What this is
- This research investigates the genetic connections between obesity and lipid levels, focusing on their impact on cardiovascular disease (CVD).
- It identifies genomic regions with contrasting genetic effects on obesity and lipid traits, termed protective and adverse BMI-lipid loci.
- The study utilizes data from the UK Biobank and diverse populations to explore these genetic associations and their implications for CVD risk.
Essence
- The study identifies distinct genomic loci linking body mass index (BMI) with lipid levels, revealing protective and adverse effects on cardiovascular disease risk. Protective loci are associated with favorable lipid profiles and reduced CVD risk.
Key takeaways
- 789 high-density lipoprotein (HDL) loci, 26 low-density lipoprotein (LDL) loci, and 494 triglyceride (TG) loci show significant genetic correlation with BMI. This indicates a complex genetic relationship between obesity and lipid levels.
- The protective polygenic risk score (PRS) is associated with higher BMI but favorable lipid profiles and reduced CVD risk in the Population Architecture using Genomics and Epidemiology (PAGE) study.
- Mendelian randomization supports the causal impact of protective loci on cardiometabolic outcomes, suggesting that certain genetic profiles can mitigate the adverse effects of obesity on lipid metabolism.
Caveats
- The study's findings may not be generalizable to non-European ancestry populations, as the genomic partitioning was based on a European-ancestry structure.
- BMI is a crude proxy for adiposity, which may limit the identification of important pleiotropic loci linking obesity to lipid traits.
Definitions
- pleiotropy: The phenomenon where one gene influences multiple traits or outcomes, such as obesity affecting lipid levels and cardiovascular disease risk.
Simplified
Background
Obesity imposes an enormous global public health burden [1, 2] and increases the burden of cardiovascular diseases (CVD) and many other downstream sequelae [3] through its impact on CVD risk factors (e.g., dyslipidemia, hypertension, and type 2 diabetes) [4–7]. However, major research gaps remain. Few studies have evaluated the often described but poorly understood heterogeneity in CVD outcomes observed among populations with obesity [8]. This heterogeneity may be due to heterogeneous relationships between obesity and CVD risk factors, especially lipid levels [9, 10]. One plausible but largely unexplored source of heterogeneity in the obesity-lipid level relationship is the shared genetic architecture across obesity and lipid traits. Better characterization of the shared genetic underpinnings may help us better understand the heterogeneous impact of obesity on lipid traits and CVD.
Recent studies have suggested that pleiotropic obesity loci, especially those with counterintuitively protective associations with CVD traits, could help explain the observed heterogeneous impact of obesity on CVD [11–17]. For example, two recent studies identified 36 [15] and 62 [11] variants that were linked to both increased adiposity and favorable metabolic profiles, respectively. Several variant-level approaches have been implemented to identify pleiotropic obesity variants [11]; however, no previous studies have used locus-level approaches and local genetic correlation analysis, an emerging genomic analysis tool to explore pleiotropy. In addition, previously identified pleiotropic loci have not been validated in populations with diverse ancestries. As is common with other genetic research, these loci were discovered in European ancestry populations, and it is unknown whether the identified bivariate loci show comparable influences on obesity and cardiometabolic traits in other ancestries.
We therefore aimed to (1) identify genomic regions with significant shared genetic signals between body mass index (BMI) and lipid traits (BMI-lipid bivariate loci) in opposing directions and investigate the potential causal genes underlying counterintuitive pleiotropy between BMI and lipid levels, and (2) examine the potential influence of BMI-lipid bivariate loci on BMI, lipid levels, and downstream CVD and other disease outcomes in diverse populations.
Methods
Study cohorts and data sources
We utilized GWAS summary statistics from the UK Biobank (UKBB) for the identification of BMI-lipid bivariate loci. In addition, individual-level data from the Population Architecture using Genetics and Epidemiology (PAGE) study, All of Us (AoU) Research Program, and UKBB were used for testing the associations of the BMI-lipid bivariate loci with relevant phenotypes. Specifically, PAGE data were used to evaluate associations with cardiometabolic traits, AoU data were used to perform a phenome-wide association (PheWAS) analysis, and UKBB data were used to characterize the loci in relation to MRI-derived measures of fat distribution.
UKBB
The UKBB is a large-scale prospective study of More than 500,000 individuals Living in the UK. Participants aged 40–69 were recruited from 2006 to 2010, and their phenotypic and genotypic information, including questionnaires, physical and blood measures, genome-wide genotyping data, imaging data, and health outcomes, has been collected [18].
In the current study, we utilized the publicly available Pan-UKBB GWAS summary statistics [19, 20] for BMI and lipid traits in individuals of European ancestry as the discovery sample for identifying BMI-lipid bivariate loci (N = 419,163 for BMI, N = 367,021 for HDL, N = 400,223 for LDL, and N = 400,639 for TG). Additional descriptions on the GWAS analysis conducted by the Pan-UKBB team can be found in Additional file 1. We also used MRI-derived regional body fat measures from the UKBB imaging substudy extension [21] to characterize the identified BMI-lipid bivariate loci (maximum available N = 22,988).
PAGE
The PAGE consortium was launched in 2008 through the NHGRI’s effort to expand the ancestral diversity in genomic studies [22, 23]. PAGE cohort studies include the Atherosclerosis Risk in Communities (ARIC) study, Coronary Artery Risk Development in Young Adults study (CARDIA), Hispanic Community Health Study / Study of Latinos (HCHS/SOL), Women’s Health Initiative (WHI), Multiethnic Cohort Study (MEC), and Icahn School of Medicine at Mount Sinai BioMe biobank. In our study, a total of 83,376 participants with relevant genetic and phenotypic data from these studies were included. Self-identified White (N = 25,418), Black (N = 25,255), Hispanic/Latino (N = 25,814), and East Asian (N = 6889) participants were included in the current analysis (Table S1) as the validation population for identified BMI-lipid bivariate loci. These cohort studies are summarized in Additional file 1.
AoU
We used the AoU Research Program as the target population for a PheWAS analysis. AoU is a diverse population-based research study that utilizes electronic health records (EHR) from participants across the US. Of the 245,394 participants in the “All of Us Controlled Tier Dataset v7” with whole genome sequencing (WGS) data at the time of this study, 147,327 had demographic data. We excluded those with insufficient EHR data (an EHR length of less than 3 years, with fewer than three independent visit dates), leaving 99,409 participants available for analysis. AoU performed ancestry prediction using gnomAD v3.1 [24] to train a random forest classifier using 1000 Genomes / Human Genome Diversity Project (HGDP) data [25, 26] as a reference. Based on this classification, 61,393 participants were of European ancestry, 20,191 of African ancestry, 15,479 of American ancestry, 1182 of East Asian ancestry, 841 of South Asian ancestry, and 323 of Middle Eastern ancestry. The mean age of the study population was 56.0 (SD: 16.8) years (Table S2).
Measurement
Genetic data
UKBB
A total of 488,377 participants from the UKBB were genotyped on the Applied Biosystems UKBB Lung Exome Variant Evaluation (UK BiLEVE) Axiom Array (N = 49,950) or the UKB Axiom Array (N = 438,427) [27]. Imputation was performed using IMPUTE4 with the Haplotype Reference Consortium, UK10K, and 1000 Genome Phase 3. Detailed methods have been described previously [27].
PAGE
Participants were genotyped on the Multi-Ethnic Genotyping Array (MEGA) at the Center for Inherited Disease Research as part of the PAGE study [23, 28]. Additionally, some participants from ARIC, BioMe, CARDIA, MEC, and WHI were genotyped separately on Illumina or Affymetrix arrays by each study or ancillary study. The number of samples included in our analyses by study, self-reported race/ethnicity, and genotyping platform is shown in Table S3. A total of 34,373 samples that were included in the current analysis were Genotyped on the MEGA array, and the remaining 49,003 samples were genotyped on alternative arrays. Table S4 summarizes the genotyping platform, QC criteria, imputation methods, and reference panel that each study and ancillary study implemented.
AoU
WGS for all consented participants was performed with NovaSeq 6000. The DRAGEN (Illumina) pipeline (v3.4.12) [29] was utilized to generate the metrics for the data processing steps and to perform genome Mapping, alignment, and variant calling. Over 702 million variants were reported, with an average coverage of ≥ 30× and over 90% of bases at 20× coverage. A consistent sample and data processing protocol was implemented to minimize possible batch effects across centers. Further details on WGS can be found in the AoU Research Program Genomic Research Data Quality Report [30] and other sources [31].
Phenotype data
UKBB
BMI (weight (kg)/[height (m)]2) was calculated from standing height and weight measured at the baseline visit [18]. Participants’ blood samples were also collected at the same visit, and various biomarkers, including three lipid traits (high-density lipoprotein cholesterol (HDL), low-density lipoprotein cholesterol (LDL), and triglycerides (TG)), were measured as previously described [32].
We utilized magnetic resonance imaging (MRI)-derived body fat distribution data to investigate the potential implications of the identified heterogeneous BMI-lipid bivariate loci. As part of the UKBB imaging substudy extension [21], regional fat depositions were quantified using a Siemens MaGNETOM Aera 1.5-T MRI scanner (Siemens Healthineers, Erlangen, Germany) with the dual‐echo Dixon Vibe protocol employed [33, 34]. The sample size and the distribution of variables included in the current analysis are presented in Table S5. The fat distribution variables used in the current study included abdominal visceral adipose tissue volume (VAT), abdominal subcutaneous adipose tissue volume (ASAT), total adipose tissue volume (TAT; defined between the bottom of the thigh muscles and the top of the vertebrae T9), muscle fat infiltration (MFI), and liver fat (assessed using the 10-point symmetric chemical-shift encoded acquisition (10P) method, quantified as the average liver proton density fat fraction (PDFF) across three to nine regions of interest). Additionally, gluteofemoral adipose tissue (GFAT) was derived by subtracting VAT and ASAT from the TAT, as suggested by a previous study [35]. We also computed adipose tissue ratios, including VAT/ASAT, VAT/GFAT, ASAT/GFAT, VAT/TAT, ASAT/TAT, and GFAT/TAT.
PAGE
BMI was used as a continuous proxy measure of overall adiposity, and obesity status was defined as BMI ≥ 30 kg/m2 for non-East Asians and BMI ≥ 25 kg/m2 for East Asians [36, 37]. Blood samples were drawn following an 8-h fast for lipid, glucose, and insulin measures. Three lipid measures (HDL, LDL, and TG) were used as continuous proxy measures of dyslipidemia. HDL and TG levels were directly quantified. Since directly measured LDL was not consistently available across all PAGE participants, LDL levels were instead computed using the Friedewald Eq. [38], excluding individuals whose TG levels were > 400 mg/dL, as in previous literature [39]. We determined participants’ diabetes status if they met any of the following American Diabetes Association criteria [40]: diabetes medication, self-reported diagnosis, fasting glucose ≥ 7 mmol/L or HbA1c ≥ 48 mmol/mol, or random glucose > 11.11 mmol/L, and aged ≥ 25 years at the time of diagnosis (to avoid potential misclassification between T1D and T2D). Blood pressure was measured with a standardized protocol. We classified participants as hypertensive if they met any of the following criteria: (1) systolic blood pressure (SBP) ≥ 140 mmHg, (2) diastolic blood pressure (DBP) ≥ 90 mmHg, (3) self-reported use of any antihypertensive medication, or (4) ICD-9 codes 401.x or ICD-10 codes I10.x—I15.x [23]. Additionally, a subset of PAGE cohorts collected CVD data, including prevalence, incidence, or related deaths. Detailed descriptions of phenotypic measures for each cohort in PAGE are provided in Additional file 1.
Statistical analyses
Global SNP-based heritability and genetic correlations
Prior to performing local genetic correlation analyses, we estimated the global single nucleotide polymorphism (SNP)-based heritability of BMI and the three lipid traits, as well as the genetic correlation between three BMI-lipid pairs (BMI-HDL, BMI-LDL, and BMI-TG) in the UKBB by performing linkage disequilibrium (LD) score regression [41] based on GWAS summary statistics from UKBB.
Bivariate loci identification
We classified significant local genetic correlation estimates [p < (0.05 / the number of significant univariate loci for both traits)] into two different BMI-lipid bivariate locus groups, adverse or protective, based on their directions of effect with BMI and lipid levels (Table S6). A bivariate locus was considered adverse if it showed a positive local genetic correlation coefficient (rg) between BMI and LDL or TG, or a negative local rg between BMI and HDL (BMI( +)LDL( +) locus, BMI( +)TG( +) locus, or BMI( +)HDL(−) locus). Conversely, a bivariate locus was considered protective if it showed a negative local rg between BMI and LDL or TG, or a positive local rg between BMI and HDL (BMI( +)LDL( −) locus, BMI( +)TG( −) locus, or BMI( +)HDL( +) locus). We considered the protective BMI-lipid bivariate loci counterintuitive, as the phenotypic correlations were in opposite directions (i.e., a phenotypic correlation coefficient (r) < 0 for BMI-HDL, r > 0 for BMI-LDL and BMI-TG). Each BMI-lipid pair was tested separately, so there could be overlapping loci for multiple BMI-lipid pairs, even with different directions (e.g., a locus can be adverse for one BMI-lipid pair and protective for another BMI-lipid pair).
Overview of analytical framework.We first identified pleiotropic genomic loci showing opposite directions of association between BMI and lipid traits using local genetic correlation analyses (implemented through LAVA).To prioritize candidate causal genes within the BMI-lipid bivariate loci, we performed transcriptome-wide association studies (TWAS) using FUSION (TWAS-FUSION) and summary-based Mendelian randomization (SMR). Significant genes were filtered based on consistent associations with both BMI and the corresponding lipid trait in the hypothesized direction.Clinical implications of the protective and adverse BMI-lipid bivariate loci were evaluated through four analyses: (1) association testing of bivariate loci-stratified polygenic risk scores (PRS) with cardiometabolic traits in the PAGE study; (2) Phenome-wide association study (PheWAS) in All of Us using PRSbased on protective loci only; (3) Mendelian randomization to infer causal relationships between protective or adverse adiposity and health outcomes; and (4) association of bivariate loci-stratified PRSwith MRI-measured fat distribution traits in UK Biobank A B C BMI BMI BMI
Gene prioritization for the BMI-lipid bivariate loci
To investigate biological implications and prioritize the genes related to the protective BMI-lipid bivariate loci, we conducted TWAS-FUSION, following the recommended protocol with default settings (http://gusevlab.org/projects/fusion/↗) [44] and prioritized potential genes whose genetically predicted expression levels were associated with both BMI and a corresponding lipid trait (Fig. 1B). We integrated each GWAS summary result (BMI, HDL, LDL, and TG) with reference gene expression levels and identified genes whose genetically predicted expression levels were significantly associated with BMI or lipid traits (multiple testing corrected by the number of genes tested). Then, we filtered to the protein-coding genes where the whole gene was located within the bivariate loci and identified the overlapping genes from the BMI GWAS-based results and corresponding lipid trait GWAS-based results. We also examined directional consistency by comparing TWAS-FUSION Z scores for BMI and the corresponding lipid trait. For example, we verified if a gene within BMI( +)HDL( +) loci had the same direction of TWAS-FUSION Z-score for both BMI and HDL in the same tissue. Based on the known roles of the overlapping genes (as reported in public databases (e.g., GeneCards [45])), we inferred potential pathways simultaneously influencing BMI and lipid traits.
We additionally implemented the summary-based Mendelian randomization (SMR) method (https://yanglab.westlake.edu.cn/software/smr/↗) [46], a transcriptome-wide gene-based MR approach that integrates GWAS and eQTL data to prioritize genes whose expression levels may be linked to a trait of interest. The current analysis used eQTL data for More than 15,000 genes from the GTEx database [47]. SMR utilizes the lead cis-eQTL SNP—pre-identified by the developers—for each gene as an instrumental variable (IV) for MR to test for causal effects on the trait of interest. The input GWAS summary statistics were from UKBB European ancestry populations, which were the same as those used in the local genetic correlation analysis. We first identified genes significantly associated with both BMI and the corresponding lipid trait and then selected protein-coding genes located within the protective BMI-lipid loci. Among these, we retained genes whose associations were directionally aligned with our hypothesized effects (i.e., increased BMI and decreased LDL or HDL or increased HDL), consistent with our TWAS-FUSION identification strategy.
Both TWAS-FUSION and SMR analyses were conducted across six relevant tissues from GTEx v8 [48]: five tissues prioritized in a previous study [49]—whole blood (WB), VAT, SAT, liver, and artery—along with skeletal muscle tissue, which was additionally included due to its relevance in body mass and fat metabolism.
Associations of BMI-lipid-stratified PRSwith cardiometabolic traits in PAGE BMI
We examined the potential influence of the BMI-lipid bivariate loci in ancestrally diverse PAGE participants (Fig. 1C). We hypothesized that the protective and adverse bivariate loci would be involved in distinct biological pathways, linking adiposity with protective and detrimental roles in lipid metabolism, respectively, and that polygenic risk scores for BMI (PRSBMI) constructed with variants restricted to the identified bivariate loci would capture the genetic predisposition to distinct subtypes of adiposity. Based on these assumptions, we constructed three protective BMI-lipid bivariate loci-stratified PRSBMI—i.e., PRSBMI(+)HDL(+) constructed with the variants within BMI( +)HDL( +) loci, PRSBMI(+)LDL(−) constructed with the variants within BMI( +)LDL( −) loci, and PRSBMI(+)TG(−) constructed with the variants within BMI( +)TG( −) loci—and three adverse BMI-lipid bivariate loci-stratified PRSBMI (PRSBMI(+)HDL(−), PRSBMI(+)LDL(+), and PRSBMI(+)TG(+)) by restricting to the variants within corresponding bivariate loci.
We utilized publicly available PRS weights for BMI based on UKBB summary statistics for BMI, estimated using PRS-CS [50] methods with the auto parameter option (validation set is not needed) [51] (PGS Catalog ID: PGS002844). The Pearson correlation between PRS and BMI in the testing sample was 0.321 [51]. The original weights included 1,113,832 SNPs, but for our analysis, we restricted to the variants within the identified BMI-lipid bivariate loci and applied the corresponding weights to our target population, the PAGE study. Variants that were located outside of the bivariate loci were not included in the PRS calculation. The PRS was calculated using the “–score” function in PLINK (version 2.00a3LM) [52, 53]. The number of SNPs included in each BMI-lipid loci-stratified PRSBMI is shown in Table S7.
We then assessed the associations of the bivariate loci-stratified PRSBMI with BMI and obesity status, lipid traits (HDL, LDL, TG, total cholesterol, and dyslipidemia), CVD risk factors (fasting glucose, fasting insulin, homeostatic model assessment for insulin resistance (HOMA-IR), hemoglobin A1c (HbA1c), type 2 diabetes (T2D) status, SBP, DBP, and hypertension), and CVD outcomes (myocardial infarction (MI) and stroke). We hypothesized that higher protective loci-based PRSBMI would be associated with increased BMI or obesity but, counterintuitively, with protective cardiometabolic profiles. Conversely, higher adverse loci-based PRSBMI would be associated with increased BMI and an adverse cardiometabolic profile. We applied linear regression models for continuous outcomes and logistic regression models for binary outcomes. Included covariates were age, sex, ten genetic principal components (PCs) of ancestry, study, genotype panel, and self-reported race/ethnicity as a social construct associated with the social determinants of health, racism, discrimination, and environmental factors [54, 55]. The PRS analyses were conducted in R (version 4.1.0).
Phenome-wide association study of BMI-lipid bivariate loci in AoU
We further explored the potential influence of protective loci-based PRSBMI on various groups of disease outcomes in the AoU study (Fig. 1C). We adopted the PheWAS pipeline used in the previous report [56] and hosted on AoU as a demonstration workspace (“Demo—PheWAS Smoking”). For mapping ICD-9 and ICD-10 codes to phecodes, we utilized Phecode map v1.2. In our study, we utilized different covariates compared to Ramirez et al. [56]: age at last code, sex at birth/gender, 10 ancestry PCs, record depth, and visit frequency. Record depth was estimated as the sum of all visits where an observation or condition code appeared in the EHR. Visit frequency was subsequently calculated by dividing the record depth by the length of the EHR. In our primary analysis, we defined a case as having two instances of a phecode and included all ancestry groups. Under this Model, individuals with only one instance of a phecode were excluded as controls. We also excluded phecodes with less than 100 cases. We conducted sensitivity analyses in which disease status was determined by having one or two phecodes. In another sensitivity analysis, we stratified individuals by populations—European (as defined using HGDP/1000G clusters as described above) and African—to mitigate any potential confounding effects of population structure.
Causal effects of BMI-lipid bivariate loci on cardiometabolic outcomes
We conducted two-sample Mendelian randomization (MR) analyses to investigate the causal effects of hypothesized protective and adverse adiposity on 17 health outcome traits. Genetic IVs for protective adiposity were defined as independent lead SNPs (LD r2 < 0.01) associated with BMI at p < 5 × 10−6 and located within protective BMI-lipid loci (i.e., BMI( +)HDL( +), BMI( +)LDL(–), and BMI( +)TG(–) loci). Similarly, IVs for adverse adiposity were selected from adverse BMI-lipid loci. To reduce weak instrument bias, SNPs with F-statistics < 10 were excluded. The 17 outcome traits were selected based on traits analyzed in the PRS analysis of the PAGE study or closely related phenotypes, supplemented with two additional traits identified through PheWAS, atrial fibrillation, and dementia. These outcomes included the following: obesity (as a confirmatory trait), HDL, LDL, TG, hypercholesterolemia, SBP, DBP, hypertension, fasting glucose, fasting insulin, HbA1c, HOMA-IR, T2D, MI, stroke, atrial fibrillation, and dementia. We utilized publicly available GWAS summary statistics: FinnGen GWAS [57, 58] for atrial fibrillation, dementia, hypercholesterolemia, hypertension, MI, obesity, stroke, and T2D; Global Lipids Genetics Consortium (GLGC) GWAS [59, 60] for HDL, LDL, and TG; Meta-Analyses of Glucose and insulin-related traits Consortium (MAGIC) GWAS [61–64] for fasting glucose, fasting insulin, HOMA-IR, and HbA1c; and Million Veteran Program (MVP) GWAS [65, 66] for DBP and SBP. Detailed data sources for each outcome are provided in Table S8. For the exposure, we used UKBB GWAS summary statistics for BMI, consistent with those used in the local genetic correlation analyses. MR analyses were performed using the TwoSampleMR R package (version 0.6.9) [67]. The primary causal estimates were derived using the inverse-variance weighted (IVW) method. Statistical significance is reported using both a Bonferroni-corrected threshold (p < 0.05/17) and a nominal threshold (p < 0.05). Results meeting either threshold are reported.
To assess robustness and potential violations of MR assumptions, we conducted sensitivity analyses using the weighted median, weighted mode, and MR-Egger methods. The MR-Egger intercept was used to evaluate directional pleiotropy, and MR-PRESSO was applied to detect and correct for outlier variants that could potentially bias the estimates. We further performed multivariable MR (MVMR) analyses to evaluate whether the observed effects of protective or adverse adiposity on outcomes were mediated through corresponding lipid traits. MVMR was implemented using the MVMR R package (version 0.4) [68]. We used the same set of BMI-associated IVs as in the univariable MR analyses and included SNP–lipid trait associations to estimate the direct effect of BMI while accounting for mediation via lipid metabolism. Effect estimates for the lipid traits in the MVMR analysis were obtained from UKBB GWAS summary statistics, which were also used in the local genetic correlation analysis.
Implications of BMI-lipid bivariate loci for fat distribution and body composition
We examined the associations between the BMI-lipid loci-stratified PRS and MRI-derived fat distribution traits to gain insight into the regional fat deposition patterns underlying the identified protective and adverse adiposity loci. Specifically, we tested whether PRSs constructed from BMI-lipid bivariate loci, representing genetically inferred protective and adverse adiposity, were associated with fat distribution traits, thereby providing biological context for these loci. We analyzed eight MRI-based fat distribution traits and their corresponding ratio measures, including BMI as a confirmatory outcome. PRSs were constructed following the same approach used in the PAGE study. Briefly, to generate BMI-lipid loci-stratified PRSs, we selected SNPs located within BMI-lipid (i.e., BMI-HDL, BMI-LDL, and BMI-TG loci) from the genome-wide PRSBMI weight List comprised of HapMap phase 3 SNPs. The Maximum number of participants included in the current analysis was 22,988. The number of SNPs included in each PRS is reported in Table S7. We implemented linear regression Models, adjusting for age, sex, ancestry PC 1 to 10, assessment center, and race/ethnicity. All fat-distribution traits were rank-based inverse normal transformed prior to analysis. PRS were standardized to have a mean of 0 and a standard deviation of 1.
Results
BMI-lipid bivariate loci identification in UKBB
We discovered novel signals for protective BMI-lipid pleiotropy and confirmed previous reports for other protective BMI-lipid pleiotropic signals. We compared the protective bivariate loci results with findings from five previous studies of counterintuitively protective BMI-CVD risk factor pleiotropy [11, 13, 15, 69, 70]. All five studies used variant-based approaches (e.g., multivariate adiposity and cardiovascular traits GWAS). A total of 149 distinct variants located within 104 loci (out of the 2495 genomic regions used for our local genetic correlation analyses) have been reported as obesity variants associated with protective cardiometabolic profiles (Table S12). Although our analyses were locus-based, and it is difficult to directly compare loci and variants, we identified 11 novel protective loci (7 for BMI( +)LDL( −) loci and 5 for BMI( +)TG( −) loci; 1 overlapping locus), which included no previously reported variants. All three BMI( +)HDL( +) loci, three of ten BMI( +)LDL( −) loci, and three of eight BMI( +)TG( −) loci included at least one of the previously identified protective/favorable adiposity variants. Differences observed across studies may be a result of different discovery populations (though some of the prior studies also utilized UKBB) and/or different identification strategies and methods.
There were overlapping protective bivariate loci across multiple BMI-lipid pairs. We identified four protective bivariate loci (Loc1351 in chromosome 8, Loc2351 in chromosome 19, Loc 965 in Chromosome 6, and Loc1851 in chromosome 12) across multiple BMI-lipid pairs. Loc1351 (Chr8:125,453,323–126,766,827) was protective for all three BMI-lipid pairs. Of these four overlapping loci, three included previously reported protective variants, rs2980888 [11] and rs7005992 [13] in Loc1351, rs7133378 [11, 15, 70], rs7973683 [13], and rs863750 [11] in Loc1851, and rs2075650 [11] in Loc2351 (Table 1 and Table S12).
| Locus | Chr | Start position | Stop position | Discovery pair | Prioritized genesb | BMI-HDL | BMI-LDL | BMI-TG | |||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Local genetic correlation coefficient | -valuep | Local genetic correlation coefficient | -valuep | Local genetic correlation coefficient | -valuep | ||||||
| 158 | 1 | 205,009,624 | 205,917,548 | BMI-LDL | − 0.57 | 2.01e-9 | −0.57 | 0.0000269 | 0.32 | 0.000846 | |
| 498 | 3 | 87,411,259 | 88,375,763 | BMI-LDL | − 0.61 | 0.00000634 | −0.67 | 0.00000312 | 0.48 | 0.00272 | |
| 692c | 4 | 102,544,804 | 104,384,534 | BMI-LDL | − 0.75 | 9.16e-37 | −0.47 | 0.0000385 | 0.68 | 3.08e-12 | |
| 836 | 5 | 73,314,062 | 74,245,354 | BMI-LDL | − 0.52 | 0.00000578 | −0.56 | 9.74e-9 | 0.46 | 0.000458 | |
| 837 | 5 | 74,245,355 | 75,239,302 | BMI-LDL | TWAS:BMI↑ in Artery)GCNT4 (SMR:BMI↓ inBMI↑ inHMGCR (Skeletal Muscle), ANKDD1B (VAT) | − 0.72 | 1.59e-10 | −0.69 | 5.84e-27 | 0.11 | 0.426 |
| 965a | 6 | 32,586,785 | 32,629,239 | BMI-LDL, BMI-TG | − 0.33 | 0.00897 | −0.70 | 5.83e-8 | −0.54 | 0.0000018 | |
| 1185 | 7 | 98,173,565 | 99,465,540 | BMI-LDL | − 0.46 | 0.00000568 | −0.38 | 0.00000806 | 0.35 | 0.00162 | |
| 1246a | 8 | 8,064,601 | 8,589,770 | BMI-TG | TWAS:(BMI↓ in VAT and Artery, BMI↑ in Liver)PRAG1 | − 0.39 | 0.000723 | 0.15 | 0.212 | −0.48 | 8.74e-9 |
| 1247a | 8 | 8,589,771 | 9,167,795 | BMI-TG | TWAS:(BMI↓ in Liver),(BMI↑ in Skeletal Muscle),(BMI↑ in Artery)ERI1CLDN23, MFHAS1PPP1R3BSMR:(BMI↑ in Artery)PPP1R3B | 0.04 | 0.46 | 0.26 | 0.00182 | −0.35 | 0.00000228 |
| 1248c | 8 | 9,167,796 | 9,835,863 | BMI-TG | 0.19 | 0.00101 | 0.36 | 0.00000245 | −0.62 | 1.95e-11 | |
| 1249 | 8 | 9,835,864 | 10,478,851 | BMI-TG | TWAS:(BMI↓ in Artery)MSRA | − 0.02 | 0.706 | 0.51 | 0.0000399 | −0.41 | 2.33e-8 |
| 1251a | 8 | 11,466,762 | 12,296,849 | BMI-TG | TWAS:(BMI↑ in VAT and Artery),(BMI↑ in VAT, SAT, WB, Artery),(BMI↓ in VAT, SAT, WB),(BMI↑ in SAT, WB)NEIL2FDFT1BLKFAM167A | − 0.10 | 0.255 | N/A | N/A | −0.36 | 0.00000227 |
| 1351a,c | 8 | 125,453,323 | 126,766,827 | BMI-HDL, BMI-LDL, BMI-TG | 0.33 | 0.0000133 | −0.54 | 1.51e-14 | −0.50 | 7.14e-14 | |
| 1851a,c | 12 | 123,396,635 | 124,843,768 | BMI-HDL, BMI-TG | TWAS:(BMI↑ in Skeletal Muscle),(BMI↑ in VAT),(BMI↑ in SAT, VAT, Liver, WB),(BMI↓ in WB, BMII↑ in SAT, VAT)RILPL2DNAH10CCDC92ZNF664SMR:(BMII↑ in VAT)CCDC92 | 0.37 | 1.59e-8 | −0.43 | 0.0000863 | −0.53 | 5.92e-12 |
| 2135 | 16 | 53,393,883 | 54,866,095 | BMI-LDL | TWAS:(BMI↑ in Skeletal Muscle)FTO | − 0.56 | 4.69e-30 | −0.73 | 5.36e-21 | 0.18 | 0.00998 |
| 2351c | 19 | 45,040,933 | 45,893,307 | BMI-HDL, BMI-LDL | 0.34 | 9.97e-13 | −0.46 | 1.21e-27 | 0.21 | 0.00000223 | |
Gene prioritization by TWAS-FUSION and SMR
To investigate genes involved in the protective BMI-lipid pleiotropy, we performed TWAS-FUSION and SMR, integrating GWAS summary statistics with GTEx v8 gene expression data across six relevant tissues. We prioritized genes whose genetically predicted expression levels were associated with both BMI and lipid traits in directions discordant with phenotypic correlations (Fig. 1B; Table 1; Tables S13–14). TWAS identified genes at several loci, including GCNT4 (locus 837), NEIL2, FDFT1, BLK, and FAM167A (locus 1251), PRAG1 (locus 1246), ERI1, CLDN23, MFHAS1, and PPP1R3B (locus 1247), and MSRA (locus1249), with expression patterns consistent with protective effects. At locus 1851, RILPL2, DNAH10, CCDC92, and ZNF664 were prioritized, consistent with previous studies of [11] protective adiposity. FTO at locus 2135 also showed a protective association for BMI-LDL. In parallel, SMR identified 11 genes as potential causal mediators of protective BMI-lipid associations (Table S14). Among these, 6 genes (GCNT4, MFHAS1, PPP1R3B, NEIL2, FDFT1, and CCDC92) were also detected in TWAS. Four genes, CCDC92, PPP1R3B, HMGCR, and ANKDD1B, passed the HEIDI test (PHEIDI > 0.05), suggesting the association was not driven by linkage. Notably, CCDC92, supported by both TWAS and SMR in our analysis, was previously reported as a protective adiposity gene.
Associations of BMI-lipid-stratified PRSwith cardiometabolic traits in PAGE BMI
![Associations between PRSand cardiometabolic traits in PAGE study.In the PAGE study, PRSconstructed using BMI-lipid bivariate loci demonstrated distinct associations with cardiometabolic risk. Specifically, PRSbased on protective BMI-HDL loci [PRS] was positively associated with BMI and obesity risk but inversely associated with dyslipidemia (higher HDL, lower LDL, TG, and total cholesterol) and fasting glucose.In the All of Us participants, PheWAS analysis of PRSrevealed protective associations against lipid-related disorders, atrial fibrillation, and cognitive disorders, including Alzheimer’s disease BMI(+)HDL(+) BMI BMI BMI(+)HDL(+) BMI(+)HDL(+) A B](https://europepmc.org/articles/PMC12502452/bin/13073_2025_1522_Fig2_HTML.jpg.jpg "Click to view full size")
Associations between PRSand cardiometabolic traits in PAGE study.In the PAGE study, PRSconstructed using BMI-lipid bivariate loci demonstrated distinct associations with cardiometabolic risk. Specifically, PRSbased on protective BMI-HDL loci [PRS] was positively associated with BMI and obesity risk but inversely associated with dyslipidemia (higher HDL, lower LDL, TG, and total cholesterol) and fasting glucose.In the All of Us participants, PheWAS analysis of PRSrevealed protective associations against lipid-related disorders, atrial fibrillation, and cognitive disorders, including Alzheimer’s disease BMI(+)HDL(+) BMI BMI BMI(+)HDL(+) BMI(+)HDL(+) A B
PheWAS results for the protective BMI-lipid loci-based PRSin AoU BMI
We conducted a PheWAS in the AoU to evaluate the clinical implications of protective PRSBMI. Among 1302 distinct disease outcomes tested, 12 diseases from endocrine/metabolic, mental disorders, and circulatory categories were inversely associated with PRSBMI(+)HDL(+) (Fig. 2B and Table S21). These included hyperlipidemia [OR (95% CI) 0.91 (0.89–0.92), p = 1.72 × 10−33], atrial fibrillation (circulatory) [OR (95% CI) 0.94 (0.91–0.97), p = 1.28 × 10−5], and cognitive disorders such as Alzheimer’s disease [OR (95% CI) 0.71 (0.63–0.80), p = 5.08 × 10−8]. We also observed a positive association between the PRS and an adiposity measure, BMI, (β (SE) = 0.08 (0.02), p = 4.62 × 10−4) in a separate regression analysis of non-EHR data. Furthermore, among the other disease outcomes tested in the PAGE study (MI, stroke, T2D, and hypertension), ischemic heart disease (phecode 411) (i.e., MI) [OR (95% CI) 0.97 (0.96–0.99), p = 6.70 × 10−3] and cerebrovascular disease (phecode 433) (i.e., stroke) [OR (95% CI) 0.97 (0.95–0.99), p = 1.58 × 10−2] were nominally associated with the PRS in a protective direction (Table S21). In the sensitivity analysis using a single phecode definition, no meaningful differences in associations were observed (Table S22). Since the protective BMI-HDL loci included the APOE locus known for strong pleiotropy, in another sensitivity analysis, we excluded the variants within the APOE locus from the PRS calculation. As a result, protective associations with metabolic and circulatory diseases persisted, but the association with dementia disappeared (Table S23). This suggests that the apparent protective effect on dementia is driven by the APOE locus, while the cardiometabolic effects reflect broader polygenic architecture.
For the other two protective PRSBMI [PRSBMI(+)LDL(−) and PRSBMI(+)TG(−)], we did not observe the hypothesized protective associations in AoU (Tables S24–26, S27–28), which is consistent with the results in PAGE. PRSBMI(+)LDL(−) was associated with diseases from multiple categories, including endocrine/metabolic, circulatory, and musculoskeletal; however, the direction of effect for each was non-protective, as opposed to our hypothesis, and more aligned with overall PRSBMI reported from the previous PheWAS studies [71, 72].
Putative causal effects of BMI-lipid bivariate loci on cardiometabolic outcomes (MR results)
To investigate whether lipid traits mediated the effects of protective or adverse BMI loci on cardiometabolic outcomes, we conducted MVMR analyses by including the corresponding lipid trait (HDL or TG) as an additional explanatory variable for outcomes that showed opposite and significant associations (p < 0.05) in univariate MR (Table S30). After adjusting for the corresponding lipid, none of the protective BMI loci remained significant. This suggests that the protective associations observed in univariate MR may be mediated through HDL or TG biology, rather than reflecting lipid-independent effects of adiposity on these outcomes.
Mendelian randomization (MR) results for BMI–lipid bivariate loci and cardiometabolic traits.MR was conducted using BMI-associated SNPs from protective and adverse loci separately for 17 cardiometabolic traits.Protective BMI–HDL loci showed nominal associations with higher HDL, lower TG, and reduced risk of stroke. Protective BMI–TG loci were nominally associated with lower TG, fasting insulin, HOMA-IR, and reduced risk of atrial fibrillation. Adverse loci showed directionally opposite associations, supporting etiologic heterogeneity A B
Implications of BMI-lipid bivariate loci for fat distribution and body composition
Characterization of BMI-lipid loci using MRI-derived regional fat measures. Associations between bivariate loci-stratified PRSand MRI-based fat distribution traits in the UKBB revealed that the PRSbased on protective BMI-TG loci was uniquely and inversely associated with the VAT/ASAT, suggesting reduced visceral fat relative to abdominal subcutaneous fat. No such pattern was observed for other PRS BMI BMI BMI
Discussion
In this study, using large-scale GWAS summary statistics derived from the UKBB, we identified 16 genomic regions with shared genetic underpinnings between BMI and Lipid levels, which were associated with increased obesity risk but decreased risk for dyslipidemia. We further explored the potential causal Genes underlying the protective BMI-lipid bivariate loci using Gene-based TWAS-FUSION and SMR results and prioritized 18 compelling candidate genes for further consideration. Using bivariate loci-stratified PRSBMI, specifically PRSBMI(+)HDL(+), we observed protective associations with lipid-related traits, CVD, and cognitive disorders in independent populations. MR further supported a distinct causal link between protective adiposity (measured through BMI-associated SNPs within the protective BMI-lipid loci) and decreased risk of cardiometabolic outcomes such as atrial fibrillation and stroke. Finally, associations with MRI-based fat distribution measures suggested that these protective adiposity loci may reflect a phenotype characterized by reduced visceral to abdominal subcutaneous fat ratio.
In the loci identification stage using UKBB GWAS summary statistics, the smaller global genetic correlations between BMI and LDL in comparison to BMI-HDL and BMI-TG have been consistently reported in the literature [73, 74]. By performing local-level genetic correlation analysis for BMI and LDL, we investigated whether there is a true lack of genetic correlation (both locally and globally) between BMI and LDL or whether the lack of global genetic correlation is due to the presence of local-level correlations in opposite directions that globally nullify the effects of the other regions. The current study supported both possibilities—(i) a much smaller number of BMI-LDL correlated loci was identified, implying a lack of genetic correlation as compared to the BMI-HDL or BMI-TG pairs, and (ii) the numbers of protective [BMI( +)LDL( −)] and adverse [BMI( +)LDL( +)] loci are comparable. Indeed, many more adverse bivariate loci were discovered compared to protective loci for BMI-HDL and BMI-TG, as expected from the high phenotypic positive correlation between obesity and dyslipidemia [75]. It is also true that, unlike BMI-HDL or BMI-TG results, a similar number of adverse loci and protective loci were identified among BMI-LDL, and they might have nullified each other’s effects, resulting in a small magnitude of global genetic correlation between BMI and LDL. These differences in BMI-lipid pairs (BMI-TG, BMI-HDL vs. BMI-LDL) may suggest the presence of distinct obesity-lipid inter-relationships for HDL and TG versus LDL [76, 77].
By integrating TWAS-FUSION and SMR results with the current local genetic correlation analysis, we prioritized potential causal genes, both novel and known genes, underlying the counterintuitively protective genetic correlations between BMI and lipid traits. As an example of the plausible novel genes, we prioritized the NEIL2 gene within the BMI( +)TG( −) locus (Loc1251). NEIL2, Nei-like DNA Glycosylase 2, is involved in Autosomal Dominant Adult-Onset Proximal Spinal Muscular Atrophy [78] which is relevant for both reduced body weight and an adverse CVD risk profile. According to the GWAS catalog, variants in/near the NEIL2 gene have been associated with TG levels and waist-to-hip ratio adjusted for BMI and other CVD traits, further supporting the NEIL2 gene as a potential causal gene influencing adiposity and CVD. The FDFT1 gene in the same locus was also a plausible candidate. As FDFT1 encodes squalene synthase, it is closely related to cholesterol synthesis, and high serum squalene levels have been associated with abdominal obesity [79]. Furthermore, another gene identified in the locus, BLK, is associated with maturity-onset diabetes of the young (MODY), a subtype of diabetes. In Loc1247, the PPP1R3B gene was prioritized by both TWAS-FUSION and SMR. It is known to play a crucial role in liver glycogen metabolism [80] and may influence lipid metabolism through this pathway. As consecutive loci from Loc1246 to Loc1251, except for Loc1250, were identified as protective BMI-TG bivariate loci, the genomic loci spanning Loc1246 to Loc1251 may harbor protective adiposity loci. The fact that Loc1248 included known protective adiposity variants further supported the current finding. In addition, FTO, a novel potential causal gene within the BMI( +)LDL( −) locus (Loc2135), is a well-established obesity-associated gene, and BMI-increasing risk alleles of the SNPs in FTO have been associated with an adverse cardiometabolic profile [81–83]; however, some previous studies reported paradoxically favorable influences of FTO on cardiometabolic risk profiles [84, 85]. We also identified HMGCR as one potential causal gene for protective BMI-LDL loci. HMGCR is a pharmaceutical target of statins—i.e., inhibition of HMGCR reduces cholesterol synthesis, thereby lowering blood cholesterol levels [86]. Although its role in body weight or adiposity is not fully understood, a previous study reported an association between a SNP in HMGCR and LDL and body weight in opposite directions [87]. As described above, the genes prioritized in this study are each directly or indirectly related to adiposity and lipid metabolism, supporting their potential roles as causal genes. However, further studies are needed to complete functional follow-up and elucidate the mechanisms by which these genes exert their protective influence on lipid and cardiometabolic traits.
In addition to novel genes, several of our findings strengthen the findings from previous studies. For example, in Loc1851, we identified three genes (ZNF664, DNAH10, and CCDC92) associated with BMI, TG, and HDL. Notably, CCDC92 was prioritized by both TWAS-FUSION and SMR analyses. ZNF664, DNAH10, and CCDC92 have previously been prioritized as candidate causal genes for the adiposity variants associated with a protective cardiometabolic profile [11]. Indeed, DNAH10 and CCDC92 have been identified as insulin resistance-related genes, and gene knockdown experiments have supported their associations with the decreased peripheral adipose deposition capacity [13].
Our findings of heterogeneous BMI-lipid bivariate loci support a potential clinical utility of the stratified PRSBMI. From the BMI-HDL bivariate loci, we provide evidence of a generalizable heterogeneous genomic relationship between obesity and dyslipidemia. Notably, in the PAGE study, the increase in PRSBMI(+)HDL(+) was associated with protective lipid values (lower risk of dyslipidemia), while it was associated with increased BMI and increased risk of obesity. These results suggest that shared genetic underpinnings between obesity traits and lipid traits may partly explain the heterogeneous impact of BMI on CVD risk. Of note, despite the small number of variants included in the PRSBMI(+)HDL(+)—i.e., ~ 1500 variants (from three loci) in PRSBMI(+)HDL(+) vs. ~ 1.11 M variants in overall PRSBMI and ~ 373 K variants in PRSBMI(+)HDL(−), its effect size on dyslipidemia was greater than that of from overall PRSBMI and PRSBMI(+)HDL(−) in opposing directions.
PheWAS findings provide an exhaustive list of disease outcomes associated with the protective PRSBMI, which is distinct from the associations with overall PRSBMI or phenotypic association with BMI. First, the PheWAS results confirmed the protective associations of PRSBMI(+)HDL(+) with lipid metabolism and CVDs (lower risk for atrial fibrillation with Bonferroni-corrected significance and ischemic heart disease and cerebrovascular disease with nominal significance) as observed in the PAGE study. This is in contrast to a previous phenome-wide Mendelian randomization (MR) study [71] where overall PRSBMI was adversely associated with lipid-related disorders—e.g., hyperlipidemia [OR (95% CI) 1.47 (1.41–1.54)], atrial fibrillation [OR (95% CI) 1.61 (1.52–1.71)], and ischemic heart disease [OR(95% CI) 1.67 (1.60–1.74)]. The elevated risk of the abovementioned CVDs with the increase in PRSBMI has also been reported in eMERGE [72]. Thus, our results support the presence of distinct adiposity-associated genetic loci heterogeneously influencing CVD risk.
In addition, we observed protective associations of PRSBMI(+)HDL(+) with several cognitive disorders, such as mild cognitive disorders, memory loss, Alzheimer’s disease, and dementia. The phenotypic association between BMI and cognitive disorders has been inconclusive; some studies reported that lower BMI was associated with an increased risk of Alzheimer’s disease [88–90], whereas other study reported obesity and overweight as a risk factor for dementia [91]. To our knowledge, most MR studies did not find evidence of any association between genetically influenced BMI and dementia [71, 92, 93] or reported a positive association between genetically predicted BMI and Alzheimer’s disease [94]. Unlike previous studies based on overall BMI genetics, the current findings suggest that specific local-level BMI genetics may play a protective role against cognitive disorders, potentially mediated through HDL levels, contrasting with global-level BMI genetics. The current results may suggest potential roles of favorable adiposity in protecting against cognitive disorders, which may be reflected in the often-reported inverse phenotypic association between BMI and cognitive disorders. We also note that our PRSBMI(+)HDL(+) includes the well-known lipid- and dementia-associated APOE locus. In a sensitivity analysis excluding the APOE locus from the PRS, the association with dementia is no longer observed. Thus, the observed protective effects of the PRSBMI(+)HDL(+) on dementia risk are likely driven by APOE variants. Further studies are warranted to replicate the findings and elucidate the underlying biological mechanisms.
MR analyses highlight a distinct causal relationship between BMI-associated SNPs within the protective BMI-HDL and BMI-TG loci and favorable cardiometabolic outcomes, including increased HDL, decreased TG, fasting insulin, and HOMA-IR, and reduced risk of stroke and atrial fibrillation. In contrast, BMI-associated SNPs within the adverse BMI-HDL and BMI-TG loci showed significant associations with these conditions in the opposite direction. These findings also contrast with traditional MR results using genome-wide BMI-associated SNPs, as well as our reference analysis, in which BMI SNPs were generally associated with adverse cardiometabolic profiles [95–97]. Further MVMR analyses suggested that the protective effects on these cardiometabolic conditions appear to be mediated through HDL or TG. Together with PRS analyses and PheWAS findings, our MR results support the existence of heterogeneous forms of adiposity, some with protective and others with adverse downstream cardiometabolic effects. Our findings suggest that the identified loci may contribute to the genetic architecture underlying protective adiposity.
Using MRI-based fat distribution measures, we observed a distinct inverse association between protective PRSBMI(+)TG(−) and VAT/ASAT, compared to both the reference PRSBMI and adverse PRSBMI(+)TG(+) groups. This unique association suggests that the protective adiposity captured by BMI-TG loci may reflect a lower proportion of VAT, which is considered metabolically detrimental [98], relative to ASAT. These findings suggest that the identified protective adiposity loci may contribute to heterogeneous fat distribution patterns, consistent with prior evidence indicating that the cardiometabolic consequences of adiposity depend on the location of fat depots [99].
The present study has some limitations. First, the genomic partitioning was based on a European-ancestry LD structure (1000 Genomes European population [25]); thus, the partitioned genomic regions may not apply well to non-European ancestry populations. We may, therefore, have missed ancestry-differentiated genetic correlations. In addition, since BMI is a crude proxy for adiposity, our approach may have missed important pleiotropic loci that link obesity to lipid traits. However, BMI is widely used and readily measured in large-scale cohorts, allowing for substantially larger sample sizes. These larger sample sizes, in turn, substantially increase statistical power to detect genetic associations, particularly for variants with modest effect sizes. This enhanced power helps to identify robust signals that may be relevant to more refined obesity-related traits. Thus, the gains in power afforded by large sample sizes may help to offset the limitations related to its lack of specificity in capturing diverse aspects of adiposity. Nonetheless, we acknowledge that alternative anthropometric measures such as waist-to-hip ratio or waist circumference, as well as imaging-based adiposity measures, may better capture fat distribution and its metabolic consequences. Future studies incorporating these more specific indices will be important to validate and extend our current findings.
The current study has notable strengths as well. First, the total sample size of the PAGE study was large, and we were able to evaluate the relationships between BMI-lipid bivariate loci and various CVD profiles with individual-level phenotype data. In addition, the distribution of self-identified race/ethnicity in the PAGE study, especially across White, Black, and Hispanic/Latino self-identified populations, was well-balanced, thus equally contributing to the population-pooled results. Furthermore, this study implemented a novel locus-based approach to identify BMI-lipid bivariate loci and proposed a novel application of locus-restricted PRS to evaluate the influence of certain genomic loci on phenotypes. We were also able to leverage an independent, large, and diverse US population (All of Us) for PheWAS analyses.
Conclusions
We identified two distinct types of BMI-lipid bivariate loci in opposing directions (protective versus adverse) and suggested potential causal genes (e.g., NEIL2, FDFT1, or BLK) underlying protective BMI-lipid loci. Notably, the bivariate loci-stratified PRSBMI, specifically PRSBMI(+)HDL(+), provided evidence of protective influences on lipid-related traits, CVD, and cognitive disorders in independent populations. MR analyses further supported causal relationships between protective BMI loci and multiple cardiometabolic outcomes. The current findings suggest the presence of heterogeneous obesity-related genetics at the local level, which cannot be observed at global-level obesity genetics. With much larger sample sizes, disease-risk stratification by integrating both protective PRSBMI and adverse PRSBMI could be clinically meaningful. Specifically, individuals with comparable genetic risk for overall adiposity can be further stratified based on their genetic predisposition to metabolically favorable and unfavorable adiposity. This stratification may enable more tailored clinical monitoring or preventive interventions, for example, prioritizing intensive lipid management or early lifestyle interventions in individuals with a high genetic predisposition to adverse adiposity or a low genetic predisposition to protective adiposity.
Supplementary Information
Additional file 1: Supplementary methods. Detailed description of the methods used in this study. Includes: (1) overview of PAGE-participating cohort studies, (2) phenotype definitions and measurement procedures, (3) cardiovascular disease ascertainment protocols used across PAGE studies, (4) GWAS of BMI and lipid traits in the UKBB, and (5) analytical framework for local genetic correlation analysis. Additional file 2: Supplementary Tables.