A Bayesian joint pQTL study sheds light on the genetic architecture of obesity

The genetic contribution to obesity has been widely studied, yet the functional mechanisms underlying metabolic states remain elusive. This has prompted analysis of endophenotypes via quantitative trait locus studies, which assess how genetic variants affect intermediate gene (eQTL) or protein (pQTL) expression phenotypes. However, most such studies rely on univariate screening, which entails a strong multiplicity burden and does not leverage shared regulatory patterns. We present the first multivariate pQTL analysis with our highly scalable Bayesian framework LOCUS, on plasma protein levels from a dual mass-spectrometry and SomaLogic assay, and show that it is more powerful than a standard univariate procedure on this data. We identify 136 pQTL associations in the Ottawa obesity cohort, of which > 80% replicate in the independent DiOGenes cohort and have significant functional enrichments; 15% of the hits would be missed by univariate analysis. By exploiting clinical data, we reveal the implication of proteins under genetic control in low-grade inflammation, insulin resistance, and dyslipidemia, opening new perspectives for diagnosing and treating metabolic disorders. All results are freely accessible online from our searchable database.


Introduction
Genome-wide association studies (GWAS) have identified hundreds of loci associated with obesity susceptibility, 1 yet their functional impact on metabolism remains poorly understood. The analysis of As SNPs were imputed, the 0.95-r 2 pruning led to a drastic cut of "redundant" markers: without this pruning step, the number of SNPs was ≈ 4M. Such a reduction is not surprising considering the nature of the underlying SNP arrays (essentially based on tag SNPs) and indicates that little information was discarded. In the Ottawa cohort n = 376 subjects had both genotype and MS proteomic data, and n = 394 subjects had both genotype and MS proteomic data. In the DiOGenes cohort, these numbers were n = 400 and 548.
Clinical data. Both cohorts had records on age, gender, anthropometric traits (weight and BMI), glycemic variables (fasting glucose, fasting insulin, HOMA-IR), and total lipid levels obtained from blood biochemistry (total cholesterol, triglycerides, HDL). We derived LDL values using the Friedewald formula, 19 and obtained gender-specific visceral adiposity index (VAI) values using the formula of Amato et al. 20 In each cohort and for each clinical variable, we removed a few samples with extreme measurements, similarly as for the proteomic data quality control.
LOCUS: fast Bayesian inference for multivariate QTL analysis. LOCUS 11 is a variational inference approach for joint mapping analysis at the scale required by current molecular QTL studies ( Figure 1A). It implements a hierarchical sparse regression model that involves a collection of highdimensional regressions, where y = (y 1 , . . . , y q ) is an n × q matrix of q centered outcomes (e.g., genomic, proteomic, or metabolomic levels), and X is an n × p matrix of p centered candidate predictor SNPs, for each of n samples. Each outcome, y t , is related linearly to all p candidate SNPs, and has a specific residual precision, τ t , to which we assign a Gamma prior, τ t ∼ Gamma(η t , κ t ). As p, q n, sparsity of the p×1 regression parameters β t is enforced by placing a spike-and-slab prior on each of their components, namely, for s = 1, . . . , p, where δ 0 is the Dirac distribution. Hence, to each regression parameter β st corresponds a binary latent parameter γ st , which acts as a "predictor-outcome association indicator": the predictor X s is associated with the outcome y t if and only if γ st = 1. The parameter σ represents the typical size of nonzero effects and is modulated by the residual scale, τ −1/2 t C++ subroutines and is publicly available at https://github.com/hruffieux/locus. Our MS and SomaLogic analyses completed in a few hours for 275K tag SNPs representing information from about 5M common markers, yet larger SNP panels can be considered as our method scales linearly in terms of memory and CPU usage. For instance, analyses of 2M SNPs and 1000 proteins run in less than 40 hours (see profiling in the Appendix A).
Simulation study design. We evaluated the performance of LOCUS expected on our data by conducting two simulation studies. We compared its statistical power to detect pQTL associations with that of the linear mixed model approach GEMMA, 24 which estimates the associations between each SNP and each outcome in a univariate fashion. We used the R package echoseq to generate synthetic data that emulate real data.
For the first simulation, we ran LOCUS and GEMMA on the SNPs of all n = 376 Ottawa subjects, and on simulated expression outcomes with residual dependence replicating that of the q = 133 MS proteomic levels. We used the SNPs from chromosome one (p = 20, 900), and generated associations between 20 SNPs and 25 proteins chosen randomly, leaving the remaining variables unassociated.
Some proteins were under pleiotropic control; we drew the degree of pleiotropy of the 20 SNPs from a positively-skewed Beta distribution, so only a few SNPs were hotspots, i.e., were associated with many proteins. We generated associations under an additive dose-effect scheme and drew the proportions of outcome variance explained by a given SNP from a Beta (2,5) distribution to give more weight to smaller effect sizes. We then rescaled these proportions so that the variance of each protein attributable to genetic variation was below 35%. These choices led to an inverse relationship between minor allele frequencies and effect sizes, which is to be expected under natural selection. We generated 50 replicates, re-drawing the protein expression levels and effect sizes for each.
For the second simulation, we re-assessed the performance of LOCUS for a grid of data generation scenarios. We considered a wide range of sparsity levels (numbers of proteins under genetic control) and effect sizes (proportions of outcome variance explained by the genetic variants). Given the large number of configurations (130), and in order to limit the computational burden, we used the first p = 2, 000 SNPs, and ran LOCUS and GEMMA on 20 replicates for each configuration.
Proteomic quantitative trait locus analyses. We performed pQTL analyses separately for each platform, i.e., one analysis for the MS proteomic dataset, and another for the SomaLogic proteomic dataset. Each analysis comprised two stages: a discovery stage using the Ottawa cohort and a replication stage based on the DiOGenes cohort.
For discovery, we used LOCUS on both the MS and the SomaLogic datasets, with an annealing schedule of 50 geometrically-spaced temperatures and initial temperature of 20; pilot experiments indicated that estimation was not sensitive to these choices. We used a convergence tolerance of 10 −3 on the absolute changes in the objective function as the stopping criterion. The algorithm can handle missing data in the outcome matrix, so no imputation was necessary for the MS proteomic data.
We adjusted all analyses for age, gender, and BMI at baseline. No important stratification was observed in the genotype data; the first ten principal components together explained little of the total variance (< 4%), so we did not include them as covariates. We derived FDR values from the posterior probabilities of association obtained between each SNP and each protein, and reported pQTL associations using an FDR threshold of 5%.
We performed a validation study of the pQTLs discovered using the DiOGenes cohort with GEMMA, 25 with centered relatedness matrix (default) and p-values from (two-sided) Wald tests.
once. We made both general queries and queries asking whether a pQTL uncovered by LOCUS was an eQTL for the gene coding for the controlled protein.
We retrieved known associations between the validated sentinel pQTLs and diseases or clinical traits, based on the GWAS catalog 29 (v1.0 release e92), and also using an LD proxy search (r 2 > 0.8).
We evaluated enrichment for eQTL and risk loci using one-sided Fisher exact tests based on the 104 validated sentinel pQTLs.
Associations with clinical variables. We tested associations between the proteins under genetic control and clinical parameters separately in each cohort. For the DiOGenes data, we used linear mixed-effect models, adjusting for age, gender as fixed effects, and center as a random effect.
For the Ottawa data, we used linear models, adjusting for age and gender. Except when testing associations with anthropomorphic traits, all analyses were also adjusted for BMI. For the clinical variables available in the two cohorts (total cholesterol, HDL, LDL, fasting glucose, fasting insulin, HOMA-IR, triglycerides and VAI), we performed meta-analyses using the R package metafor. We used random-effects models to account for inter-study variability, which may in part result from geographical differences, and employed two-sided Wald tests for fixed effects, and Cochran Q-tests for measuring residual heterogeneity; we did not interpret the results if between-study heterogeneity estimates were high (I 2 > 80%), and evaluated the directional consistency of the effects between Ottawa and DiOGenes. We adjusted for multiplicity using Benjamini-Hochberg correction across all tests, i.e., involving the 88 tested proteins and the two proteomic technologies, and reported associations using a 5% FDR threshold.
We assessed whether the proteins under genetic control were enriched in associations with the clinical variables. We randomly selected 10 5 sets of 88 proteins from the panel used for the pQTL analyses and derived an empirical p-value by counting, for each set, the number of proteins with at least one clinical association at FDR 5%.  Table S4).
Comparison with a standard univariate pQTL analysis. The high replication rates and the novel discoveries uncovered by LOCUS are largely attributable to its flexible hierarchical sparse regression model which exploits shared association patterns across all SNPs and proteomic levels ( Figures   1A-C), as extensively shown in previous numerical experiments. 11 Here, we provide additional evidence for our specific study and data by comparing LOCUS with the univariate method GEMMA in two ways. First, we evaluate variable selection performance in two simulation studies and, second, we confront the hits of LOCUS real data analysis to those found by re-analysing the Ottawa and DiOGenes datasets with GEMMA.
Colocalization with eQTLs and evidence for regulatory impact. We assessed the overlap of the 113 validated pQTLs with known eQTLs (Supplementary Table S6). Seventy-seven of the 104 sentinel SNPs involved in our pQTL associations had one or more eQTL associations in at least one tissue. These SNPs have been implicated in 83 eQTL associations, representing a significant enrichment (p < 2.2 × 10 −16 ). Forty-nine of these 77 SNPs were eQTL variants for the gene coding for the protein with which they were associated in our datasets. Our pQTLs were also enriched in epigenome annotation marks (p = 9.20 × 10 −4 ) and significantly closer to transcription start sites compared to randomly chosen SNP sets (p = 9.99 × 10 −6 ). These observations suggest potential functional consequences for our pQTL hits.
Colocalization with GWAS risk loci. A total of 217 previously reported genome-wide associations overlapped our validated pQTL loci, corresponding to 139 unique traits mapping to 68 distinct regions (based on LD r 2 > 0.8). Nineteen SNPs were directly involved in these associations (Supplementary   Table S7) representing a significant enrichment (p < 2.2 × 10 −16 ).
Some of these results generate useful hypotheses to be explored in future research. For instance, a HGFL cis pQTL, rs1800668, is in strong LD (r 2 > 0.95) with rs9858542 and rs3197999, which are known to associate with Crohn's disease. 33,34 Our pQTL finding may be of clinical relevance given the prevalence of Crohn's disease in overweight and obese subjects; 35 the region would merit follow-up in inflammatory bowel disease cohorts.
Another example concerns an association between rs3865444 and the Siglec-3 protein, whose coding gene, CD33, has been reported as a risk factor for Alzheimer's disease. 36 As subjects obese in midlife are more at risk of developing late-life Alzheimer's, 37 this pQTL may help to better understand the genetic bases of Alzheimer's disease and dementia; its potential as a prognosis biomarker should be studied in Alzheimer's cohorts, ideally using weight records.
Proteins as endophenotypes to study the genetics of obesity. Annotation from public databases suggested that most pQTLs had implications in inflammation, insulin resistance, lipid metabolism or cardiovascular diseases. We performed a more systematic evaluation of their clinical relevance in a meta-analysis of the DiOGenes and Ottawa clinical and proteomic data, and found   that 35 of the 88 proteins under genetic control had associations with dyslipidemia, insulin resistance or visceral fat-related measurements at FDR 5%; these associations should be attributable metabolic factors independently of overall adiposity, as we controlled for BMI as a potential confounder. They are displayed as a network in Figure 4A and are listed in Table 1. We observed consistent directions of effects in the two cohorts (see Forest plots of Figure 5, and Supplementary Table S8 for full details).
Remarkably, we found that the 88 genetically-driven proteins are significantly more associated with the clinical variables than randomly chosen protein sets (p = 0.014); this enrichment suggests that the primary pQTL analyses can help uncover potential proteomic biomarkers for the Metabolic Syndrome and other obesity-related complications.
As shown in the network of Figure Table S8), which is consistent with previously described pleiotropic associations of RARRES2 variants with circulating RARR2, triglyceride levels and diverse measurements related to inflammation, 46 and findings from animal models. 47,48 Moreover, MS and SomaLogic RARR2 levels were strongly associated with visceral fat, even when controlling for BMI ( Figure 5; Supplementary Table S8), further strengthening the relevance of this protein in the development of the Metabolic Syndrome. 45,49 Pleiotropic effects from the ABO locus onto CADH5, CD209, INSR, LYAM2 and TIE1.
ABO is a well-known pleiotropic locus associated with coronary artery diseases, type 2 diabetes, liver enzyme levels (alkaline phosphatase) and lipid levels. [2][3][4] Our analyses highlighted two independent sentinel SNPs in the ABO region: rs2519093 and rs8176741 (r 2 = 0.03). The former SNP is trans-acting on E-selectin (protein LYAM2 encoded by SELE), the Insulin Receptor and the CD209 antigen.
Both SNPs were reported as cis-acting eQTL variants for ABO, OBP2B and SURF1, and further queries in public databases indicated that rs8176741 may affect the binding sites for three transcription factors (Myc, MYC-MAX and Arnt), suggesting a complex gene regulation circuitry.
Since the Ottawa subjects are more insulin-resistant than the DiOGenes subjects (average HOMA-IR with standard deviation: 4.97(3.88) versus 3.00(1.71), p = 2.52 × 10 −18 ; Supplementary Table S1), LYAM2 might represent a marker of insulin-resistance severity. Consistent with this hypothesis, the plasma levels of LYAM2 are employed as a biomarkers of endothelial dysfunction and risk of type 2 diabetes. 50 We found that CD209 circulating levels were positively associated with HDL, negatively with triglyceride levels, and, consistently with these effects, negatively with visceral fat index, suggesting beneficial effects of high CD209 levels. Further investigation using deconvolution of adipose tissue gene expression profiles showed that CD209 is predominantly secreted by M2 macrophages. 51 These cells are involved in extracellular matrix remodelling and secrete cytokines with an antiinflammatory role, counteracting the effect from pro-inflammatory macrophages M1. Interestingly, other adipose cell types, including M1 macrophages display little, if any, CD209 expression. M1 and M2 macrophages have been extensively discussed in the context of obesity, and it is well established that M2 macrophages have a protective role against obesity and insulin resistance 52 and increase fatty oxidation and oxidative phosphorylation. 53 In our data (both the Diogenes and Ottawa cohorts, and at FDR 5%), we found that CD209 levels were positively correlated with M2 secreted proteins (IL10, CSF1, ARGI1) and negatively correlated with M1 pro-inflammatory markers, such as TGF-beta, IL6 and interferon-gamma (Supplementary Figure C.1). Importantly, CD209 was positively associated with circulating levels of adiponectin, an hormone secreted in adipose tissue which plays a key role in glucose regulation, fatty acid oxidation and triglycerides clearance. This lends support that CD209 could be a secreted protein, released by M2 macrophages from adipose tissue, with a beneficial role in controlling lipid levels, thereby possibly protecting from developing dyslipidemia and related metabolic complications.
XRCC6, a DNA repair protein as putative biomarker for metabolic disorders. We identified rs4756623, as a novel trans pQTL for XRCC6 (X-Ray Repair Complementing Defective Repair In Chinese Hamster Cells; also known as Ku70). Proxy searches (down to r-square 0.5 in European ancestry panels) did not reveal any tag SNP previously reported as a QTL (including e-, p-, or m-QTL).
The XRCC6 gene activates DNA-dependent protein kinases (DNA-PK) to repair double-stranded DNA breaks by nonhomologous end joining. DNA-PKs have been linked to lipogenesis in response to feeding and insulin signaling. 54 DNA-PK inhibitors may reduce the risk of obesity and type 2 diabetes by activating multiple AMPK targets. 55 A recent review discussed the role of DNA-PK in energy metabolism, and in particular, the conversion of carbohydrates into fatty acids in the liver, in response to insulin. 56 It described increased DNA-PK activity with age, and links with mitochondrial loss in skeletal muscle and weight gain. Finally, XRCC6 functions have been reported as associated with regulation of beta-cell proliferation, islet expansion, increased insulin levels and decreased glucose levels. 55,57 We observed significant associations between the XRCC6 protein levels and several clinical variables in the Ottawa cohort (FDR < 5%). Higher expression was associated with decreased HDL (p = 5.83×10 −4 ), as well as with higher triglycerides (p = 4.39×10 −4 ), insulin levels (p = 4.50×10 −4 ) and visceral adiposity (p = 5.94 × 10 −5 ; Figure 5). We only found marginal associations using the DiOGenes data for insulin levels (nominal p = 0.02, corrected p = 0.14) and HOMA-IR (nominal p = 0.02, corrected p = 0.16). The directionality of these effects was consistent in both cohorts. As the Ottawa subjects were more severely obese, the effects might be larger for subjects with pronounced Metabolic Syndrome.
Our pQTL sentinel SNP, rs4756623, is intronic and located within the LRRC4C gene, a binding partner for Netrin G1 and member of the axon guidance. 58 To our knowledge, LRRC4C has not been previously described in the context of obesity, insulin resistance or type 2 diabetes. However, its partner Netrin G1 is known to promote adipose tissue macrophage retention, inflammation and insulin resistance in obese mice. 59 The underlying regulatory mechanisms between rs4756623 and the XRCC6 locus should be clarified, and functional studies will be required to understand their physiological impact.

Discussion
Despite important technological advances, large-scale pQTL studies remain infrequent, owing to their high costs. [2][3][4][5][6]30 To date, most studies have focused on data from the general population with limited access to clinical parameters and have assessed links with diseases by relying on information from different studies.
Here we described the first integrative pQTL study that relates the associations discovered to metabolic disorders, such as insulin resistance and dyslipidemia, in the obese population considered.
Our Bayesian method LOCUS confirmed 93 pQTLs (75 distinct proteins) highlighted in previous studies, 2,4-6,30 despite our sample sizes 2.5 to 18 times smaller, and revealed 20 novel pQTLs (18 distinct proteins, see Supplementary Table S4), with sound evidence for functional relevance and implications for the development of the Metabolic Syndrome. Our two-stage approach achieved very high replication rates (> 80%), and validated findings which standard univariate designs would have missed (e.g., the trans associations with INSR, PROC, SEM3A, TENA and XRCC6 would have been missed). This corroborates our simulation study, which demonstrated the increased statistical power of LOCUS over univariate approaches on synthetic data mimicking the real data. Owing to its joint modelling of all proteins and genetic variants, LOCUS both accounts for linkage disequilibrium and exploits the shared regulatory architecture across molecular entities; this drastically reduces the multiplicity burden and enhances the detection of weak effects. Finally, our analyses indicated that proteins under genetic control are enriched in associations with clinical parameters pertaining to obesity co-morbidities, which further supports a genetic basis of these parameters and emphasizes the advantages of pQTL studies for elucidating the underlying functional mechanisms. Our complete pQTL and clinical association results offer opportunities to generate further hypotheses about therapeutic options; they are accessible from the searchable online database https://locus-pqtl.epfl.ch.
The work presented in this paper is at the interface of methodological developments for pQTL mapping and concrete biological findings that take advantage of our tailored statistical approach in a thorough analyses of two obesity cohorts and two independent proteomic technologies. A central ambition was to showcase that LOCUS can bridge the gap between Bayesian multivariate inference and its practical use for analyzing current molecular QTL data. Indeed, the applicability of LOCUS Author contributions HR designed and developed the LOCUS method with input from AD and JH. AV supervised the omics data generation and preprocessing. JC contributed to data processing. HR and AV designed the pQTL study. HR implemented statistical analyses with input from AV and AD. WS and AA designed the DiOGenes clinical study; BD and MEH designed the Canadian program. HR and AV interpreted the results, wrote the manuscript with input from all authors, and have primary responsibility for final content.

A Computational performance of LOCUS
The runtime of LOCUS for the simulations presented in the paper was similar to that of GEMMA: on average, for one replicate, LOCUS took 5 minutes and 26 seconds to complete, while GEMMA took 7 minutes and 4 seconds, running in parallel on four cores of an Intel Xeon CPU, 2.60 GHz.

B Further examples of pQTL loci with probable implications in metabolic disorders
The importance of IL1AP for Metabolic Syndrome. The IL-1 pathway plays a critical role in the immune-response associated with obesity and type 2 diabetes; 60 other IL-1 related cytokines, such as IL-1ra, are also well documented in the context of type 1 and type 2 diabetes. 61 The IL1AP (IL-1 receptor accessory) protein is a co-receptor of the IL-1 receptor, and its soluble levels were found reduced in obese subjects. 62 Our analyses found an association between rs724608 and IL1AP, corroborating previously identified associations with SNPs in LD (r 2 = 0.93). 62 We found associations between IL1AP expression and measures of fasting insulin levels (p = 3.88× WFKN2, a TGFβ-activity protein with protective effect against metabolic disorders.
The role of the WFKN2 protein and of its coding gene, WFIKKN2, in regulating TGFβ activity has been extensively studied in muscle and skeletal muscle, 64 but, to our knowledge, not in other tissues.
We describe it for the first time in the context of obesity and metabolic disorders. We found that higher protein levels were associated with lower levels of fasting insulin, triglycerides, HOMA-IR and visceral fat (Figure 5), suggesting a protective role against metabolic dysregulation.
Our analyses suggested that the WFKN2 levels are controlled by rs9303566, which is consistent with other p-and eQTL studies (Supplementary Tables S4-S5). This SNP was found to be associated with DNA methylation and histone marks, 65,66 and is located within 100 base pairs of a transcription factor binding site, with numerous factors such as MYBL2, NFIC, EP300 and MXI1. It is in strong LD with other SNPs with potential regulatory impact; for instance, it is located 9Kb upstream to rs8072476 (r 2 = 0.97), which overlaps another cluster of transcription factor binding sites (FOXA1,   ESR1, USF1 & 2, TFAP2A & 2C).
Inflammation mediated proteins and their role in insulin resistance. We found a cis effect of rs6741488 on KYNU (Kynureninase) plasmatic levels. KYNU is an enzyme involved in the biosynthesis of nicotinamide adenine dinucleotide (NAD) cofactors from tryptophan. This protein and its pathway have been found to be particularly relevant for obesity and associated metabolic disorders. KYNU was found to be up-regulated by pro-inflammatory cytokines in human primary adipocytes, and more so in the omental adipose tissue of obese compared to lean control subjects. 67 Other studies indicated that the kynurenine pathway (KP) may act as an inflammatory sensor, and that increased levels of its catabolites may be linked with several cardiometabolic defects, including cardiovascular disease, diabetes and obesity. 68 In our cohorts, higher KYNU levels were associated with decreased HDL levels (p = 6.66 × 10 −4 ), and increased triglycerides levels (p = 3.43 × 10 −8 ), visceral fat (p = 2.51 × 10 −8 ) and insulin resistance (marginally, nominal p = 2.53 × 10 −2 , corrected p = 0.17), see Figure 5; as expected, higher protein levels were associated with a worsened Metabolic Syndrome score (Ottawa p = 8.23 × 10 −5 ; DiOGenes p = 3.62 × 10 −6 ).

Inflammation mediated proteins
Ottawa DiOGenes Recent work suggested a causal link between obesity and cancer, mediated by KP activation through inflammatory mechanisms. 69 Interestingly, our analyses highlighted two soluble interleukin receptor antagonist proteins, namely IL6RA and I17RA, that were both under genetic control and associated with insulin resistance ( Figure 4A). We did not find significant correlation between the I17RA and KYNU protein levels, but we did observe a significant negative correlation between IL6RA and KYNU (Ottawa p = 0.01 and DiOGenes p = 4×10 −3 ). We found a link between the plasma levels of KYNU and pro-inflammatory molecules, namely, IL6, IFNG and TNFα. In the Ottawa cohort, where subjects displayed high low-grade inflammation status, KYNU was positively associated with IL6 and IFNG at FDR 5%, while in DiOGenes, we found a positive association with IFNG only (see Figure B.1). Finally, metabolic dysfunctions mediated via KP may relate to another inflammatory pathology, namely, psoriasis, 70 a skin disease aggravated by obesity and improved by weight loss. 71,72 Our results thus highlighted pQTLs with probable roles in inflammation and subsequent metabolic dysfunctions, reinforcing previous discussion 68, 73 of the potential of KP therapeutic inhibitors against cardiovascular disease and metabolic disorders.
Complement/coagulation: a trans-acting insertion linking PROC and its receptor. PROC (Protein C, coding gene PROC on chromosome 2) and its paralog protein FA7 (Coagulation Factor 7, coding gene F7 on chromosome 13) regulate the complement and the coagulation systems. Both systems promote inflammation 74 and contribute to metabolic dysfunction in the adipose tissue and liver. 75 Our analyses suggested novel pQTLs for these proteins (Supplementary Table S3): FA7 was associated with rs3093233, which is a known eQTL of F7 and F10 in several tissues (Supplementary Table S5). PROC may be controlled by trans-regulatory mechanisms, initiated in its receptor gene, PROCR, on chromosome 20; it was indeed associated with an insertion, rs141091409, located 20Kb upstream of PROCR, an association observed with both our proteomic platforms. Previous studies found associations between cardiovascular disease and variants located in the PROC or PROCR genes. 76,77 Interestingly, our hit, rs141091409, was in strong LD (r 2 > 0.95) with the missense variant rs867186, previously identified as associated with coronary heart disease. 77 Our clinical analyses support the relation of PROC and FA7 levels with lipid traits: both were positively associated with cholesterol, triglycerides and visceral fat ( Figures 4A and 5). PROC levels were quantified by both platforms, and displayed consistent results. The SomaLogic measurements of PROC were positively associated with LDL (p = 5.39 × 10 −5 ). The role of these proteins for cardiovascular disease and NAFLD diseases in the overweight/obese population would merit further investigation.
C Correlation between CD209 and other macrophage protein levels