Co-regulatory networks of human serum proteins link genetics to disease

Abstract

Proteins circulating in the blood are critical for age-related disease processes; however, the serum proteome has remained largely unexplored. To this end, 4137 proteins covering most predicted extracellular proteins were measured in the serum of 5457 Icelanders over 65 years of age. Pairwise correlation between proteins as they varied across individuals revealed 27 different network modules of serum proteins, many of which were associated with cardiovascular and metabolic disease states, as well as overall survival. The protein modules were controlled by cis- and trans-acting genetic variants, which in many cases were also associated with complex disease. This revealed co-regulated groups of circulating proteins that incorporated regulatory control between tissues and demonstrated close relationships to past, current, and future disease states.

Fig. 1.. The serum protein network structure.

(A) Hierarchical clustering dendrogram using dynamic tree cut (29), revealing 27 serum protein modules. Each branch of the dendrogram represents a single protein, and the colored bar below denotes its corresponding protein module, as annotated in the legend to the right. The dendrogram height is the distance between proteins (14). (B) The cohort was randomly split into two equal parts, one for a training set and another for the test set, and a summary Z score statistics (15), plotted for each module presented as colored data points. The summary Z score <2 (blue dotted line) indicates no preservation; 2< summary Z score <10 (between the blue and green dotted lines) indicates moderate evidence of preservation; and a summary Z score >10 (green dotted line) indicates strong evidence of preservation. See fig. S10 for the ⅔ versus ⅓ split of the cohort.

Fig. 2.. The relationship between module’s E(q) to disease-related measures.

(A) The module PM1 is a single cluster of 31 proteins. (B) Positive associations of E^(PM1) quintiles to variation (cm²) in visceral adipose tissue (VAT), incident coronary heart disease (inc CHD), type 2 diabetes (T2D), and the metabolic syndrome (MetS), ***P < 1 × 10⁻¹⁰. (C) Overall survival, i.e., with respect to all-cause mortality, was reduced for high E^(PM1) levels (red curve) compared to low E^(PM1) levels (cyan curve). (D) The modules PM6, PM9, and PM10 are members of supercluster II. (E) Inverse association of the E^(PM6) and E^(PM9) to prevalent CHD (prev CHD) and prevalent heart failure (prev HF), ***P ≤ 1 × 10⁻⁹. (F) Reduced overall survival for low E^(PM10) levels (cyan curve) compared to high E^(PM10) levels (red curve). (G) The PM17 is in supercluster IV. (H) Positive association of module’s E^(PM17) to incident CHD and HF as well as prevalent CHD and HF, ***P < 1 × 10⁻¹⁷. (I) Reduced postincident CHD survival as well as overall survival for high E^(PM17) levels (red curve) compared to low E^(PM17) levels (cyan curve). (J) The module PM23 is a member of supercluster V. (K) Positive associations of the module E^(PM23) quintiles to VAT and subcutaneous adipose tissue (SAT), and prevalent CHD, T2D, and MetS, ***P < 1 × 10⁻¹³. Data were analyzed using forward linear or logistic regression or Cox proportional hazards regression, depending on the outcome being continuous, binary, or a time to an event. Kaplan-Meier plots were used to display survival probabilities. The number of proteins per module is denoted at the branches of the dendrogram. Controls are individuals free of the disease in question.

Fig. 3.. The relationship between connectivity of proteins and disease-related measures.

(A) Circle graph of PM1 highlighting the hub protein CMPK1. (B) Positive correlation between within module connectivity (Ki) (x axis) and the absolute value of the effect size of the association of proteins to type 2 diabetes (T2D) (y axis). (C) Positive association of CMPK1 to T2D, and reduced overall survival associated with high serum CMPK1 levels (red curve). (D) Spring graph of PM9, highlighting the hub SYTL4. (E) Positive correlation between Ki and the association of proteins to prevalent heart failure (prev HF). (F) Inverse association of SYTL4 to HF, P < 1 × 10⁻³⁰, and reduced overall survival associated with low serum SYTL4 levels (cyan curve). (G) A circle graph of PM17 highlighting the hub protein SUMO3. (H) Positive correlation between the Ki of proteins and their association to prevalent HF. (I) SUMO3 is positively associated with prevalent HF, and high levels of SUMO3 (red curve) predict reduced survival postincident CHD. (J) A circle graph of the PM23 highlighting the hub ACY1. (K) Positive correlation between the Ki of proteins and their association to both MetS and T2D. (L) Strong positive association of ACY1 to MetS. Network visualization was performed with the igraph package in R (30). Pearson’s r was estimated for correlation between Ki of proteins and their strength of association to disease measures. See Fig. 2 for other relevant statistics.

Fig. 4.. The network-associated pSNP rs704 regulates modules related to immune functions.

(A) A circular Manhattan plot highlights the GWAS results for E^(PM7), revealing a single highly significant association at rs704, a missense variant (NP_000629.3: p.Thr400Met) in VTN. Four other modules within supercluster II were affected by rs704 (table S20). (B) The rs704 variant, and many other linked SNPs in the region, exert a strong cis-acting effect on VTN within the 300-kb genomic region across the VTN gene. The black triangle demonstrates linkage disequilibrium (r²) patterns in the region derived from the AGES cohort data. (C) The bimodal population distribution of the VTN protein is explained by the drastic reduction in VTN levels in individuals homozygous for the rs704 C minor allele. (D) A schematic presentation of the single cis and many trans effects mediated by the rs704 in VTN on serum proteins. (E) Proteins affected by rs704 in trans (and VTN in cis) cluster in modules of immune-related functions (table S20). The percentage denotes the fraction of proteins within a given module regulated by the rs704 locus.

Fig. 5.. Network-associated pSNPs at the APOE and BCHE loci regulate proteins of module PM11.

(A) A circular Manhattan plot for the E^(PM11), revealing two distinct genomic loci at chromosomes 3 (BCHE) and 19 (APOE / TOMM40). The three npSNPs at the APOE/ TOMM40 locus are not correlated (r² = 0). (B) The npSNPs at the two genomic regions exert a strong cis-acting effect on the serum levels of APOE (left) and BCHE (right). (C) The npSNPs at the APOE locus affected APOE in cis and mediated trans effects on 38 proteins, whereas the npSNP rs1803274, a missense variant (NP_000046.1: pAla567Thr) in BCHE, affected BCHE in cis and 20 other proteins in trans (table S20). (D) The distinct npSNPs at the APOE and BCHE loci regulate 88.9% of all proteins that constitute PM11, Fisher’s exact test P = 3 × 10⁻³⁴, as demonstrated in the Venn diagram to the right.The number 3 refers to proteins in PM11 that are not regulated by the npSNPs.