Study cohort

A total of 3,000 individuals (1,725 T2D cases and 1,275 controls) were included in this study from the Asian Indian Diabetic Heart Study (AIDHS)/Sikh Diabetes Study (SDS) [15–17]. The Sikh population is a relatively homogenous endogamous community from Northern India. Sikhs are primarily non-smokers, and ~50% of them are vegetarians. However, the incidence of cardiometabolic diseases in Sikhs and SAs has markedly increased over the past two decades [1,18]. T2D was diagnosed based on their medical records for symptoms and use of diabetic medications, and following the American Diabetes Association guidelines as described earlier [16,17]. Non-diabetic controls (N = 1,275) were selected based on a fasting blood glucose (FBG) <100.8 mg/dl (5.6 mmol/l) or a 2-hour glucose <141.0 mg/dl (7.8 mmol/l). All blood samples were obtained at the baseline visit. Subjects with type 1 diabetes, those with a family member with type 1 diabetes, or rare forms of T2D subtypes (maturity-onset diabetes of the young [MODY]) or secondary diabetes (from, e.g., hemochromatosis or pancreatitis) were excluded from the study based on clinical reports, as previously described [15,17,19]. Coronary artery disease (CAD) was considered if there was use of nitrate medication (nitroglycerine), electrocardiographic evidence of angina pain, coronary angiographic evidence of severe (greater than 50%) stenosis, or echocardiographic evidence of myocardial infarction. The diagnosis was based on the date of coronary artery bypass graft (CABG) or angioplasty and medication usage obtained from patient records, as described previously [17,20]. All participants in this study were recruited following the written informed consent procedures approved by the institutional review boards (IRBs). All AIDHS/SDS protocols and consent documents were reviewed and approved by the University of Oklahoma Health Science Center’s (IRB #2911 and IRB # 13302), the University of Pittsburgh (IRB # 021234), as well as the Human Subject Protection (Ethics) Committees at the participating hospitals and institutes in India, as described previously [14,21–23]. The local institutions and hospitals also separately obtained Federal Wide Assurance (FWA) from the Office of Human Research Protection, US Department of Health and Human Services (S1 File). All human studies reported in this manuscript abided by the Declaration of Helsinki principles.

BMI was calculated as [weight (kg)/height (m²)]. A tape measure of the waist and hip circumferences at the abdomen and the hip, respectively, were recorded. For using BMI thresholds for obesity, we used the World Health Organization’s (WHO) guidelines [24]. Blood pressure (BP) was measured twice after a 5-minute seated rest period with the participant’s feet flat on the floor. Serum lipids [total cholesterol (TC), TGs, high‐density lipoprotein cholesterol (HDL‐C), and low‐density lipoprotein cholesterol (LDL‐C)] were measured using standard enzymatic methods (Roche, Basel, Switzerland) as described previously [17,20,25–27]. The design and the workflow for the study are shown in Fig 1.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Summary of the workflow detailing the study design, lipidomics analysis, and the outcomes.

https://doi.org/10.1371/journal.pmed.1005039.g001

Inclusivity in global research

Additional information regarding the ethical, cultural, and scientific considerations specific to inclusivity in global research is included in the Supporting Information (S1 Text).

Metabolomics/lipidomics profiling, quality control (QC), and analysis

Aliquots of 50 µL serum/plasma were shipped on dry ice to the UC Davis West Coast Metabolomics Center to measure lipidomics profiles using liquid chromatography–tandem mass spectrometry (LC–MS/MS) in untargeted mode. The serum and plasma aliquots were extracted with a degassed ternary mixture of methanol, MTBE, and water, as described in detail previously [28]. The data produced by LC–MS/MS is a relative quantification method that utilizes internal standards as surrogate markers for quantification. The LC–MS/MS analyses were conducted using an Agilent 1290 Infinity LC system (Agilent Technologies, Santa Clara, CA, USA) equipped with a Waters Acquity CSH C18 2.1 × 100 mm, 1.7 μm particle size column, maintained at 65 °C at a flow rate of 0.6 mL/min. The mobile phases consisted of (A) 60:40 (v/v) acetonitrile:water with 10 mM ammonium formate/0.1% formic acid and (B) 90:10 (v/v) isopropanol:acetonitrile with 10 mM ammonium formate/0.1% formic acid. Chromatography gradient was: 0 min 15% (B); 0–2 min 30% (B); 2–2.5 min 48% (B); 2.5–11 min 82% (B); 11–11.5 min 99% (B); 11.5–12 min 99% (B); 12–12.1 min 15% (B); and 12.1–15 min 15% (B). Mass spectrometric detection of lipids was performed on Agilent 6530 and 6546 QTOF mass spectrometers [28] in data-dependent MS/MS mode using both electrospray ionization positive and negative ion modes. Raw data were processed using MS-DIAL (v. 4.90) software [29] from 280 to 1,500 Da from 0.3 to 12.6 min with a centroiding tolerance of 10 mDa and a minimum peak height amplitude of 500 with a mass slice width of 50 mDa, an accurate mass tolerance of 10 mDa, and 6s alignment retention time tolerance as described [29].

We used three types of internal standards for the analysis: 43 Biorec (commercial), 43 blank, and 43 pooled samples (derived from representative samples of the AIDHS/SDS population) to assess the technical variance and robustness of the analytical method, as described previously [30,31]. Lipids were identified by MS/MS matching to LipidBlast within MS-DIAL at MS1 difference <10 mDa and MS/MS similarity >700 with manual inspection of all similarity matches. All peaks had signal-to-noise ratios greater than 3:1 for annotated compounds and greater than 10:1 for unknowns, using blank samples as negative controls. All data peaks were processed using MS-DIAL 4 and the concentrations (pmol per μl serum or plasma) were calculated by using internal standards, where the tag of level 2 was assigned if the lipid was quantified by an internal standard of the same lipid polar head class, and the tag of level 3 was assigned if the lipid was quantified by an internal standard of a similar lipid class or representative standard compound based on the Lipid Standard Initiative (LSI) guidelines (https://lipidomics-standards-initiative.org/) as described previously [29]. The identification of unknown MS/MS spectra was elucidated by using internal standards, consulting the literature, or predicting the putative structure based on fragment ion evidence.

Batch correction

The samples were sent in two batches, each containing 1,500 samples, to measure the lipidomics profile. Each batch was sent separately at a different time, with a gap of approximately a year between processing. QC was assured by randomizing injection sequences and injecting 10 QC pool samples before the actual sequence of samples, injecting QC pool samples at the beginning and the end of each batch sequence and between each 10 actual samples, injecting QC method blanks after each set of 10 actual samples, checking the peak shape and the intensity of spiked internal standards and the internal standard added before injection. The QC pool samples were used to correct for batch differences, longitudinal drift, and instrument variation, causing technical data variance by random forest machine learning in the SERRF algorithm [32]. SERRF defines systematic errors as errors not only associated with batch effects and injection order but also considers patterns of related compounds. The cross-validated relative standard deviation (coefficient of variation, %CV) is used to evaluate the data performance. Per batch, reproducibility was achieved with a 7.3% CV as the median overall acquired lipids in positive electrospray ionization (ESI) mode using the sample pool QCs. For the negative ESI mode, reproducibility was measured at a median 8.6% CV across all lipids for the sample pool QCs.

Genotyping, imputation, and quality controls

Genomic DNA was extracted from buffy coats using QIAamp blood kits (Qiagen, Chatsworth, CA) or by the salting-out procedure [33]. Samples were genotyped using the Illumina 660W Quad BeadChip, Illumina Global Screening Arrays (GSA), and GSA with multi-disease content (GSA+) arrays as described previously [19,20,34]. Both discovery (training) and test sets for Punjabi cohorts originated from the AIDHS/SDS, excluding families due to relatedness. The discovery set was genotyped using Illumina’s 660 Quad and the training cohort with GSA+ arrays. There was no sample overlap in the training and test datasets. Also, samples with genotyping call rate <95%, cryptic relatedness, population outliers, departures from Hardy–Weinberg equilibrium (HWE) (p < 10⁻⁷) or minor allele frequency (MAF) <5% were excluded before association testing. To increase genome coverage, data were imputed using Minimac4 [35] (https://imputationserver.sph.umich.edu/) with TOPMED r3 multiethnic reference panel in NCBI Build 38 (hg38) coordinates as reported previously [3,36]. Of a total of 23,739,260 variants, we removed variants with an imputation certainty info score <0.5, MAF < 0.001, and with HWE in controls (p < 1 × 10⁻⁶) before further analysis. The genetic principal components (PCs) were estimated from our Sikh population, as the existing HapMap2, HapMap3, and 1000 Genomes data do not include data from Punjabi Sikhs, as described previously [17,34]. The lipid metabolite PCs were calculated from the metabolite data generated for the 3,000 AIDHS/SDS individuals for serum and plasma, respectively, using a standard PCA algorithm, which is an orthogonal transformation that captures the maximum variance in the data. This process involves centring the data, calculating the covariance matrix, and then finding the eigenvectors and eigenvalues of that matrix [13,37].

Statistical analysis

The clinical and demographic variables were summarized using the mean for continuous variables and percentages for categorical variables using the SPSS software version 29 (IBM, New York City, USA). Multivariate linear regression analyses were performed to assess the impact of individual metabolite markers on T2D, CAD, and other cardiometabolic risk factors (e.g., BMI, waist, FBG) after adjusting for covariates such as age, sex, and medications as described previously [13]. We conducted a lipid metabolite GWAS on 385 serum and 321 plasma metabolite species that are associated with T2D and other cardiometabolic traits. Genome-wide analysis was performed after adjusting for age, sex, BMI, T2D, CAD, 5 genetic PCs, and 5 metabolite PCs in Model 1 and including blood lipids in Model 2, along with covariates of age, sex, BMI, T2D, CAD, 5 genetic, and metabolite PCs [38]. We further performed association analysis for top mQTLs after adjusting for age, sex, BMI, T2D, CAD, 5 genetic PCs, and blood lipids in Model 3. For selecting the independent signals, LD clumping was used with an LD threshold (R² > 0.6). All regression analyses were performed using PLINK 2.0 [39]. We assessed the pairwise correlation between selected lipids and phenotypic traits such as blood glucose, TG, systolic blood pressure (SBP), and diastolic blood pressure (DBP) using SPSS software version 29 (IBM, New York City, USA). Logistic regression analyses were performed to assess the impact of metabolite-associated SNPs on T2D and CAD after adjusting for covariates using SVS version 8.9.1 (Golden Helix, Bozeman, MT, USA). The heritability of each lipid metabolite was calculated using the genome-wide complex trait analysis (GCTA) software package [40] and selecting genome-wide significant variants. All analyses were performed using PLINK 2.0 [39], SVS version 8.9.1 (Golden Helix, Bozeman, MT, USA), the SPSS software version 29 (IBM, New York City, USA), and Metaboanalyst 6.0 (Quebec, Canada).

Independent validation cohorts

Genome-wide polygenic risk score (PRS) construction and analysis

Asian Indian ancestry-specific PRS (PRSAI) for T2D and CAD were constructed using candidate variants derived from genome-wide genotypes of AIDHS/SDS and other robustly associated variants reported for GWAS studies of CAD and T2D for people from SA studies [17,20,34,49,50]. The details of SNP selection criteria from 46,985,978 common and rare variants for the PRS construction have been described previously [3,36]. After regression analysis (1) we selected all significant SNPs (p < 10⁻²), (2) QC was performed (a) excluding INDELs, duplicate and multiallelic SNPs, (b) SNPs with info score <0.80, (c) including SNPs with MAF > 0.01 and MAF < 0.45, (3) Remaining SNPs with p < 10⁻⁴ were chosen. (4) Then we removed clumping-based linkage disequilibrium (LD) using R² (LD) <= 0.25 and 500 Kb distance, a total of 2,921 and 925 significant SNPs were used for the construction of the PRSAI in T2D and CAD, respectively [3,36,51]. To construct European PRS (PRSEU), we used the summary statistics data from O’Connor and colleagues, [47] for T2D PRS and 113,320 variants from the Myocardial Infarction Genetics and CARDIoGRAM Exome chip meta-analysis [52]. The selection criteria of SNPs for constructing European PRSEU were the same as PRSAI. A total of 1847 and 405 significant SNPs were selected for the construction of the PRSEU in T2D and CAD, respectively. The individual-level regression coefficients were multiplied by the number of risk alleles to compute the PRS in training and test sets as described previously [3,36,53]. The weighted PRS was calculated using the following equation:

where N is the number of SNPs in the score, β_i is the effect size (or beta) of variant i, and dosage is the number of copies of SNP in the genotype of individual j [54].

For exploratory analysis, we also constructed the genome-wide PRS for selected top lipid species, which showed a strong causal association with cardiometabolic traits using 2,577 and 2,538 SNPs, respectively, for LPC O-16:0 and PC 38:4 (C), using the same QC details described above for T2D and CAD PRS. The goal was to compare the genetic contribution of a single lead variant related to lipid traits with that of combined polygenic variants concerning the disease.

MR analysis

We performed a two-sample MR analysis [55] to investigate the causal effect of metabolite-associated SNPs on T2D and CAD. The associations between the instrumental variables (gene variants) and the exposure (metabolite) and the outcome (T2D, CAD, or risk factor phenotypes) are estimated from different studies, mainly AIDHS/SDS, SA (UKBB)), EU (UKBB), data from O Connor and colleagues [47], and DIAMANT consortium (mixed ancestries). Three basic hypotheses were considered while conducting two-sample MR: (1) genetic instrument variables (IVs) should be robustly associated with the exposure; (2) IVs should not be directly correlated to the outcome and affect the outcome merely via the exposure without any gene pleiotropy; and (3) IV should be independent of any potential confounders. The combined SNP-specific estimates were calculated using the inverse-variance weighted (IVW) method when more than 2 associated SNPs were used as IVs. The odds ratios (ORs) of T2D and CAD were calculated per 1-SD increase in genetically predicted metabolite levels. Sensitivity analyses were performed using the MR Egger method of Burgess and Thompson [56], which is based on the hypothesis that the pleiotropic effects are independently distributed from the genetic associations with the exposure. A non-zero intercept is meaningful to MR Egger, signifying that gene pleiotropy is considered to exist. MR analyses were performed using the Two-sample MR package [57] in R version 4.3.3. We ensured that the genetic instrument was strongly associated with the exposure in the target (ancestry similar or non-similar) population based on regression (beta) coefficients and p value/F statistics accounting for the LD and allele frequency. Based on the differences in LD and MAF, the MR sensitivity analysis selected and excluded the variants from each ancestry to ensure data harmonization and reduce pleiotropy.

Colocalization and pathway analyses

We further investigated whether a given genetic association of exposure (metabolite) was causally influencing the disease or trait outcomes by performing colocalization analysis using FUMA [58]. We analyzed each genetic region (maximum p value of lead SNPs < 10⁻⁸ and R² threshold for independent SNPs < 0.5) for each associated QTL of interest by expanding 5Kb on each 5′ and 3′ site to explore if the genetic associations of exposure and the disease share the same underlying variants in a specific region. This analysis validates the causal association determined by MR and helps narrow down the potential causal genes. We also investigated whether the genetic association with the exposure or the disease is influenced by gene expression QTLs (eQTLs) in a particular tissue using FUMA. We identified chromatin interaction between eQTLs and disease or metabolite QTLs (mQTLs) using gene set enrichment analysis (GSEA). To gain insight into the underlying biological mechanisms of the identified set of significant mQTLs and disease associations, we performed over-representation analysis (ORA) implemented in integrated molecular pathway level analysis (IMPaLA) software [59] and using the data sources from Reactome, Kyoto Encyclopaedia of genes and genomes (KEGG), Encyclopaedia of human genes and metabolism (HumanCyc), and Wiki pathways as described previously [60].

Fine mapping

Fine mapping analysis of the selected mQTLs were performed using a Bayesian fine-mapping method implemented in SuSiE using the GWAS summary statistics [61]. The sum of single effects model or SuSiE uses the sparse vector of regression coefficients as a sum of single-effect vectors, each with one non-zero element, and employs an iterative Bayesian stepwise selection (IBSS) that, instead of selecting a single variable at each step, computes a distribution on variables that captures uncertainty in which variable to select. IBSS thus helps in optimizing a variational approximation to the posterior distribution or posterior inclusion probability (PIP). The PIP is a value of each variable’s probability to be included in the true model in Bayesian analysis providing a credible set of variables for each selection [61].

Replication studies and the discovery of population-specific mQTLs

For validation studies, we examined the association of all 236 mQTLs (p </= 5 × 10⁻⁸) identified in AIDHS/SDS in other independent cohorts available with lipidomics profiles. We considered our mQTL variant(s) to be new based on their associations with the metabolite or related metabolites, and CV disease and risk factors were not listed in the public repositories in context to the metabolite or the trait. For instance, the new mQTLs for LPC O-16:0 (ELOVL4, ELP4, PTPRC/CD45, TMEM108, RORA, CDH23, HELZ2, and ROBO2), were only detected in our study and no association was observed in any other published study. We further extended our search ± 1.5 Mb from our mQTLs in other studies and did not find any variant in LD with our lead SNP or proxy SNP showing association for the specific mQTL or other metabolites. However, extending this search beyond 6Mb from our lead variant in the TNS3 (rs371318240) associated with LPC O-16:0, we identified two variants located upstream to the TNS3 variant encoded by SUGCT gene (rs554961731 and rs570248035) showing association with LPC O-16:0 in Cadby and colleagues [42]. However, these variants were not in LD with our lead SNP in the TNS3 region. Further evaluation of the SUGCT gene in AIDHS/SDS did not yield any significant association with LPC O16:0 or any other metabolite (S19 Table).

The mQTLs identified on the chromosome 11 region for FAs, PCs, and TGs were independently replicated in other external cohorts, SA and EU studies. In contrast to studies conducted in the EU, which have predominantly identified the chromosome 11 region (FADS1/2/3) as harboring loci linked to FAs, our research found no significant associations with any FA reaching the GWAS empirical p-value threshold for this region, except for arachidonic acid (FA 20:4) in Asian Indians that disappeared after adjusting for clinical lipids (S2 Table). In contrast, we identified a mQTL signal associated with the FADS1, FADS2, TMEM258, FEN1, and MYRF genes in this region with PC 38:4 (C) (S5 Table). The same variants rs174547 (FADS1) (Beta ± SE = −0.15 ± 0.01; p </= 7.2 × 10⁻¹¹⁵) and rs102275 (TMEM258) (Beta ± SE = −0.14 ± 0.01; p </= 1.5 × 10⁻¹⁰⁵) were significantly associated with PC 38:4 (C) in Pakistani cohort, as seen in the AIDHS/SDS and were in the same direction. However, the same variants rs174547 (FADS1) (Beta ± SE = −0.47 ± 0.02; p </= 5.8 × 10⁻⁹⁸) and rs102275 (TMEM258) (Beta ± SE = −0.47 ± 0.02; p </= 3.9 × 10⁻⁹⁹) were associated with another related PC (PC 38:5) in the BHS cohort in Australians and PC18.2.0_20.4.0/PC38:6 in Finnish (Beta ± SE −0.17 ± 0.02, p </= 9.6 × 10⁻²⁴) (S5 Table). Additionally, we also detected many robustly associated variants from different genomic regions with FA18:0;(2OH), represented by ADGRB2, PPARGC1α, SPP1, HDAC5, PTK7, and TRIP4 genes, which have not been reported in any genome-wide study for lipid metabolites (S3 Table).

Phenotypic correlation of serum metabolites with other cardiometabolic traits

We further determined the phenotypic correlation between the top mQTLs with other cardiometabolic traits, and significant results are presented in heat maps (S3 Fig). Circulating levels of all FAs correlated positively and significantly with FBG in AIDHS/SDS. Similarly, glycerophospholipids such as LPCs showed a significant positive correlation (ranging from r = 0.14 to 0.24 and p </= 2.7 × 10⁻¹³ to 1.1 × 10⁻³⁸) with FBG except for LPC 16:1. On the other hand, most of the PCs correlated negatively with FBG (ranging from r = −0.04 to −0.14 and p </= 0.04 to 1.7 × 10⁻¹⁴). Likewise, most high-carbon-containing sphingolipids (SMs), SM d36:1 and SM d41:2 Isomer B, were reduced with FBG and TG, while high-carbon-containing neutral lipids, TG 50:2 and TG 50:6, were significantly reduced with increased FBG (S3 Fig).

Association of mQTL variants with cardiometabolic phenotypes in UKBB

We further explored the association of all 236 mQTL variants with blood glucose, glycated hemoglobin (HbA1c), creatinine, C-reactive protein (CRP), apolipoprotein B (ApoB), atherosclerosis, arterial embolism and thrombosis, cerebral infarction, intracranial hemorrhage, and subarachnoid hemorrhage using phenotype data from UKBB. We identified 205 significant associations of top mQTL variants with cardiometabolic traits. The vast majority of the variants were from the FADS region represented by FADS1/2, MYRF, FEN1, and TMEM258 genes, which showed strong association with ApoB, CRP, creatinine, glucose, and HbA1C levels with p values ranging between 5.0 × 10⁻⁸ to 2.5 × 10⁻⁸⁹. These genes represented the top mQTLs for PC 38:4 (C), TG 53:4, CE 20:4, FA 20:4, FA 20:5. Another mQTL for SM 35:2;2O|SM 18:2;2O/17:0 in ELP1 (chromosome 9) showed a strong association with glucose levels (p </= 3.1 × 10⁻⁰⁸) and HbA1C (p </= 2.8 × 10⁻⁰⁷). Additionally, the mQTL signals identified for FA 18:0;(2OH) in PTK7, and PLCG2 for LPC O-16:0 were associated with ApoB, creatinine, and CRP levels at a modest p </= 1.3 × 10⁻³ (S6 Table).

Univariable Mendelian randomization (MR) and sensitivity analysis

To further evaluate the causal link between mQTLs and T2D or CAD, we utilized genetic instruments in the application of two-sample MR. Using univariable MR and sensitivity analysis applying the IVW, weighted median, weighted mode, maximum likelihood, and MR-Egger methods for fixed effects (FE) and random effects (RE), we tested the associations of all significantly associated mQTL on T2D or CAD and other cardiometabolic risk factors in other ancestries. We followed a strict quality control and systematic approach to ensure that a genetic instrument derived from one ancestry is valid for other ancestry using univariable 2-sample MR. We ensured that the genetic instrument was strongly associated with the exposure in the target (ancestry similar or non-similar) population based on regression (beta) coefficients and p-value/F statistics accounting for the LD and allele frequency. LPC O-16:0 showed significant positive associations with T2D and CAD. The MAFs for LPC O-16:0 associated mQTLs used for MR analysis in AIDHS, SAs, and EUs are mentioned in S7a Table. Genetically instrumented per 1SD increment of blood LPC O-16:0 level would increase T2D risk to over ~2-fold in SAs from UKBB OR 1.75 (95% CI [1.34,2.16]; p </= 0.007). A similar but less significant trend was seen in EUs, because the 7 top variant mQTLs of LPC O-16:0 detected in Asian Indians in this study were monomorphic in EUs. Also, the increased concentration of LPC O-16:0 would increase the risk of CAD in UKBB SA OR 1.75 (95% CI [1.22,2.28]; p </= 0.04) and UKBB EU OR 1.17 (95% CI [1.02,1.33]; p </= 0.03) (Fig 4; S8 and S9 Tables). Both heterogeneity and pleiotropy were minimal and not significant for both the LPC O-16:0 effect on T2D and CAD (S10a and S10b Table).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Scatter plots and forest plots showing genetic association between LPC O-16:0, T2D, and CAD.

Using univariable MR and sensitivity analysis applying the IVW, weighted median, weighted mode, maximum likelihood, and MR-Egger methods for fixed effects (FE) and random effects (RE) (A) Scatter plots on top left to bottom show the effect of LPC O-16:0-associated SNPs on the exposure (metabolite) on the x-axis and outcome (T2D) on the y-axis with each dot represent individual SNP in SA, and EU. (B) Forest plot shows the IVW ORs and p-values of LPC O-16:0-associated SNPs on T2D in SA and EU. (C) Scatter plots on top left to bottom right show the effect of LPC O-16:0-associated SNPs on the exposure (metabolite) on the x-axis and outcome (CAD) on the y-axis, in SA, and EU. (D) Forest plot shows the IVW ORs and p-values of LPC O-16:0-associated SNPs on CAD in SA, and EU. MR, Mendelian randomization; T2D, type 2 diabetes; CAD, coronary artery disease; SA, South Asian; EU, European. The p-value was calculated using IVW regression.

https://doi.org/10.1371/journal.pmed.1005039.g004

Genetically instrumented decrease in PC 38:4 (C) due to variation in FADS1/2 and FEN1, variants showing negative ORs (0.90 to 0.77; p </= 8.1 × 10⁻¹² to 1.2 × 10⁻³⁵) was associated with a significantly increased risk for CAD in AIDHS/SDS with ORs and p values ranging from (1.28 to 1.36; p </= 0.007 to 9.6 × 10⁻⁴) (S5 Fig). The effects of these variants on CAD in SA from UKBB were in the same direction but were not significant (S11 Table). The MAFs for PC 38:4 (C) associated mQTLs used for MR analysis in AIDHS, SAs, and EUs are mentioned in S7b Table. On the other hand, in the UKBB EUs, the beta directional effects were opposite. Aside from the two metabolites (LPC O-16:0 and PC 38:4 (C)), no other lipid metabolites could be confirmed to have a causal association with T2D, CAD, or other cardiometabolic risk traits such as glucose, TG, SBP, and DBP using the MR approach. The genetic variance accounting for phenotypic variance in LPC O-16:0, measured by R-squared (R²), varied from 0.01% to 4.5% for each independent lead variant, while the aggregate R² of the independent variants ranged from 15.1% to 20.9%. However, the total genetic variance (R²) explained by these variants for increasing the risk of T2D and CAD was 1.8% and 1.6%, respectively.

Association between lipid metabolites with PRS

We next tested the association of the ancestry-specific and EU-derived PRS for T2D and CAD with all 262 mQTLs using a multivariate linear regression analysis adjusted for age, gender, and BMI. None of the top lipid metabolites revealed significant and positive associations with T2D PRS in both ancestry-specific or EU-derived polygenic scores (S12a Table). On the other hand, several polyunsaturated fatty acids (PUFAs) were positively associated with CAD PRS. Specifically, FA 18:1 (Beta ± SE = 0.007 ± 0.001; p </= 1.26 × 10⁻¹⁴), FA 18:2 (Beta ± SE = 0.006 ± 0.0009; p </= 3.77 × 10⁻¹³), FA 20:3 (Beta ± SE = 0.006 ± 0.0007; p </= 3.18 × 10⁻¹⁷), and FA 20:4 (Beta ± SE = 0.005 ± 0.0008; p </= 1.99 × 10⁻¹¹) were significantly increased with increased PRS (S12b Table), showing the R² range between 4.0% to 4.9% and 2.4% to 3.1% in SA and EU PRS, respectively. The genome-wide metabolite PRS (for the individual metabolite) of LPC O-16:0 (using 2,577 significant SNPs with p </= 10⁻⁴) revealed a strong association of PRS with LPC O-16:0, showing Beta ± SE 0.51 + 0.003; p </= 2.72 × 10⁻⁸⁰, after adjusting for age, gender, BMI, T2D, CAD, and clinical lipids (S13a Table) explaining the genetic variance of 23.1% (R²). Similarly, the genome-wide metabolite-PRS for PC 38:4 (C) (using 2,538 significant SNPs with p </= 10⁻⁴) showed a robust association with PC 38:4 (C) metabolite (Beta + SE 0.70 + 0.002; p </= 1.84 × 10⁻²⁶⁵) after adjusting for covariates (S13a Table), and with the R² of 36.7%. However, despite the high genetic variance, neither of the metabolite PRS for LPC O-16:0 nor PC 38:4 (C) could significantly predict the risk of T2D or CAD. LPC O-16.0 PRS showed an OR 0.98 (95% CI [0.58,1.39]; p </= 0.94) for T2D and OR 1.03 (95% CI [0.48,1.58]; p </= 0.91) for CAD, and showed an R² range of 0.2% to 0.9%, respectively. Similarly, PC 38:4 (C) PRS showed an OR 0.97 (95% CI [0.52,1.42]; p </= 0.89) for T2D and OR 1.00 (95% CI [0.39,1.61]; p </= 0.99) for CAD, having an R² range of 0.4% to 0.2%, respectively (S13b Table). Using the polygenic score only from the top 44 SNPs of LPC O-16:0, the PRS analysis predicted that increased LPC O-16:0 would significantly increase the T2D risk in AIDHS/SDS showing an OR 1.06 (95% CI [1.02,1.09]; p </= 0.004) in Models 1 and OR 1.05 (95% CI [1.02,1.09]; p < / = 0.005) in Model 2 with clinical lipids. However, the R² value for T2D was marginally increased from 0.2% of the original metabolite to 0.7% (using the high-impact 44 SNPs) (S13b Table).

Colocalization analysis of mQTLs with disease outcomes, co-sharing, and pathways

Further fine mapping analysis by FUMA of chromosome 1q31.1-q31.3 associated with LPC O-16:0 (containing PTPRC, CRB1, and DENND1B genes) demonstrated strong chromatin interactions with several other genes, including CFHR5, KIF14, ZNF281, ZBTB41, NEK7, and KCNT2, reported previously to be associated with CAD and CV traits in the GWAS catalogue (Fig 5A and S14 Table). Colocalization analysis also identified the region within PTPRC (rs558096054), linked to LPC O-16:0 (in regression analysis) to be associated with an increased risk of CAD OR 3.35 (95% CI [2.62,4.07]; p </= 2.3 × 10⁻⁶) (Fig 5B). Moreover, the top lead SNP in the PTPRC and other variants associated with LPC O-16:0 (rs142028924; RORA) and rs145555444 (ROBO2), and (rs540305535; LEPR) have been previously reported in GWAS of CRP, Asthma, blood lipid phenotypes, Inflammatory bowel disease (IBD), SBP, T2D, BMI, waist-to-hip ratio (WHR), apolipoprotein A1 (ApoA1) with p-values ranging from 3.0 × 10⁻⁰⁸ to 1.0 × 10⁻⁵⁶ (Table 2).

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Genomic regions identified by the top lead variant showing co-sharing with lipid species and other diseases.

https://doi.org/10.1371/journal.pmed.1005039.t002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Colocalization and functional mapping to characterize genomic loci using FUMA.

(A) Circos plot depicting the colocalization of the eQTL CRB1, DENND1B, and PTPRC region (chromosome 1q31.1-q32.1) showing cis-eQTL interaction and chromatin interaction with other genes on chromosome 1. The outermost layer of the Circos plot is the Manhattan plot showing SNPs with p values <0.05, and the SNPs with colors show LD (r2) with the top independent SNP: red r2 > 0.8, orange r2 > 0.6, blue r2 > 0.3, and other SNPs are colored gray. (B) Regional association plot of Chromosome 1 (197,201,504−198,757,476) region showing association with LPC O-16:0 and CAD identified by the top variant (rs558096054) in the PTPRC by FUMA. The p-value was calculated using multivariate regression.

https://doi.org/10.1371/journal.pmed.1005039.g005

Interactive analysis of eQTL association and chromatin interaction mapping in FUMA, we detected strong evidence of cis-eQTL and chromatin interactions within the 11q12.2 region encompassing FADS1/2, MYRF, FADS3, DAGLA, with other loci like LRRC10B, APIP, PDHX, CXCR5, SLC37A4, and GRAMD1B, etc., previously reported to be associated with CAD and CV traits in the GWAS catalogue (Fig 6A and S15 Table). The FADS1 polymorphism (rs174544), strongly associated with PC 38:4 (C), TG 53:4, and FA 20:4 in this study, was also linked to CAD risk in the AIDHS, as shown by colocalization analysis in FUMA (Fig 6B). Interestingly, the individuals with high CAD PRS exhibited significantly lower levels of PC 38:4 (C). For example, the mean levels of PC 38:4 (C) in patients within the highest quartile of PRS were significantly lower (0.91 ± 0.053 µM) compared to those in the lowest quartile of PRS (1.09 ± 0.039 µM) (F = 14.22, two-tailed p </= 0.005). Regional plots show well-defined peaks for the metabolite and CAD association represented by a 3’UTR variant (rs174544) within the FADS1 region OR 1.36 (95% CI [1.18,1.51]; p </= 9.6 × 10⁻⁶) with a CADD score of 7.993 (Fig 6B). Many variants within this region also showed high CADD and RegulomeDB scores. This same FADS1 rs174544 variant has been linked with comorbidities such as Crohn’s disease, asthma, rheumatoid arthritis, and IBD, and has been shown to regulate clinical parameters such as TG, TC, LDL-C, Apo B, and Apo A1 levels (Table 2). The heritability of the top metabolites achieving GWAS significance is detailed in S16 Table. Notably, FA 18:0;(2OH) showed the highest heritability of 0.78. Both LPC O-16:0 and PC 38:4 (C) exhibited heritability of 0.16 and 0.13, respectively.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Colocalization and functional mapping to characterize genomic loci using FUMA.

(A) Circos plot depicting the colocalization of the eQTL FADS region (chromosome 11q12-q13.1) showing cis-eQTL interaction and chromatin interaction with other genes on chromosome 11. The outermost layer of the Circos plot is the Manhattan plot showing SNPs with p values <0.05, and the SNPs with colors show LD (r2) with the top independent SNP: red r2 > 0.8, orange r2 > 0.6, blue r2 > 0.3, and other SNPs are colored gray. (B) Regional association plot of the FADS region Chromosome 11 (61,752,636−61,788,518) shows evidence of association with PC 38:4 (C) and CAD identified by rs174544 by FUMA. (C) CADD and RegulomeDB score plots of the SNPs located under the FADS region. The p-value was calculated using multivariate regression.

https://doi.org/10.1371/journal.pmed.1005039.g006

The interactive visualization using GSEA validated established candidate genes and identified other prospective genes through the use of eQTL mapping and chromatin interaction mapping. Using GSEA, we identified several crucial pathways that were significantly over-represented in the mQTLs for the LPCs and specifically the ether-containing LPC O-16:0. Overrepresentation of SLIT-ROBO signaling, AMP signaling, leptin and adiponectin signaling, RORA and PGC1α signaling were the most relevant pathways due to genetic variation in ROBO2, DENN1B, LEPR, ENPP1, and CRB1 genes, connected with circulating LPC O-16:0 (Table 3). ROBO2 proteins bind with SLIT proteins in the cell signaling protein complex, which is involved in many diverse physiological and pathological functions, including axon guidance and angiogenesis [62]. The LEPR gene transcribes the leptin receptor protein, which regulates the leptin hormone responsible for controlling appetite and body fat [63,64]. Transcription factor RORA recruits PGC1α and histone acetylase p300 by binding to DNA response elements, activating transcription of the downstream genes [65]. Dysregulation of these pathways collectively contributes to metabolic diseases, immune system dysfunction, cancer, and inflammation, neural development, obesity, diabetes, and vascular dysfunction (Table 3).

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Over-represented pathways regulated by genes associated with metabolites.

https://doi.org/10.1371/journal.pmed.1005039.t003

Fine mapping analysis to calculate PIP and Z-scores

The fine mapping analysis of the mQTL regions associated with LPC O-16:0 yielded stronger PIP scores for the top mQTL (rs558096054; PTPRC), showing a PIP = 0.99; Z-score = −6.38 (S17a Table). This same SNP was associated with increased risk of CAD OR 3.35 (95% CI [2.62,4.07]; p </= 2.3 × 10⁻⁶) through colocalization analysis. Similarly, fine mapping of the FADS region showed a strong association with PC 38:4 (C) (regression) for a variant (rs174564; FADS2) with a PIP of 0.61; Z-score = −9.03. This SNP was in strong LD (R² = 0.98) with 3′UTR variant (rs174544) within the FADS1 region which showed increased CAD risk through colocalization analysis (S17b Table).

We also performed regression analysis excluding metabolite PCs from the model (Model 3) and compared results across adjusted and unadjusted models. Overall, in the combined serum and plasma, 29.5% of GWAS significant SNPs became less significant while 71.5% associations remained significant at the GWAS p value or higher. At the same time, the average change in the β values was <2 units between models 2 and 3. Average β varied significantly between model 1 and the other two models (2 and 3) because of the addition of clinical lipids in models 2 and 3 (S18 Table). We have presented the change of beta between models 2 and 3 in S6 Fig. Despite these differences, the (β) direction of associations remained consistent in all associated SNPs across all models. It was interesting to note that the major QTLs identified for LPC O-16:0, PC34:1, FA18:0(2OH), DG O-37:1|DG O-21:0_16:1, and PC38:4 remained statistically significant in all three models. However, the QTL association for TG 53:4 and CE20:4 in the FADS1/2 region did not achieve GWAS p-values in Model 3.