We aimed to unravel the mechanisms connecting adiposity to type 2 diabetes. We used MR-Clust to cluster independent genetic variants associated with body fat percentage (388 variants) and BMI (540 variants) based on their impact on type 2 diabetes. We identified five clusters of adiposity-increasing alleles associated with higher type 2 diabetes risk (unfavorable adiposity) and three clusters associated with lower risk (favorable adiposity). We then characterized each cluster based on various biomarkers, metabolites, and MRI-based measures of fat distribution and muscle quality. Analyzing the metabolic signatures of these clusters revealed two primary mechanisms connecting higher adiposity to reduced type 2 diabetes risk. The first involves higher adiposity in subcutaneous tissues (abdomen and thigh), lower liver fat, improved insulin sensitivity, and decreased risk of cardiometabolic diseases and diabetes complications. The second mechanism is characterized by increased body size and enhanced muscle quality, with no impact on cardiometabolic outcomes. Furthermore, our findings unveil diverse mechanisms linking higher adiposity to higher disease risk, such as cholesterol pathways or inflammation. These results reinforce the existence of adiposity-related mechanisms that may act as protective factors against type 2 diabetes and its complications, especially when accompanied by reduced ectopic liver fat.
The relationship between excess adiposity and type 2 diabetes is complex.
Can genetic subtypes of adiposity reveal distinct pathways linking adiposity with type 2 diabetes?
Higher adiposity increases type 2 diabetes risk via different mechanisms (e.g., cholesterol pathways or inflammation) but decreases risk via other mechanisms (lower liver fat and improved insulin sensitivity, or increased body size and enhanced muscle quality).
These insights could improve precision medicine for type 2 diabetes via treating adiposity.
Introduction
The strong link between excess weight (adiposity) and type 2 diabetes emphasizes the crucial role of weight management in prevention and treatment (1). However, the complex nature of type 2 diabetes and adiposity, influenced by genetics and lifestyle, poses challenges. This complexity leads to variations in insulin resistance, production, and fat accumulation in ectopic places (liver, skeletal muscles, and pancreas) (2), making tailored weight management for diabetes challenging (3,4). While weight loss benefits glycemic control and health, responses vary among individuals (5,6), underscoring the need for personalized interventions.
Individuals with the same overall adiposity also have different risks of developing cardiometabolic disease (7,8). Reporting adiposity using surrogates, such as BMI, has limitations in distinguishing fat and lean mass or accounting for variations in fat distribution, for example, between the metabolically benign subcutaneous fat and more metabolically harmful visceral fat, and across different ethnicities (9,10). The current strategy for managing obesity in individuals with type 2 diabetes relies on using crude cutoffs for BMI and metabolic measures such as HbA1c or blood pressure. There is a need to create a reliable subtype classification system that accounts for the underlying causal factors that connect adiposity and type 2 diabetes to allow more accurate predictions of the benefits of intentional weight loss.
Research on adiposity subtype classification has primarily focused on metabolically healthy obesity, a condition with multiple definitions where individuals with obesity may not immediately exhibit metabolic dysfunction (11,12). Other approaches have involved behavioral traits, BMI, HbA1c, cardiometabolic traits and machine-learning techniques (13,14). However, these studies often relied on traits secondary to obesity or diabetes, introducing potential confounding from correlated factors and limiting their biological or clinical significance. In contrast, approaches that integrate genetic data allow clustering based on risk factors present at birth and unaffected by treatment, distinct from clinical biomarkers. In our previous work, we combined genetics with machine learning to identify two adiposity phenotypes with opposing effects on type 2 diabetes risk (15). Yet, including metabolic biomarkers, such as liver-specific enzymes, in our model might introduce circular arguments, potentially biasing findings toward specific aspects, such as variants influencing liver fat.
In this study, we hypothesized that distinct biological pathways link higher adiposity with type 2 diabetes risk. We first selected variants associated with measures of adiposity. We next used MR-Clust (16) to categorize adiposity variants based on their causal links to type 2 diabetes. MR-Clust groups variants with similar effect estimates, operating on the premise that an exposure (e.g., adiposity) can impact an outcome (e.g., type 2 diabetes) through diverse causal mechanisms with varying degrees. MR-Clust includes a provision to address potential spurious clusters by classifying variants with uncertain causal effect estimates into “null” or “junk” clusters. This methodology was previously applied to cluster IGF-1–associated variants based on their causal associations with type 2 diabetes (17). We then used different biomarkers, including metabolites, lipids, insulin sensitivity and secretion measures, and inflammatory cytokines, to characterize metabolic signatures of each cluster. To further investigate the difference between clusters, we quantified the genetic effect of each cluster on body composition and adipose tissue distribution measured using MRI. Finally, we estimated the causal effect of higher adiposity through each cluster on different diseases, including those common in people with type 2 diabetes, using Mendelian randomization (MR).
Research Design and Methods
Study Design
Supplementary Fig. 1 summarizes our study design. To identify distinct causal pathways that link adiposity to type 2 diabetes, we first used independent genetic variants associated with two measures of adiposity—body fat percentage (BFP) and BMI. Although BMI does not represent adiposity accurately (9), it is by far the most commonly used metric to categorize people with obesity; therefore, it is a useful measure to compare with BFP. Second, we clustered these genetic variants based on their effect on type 2 diabetes risk (18). Each cluster represents a different causal pathway from adiposity to type 2 diabetes risk. Third, we validated the effect of each cluster on type 2 diabetes risk using FinnGen (Data Freeze 8 [19]) as an independent cohort. Fourth, to find the metabolic signature of each cluster, we calculated cluster-specific genetic risk score effects on different biomarkers. Fifth, we calculated the causal effect of higher adiposity using MR through each cluster on different diseases, including those prevalent in type 2 diabetes.
Identification of Distinct Causal Pathways
To identify distinct causal pathways linking adiposity to type 2 diabetes, we used MR-Clust (16). This method calculates the MR estimate for each genetic variant as the ratio of the genetic association with the outcome (type 2 diabetes) divided by the genetic association with the exposure measure (adiposity) and seeks to find clusters of variants with similar estimates by maximizing the likelihood of a mixture of normal distributions. By convention, a genetic variant is only assigned to a cluster if the estimated probability of cluster membership is >80%; if lower than this, then the variant is not assigned to any cluster. The motivation is that variants with similar MR estimates are likely to influence the outcome via similar mechanisms.
Data Source
We used published genome-wide association studies (GWAS) summary statistics from the largest and latest studies for traits of interest (anthropometric traits, clinical biomarkers, insulin sensitivity and secretion measures, metabolites, and inflammatory markers and cytokines), focusing on European-specific data (Table 1). For measures of adiposity, we accessed the GWAS of BFP from the Integrative Epidemiology Unit OpenGWAS project (20), where BFP had been estimated by impedance measurement in the UK Biobank (21), using the R package ieugwasr (n = 454,633). For BMI, we used the latest meta-analysis of the Genetic Investigation of ANthropometric Traits (GIANT) consortium and UK Biobank (n = 806,834) (22). To determine adiposity variant clusters, we used European-specific data from the DIAbetes Meta-ANalysis of Trans-Ethnic (DIAMANTE) type 2 diabetes GWAS (80,154 case subjects vs. 853,816 control subjects) (18). For the second type 2 diabetes data set and disease outcomes, we used data from FinnGen Data Freeze 8 or 7 (19).
Trait/disease . | PubMed ID . | Sample size (case subjects/control subjects for disease if available) . | Ethnicity . | Reference . |
---|---|---|---|---|
Cytokines and growth factors | 27989323, 33491305 | 8,293 | EUR | Ahola-Olli AV, et al., AJHG, 2017; Kalaoja, M et al., Obesity, 2021 |
Metabolites | 35692035 | 115,078 | EUR | Borges CM, et al., BMC Medicine, 2022. Accessed via IEU OpenGWAS ID: met-d-* |
Childhood obesity | 31504550 | 24,160 | EUR | Bradfield JP, et al., Human Molecular Genetics, 2019 |
Childhood BMI | 26604143 | 35,668 | EUR | Felix JF, et al., Human Molecular Genetics, 2016 |
HbA1c | 34059833 | 281,416 | EUR | Chen J, et al., Nature Genetics, 2021 |
Adiponectin | 22479202 | 45,891 (AA n = 4,232, EAS n = 1,776, EUR n = 29,347) | AA, EAS, EUR | Dastani Z, et al., PLoS Genetics, 2012 |
HOMA-B, HOMA-IR | 20081858 | 46,186 | EUR | Dupuis J, et al., Nature Genetics, 2010 |
HDL, LDL, and non-HDL cholesterol, total cholesterol, triglycerides | 34887591, 36575460, 35931049 | 1,320,000 | EUR | Graham SE, et al., Nature, 2021; Kanoni S, et al., Genome Biology, 2022; Ramdas S, et al., AJHG, 2022 |
Leptin | 26833098 | 32,161 | EUR | Kilpeläinen TO, et al., Nature Communications, 2016 |
Fasting glucose, fasting insulin | 33558525 | 140,595, 98,210 | EUR | Lagou V, et al., Nature Communications, 2021 |
Type 2 diabetes | 35551307 | 80,154/853,816 | EUR | Mahajan A, et al., Nature Genetics, 2022 |
Liver enzymes (ALP, ALT, GGT) | 33972514 | 437,438, 437,267, 437,194 | EUR | Pazoki R, et al., Nature Communications, 2021 |
Disposition index, corrected insulin response, insulin at 30 min, incremental insulin at 30 min | 24699409 | 5,318 | EUR | Prokopenko I, et al., PLoS Genetics, 2014 |
Adult BMI, waist-to-hip ratio (female), waist-to-hip ratio (male) | 30239722 | 806,834, 379,501, 315,284 | EUR | Pulit SL, et al., Human Molecular Genetics, 2019 |
Fasting proinsulin | 21873549 | 27,079 | EUR | Strawbridge RJ, et al., Diabetes, 2011 |
Insulin sensitivity index | 27416945 | 16,753 | EUR | Walford GA, et al., Diabetes, 2016 |
Birth weight | 31043758 | 298,142 | EUR | Warrington NM, et al., Nature Genetics, 2019 |
Adult height | 36224396 | 4,080,687 | EUR | Yengo L, et al., Nature, 2022 |
BFP | NA | 454,633 | EUR | Elsworth B, 2018. Accessed via IEU OpenGWAS ID: ukb-b-8909 |
C-reactive protein | 30388399 | 204,402 | EUR | Ligthart S, AJHG, 2018. Accessed via IEU OpenGWAS ID: ieu-b-35 |
Whole-body fat-free mass | NA | 454,850 | EUR | Elsworth B. 2018. Accessed via IEU OpenGWAS ID: ukb-b-13354 |
Sex hormone-binding globulin (female) | NA | 214,989 | EUR | Richmond R., 2020. Accessed via IEU OpenGWAS ID: ieu-b-4870 |
Sex hormone-binding globulin (male) | NA | 185,221 | EUR | Richmond, R., 2020. Accessed via IEU OpenGWAS ID: ieu-b-4871 |
FinnGen Data Freeze 8 disease outcomes | 36653562 | 342,499 | EUR | Kurki MI, et al., medRxiv, 2022 |
Type 2 diabetes | 49,114/283,207 | |||
Diabetic retinopathy | 8,942/283,545 | |||
Diabetic nephropathy | 3,676/283,456 | |||
Diabetic neuropathy | 2,444/249,480 | |||
Hypertension | 81,138/243,756 | |||
Polycystic ovary syndrome | 1,196/181,796 | |||
Nonalcoholic fatty liver disease | 1,908/340,591 | |||
Ischemic heart disease | 56,730/285,769 | |||
Stroke | 34,560/249,480 | |||
Atherosclerosis (excl. cerebral, coronary, and PAD) | 13,434/317,899 | |||
Heart failure | 23,622/317,939 | |||
Atrial fibrillation | 40,594/168,000 | |||
Chronic kidney disease | 7,916/330,300 | |||
Venous thromboembolism | 17,048/325,451 | |||
Deep vein thrombosis | 8,077/295,014 | |||
Pulmonary embolism | 8,170/333,487 | |||
Aortic aneurysm | 7,603/317,899 | |||
Gout | 7,461/221,323 | |||
Osteoarthritis (knee) | 39,343/221,323 | |||
Osteoarthritis (hip) | 17,536/324,963 | |||
Osteoporosis | 6,303/325,717 | |||
Rheumatoid arthritis | 11,178/221,323 | |||
Gallstones | 32,894/301,383 | |||
Gastroesophageal reflux disease | 22,867/292,256 | |||
Depression | 38,225/299,886 | |||
Psoriasis | 8,075/330,975 | |||
Asthma | 37,253/187,112 | |||
Intrahepatic liver and bile duct cancer | 648/259,583 | |||
Colorectal cancer | 5,458/259,583 | |||
FinnGen Data Freeze 7 disease outcomes | ||||
PAD | 11,924/288,638 |
Trait/disease . | PubMed ID . | Sample size (case subjects/control subjects for disease if available) . | Ethnicity . | Reference . |
---|---|---|---|---|
Cytokines and growth factors | 27989323, 33491305 | 8,293 | EUR | Ahola-Olli AV, et al., AJHG, 2017; Kalaoja, M et al., Obesity, 2021 |
Metabolites | 35692035 | 115,078 | EUR | Borges CM, et al., BMC Medicine, 2022. Accessed via IEU OpenGWAS ID: met-d-* |
Childhood obesity | 31504550 | 24,160 | EUR | Bradfield JP, et al., Human Molecular Genetics, 2019 |
Childhood BMI | 26604143 | 35,668 | EUR | Felix JF, et al., Human Molecular Genetics, 2016 |
HbA1c | 34059833 | 281,416 | EUR | Chen J, et al., Nature Genetics, 2021 |
Adiponectin | 22479202 | 45,891 (AA n = 4,232, EAS n = 1,776, EUR n = 29,347) | AA, EAS, EUR | Dastani Z, et al., PLoS Genetics, 2012 |
HOMA-B, HOMA-IR | 20081858 | 46,186 | EUR | Dupuis J, et al., Nature Genetics, 2010 |
HDL, LDL, and non-HDL cholesterol, total cholesterol, triglycerides | 34887591, 36575460, 35931049 | 1,320,000 | EUR | Graham SE, et al., Nature, 2021; Kanoni S, et al., Genome Biology, 2022; Ramdas S, et al., AJHG, 2022 |
Leptin | 26833098 | 32,161 | EUR | Kilpeläinen TO, et al., Nature Communications, 2016 |
Fasting glucose, fasting insulin | 33558525 | 140,595, 98,210 | EUR | Lagou V, et al., Nature Communications, 2021 |
Type 2 diabetes | 35551307 | 80,154/853,816 | EUR | Mahajan A, et al., Nature Genetics, 2022 |
Liver enzymes (ALP, ALT, GGT) | 33972514 | 437,438, 437,267, 437,194 | EUR | Pazoki R, et al., Nature Communications, 2021 |
Disposition index, corrected insulin response, insulin at 30 min, incremental insulin at 30 min | 24699409 | 5,318 | EUR | Prokopenko I, et al., PLoS Genetics, 2014 |
Adult BMI, waist-to-hip ratio (female), waist-to-hip ratio (male) | 30239722 | 806,834, 379,501, 315,284 | EUR | Pulit SL, et al., Human Molecular Genetics, 2019 |
Fasting proinsulin | 21873549 | 27,079 | EUR | Strawbridge RJ, et al., Diabetes, 2011 |
Insulin sensitivity index | 27416945 | 16,753 | EUR | Walford GA, et al., Diabetes, 2016 |
Birth weight | 31043758 | 298,142 | EUR | Warrington NM, et al., Nature Genetics, 2019 |
Adult height | 36224396 | 4,080,687 | EUR | Yengo L, et al., Nature, 2022 |
BFP | NA | 454,633 | EUR | Elsworth B, 2018. Accessed via IEU OpenGWAS ID: ukb-b-8909 |
C-reactive protein | 30388399 | 204,402 | EUR | Ligthart S, AJHG, 2018. Accessed via IEU OpenGWAS ID: ieu-b-35 |
Whole-body fat-free mass | NA | 454,850 | EUR | Elsworth B. 2018. Accessed via IEU OpenGWAS ID: ukb-b-13354 |
Sex hormone-binding globulin (female) | NA | 214,989 | EUR | Richmond R., 2020. Accessed via IEU OpenGWAS ID: ieu-b-4870 |
Sex hormone-binding globulin (male) | NA | 185,221 | EUR | Richmond, R., 2020. Accessed via IEU OpenGWAS ID: ieu-b-4871 |
FinnGen Data Freeze 8 disease outcomes | 36653562 | 342,499 | EUR | Kurki MI, et al., medRxiv, 2022 |
Type 2 diabetes | 49,114/283,207 | |||
Diabetic retinopathy | 8,942/283,545 | |||
Diabetic nephropathy | 3,676/283,456 | |||
Diabetic neuropathy | 2,444/249,480 | |||
Hypertension | 81,138/243,756 | |||
Polycystic ovary syndrome | 1,196/181,796 | |||
Nonalcoholic fatty liver disease | 1,908/340,591 | |||
Ischemic heart disease | 56,730/285,769 | |||
Stroke | 34,560/249,480 | |||
Atherosclerosis (excl. cerebral, coronary, and PAD) | 13,434/317,899 | |||
Heart failure | 23,622/317,939 | |||
Atrial fibrillation | 40,594/168,000 | |||
Chronic kidney disease | 7,916/330,300 | |||
Venous thromboembolism | 17,048/325,451 | |||
Deep vein thrombosis | 8,077/295,014 | |||
Pulmonary embolism | 8,170/333,487 | |||
Aortic aneurysm | 7,603/317,899 | |||
Gout | 7,461/221,323 | |||
Osteoarthritis (knee) | 39,343/221,323 | |||
Osteoarthritis (hip) | 17,536/324,963 | |||
Osteoporosis | 6,303/325,717 | |||
Rheumatoid arthritis | 11,178/221,323 | |||
Gallstones | 32,894/301,383 | |||
Gastroesophageal reflux disease | 22,867/292,256 | |||
Depression | 38,225/299,886 | |||
Psoriasis | 8,075/330,975 | |||
Asthma | 37,253/187,112 | |||
Intrahepatic liver and bile duct cancer | 648/259,583 | |||
Colorectal cancer | 5,458/259,583 | |||
FinnGen Data Freeze 7 disease outcomes | ||||
PAD | 11,924/288,638 |
AA, African American; ALP, alkaline phosphatase; EAS, East Asian; EUR, European; GGT, γ-glutamyl transferase; IR, insulin resistance; NA, not applicable; PAD, peripheral artery disease; met-d-*, *represents unique metabolite GWAS identification number.
Studies of MRI Scans
The UK Biobank MRI abdominal protocol has previously been reported (23). We used the neck-to-knee Dixon MRI and single-slice multiecho MRI in the abdomen. Dedicated image processing using deep learning models trained on 100+ manually annotated structures, achieved DICE scores >0.8 for each organ (24–27). Image-derived phenotypes (IDPs) from these segmentations include volume, and median proton density fat fraction (PDFF), which was calculated from the Phase Regularized Estimation using Smoothing and Constrained Optimization (PRESCO) method (28). Quality control involved evaluating univariate distributions and visually inspecting scans with extreme values.
Supplementary Table 1 summarizes the 15 IDPs used in this study, including subcutaneous adipose tissue (SAT) volumes (abdominal and thigh), visceral adipose tissue (VAT) volumes, internal fat and thigh intermuscular adipose tissue volumes (corrected for muscle volume), iliopsoas and total muscle volumes (indexed to height2), and organ volumes (kidney, pancreas, liver, and spleen). We computed the VAT-to-abdominal SAT ratio. We also obtained a measure of fat (PDFF) stored in the liver, pancreas, and the paraspinal muscles (intramyocellular fat) from the single-slice multiecho acquisition.
GWAS for the IDPs were performed using REGENIE version 3.1.1 (29). We included participants self-identified as “White British” and clustering as such in principal components analysis, excluding anomalies related to sex, heterozygosity, missingness, and genotype call rate (21). Sample sizes ranged from 28,587 to 37,589. Age, age squared, sex, genotyping array, imaging center, and the first 10 principal components of the genotype relatedness matrix were included. Phenotypes were inverse normal transformed. Imputed single nucleotide polymorphisms (SNPs) were filtered to minor allele frequency >0.01 and INFO score (a measure of imputation quality) >0.9, leaving 9,788,243 SNPs included in the final association study.
Genetic Risk Score Analysis
To calculate genetic risk score effects, we the extracted effect size estimate (β) and its corresponding SE for each variant from trait GWAS summary statistics. For missing variants, we obtained proxies (r2 ≥ 0.8) using the European reference panel from the 1000 Genomes Project Phase 3 (1000G EUR). We aligned all effects for the adiposity-increasing alleles. We performed a random-effect meta-analyses approach using the “rma” function in the R package metafor to calculate the effect of each genetic risk score as previously described (30). To account for multiple testing, we used a Benjamini-Hochberg–adjusted P value <0.05 to highlight significant associations.
MR Analysis
To best estimate the causal effects of each cluster on disease outcomes, we performed MR analyses in R 4.2.2 using the TwoSampleMR package (31,32). The inverse variance weighted method (IVW) was our main test. We used MR-Egger as a sensitivity analysis method to identify horizontal pleiotropy based on the Egger intercept. Additionally, we used weighted median, simple mode, and weighted mode (33). For missing variants, we calculated proxies (r2 ≥ 0.8) using 1000G EUR Phase 3. To account for multiple testing, we used a Benjamini-Hochberg–adjusted P value <0.05 to highlight significant causal associations.
Pathway Enrichment Analysis
For each cluster, we first used the SNP2GENE function in Functional Mapping and Annotation (FUMA) (34) to identify expression quantitative trait loci (eQTL) using Genotype-Tissue Expression (GTEx) v8 (35) and default settings. Genes identified through SNP2GENE were input into the Protein ANalysis THrough Evolutionary Relationships (PANTHER) v17.0 tool for pathway enrichment analysis (36).
eQTL Comparison in Adipose and Brain Tissue
To compare the number of independent eQTLs within each cluster in subcutaneous adipose, visceral adipose, and brain tissue, eQTLs were identified using FUMA and then clumped using 1000G EUR Phase 3, using a moderate cut of r2 ≥ 0.1 within 10,000-kilobase windows. Data sources for tissues were Multiple Tissue Human Expression Resource (MuTHER) and GTEx v8.
Data and Resource Availability
All data supporting the findings of this study are available within the paper and its supplementary information. Publicly available GWAS summary statistics are available online.
Results
Clusters of Adiposity Genetic Variants
The adiposity-increasing alleles had a considerable heterogeneous effect on type 2 diabetes risk (Supplementary Fig. 2). There was also significant heterogeneity in causal effects from MR-IVW results among instruments for both BFP and BMI (Cochran's Q statistic P value <1e−150 and 1.29e−140, respectively), suggesting that distinct causal pathways exist between adiposity and type 2 diabetes.
Using MR-Clust, we identified five clusters of BFP-increasing alleles representing five different causal pathways (Fig. 1A). Three clusters, comprising 7 variants in BFP-C1, 101 in BFP-C2, and 14 in BFP-C3, indicated a positive causal effect on type 2 diabetes risk, aligning with “unfavorable adiposity” (higher adiposity, adverse metabolic profile, higher disease risk (15) (Supplementary Table 2). Conversely, two BFP clusters (BFP-C4 with 13 variants and BFP-C5 with 9 variants) suggested a strong negative causal effect, consistent with “favorable adiposity” (higher adiposity, favorable metabolic profile, lower disease risk [15]). Among BFP-C1, BFP-C2, and BFP-C3, two, five, and three variants, respectively, were previously associated with unfavorable adiposity (15). Among BFP-C4 and BFP-C5, four variants in each cluster were previously associated with favorable adiposity (Table 2) (15). The higher number of previously known favorable and unfavorable adiposity variants among BFP clusters is anticipated, as the earlier study exclusively used variants associated with BFP to identify these groups.
Chr:pos (b37) . | rsID . | Adiposity-increasing allele . | Other allele . | Cluster . | Novel? (yes/no) . | Nearest gene . |
---|---|---|---|---|---|---|
1:203527812 | rs2802774 | A | C | BFP-C4 | No | OPTC–[]–ATP2B4 |
2:135597628 | rs10496731 | T | G | BFP-C4 | Yes | ACMSD |
3:123062657 | rs9814758 | T | G | BFP-C4 | Yes | ADCY5 |
3:171833266 | rs4894808 | G | C | BFP-C4 | Yes | FNDC3B |
9:136929586 | rs55924785 | C | T | BFP-C4 | Yes | BRD3 |
11:27487992 | rs11030016 | T | C | BFP-C4 | Yes | LGR4 |
12:121709430 | rs75412871 | C | T | BFP-C4 | Yes | CAMKK2 |
12:124409502 | rs7133378 | A | G | BFP-C4 | No | DNAH10 |
15:31689543 | rs12441543 | A | G | BFP-C4 | No | KLF13 |
18:2846812 | rs11664106 | T | A | BFP-C4 | No | SMCHD1–[]–EMILIN2 |
19:34008600 | rs33836 | C | T | BFP-C4 | Yes | PEPD |
19:46182304 | rs10423928 | T | A | BFP-C4 | Yes | GIPR |
22:38599767 | rs4820323 | C | G | BFP-C4 | Yes | MAFF/PLA2G6 |
1:219744138 | rs2785988 | A | C | BFP-C5 | Yes | []–ZC3H11B |
2:165528876 | rs13389219 | T | C | BFP-C5 | No | COBLL1 |
3:12393125 | rs1801282 | G | C | BFP-C5 | Yes | PPARG |
3:64718258 | rs2371767 | C | G | BFP-C5 | Yes | ADAMTS9–[] |
4:89726283 | rs2276936 | A | C | BFP-C5 | Yes | FAM13A |
6:43757896 | rs998584 | C | A | BFP-C5 | No | VEGFA |
6:127003464 | rs853961 | T | G | BFP-C5 | Yes | CENPW–[]–RSPO3 |
7:130466854 | rs972283 | A | G | BFP-C5 | No | KLF14–[]–MKLN1 |
7:150542711 | rs6977416 | G | A | BFP-C5 | No | AOC1 |
1:11284336 | rs10779751 | A | G | BMI-C3 | Yes | MTOR |
3:48085349 | rs11919665 | A | T | BMI-C3 | Yes | MAP4 |
6:130384187 | rs9375702 | C | T | BMI-C3 | Yes | L3MBTL3 |
7:93085722 | rs2283006 | A | G | BMI-C3 | Yes | CALCR |
12:122963550 | rs12369179 | C | T | BMI-C3 | No | ZCCHC8 |
14:91512339 | rs1951455 | C | T | BMI-C3 | Yes | RPS6KA5 |
19:46180184 | rs11672660 | C | T | BMI-C3 | Yes | GIPR |
20:62691550 | rs6512302 | C | G | BMI-C3 | Yes | TCEA2 |
Chr:pos (b37) . | rsID . | Adiposity-increasing allele . | Other allele . | Cluster . | Novel? (yes/no) . | Nearest gene . |
---|---|---|---|---|---|---|
1:203527812 | rs2802774 | A | C | BFP-C4 | No | OPTC–[]–ATP2B4 |
2:135597628 | rs10496731 | T | G | BFP-C4 | Yes | ACMSD |
3:123062657 | rs9814758 | T | G | BFP-C4 | Yes | ADCY5 |
3:171833266 | rs4894808 | G | C | BFP-C4 | Yes | FNDC3B |
9:136929586 | rs55924785 | C | T | BFP-C4 | Yes | BRD3 |
11:27487992 | rs11030016 | T | C | BFP-C4 | Yes | LGR4 |
12:121709430 | rs75412871 | C | T | BFP-C4 | Yes | CAMKK2 |
12:124409502 | rs7133378 | A | G | BFP-C4 | No | DNAH10 |
15:31689543 | rs12441543 | A | G | BFP-C4 | No | KLF13 |
18:2846812 | rs11664106 | T | A | BFP-C4 | No | SMCHD1–[]–EMILIN2 |
19:34008600 | rs33836 | C | T | BFP-C4 | Yes | PEPD |
19:46182304 | rs10423928 | T | A | BFP-C4 | Yes | GIPR |
22:38599767 | rs4820323 | C | G | BFP-C4 | Yes | MAFF/PLA2G6 |
1:219744138 | rs2785988 | A | C | BFP-C5 | Yes | []–ZC3H11B |
2:165528876 | rs13389219 | T | C | BFP-C5 | No | COBLL1 |
3:12393125 | rs1801282 | G | C | BFP-C5 | Yes | PPARG |
3:64718258 | rs2371767 | C | G | BFP-C5 | Yes | ADAMTS9–[] |
4:89726283 | rs2276936 | A | C | BFP-C5 | Yes | FAM13A |
6:43757896 | rs998584 | C | A | BFP-C5 | No | VEGFA |
6:127003464 | rs853961 | T | G | BFP-C5 | Yes | CENPW–[]–RSPO3 |
7:130466854 | rs972283 | A | G | BFP-C5 | No | KLF14–[]–MKLN1 |
7:150542711 | rs6977416 | G | A | BFP-C5 | No | AOC1 |
1:11284336 | rs10779751 | A | G | BMI-C3 | Yes | MTOR |
3:48085349 | rs11919665 | A | T | BMI-C3 | Yes | MAP4 |
6:130384187 | rs9375702 | C | T | BMI-C3 | Yes | L3MBTL3 |
7:93085722 | rs2283006 | A | G | BMI-C3 | Yes | CALCR |
12:122963550 | rs12369179 | C | T | BMI-C3 | No | ZCCHC8 |
14:91512339 | rs1951455 | C | T | BMI-C3 | Yes | RPS6KA5 |
19:46180184 | rs11672660 | C | T | BMI-C3 | Yes | GIPR |
20:62691550 | rs6512302 | C | G | BMI-C3 | Yes | TCEA2 |
Variants not previously identified as favorable adiposity in previous work (15) are considered novel (yes).
We also identified three clusters of BMI-increasing alleles (Fig. 1B). Two clusters (BMI-C1, 39 variants; and BMI-C2, 82 variants) indicated a positive causal effect on type 2 diabetes risk (consistent with unfavorable adiposity), while one cluster (BMI-C3, 8 variants) suggested a negative causal effect (consistent with favorable adiposity) (Supplementary Table 3). Among BMI-C1 and BMI-C2, one and two variants, respectively, were previously associated with unfavorable adiposity (15). One variant in BMI-C3 was previously associated with favorable adiposity (Table 2) (15). Correlated variants (r2 ≥ 0.8) were observed between BFP and BMI clusters, reflecting shared genetic architecture. Importantly, no correlation was noted between unfavorable and favorable adiposity clusters (Supplementary Table 6 and Supplementary Fig. 3).
We validated the causal effect of adiposity through these clusters on type 2 diabetes (in the unfavorable and favorable direction) using FinnGen (19) as an independent cohort. MR-IVW results against type 2 diabetes risk (odds ratios [95% confidence intervals]) were as follows: BFP-all, 2.20 (1.89–2.56); BFP-C1, 11.20 (6.90–18.21); BFP-C2, 4.42 (3.72–5.25); BFP-C3, 1.41 (1.07–1.86); BFP-C4, 0.29 (0.18–0.48); and BFP-C5, 0.05 (0.030–0.080) per 1-SD increase in BFP. For BMI, results were BMI-all, 2.35 (2.19–2.53); BMI-C1, 4.23 (3.53–5.07); BMI-C2, 2.40 (2.13–2.71); and BMI-C3, 0.47 (0.23–0.95) per 1-SD increase in BMI (Supplementary Table 5). The F statistic (a representation of instrument strength for MR-IVW) was >50 for all BFP and BMI clusters (Supplementary Table 7).
The Effect of Clusters on Adiposity-Related Traits
To investigate differences in cluster metabolic signatures, we generated cluster-specific genetic risk scores and compared the effects of these scores on different adiposity-related traits. We included metabolic biomarkers, anthropometric traits, metabolites, and inflammatory cytokines (Figs. 2–4 and Supplementary Table 4).
The genetic risk scores for all BFP and BMI clusters were associated with higher adult BMI and leptin, regardless of their favorable or adverse metabolic effect. BMI clusters showed more significant associations with higher adiposity from early life (birth weight, childhood obesity, and childhood BMI) than BFP clusters. This could be explained by the fact that BMI reflects overall body size, while BFP, focused on the proportion of body weight composed of fat, may be more influenced by factors related to fat distribution and metabolic processes. Comparisons would be more readable if we had a GWAS for childhood BFP. All of the unfavorable adiposity clusters (BFP-C1, C2, and C3 and BMI-C1 and C2) were associated with an adverse metabolic profile (higher triglycerides, C-reactive protein, liver enzymes, insulin resistance, and lower HDL-cholesterol and sex-hormone binding globulin) while favorable adiposity clusters (BFP-C4 and C5 and BMI-C3) were associated with a favorable metabolic profile (Fig. 2).
Genetic risk scores for unfavorable adiposity clusters were associated with insulin resistance-correlated amino acids (37) (with a weaker effect for BFP-C3 but directionally consistent), including phenylalanine, tyrosine, isoleucine, leucine, and valine. There was also an association with higher glycoprotein acetyls levels, suggesting these clusters affect inflammation (38), and lower glutamine and glycine levels, which are metabolites linked to improved glucose regulation (37) (Fig. 3 and Supplementary Table 5).
Favorable adiposity clusters had a significant association with lower n-3 levels, and a higher n-6–to–n-3 ratio, whereas unfavorable adiposity clusters had no association with n-3 or n-6 (Fig. 3). Although observational studies link high n-6–to–n-3 ratios with obesity (39), evidence from randomized controlled trials and MR studies remains inconclusive regarding their causal effects on metabolic outcomes such as type 2 diabetes, glucose metabolism, or cardiovascular disease (40,41). Inconsistencies in trial results may stem from factors such as study duration, cooking methods, ethnicity, sample size, and fatty acid source.
To further investigate the cluster-specific role of inflammation, as inflammation has been suggested as a mechanism that increases type 2 diabetes risk in people with obesity (42), we used data on pro- and anti-inflammatory cytokines. Genetic risk scores for unfavorable adiposity clusters were associated with higher cytokine levels (tumor necrosis factor-related apoptosis-inducing ligand, tumor necrosis factor-β, interleukin [IL] 7, hepatocyte growth factor, chemokine [CC-motif] ligand 2/MCP-1 for BFP-C2, and IL-2, IL-5, IL-7, and hepatocyte growth factor for BMI-C1). The favorable adiposity cluster BFP-C5 was associated with lower inflammatory cytokine levels (e.g., IL-12) (Fig. 4 and Supplementary Table 5).
The Effect of Clusters on MRI-Derived Measures of Fat Distribution and Body Composition
We used precision MRI-derived measures of fat and body composition to investigate differences in fat distribution patterns of our adiposity clusters. Genetic risk scores for all clusters were associated with higher abdominal and thigh SAT. Unfavorable adiposity clusters were also associated with increased ectopic fat accumulation in pancreas, liver, and paraspinal muscle, VAT, internal fat, and thigh intermuscular adipose tissue. They were also associated with higher muscle index and organ volume (kidney, liver, and spleen), with some cluster specific effects (Fig. 5 and Supplementary Table 5).
Favorable adiposity clusters had unique and distinct patterns of association with MRI-derived measures. BFP-C4 was associated with higher paraspinal muscle PDFF and higher thigh intermuscular adipose tissue, but no association with liver PDFF, pancreas PDFF, VAT, muscle index measures, or organ volume. BFP-C5 was associated with lower liver PDFF, lower VAT-to-abdominal SAT ratio, lower muscle index measures, and lower kidney and spleen volume. BMI-C3 was associated with higher muscle index and higher kidney and liver volume. These results were consistent in data from men and women.
The Causal Effect of Adiposity Clusters on Risk of Type 2 Diabetes-Related Disease Outcomes
Since the genetically predicted favorable and unfavorable adiposity clusters had distinct effects on different clinical and MRI biomarkers, we used two-sample MR to investigate differences in causal effect of each cluster on disease risk, including those related to type 2 diabetes (Fig. 6 and Supplementary Table 5). We detected evidence of heterogeneity from MR estimates when we studied the effect of higher adiposity using all BFP and BMI variants (BFP-all and BMI-all) (Supplementary Table 5). However, there was no evidence of heterogeneity in the causal estimates when using each cluster. Unfavorable adiposity clusters BFP-C1, BFP-C2, BMI-C1, and BMI-C2 were associated with higher disease risk, including diabetic nephropathy, retinopathy, and neuropathy, hypertension, polycystic ovary syndrome, nonalcoholic fatty liver disease, ischemic heart disease, stroke, peripheral artery disease, atherosclerosis, heart failure, atrial fibrillation, chronic kidney disease, thrombotic events, aortic aneurysm, gout, osteoarthritis, gallstones, and asthma. BFP-C3 was only associated with higher risk of peripheral artery disease, atherosclerosis, and aortic aneurysm. We also observed some cluster-specific effects among unfavorable adiposity clusters; for example, BFP-C2 and BMI-C2 were associated with higher psoriasis risk.
Among favorable adiposity clusters, BFP-C5 was associated with lower disease risk, including diabetic nephropathy, retinopathy, and neuropathy, hypertension, nonalcoholic fatty liver disease, ischemic heart disease, stroke, peripheral artery disease, and atherosclerosis, but it was associated with higher risk of thrombotic events and osteoarthritis. BFP-C4 was associated with lower diabetic retinopathy risk and higher osteoarthritis risk, and BMI-C3 was associated with higher risk of osteoarthritis and gallstones. All results were directionally consistent with those from sensitivity tests (Supplementary Table 5).
eQTL and Pathway Enrichment Analysis
To explore differences in tissue-specific gene expression for unfavorable and favorable adiposity variant clusters, we counted the number of independent eQTLs in brain and adipose (subcutaneous and visceral) tissue per cluster. When the ratio of independent eQTLs in adipose to brain tissue was compared, unfavorable adiposity clusters BFP-C2 and BMI-C2 were more enriched for eQTLs in the brain, and favorable adiposity clusters were more enriched in adipose tissue (Supplementary Table 8).
All clusters were enriched for different pathways (Supplementary Table 9). Notable pathways for unfavorable adiposity clusters comprised cytoskeletal regulation by Ras homolog guanosine triphosphatase (BFP-C1), Janus kinase/STAT signaling pathway (BFP-C2 and BMI-C2), endothelin signaling pathway (BFP-C3), and ubiquitin proteasome pathway (BMI-C1). For favorable adiposity clusters, the Alzheimer disease-amyloid secretase pathway was highlighted (BFP-C4). Of these, only BFP-C3 and BMI-C2 remained significant after correction for multiple testing (false discovery rate <0.05).
Discussion
We performed clustered MR analyses to identify distinct causal mechanisms linking higher adiposity with type 2 diabetes risk. We identified evidence for multiple causal mechanisms by which adiposity influences type 2 diabetes risk. While most biological mechanisms associated with higher adiposity lead to increased type 2 diabetes risk (e.g., inflammation), there may also be some pathways associated with higher adiposity that lead to lower type 2 diabetes risk. These potentially protective mechanisms relate to lower liver fat and improved insulin sensitivity or to increased body size and enhanced muscle quality.
Association Patterns Common to All Adiposity Clusters
Shared associations across adiposity clusters, irrespective of their favorable or unfavorable metabolic effect, suggest consequences of higher adiposity beyond metabolic impact. For example, association with higher leptin for all clusters was expected, because leptin is produced by adipose tissue. The associations with higher osteoarthritis risk are consistent with previous findings stating that the metabolic effect of adiposity might not be the primary driver of this condition. The higher thrombotic event risk is also in agreement with previous results confirming the causal role of nonmetabolic components of higher adiposity (e.g., the mechanical effect of higher weight on blood flow in lower limbs) (43).
The Difference Between Unfavorable and Favorable Adiposity Clusters
Overall, the unfavorable adiposity clusters were associated with an adverse metabolic profile encompassing higher insulin resistance and inflammatory markers, adverse liver profile, and increased ectopic fat deposition (liver, pancreas, paraspinal, and thigh muscle). The favorable adiposity clusters were overall associated with a healthy metabolic profile, with an association pattern opposite to the unfavorable adiposity clusters.
The association between unfavorable adiposity clusters and higher organ volume, especially the liver, could be due to increased ectopic fat. No cluster showed an association with pancreas volume, suggesting limited power or a lack of involvement in adiposity-to-diabetes pathways, although pancreatic volume tends to decline in diabetes, suggesting volume changes in this organ are more difficult to contextualize. The overall associations with fat distribution were consistent with previous work, where unfavorable adiposity was associated with higher liver, pancreatic and visceral fat, and favorable adiposity was associated with lower liver fat and had no significant effect on pancreatic fat (15,43).
Recent findings show that intentional weight loss in type 2 diabetes reverses many associated amino acid changes (44). Therefore, the opposite effect of favorable and unfavorable adiposity clusters on amino acid levels previously associated with lower insulin sensitivity and higher insulin resistance and type 2 diabetes risk (37) could suggest these amino acids are not causal risk factors but are biomarkers of metabolically healthy or unhealthy adiposity.
Differences Between Unfavorable Adiposity Clusters
Differences among unfavorable adiposity clusters in associations with biomarkers suggest diverse mechanisms by which higher adiposity leads to adverse metabolic outcomes. BFP-C1 demonstrated a more unfavorable metabolic effect, with the strongest impact on type 2 diabetes risk, circulatory lipids, and surrogates of insulin resistance, with no effect on inflammatory cytokines. Cytoskeletal regulation by Ras homolog guanosine triphosphatase was highlighted for BFP-C1, for which there is emerging evidence to implicate a role in metabolic homeostasis by regulating glucose uptake into skeletal muscle and adipose tissue (45). This cluster also had more significant associations with measures of fat distribution and body composition in data from women.
BFP-C2 and BMI-C2 were associated with cytokines and inflammatory markers and were both enriched for pathways related to inflammation, suggesting that inflammation is strongly associated with the mechanisms these clusters may represent. Higher adiposity through these clusters was associated with higher risk of psoriasis, possibly through higher inflammation as an underlying mechanism. BFP-C3 was only associated with vascular outcomes, including peripheral artery disease, atherosclerosis, and aortic aneurysm, aligning with the highlighted endothelin signaling pathway for this cluster.
Differences Between Favorable Adiposity Clusters
Similarly, the differences between favorable adiposity clusters associations with metabolic and imaging biomarkers suggest that there is more than one mechanism of adiposity leading to favorable metabolic outcomes. BFP-C5 was more protective against disease risk compared with BFP-C4 and BMI-C3. BFP-C5 was associated with higher insulin sensitivity and lower inflammatory marker levels, while BFP-C4 and BMI-C3 were not associated with these measures.
The favorable adiposity clusters also had unique association patterns with measures of fat distribution and body composition. BFP-C5 was associated with lower liver PDFF, while BFP-C4 and BMI-C3 had no association with liver fat. BFP-C4 was associated with higher subcutaneous fat and paraspinal muscle PDFF but had no association with any other ectopic fat depot.
BMI-C3 could represent an adiposity subtype associated with increased body size regardless of fat, as it was associated with higher measures of early-life obesity, muscle index, kidney volume, and liver volume, but had no association with any ectopic fat measures. The favorable effect of BMI-C3 could be through increasing skeletal muscle mass, which has been associated with decreased type 2 diabetes risk potentially via increased insulin sensitivity, improved glucose metabolism, or acting as a sink for glucose (46,47).
None of the favorable adiposity clusters were associated with pancreatic fat, although this is harder to measure accurately. This is consistent with result of the “twin-cycle” hypothesis finding that liver fat is more likely to mediate glycemic control in type 2 diabetes than pancreatic fat (15,48).
Strengths and Limitations
We leveraged a range of publicly available GWAS data sets to investigate the complexity between adiposity and type 2 diabetes risk. This research can be expanded as sample sizes and data accessibility improve. We also used gold standard measurements of MRI scans of sex-specific fat and organ content within the UK Biobank to strengthen our analysis and consider sexual dimorphism in body fat distribution.
The GWAS data sets we chose were focused on European populations due to large sample size, potentially limiting the generalizability of our findings to people of other ethnicities and fat distributions (9). Nevertheless, we have shown that previously identified favorable and unfavorable adiposity clusters have a consistent effect across different ethnic groups (49).
Second, the biological interpretation of our adiposity cluster variants will require further exploration, because most GWAS variants reside within noncoding regions and often exert their effects alongside correlated variants (50).
Third, using genetic associations as a starting point may downplay the influence of environmental factors. This approach necessitates accurate effect estimates, well-established genetic foundations for traits, and large sample sizes; hence, why we selected the most current and expansive GWAS studies available.
Fourth, in our clustering algorithm, we prioritized the minimization of false-positive findings. While this cautious approach bolsters reliability of our findings, it may leave certain associations unexplored if we overlooked variants that might belong to adiposity clusters.
Finally, one key consideration is the strength and distinctiveness of the identified clusters. The interpretation of “distinct” clusters is contingent upon effect size ratios, and we recognize the need for a nuanced evaluation of their robustness. We acknowledge that the observed differences in associations with various traits among clusters may, in some instances, represent differences in magnitude rather than distinct mechanistic pathways.
Conclusion
Using genetically predicted measures of adiposity and diverse traits, we found evidence for different underlying pathways and subtypes of higher adiposity with contrasting risks for type 2 diabetes and related complications. These novel insights hold potential for advancing precision medicine strategies for type 2 diabetes and related conditions through targeted adiposity management.
This article contains supplementary material online at https://doi.org/10.2337/figshare.25460608.
Article Information
Acknowledgments. This research was conducted using the UK Biobank Resource under Application Number 44584. Data on glycemic traits were contributed by Meta-Analyses of Glucose and Insulin-related traits Consortium (MAGIC) investigators and were downloaded from www.magicinvestigators.org. Data on birth weight, childhood obesity, and childhood BMI were contributed by the Early Growth Genetics (EGG) Consortium using the UK Biobank Resource and were downloaded from www.egg-consortium.org. The authors also want to acknowledge the participants and investigators of the FinnGen study.
Funding. H.Y. is funded by the Diabetes UK RD Lawrence Fellowship (grant 17/0005594). N.S. is supported by the British Heart Foundation Research Excellence Award (RE/18/6/34217). S.B. is supported by the Wellcome Trust (225790/Z/22/Z) and the United Kingdom Research and Innovation Medical Research Council (MC_UU_00002/7, MC_UU_00040/01).
Duality of Interest. N.S. has received grants and personal fees from AstraZeneca, Boehringer Ingelheim, Novartis, and Roche Diagnostics, and personal fees from Abbott Laboratories, AbbVie, Amgen, Eli Lilly, Hanmi Pharmaceuticals, Janssen, Menarini-Ricerche, Novo Nordisk, Pfizer, and Sanofi outside the submitted work. M.C. and E.P.S. are employees of Calico Life Sciences LLC. No other potential conflicts of interest relevant to this article were reported.
Author Contributions. A.A., M.C., M.T., N.B., M.A.H., E.P.S., B.W., S.B., J.D.B., N.S., E.L.T., and H.Y. contributed to discussion and reviewed and edited the manuscript. A.A., M.C., M.T., N.B., M.A.H., E.P.S., and B.W. ran investigation for the study. A.A. and H.Y. wrote the original draft. H.Y. conceptualized the study, acquired funding for the study, and supervised the study. 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.