Multiplex proteomics for prediction of major cardiovascular events in type 2 diabetes

Participants provided written informed consent, and the study was conducted according to the Declaration of Helsinki. Ethical permission was granted by the ethics committees of Linköping University (Dnr. 26–05; CARDIPP), Uppsala University (Dnr. 251/90 and 97/329 for ULSAM; Dnr. 00419 and 2005/M-079 for PIVUS; Dnr. 2005:382 for SAVa/PADVa) and the Dante Pazzanese Institute of Cardiology (São Paolo, Brazil).

In CARDIPP, MIVC, SAVa-control and PADVa, type 2 diabetes was defined as a physician diagnosis of type 2 diabetes according to national guidelines (at least two separate fasting glucose levels ≥7.0 mmol/l, or at least two separate HbA_1c concentrations >48 mmol/mol [>6.5%; in MIVC], or prescription of diabetes medication). In ULSAM, type 2 diabetes was defined as HbA_1c >48 mmol/mol (>6.5%), prescription of diabetic medication or a fasting plasma glucose level ≥7.0 mmol/l. In 25 out of 86 participants included in ULSAM, diabetes was diagnosed by elevated fasting glucose alone. In PIVUS, type 2 diabetes was defined as a physician diagnosis, prescription of glucose-lowering medication or a fasting plasma glucose level ≥7.0 mmol/l. In the PIVUS group, diabetes was diagnosed by elevated fasting glucose alone in 21 out of the 98 included participants. Individuals without available fasting frozen plasma or serum samples, or with missing outcome data, were excluded. MACE was defined as a new episode of fatal or non-fatal myocardial infarction (I21 in ICD-10; www.​who.​int/​classifications/​icd/​en/​) or fatal/non-fatal stroke (I60–I63), whichever occurred first, and was from obtained from hospital and death register linkage.

To adjust for established risk factors, we selected all variables included in the Swedish National Diabetes Register (NDR) calculator for 5 year risk of MACE in individuals with type 2 diabetes [13]: sex, systolic blood pressure (mmHg), BMI (kg/m²), current smoking, diagnosis of atrial fibrillation, history of myocardial infarction or stroke, HbA_1c (mmol/mol, %), HDL-cholesterol and total cholesterol (mmol/l), duration of type 2 diabetes (days), microalbuminuria (3–30 mg/mmol urinary creatinine) and macroalbuminuria (>30 mg/mmol urinary creatinine). Additional covariates included current antihypertensive, statin or diabetes medication, LDL-cholesterol (mmol/l) and eGFR (ml min⁻¹ [1.73 m]⁻²), calculated with plasma creatinine according to sex, age and ethnicity). Missing values in covariates were imputed by multivariate imputation by chained equations with predictive mean matching using all other covariates and averaged across five iterations. Imputed values were compared against recorded values to assess for aberrations.

Blood samples were obtained from individuals instructed to fast overnight, and were then spun down and stored as serum (ULSAM) or EDTA plasma samples (all other cohorts) at −70°C until analysis. The Proseek CVD Multiplex 96×96 (Olink, Uppsala, Sweden) measures 92 cardiovascular or inflammatory proteins and four internal control samples using the proximity extension assay method (details on quality control, validation and content of the assay are available in electronic supplementary material [ESM] Table 1 and ESM Methods). It has previously been applied to discovering biomarkers for cardiometabolic traits [16‐18]. In brief, approximately 10 μl of sample were assayed on a 96-well plate, and protein abundance was measured by PCR based on the binding of two specific antibodies for each protein. Log₂-scaled abundance values adjusted for technical variation with internal controls were transformed to a mean of zero and an SD of 1. Proteins with >15% missing values were excluded. Other missing values were imputed by the lower limit of the detection threshold divided by two. The numbers of missing values are given in ESM Table 2. A total of 12 proteins had >15% missing values in at least one cohort and were excluded, leaving 80 proteins for inclusion in the study.

Figure 1 illustrates the study flow chart, and Table 1 lists the baseline characteristics of all participants. The discovery sample (CARDIPP) recorded 71 MACE events in 708 participants over a mean (±SD) of 7.3 ± 1.8 years (range 0.1–9.65). At a 5% FDR, we estimated 80% power to detect an HR of 1.41 per 1 SD change in protein signal. The replication sample combined participants with type 2 diabetes in ULSAM (n = 86; 37 events over 6.8 ± 3.8 years), PIVUS (n = 98; 29 events over 8.1 ± 2.9 years), MIVC (n = 140; 38 events over 2.9 ± 1.2 years), SAVa-control (n = 80; ten events over 4.9 ± 1.6 years) and PADVa (n = 99; 26 events over 4.5 ± 2.0 years). The replication set thus included 503 diabetic individuals, 140 of whom experienced a MACE during 5.2 ± 3.1 years (range 0.01–12.83), with 80% power to detect an HR of 1.28 per SD unit of protein. None of the models violated the proportional hazards assumption (p > 0.05).

In the discovery sample, 35 out of 80 proteins were associated with prospective risk of MACE at a 5% FDR after adjustment for age and sex (ESM Table 3). Eight associations were replicated at <5% FDR in the separate replication sample (ESM Table 4). In order to test for associations between biomarkers and MACE independent of established risk factors, we combined all cohorts and tested the eight replicated biomarkers in models adjusted for cardiovascular risk factors. Figure 2 shows the results for the eight biomarkers. In the fully adjusted models, increased levels of the following were associated with incident MACE: matrix metalloproteinase (MMP)-12 (HR per SD increase in protein abundance 1.31, 95% CI 1.17, 1.47); TNF-related apoptosis-inducing ligand receptor 2 (TRAIL-R2, also known as death receptor 5) (HR 1.44, 95% CI 1.19, 1.74); IL-27 subunit α (IL-27a; HR 1.47, 95% CI 1.21, 1.78); kidney injury molecule (KIM)-1 (HR 1.23, 95% CI 1.10, 1.36); fibroblast growth factor (FGF)-23 (HR 1.20, 95% CI, 1.05, 1.37); TNF receptor (TNFR)-1 (HR 1.17, 95% CI 1.06, 1.28); TNFR-2 (HR 1.37, 95% CI 1.18, 1.59); and protein S100-A12, also known as extracellular newly identified receptor of advanced glycation end products (RAGE)-binding protein (EN-RAGE; HR 1.30, 95% CI 1.14, 1.48). The exclusion of 12 individuals who had had a haemorrhagic stroke (in MIVC, all five individuals with any type of stroke were excluded as subtypes had not been recorded) had little effect on the effect sizes of the eight biomarkers and replicated the same eight proteins as in the main analysis plus three additional ones (ESM Results, ESM Tables 5–7). A sensitivity analysis with additional adjustment for circulating levels of N-terminal pro-brain natriuretic peptide in those cohorts with available measurements resulted in somewhat increased p values, but essentially left the associations between biomarkers and risk of MACE unchanged (ESM Results, ESM Table 8). Correlations between the biomarkers are shown in ESM Table 9.

We trained a baseline GBM model including all NDR variables and an NDR-plus-protein model adding all 80 proteins. In the training set of 136 people who experienced a MACE event and 698 people who did not, discrimination improved from C = 0.738 (95% CI 0.735, 0.740) at baseline to C = 0.825 (95% CI 0.824, 0.827) with added proteins (p for difference, p_diff = 3.33 × 10⁻⁵⁴). Discrimination in the separate test sample of 49 people with and 229 without a MACE event improved from C = 0.686 (95% CI 0.682, 0.689) to C = 0.748 (95% CI 0.746, 0.751; p_diff = 8.48 × 10⁻²¹). Sensitivity and specificity in the test sample for the upper 50th risk percentile were 70.9% and 54.0%, respectively, for baseline, and 79.1% and 55.8% with added proteins. Sensitivity and specificity for the upper 25th risk percentile in the test set were 48.1% and 78.9%, respectively, for baseline, and 53.6% and 80.4% for added proteins. Lasso Cox regression in the training sample selected 20 proteins in addition to the NDR risk factors. The prediction performance in the independent test sample had a C statistic of 0.747 (95% CI 0.653, 0.842). A model that included the eight replicated biomarker proteins discovered in part 1 resulted in a C statistic of 0.736 (95% CI 0.641, 0.829).

We identified eight circulating biomarkers, including four novel ones, for incident cardiovascular events after adjustment for established risk factors. Our results replicate previous findings in individuals with type 2 diabetes of associations of increased levels of MMP-12 [17], FGF-23 [28], TNFR-1 and TNFR-2 [29] with incident MACE. For the other four biomarkers, we found no previous studies of prospective associations with MACE in type 2 diabetes, although all have been implicated in cardiometabolic disease in other settings.

Protein S100-A12 (EN-RAGE), the ligand for RAGE, has been associated with incident type 2 diabetes [30] and risk of coronary heart disease [31]. Interaction between RAGE and EN-RAGE triggers an inflammatory cascade, and it has been shown that expression of protein S100-A12 in vascular smooth muscle cells induces oxidative stress, inflammation and vascular remodelling [32].

KIM-1 is mainly expressed in the apical membrane of the renal proximal tubule, and raised circulating levels of KIM-1 are associated with progressive stages of chronic kidney disease in individuals with type 2 diabetes [33, 34]. Associations between raised plasma levels of KIM-1 and adverse cardiovascular risk factors in the general population have recently been reported [35]. Our results in analyses adjusted for kidney function support a potential role of circulating KIM-1 as a cardiovascular risk marker independent of its association with renal function. Our study cannot address the pathogenic mechanisms or potential causality linking KIM-1 to cardiovascular risk in type 2 diabetes, and future experimental studies are indicated.

TRAIL-R2 is a cell surface receptor for TNF-related apoptosis-inducing ligand (TRAIL), involved in apoptosis. Raised circulating TRAIL-R2 levels have been linked with cerebral atherosclerosis [36] and increased mortality in acute myocardial infarction [37]. Possible mechanisms linking the TRAIL/TRAIL-R2 pathway to atherosclerotic disease involve the endothelial response to cholesterol deposits [37, 38] and the composition of circulating fatty acids, as a study in an Alaskan Inuit population found an association between plasma fatty acid levels and genetic variants of the TRAIL-R2 gene TNFRSF10B [39].

IL-27 has complex pro- and anti-inflammatory effects that include direct modification of CD4⁺ and CD8⁺ T cells, as well as roles in both innate and antibody-mediated immunity [40]. It has been linked, for instance, to type 1 diabetes [41] and improved atherosclerosis in mice [42], yet functional genetic variants of IL27 were not associated with cardiovascular outcomes in a sample of Chinese individuals [43]. The roles of the four new biomarkers in inflammatory pathways point to an important role of the immune system in cardiovascular pathology in type 2 diabetes. Whether the novel biomarkers might serve as treatment targets remains to be assessed in future studies.

The addition of proteins to the variables included in the NDR risk model significantly improved cardiovascular risk prediction. In our test sample, added biomarkers improved discrimination from 68.6% to 74.8%, compared with 72.0% reported in the original publication of the NDR model [13]. The model containing the NDR risk factors plus proteins also improved sensitivity and specificity for the upper half (79.1% and 55.8%, compared with 76.2% and 52.9%, respectively, in the original NDR model [13]) and the upper quarter of predicted risk (53.6% and 80.5%, compared with 51.2% and 77.9%, respectively). Importantly, direct comparisons with the NDR calculator are not possible as we used a different statistical method and study design, as well as a smaller test sample and a somewhat longer follow-up of approximately 6 years. The crucial comparisons are therefore the test set performances in our own sample. Predictor selection with lasso regression retained a subset of 20 proteins in addition to risk factors and achieved a near-identical discrimination performance (C = 0.747) as the model including all 80 proteins. Our results demonstrate that adding proteomics data to known risk factors might aid decision-making for cardiovascular prevention in individuals affected by type 2 diabetes. The protein assay used in this study analyses small sample volumes in under 48 h, making it potentially useful for clinical practice. The accessibility of proteomics platforms is likely to increase in the coming years, and a number of studies have demonstrated how proteomics can discover new biological insights [16‐18].

Clinical decisions about whether more aggressive cardiovascular prevention with newer drugs will benefit individuals with type 2 diabetes are difficult, given the progressively smaller benefits, risk of side effects and treatment costs [7‐9, 11, 14]. In this study, we demonstrate how a multi-biomarker assay can improve risk prediction; future studies in an embedded healthcare setting are indicated to assess the value of ‘-omics’ methods in day-to-day practice. Targeted cardiovascular proteomics might also be useful for streamlining clinical trials of cardiovascular prevention by risk-stratifying participants for cardiovascular prevention, which could lead to improved power to detect clinically meaningful effects and limit expenses [44]. Any application of proteomics with clinical consequences, however, first requires careful validation in future studies.

Strengths of our study include the prospective community samples, a discovery/replication design and the use of a low sample-volume assay with high-specificity antibody doublets. We limited the risk of overfitting by replicating results in a separate random test subsample and averaging across 1000 iterations, but the bootstrapped CIs have to be interpreted with caution, and our model should be replicated in an independent study. The C statistic of the baseline model was somewhat lower than expected, which may have led to overoptimistic results after adding proteins. On the other hand, the C statistic is usually rather insensitive to added predictors, and we showed convincing improvement [45]. Limitations include the moderate sample size, lack of power to assess the components of MACE as separate outcomes, and failure of 12 proteins in quality control because of missing values. Generalisability is limited to middle-aged to elderly adults (30–77 years of age). Analyses accounted for heterogeneity between cohorts and, rather than limiting variability and effective sample size by tightening the inclusion criteria, we attempted to increase external validity by including a broader range of individuals with type 2 diabetes that would reflect clinical reality.