Deciphering the Plasma Proteome of Type 2 Diabetes

Type 2 diabetes is a significant cause of morbidity and mortality, with an estimated worldwide prevalence of 463 million patients, one-half of whom are undiagnosed (1). It is a complex, multifactorial disease characterized by an interplay of both genetic and nongenetic factors that lead to insulin resistance and hyperinsulinemia (1,2). Moreover, type 2 diabetes causes widespread microvascular and macrovascular complications, resulting in significant health care expenditure (1).

The proteomics of type 2 diabetes, the investigation of a set of proteins within different tissues of diabetic animal models, and the comparison of patients with diabetes with healthy control subjects have enabled the discovery of new protein–type 2 diabetes associations (3–5). Examples of associations include adiponectin (3), leptin (5), and insulin-like growth factor–binding protein 2 (IGFBP-2) (4). Of particular clinical interest is the study of type 2 diabetes associations with plasma proteins, which reflect systemic effects and may serve as predictive biomarkers (3,5–7).

The integration of genetic and proteomic knowledge has provided new insight into the pathophysiology of type 2 diabetes. The best example is Mendelian randomization (MR), a method used to infer causality in observational study settings (4,8). Previous MR studies of biomarkers and type 2 diabetes have suggested causal protective roles for proteins like adiponectin, β-carotene, N-terminal proB-type natriuretic peptide, and sex hormone–binding globulin (SHBG) as well as causal harmful roles of delta-6 desaturase and ferritin (7).

Here, we use a highly multiplexed aptamer-based proteomics platform to analyze the associations between prevalent type 2 diabetes and 1,095 plasma proteins in the Cooperative Health Research in the Region of Augsburg (KORA) study. We replicate our results in the independent Nord-Trøndelag Health Study (HUNT) study and investigate associations with incident type 2 diabetes using follow-up data from KORA and HUNT. Moreover, we test the performance of our newly discovered biomarkers to predict incident type 2 diabetes when added to an adapted version of the updated German Diabetes Risk Score (GDRS) (9). We then evaluate these newly identified proteins using the protein-protein interaction resource STRING (10). Finally, we applied two-sample bidirectional MR analysis (11) to assess causality and prioritize the newly discovered relationships.

The KORA study comprises independent samples from Augsburg in southern Germany (12). In the current study, we used a subsample of 1,000 individuals randomly selected from the participants of the KORA F4 survey (N = 3,080, performed 2006–2008) with deep phenotyping data (n = 1,800) (13). Detailed clinical and sociodemographic information was collected. Data from the KORA FF4 survey (performed 2013–2014) represents the 7-year follow-up of KORA F4. The ethics committee of the Bavarian Medical Association (Berlin, Germany) reviewed and approved the study, and all participants gave written informed consent.

HUNT is a prospective population-based cohort from Nord-Trøndelag County in Norway (14). We used the HUNT3 survey (n = 1,117 with proteomics measurements, performed 2006–2008) for the validation of the KORA study results. The HUNT study collected detailed sociodemographic and clinical information. We used linked primary care and hospital registries for information on diabetes status at 9 years follow-up. All study participants provided written informed consent.

Proteins were measured in fasting and nonfasting plasma samples in KORA and HUNT, respectively, using the SOMAscan platform as described previously (13,15). In summary, plasma and bead-coupled aptamers, each of which has a high affinity toward a specific protein, were incubated. After washing steps, bead-bound proteins were biotinylated, and complexes comprising biotinylated target proteins and fluorescence-labeled aptamers were photocleaved off the bead support and pooled. Following recapture on streptavidin beads and further washing steps, aptamers were eluted and quantified as a proxy to protein concentration by hybridization to custom arrays of aptamer-complementary oligonucleotides. On the basis of standard samples included on each plate, the resulting raw intensities were processed using a data analysis workflow that included hybridization normalization, median signal normalization, and signal calibration to control for interplate differences (16). Raw intensities are reported in relative florescence units.

In KORA, one sample failed SOMAscan quality control, leaving 999 samples for analysis. Of the 1,129 SOMAmer probes (SOMAscan assay version 3.2), 29 failed SOMAscan quality control. We also removed the five probes recommended by the SOMAscan assay change log issued on 22 December 2016, leaving 1,095 probes for analysis. For replication, we used the HUNT probes that passed quality control.

In KORA, type 2 diabetes was defined as self-reported disease validated by the responsible physician or medical chart review or as current use of glucose-lowering medication. All participants without known diabetes were assigned to receive a standard 75-g oral glucose tolerance test (3). Prevalent type 2 diabetes refers to participants with the disease at the time of blood sample collection, and incident refers to those developing type 2 diabetes after that time point within a 7- and 9-year follow-up period in KORA and HUNT, respectively.

In HUNT, prevalent type 2 diabetes was self-reported, which we validated using clinical data from hospitals and primary care registries using the ICD-10 code E11 and the International Classification of Primary Care, Second Edition, code T90. We identified incident cases of type 2 diabetes from the same registries using identical codes.

We classified participants of both cohorts who participated in leisure time physical activity for at least 1 h/week as physically active (more details are available in the Supplementary Material). Current hypertension was defined in KORA as having a systolic blood pressure ≥140 mmHg, diastolic blood pressure ≥90 mmHg, and/or use of antihypertensive medication. In HUNT, we used hospital and primary care data and ICD-10 codes I10–I15 and International Classification of Primary Care, Second Edition, codes K86 or K87 to identify participants with hypertension.

Drugs were assessed in KORA by asking the participants to bring the packages of their medication and supplements with them to their study center visit. Using database software (17), medications were identified using Anatomical Therapeutic Chemical codes, medication identifier bar code, or product name.

Preprocessing of the quality controlled SOMAscan data was the same in both cohorts and involved log₂ transformation and (0, 1) standardization by subtracting the per-cohort mean and dividing by the per-cohort SD to allow easier interpretation of the odds ratios (ORs) per SD of the protein.

Using logistic regression, we ran two proteome-wide analyses in KORA: associations between proteins with prevalent type 2 diabetes and with incident type 2 diabetes. For each of the two outcomes, we ran one model per protein (i.e., 1,095 models per outcome) and adjusted for the potential confounders age, sex, BMI, smoking status, and current hypertension at baseline. We then replicated the results in HUNT using the same model. We excluded participants from both cohorts with missing values for the confounders, which led to the sample sizes of 993 and 940 for KORA and HUNT, respectively. For the analysis with incident type 2 diabetes, we further excluded all participants with prevalent type 2 diabetes, resulting in sample sizes of 881 and 794 for KORA and HUNT, respectively. We used the false discovery rate (FDR) Benjamini-Hochberg method separately for the outcomes to account for multiple testing. An association was considered statistically significant at FDR <0.05.

We replicated significant results in HUNT using the same model. We considered proteins replicated at FDR <0.05, with FDR calculated on the basis of the number of significant proteins in KORA. To examine whether antidiabetic drug intake influenced the replicated associations, we ran sensitivity analyses by including the drugs of interest as confounders one at a time.

The candidate proteins were processed through the Pharos (18) platform, experimental Gene Ontology (19) terms, Kyoto Encyclopedia of Genes and Genomes (KEGG) (20) pathways, human disease association data from the GWAS Catalog (21), Online Mendelian Inheritance in Man (22), and the text-mining DISEASES platform (23) as well as through phenotype data from the corresponding mouse ortholog knockouts (24). We mined these resources for data on the potential associations between the candidate proteins and type 2 diabetes.

We applied a biomarker discovery strategy to investigate whether proteins significantly associated with incident type 2 diabetes in KORA could be used for prediction of incident type 2 diabetes. These were 10 proteins, of which 1 failed quality control in HUNT. We used KORA as a training data set and HUNT as a test data set and used an adapted version of GDRS, a diabetes risk score that was trained in 21,845 participants of the European Prospective Investigation into Cancer and Nutrition Potsdam (EPIC-Potsdam) study with a mean follow-up time of 7 years, as a benchmark (9). More details on the GDRS are available in the Supplementary Material.

Adaptation of the GDRS was necessary because some of the variables were missing from one or both cohorts. We defined smoking status using only information on current and former smoking per se without regard to the number of cigarettes smoked. We used the average of the original GDRS score weights for each smoking category to represent our combined categories (former: [15 + 45] / 2 = 30; current: [23 + 77] / 2 = 50). Family history of diabetes was defined in KORA as having at least one parent or sibling with diabetes and in HUNT, as having at least one parent, sibling, or child with diabetes. We calculated the risk of a positive family history by averaging the original GDRS of having one parent, both parents, or at least one sibling with type 2 diabetes ([56 + 106 + 48] / 3 = 70). Our final adapted GDRS was calculated as follows: 5.1 × age in years + 7.6 × waist circumference in cm – 2.7 × height in cm + 47 × hypertension status − 2 × physical activity (at least 1 h/week) + 30 × former smoking + 50 × current smoking + 70 × family history of type 2 diabetes.

We tested prediction performance of all models using the receiver operating characteristic area under the curve (AUC) and applied the DeLong test to compare AUCs of nested models. First, we added the proteins to the adapted GDRS model and applied the least absolute shrinkage and selection operator (LASSO) (25) for model selection in KORA. LASSO shrinks the sum of the absolute values of the regression coefficients, forcing some to be set to 0, thus performing a form of model selection. The LASSO λ was chosen by cross-validation using the squared error for Gaussian models. The GDRS was calculated in each cohort and used as a score that was fixed by setting its penalty factor to 0 to prevent any shrinkage by LASSO. We then compared the performance of the LASSO protein-extended GDRS model to the adapted GDRS model. We assessed the calibration of the LASSO-selected model using calibration plots (26). Moreover, we tested the performance of the proteins as single predictors on top of the adapted GDRS model.

We attempted to infer causality of the replicated proteins associated with type 2 diabetes by applying two-sample bidirectional MR. Figure 1 shows the summary of the pipeline for the causal inference analysis. In summary, we extracted single nucleotide polymorphisms (SNPs) as instrumental variables (IVs) from published genome-wide association study (GWAS) summary statistics of European ancestry if they passed the Bonferroni threshold of P < 5e-8. We extracted the IVs from the meta-analysis of type 2 diabetes GWAS studies by Xue et al. (27) (N = 455,607) and the GWAS studies of SOMAscan-measured proteins by Sun et al. (28) (N = 3,301), Suhre et al. (13) (N = 1,000), and Emilsson et al. (29) (N = 5,457) for proteins. We identified ambiguous palindromic SNPs, defined as SNPs with A/T or G/C alleles and an effect allele frequency of ∼0.5 using the cutoff points defined by the TwoSampleMR package in R (30). We replaced these with a proxy SNP, defined as a SNP with r² >0.85 with the SNP in question, when available, or excluded them from further analyses (31). We then clumped the SNPs, which implies removing SNPs in linkage disequilibrium with the lead SNP using the r² cutoff 0.001. We did not manually prune the final list of IVs. Furthermore, IVs selected for proteins needed to be in cis (i.e., within 1 Mb of the protein-coding gene as per Human Genome Assembly GRCh37.p13).

We proceeded to extract the results of these IVs or of one of their proxies from the outcome’s GWAS. For proteins, priority was given to results from Sun et al. (28) because of the larger sample size, followed by Suhre et al. (13) dependent on availability.

We used the Wald ratio to check for causality (32). In cases of more than one IV, we used the random effects model of the inverse variance–weighted meta-analysis to combine the Wald ratio estimates of all IVs (8,32). For sensitivity analyses, whenever there was more than one IV, we ran the MR‐Egger regression model to look for horizontal pleiotropy in our causal models (33) and leave-one-out analysis and forest plots to identify outliers among these IVs that would be driving the results in a certain direction and examined scatter plots to check for outliers.

Analytical steps are summarized in Supplementary Fig. 1. All analyses were done using R version 3.5.1 software (The R Foundation for Statistical Computing). For MR analysis, the TwoSampleMR package of R version 0.4.22 was used (30).

Informed consents given by KORA study participants do not cover data posting in public databases. However, the KORA data are available given approval of online requests at the KORA Project Application Self-Service Tool (https://epi.helmholtz-muenchen.de). The HUNT data can be accessed given approval of applications to the HUNT Research Centre (https://www.ntnu.edu/hunt/data). The data used in the MR analysis are publicly available and can be accessed through https://cnsgenomics.com/content/data (Xue et al. [27]), https://www.phpc.cam.ac.uk/ceu/proteins (Sun et al. [28]), https://metabolomics.helmholtz-muenchen.de/pgwas/ (Suhre et al. [13]), and www.sciencemag.org/content/361/6404/769/suppl/DC1 (Emilsson et al. [29]). Example code for the analytic steps of the article can be accessed at https://github.com/maelhadad/T2D_SOMAscan_Proteomics.

Table 1 and Supplementary Table 1 show the baseline characteristics of both cohorts and their follow-up subsets, respectively. HUNT participants were on average older and comprised more men.

The proteome-wide analysis with prevalent type 2 diabetes yielded 85 FDR-significant proteins (Supplementary Table 2), of which 24 successfully replicated in HUNT (Table 2 and Fig. 2A). Of these, osteomodulin was most strongly associated (on the basis of KORA P value) with an OR-per-SD increase in protein level of 0.61 (95% CI 0.47–0.77) in KORA and of 0.65 (0.53–0.79) in HUNT. Among the positively associated proteins, peptide YY (PYY) had the strongest association (1.34 [1.1–1.62] in KORA and 1.58 [1.32–1.92] in HUNT).

To assess whether the proteome panel was associated with future type 2 diabetes, we performed a proteome-wide analysis with incident type 2 diabetes using the same model, which yielded 10 FDR-significant protein associations (Supplementary Table 3). Of these, aminoacylase-1, growth hormone receptor, and IGFBP-2 replicated in HUNT (Table 3 and Fig. 2B). Adiponectin failed quality control in HUNT, and thus, replication was not possible. Among the replicated proteins, aminoacylase-1 showed the strongest association (OR 1.78 [95% CI 1.34–2.37] in KORA and 1.6 [1.26–2.05] in HUNT). Interestingly, aminoacylase-1 overlapped between the replicated results of both prevalent and incident type 2 diabetes.

Additionally, we assessed the concordance of the effect estimates across the cohorts. Of 85 KORA FDR-significant proteins associated with prevalent type 2 diabetes, only 7 had different effect directions, but none of these was nominally significant in HUNT (Fig. 3A). For incident type 2 diabetes, two proteins showed opposite effect directions, with neither of these reaching nominal statistical significance (Fig. 3B).

To assess the overlap between our results and known type 2 diabetes associations, we compared our results to gene-based results described by Xue et al. (27). Furthermore, we compared our replicated proteins with protein lists of interest published by the Human Diabetes Proteome Project, namely the 1,000 diabetes-related proteins, the human islet of Langerhans proteome, the rodent β-cell proteome, and the human blood glycated proteome (34). Of the 26 unique replicated proteins, 18 overlapped with at least one list. Eight proteins have not been previously found to be related to type 2 diabetes (Supplementary Table 4).

Supplementary Table 5 shows information extracted from Pharos for our replicated proteins. α-l-Iduronidase, cathepsin A, and cathepsin Z shared the same lysosomal pathway association according to KEGG (20).

None of the replicated protein-incident type 2 diabetes associations showed loss of significance when adjusting for any of the investigated drugs. On the other hand, three of the replicated associations with prevalent type 2 diabetes lost statistical significance when adjusting for antidiabetic medication intake (Supplementary Table 6 and Supplementary Fig. 2). All the associations retained the same direction of effect apart from PYY, which showed an opposite effect after adjusting for antidiabetic medication and, more specifically, metformin.

Starting with the nine proteins associated with incident type 2 diabetes in KORA available in HUNT, we evaluated whether a subset of them selected using LASSO would improve the predictive performance of the adapted GDRS benchmark model (9). LASSO selected five proteins, namely transforming growth factor-β receptor type 3 (TGFbR3), tartrate-resistant acid phosphatase type 5, pappalysin-1, afamin, and scavenger receptor cysteine-rich type 1 protein M130 (sCD163). The LASSO-selected protein-enhanced model showed improvement in both KORA and HUNT (GDRS protein-extended AUC 0.84 [95% CI 0.79–0.89] and 0.67 [0.61–0.72]; GDRS-only AUC 0.77 [0.71–0.83] and 0.66 [0.60–0.72], respectively); however, according to the Delong test, the AUC improvement in HUNT was not statistically significant (P = 0.72) (Supplementary Fig. 3). The calibration plot of the LASSO-selected model in HUNT yielded an intercept of 0.23 and a slope of 0.53 (Supplementary Fig. 3). The intercept of the calibration plot examines the difference of means of predicted and observed risk. In HUNT, it is >0, thus showing higher observed type 2 diabetes cases in HUNT as those predicted. This could be attributed to longer follow-up in HUNT (9 years vs. 7 years in KORA) and to the fact that HUNT is older than KORA and therefore has more cases of type 2 diabetes. The slope of calibration is 0.53 in HUNT, which indicates a possible overfitting of the model or the need for coefficient shrinkage in HUNT that could also be attributed to the heterogeneity between the study populations in terms of patient characteristics and outcome definition. The training data set used gold standard screening to define type 2 diabetes, where HUNT did not apply a similar definition and would therefore have hidden cases and measurement error. Therefore, the outcome being predicted for HUNT (and defined by KORA) is slightly different from the outcome observed.

We further tested the performance of individual proteins as predictors of incident type 2 diabetes in KORA and validated our models in HUNT (Supplementary Fig. 4). The following proteins showed relatively similar performance in both cohorts: aminoacylase-1 (KORA AUC 0.78 [95% CI 0.73–0.84]; HUNT AUC 0.71 [0.65–0.77]), growth hormone receptor (KORA AUC 0.77 [0.71–0.83]; HUNT AUC 0.70 [0.64–0.76]), and IGFBP-2 (KORA AUC 0.78 [0.72–0.84]; HUNT AUC 0.73 [0.68–0.79]).

Using up to 120 SNPs as genetic instruments, we investigated whether type 2 diabetes had a causal effect on the 26 replicated proteins from both the prevalence and the incidence analyses (Supplementary Table 7 and Supplementary Fig. 5). For cathepsin Z (MR inverse variance–weighted β = 0.13; P = 2.00e-03) and renin (0.08; P = 3.15e-02), a nominally significant causal effect of prevalent type 2 diabetes was observed, each with the same direction of effect as its observational results. MR-Egger analyses to test for the presence of horizontal pleiotropy showed no significant results for either protein (intercept P = 0.17 and 0.1 for cathespin Z and renin, respectively). Tests and plots to check for outliers in the IVs showed no significant aberrations (Supplementary Figs. 6 and 7).

We also ran MR to investigate whether any of the proteins had a causal effect on type 2 diabetes. We analyzed 13 proteins for which we found independent cis-acting IVs (Supplementary Table 8 and Supplementary Fig. 5). We observed a nominally significant causal effect of SHBG on type 2 diabetes, with the same direction of effect as its observed association (MR Wald β = −0.09; P = 2.95e-02). None of the associations for either direction survived Bonferroni multiple testing correction.

We report a proteome-wide analysis of type 2 diabetes in KORA and replication in HUNT using aptamer-based affinity proteomics. Our analysis yielded 26 unique replicated significant protein associations. Of these, 24 replicated exclusively with prevalent type 2 diabetes, 2 replicated exclusively with incident type 2 diabetes, and aminoacylase-1 replicated with both.

Aminoacylase-1 is a zinc-dependent peptidase involved in amino acid metabolism (35). The protein has not been described in the context of type 2 diabetes before but has been reported to be overexpressed in obese liver tissue, thus linking it to obesity and inflammation (36). A further study found aminoacylase-1 to be downregulated in obese omental fat, which the authors hypothesized to be due to adipocyte dysfunction caused by obesity (35). Moreover, aminoacylase-1 is associated with arginine production according to KEGG (20). Plasma levels of arginine were found to be higher in patients with type 2 diabetes (37).

In addition to aminoacylase-1, incident type 2 diabetes results included an inverse association with IGFBP-2 and a positive association with growth hormone receptor. IGFBP-2 was reported to have type 2 diabetes protective effects and has been shown to reverse hyperglycemia in insulin and leptin deficiency (38). These associations highlight the role of the growth hormone axis in the early pathophysiology of type 2 diabetes. Both growth hormone and IGF-I are known to play roles in the insulin receptor cascade, leading to insulin resistance (39).

The analysis of prevalent type 2 diabetes confirmed previously known proteomic associations like gelsolin (40), renin (41), SHBG (42), and hepatocyte growth factor receptor and revealed promising new candidate proteins, including osteomodulin, matrilin-2, Wnt inhibitory factor-1 (WIF1), tumor necrosis factor–inducible gene 6 protein (TNFAIP6), cerebral dopamine neurotrophic factor (CDNF), RGM domain family member B, TGFbR3, and SLIT and NTRK-like protein 5, which were downregulated in type 2 diabetes cases, and lysosomal protective protein, galectin-3 binding protein (LGALS3BP), and PYY, which were upregulated.

Our results overlap and complement results of mass spectrometry studies on obesity. Plasma levels of apolipoprotein B, LGALS3BP, and SHBG were found to be altered by sustained weight loss (43) and gastric bypass surgery–induced weight loss (44), with the latter affecting also plasma protease C1 inhibitor, complement C2, and gelsolin.

New protein associations with prevalent type 2 diabetes included proteins previously reported in association with complications of type 2 diabetes. Increased circulating levels of LGALS3BP were linked to nonalcoholic fatty liver disease (45) and acute venous thrombosis (46), and TGFbR3 was reported to be associated with diabetic nephropathy (47). TNFAIP6 and CDNF were shown to have protective effects, while WIF1, TGFbR3, and PYY were reported to have harmful effects, in the development and progress of cardiovascular atherosclerotic diseases (48–52). Along this line, members of the complement family like plasma protease C1 inhibitor and complement C2 were downregulated and upregulated, respectively, in our results, and proteins from the renin-angiotensin and kallikrein-kinin systems included the upregulated renin and downregulated kallikrein-7.

Although our study cohorts were different regarding the fasting status of their samples, most proteins (78 of 85 for prevalent and 8 of 10 for incident type 2 diabetes) showed concordant effects between cohorts, while none of the nonconcordant proteins were statistically significant in the replication (Fig. 3). Nonetheless, fasting has significant metabolic consequences that are expected to be reflected in the plasma proteome and could have contributed to nonreplication in HUNT. However, there are multiple other potential explanations for the nonreplication, perhaps differing from one protein to another. Importantly, while plasma protein levels differ between fasting and nonfasting samples, this does not necessarily match the variance in the protein levels caused by the disease status. As such, disease-related variance would still be apparent despite differences in fasting status. For example, some of our examined proteins were reported to show differences in their levels according to fasting status, like SHBG (53), PYY (53), and soluble CD163 (54), yet their associations with disease status were replicated in our study. Of the proteins that failed replication in HUNT, MMP2 (53) and pappalysin-1 (55) have been found to be affected by food intake. However, their effect sizes were similar in both cohorts, suggesting that fasting status may not be the primary reason for nonreplication for most proteins.

Sensitivity analyses into the potential effect of drugs on the type 2 diabetes–plasma protein associations showed loss of significance of some associations after adjusting for antidiabetic medication. The effect direction of all the resultant associations remained the same except for PYY, which showed a change of direction after adjusting for metformin intake; however, because its effect estimate was not significant after adjustment, it is difficult to draw any conclusions from this.

Additionally, we evaluated the significant proteins’ ability to predict incident type 2 diabetes. The protein-extended models showed improved performance over the adapted GDRS benchmark model (9) in both the KORA discovery and the HUNT replication, although the improvement was very small and not statistically significant (P = 0.72) for the latter. Moreover, we tested the performances of individual proteins on top of the adapted benchmark model. The best performances in the replication cohort came from aminoacylase-1, growth hormone receptor, and IGFBP-2, each of which achieved approximately equal performance in HUNT compared with KORA, results that warrant validation in clinical trials using commercially available ELISA kits. Because the KORA samples were taken from individuals in a fasting state (≥8 h) and HUNT samples were taken nonfasting, these results seem to indicate that fasting status is largely irrelevant with regard to type 2 diabetes prediction for these candidate biomarkers. However, fasting may potentially be relevant for other markers, since the AUC was much smaller in HUNT compared with KORA for some of the other measured biomarkers in combination with the GDRS.

Our investigations into the causal framework governing the relation between plasma proteins and type 2 diabetes showed suggestive harmful causal effects of SHBG on type 2 diabetes. SHBG has been previously reported to be associated with type 2 diabetes (42) and may be implicated in the development of insulin resistance (42). We demonstrated it to be negatively associated with type 2 diabetes, a causal direction suggested by the MR analysis as well.

Causal inference analysis showed suggestive causal effects of type 2 diabetes on both cathepsin Z and renin. In line with previous observations, we demonstrated renin to be positively associated with type 2 diabetes in both observational and MR analysis results (41). The association is an indicator of the upregulated renin-angiotensin-aldosterone system, which is activated in obesity and type 2 diabetes, thus contributing to cardiovascular disease complications (41,56). Cathepsin Z is a member of the peptidase C1 family that plays a role in lysosomal function, which might explain its connection to diabetes through β-cell failure driven by lysosomal degradation (57).

We applied a high-throughput proteomics platform on samples from population-based cohorts for our analyses, which enabled us to test a large number of proteins with a wide concentration range and to generalize our results to our samples’ respective populations. We used samples from plasma, which is easily accessible and is the usual medium of biomarkers. Additionally, the plasma proteome reflects on the levels of proteins originating from a broad range of tissues, thus giving us insight into systemic pathways. Finally, we were able to test for the causal relationship in both directions using publicly available data on genetic associations with both type 2 diabetes and proteins.

We are aware of several limitations to our study. First, aptamer-based proteomics is susceptible to potential probe cross-reactivity and nonspecific binding (28,29). However, we verified that none of the proteins identified have been flagged for such issues (validation data presented in Supplementary Material, Supplementary Table 10, and Supplementary Figs. 10 and 11). Because of the lack of oral glucose tolerance test data in HUNT, the rigorous definition of type 2 diabetes used in KORA could not be extended, and the discrepancy in fasting status between the cohorts may have contributed to the limited replication of our results. Our prediction models do not reflect the dynamic changes in the proteome, which would require a more detailed investigation. This is also true for the MR results, which reflect the lifelong genetic risk rather than point change in single protein levels in relation to disease status. Although, there is an overlap between the participants of the genetic data sets used for type 2 diabetes and proteins through KORA, none of the associations tested using such data were significant.

Our proteome-wide analysis of type 2 diabetes replicated known associations and revealed novel candidate proteins. Associations with incident type 2 diabetes included aminoacylase-1, which overlapped with prevalent type 2 diabetes associations. New associations with prevalent type 2 diabetes included TNFAIP6, CDNF, WIF1, TGFbR3, and PYY, all of which are believed to play a role in the development of cardiovascular complications, like atherosclerosis. MR suggested a causal role of SHBG on type 2 diabetes, which is in line with previous observational and MR analysis results. It also suggested a causal effect of type 2 diabetes on cathepsin Z and renin, both of which are known to play a role in type 2 diabetes complications. Our results offer insight into proteins involved in the pathogenesis of type 2 diabetes and its complications, proteins that could be valuable drug targets for all levels of prevention.

Funding. The KORA study was initiated and financed by the Helmholtz Zentrum München – German Research Center for Environmental Health, which is funded by the German Federal Ministry of Education and Research and by the State of Bavaria. This work was supported by Deutsche Forschungsgemeinschaft grant WA 4081/1-1 and by the Bundesministerium für Bildung und Forschung within the framework of the European Union Joint Programming Initiative “A Healthy Diet for a Healthy Life” (DIMENSION grant number 01EA1902A). This work was further supported by the Biomedical Research Program at Weill Cornell Medicine in Qatar, a program funded by the Qatar Foundation. The HUNT3 study was funded by the Norwegian Ministry of Health, Norges Teknisk-Naturvitenskapelige Universitet, Norges Forskningsråd, Helse Midt-Norge, the Nord-Trøndelag County Council, and the Norwegian Institute of Public Health. M.A.E. is supported by the Deutsches Zentrum für Herz-Kreislaufforschung. T.I.O. is supported by National Institutes of Health grants U24-CA-224370, U24-TR-002278, and U01-CA-239108. K.S. is supported by Qatar National Research Fund grant NPRPC11-0115-180010.

Duality of Interest. C.J. has received personal fees for research consultancy work from Pfizer and Bayer outside the submitted work. T.I.O. has received honoraria or consulted for Abbott, AstraZeneca, Chiron, Genentech, Infinity Pharmaceuticals, Merz Pharmaceuticals, Merck Darmstadt, Mitsubishi Tanabe, Novartis, Ono Pharmaceuticals, Pfizer, Roche, Sanofi, and Wyeth outside the submitted work. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. M.A.E. analyzed the data, interpreted the results, and wrote and revised the manuscript. M.A.E. and M.W. designed the study. R.W., V.G.-D., and T.I.O. helped with the analyses. C.J., C.H., R.W., C.G., P.M., H.G., J.G., W.R., C.v.T., S.M.H., W.K., M.F.S., K.S., B.T., K.H., A.P., and M.W. were involved in the data collection, data management, and preparation of their respective cohorts. All authors contributed to the writing of the article, critically reviewed it, and approved the final version for submission. M.A.E. and M.W. are the guarantors of this work and, as such, had full access to all the data in the study and take full responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. Parts of this study were presented in abstract form at the AHA Epidemiology and Prevention–Lifestyle and Cardiometabolic Health 2020 Scientific Sessions, Phoenix, AZ, 3–6 March 2020. This study has been published previously in abstract form in the following publication: Elhadad MA, Jonasson C, Huth C, et al. Deciphering the plasma proteome of type 2 diabetes. Circulation 2020;141(Suppl. 1):A21.