Targeted Proteomics Reveals Functional Targets for Early Diabetes Susceptibility in Young Adults

Abstract

Background:

The circulating proteome may encode early pathways of diabetes susceptibility in young adults for surveillance and intervention. Here, we define proteomic correlates of tissue phenotypes and diabetes in young adults.

Methods:

We used penalized models and principal components analysis (PCA) to generate parsimonious proteomic signatures of diabetes susceptibility based on phenotypes and on diabetes diagnosis across 184 proteins in >2000 young adults in the Coronary Artery Risk Development in Young Adults study (CARDIA; mean age 32 years, 44% women, 43% Black, mean BMI 25.6+/−4.9 kg/m²), with validation against diabetes in >1800 individuals in the Framingham Heart Study (FHS) and Women’s Health Initiative.

Results:

In 184 proteins in >2000 young adults in CARDIA, we identified two proteotypes of diabetes susceptibility—a pro-inflammatory fat proteotype (visceral fat, liver fat, inflammatory biomarkers) and a muscularity proteotype (muscle mass), linked to diabetes in CARDIA and WHI/FHS. These proteotypes specified broad mechanisms of early diabetes pathogenesis, including trans-organ communication, hepatic and skeletal muscle stress responses, vascular inflammation and hemostasis, fibrosis, and renal injury. Using human adipose tissue single cell/nuclear RNA-seq, we demonstrate expression at transcriptional level for implicated proteins across adipocytes and non-adipocyte cell types (e.g., fibro-adipogenic precursors, immune and vascular cells). Using functional assays in human adipose tissue, we demonstrate association of expression of genes encoding these implicated proteins with adipose tissue metabolism and inflammation, and insulin resistance.

Conclusions:

A multi-faceted discovery effort uniting proteomics, underlying clinical susceptibility phenotypes, and tissue expression patterns may uncover potentially novel functional biomarkers of early diabetes susceptibility in young adults for future mechanistic evaluation.

Introduction

Diabetes is an end-result of complex, subclinical metabolic remodeling over decades. Given the impact of prevention on downstream morbidity, identifying susceptibility to diabetes through functional “omic” biomarker analyses (e.g., proteomics, metabolomics, genomics) has been a mainstay. While these efforts culminated in seminal discoveries (e.g., TCF7L7 gene; branched chain/aromatic amino acids), most studies involved older Europeans (later along subclinical phase of diabetes development) and relied on clinical diabetes diagnosis to focus discovery^1-6. Life-course epidemiology of diabetes underscores the importance of adiposity and inflammation in its early pathogenesis, potentially accounting for why adjustment for body mass index (BMI) attenuates association of biomarkers with diabetes in large cohorts⁵. Nevertheless, it is increasingly clear that studying the biology of tissue-based physiologies (e.g., adipose tissue) relevant to diabetes early in its course may uncover novel, mechanistically relevant, modifiable biomarkers of disease⁷. A key next step in this integrative approach is tying what is found in circulation (proteomic or transcriptional) with what may be occurring at a transcriptional level in adipose tissue, specifically in pro-inflammatory adipose-resident cell types.

Here, we use a multi-dimensional approach to embark on this integrative approach. First, we prioritize functional diabetes biomarkers (proteins) in >2000 young adults with precise measures of diabetes susceptibility (e.g., fat, liver, muscle structure and inflammation) over two decades of follow-up, followed by large cohort clinical validation. We follow these clinical associations with cell-specific adipose tissue expression of gene targets with phenotypes of adipose tissue metabolism and inflammation. This approach unites clinical proteomics and tissue-based functional studies and prioritizes novel targets relevant to early diabetes pathogenesis applicable in young adults at high lifetime risk. Ultimately, our goal was to offer a new translational roadmap to understand metabolic risk early in its development.

Methods

Data underlying this study is available from CARDIA (www.cardia.dopm.uab.edu), WHI (https://www.whi.org), or FHS (www.framinghamheartstudy.org) studies directly. Tissue data from our bulk transcriptomics has been deposited at GEO under the accession number GSE25402. Sequencing data from single cell/single nuclear analyses is sourced as in the parent manuscript⁸. The analysis was conducted under IRB approval from the Vanderbilt University Medical Center and University of Michigan or by investigators at parent study sites (FHS or WHI). Full methods are available in the Supplemental Material.

Results

Clinical characteristics

The flow of our study is in Figure 1. The characteristics of the CARDIA sample are shown in Supplementary Table I, and WHI and FHS characteristics are shown in Supplementary Table II. At the time of proteomics, CARDIA was a young adulthood cohort (mean age 32 years, 44% women, 43% Black) with average BMI 25.6±4.9 kg/m² (≈15% categorized as obese) and average fasting glucose 89.9±7.3 mg/dl. WHI and FHS offered the opportunity for validation in older samples at higher risk: participants in the FHS and WHI were significantly older (FHS: mean 69 years; WHI: mean 81 years), with average BMI across both studies in the overweight range, and overall higher cardiometabolic risk (by blood pressure, lipid and glucose assessment). Characteristics of 56 women in the human adipose tissue functional studies (stratified by obesity status) is shown in Supplementary Table III. As reported previously, individuals with obesity had a greater per-adipocyte inflammatory response (by CCL2, TNFa, and IL6 secretion), more adipocyte hypertrophy, and higher insulin resistance (as noted by higher HOMA-IR and lower insulin-stimulated lipogenesis) compared with non-obese subjects.

Figure 1:

Study scheme.

Identifying unique diabetes susceptibility proteotypes in young adulthood

Our distinct clinical outcome-driven (hereafter referred to as “outcome-driven") and phenotype-driven approaches allowed us to compare traditional discovery efforts identifying proteins associated with clinical diabetes (of heterogeneous origin) against a potentially more sensitive approach utilizing intermediate endophenotypes. The result of the outcome-driven approach (regression against diabetes itself) is shown in Figure 2 (regression summaries in Online-Only Supplementary Data File). At a median follow-up 23 years (IQR 18-23; 238 cases of incident diabetes), our results move beyond reproducing already demonstrated associations in older cohorts (e.g., IGFBP-2, leptin, PAI, PON3) to uncover several proteins not widely described at this early stage in diabetes pathogenesis (detailed ontology/mechanism in Table 1), including proteins implicated in trans-organ communication (e.g., FABP4, GDF-15), hepatic and skeletal muscle stress responses (e.g., FGF-21, PRSS8), vascular inflammation and hemostasis (e.g., soluble U-PAR, t-PA, IL-1/IL-6, CCL3, MCP-1), fibrosis (e.g., MMP7, Gal-3), and renal injury (KIM1), among others. Notably, unlike previous studies in this space⁵, a substantial fraction was significant after adjustment for fasting glucose, demographics, BMI, and parental history of diabetes (Figure 2C).

Figure 2:

Outcome-driven approach. Panel A shows a volcano plot of proteins associated with incident diabetes in Cox regression, adjusted for age, sex, and race. Panel B shows the proteins significant at a 5% false discovery rate (FDR) from Cox regressions in Panel A, sorted from the largest negative to the largest positive standardized beta coefficient. Text labels on each bar indicate standardized hazard ratios. Positive beta coefficients are shown in red while negative beta coefficients are in blue. Panel C shows the number of proteins significantly associated (at a 5% FDR) with incident diabetes in (1) unadjusted, (2) age, sex, and race adjusted, and (3) fully adjusted models (adjustments in text). The dashed line indicates proteome-wide significance with Bonferroni correction.

Table 1.

Disease ontology and suspected mechanisms for selected proteins prioritized by association studies in young adults. This table is not meant to reflect a comprehensive curation of proteins identified by this work.

Protein	Potential mechanisms and diabetes association
Insulin-like growth factor binding proteins (IGFBP-1, −2)	Metabolic signaling; circulating IGFBP-1 implicated in improving insulin sensitivity via multiple mechanisms (effects on NO signaling9; beta-cell expansion10). Circulating IGFBP-2 is leptin-regulated, can reverse diabetes in murine models11, and increases with bariatric surgery12
CD163	Inflammation; Haptoglobin-hemoglobin receptor; soluble CD163 is released into circulation during LPS-mediated macrophage activation13; high expression in human adipose tissue macrophages14, related to diabetes in older individuals15
Paroxonase 3 (PON3)	Mechanism unclear, possible mitochondrial stress; protective against diabetes in older individuals2, 3; suspected role in attenuating mitochondrial oxidative stress, dysfunction, and apoptosis16; liver PON3 expression increased by GLP-1 receptor agonist17
Growth differentiation factor-15 (GDF-15)	Metabolic signaling and inflammation; expressed during oxidative stress in adipocytes, in proportion to body fat, insulin resistance, and inflammation18; association with diabetes appears more significant in age under 55 years19
Soluble urokinase plasminogen activator receptor (U-PAR)	Inflammation; associated with diabetes20; most studies in renal dysfunction and diabetes complications
Leptin (LEP)	Metabolic signaling; adipokine central to appetite regulation and fuel substrate utilization; leptin resistance is a well-known metabolic predisposition to diabetes
Adrenomedullin (ADM)	Inflammation; produced by adipocytes21, can modulate post-glucose pancreatic insulin secretion22, and regulates IL-6 expression23
Fibroblast growth factor-21 (FGF-21)	Metabolic signaling; circulating FGF-21 are associated with diabetes in middle aged adults24; FGF-21 treatment in murine models are protective (via fat browning/energy expenditure and islet cell survival25)
Fatty acid binding protein-4 (FABP4)	Metabolic signaling; elevated levels related to diabetes3, 26; new evidence suggesting adipocyte-derived FABP4 signals to the pancreatic islet to improve islet function and survival27
Galectins (Gal-3, Gal-9)	Inflammation; gal-3 associated with diabetes in large cohort studies28, and gal-9 increased in diabetes, inversely proportional to renal function29; gal-9 involved in T cell immune responses30
Pancreasin (marapsin; serine protease 27; PRSS27)	Mechanism unknown; expressed in pancreas, and induced by inflammatory cytokines (IL-20)31; expression in muscle downregulated by caloric restriction32
Secretoglobulin family 3A2 (SCGB3A2)	Involved in LPS-mediated pyroptosis (inflammatory cell death)33
Stem cell factor (SCF)	Interaction with c-kit involved in stem cell maintenance; may be involved in pancreatic islet development34; mice overexpressing c-kit in the pancreas may enhance beta-cell function and resist diet-induced diabetes35
Kallikrein 6 (KLK6)	Mechanism unknown; protective association with diabetes in one study36; predominantly implicated in cancer and neurodegenerative disease
Pentraxin 3 (PTX3)	Mechanism unknown; acute phase reactant expressed in vascular atherosclerotic lesions, and has had divergent effects on insulin sensitivity and cardiovascular function in human and rodent studies37-39
Decorin (DCN)	Involved in regulation of TGF-beta (potential roles in inflammation and fibrosis); may be protective against diabetic cardiomyopathy40 and nephropathy41
Poly ADP ribose polymerase (PARP-1)	PARP-1 increased in diabetic cardiomyopathy; inhibition of PARP-1 increases sirtuin 1 and PGC-1α, reducing inflammation/fibrosis42; inhibition can ameliorate diabetic nephropathy43
Proprotein convertase subtilisin/kexin type 9 (PCSK9)	Mechanism unclear; PCSK9 levels higher in patients with diabetes44; reduction in PCSK9 improves dysglycemia45
Prostasin (PRSS8)	May connect endotoxemia with hepatic inflammation. High-fat diet suppresses hepatic PRSS8, increases hepatic toll-like receptor 4, and potentiates insulin resistance; rescuable with increasing PRSS8 expression46
Matrix metalloproteinases (MMP-3, MMP-7)	Matrix remodeling; MMP-3 associated with vascular stiffness47, and MMP-7 associated with diabetic complications (diastolic dysfunction, renal disease)48 and decline in renal function in general population49
Kidney injury molecule-1 (KIM1)	Marker of renal tubular injury that forecasts decline in renal function over nearly 2 decades50; responsive to SGLT2 inhibition51
Delta like-1 (Dlk-1; preadipocyte factor-1)	Interacts with Notch1 to affect signaling52; related to decreased adipose tissue inflammation, hepatic steatosis and glucose output in mice (potentially via AMP kinase activation)52
Growth differentiation factor-2 (GDF-2, also BMP9)	Limited reports in humans suggesting lower circulating and adipose/muscle tissue BMP9 expression in diabetes; decrease in BMP9 over time prevented by GLP-1 receptor agonist therapy53, 54; central role in limiting vascular permeability in diabetic macular edema55

As a second, phenotype-driven approach (Figure 3), we identified proteins associated with intermediate endophenotypes of diabetes susceptibility. As expected from additional statistical power in evaluating continuous phenotypes, we observed many more associations between the proteome and each phenotype (Supplementary Data File), many of which were not associated with incident diabetes. Selected results of multivariable selection (LASSO) regressions for each endpoint are summarized in Figure 3A (PCA weights for proteotypes in Supplementary Table IV). Two PC “proteotypes” explained ≈62% of the overall variance in the proteome-phenotype relation: a “pro-inflammatory adiposity” proteotype (loaded on adiposity and inflammatory phenotypes) and an “muscularity” proteotype (loaded on muscle phenotypes; Figure 3B). The muscularity proteotype displayed significant heterogeneity by sex (men higher than women), and the pro-inflammatory proteotype and Cox LASSO scores built on incident diabetes directly displayed significant heterogeneity by race (Black higher than white participants) and association with BMI (Supplementary Figure I). While several proteins highly weighted in the pro-inflammatory proteotype represented known involvement in inflammation and diabetes risk (e.g., FABP4, IL-6, leptin), several proteins with “protective” weights had not been previously widely implicated in human diabetes, despite physiologically plausible mechanisms in model systems (e.g., PRSS27, DLK-1, GDF-2, RARRES2, KLK-6, among others; Table 1). As expected, proteins highly weighted in the muscularity proteotype were central to muscle function (e.g., myoglobin [MB], brother of CDO [BOC; involved in muscle cell differentiation]) or involved in diverse processes central to insulin sensitivity and diabetes (GDF-2, IGFBP-1/2; Table 1). Of note, several proteins highly weighted in the muscularity proteotype had previously not been noted involved in muscle cell health (e.g., contactin-1 [CNTN1; neuronal function], superoxide dismutase 2 [SOD2; involved in oxidative stress]), suggesting potential confounding with muscle mass (e.g., by sex) or widespread pleiotropy across multiple proteins. Each of these proteotype scores (and the Cox diabetes score) were related to incident diabetes in CARDIA (Figure 3C).

Figure 3:

Phenotype-driven approach defines proteotypes (scores) of diabetes susceptibility. Panel A shows the loadings for each phenotype in PCA. This is overlaid over a heatmap showing selected proteins associated with subclinical endpoints in LASSO regressions. The 50 proteins with the largest cumulative magnitude of standardized beta coefficients across all endpoints are displayed for purposes of visualization. Bar graphs at the top show weights for the same selected metabolites in the two phenotype-derived scores with positive weights in blue and negative weights in shown in red. Score weights are on an arbitrary scale as scores are scaled to zero mean and unit variance after taking dot product with protein measures. Color legend applies only to heatmap. Panel B shows the PC loadings for each phenotype in the overall proteotypes. Positive loadings are shown in blue and negative loadings in red. Panel C shows standardized hazard ratios for incident diabetes in CARDIA for each of the three scores (2 phenotype-driven and 1 outcome-driven) across multiple adjustments (see Methods).

To increase external generalizability of this approach (and because the Cox-based scores were overfit in CARDIA sample), we present results of replication efforts in Table 2. Of note, glucose was an adjustment only for incident (not prevalent) diabetes models, given its role in diagnosis of diabetes. Of 522 individuals at baseline in the FHS, 113 had prevalent diabetes. After adjustment for age, sex, and BMI, the Cox diabetes scores (but neither of the phenotype-based scores) was associated with prevalent diabetes (OR = 1.70, 95% CI 1.32-2.18, P<0.0001). Between Exam 7 and Exam 8 in FHS (mean duration 6 years), we observed 33 new cases of diabetes (out of 268 + 33 individuals at risk). The pro-inflammatory and Cox diabetes scores were both related to incident diabetes in this group after age and sex adjustment (with odds ratios > 1.9), though these associations were mitigated after full multivariable adjustment. We observed generally similar results in the WHI (prevalent cases in minimally adjusted models: 280 diabetes/1056 non-diabetes; incident cases in minimally adjusted models: 122 diabetes/934 non-diabetes), though with smaller effect sizes across both prevalent and incident diabetes.

Table 2.

Association with incident and prevalent diabetes in replication cohorts (FHS and WHI). *Sex adjustment was not performed in WHI (given study sample was women only), and race adjustment was not performed in FHS.

Prevalent diabetes
		Framingham Heart Study		Women’s Health Initiative
Proteomic score	Adjustments*	Odds Ratio	P	Odds Ratio	P
Pro-inflammatory proteotype score	Age + Sex + Race	1.38 (1.10-1.72)	0.005	1.24 (1.08-1.43)	0.003
	Age + Sex + Race + BMI	1.07 (0.81-1.41)	0.64	1.02 (0.84-1.23)	0.87
Muscularity proteotype score	Age + Sex + Race	1.33 (0.98-1.79)	0.06	0.90 (0.79-1.03)	0.12
	Age + Sex + Race + BMI	1.34 (0.99-1.83)	0.06	0.91 (0.80-1.05)	0.2
Diabetes (Cox) score	Age + Sex + Race	1.87 (1.49-2.36)	<0.0001	1.27 (1.10-1.46)	0.0008
	Age + Sex + Race + BMI	1.70 (1.32-2.18)	<0.0001	1.14 (0.97-1.34)	0.11
Incident diabetes
		Framingham Heart Study		Women’s Health Initiative
Proteomic score	Adjustments*	Odds Ratio	P	Hazard Ratio	P
Pro-inflammatory proteotype score	Age + Sex + Race	1.98 (1.28-3.06)	0.002	1.21 (1.01-1.45)	0.04
	Age + Sex + Race + BMI + Fasting glucose	1.39 (0.77-2.51)	0.27	1.04 (0.81-1.34)	0.75
Muscularity proteotype score	Age + Sex + Race	1.38 (0.80-2.37)	0.25	1.06 (0.88-1.27)	0.56
	Age + Sex + Race + BMI + Fasting glucose	1.37 (0.77-2.46)	0.29	1.00 (0.82-1.20)	0.97
Diabetes (Cox) score	Age + Sex + Race	1.91 (1.28-2.85)	0.002	1.26 (1.05-1.50)	0.01
	Age + Sex + Race + BMI + Fasting glucose	1.24 (0.77-2.01)	0.37	1.06 (0.86-1.32)	0.57

The diabetes susceptibility proteome specifies genes with cell-specific expression patterns within human adipose tissue

Given the high effect sizes for pro-inflammatory adiposity, we next examined the cell type-specific expression within human adipose tissue of genes encoding proteins associated with subcutaneous or visceral fat or incident diabetes in CARDIA (in age/sex/race-adjusted models at a 5% FDR; Figure 4). Of the prioritized genes (see Methods), we observed a remarkable heterogeneity across cell types at the single cell level. Most prioritized transcripts were expressed in non-adipocyte cell types, including FAPs, vascular cells, and innate immune cells, consistent with involvement of the adipose stromal tissue in early insulin resistance (adipose tissue inflammation/ remodeling), many of which serve unique, potentially novel functions in early diabetes pathogenesis (Table 1).

Figure 4.

Single cell/single nuclear adipose tissue expression of genes overlapping with proteins associated with subcutaneous fat, visceral fat or incident diabetes in CARDIA. Top: Beta coefficient for the clinical association; Middle: heatbars (black/white) designate which cell type the gene was considered a marker gene for; Bottom: the dot plots showing the expression of these genes. Expression unit is natural log(UP10K + 1), where UP10K is the unique molecular identifiers (UMIs) per 10000. “All” signifies subcutaneous (“sc”) and visceral (“vis”) fat sample data together.

Expression of genes encoding the diabetes susceptibility proteome is related to adipose tissue morphology and function, insulin resistance, and adiposity

We next studied the relation of adipose tissue expression of transcripts encoding proteins prioritized based on our cross-sectional cohort of 56 women with adipose tissue morphology and function as well as insulin resistance and adipose distribution (Figure 5, correlation data and significance testing in Supplementary Data File). We observed consistent associations at the transcriptional level between prioritized transcripts and adiposity (by BMI, waist-to-hip ratio, and DEXA-derived tissue distribution), adipose tissue morphologic and functional phenotypes (adipocyte size, secretion of pro-inflammatory cyto-/chemokines and lipid turnover, i.e. lipolysis and insulin-stimulated lipogenesis), and measures of IR (HOMA-IR), including macrophage activation and inflammation (CCL2/3, SLAMF7⁵⁶, CD163⁵⁷, PLAUR⁵⁸, LGALS3⁵⁹), adipocyte iron metabolism (TFRC⁶⁰, HMOX1⁶¹), insulin resistance (IL1RN⁶²), and adipogenesis and adipocyte differentiation (RARRES2⁶³, NOTCH3⁶⁴; Table 3).

Figure 5.

Spearman correlation of whole-body and adipose tissue metabolic and inflammatory phenotypes with RNA expression (by microarray) in adipose tissue. Correlations and FDR values are in Online-Only Supplementary Data File. Abbreviations: BMI = body mass index; TG = triglyceride; HOMA = homeostatic model of insulin resistance; Fat % = total body fat percentage; TNFa secr = adipose tissue TNF alpha secretion; MCP1 secr = adipose tissue MCP1 alpha secretion; IL6 secr = adipose tissue IL6 alpha secretion; WHR = waist to hip ratio; CRP = C-reactive protein; lipolysis = isoprenaline-stimulated lipolysis; lipogenesis = insulin-stimulated 3H-glucose incorporation into lipid; FFA = free fatty acids.

Table 3.

Selected targets prioritized by population-based correlates of diabetes susceptibility phenotypes and adipose tissue expression functional studies. Up-arrow indicates positive correlation; down-arrow indicates negative correlation. This table is not meant to reflect a comprehensive curation of proteins identified by this work.

Transcript	Correlation With Adiposity	Correlation With HOMA	Potential mechanisms and diabetes association
SLAM family member 7 (SLAMF7)	↑	↑	IFN-γ inducible molecule that drives inflammatory responses56
Urokinase plasminogen activator receptor (PLAUR)	↑	↑	Expressed by adipose tissue macrophages; soluble levels (suPAR) elevated in obesity58
Transferrin receptor (TFRC)	↑	↑	Adipocyte-specific ablation of the transferrin receptor is protective against high-fat diet induced dysmetabolism via reduction in adipocyte iron and subsequent changes in lipid uptake from the intestine60
Heme oxygenase-1 (HMOX1)	↑	↑	Higher in adipose tissue in obese individuals and related to adipose iron overload, inflammation, and mitochondrial dysfunction; reversible with weight loss61
Interleukin-1 receptor antagonist (IL1RN)	↑	↑	IL1RN knock-out mice display greater insulin sensitivity62
Chimerin, retinoic acid receptor responder 2 (RARRES2)	↓	↓	Adipokine involved in adipocyte function and immunity65; elevated in murine models of obesity and diabetes and may be involved in glucose intolerance66; may have depot specific effects on adipogenesis63, with divergent levels in circulation versus in adipose tissue; expression in subcutaneous tissue related to more insulin sensitivity67
NOTCH3	↑	↑	Implicated in early differentiation of adipocytes64; potential different effects of different NOTCHs on brown versus white fat specification (with NOTCH3 supporting a white adipose tissue phenotype)68
Tartrate-resistant acid phosphatase 5 (ACP5)	↑	↑	Adipokine; stimulates adipocyte proliferation69
Cathepsins	↑	↑	Inflammation; Highly expressed in liver (in macrophages)70; associated with insulin resistance1, hepatic steatosis71
Thrombospondin-2	↑	↑	Implicated in liver fibrosis, delayed wound healing72, and interactions with extracellular matrix 73
Perlecan, basement membrane-specific heparan sulfate proteoglycan core protein (HSPG2)	↑	↑	HSPG2 knock-out mice display smaller adipocyte cell size, greater fatty acid oxidation, and increased mitochondrial biogenesis (driven by PGC1α)74
CD93	↑	↑	Expressed by activated endothelial cells, with increased expression in response to inflammation75; however, some data to suggest CD93 null mice have impaired insulin sensitivity76

Discussion

The ability to identify functional biomarkers of tissue physiologies classically related to diabetes can characterize early stages and mechanisms of diabetes susceptibility, before fasting or post-prandial dysglycemia. Here, we link quantitative phenotypes of diabetes susceptibility across multiple organs (e.g., adipose, muscle, inflammatory traits) with a circulating proteome in >2000 young adults across race (mean age 32 years) to identify inflammatory and muscle-based proteotypes related to long-term diabetes over 20 years, independent of classical diabetes risk factors, with replication in older cohorts (with an effect size mitigated by adjustment for BMI and dysglycemia, as in other studies⁵). The identified proteins span known (e.g., IGFBP-1/2, CD163, leptin, FGF-21) and novel tissue-specific mediators of inflammation, matrix remodeling, and organ injury (e.g., pancreasin, GDF-2, prostasin, KIM-1, MMPs, among others). Given these links across the circulating proteome, inflammatory adiposity, and long-term diabetes in young adulthood, we further investigated the relation of these proteins at the transcriptional level with tissue- and organism-wide metabolic and inflammatory states. These studies extended our epidemiologic observations, demonstrating (1) adipose tissue expression of key transcripts mostly in human adipose stroma and (2) strong associations between these transcripts and functional phenotypes of adipose tissue function, morphology, distribution, and IR. Of note, many identified targets using this precision approach are not widely described in human diabetes. Collectively, these results display a fully translational approach to biomarker discovery in diabetes, leveraging one of the largest multi-racial samples with quantitative phenotypes of diabetes susceptibility in a young, at-risk population with single-cell adipose tissue expression and functional studies to prioritize novel targets across a wide array of potential mechanisms relevant to human diabetes.

Most population-based proteomic studies in diabetes have simultaneously attempted to identify early risk biomarkers before frank dysglycemia and to use these to prioritize functional pathways of insulin resistance using traditional genetic approaches (Table 4). In one of the largest, seminal reports to date, Gudmundsdottir and colleagues quantified >4000 proteins in 5438 individuals in the Age, Gene/Environment Susceptibility (AGES)-Reykjavik Study (age ≈77 years)⁵, demonstrating 99 proteins significantly associated with incident diabetes, none of which survived adjustment for age, sex, glucose and BMI. In Mendelian randomization, several proteins (some shared with CARDIA) demonstrated genetic evidence of a causal role in diabetes, some supported by model systems (e.g., adipose-pancreatic communication with FABP-4²⁷). Functional genetic studies have also been used to prioritize functionally important biomarkers: Ngo and coworkers demonstrated an association between aminoacylase-1, a metabolic enzyme responsible for amino acid metabolism, with incident diabetes, with modulation of this enzyme in vivo related to insulin hypersecretion (a potential antecedent of beta-cell exhaustion)⁶. Nevertheless, most studies still have key limitations in the population studied (focused on older individuals), diversity (primarily European), and discovery endpoint, relying on a clinical diagnosis of diabetes^1-6. Approaches to functional validation—reliant on knowing where in key metabolic tissues transcripts or proteins are expressed; their relation to key phenotypes; and effects with perturbation—remain difficult, restricted largely to animal models or in vitro validation, and are expensive. While our results from CARDIA and limited replication after adjustment for traditional risk factors (e.g., BMI) are in line with these studies, the current study provides novelty in leveraging methodologies in human tissue, large cohort epidemiology and functional studies to prioritize targets for experimental validation.

Table 4.

Selected recent proteomic studies in diabetes.

Study	Population	Proteome Quantified	Findings
Uppsala Longitudinal Study of Adult Men (USLAM) 1 Prospective Investigation of the Vasculature in Uppsala Seniors (PIVUS)	Incident diabetes; both cohorts European USLAM: 540 individuals (age 77.6 years, 0% female; incident diabetes identified to age 88) PIVUS: 827 individuals (age 70.2 years, 51% female; incident diabetes defined to age 80)	96 cardiovascular proteins (proximity extension assay, Olink)	Associations with homeostatic model of insulin resistance, many overlapping with CARDIA (leptin, t-PA, IL-1ra, FABP4, cathepsin D) In models for incident diabetes, t-PA and IL-1ra were significant independent of clinical risk; with minimal improvement in discrimination and attenuation of effect estimates with adjustment for glucose
EpiHealth 2	Prevalent diabetes; European cohort 2467 individuals (for discovery sample: age 60.4 years, 49% female)	249 cardiovascular proteins (proximity extension assay, Olink)	29 proteins associated with prevalent diabetes, including several found in CARDIA, including cathepsin D (↑), HAO1 (↑), galectin-4 (↑), GDF-15 (↑), lipoprotein lipase (↓), PAI-1 (↑), ACE-2 (↑), IGFBP-2 (↓), FABP4 (↑)
Malmo Preventative Project 3	Prevalent and incident diabetes 1707 individuals (European; age 67.4 years; 29% female, 8 year follow-up)	96 cardiovascular proteins (proximity extension assay, Olink)	Incident diabetes: Several associations with lower (PON3, IGFBP-2) and higher (FABP4, PAI1, CD163, cathepsin D, galectin-4) hazard of diabetes overlapped with CARDIA. Largely similar findings for prevalent diabetes; small increase in C-statistic over fully adjusted models for diabetes prediction
Cooperative Health Research in the Region of Augsburg (KORA) F4 4 Nord-Trondelag Health Study (HUNT3)	Prevalent and incident diabetes KORA: 993 individuals (European, age 59.3 years, 52% female; 7 year follow-up) HUNT: 940 individuals (age 69.0 years, 26% female; 9 year follow-up)	≈1095 proteins (aptamer-based proteomics, SOMAScan)	Prevalent diabetes: 85 associated proteins (at 5% false discovery rate) Incident diabetes: several replicated across cohorts using either single protein-outcome regression or LASSO, some overlap with CARDIA (IGFBP-2, CD163, among them); Mendelian randomization found nominally significant evidence of causality
Age, Gene/Environment Susceptibility (AGES)-Reykjavik Study 5	Prevalent and incident diabetes 5,438 individuals (European, age 76.6 years, ≈40-60% female; 5 year follow-up)	≈4,137 proteins (aptamer-based proteomics, SOMAScan)	Prevalent diabetes: 520 associated proteins; attenuation of number of significant proteins after adjustment for insulin or BMI Incident diabetes: 99 proteins significantly associated; none survive adjustment in models adjusted for age, sex, glucose, and BMI; highest effect sizes include some overlapping proteins with CARDIA: IGFBP-2 (protective; odds ratio OR 0.35) and leptin (OR 2.42). Proteins associated were implicated in metabolic processes and insulin physiology Other findings included sex-specific differences in protein expression, and Mendelian randomization supported causal role of select proteins in diabetes (e.g., FABP4, MMP12, GDF-15, among them)
Framingham Heart Study (FHS) 6 Malmo Diet and Cancer Study (MDCS)	2839 total individuals, case-control FHS: 1618 individuals (for cases: age 58 years, 48% female) Malmo: 1221 individuals (for cases: age 58 years, 49% female)		After multivariable adjustment, 19 proteins associated with incident diabetes, many of which were novel (e.g., WFIKKN2, gelsolin, THBS2); minimally adjusted models also showed similar results as in CARDIA (e.g., IGFBP-1/2, leptin, RARRES2)

The first innovation of our approach was the use of phenotypes for discovery, which facilitated discovery beyond known pathways of metabolic regulation to those not widely described in human diabetes. Proteins across a wide array of mechanisms—host response to endotoxemia (secretoglobulin family 3A2, prostasin), renal and vascular injury (KIM-1, MMPs, pentraxin 3), pancreatic beta-cell function (SCF), and inflammatory responses (pancreasin)—not widely linked to human diabetes were uncovered with this approach. For example, lower adipose or muscle tissue expression of growth differentiation factor-2 (GDF-2, or BMP9)—favorably weighted in the pro-inflammatory proteotype in CARDIA—is observed in individuals with diabetes and is responsive to GLP-1 receptor agonist therapy^{53, 54}. Similarly, delta like-1 (preadipocyte factor-1; negatively loaded in pro-inflammatory proteotype) is involved in regulation of Notch signaling⁵² and is related to decreased adipose tissue inflammation, hepatic steatosis and glucose output in mice (potentially via AMP kinase activation)⁵². Similarly, inhibition of PARP-1 (poly ADP ribose polymerase-1)—associated with increased diabetes risk in CARDIA—increases sirtuin 1 and PGC-1α, reducing inflammation/fibrosis⁴² and ameliorating diabetic nephropathy⁴³.

A second key innovation of our approach was the integration with human adipose tissue, both with cell-specific expression (using single cell technologies) and functional-morphologic assessments of metabolism. Strikingly, genes prioritized by proteomic association with adipose tissue measures or diabetes were expressed in non-adipocyte adipose tissue populations, prominently inflammatory, vascular, and progenitor cells. Expression of relevant genes in non-adipocyte cell populations is consistent with the importance of adipose tissue stroma in early IR mechanisms. Importantly, we observed association between adipose tissue expression of key prioritized genes and both insulin resistance and adiposity measures, as well as adipocyte metabolic-inflammatory function. Ultimately, the genes prioritized from population-level evidence and functional-morphologic data identified several highly novel targets within human diabetes, including iron metabolism (HMOX1 and TFRC), adipocyte metabolic dysregulation and mitochondrial function (HSPG2), and several novel inflammatory targets. Of note, whether RNA expression in subcutaneous tissue is translated to adipose tissue protein or even circulating levels remains an issue^{67, 77}, as well as potential differences in directionality (e.g., CSTB⁷⁸; potentially addressed by future genetic or model studies). In addition, while further filtering of targets would benefit from laboratory studies (e.g., gain and loss of function studies in obesity models), the results of this translational paradigm—from humans to tissues to cells—are striking, providing a critical connection toward that laudable goal.

Several important caveats merit mention. Absence of validation in FHS and WHI independent of glucose and BMI is consistent with prior reports⁵, likely due to glucose defining diabetes itself and age and follow-up of the cohorts. However, we did not observe this in CARDIA, likely related to the lower blood glucose and BMI at study entry (greater potential for the proteome to explain variance in diabetes). Also, the selection of these validation samples was to maintain consistency with a prevention population with the same proteomics platform, given clear cross-platform differences⁷⁹. While we did not explore associations by sex and race, several associations may be different (potentially driven by social determinants of health), and we recommend larger studies across diverse populations and proteomes with phenotyping of social determinants to address this. In addition, proteomic coverage here was not as broad as in other reports, limiting traditional pathway enrichment and leading to a selected literature-based approach to curation (not a comprehensive examination across all associations). Nevertheless, phenotypic and functional characterization were highly unique, and identified plausible known and novel pathways of diabetes susceptibility. Certainly, broader proteomic coverage (including MS-based methods for post-translational modifications) are likely to yield even greater insights. Future studies in large cohorts (e.g., UK Biobank) will offer substantial power for a broader proteome.

The approach here is a fundamental departure from published proteomic studies in diabetes, demonstrating the utility of a multi-faceted discovery effort uniting proteomics, clinical susceptibility phenotypes, and cell-level expression patterns to uncover novel functional biomarkers of early diabetes susceptibility for future mechanistic evaluation. As the “omic” space expands, a union of tissue-, population-, and functional biology-based approach is likely to prioritize targets for therapeutic and mechanistic evaluation.

Nonstandard Abbreviations and Acronyms

BMI

Body mass index

CARDIA

Coronary Artery Risk Development in Young Adults

WHI

Women’s Health Initiative

FHS

Framingham Heart Study

IQR

interquartile range

LASSO

Least absolute shrinkage and selection operator

PCA

principal component analysis

FDR

false discovery rate

IR

insulin resistance

DEXA

dual x-ray absorptiometry

GLP-1

glucagon like peptide-1