment), from control participants with no cardiovascular background (n = 13) and CVD patients (n = 19) in Medical University of Graz, Graz, Austria. All control and over 40% of the CVD (n = 8) samples were collected from deceased donors. The study was approved by the respective local ethics committee (approval number: 26-355 ex 13/14). CVD subjects were characterized by established vascular disease and had at least one of the following conditions: peripheral artery disease; coronary artery disease; arterial hypertension; cerebral hemorrhage; or carotid stenosis. The characteristics of the participants in this study are presented in Table 1.

Plasma samples were processed with the filter-aided sample preparation protocol, as previously described by Wisniewski et al. [27]. Specifically, protein concentration of plasma was determined by Bradford assay and 200 µg of protein per sample was mixed with 0.2 ml of 8 M urea in 0.1 M Tris/HCl, pH 8.5 (urea buffer). Reduction was performed with 0.1 M dithioerythritol (DTE) for 20 min (room temperature—RT) and the samples were loaded into 30 kDa Amicon Ultra Centrifugal Filters, Ultracel (Merck Millipore). The filters were then centrifuged at 16,000g for 15 min and washed twice with 0.2 ml of urea buffer. The concentrates were then mixed with 0.1 ml of 50 mM iodoacetamide in urea solution and incubated in the dark, at RT for 20 min (alkylation), followed by centrifugation for 10 min. Then, the filters were washed once with 0.1 ml of urea solution and twice with 0.1 ml of 50 mM NH₄HCO₃, pH 8.5. Finally, samples were digested with trypsin (proteomics grade) overnight, in the dark at RT (trypsin/protein ratio: 1/100) and the peptides were collected by centrifugation of the filter units for 10 min at 16,000g in clean tubes. The filters were washed once more with 40 µl of 50 mM NH₄HCO₃ and eluted peptides were collected in the same tube. The eluates were lyophilized and stored at − 20 °C until further processing.

Five microgram of protein digest was loaded onto a Dionex Ultimate 3000 RSLS nano flow system (Dionex, Camberly UK). After loading onto a Dionex 0.1 × 20 mm 5 μm C18 nano trap column at a flow rate of 5 μl/min in 0.1% formic acid and 2% acetonitrile, samples were applied onto an Acclaim PepMap C18 nano column 75 μm × 50 cm (Dionex, Sunnyvale, CA, USA), 2 μm 100 Å at a flow rate of 0.3 μl/min. The trap and nano flow column were maintained at 35 °C. The samples were eluted with a gradient of solvent A: 0.1% formic acid and 2% acetonitrile versus solvent B: 0.1% formic acid and 80% acetonitrile starting at 1% B for 5 min rising to 5% B at 10 min then to 25% B at 360 min and 65% B at 480 min. The column was then washed and re-equilibrated prior to injection of the next sample. The eluent was ionized using a Proxeon nano spray ESI source operating in positive ion mode into an Orbitrap Velos FTMS (Thermo Finnigan, Bremen, Germany). Ionization voltage was 2.6 kV and the capillary temperature was 275 °C. The mass-spectrometer was operated in MS/MS mode scanning from 380 to 1600 amu. The resolution of ions in MS1 was 60,000 and 7500 for higher-energy collisional dissociation (HCD) MS2. The top 20 multiply charged ions were selected from each scan for MS/MS analysis using HCD at 40% collision energy. AGC settings were 1,000,000 for full scan in the FTMS and 200,000 for MSn. Dynamic exclusion was enabled with a repeat count of 1, exclusion duration of 30 s.

Protein identification was performed with Proteome Discoverer 1.4 (Thermo Scientific) using the SEQUEST search engine [28]. Protein search was performed against the SwissProt human protein database [29] (30.05.2016) containing 20,197 reviewed entries. The following search parameters were applied: (i) precursor mass tolerance: 10 ppm and fragment mass tolerance: 0.05 Da; (ii) full tryptic digestion; (iii) max missed cleavage sites: 2; (iv) static modifications: carbamidomethylation of cysteine; (v) dynamic modifications: oxidation of methionine; (vi) event detector mass precision: 2 ppm; (vii) precursor mass: 600–5000 Da; (viii) collision energy: 0–1000 eV; (ix) target FDR (strict): 0.01; (x) target FDR (relaxed): 0.05; (xi) FDR validation based on: q-value. Obtained results were further processed by applying the following filters: (i) high, medium and low confidence peptides (FDR < 10%); (ii) peptide rank up to 5; (iii) peptide grouping was enabled and protein grouping was disabled. The list of peptides was exported and processed using a clustering approach that allows combining the data sets acquired in the course of multiple MS runs. This approach groups features (as defined by calibrated retention time and mass) into clusters, followed by the sequence assignment. In this way, data are harmonized, and consistency in sequence assignment when multiple samples are analyzed in individual experiments is achieved. In brief, data were processed in the following steps: (1) calibration: to adjust for sample to sample LC retention time (rt) shifts, rt was calibrated using LOWESS (locally weighted scatterplot smoothing) non-parametric regression method [30]. To identify and calibrate respective peptides across the whole study, a peptide list from one sample (i.e. sample 27—control) covering the full mass and rt range was selected as reference, based on which alignment/calibration of rt of all samples was performed. (2) Clustering: peptides, as defined by mass and calibrated rt, from all exports from Proteome Discoverer were compiled into a single file. This compiled list was used as an input file for clustering. Clustering was then performed according to the following algorithm: all data points (pairs of mass and rt) were placed in a 2-dimensional plane. Clusters (mass window of ± 5 ppm and retention time window of ± 5% of the feature’s rt) were defined and moved until capturing most data points. The center of these clusters (pairs of mass and rt) were compiled in a “cluster list”, with each of the clusters assigned with a unique identifier. Only clusters with at least 2 members were considered for further analysis. (3) Matching and sequence assignment: as described above, at this stage each cluster includes a set of features (peptides) within a certain “window” of mass and rt. For each cluster, the sequence with the highest frequency across the study was selected as representative. In case of a tie, the sequence with the highest Xcorr was selected. (4) Protein annotation: for each assigned sequence, Uniprot ID was annotated with protein name and gene symbol, based on the respective Uniprot database. In the case of one peptide being assigned to multiple proteins, the protein represented with highest frequency across the cluster list was selected. (5) Retrieve peptide information for further quantification: peptide area, confidence, Xcorr and mass/retention time information were retrieved from the peptide lists of each sample according to their cluster IDs. Only peptides < 5 ppm difference between experimental and theoretical mass were considered. (6) Quantification: Samples were randomly divided into discovery and validation set based on the rule 2/3 (n = 22, 13 cases and 8 controls) and 1/3 (n = 11, 6 cases and 5 controls), respectively. Each set of samples was analyzed separately as follows: for a limited number of sequences for which no peptide area could be retrieved by proteome discoverer (this is a well-known, but not yet corrected problem of this software), the missing values were replaced by the mean area values of that group. When the peptide was not identified in the particular sample, the missing values were replaced with zero. For the discovery set, only peptides reported in more than 50% of the samples of at least one group were included in the differential expression analysis, while no filters were applied in the test set analysis. Subsequently, part per million (ppm)-normalization of the peptide peak areas was conducted according to the following formula: normalized peak area = (peptide peak area/total peak area) × 10⁶. Protein abundance in each sample was calculated as the sum of all normalized peptide areas for a given protein, as described previously [31]. Statistical analysis was based on the Mann–Whitney test performed using R package. Proteins with p-value ≤ 0.05 were considered as statistically significant.

Functional analysis was performed with the ClueGO plug-in [32] in Cytoscape 3.4.0 [33] as well as the Ingenuity Pathway Analysis software (QIAGEN Inc.) [34]. In regard to the ClueGO analysis, ontologies were retrieved from REACTOME pathways database (updated on September 4, 2017) and only statistical significant pathways (Bonferroni corrected p-value ≤ 0.05, two-sided hypergeometric test) were taken into account. For the remaining parameters, default settings were used. Results were simplified based on biological relevance and only the leading term from each group is presented.

Spearman rank correlation analysis was performed using IBM SPSS Statistics for Windows, Version 22.0 after logarithmic transformation of the values.

In total, 32 plasma samples, 13 controls and 19 CVD cases, were analyzed by high-resolution LC–MS/MS. To establish confidence in the validity of the results from the differential expression analysis, the samples were randomly divided into discovery (2/3 of the samples; corresponding to 21 samples—13 cases and 8 controls) and test (1/3 of the samples; corresponding to 11 samples—6 cases and 5 controls) sets. In the discovery set, only peptides detected in at least 50% of samples per group (cases or controls) were considered, resulting in a total of 3796 identified proteins. Of these, 838 protein identifications were based on at least 2 peptides. The full list of identified peptides and proteins is provided in Additional file 1. To confirm the biological relevance of this dataset, pathway annotation was performed using the ClueGO plug-in of Cytoscape platform (Additional file 2). As shown, the detected proteins segregated into (plasma) biologically relevant pathways such as complement cascade, scavenging heme from plasma, fibrin clot formation and platelet degranulation (these being among the statistically significant formed pathways—Additional file 2). When only identifications based on at least 2 peptides were included in the analysis, HDL assembly and TGF-b signaling also emerged among the statistically significant pathways (Additional file 2). Collectively, this phenotyping of the findings via pathway annotation verified that the proteins identified via the applied proteomics approach are biologically relevant and the ones expected to be found in plasma based on the current knowledge.

To assess the reliability of the applied quantification approach, the relative abundance of proteins from the discovery set was compared to the respective absolute concentrations reported in the plasma proteome database (PPD) (http://​plasmaproteomeda​tabase.​org/​index.​html) [35]. For this comparison, PPD entries originating from spectral counting experiments on plasma were considered. This corresponded to 299 proteins (286 proteins in the controls, 288 proteins in the cases) reported in both PPD and our study (listed in Additional file 3). Spearman rank correlation analysis revealed significant correlation between the reported PPD concentrations and the observed protein abundances in our datasets (r_s = 0.521, p-value < 0.0001 for controls; and r_s = 0.536, p-value < 0.0001 for cases). Scatter plots of these comparisons are presented in Fig. 2. This result supports that the applied quantification strategy is appropriate incorporating state of the art knowledge.

To increase reliability in the differential expression analysis in the discovery set, given also the relatively small sample size, only proteins identified based on at least 2 peptides were considered, and differential expression was defined based on statistical significance (Mann–Whitney p-value ≤ 0.05) and change by at least twofold in the cases versus controls. A total of 100 proteins were found to meet these criteria (listed in Additional file 1). These findings were subsequently examined for validity in an independent test set of 11 samples, including 5 controls and 6 cases (the full list of peptides and proteins identified in the test set is provided in Additional file 3). Spearman rank correlation analysis revealed significant correlation between the protein abundances of discovery and test sets (r_s = 0.850, p-value < 0.0001 for controls; and r_s = 0.893, p-value < 0.0001 for cases). Out of the 100 differentially expressed proteins, 39 proteins were detected at the same expression trend (up- or down-regulated) and with a fold change of at least 2, when comparing cases and controls in the test set (Table 2, Additional file 3). Of these 39 proteins, 9 showed statistically significant changes (Mann–Whitney p-value ≤ 0.05) also in the test set (Table 2, Additional file 3) and thirteen were described in the past as associated with cardiovascular disease, further supporting the validity of our approach. As shown in Table 2, the latter proteins included alpha-2-macroglobulin; apolipoprotein B; heparin cofactor 2; lipopolysaccharide-binding protein; latent-transforming growth factor beta-binding protein 2; and others.

Interestingly, molecular pathways that these 39 proteins participate in, based on the Ingenuity Pathway Analysis (IPA), included acute phase signaling; LXR/RXR activation; interleukin-6 signaling; coagulation system; and iNOS signaling (Additional file 4; canonical pathways), suggesting potential impact on these processes based on the observed protein abundance changes in CVD. In fact, and despite the small input dataset (39 proteins), most of the aforementioned pathways (with the exception of iNOS signaling) were predicted to be changing at statistically significant levels in the CVD cases versus controls (Additional file 4).

Along the same lines, IPA predicted significant enrichment of various cardiovascular system- and lipid metabolism-related processes based on the 39 verified proteins; these included uptake of lipid; activation of vascular endothelial cells; atherogenesis; hypercholesterolemia; vasculogenesis; dilated cardiomyopathy; all emerging as statistically significant functions reflected in the verified proteins (Additional file 4; biological functions).