Proteomic Architecture of Human Coronary and Aortic Atherosclerosis

Abstract

Backgound: The inability to detect premature atherosclerosis significantly hinders implementation of personalized therapy to prevent coronary heart disease. A comprehensive understanding of arterial protein networks and how they change in early atherosclerosis could identify new biomarkers for disease detection and improved therapeutic targets.

Methods: Here we describe the human arterial proteome and proteomic features strongly associated with early atherosclerosis based on mass spectrometry analysis of coronary artery and aortic specimens from 100 autopsied young adults (200 arterial specimens). Convex analysis of mixtures, differential dependent network modeling, and bioinformatic analyses defined the composition, network rewiring, and likely regulatory features of the protein networks associated with early atherosclerosis and how they vary across 2 anatomic distributions.

Results: The data document significant differences in mitochondrial protein abundance between coronary and aortic samples (coronary>>aortic), and between atherosclerotic and normal tissues (atherosclerotic<<normal), and major alterations in tumor necrosis factor, insulin receptor, peroxisome proliferator-activated receptor-α, and peroxisome proliferator-activated receptor-γ protein networks, as well, in the setting of early disease. In addition, a subset of tissue protein biomarkers indicative of early atherosclerosis was shown to predict anatomically defined coronary atherosclerosis when measured in plasma samples in a separate clinical cohort (area under the curve=0.92 [0.83-0.96]), thereby validating the use of human tissue proteomics to discover relevant plasma biomarkers for clinical applications. In addition to the specific proteins and pathways identified here, the publicly available data resource and the analysis pipeline used illustrate a strategy for interrogating and interpreting the proteomic architecture of tissues that may be relevant for other chronic diseases characterized by multicellular tissue phenotypes.

Conclusions: The human arterial proteome can be viewed as a complex network whose architectural features vary considerably as a function of anatomic location and the presence or absence of atherosclerosis. The data suggest important reductions in mitochondrial protein abundance in early atherosclerosis and also identify a subset of plasma proteins that are highly predictive of angiographically defined coronary disease.

Keywords: atherosclerosis; biomarkers; mitochondria; proteomics.

Figure 1. Weighted Co-Expression Network Analysis of Human LAD Proteins

a. Adjacency map of LAD proteins color coded by module assignment based on hierarchical clustering of the topological overlap matrix (TOM)-based dissimilarity measure. For clarity of presentation only nodes (proteins) with at least one edge (adjacency measure (k)) > 97.5%tile are shown (n-544). b. Left panel shows that scale-free topology is best approximated when the adjacency power parameter β= 10. Right panel shows the log-log plot of adjacency (k) vs prob(k) with β= 10, confirming the power-law relationship in the connectivity of the expressed proteins. c. Module assignment for fifty 90% random samples of the data illustrating the overall stability of the modular structure of the protein expression patterns. Colors are assigned according cluster size which may vary with each random sample. As a result actual color assignment may vary from run to run, but module membership remains relatively stable. d. Top non-redundant GO Terms with Bonferroni corrected p-values and an exemplar protein for each module.

Figure 2. Comparison of Mitochondrial Proteins in Normal LAD and AA Samples

DIA-MS analysis of completely normal LAD and AA samples (n = 30 from each anatomic region) with adjustment for age, sex, MYH11, RABA7A, TERA, G6PI. LAD vs AA MANOVA p-values by mitochondrial protein group: fatty acid metabolism, p = 0.04; oxidative phosphorylation, p < 0.0001; TCA, p < 0.0001; mito biogenesis, p < 0.0001.

Figure 3. Convex Analysis of Mixtures of LAD Protein Data

a. Heatmap of mixed expressions of upregulated marker proteins (UMPs) in 99 LAD samples. b. Estimated proportions of NL1, NL2, FS, and FP across 99 LAD samples. c. Heatmap of subpopulation-specific expressions of UMPs. d. Mathematical description on the i-th protein expression readout ‘x’ as a weighted sum of the protein expressions in the distinctive tissue types ‘s’ present in the heterogeneous samples, weighted by the mixing proportions ‘a’. e. Geometry of the mixing operation in scatter space that produces a compressed and rotated scatter simplex whose vertices host subpopulation-specific UMPs and correspond to mixing proportions.

Figure 4. Patho-Proteomic Phenotyping

a. Hierarchical clustering and principal component score plot of the pathologist grading of extent of FP, FS, and normal tissue combined with CAM-derived estimates of proportion of four different empirical tissue types. Dashed circles indicate samples (2:1) from the extremes of the first principal component which separates fibrous plaques and normal arterial tissue. b. Volcano plot of log(fold change:FP/Normal) vs. -log(t-test q-value) for 944 arterial proteins. Black data points indicate proteins with a fold-change of >1.7 (or < 1/1.7) and a q-value of <0.05. (Two proteins with –log(q-value)>10 not shown.) c. Heatmap of the spectral count values for the N=88 proteins meeting the fold-change and q-value criteria noted in panel b. The individual proteins are listed in Supplementary Tables 8 and 9. d. A combined network map of the significantly up- and down-regulated proteins that are consistent with effects of an upstream master regulator. The observed pattern of proteins in FP are highly suggestive of TNF activation (p=1.64E-6, z-score=3.19), and inhibition of PPAR-α (p=3.55E-10, z-score=-3.06), PPAR-γ (p=2.87E-10, z-score=-3.02), and the insulin receptor (1.41E-12, z-score=-2.16). PPAR-α, PPAR-γ, and the insulin receptor pathways are themselves inhibited by TNF activation. When using >1.5 fold-change criteria, several additional regulatory networks were also identified (e.g. TGFB1, TP53, SP1, MYC). e. Predicted disease processes and their affiliated proteins significantly overrepresented by the up- and down-regulated proteins in fibrous plaques. The data are consistent with major alterations in cell-cell adhesion/interaction (p = 3.64E-03), lipid uptake (p = 1.67E-03), glucose homeostasis (p = 2.83E-06) and blood pressure regulation (p = 3.68E-03).

Figure 5. Differential Dependent Network Analysis of LAD Fibrous Plaque Proteins

a. Plot depicts re-wiring of the protein network between FP and normal samples. Green nodes are down-regulated and red nodes are up-regulated in FP samples. Green edges indicate significant correlation in normal samples, but not FP samples. Red edges indicate significant correlation in FP samples. Black squares indicate differential network “hub proteins” (ie. proteins with different couplings to network partners in FP and normal samples.) GO term analysis of the differential network hub proteins indicated significant enrichment of TCA proteins (p=4.8×10^-6). b. TCA cycle proteins with MS data available for additional anlaysis. Every protein indicated by a red box was quantitatively lower in FP samples than in normal samples after adjustment for housekeeping proteins, age and sex. c. Statistical comparison of TCA proteins in FP vs normal LAD samples.

Figure 6. Comparison of Mitochondrial Proteins in FP vs NL Samples from the LAD and AA

DIA-MS analysis was used to compare a targeted set of mitochondrial proteins in n=15 FP and n=30 NL LAD samples after adjustment for age, sex, MYH11, RABA7A, TERA, G6PI. Histogram bars indicate relative difference ebetween FP and NL samples in each anaotomic location. LAD MANOVA p-values for each mitochondrial protein group: p < 0.0001 for each group; AA MANOVA p-values for each mitochondrial protein group: p= n.s for each group.

Figure 7. Clinical Validation of Fibrous Plaque Proteins

A. Solution path for the elastic net model (α = 0.9) with parameter estimates as a function of increasing number of variables entered into the model as reflected by the sum of the Scaled Parameter Estimates. The generally monotonically increasing solution paths suggest relative stability and no major interactions among the proteins included in the model. B. The-LogLikelihood derived from the leave-one-out validation samples as a function of increasing number of variables entered into the model. The negative inflection point indicates the optimal feature selection after which adding additional variables lead to worsening (increasing) –LogLikelihoods due to overfitting of the data. C. The ROC curve for the optimal model predicting presence of angiographically proven coronary disease. Bootstrap estimates of the AUC and its 95% CI based on 10,000 samples were 0.93 and 0.85-0.97 respectively. D. Parameter estimates for proteins included in the plasma protein biomarker panel. Six proteins (indicated in bold) had the strongest evidence for contributing to model performance. However, elastic net models permit inclusion of additional variables that may be correlated with other terms in the model. These proteins may not individually contribute greatly to the overall prediction performance, but nevertheless participate in underlying mechanistic pathways worthy of further consideration.