Assessment of multi-population polygenic risk scores for lipid traits in African Americans

Study population

We accessed an existing and previously described dataset (Crawford et al., 2015) of de-identified EHRs linked to biospecimens genotyped on the Illumina Metabochip (Buyske et al., 2012; Voight et al., 2012). These data are a subset of the larger BioVU, Vanderbilt University Medical Center’s (VUMC) biobank established in 2007 (Roden et al., 2008).

BioVU links discarded biospecimens collected for clinical purposes during outpatient visits to a de-identified version of the patient’s EHR. The de-identification process includes the scrubbing of identifiers, data shifting, and linkage to biospecimens via a one-way hash that combined make re-identification difficult and unlikely (Roden et al., 2008). BioVU as a VUMC program has received approval from the Vanderbilt Institutional Review Board (IRB) and was reviewed in detail by the federal Office for Human Research Protections (OHRP), who agreed with the non-human subjects regulatory designation for both the resource and subsequent research (Roden et al., 2008).

The data accessed for the current study were collected under an opt-out participation model described to the patient annually as part of the consent to treat process and represent DNA samples extracted from blood collected between 2007 and 2011 and clinical data collected through 2011. As a study site of the PAGE Study, a consortium focused on genetic association studies in diverse populations (Matise et al., 2011), we selected at the time all DNA samples in BioVU from non-European descent patients for Illumina Metabochip genotyping (n = 15,863), the majority of which are from African American or Black patients (73%; Crawford et al., 2015). As we have previously described (Crawford et al., 2015), BioVU is representative of the Nashville, Tennessee VUMC patient population, a population that is on average older and of European descent compared with the surrounding Davidson County population. The dataset described here is de-identified and is considered to be non-human subjects research (Vanderbilt University IRB #110579) (Pulley et al., 2010).

Electronic phenotyping

Analyses were limited to African American adults aged 18 years or older. All demographic and clinical data were extracted using a combination of structured (e.g., International Classification of Diseases 9th edition, Clinical Modification (ICD-9-CM) codes, Current Procedural Terminology (CPT) codes, problems lists, and laboratory values) and unstructured (clinical notes) data available in the EHR.  Race/ethnicity in BioVU was administratively assigned, and these assignments are highly correlated with genetic ancestry for some (e.g., European and African Americans) but not all (e.g., Hispanics and East Asians) groups (Dumitrescu et al., 2010; Hall et al., 2014; Farber-Eger et al., 2017). Age was calculated as the age of the patient in 2011 based on the patient’s year of birth. First mention (e.g., the first clinic visit associated with a lipid lab result) of each patient’s serum HDL-C, LDL-C, TG, and TC was extracted and recorded as either a “pre-medication” or “post-medication” value. Pre-medication values were defined as free of evidence of concurrent lipid lowering medication usage. As previously described (Crawford et al., 2015), lipid lowering medications included statins (also known as HMG CoA reductase inhibitors, atorvastatin (Lipitor^®), fluvastatin (Lescol^®), lovastatin (Mevacor^®, Altoprev™), pravastatin (Pravachol^®), rosuvastatin calcium (Crestor^®), simvastatin (Zocor^®), lovastatin + niacin (Advicor^®), atorvastatin + amlodipine (Caduet^®), and simvastatin + ezetimibe (Vytorin™)), selective cholesterol absorption inhibitors (ezetimibe (Zetia^®)), resins (cholestyramine (Questran^®, Questran^® Light, Prevalite^®, Locholest^®, Locholest^® Light), colestipol (Colestid^®), colesevelam Hcl (WelChol^®)), fibrates (gemfibrozil (Lopid^®), fenofibrate (Antara^®, Lofibra^®, Tricor^®, and Triglide™), clofibrate (Atromid-S)), and niacin. Fasting status is unknown. Heights and weights closest to the first mention of the lipid trait were extracted to calculate body mass index (BMI; kg/m²). Prevalent cardiovascular disease was defined as mention of heart attack keywords in the problems list (“MI”, “myocardial infarction”, or “heart attack”), CPT codes (Crawford et al., 2018) for coronary artery bypass graft (33510–33514; 33515; 33516; 33517–33519; 33520; 33521–33523; and 33534–33536) and angioplasty and/or coronary stents (92980–92981; 92982; 92984; 92995; and 92996), or three mentions of International Classification of Diseases (ICD)-9-CM codes (410 or 410.*) (Dumitrescu et al., 2015a) for myocardial infarction. Type 2 diabetes cases status was defined using a combination of ICD codes, labs, and medications from the previously described and validated electronic Medical Records & Genomics (eMERGE) Network algorithm (Kho et al., 2012).

Genotyping and quality control

The Illumina Metabochip was genotyped by Vanderbilt Technologies for Advanced Genomics (VANTAGE), formerly the Vanderbilt University Center for Human Genetics Research DNA Resources Core, for the Epidemiologic Architecture for Genes Linked to Environment (EAGLE) study, a study site of PAGE I (Matise et al., 2011). Genotyping was conducted following the manufacturer’s protocol (Illumina, Inc.; San Diego, CA.). Per PAGE I protocol, 360 HapMap samples were also genotyped for quality control (Crawford et al., 2012). Genotype calling and standard quality control measures (call rates, duplicate checks, Hardy-Weinberg Equilibrium, etc.) have been described by Buyske et al. 2011 and more recently by Kaur et al. (2021). Metabochip genotypes linked to the EAGLE BioVU de-identified clinical data are available through the database of Genotypes and Phenotypes (dbGaP) accession number phs002767.v1.p1.

Polygenic risk score calculations

Unweighted and weighted multi-population PRS for HDL-C, LDL-C, TG, and TC were calculated for each patient based on published summary statistics for diverse populations from the PAGE Study (Hu et al., 2020). We first extracted Metabochip genotypes available for SNPs previously associated HDL-C (44), LDL-C (36), TG (51), TC (48) reported by the PAGE Study at genome-wide significance (p ≤ 5  ×  10⁻⁸) (Table S7 from Hu et al., 2020; Supplementary Table). After removing SNPs out of Hardy Weinberg Equilibrium at p < 10⁻⁴, we calculated unweighted PRS for each patient and each lipid trait by counting the number of risk alleles (N) for 42, 34, 50, and 46 HDL-C, LDL-C, TG, and TC-associated SNPs (k), respectively. PRS were weighted (PRS_w) using the absolute value of the betas (|β|) for the corresponding associations published by the PAGE Study (Hu et al., 2020).

${\displaystyle {\text{PRS}}_{\text{w}}=\sum _{i=1}^k\left|{\beta }_i\right|N_i.}$

Statistical methods

Principal components (PCs) were generated as previously described (Buyske et al., 2012; Kaur et al., 2021). We tested each SNP for an association with the lipid trait previously associated as described in Hu et al. (2020). Single SNP tests of association were performed in PLINK version 1.90 using linear regression, unadjusted and adjusted for sex, age, BMI, and the first ten PCs. TG levels were transformed (natural log) prior to tests of association. Summary statistics were visualized using Synthesis View (Pendergrass et al., 2010b; Pendergrass et al., 2010a; Supplementary File). We also tested for associations between the lipid traits and their respective unweighted and weighted PRS using linear regression, unadjusted and adjusted for age, sex, BMI, and the first ten PCs. Tests of association between CVD and T2D and unweighted and weighted PRS were performed using 2x2 tables (unadjusted) and logistic regression (CVD and LDL-C weighted PRS adjusted by LDL-C). A limited phenome-wide association study (PheWAS) was performed for unweighted and weighted PRS and platelet counts, glucose, serum creatinine, serum albumin, serum albumin creatinine ratio, blood urea nitrogen (BUN), uric acid, urine creatinine, urine albumin, and estimated glomerular filtration rate (eGFR). eGFR was calculated using EHR-extracted serum creatinine, age closest to serum creatinine, sex, and race and the CKD-EPI equation (Levey et al., 2009). eGFR was then categorized as stage 1 (≥90 mL/min per 1.73 m²), stage 2 (60–89 mL/min per 1.73 m²), and stage 3 (<60 mL/min per 1.73 m²) for the limited PheWAS. When necessary, laboratory values were transformed as the natural log of the quantitative value plus one per the PAGE I study PheWAS protocol (Pendergrass et al., 2011). All PheWAS tests of association were performed using linear regression except for eGFR where multinomial logistic regression was used to test for association using stage 1 as the referent. All statistical analyses were performed using R version 4.1.0 (R Core Team, 2021) and PLINK 1.90.

Study population

A total of 11,521 African American patients were genotyped on the Illumina Metabochip as part of EAGLE BioVU (Crawford et al., 2015). Among these patients, 3,254 were adults with height, weight, and Metabochip genotypes (Table 1). Many (61.16%) were female with an average age of 46.95 years (±15.32 years standard deviation) and an average BMI (31.68 kg/m²; ±8.52 kg/m² standard deviation) considered obese. Average pre-medication and post-medication lipid lab values were within the range expected for a general US adult African American population (Cohen et al., 2010; Carroll et al., 2012). Each patient had at least one risk allele for each lipid trait (Fig. S1). On average, African American adult patients had a higher TG PRS (5.95) compared with TC (3.52), HDL-C (3.39) and LDL-C (2.01) (Table 1).

Lipid trait genetic associations

We first performed unadjusted (Fig. S2) and adjusted single SNP tests of association for each lipid trait, stratified by lipid lowering medication exposure as described in Materials & Methods. Among African Americans with pre-medication lipid labs, four out of 42 HDL-C, three out of 34 LDL-C, five out of 50 TG, and eight out of 46 TC SNPs were associated with their respective lipid trait at p < 0.05 adjusted for sex, age, BMI, and PCs (Figs. 1–4). Compared with the literature and accounting for coded alleles (Hu et al., 2020), all associations identified here at p < 0.05 for LDL-C were in the expected directions of effect. Most of the associations identified here for HDL-C were in the opposite direction compared with Hu et al. (2020), with only rs1800775 associated in the expected direction. For transformed triglycerides (rs1077834) and total cholesterol (rs8106922) each, all but one of the associated SNPs were in the expected direction of effect.

Among African Americans with post-medication lipid labs, two out of 42 HDL-C, four out of 34 LDL-C, and one out of 49 TG SNPs were associated with their respective lipid trait at p < 0.05 adjusted for sex, age, BMI, and PCs (Fig. S3). None of the SNPs tested were associated with post-medication TC at p < 0.05. Overall, fewer associations at p < 0.05 were identified among post-medication lipid labs (eight) compared with pre-medication lipid labs (20), most likely due to differences in sample size (Table 1). Despite the smaller sample sizes, three associations were identified among African Americans with post-medication labs but not among African Americans with pre-medication labs: rs1864163 (HDL-C), rs688 (LDL-C), and rs4407894 (transformed TG) (Figs. 1–4 versus Fig. S3). Of these three associations, rs1864163 was associated in the opposite direction compared with Hu et al. (2020). Among the five overlapping pre- and post-medication associations, all were consistent in direction of effects with the exception of rs1800775, which was associated with decreased HDL-C levels among African Americans with pre-medication labs but with increased levels among African Americans with post-medication labs.

Lipid PRS, CVD, and T2D

We calculated unweighted and weighted PRS for each lipid trait using genomic discovery data specific for diverse populations as described in Materials & Methods. Among unadjusted and adjusted tests of association, none were associated with their respective lipid traits at p < 0.05 (Table 2). All PRS tested here, regardless of significance, were associated with decreased lipid lab values.

To explore the potential clinical utility of lipid PRS, we tested each lipid weighted PRS for association with CVD and T2DM (Table 3). In this African American patient population, 29.5% had one or more CVD event(s) recorded in their EHR (MI (1.7%), heart attack (29.4%). CABG (1.0%), CSA (1.5%)). Similarly, almost one-third of African American adult patients in the present study had T2D (28.4%). We did not detect significant associations between any of the lipid weighted PRS and T2D at p < 0.05 (Table 3). For LDL-C PRS, CVD cases had a higher proportion of extreme LDL-C weighted PRS (10.9%) compared with controls (9.6%) (OR = 1.15, 95% Cl [1.01–1.32]; p = 0.04). The association was no longer nominally significant when adjusted for LDL-C levels (p > 0.05).

Limited PheWAS

To identify possible pleiotropic relationships between genetic determinants of lipid traits and other commonly ordered clinical labs, we performed a limited PheWAS for each lipid PRS and EHR-extracted labs related to biochemistry, liver, and kidney functions. Nominal associations (p < 0.05) were identified for HDL-C PRS and albumin creatinine ratio; serum creatinine and LDL-C and TC PRS; and stage 2 eGFR and TC PRS (Table 4). However, after adjusting significance thresholds for multiple testing (four lipid trait PRS ×10 labs), none of the limited PheWAS associations remained significant.