Genetic Evidence for Different Adiposity Phenotypes and Their Opposing Influences on Ectopic Fat and Risk of Cardiometabolic Disease

Obesity is a significant risk factor for various conditions including type 2 diabetes, heart disease, and hypertension—a cluster of events often referred to as the metabolic syndrome (1). However, in the general population, ∼15–40% of individuals categorized as obese do not present any obesity-related metabolic conditions or diseases and are “metabolically benign” at the specific time point of measurement, supporting the existence of metabolically benign obesity (2,3).

Previously we showed that a genetic predisposition to storing excess fat in subcutaneous adipose tissue (SAT) is associated with a reduced propensity to store fat in the liver, consequently reducing risk of disease (4). The identification of “favorable adiposity” variants, with their adiposity-increasing alleles paradoxically associated with lower risk of type 2 diabetes, heart disease, and hypertension (4–7), provided genetic evidence for the paradox of metabolically benign obesity. These genetic findings suggest that there are at least two types of variants associated with higher adiposity: one with favorable metabolic profile (favorable adiposity [FA]) and the other with an unfavorable metabolic profile (unfavorable adiposity [UFA]).

Although our previous studies suggested an important role for liver fat, we have been unable to determine whether pancreatic fat deposition or liver and pancreas volumes were similarly implicated due to lack of data, and it has not been possible to investigate mechanisms imposed by each variant individually. Clarification of the underlying pathophysiologic mechanisms that link adiposity to higher risk of type 2 diabetes and other cardiometabolic disease is critical to understanding disease progression and remission, especially given the rising prevalence of obesity and the rapid rise of type 2 diabetes in an aging population. The availability of both metabolic markers and MRI scan data in the UK Biobank (8) has enabled us to test in more detail the characteristics of adiposity variants and the role of ectopic fat in disease mechanism.

In this study, we focused on how higher adiposity is associated with ectopic fat, metabolic derangements, and cardiometabolic risk. Specifically, we aimed to 1) identify distinct clusters of FA and UFA variants, 2) investigate the relation between FA and UFA variants and ectopic fat deposition in the liver and pancreas, 3) examine how FA and UFA variants are associated with circulating markers of inflammation, 4) determine whether sexual dimorphism is a factor in the association between the clusters and metabolic biomarkers, fat distribution, and disease risk; and 5) use Mendelian randomization (MR) to determine the potential causal role of “favorable” and “unfavorable” adiposity in different components of metabolic syndrome.

UK Biobank recruited >500,000 individuals aged 37–73 years (99.5% were between 40 and 69 years of age) between 2006 and 2010 from across the U.K. (8) (Supplementary Table 1). The UK Biobank has approval from the North West Multicenter Research Ethics Committee (https://www.ukbiobank.ac.uk/ethics/), and these ethics regulations cover the work in this study. Written informed consent was obtained from all participants.

The steps performed to identify variants associated with adiposity but with different effects on metabolic traits are outlined in Supplementary Fig. 1 and, briefly, are as follows.

We performed a multivariate genome-wide association study (GWAS) of relevant metabolic biomarkers that were available in individuals of European ancestry from the UK Biobank, including HDL cholesterol (HDL) (n = 392,965), sex hormone–binding globulin (SHBG) (n = 389,354), triglycerides (n = 429,011), AST (n = 427,778), and ALT (n = 429,203), using BOLT-LMM v2.3.4 (9) and metaCCA software (10) as described previously (4). Specifically, metaCCA uses canonical correlation analysis to identify the maximal correlation coefficient between genome-wide genetic variants and a linear combination of the above phenotypes, based on the computed phenotype-phenotype Pearson correlation matrix. We chose these specific metabolic biomarkers to be consistent with our previous approach (4). These biomarkers are used to discriminate between three monogenic forms of insulin resistance: lipodystrophy (disorders of fat storage), monogenic obesity, and insulin signaling defects (6,11).

We identified 254 variants at P < 5 × 10⁻⁸ associated with both our univariate GWAS of body fat percentage (n = 620 variants previously published [4]) and our composite metabolic phenotype as estimated by the above multivariate GWAS model. This represents an increase in 221 signals compared with the 33 previously reported using a similar approach (4). This increase was largely attributable to the availability of the metabolic biomarkers in 451,099 individuals of European ancestry from UK Biobank, whereas previous studies were limited to smaller separate data (e.g., 100,000 with HDL and triglycerides, 21,800 with SHBG, and 55,500 with ALT).

We applied a k-means algorithm on the 254 variants and their effects on the values of the six phenotypes from the first step and used the parameter k = 3 to group them into FA and UFA. We considered a third cluster of “conflicting” to group any variants that do not belong to the FA or UFA clusters and did not pursue these variants in the rest of the analyses to minimize false discovery. Within UFA and FA clusters, we inspected whether the loci are driven by colocalization of signals from a combination of traits or represent a strong univariate signal.

To validate FA and UFA variants, we assessed their effects on risk of type 2 diabetes using data from GWAS of 31 studies, excluding UK Biobank, which included 55,005 case and 400,308 control subjects of European ancestry (12). We expected to observe adiposity-increasing alleles as associated with lower risk of type 2 diabetes for FA variants and higher risk of type 2 diabetes for UFA variants.

A subcohort of 100,000 subjects were selected for the imaging enhancement of the UK Biobank, currently at 49,938. Abdominal MRI scans were obtained with a MAGNETOM Aera 1.5T scanner (software version syngo MR D13) (Siemens Healthineers, Erlangen, Germany) (13). Image-derived phenotypes were generated from the three-dimensional Dixon neck-to-knee acquisition, the high-resolution T1-weighted three-dimensional pancreas acquisition, and liver and pancreas single-slice multiecho acquisitions. Images for this study were obtained through UK Biobank application no. 44584. Following automated preprocessing of the different sequences, volumes of organs of interest (including the liver, pancreas, and SAT and visceral adipose tissue [VAT]) were segmented using convolutional neural networks (14). Fat content of the liver and pancreas was obtained from the multiecho acquisitions after preprocessing where the proton density fat fraction was estimated (15).

We used the UK Biobank Imputed Genotypes v3 (16), excluding single nucleotide polymorphisms with minor allele frequency <1% and imputation quality <0.9. We excluded participants not recorded as European, exhibiting sex chromosome aneuploidy, with a discrepancy between genetic and self-reported sex, heterozygosity outliers, and genotype call rate outliers. We used BOLT-LMM (9) v2.3.2 to conduct the genetic association study. We included age at imaging visit, age squared, sex, imaging center, and genotyping batch as fixed-effects covariates, in addition to scan date scaled and scan time scaled and genetic relatedness derived from genotyped single nucleotide polymorphisms as a random effect to control for population structure and relatedness. We performed GWAS for VAT (n = 32,859), SAT (n = 32,859), VAT-to-SAT ratio (n = 32,859), pancreatic fat (n = 24,673), liver fat (n = 32,655), pancreas volume (n = 31,758) and liver volume (n = 32,859) (Supplementary Table 1).

To replicate the effect of FA and UFA variants on measures of adiposity, metabolic biomarkers, and C-reactive protein (CRP), we used summary statistics from published GWAS, which were independent of UK Biobank (Supplementary Table 1). To replicate the effects on subcutaneous and ectopic fat, we used data from a combined multiethnic sample size–weighted fixed-effects meta-analysis of SAT (n = 18,247), VAT (n = 18,332), VAT-to-SAT ratio (n = 18,191), and pericardial adipose tissue (n = 12,204) measured by computed tomography (CT) or MRI (17) (Supplementary Table 1).

We studied the association of individual variants and of genetic scores with cardiometabolic traits and diseases in the UK Biobank using our GWAS results. We performed GWAS with BOLT-LMM to account for population structure and relatedness using covariates such as age, sex, genotyping platform, and study center in the model. For genetic score analysis, we used the inverse variance–weighted method (IVW), assigning a weight of 1 to each variant. This method approximates the association of an unweighted genetic score (18).

We investigated the causal associations between FA and UFA using FA and UFA clusters as instruments and six cardiometabolic disease outcomes (type 2 diabetes, heart disease, hypertension, stroke, nonalcoholic fatty liver disease [NAFLD], and polycystic ovary syndrome [PCOS]) by performing two-sample MR analysis (19). We used IVW as our main analysis and MR-Egger and weighted median as sensitivity analyses in order to detect unidentified pleiotropy of our genetic instruments. We used two sources of data: FinnGen GWAS summary results (20) and published GWAS of the same diseases, excluding UK Biobank to separate it from our discovery data set and allow us to run two-sample MR (Supplementary Table 1). We performed MR within each data source and then meta-analyzed the results across the two data sets using a random-effects model with the R package metafor (21). We ran the same models in UK Biobank data for comparison. For more information on definition of diseases and ICD codes, please see Supplementary Material.

We used DEPICT (Data-Driven Expression-Prioritized Integration for Complex Traits) v.1 rel194 β (22) to identify tissues and cells in which the genes from associated loci are highly expressed. Using 37,427 human Affymetrix HGU133a2.0 platform microarrays, DEPICT assesses whether genes at the relevant loci are highly expressed in any of the 209 tissues, cell types, and physiological systems annotated by Medical Subject Headings (MeSH).

The data sets analyzed during the current study are available from FinnGen (20) and the relevant published GWAS (Supplementary Table 1). The data that support the findings of this study are available from UK Biobank, but restrictions apply to the availability of these data, which were used under license for the current study (UK Biobank project application nos. 9072, 9055, and 44584) and therefore are not publicly available. No applicable resources were generated or analyzed during the current study.

Among 254 variants associated (at P < 5 × 10⁻⁸) with both body fat percentage (4) and a composite metabolic phenotype, we identified two distinct clusters of adiposity variants: 1) 38 variants grouped as UFA and 2) 36 variants grouped as FA (Supplementary Tables 2–4 and Supplementary Fig. 2). UFA genetic score was associated with higher body fat percentage and higher BMI and an adverse metabolic profile including lower HDL and SHBG and higher triglycerides, ALT, and AST. FA genetic score was also associated with higher body fat percentage and BMI but, in contrast, a favorable metabolic profile including higher HDL and SHBG and lower triglycerides, ALT, and AST (Table 1 and Fig. 1A). There was no sex difference in the association of FA and UFA genetic scores with adiposity measures or biomarkers at the multiple testing–corrected significance threshold (0.05 of 44 tests = 0.0011) (Supplementary Table 5 and Supplementary Fig. 3). The association between UFA adiposity-increasing alleles and an adverse metabolic profile and FA adiposity-increasing alleles and a favorable metabolic profile was replicated in independent published GWAS of these biomarkers (Supplementary Table 6).

The mean (SD) UFA and FA genetic scores in the UK Biobank were 36.99 (3.83) and 37.32 (3.75) respectively. The distributions of UFA and FA genetic scores among the UK Biobank participants with and without type 2 diabetes are shown in Supplementary Fig. 4. The UFA and FA variants explained 0.6% and 0.2% variance in body fat percentage in the UK Biobank, respectively.

We used data from the latest GWAS of type 2 diabetes excluding UK Biobank (12) to validate the paradoxical association between the adiposity-increasing alleles at FA and UFA variants and risk of type 2 diabetes. Among 38 UFA variants, 33 adiposity-increasing alleles were correlated with higher risk of type 2 diabetes (P_{two-tailed binomial} = 4E−6), with 24 at P < 0.05. Among 36 FA variants, all adiposity-increasing alleles were correlated with lower risk of type 2 diabetes (P_{two-tailed binomial} = 3E−11), including 23 variants at P < 0.05 (Fig. 2). This paradoxical association was consistent with the pattern of association with type 2 diabetes with use of data from the UK Biobank (Supplementary Fig. 5).

To explore whether the UFA and FA variants represent biologically meaningful entities, we searched whether genes at the relevant loci were enriched for expression in certain tissues or pathways. FA variants were enriched (at false discovery rate <5%) in adipocyte-related cells and tissues and in physiological systems labeled as “digestive” (small intestine, esophagus, pancreas, upper gastrointestinal tract, and ileum) and “cardiovascular” (arteries), while UFA variants were enriched in mesenchymal stem cells and in physiological systems labeled as cardiovascular (aortic valve, heart valves) (Fig. 3 and Supplementary Tables 7 and 8).

To investigate the relation between UFA and FA variants and abdominal fat distribution, we looked at the effect of FA and UFA variants on SAT, VAT, and ectopic fat in the liver and pancreas in addition to liver and pancreas volume in 32,859 individuals of European ancestry from the UK Biobank. While both UFA and FA genetic scores were associated with higher SAT with similar effect size, UFA score was associated with higher liver and pancreatic fat, higher VAT, and increased liver volume, but FA score was associated with lower liver fat and smaller liver and pancreas volume and had no effect on pancreatic fat (Table 1 and Fig. 1B). Both UFA and FA genetic scores were associated with lower VAT-to-SAT ratio.

We replicated these associations using independent data from the published GWAS of abdominal fat as measured by CT or MRI in up to 18,332 individuals for some of the measured phenotypes. UFA genetic score was associated with higher SAT (P = 3E−8), higher VAT (P = 6E−7), and higher pericardial adipose tissue (P = 0.003) but had no effect on VAT-to-SAT ratio (P = 0.70), while FA genetic score was associated with higher SAT (P = 6E-6) and lower VAT-to-SAT ratio (P = 2E−6) and had no effect on VAT (P = 0.92) or pericardial adipose tissue (P = 0.50) (Supplementary Table 6).

There was no sex difference in the association between FA and UFA clusters and MRI-derived measures of fat distribution at the multiple testing–corrected significance threshold except for VAT, where FA score was associated with higher VAT in men versus lower VAT in women (P_difference = 0.0006 [Supplementary Table 5 and Supplementary Fig. 3]).

With the data from the UK Biobank MRI subcohort, among 38 UFA variants, adiposity-increasing alleles at 31 variants (P_{two-tailed binomial} = 0.0001) were correlated with higher ectopic liver fat, including 7 variants with P < 0.05 (Supplementary Fig. 6), and 31 adiposity-increasing alleles were correlated with higher pancreatic fat (P_{two-tailed binomial} = 0.0001), including 7 variants at P < 0.05 (Supplementary Fig. 7). Of the 36 FA variants, 29 adiposity-increasing alleles were correlated with lower liver fat (P_{two-tailed binomial} = 0.0003), including 9 variants at P < 0.05 (Supplementary Fig. 6). FA variants had a mixed effect on pancreatic fat, as only 14 adiposity-increasing alleles were correlated with lower pancreatic fat (P_{two-tailed binomial} = 0.24), including two alleles associated with higher and two with lower pancreatic fat at P < 0.05 (Supplementary Fig. 7).

Results on interesting individual FA variants with paradoxical effects where adiposity-increasing alleles are associated with lower risk of type 2 diabetes (from UK Biobank–independent GWAS) and lower liver fat (from the UK Biobank) at P < 0.05 are illustrated as forest plots in Supplementary Fig. 8. These include eight variants: rs4684847 (PPARG), rs12130231 (LYPLAL1/SLC30A10), rs11664106 (EMILIN2), rs13389219 (GRB14/COBLL1), rs2943653 (NYAP2/IRS1), rs30351 (ANKRD55), rs4450871 (CYTL1), and rs7133378 (DNAH10). Among these variants, the FA alleles at only two variants were associated with pancreatic fat at P < 0.05: near GRB14, with lower pancreatic fat, and near PPARG, with higher pancreatic fat.

To understand the role of inflammation in the mechanisms that link higher adiposity to risk of cardiometabolic disease, we looked at the association between FA and UFA variants and CRP levels as an inflammatory marker. In the UK Biobank, both UFA and FA genetic scores were associated with higher CRP (Table 1 and Fig. 1A). These associations were replicated using an independent GWAS of CRP (23) (Supplementary Table 6). There was no sex difference in the association between UFA and FA variants and CRP levels (Supplementary Table 5 and Supplementary Fig. 3). Of 38 UFA adiposity-increasing alleles, 35 (P_{two-tailed binomial} = 7E−8) and, of 36 FA adiposity-increasing alleles, 27 (P_{two-tailed binomial} = 0.004) were correlated with higher CRP, including 32 and 15 variants with P < 0.05, respectively (Supplementary Fig. 9). To further understand the role of higher adiposity on the association between UFA and FA genetic scores and higher CRP, we ran our statistical models, but we additionally adjusted for BMI or body fat percentage. This adjustment removed the association with higher CRP for both genetic scores, indicating that the effect was mediated by higher adiposity (Supplementary Table 9).

For UK Biobank data, UFA genetic score was associated with higher risk of type 2 diabetes (P = 6E−16), heart disease (P = 2E−7), hypertension (P = 4E−8), stroke (P = 0.0005), fatty liver disease (P = 0.004), and PCOS (P = 7E−5) (Table 1 and Fig. 1C). In contrast, FA genetic score was associated with lower risk of type 2 diabetes (P = 2E-9), heart disease (P = 3E−5), hypertension (P = 0.0001), fatty liver disease (P = 0.03), and PCOS (P = 0.02) (Table 1 and Fig. 1C). These findings were replicated with use of UK Biobank–independent GWAS data (Supplementary Table 6). There was no sex difference in the association of UFA and FA clusters with risk of disease at the multiple testing–corrected significance threshold (Supplementary Table 5 and Supplementary Fig. 3).

To understand the causal nature of these associations, we took an MR approach and used two UK Biobank–independent data sets (published GWAS and FinnGen). A 1-SD-higher genetically instrumented UFA was associated with higher risk of type 2 diabetes (IVW odds ratio 5.50 [95% CI 4.29, 7.05]; P = 4E−41), heart disease (1.66 [1.08, 2.54]; P = 0.02), hypertension (3.03 [2.18, 4.22]; P = 5E−11), stroke (1.43 [1.23, 1.67]; P = 3E−6), NAFLD (3.70 [1.22, 11.17]; P = 0.02), and PCOS (7.13 [3.66, 13.90]; P = 8E−9) (Table 2, Fig. 4, and Supplementary Table 10). In contrast, a 1-SD-higher genetically instrumented FA was associated with lower risk of type 2 diabetes (0.11 [0.08, 0.16]; P = 4E−33), heart disease (0.34 [0.25, 0.47]; P = 2E−11), hypertension (0.34 [0.21, 0.55]; P = 1E−5), stroke (0.65 [0.52, 0.83]; P = 0.0004), and NAFLD (0.14 [0.03, 0.79]; P = 0.03). There was a trend toward an association with lower odds of PCOS (0.51 [0.21, 1.23]; P = 0.13) but with wider CIs due to smaller sample size in the UK Biobank–independent GWAS (Table 2, Fig. 4, and Supplementary Table 10).

In this study, we used a unique genetic approach to understand the role of body adiposity in relation to the fat content and volumes of the liver and pancreas, as well as pathogenicity of cardiometabolic disease. We have identified two distinct clusters of variants associated with higher adiposity: one with a favorable metabolic profile (FA), consisting of 36 variants, and the other with an unfavorable metabolic profile (UFA), which included 38 variants. Although the adiposity-increasing alleles in both clusters are associated with increased SAT, the FA alleles are specifically associated with a lower liver fat and appear to provide protection against risk of cardiometabolic diseases, whereas the UFA alleles are associated with higher deposition of all fat depots including liver, pancreas, and visceral fat and are associated with higher risk of cardiometabolic disease.

The results of our genetic analysis support the observations from phenotyping studies that have proposed different obesity phenotypes related to the distribution of body fat (24,25). The two adiposity phenotypes that we have described in the current study highlight the role of SAT as a metabolic sink in obesity. In FA, this metabolic sink can accommodate excess triglycerides to specifically protect the liver from ectopic fat accumulation and prevent or delay pathogenic processes; in UFA, the excess triglycerides appear to exceed the capacity of the SAT metabolic sink and are consequently being deposited in alternate sites, including the VAT depot, liver, pancreas, and pericardial adipose tissue (26).

Our data, consistent with previous findings, provide evidence that accumulation of fat in the liver, which is an organ integral to glucose, insulin, and metabolism, directly contributes to the development of metabolic derangements associated with higher adiposity (27,28). Using a small subset of FA variants and a limited sample size with MRI scans (n = 9,510), we previously showed that FA alleles were associated with lower ectopic liver fat in women but not men (4). The availability of MRI scans of liver fat in 32,859 UK Biobank participants allowed us to demonstrate that there is no sex-specific association with liver fat. However, both FA and UFA variants had a greater overall effect on liver fat in women than men, which could indicate the confounding effects of other factors in the measured liver fat in men. The association of both FA and UFA clusters with, respectively, smaller and bigger liver size is most likely biased by the accumulation of liver fat (29).

The pattern of association between FA and UFA variants and MRI-derived measures of ectopic fat can help with understanding the role of each ectopic fat depot in the pathophysiology of cardiometabolic diseases. While the FA cluster is associated with lower ectopic liver fat, it has no effect on VAT. This is consistent with previous studies showing the effect of thiazolidinediones, a class of medicines that improve insulin sensitivity, on promotion of differentiation of new adipocytes in SAT without changing VAT (30). In the light of strong association between VAT and development of metabolic dysfunction (31), our data suggest that VAT may reflect the ectopic fat deposition in the liver (r between the two measures = 0.5 [14]) but itself may not be causally related to the development of cardiometabolic diseases. The sex-specific association between the FA cluster and lower VAT in women and higher VAT in men is also consistent with the more general sex-specific pattern of VAT distribution as previously shown by genetic studies of waist-to-hip ratio (WHR) (32). Furthermore, while FA alleles are associated with lower VAT in women and higher VAT in men, they are associated with protection from cardiometabolic diseases with similar effect size between men and women.

Although the UFA cluster was associated with higher pancreatic fat, we did not detect any association between the FA cluster and this fat depot. FA variants individually had a mixed effect on pancreatic fat with the FA allele near PPARG, which is the most prominent example of FA variants mimicking the effect of thiazolidinediones, being associated with higher pancreatic fat. The role of pancreatic fat in the pathogenicity of type 2 diabetes is currently not clear-cut. Although many cross-sectional studies have reported higher pancreatic fat in subjects with type 2 diabetes compared with age-matched control subjects (33–36), there are conflicting views regarding whether pancreatic fat is itself a driver of type 2 diabetes, with some studies showing no association between type 2 diabetes and pancreatic fat using either CT or histology at autopsy (37–40). Moreover, a recent study testing the so-called “twin cycle model” (liver and pancreatic fat) before and after the onset of type 2 diabetes showed that liver fat was the main mediator associated with glycemic control (41). Furthermore, a recent genetic study of pancreatic fat and liver fat in the UK Biobank showed that genetic variants associated with pancreatic fat did not have a significant impact on metabolic disease (14), suggesting that pancreatic fat has no direct role in pathogenicity of type 2 diabetes and other metabolic disease. Longitudinal imaging studies of individuals prior to and after clinical disease onset in addition to MR studies of pancreatic fat in type 2 diabetes and other metabolic diseases will help to unravel the cause and consequence of this relationship.

The FA cluster was associated with a smaller pancreas volume, whereas the UFA cluster had no association with pancreas volume. Studies of individuals with type 2 diabetes using CT or ultrasound have shown 7–33% lower pancreas volume compared with that of control subjects (39,42–44). Given that only 1–2% of the adult pancreas is composed of endocrine islets, changes in exocrine cell number may contribute more to lower pancreas volume as shown in studies of type 1 diabetes (45). Since insulin also acts as a growth stimulation hormone and maintains tissue mass (46,47), a decline in pancreas size in diabetes could be due to a loss of the trophic effect of insulin on exocrine cells (48,49). We previously showed that variants associated with FA are associated with lower fasting insulin levels (4–6), which could explain why the FA cluster was associated with a smaller pancreas volume.

Subclinical inflammation is another factor that has been shown to be associated with components of metabolic syndrome and vascular disease (50–54). Our data provided no evidence for any direct role of CRP in mechanisms that link FA and UFA clusters to, respectively, lower and higher risk of disease, since both clusters were associated with higher CRP levels, consistent with the findings of the largest MR study of CRP and risk of metabolic disease (23). We observed that the association between FA and UFA genetic scores and higher CRP levels disappeared after adjustment for adiposity (BMI or body fat percentage), indicating that their effect on higher CRP was mediated by higher adiposity. This pattern of association could suggest that higher CRP levels are secondary to higher adiposity. Data on other markers of inflammation, including tumor necrosis factor-α and interleukin-6, could clarify further the role of inflammation in cardiometabolic disease mechanisms.

Our tissue enrichment analysis provided further evidence that UFA and FA are biologically two different subtypes of adiposity. FA loci were enriched for genes expressed in adipose tissue and adipocytes, while UFA loci were enriched for genes expressed in mesenchymal stem cells. The enrichment of genes in adipose-related tissues and cells was previously shown for loci associated with WHR (55) in contrast to BMI loci enriched in the central nervous system (56). However, this is the first time mesenchymal stem cells have been highlighted in tissue enrichment analysis to be associated with adiposity. Mesenchymal stem cells are major sources of adipocyte generation, and in addition to adipose tissue, they also exist in skeletal muscle, the liver, and pancreas, which could suggest that they are responsible for ectopic fat formation in these organs (57,58). Further experiments and data are necessary to determine the relationship between mesenchymal stem cells and ectopic lipid accumulation.

There have been few approaches to identify variants associated with FA and UFA using different combinations of traits. Winkler et al. (59) used 159 variants associated with BMI, WHR, or WHR adjusted for BMI and described 24 FA variants as those associated with both lower WHR and higher BMI and 82 UFA variants as those associated with both higher WHR and higher BMI. Pigeyre et al. (60) used polygenic correlation between BMI and type 2 diabetes to identify genetic regions where BMI-increasing effect was linked to a corresponding increase, decrease, or neutral effect on type 2 diabetes risk. Our FA and UFA variants that overlap with these studies are listed in Supplementary Table 11. The main difference between our approach and these two studies is that they have started with variants associated with BMI or WHR as measures of adiposity. Recent studies have demonstrated that BMI is a poor proxy for body adiposity (61), particularly at an individual level, providing limited insight into body fat distribution, VAT, or nonadipose deposition of fat (25). For example, only 7 and 12 of 36 FA variants are associated with BMI and WHR, respectively. Similarly, while UFA variants are enriched for BMI variants, 7 variants are not associated with BMI and only 14 of 38 UFA variants are associated with WHR (Supplementary Fig. 10). Furthermore, using only BMI and type 2 diabetes to identify variants with opposite effects can induce index event bias (62), e.g., TCF7L2 (60).

There are several limitations to our study. First, we did not have independent studies to replicate the association with liver and pancreas fat and volume measurements. However, we used the largest data set on MRI phenotypes available from the UK Biobank, with 32,859 samples, and replicated the association with some fat depots available from a published GWAS (17). Second, our study population was limited to Europeans only; it is unclear how our findings can be generalized to other populations and whether they can explain the excess risk of cardiometabolic disease in non-Europeans (63). Third, we lacked data on ectopic fat accumulation in muscle in our samples; future studies of MRI-derived muscle fat in the UK Biobank will enable the role of this ectopic fat in the pathophysiology of cardiometabolic disease to be investigated. Fourth, lower-body subcutaneous fat mass in the gluteofemoral or leg region has previously been associated favorably with obesity-related cardiometabolic diseases (64). It would also have been of interest to study whether the FA cluster is protective of cardiometabolic diseases, particularly via increasing gluteofemoral and leg fat mass. Fifth, we used ALT and AST in our discovery pipeline to identify FA and UFA variants, which could have biased our findings toward those variants that influence liver fat more than pancreatic fat. Finally, in comparison with our previous study (4), we did not have fasting insulin and adiponectin in our composite metabolic phenotype, since these two biomarkers are not available in the UK Biobank. However, our 36 FA variants include all 14 variants previously identified as FA (4) and the comparison of the multivariate GWAS P values for these variants (Supplementary Fig. 11) indicates additional power gained in the current study largely attributable to the availability of other metabolic biomarkers in 451,099 individuals in a single cohort, the UK Biobank.

One of the major strengths of this study was the unique approach to understanding different mechanisms underlying the association between adiposity and risk of cardiometabolic diseases. This unique approach is coupled with gold standard measurements of organ volume and content from MRI scans for understanding the role of different fat depots in pathogenicity. The availability of the UK Biobank made it possible to study the sex-specific effects of our variants against metabolic biomarkers, MRI measures of fat distribution and ectopic fat, and risk of disease. We used the largest published GWAS and FinnGen study and independently replicated our results against risk of disease and performed MR studies. Unlike previous studies that examined the role of ectopic fat and pancreas size in small groups of participants categorized by diabetes status, we investigated the role of these phenotypes in a population-based study of 32,859 participants, which minimizes the effect of confounding factors and statistical bias. Finally, our sets of FA and UFA variants provide two important genetic instruments for any MR study to examine the causal role of adiposity on risk of disease uncoupled from its metabolic effect.

This study provides genetic evidence for two types of adiposity: one coupled with a favorable metabolic profile and the other with an unfavorable profile. Both FA and UFA variants were associated with higher CRP levels. We demonstrated that reduced liver fat, but not VAT or pancreatic fat, is on the pathway that links FA to lower risk of diseases related to metabolic syndrome. We determined no sexual dimorphism in the way the FA and UFA variants are associated with metabolic profile, abdominal fat distribution, and risk of disease. Future MR, longitudinal, and independent studies are required to elucidate whether higher pancreatic fat and smaller pancreas volume are a consequence of the ongoing pathological processes or causal of cardiometabolic disease outcomes. Better understanding of the difference between FA and UFA may lead to new insights in preventing and predicting, and treating people who suffer from, cardiometabolic diseases.

Acknowledgments. This research has been conducted using data from the UK Biobank resource and carried out under UK Biobank project application numbers 9072, 9055, and 44584. UK Biobank protocols were approved by the National Research Ethics Service Committee. The authors acknowledge the participants and investigators of the FinnGen study. The authors thank Amoolya Singh and Kevin Wright, Calico Life Sciences LLC, for their feedback on the manuscript. The authors would like to acknowledge the use of the University of Exeter High-Performance Computing (HPC) facility in carrying out this work. We acknowledge use of high-performance computing funded by an MRC Clinical Research Infrastructure award (MRC Grant: MR/M008924/1).

Funding. S.M. is funded by the Medical Research Council. H.Y. is funded by a Diabetes UK RD Lawrence fellowship (17/0005594). J.T. is supported by an Academy of Medical Sciences (AMS) Springboard Award, which is supported by the AMS, the Wellcome Trust, the Global Challenges Research Fund, the U.K. Government Department of Business, Energy & Industrial Strategy, the British Heart Foundation, and Diabetes UK (SBF004\1079).

Duality of Interest. M.C., E.S., and Y.L. are employees of Calico Life Sciences LLC. M.C., E.S., and Y.L. are funded by Calico Life Sciences LLC. No other potential conflicts of interest relevant to this article were reported.

The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Authors Contributions. S.M. performed the statistical analyses. S.M. and H.Y. designed the study and wrote the first draft of the manuscript. N.B., B.W., Y.L., J.D.B., and E.L.T. created the MRI-derived phenotypes and contributed to the writing of the manuscript. M.C. and E.S. performed the GWAS of MRI-derived phenotypes. J.T., R.N.B., A.R.W., and T.M.F. contributed to the analysis of biomarkers from the UK Biobank. H.Y. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.