Skip to main content
Download PDF

Clinical laboratory test-wide association scan of polygenic scores identifies biomarkers of complex disease

Genome Medicine volume 13, Article number: 6 (2021) Cite this article

Abstract

Background

Clinical laboratory (lab) tests are used in clinical practice to diagnose, treat, and monitor disease conditions. Test results are stored in electronic health records (EHRs), and a growing number of EHRs are linked to patient DNA, offering unprecedented opportunities to query relationships between genetic risk for complex disease and quantitative physiological measurements collected on large populations.

Methods

A total of 3075 quantitative lab tests were extracted from Vanderbilt University Medical Center’s (VUMC) EHR system and cleaned for population-level analysis according to our QualityLab protocol. Lab values extracted from BioVU were compared with previous population studies using heritability and genetic correlation analyses. We then tested the hypothesis that polygenic risk scores for biomarkers and complex disease are associated with biomarkers of disease extracted from the EHR. In a proof of concept analyses, we focused on lipids and coronary artery disease (CAD). We cleaned lab traits extracted from the EHR performed lab-wide association scans (LabWAS) of the lipids and CAD polygenic risk scores across 315 heritable lab tests then replicated the pipeline and analyses in the Massachusetts General Brigham Biobank.

Results

Heritability estimates of lipid values (after cleaning with QualityLab) were comparable to previous reports and polygenic scores for lipids were strongly associated with their referent lipid in a LabWAS. LabWAS of the polygenic score for CAD recapitulated canonical heart disease biomarker profiles including decreased HDL, increased pre-medication LDL, triglycerides, blood glucose, and glycated hemoglobin (HgbA1C) in European and African descent populations. Notably, many of these associations remained even after adjusting for the presence of cardiovascular disease and were replicated in the MGBB.

Conclusions

Polygenic risk scores can be used to identify biomarkers of complex disease in large-scale EHR-based genomic analyses, providing new avenues for discovery of novel biomarkers and deeper understanding of disease trajectories in pre-symptomatic individuals. We present two methods and associated software, QualityLab and LabWAS, to clean and analyze EHR labs at scale and perform a Lab-Wide Association Scan.

Background

The overarching goal of this study was to determine whether laboratory (lab) test results collected in a hospital and outpatient setting could be mined against polygenic scores (PGS) to identify known and novel biomarker associations for complex disease. Lab test results are essential to routine clinical care. These targeted biochemical measurements facilitate disease diagnosis and influence health care delivery. Clinical lab values are also monitored as mediators of disease risk and are targeted by interventions to reduce disease incidence (e.g., cholesterol-lowering medication to reduce the risk of heart disease). Lab test results in the electronic health record (EHR) are a vast and growing resource for novel biomarker discovery, especially as EHRs are increasingly linked to patient DNA samples (e.g., the eMERGE consortium (https://emerge.mc.vanderbilt.edu)), the All of Us Program (https://allofus.nih.gov), and the Million Veteran’s Program (https://www.research.va.gov/mvp/)). Genetic studies of EHR-based labs could reveal novel biomarker-disease or biomarker-gene associations, which in turn could lead to better understanding of biological processes in disease, improved diagnostic algorithms, and new therapeutic targets.

Despite their potential, however, EHR-based labs have been used in only a handful of prior genetic studies [1,2,3,4,5], and none have systematically interrogated an extended collection of EHR-based lab values. Barriers to studying EHR-based labs include uneven data quality, and challenges inherent to analyzing and interpreting high-dimensional health care data. Data entry errors exist, resulting in implausible recorded values [6], some labs have different units and reference ranges over time, and many individuals have multiple observations of different lab tests, each measured at varying times relative to diagnoses and treatment [7]. Moreover, previous studies demonstrate that while 99% of lab results are accurately transmitted from the testing laboratory to the EHR, only 70% of test results contain all required reporting elements, and only 91% of results are appropriately formatted [8]. Thus, while these data represent real clinical care and may accelerate translational research, there is little precedent for their analysis and interpretation in genetic studies.

To address these challenges, we present a high-throughput framework for genetic analysis of EHR-derived lab data. We have developed two methods: the QualityLab pipeline to clean, standardize, and visualize lab data and the Lab-Wide Association Scan (LabWAS) pipeline to scan for associations between any variable of interest (genetic or otherwise) and the cleaned EHR labs. The LabWAS method is similar to the Phenome-Wide Association Scan (PheWAS) which scans for association between an exposure variable (typically, a genetic risk factor) and many phenotypes [9]. The PheWAS method has replicated many known gene-disease associations [10] and has identified novel pleiotropic genetic effects [11], opportunities for drug repurposing, and unintended drug consequences [12]. QualityLab builds off the success of previous measurement quality control methods, such as CLARITE [13]. While, CLARITE focuses on minimal cleaning of survey data, QualityLab conducts extensive cleaning of quantitative lab measurements derived from EHRs.

We hypothesized that EHR-based lab values could be used to identify known and novel relationships between genetics, biomarkers, and disease. We deployed our framework in the Vanderbilt University Medical Center (VUMC) EHR and linked biobank, BioVU, and replicated it in an independent biobank, Massachusetts General Brigham Biobank. We focused on genetic analysis of blood values of high-density lipoprotein cholesterol (HDL), low-density lipoprotein cholesterol (LDL), and triglycerides (TG) and on coronary artery disease (CAD) as proof-of-principle examples to test the association between PGS for CAD and known biomarkers of disease (LDL, HDL, and TG) using the QualityLab and LabWAS methods across populations. We show that EHR-derived lipids values are genetically similar to those in population-based studies and that PGS for lipids robustly associate with their respective lab in a LabWAS. Additionally, our LabWAS revealed that PGS for CAD associated with known lipid biomarkers, even in individuals without a history of CAD, and with potentially novel immune biomarkers.

Methods

Study sample

Our primary analysis was performed at VUMC which is a tertiary care center providing inpatient and outpatient care in Nashville, TN. The VUMC EHR was established in 1994 and includes data on billing codes from the International Classification of Diseases, 9th and 10th editions (ICD-9 and ICD-10), Current Procedural Terminology (CPT) codes, laboratory values, reports, and clinical documentation. The de-identified mirror of the EHR, known as the Synthetic Derivative, includes patient records on more than 2.8 million individuals. In 2007, VUMC launched a biobank, BioVU, and the BioVU Consent form is provided to patients in the outpatient clinic environments at VUMC. The form states policies on data sharing and privacy and, upon consent, makes any blood leftover from clinical care eligible for BioVU banking [14]. The VUMC Institutional Review Board oversees BioVU and approved this project.

Genotyping and quality control

We obtained genotype information on 94,474 BioVU individuals of different ancestral and racial backgrounds genotyped on the Illumina MEGAEX array. Using PLINK v1.9 [15], genotypes were filtered for SNP and individual call rates, sex discrepancies, and excessive heterozygosity (Additional file 1). We selected individuals of European or African ancestry using principal component analysis implemented in Eigenstrat [16, 17] and confirmed the absence of genotyping batch effects through logistic regression with “batch” as the phenotype. Imputation was completed using the Michigan Imputation Server [18] using the Haplotype Reference Consortium (HRC) reference panel. SNPs were then filtered for SNP imputation quality (R2 > 0.3) and converted to hard calls. We restricted to autosomal SNPs, filtered SNPs with minor allele frequency > 0.01, or with allele frequencies that differed by more than 10% from the 1000 Genomes Project phase 3 CEU or ASW set respectively [19], and Hardy-Weinberg Equilibrium (p > 1 × 10−10). The resulting dataset contained 6,303,629 SNPs on 72,824 individuals of European genetic ancestry and 12,798,111 SNPs on 15,283 individuals of African genetic ancestry.

QualityLab pipeline

In parallel with the BioVU genotyping project, we extracted data on all lab tests collected in the routine clinical care of 1,521,125 VUMC patients, amounting to 275,991,157 observations across 11,061 lab tests (Fig. 1a). Of these lab tests, 5028 were reported in non-numeric values and 1618 had only been administered to one patient, leaving 4415 quantitative lab tests for further cleaning. Some lab tests had observations recorded in different units (e.g., Selenium reported in both mcg/L and μg/L); thus, we restricted to lab tests for which at least 70% of the observations were measured in the same unit and required that each lab have at least 100 patients and at least 1000 numeric observations, for a total of 939 labs retained for further analysis.

Fig. 1
figure 1

Selection of BioVU patients and datasets for different analyses presented in this manuscript. a BioVU patients were selected in parallel for clinical laboratory (lab) test cleaning and for genotyping. b Lab-specific quality control filters and subsetting were applied to the 939 lab tests in the 94,474 patients with clean lab data. Parallelograms denote input and output datasets. Options highlighted in green were selected for the proof-of-principle analyses of blood-based lipid lab values

Full size image

For each of these 939 labs, we applied lab-specific quality control filters (Fig. 1b). First, we filtered infinite and non-numeric values, as well as observations outside of 4 standard deviations from the overall sample mean, indicative of biologically implausible values due to technical or recording errors, monogenic disorders, or extreme environmental influence. We calculated the median lab value for each patient and extracted the patient’s age at median lab value. For patients in whom we had to calculate the median lab value (e.g., those with an even number of observations), we defined the age at median lab value as the mid-point of the patient’s ages at the two lab values used to calculate the median lab value.

The analyses presented in this manuscript use the QualityLab dataset constructed from pediatric and adult observations, in both sexes, in patients of all races (Fig. 1b). In downstream genetic analyses, however, we restrict to participants of European or African genetic ancestry and match the ancestry of the participants in the discovery GWAS used for the training the PGS.

The QualityLab pipeline also provides user with the option to stratify data (Fig. 1b), by age at observation, sex, and EHR-recorded race, for a total of 72 different data subsets. The QualityLab pipeline generates summary statistics and plots for each strata (e.g., mean, maximum, and minimum of the median lab value; Additional file 2: Table S1; Additional file 3: Fig. S1), and returns two versions of the data for downstream analyses. The first is a table of median lab values and age at median lab value for each individual. The second is an inverse normal quantile transformation (INT) of the median lab value data, to account for skewness and non-normality [20, 21]. Importantly, the choice of quality control thresholds is completely in the control of the user. The choices made here reflect the goals of this study which focus on the central tendencies of large populations. However, the outlier thresholds and normalization methods employed here would not be appropriate in a study of rare, potentially pathogenic, variation where large genetic effects and extreme phenotypes may be expected.

Lab heritability and GWAS analyses

Prior to calculating SNP-based heritability (h2SNP), we first calculated pairwise relatedness in the BioVU genotyped sample and removed one related individual from pairs with pi-hat greater than 0.05. This stringent threshold was chosen based on prior experience and previously published best practices in the application of restricted maximum likelihood (REML) approaches to the calculation of h2SNP [22]. After filtering, 45,010 individuals of European genetic ancestry (Fig. 1a) remained. We then used the genome-wide complex trait analysis (GCTA) package (version 1.92.4) [23] to create a pairwise genetic relationship matrix for all individuals, and heritabilities were calculated using REML methods. We used the median, INT-transformed lab values from the QualityLab pipeline, and of the 481 analyzed labs, 335 demonstrated non-zero heritability. For GWAS analyses, we used a less stringent relatedness filter appropriate to GWAS (pi-hat > 0.2) [24] resulting in a total available sample of 66,732 European descent individuals. Next, we subset to the heritable labs with at least 1000 individuals (n = 181) and performed GWAS of the median, INT-transformed lab values using fastGWA [25] (Fig. 1a). All h2SNP and GWAS analyses included covariates for sex, cubic splines (knots = 4) of median age across the medical record (to control for non-linear effects of age), and the top 10 principal components of ancestry.

Heritability and GWAS analyses of lipids

We benchmarked our lipid h2SNP estimates against those from two external datasets, the Global Lipids Genetics Consortium (GLGC) [26] and the Million Veterans Program (MVP). GLGC and MVP estimates of h2SNP for HDL, LDL, and TG were calculated from GWAS summary statistics using LDSC [27]. We computed h2SNP in BioVU using Linkage Disequilibrium Score regression (LDSC) applied to our fastGWA summary statistics for HDL, LDL, and TG (Additional file 3: Fig. S2). However, because LDSC can underestimate h2SNP [28], we also calculated h2SNP using GCTA. In addition to these h2SNP comparisons, we calculated the genetic correlations (rg) between the BioVU lipid GWASs and the GLGC and MVP lipid GWASs using LDSC and the pre-computed European LD scores from 1000 Genomes Phase 3 European data [29]. We also calculated genetic correlations using a new method, high-definition likelihood [30], which fully accounts for linkage disequilibrium across the genome and is more suitable for traits with lower heritability than LDSC. In sensitivity analyses, we repeated genetic correlations of LDL after controlling the BioVU GWASs for coronary atherosclerosis or diabetes diagnoses, defined as phecodes 411, “Ischemic heart disease,” and 249, “Secondary diabetes mellitus” (Additional file 1).

To validate EHR-based lipid values, we tested the robustness of HDL, LDL, and TG h2SNP estimates to different lab value and patient filters. First, we excluded lipid measurements that occurred after the first mention of lipid-altering mediation in the EHR (Additional file 1) and re-calculated each patient’s pre-medication median values of HDL, LDL, and TG. Second, we excluded patients with a diagnosis of CAD, defined by the phecode 411 (Additional file 1).

LabWAS pipeline

LabWAS uses the median, INT-transformed lab values from the QualityLab pipeline in a linear regression to determine the association with an input variable, adjusting for covariates. In these analyses, a primary goal of the LabWAS was to test common population genetic variation (e.g., PGS) for association with common population variation in lab values. We therefore only included the 335 labs with non-zero h2SNP. Additionally, we imposed a minimum sample size requirement of 100 for a lab to be included in the LabWAS analysis, bringing the number of labs tested in each scan to 315 in the European ancestry set and 226 in the African ancestry set.

Polygenic scoring

Prior to polygenic scoring, we randomly removed one related individual from pairs with pi-hat greater than 0.2, leaving 66,732 individuals of European genetic ancestry and 12,383 individuals of African genetic ancestry. (Fig. 1a). We generated lipids PGS for these individuals using PRS-CS [31] with weights derived from the transethnic MVP lipid GWAS summary statistics [4]. PGS for CAD (CADPGS) was calculated using SNP weights from CARDIoGRAMplusC4D GWAS summary statistics [32] using PRS-CS. Because the majority of the MVP transethnic sample was European, linkage disequilibrium was modeled using the pre-calculated European panel. PRS-CS is a recently developed Bayesian polygenic prediction method that imposes continuous shrinkage priors on SNP effect sizes (Polygenic Risk Score – Continuous Shrinkage) [31]. These priors can be represented as global-local scale mixtures of normals which allow the model to flexibly adapt to differing genetic architectures and provide substantial computational advantages. The shrinkage parameter was automatically learnt from the data (i.e., using PRS-CS-auto). SNP effect estimates were obtained from GWAS summary statistics and the score was calculated using a linkage disequilibrium reference panel from 503 European samples in the 1000 Genomes Project phase 3 [19]. Although PRS-CS outperformed other polygenic scoring methods across a range of traits in previous experiments, its superiority may not hold across all genetic architectures [31]. We therefore also generated PGS for the European sample using PRSice-2 [33] (Additional file 1) and have automated a pipeline to generate scores across both methods. PGS were scaled to have a mean of zero and SD of one before testing for association with any outcome variables. We validated each score by testing the proportion of trait variability explained by the PGS, controlling for sex, cubic splines of median age (4 knots) across the medical record, and the top 10 principal components to adjust for genetic ancestry (Additional file 3: Fig. S3).

LabWAS of polygenic scores

PGS for LDL (PGSLDL), HDL (PGSHDL), and TG (PGSTG) were calculated in BioVU participants using PRS-CS and applying SNP weights from the MVP GWAS summary statistics. We then ran LabWAS of PGSLDL, PGSHDL, and PGSTG to test whether lipid labs were robustly associated with the genetic scores to which they corresponded. Next, a PGS for CAD (CADPGS) was calculated using SNP weights from CARDIoGRAMplusC4D GWAS summary statistics [32] and a LabWAS of PGSCAD to test whether the score could identify lab traits associated with genetic risk for CAD, before and after controlling for a CAD diagnosis (Additional file 1). Each LabWAS was controlled for sex, cubic splines of median age across the medical record, and the top 10 principal components of ancestry. Results are reported as effect estimates and their 95% confidence intervals per SD increase in the PGS. The Bonferroni-corrected threshold for statistical significance across all tested labs was 3.97× 10−5 (0.05/(315 × 4)).

Replication in Massachusetts General Brigham Biobank

We next sought to replicate the associations between lipids PGS and referent lipids as well as the significant associations with CADPGS in an external biobank. The MGBB, previously the Partners Biobank, is an ongoing virtual cohort study of patients across the Partners HealthCare hospital system (including Brigham and Women’s Hospital, Massachusetts General Hospital, and other affiliated hospitals), which provides a large-scale resource of linked longitudinal electronic health records (EHR) data, genomic data, and self-reported survey data [34]. All patients provided informed consent before enrollment, and all study procedures were approved by the Partners HealthCare Institutional Review Board.

Lab values were extracted from EHRs and cleaned using QualityLab, resulting in 759 labs for analysis. The median value for each lab trait for each individual was selected and inverse normalized. Lab heritabilities were calculated using REML in GCTA. Of 759 labs that passed QualityLab, 241 demonstrated measurable heritability and included a sample size of at least 100 individuals.

Polygenic scores for HDL, LDL, TG, and CAD were calculated on individuals of European descent in MGBB (n = 25,698) using the same criteria as BioVU. Lipids and CAD polygenic scores were associated with each of 234 labs using LabWAS. Lastly, the associations between CADPGS and lab traits were controlled for CAD diagnosis, defined by phecode 411 (N cases = 1094, N controls = 20,405). All associations were controlled for sex, top 10 principal components, and the first two splines of median age across the medical record.

Results

QualityLab pipeline

A total of 94,474 BioVU patients with clean lab data, of whom 66,732 were also of European genetic ancestry were included in the PGS LabWAS analyses (Fig. 1a). These 66,732 patients had data on 939 labs, containing 30,421,498 observations. The median number of unique lab tests per patient was 44, and the median number of lab observations per patient was 201. Slightly more than half of the BioVU patients in the sample were female (55.6%), and the average median age across the EHR was 52.0 years. These BioVU participants included 10,015 CAD cases and 49,702 CAD controls. In the African ancestry sample, 12,383 patients had data on 925 labs, containing 5,367,062 observations. More than half the patients were female (61.6%) and the average median age was 38.5 years. The median number of unique lab tests per patient was 41, and the median number of lab observations per patient was 150 (Additional file 1; Additional file 2: Table S3). Distributions of lipids levels by genetic ancestry are shown in Additional file 3: Fig. S4.

Heritability and GWAS analyses

Out of 939 clean lab traits, 335 demonstrated non-zero h2SNP and the point estimates ranged from 2 × 10−6 to 0.98. (Additional file 2: Table S4, Additional file 3: Fig. S5). As a resource for the community, the GWAS summary statistics for the labs with calculable heritability and a minimum sample size of 1000 individuals (n = 181) are available in the GWAS Catalog (Study Number: GCP000091; accession numbers GCST90012603 - GCST90012784; accession numbers are listed in Additional file 2: Table S22).

Heritability and GWAS analyses of lipids

The h2SNP estimates in BioVU were robust to removing post-medication observations, and to removing CAD cases. The number of participants included in these analyses, however, was smaller, and so the standard errors of these h2SNP estimates were larger (Fig. 2a; Additional file 2: Table S5). Both GCTA and LDSC gave similar estimates of h2SNP in BioVU (Fig. 2b), and the LDSC estimates in BioVU were comparable to those in the GLGC and MVP for all lipids.

Fig. 2
figure 2

Heritability and GWAS analyses of lipids. a Estimates of heritability computed by GCTA in BioVU patients were robust to excluding individuals with a diagnosis of CAD and to removing post-medication observations. b Estimates of heritability computed using GWAS summary statistics and LDSC were comparable across BioVU and the Global Lipids Genetic Consortium (GLGC) and Million Veteran’s Project (MVP) samples. c Genetic correlations between lipid levels in BioVU and the Global Lipids Genetic Consortium (GLGC) or Million Veteran’s Program (MVP) calculated using LDSC or high-definition likelihood (HDL). Stars denote statistically significant correlations

Full size image

Genetic correlation between BioVU and GLGC summary statistics was strong for HDL (LDSC: rg = 0.96, SE = 0.08, p value = 2.69 × 10−35, high-definition likelihood: rg = 0.92, SE = 0.11, p value = 3.25 × 10−17) and TG (LDSC: rg = 0.94, SE = 0.05, p value = 5.86 × 10−97, high-definition likelihood: rg = 0.89, SE = 0.11, p value = 7.69 × 10−17). When comparing BioVU and MVP, the correlations for HDL (LDSC: rg = 0.99, p value = 7.51 × 10−61, high-definition likelihood: rg = 0.89, SE = 0.08, p value = 2.24 × 10−27) and TG (LDSC: rg = 0.94, p value = 2.28 × 10−99, high-definition likelihood: rg = 0.88, SE = 0.10, p value = 4.84 × 10−18) were nearly perfect. The LDL and LDL pre-medication genetic correlations between GLGC and BioVU were not calculable using LDSC due to low heritability. Using high-definition likelihood, GLGC LDL levels were significantly correlated when median LDL values across the entire EHR (rg = 0.44, SE = 0.10, p value = 1.08 × 10−5) and median pre-medication LDL values (rg = 0.50, SE = 0.08, p value = 6.38 × 10−10). The comparison between BioVU and MVP showed a stronger correlation for LDL (LDSC: rg = 0.84, SE = 0.17, p value = 1.47 × 10−6; high-definition likelihood: rg = 0.53, SE = 0.09, p value = 1.52 × 10−11). The genetic correlation with MVP increased when we restricted to pre-medication values of LDL in BioVU (LDSC: rg = 0.89, SE = 0.22, p value = 2.90 × 10−5; high-definition likelihood: rg = 0.56, SE = 0.07, p value = 2.06 × 10−15) (Fig. 2c) and increased further when we controlled for coronary atherosclerosis and diabetes diagnoses (GLGC, high-definition likelihood: rg = 0.57, SE = 0.09, p value = 8.88 × 10−9, MVP, LDSC: rg = 1.00, SE = 0.34, p value = 0.004) (MVP, high-definition likelihood: rg = 0.55, SE = 0.09, p value = 1.50 × 10−8) (Additional file 3: Fig. S6).

LabWAS of polygenic scores for lipids

A LabWAS of HDLPGS in the European sample was associated with levels of several metabolic markers (Fig. 3a, Additional file 2: Table S6), including increased HDL (p value< 2.23 × 10−308, beta = 0.31), decreased TG (p value = 2.06 × 10−171, beta = − 0.16), decreased total cholesterol to HDL ratio (p value = 2.54 × 10−44, beta = − 0.22), increased total blood cholesterol (p value = 2.51 × 10−37, beta = 0.07), and decreased blood glucose (p value = 4.62 × 10−32, beta = − 0.04), decreased blood urea nitrogen (p value = 1.48 × 10−15, beta = − 0.03), decreased glycated hemoglobin (p value = 1.52 × 10−12, beta = − 0.05), decreased bedside glucose (p value = 1.03 × 10−11, beta = − 0.07), and decreased whole blood glucose (p value = 2.49 × 10−5, beta = − 0.03). HDLPGS was also associated with four immune labs, white blood cell count (p value = 6.14 × 10−13, beta = − 0.03), absolute neutrophil count (p value = 5.69 × 10−7, beta = − 0.03), immature granulocytes (p value = 7.86 × 10−6, beta = − 0.02), and monocyte to leukocyte ratio (p value = 9.13 × 10−6, beta = 0.02). Five blood biomarkers associated with HDLPGS, mean corpuscular volume (p value = 3.48 × 10−17, beta = 0.03), blood carbon dioxide (p value = 6.69 × 10−11, beta = 0.02), mean corpuscular hemoglobin (p value = 9.53 × 10−10, beta = 0.02), international normalized ratio (p value = 1.31 × 10−6, beta = − 0.03), and red blood cell distribution width (p value = 2.21 × 10−5, beta = − 0.02). Finally, three other labs associated with HDLPGS, urate (p value = 1.13 × 10−11, beta = − 0.07), creatinine (p value = 1.42 × 10−10, beta = − 0.02), and urine pH (p value = 2.22 × 10−8, beta = 0.02). In the African ancestry group, HDLPGS significantly associated with increased HDL (p value = 1.38 × 10−74, beta = 0.23), decreased triglycerides (p value = 6.72 × 10−10, beta = − 0.08), and increased total cholesterol (p value = 4.81 × 10−9, beta = 0.08) (Fig. 3b, Additional file 2: Table S7).

Fig. 3
figure 3

LabWAS of PGSHDL in a individuals of European ancestry (EA) and b individuals of African ancestry (AA), LabWAS of PGSLDL in c EA and d AA, and LabWAS of PGSTG in e EA and f AA. The red line indicates the Bonferroni threshold for statistical significance and the blue line indicates a p value of 0.05. Upward triangles indicate that the PGS is associated with increased levels of the lab, while downward triangles indicate an association with reduced levels of the lab

Full size image

The LabWAS of LDLPGS showed associations with four lipid labs (Fig. 3c, Additional file 2: Table S8). The most significant association was increased calculated LDL (p value < 2.23 × 10−308, beta = 0.24), followed by increased total blood cholesterol (p value = 1.30 × 10−282, beta = 0.20), increased directly measured LDL (p value = 3.79 × 10−44, beta = 0.19), increased non-HDL cholesterol (p value = 1.78 × 10−31, beta = 0.19), increased total cholesterol to HDL ratio (p value = 5.27 × 10−17, beta = 0.13), and increased triglycerides (p value = 4.47 × 10−6, beta = 0.03). LDLPGS also associated with four blood biomarkers, mean corpuscular hemoglobin (p value = 5.68 × 10−8, beta = − 0.02), total protein in blood (p value = 2.18 × 10−6, beta = 0.02), total protein in serum (p value = 3.00 × 10−6, beta = 0.02), and mean corpuscular hemoglobin concentration (p value = 1.50 × 10−5, beta = − 0.02). LDLPGS in the African ancestry group associated with LDL cholesterol (p value = 5.71 × 10−63, beta = 0.24) and increased total cholesterol (p value = 1.63 × 10−53, beta = 0.21) (Fig. 3d, Additional file 2: Table S9).

The LabWAS of TGPGS was associated with several metabolic measurements (Fig. 3e, Additional file 2: Table S10), including increased TG (p value < 2.23 × 10−308, beta = 0.28), followed by decreased HDL (p value = 4.83 × 10−148, beta = − 0.14), increased total cholesterol to HDL ratio (p value = 2.95 × 10−28, beta = 0.02), increased blood glucose (p value = 1.20 × 10−22, beta = 0.04), increased lipemic index (p value = 1.57 × 10−18, beta = 0.01), increased total blood cholesterol (p value = 1.25 × 10−14, beta = 0.04), increased glycated hemoglobin (p value = 5.69 × 10−9, beta = 0.04), increased bedside glucose (p value = 2.99 × 10−7, beta = 0.04), and increased non-HDL cholesterol (p value = 1.18 × 10−6, beta = 0.08). Additionally, TGPGS showed associations with seven immune labs, white blood cells (p value = 3.90 × 10−30, beta = 0.04), immature granulocytes (p value = 1.99 × 10−14, beta = 0.03), absolute lymphocytes (p value = 2.01 × 10−11, beta = 0.03), monocyte to leukocyte ratio (p value = 5.21 × 10−10, beta = − 0.03), absolute neutrophils (p value = 1.87 × 10−9, beta = 0.03), complement C4 (p value = 1.03 × 10−8, beta = 0.09), and monocyte count (p value = 6.76 × 10−8, beta = − 0.03). Several blood associations also emerged with TGPGS, including carbon dioxide (p value = 2.57 × 10−24, beta = − 0.04), total protein in blood (p value = 4.25 × 10−16, beta = 0.03), mean corpuscular volume (p value = 9.16 × 10−13, beta = − 0.03), mean corpuscular hemoglobin (p value = 9.75 × 10−8, beta = − 0.02), anion gap (p value = 2.03 × 10−17, beta = 0.03), total protein in serum (p value = 2.61 × 10–16, beta = 0.04), and calcitriol (p value = 1.07 × 10−10, beta = − 0.05). Lastly, TGPGS associated with albumin to creatinine ratio (p value = 9.13 × 10−8, beta = 0.10), urate (p value = 6.58 × 10−9, beta = 0.06), urinary pH (7.66 × 10−7, beta = − 0.02), and urinary albumin concentration (p value = 2.99 × 10−5, beta = 0.06). In the African ancestry group, TGPGS showed significant associations with increased triglycerides (p value = 1.66 × 10−53, beta = 0.19), decreased HDL cholesterol (p value = 6.08 × 10−11, beta = − 0.08), and increased glucose (p value = 2.33 × 10−5, beta = 0.04) (Fig. 3f, Additional file 2: Table S11).

LabWAS of a polygenic score for coronary artery disease

We next sought to recapitulate the risk biomarker profile for CAD through a LabWAS of a CADPGS. The CADPGS reproduced associations, in the direction of risk, with canonical risk factors for CAD (Fig. 4a, Additional file 2: Table S12) in the European ancestry population, including decreased HDL (p value = 6.20 × 10−39, beta = − 0.07), increased TG (p value = 3.98 × 10−25, beta = 0.06), increased blood glucose (p value = 1.18 × 10−21, beta = 0.04) and glycated hemoglobin (p value = 2.36 × 10−12, beta = 0.05), and bedside glucose (p value = 1.10 × 10−6, beta = 0.03). The CADPGS also associated with other known biomarkers of cardiovascular health such as increased troponin-I (p value = 7.20 × 10−9, beta = 0.04) and brain natriuretic peptide (p value = 2.12 × 10−7, beta = 0.05). CADPGS associated with six blood composition markers, red blood cell distribution width (p value = 1.60 × 10−11, beta = 0.03), mean corpuscular hemoglobin (p value = 6.73 × 10−10, beta = − 0.02), mean corpuscular volume (p value = 1.17 × 10−9, beta = − 0.02), carbon dioxide (p value = 3.36 × 10−9, beta = − 0.02), red blood cell sedimentation rate (p value = 2.10 × 10−7, beta = 0.05), and international normalized rate (p value = 1.96 × 10−5, beta = 0.03). Finally, CADPGS associated with white blood cell count (p value = 8.75 × 10−11, beta = 0.02), creatinine (p value = 2.13 × 10−6, beta = 0.02), and blood urea nitrogen (p value = 1.09 × 10−5, beta = 0.02).

Fig. 4
figure 4

LabWAS of PGSCAD in individuals of a European ancestry and b individuals of African ancestry. LabWAS of PGSCAD after controlling for CAD diagnosis in individuals of c European ancestry and d individuals of African ancestry. The red lines indicate the Bonferroni threshold for statistical significance and the blue line indicates a p value of 0.05. Upward triangles indicate that the PGSCAD is associated with increased levels of the lab, while downward triangles indicate an association with reduced levels of the lab

Full size image

Notably, the CADPGS was not initially associated with LDL values (p value = 0.13, beta = 0.008). The lack of association, however, was attributable to lipid altering medication use and a significant association between the CADPGS and LDL levels was detected when we restricted to pre-medication values (p = 6.19 × 10−9, beta = 0.04).

To determine which biomarkers were explained by the clinical presence of CAD as opposed to the genetic risk for CAD, we adjusted the LabWAS of CADPGS for the coronary atherosclerosis phecode (411) (Fig. 4c, Additional file 2: Table S13). Four canonical biomarkers of CAD risk remained associated with CADPGS including TG (p value = 2.88 × 10−14, beta = 0.05), pre-medication LDL (p value = 2.40 × 10−13 beta = 0.05), HDL (p value = 2.55 × 10−13, beta = − 0.04), LDL-C (p value = 8.48 × 10−11, beta = 0.04), blood glucose (p value = 2.55 × 10−9, beta = 0.02), total cholesterol (p value = 4.16 × 10−9, beta = 0.03), and glycated hemoglobin (p value = 3.16 × 10−6, beta = 0.03). The CADPGS also remained associated with one immune marker, white blood cell count (p value = 6.44 × 10−6, beta = 0.02), and two other blood biomarkers, mean corpuscular volume (p value = 3.23 × 10−7, beta = − 0.02) and mean corpuscular hemoglobin (p value = 4.18 × 10−6, beta = − 0.02).

None of the associations in the initial LabWAS of CADPGS among African ancestry individuals reached phenome-wide significance; however, three of the top four associations were canonical CAD risk factors including increased glycated hemoglobin A1c (p value = 9.56 × 10−4, beta = 0.04), increased glucose (p value = 0.002, beta = 0.03), and increased LDL cholesterol (p value = 0.003, beta = 0.04) (Fig. 4b, Additional file 2: Table S14). When the LDL levels were restricted to pre-medication values, the top association with CADPGS was pre-medication LDL (p value = 8.50 × 10−5, beta = 0.06); however, this association did not pass multiple testing correction. After controlling the analysis for CAD diagnosis, the association between CADPGS and pre-medication LDL surpassed the Bonferroni correction for phenome-wide significance (p value = 3.92 × 10−5, beta = 0.06) (Fig. 4d, Additional file 2: Table S15).

Lastly, we ran a LabWAS of CAD diagnosis (i.e., using CAD cases/control status (Additional file 1) as the predictor variable) after adjusting for sex and median age across the EHR, which revealed the medical comorbidity pattern of CAD. CAD diagnosis was significantly associated with 136 out of 734 labs in our sample (Additional file 3: Fig. S7, Additional file 2: Table S15), including 34 immune, 32 blood, 24 metabolic, 17 cardiovascular, 8 urinary, 5 toxicology/pharmacology, 4 endocrine, 3 kidney, 3 liver, 1 cancer, and 5 other markers.

Replication in Mass General Brigham Biobank

In the MGBB, there were 21,499 individuals of European descent with genetic data available with recorded lab data. Slightly more than half of the sample was female (51.5%) and the average age was 56.1 years. The MGBB patients contained 1094 CAD cases and 20,405 CAD controls.

In MGBB, the HDLPGS most strongly associated with HDL cholesterol (p value < 2.23 × 10−308, beta = 0.33), followed by decreased triglycerides (p value = 2.77 × 10−109, beta = − 0.17), increased total cholesterol (p value = 4.96 × 10−31, beta = 0.09), and decreased very low-density lipoprotein (p value = 2.62 × 10−29, beta = − 0.14). HDLPGS also associated with decreased values of glucose (p value = 3.33 × 10−27, beta = − 0.07), hemoglobin A1c (p value = 3.64 × 10−18, beta = − 0.07), and mean glucose value (p value = 4.10 × 10−17, beta = − 0.07). Additional associations with HDLPGS included cardiac relative risk (p value = 5.72 × 10−17, beta = − 0.20), alanine aminotransferase (p value = 2.45 × 10−10, beta = − 0.04), white blood cell count (p value = 1.03 × 10−9, beta = − 0.04), mean corpuscular volume (p value = 6.51 × 10−8, beta = 0.03), non-HDL cholesterol (p value = 1.41 × 10−7, beta = − 0.06), red blood cell distribution width (p value = 2.60 × 10−7, beta = − 0.03), neutrophils (p value = 2.95 × 10−7, beta = − 0.03), urate (p value = 1.79 × 10−6, beta = − 0.05), and alkaline phosphatase (p value = 2.01 × 10−6, beta = − 0.03) (Additional file 3: Fig. 8a, Additional file 2: Table S17).

The LDLPGS associated with four metabolic labs including LDL-C (p value = 1.78 × 10−158, beta = 0.24), total cholesterol (p value = 2.37 × 10−158, beta = 0.20), calculated LDL cholesterol (p value = 1.28 × 10−81, beta = 0.23), and non-HDL cholesterol (p value = 2.90 × 10−68, beta = 0.19). The LDLPGS also associated with complement C4 (p value = 1.85 × 10−5, beta = 0.09), red blood cell sedimentation rate (p value = 2.60 × 10−5, beta = 0.04), and increased cardiac relative risk (p value = 3.80 × 10−5, beta = 0.10) (Additional file 3: Fig. 8b, Additional file 2: Table S18).

The TGPGS associated with twelve metabolic labs, including increased measured triglycerides (p value < 2.23 × 10−308, beta = 0.32), followed by increased very low-density lipoprotein (p value = 8.90 × 10−129, beta = 0.30), decreased HDL (p value = 1.33 × 10−123, beta = − 0.17), increased non-HDL cholesterol (p value = 8.70 × 10−28, beta = 0.12), increased glucose (p value = 4.56 × 10−14, beta = 0.05), average glucose (p value = 4.16 × 10−10, beta = 0.05), total cholesterol (p value = 1.58 × 10−9, beta = 0.05), anion gap (p value = 1.52 × 10−7, beta = 0.03), total protein (p value = 4.63 × 10−7, beta = 0.03), globulin in serum (p value = 8.80 × 10−6, beta = 0.03), aspartate aminotransferase (p value = 1.26 × 10−5, beta = 0.03), and sodium (p value = 1.27 × 10−5, beta = − 0.03). TGPGS also associated with seven immune labs, white blood cell count (p value = 3.89 × 10−17, beta = 0.05), lymphocytes (p value = 7.86 × 10−11, beta = 0.04), complement C4 (p value = 1.58 × 10−9, beta = 0.13), automated lymphocyte count (p value = 2.14 × 10−9, beta = 0.09), neutrophils (p value = 3.09 × 10−7, beta = 0.05), automated neutrophil count (p value = 5.13 × 10−7, beta = 0.03), and monocytes (p value = 3.38 × 10−6, beta = 0.05). Ten additional labs significantly associated with TGPGS, including increased cardiac relative risk (p value = 5.49 × 10−15, beta = 0.19), mean corpuscular volume (p value = 3.02 × 10−14, beta = − 0.05), glycated hemoglobin A1c (p value = 5.00 × 10−11, beta = 0.05), urinary pH (p value = 9.58 × 10−10, beta = − 0.04), red blood cell sedimentation rate (p value = 2.22 × 10−8, beta = 0.05), alanine aminotransferase (p value = 3.88 × 10−8, beta = 0.04), alkaline phosphatase (p value = 3.17 × 10−7, beta = 0.03), blood carbon dioxide (p value = 5.63 × 10−7, beta = − 0.03), mean corpuscular hemoglobin (p value = 1.49 × 10−6, beta = − 0.03), and urate (p value = 1.62 × 10−6, beta = 0.05) (Additional file 3: Fig. 8c, Additional file 2: Table S19).

Finally, the CADPGS associated with several known CAD risk factors, including decreased HDL-C (p value = 1.56 × 10−21, beta = − 0.07), increased glucose (p value = 9.91 × 10−15, beta = 0.05), increased glycated hemoglobin A1c (p value = 4.44 × 10−14, beta = 0.06), mean glucose (p value = 1.75 × 10−12, beta = 0.06), and increased triglycerides (p value = 2.09 × 10−12, beta = 0.05). The CADPGS also associated with increased red blood cell distribution width (p value = 2.42 × 10−14, beta = 0.05), increased red blood cell sedimentation rate (p value = 4.11 × 10−9, beta = 0.05), increased alanine aminotransferase (p value = 2.59 × 10−8, beta = 0.04), decreased hemoglobin (p value = 1.45 × 10−6, beta = − 0.03), increased alkaline phosphatase (p value = 2.26 × 10−6, beta = 0.03), increased white blood cell count (p value = 6.77 × 10−6, beta = 0.03), decreased albumin (p value = 1.06 × 10−5, beta = − 0.03), increased globulin (p value = 1.29 × 10−5, beta = 0.03), decreased iron (p value = 3.36 × 10−5, beta = − 0.04), and decreased hematocrit (p value = 3.77 × 10−5, beta = − 0.03) (Additional file 3: Fig. 9a, Additional file 2: Table S20).

After adjusting for CAD diagnosis, CADPGS remained associated with several heart disease risk factors including decreased HDL-C (p value = 8.23 × 10−16, beta = − 0.06), increased glucose (p value = 6.80 × 10−11, beta = 0.04), increased hemoglobin A1c (p value = 2.08 × 10−10, beta = 0.05), increased mean glucose (p value = 2.29 × 10−9, beta = 0.05), and increased triglycerides (p value = 3.48 × 10−9, beta = 0.05). Additionally, associations with red blood cell distribution width (p value = 1.40 × 10−12, beta = 0.04), alanine aminotransferase (p value = 4.01 × 10−8, beta = 0.04), red blood cell sedimentation rate (p value = 5.50 × 10−8, beta = 0.05), alkaline phosphatase (p value = 5.44 × 10−6, beta = 0.03), serum globulin (p value = 4.51 × 10−5, beta = 0.03), and white blood cell count (p value = 4.67 × 10−5, beta = 0.03) remained (Additional file 3: Fig. 9b, Additional file 2: Table S21). In MGBB, the CADPGS was not associated with levels of LDL-C (p value = 0.06, beta = − 0.03), and we were unable to investigate the effects of cholesterol lowering medications on the association.

Discussion

The results of our study add to a growing body of evidence indicating that lab values from EHRs with linked genetic data can be mined at scale to identify biomarkers for complex disease [1,2,3,4,5]. Our proof-of-principle analyses focused on lipids and CAD in 94,747 genotyped BioVU patients and revealed that EHR lipid values cleaned using our QualityLab pipeline were genetically comparable to those measured in samples ascertained for research. Here, we describe two proof concept studies that demonstrate the power of our proposed discovery paradigm. First, we show that PGS for lipids (HDL, LDL, and triglycerides) associate robustly to their referent lipid across ancestries (Fig. 3). Moreover, the CADPGS recapitulated associations with known biomarkers in individuals of European ancestry in two biobanks. Unlike CAD, many complex diseases do not yet have bona fide biomarkers, but do have well-powered GWAS that can be used to mine large biobanks and identify quantitative labs which may be correlated, even weakly, with genetic risk for disease. Importantly, the association between CADPGS and canonical risk factors was significant even among those who did not have a CAD diagnosis. In analyses in Mass General Brigham Biobank, several of the associations with CADPGS replicated, helping to validate our approach to cleaning and analyzing EHR laboratory data. Interestingly, CADPGS also associated with white blood cell count, an inflammatory marker that is not currently used to diagnose or monitor heart disease. This association remained after controlling for CAD diagnosis, indicating that CAD genetics could play a role in increasing inflammation. These results highlight the usefulness of our approach which takes advantage of the entire patient population regardless of disease status. This approach offers a potential path forward for the detection of novel biomarkers and for improved understanding of biomarker activity during the prodromal phase of disease. Furthermore, while disease PGS are not diagnostic, they may be useful in identifying pre-symptomatic individuals whose lab values should be monitored more closely.

Furthermore, we show that treatments (in this example, lipid-altering medications) can influence the detection of risk biomarkers at the genetic level. For example, we found that the genetic correlation between LDL measurements in BioVU and MVP increased considerably when we restricted to pre-medication LDL measurements and controlled for CAD or diabetes diagnosis. Additionally, the CADPGS was strongly associated with pre-medication median LDL values, but was not associated with combined pre- and post-medication median LDL values. This finding also has important and complex implications for the clinical use of PGS recently discussed in the literature [35, 36]. These results indicate that as preventative treatments for complex diseases are adopted (e.g., lipid-altering medications), the risk factors targeted by those treatments (e.g., lipids) are less likely to play a role in the development of subsequent disease (e.g., CAD) in current and future treated populations. Thus, today’s PGS will no longer identify at-risk individuals in future generations who are routinely treated for risk factors which are only now being discovered. Moreover, cases ascertained today for GWAS of diseases with available preventative treatments will be enriched for a different set of genetic (and environmental) risk factors because those individuals with risk factors that can be treated are less likely to develop the disease. PGS, while incredibly valuable, provide only a snapshot of the human genetic profile of complex disease and thus are highly susceptible to these types of cohort effects in addition to other known sources of technical and experimental artifacts [37, 38].

Though the results and approach presented provide an exciting path forward for genetic analysis of EHR-lab data, important limitations should be acknowledged. First, our analyses yielded more associations in patients of European ancestry compared to patients of African ancestry. This is likely to due to decreased power from both the discovery GWASs and the target sample. BioVU has considerably fewer patients of African ancestry than European ancestry, impacting our statistical power to find associations. The polygenic scores of lipids, which were trained on trans-ancestry GWAS summary statistics including individuals of African descent, strongly associated with the referent lipid in the African ancestry sample with effect estimates similar to those found in the European sample. However, the CAD polygenic score, which was trained on a trans-ancestry GWAS that did not include African ancestry samples, yielded far fewer significant associations. These results highlight the critical importance of diversity in GWAS as the downstream applications of such studies are dramatically impacted by representation. As the number of ancestrally diverse GWAS increase, so too will our ability to identify novel biomarkers in different ancestral groups, and the QualityLab pipeline is poised to deliver on these analyses. The QualityLab pipeline could also have more immediate clinical impact for diverse populations. Genetic ancestry, race, sex, age, and ethnicity strongly influence the distribution of lab tests results in healthy people [39], but many current reference ranges were developed using White middle-aged men and are applied to patients irrespective of these differences. This could result in under- or over-diagnosis in some patient groups, and developing lab reference ranges appropriate for diverse demographics is low-hanging fruit for precision medicine. The QualityLab pipeline provides summary metrics based on demographic features which allows the user to evaluate lab distributions across populations, sexes, and ages.

Second, polygenic scores are based on GWAS summary statistics that are typically unadjusted for phenotypic comorbidities. While this approach is optimal in GWAS for many reasons, it introduces the possibility of “phenotypic hitchhiking” in which a comorbid trait is unintentionally selected during the ascertainment of the index trait. Thus, two heritable phenotypes that might share common environmental risk factors but no genetic risk factors can subsequently appear correlated in PGS analysis, even in independent samples. We therefore emphasize that this genetic approach is still fundamentally correlational.

Third, high-throughput analysis of 939 lab traits in our LabWAS required us to prioritize statistical model performance over coefficient interpretability. In our primary analysis, we transformed lab values to fit the normal distribution to improve the performance of the linear regression models [21]. We applied the rank-based inverse normal quantile transformation to all labs, which ensured trait normality by replacing the value of each observation with its quantile from the standard normal distribution. The inverse normal quantile transformation thus preserved the rank ordering of observations, but not the values themselves, and model coefficients therefore are uninterpretable on the original scale. For example, based on our LabWAS results, we are unable to report the change in LDL levels in mg/dL per SD increase in the CADPGS. Multiple testing correction was another statistical challenge inherent to the high-throughput analysis of lab traits. We used the Bonferroni threshold for statistical significance, but this threshold is likely to be overly strict because it ignores the correlation between lab tests.

Conclusions

Here, we propose that PGS for complex disease can be used to discover genetically related biomarkers of disease by mining quantitative physiological measurements collected during routine clinical testing, but caution that mindful interpretation of correlational results is paramount to progress. We demonstrate the robustness of this discovery paradigm in a proof of principal analysis focused on CAD. As EHR resources grow in size, standardized quality control and analysis pipelines will be necessary to compare results across samples. QualityLab and LabWAS provide a starting point for consistent analysis of lab results stored in various EHR systems. Furthermore, we demonstrated that EHR-derived lipids are similar to measurements ascertained in traditional cohort studies, providing additional rationale for analyses of EHR labs [40]. QualityLab and LabWAS are scalable programs that can be used to confirm clinical paradigms and discover new genetic and environmental relationships between biomarkers and complex traits. We propose that future studies will leverage this discovery paradigm for analysis of rare or understudied complex traits with no known biomarker associations (e.g., psychiatric disorders).

Availability of data and materials

The data that support the findings of this study are available from Vanderbilt University Medical Center but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission of Vanderbilt University Medical Center.

GWAS data generated are available on the GWAS Catalog (accession numbers GCST90012603 - GCST90012784 are listed in Additional file 2: Table S22) [41] and on https://www.dropbox.com/sh/w1pbe0jq1bjkpc5/AAAUIdtBgUybE6iHraE8jvp8a?dl=0.

Code for QualityLab and LabWAS software used to generate the results presented in this paper can be found here (https://bitbucket.org/straubp_vandy/quality_labs/) [42] and here (https://bitbucket.org/juliasealock/labwas/) [43].

LabWAS plots can be viewed interactively here: https://dennislab.ca/labwas-in-electronic-health-records/

Abbreviations

Lab:

Laboratory

EHR:

Electronic health record

PGS:

Polygenic score

LabWAS:

Lab-wide association scan

PheWAS:

Phenome-wide association scan

VUMC:

Vanderbilt University Medical Center

HDL:

High-density lipoprotein

LDL:

Low-density lipoprotein

TG:

Triglycerides

CAD:

Coronary artery disease

ICD:

International Classification of Disease

CPT:

Current Procedural Terminology

HRC:

Haplotype Reference Consortium

GWAS:

Genome-wide association scan

INT:

Inverse normal quantile transformation

h2 SNP :

SNP-based heritability

GCTA:

Genome-wide complex trait analysis

REML:

Restricted maximum likelihood

GLGC:

Global Lipids Genetics Consortium

PRS-CS:

Polygenic Risk Score – Continuous Shrinkage

References

  1. Shameer K, Denny JC, Ding K, Jouni H, Crosslin DR, de Andrade M, et al. A genome- and phenome-wide association study to identify genetic variants influencing platelet count and volume and their pleiotropic effects. Hum Genet. 2014;133:95–109.

    Article  Google Scholar 

  2. Hoffmann TJ, Theusch E, Haldar T, Ranatunga DK, Jorgenson E, Medina MW, et al. A large electronic-health-record-based genome-wide study of serum lipids. Nat Genet. 2018. https://doi.org/10.1038/s41588-018-0064-5.

  3. Verma A, Lucas A, Verma SS, Zhang Y, Josyula N, Khan A, et al. PheWAS and beyond: the landscape of associations with medical diagnoses and clinical measures across 38,662 individuals from Geisinger. Am J Hum Genet. 2018. https://doi.org/10.1016/j.ajhg.2018.02.017.

  4. Klarin D, Damrauer SM, Cho K, Sun YV, Teslovich TM, Honerlaw J, et al. Genetics of blood lipids among ~300,000 multi-ethnic participants of the million veteran program. Nat Genet. 2018;50:1514–23.

    Article  CAS  Google Scholar 

  5. Verma A, Leader JB, Verma SS, Frase A, Wallace J, Dudek S, et al. Integrating clinical laboratory measures and ICD-9 code diagnoses in phenome-wide association studies. Pac Symp Biocomput. 2016. https://doi.org/10.1142/9789814749411_0016.

  6. Estiri H, Klann JG, Murphy SN. A clustering approach for detecting implausible observation values in electronic health records data. BMC Med Inform Decis Mak. 2019. https://doi.org/10.1186/s12911-019-0852-6.

  7. Pivovarov R, Albers DJ, Sepulveda JL, Elhadad N. Identifying and mitigating biases in EHR laboratory tests. J Biomed Inform. 2014. https://doi.org/10.1016/j.jbi.2014.03.016.

  8. Perrotta PL, Karcher DS. Validating laboratory results in electronic health records: a college of American pathologists Q-probes study. Arch Pathol Lab Med. 2016. https://doi.org/10.5858/arpa.2015-0320-CP.

  9. Denny JC, Bastarache L, Roden DM. Phenome-wide association studies as a tool to advance precision medicine. Annu Rev Genomics Hum Genet. 2016;17:353–73.

    Article  CAS  Google Scholar 

  10. Denny JC, Bastarache L, Ritchie MD, Carroll RJ, Zink R, Mosley JD, et al. Systematic comparison of phenome-wide association study of electronic medical record data and genome-wide association study data. Nat Biotechnol. 2013;31:1102–10.

    Article  CAS  Google Scholar 

  11. Pendergrass SA, Brown-Gentry K, Dudek S, Frase A, Torstenson ES, Goodloe R, et al. Phenome-wide association study (PheWAS) for detection of pleiotropy within the Population Architecture using Genomics and Epidemiology (PAGE) network. Plos Genet. 2013;9. https://doi.org/10.1371/journal.pgen.1003087.

  12. Robinson JR, Denny JC, Roden DM, Van Driest SL. Genome-wide and phenome-wide approaches to understand variable drug actions in electronic health records. Clin Transl Sci. 2018;11:112–22.

    Article  Google Scholar 

  13. Lucas AM, Palmiero NE, McGuigan J, Passero K, Zhou J, Orie D, et al. CLARITE facilitates the quality control and analysis process for EWAS of metabolic-related traits. Front Genet. 2019. https://doi.org/10.3389/fgene.2019.01240.

  14. Roden DM, Pulley JM, Basford MA, Bernard GR, Clayton EW, Balser JR, Masys DR. Development of a large-scale de-identified DNA biobank to enable personalized medicine. Clin Pharmacol Ther. 2008. https://doi.org/10.1038/clpt.2008.89.

  15. Purcell S, Neale B, Todd-Brown K, Thomas L, Ferreira MAR, Bender D, et al. PLINK: a tool set for whole-genome association and population-based linkage analyses. Am J Hum Genet. 2007;81:559–75.

    Article  CAS  Google Scholar 

  16. Price AL, Patterson NJ, Plenge RM, Weinblatt ME, Shadick NA, Reich D. Principal components analysis corrects for stratification in genome-wide association studies. Nat Genet. 2006;38:904–9.

    Article  CAS  Google Scholar 

  17. Patterson N, Price AL, Reich D. Population structure and eigenanalysis. Plos Genet. 2006;2:2074–93.

    Article  CAS  Google Scholar 

  18. Das S, Forer L, Schönherr S, Sidore C, Locke AE, Kwong A, et al. Next-generation genotype imputation service and methods. Nat Genet. 2016;48:1284–7.

    Article  CAS  Google Scholar 

  19. Auton A, Abecasis GR, Altshuler DM, Durbin RM, Bentley DR, Chakravarti A, et al. A global reference for human genetic variation. Nature. 2015;526:68–74.

    Article  Google Scholar 

  20. Sofer T, Zheng X, Gogarten SM, Laurie CA, Grinde K, Shaffer JR, et al. A fully adjusted two-stage procedure for rank-normalization in genetic association studies. Genet Epidemiol. 2019. https://doi.org/10.1002/gepi.22188.

  21. McCaw ZR, Lane JM, Saxena R, Redline S, Lin X. Operating characteristics of the rank-based inverse normal transformation for quantitative trait analysis in genome-wide association studies. Biometrics. 2019. https://doi.org/10.1111/biom.13214.

  22. Davis LK, Yu D, Keenan CL, Gamazon ER, Konkashbaev AI, Derks EM, et al. Partitioning the heritability of Tourette syndrome and obsessive compulsive disorder reveals differences in genetic architecture. Plos Genet. 2013. https://doi.org/10.1371/journal.pgen.1003864.

  23. Yang J, Lee SH, Goddard ME, Visscher PM. GCTA: a tool for genome-wide complex trait analysis. Am J Hum Genet. 2011;88:76–82.

    Article  CAS  Google Scholar 

  24. Zeng P, Zhao Y, Qian C, Zhang L, Zhang R, Gou J, et al. Statistical analysis for genome-wide association study. J Biomed Res. 2015. https://doi.org/10.7555/JBR.29.20140007.

  25. Jiang L, Zheng Z, Qi T, Kemper KE, Wray NR, Visscher PM, Yang J. A resource-efficient tool for mixed model association analysis of large-scale data. Nat Genet 2019;51:1749–55. https://doi.org/10.1038/s41588-019-0530-8.

  26. Willer CJ, Schmidt EM, Sengupta S, Peloso GM, Gustafsson S, Kanoni S, et al. Discovery and refinement of loci associated with lipid levels. Nat Genet. 2013;45:1274–85.

    Article  CAS  Google Scholar 

  27. Bulik-Sullivan B, Loh PR, Finucane HK, Ripke S, Yang J, Patterson N, et al. LD score regression distinguishes confounding from polygenicity in genome-wide association studies. Nat Genet. 2015. https://doi.org/10.1038/ng.3211.

  28. Evans LM, Tahmasbi R, Vrieze SI, Abecasis GR, Das S, Gazal S, et al. Comparison of methods that use whole genome data to estimate the heritability and genetic architecture of complex traits. Nat Genet. 2018. https://doi.org/10.1038/s41588-018-0108-x.

  29. Bulik-Sullivan B, Finucane HK, Anttila V, Gusev A, Day FR, Loh PR, et al. An atlas of genetic correlations across human diseases and traits. Nat Genet. 2015;47:1236–41.

    Article  CAS  Google Scholar 

  30. Ning Z, Pawitan Y, Shen X. High-definition likelihood inference of genetic correlations across human complex traits. Nat Genet. 2020. https://doi.org/10.1038/s41588-020-0653-y.

  31. Ge T, Chen CY, Ni Y, Feng YCA, Smoller JW. Polygenic prediction via Bayesian regression and continuous shrinkage priors. Nat Commun. 2019;10:1–10.

    Article  Google Scholar 

  32. Nikpay M, Goel A, Won HH, Hall LM, Willenborg C, Kanoni S, et al. A comprehensive 1000 Genomes-based genome-wide association meta-analysis of coronary artery disease. Nat Genet. 2015;47:1121–30.

    Article  CAS  Google Scholar 

  33. Choi SW, O’Reilly PF. PRSice-2: Polygenic Risk Score software for biobank-scale data. Gigascience. 2019;8:1–6.

    Article  Google Scholar 

  34. Karlson EW, Boutin NT, Hoffnagle AG, Allen NL. Building the partners healthcare biobank at partners personalized medicine: informed consent, return of research results, recruitment lessons and operational considerations. J Pers Med. 2016;6:1–11.

    Article  Google Scholar 

  35. Khera AV, Chaffin M, Aragam KG, Haas ME, Roselli C, Choi SH, et al. Genome-wide polygenic scores for common diseases identify individuals with risk equivalent to monogenic mutations. Nat Genet. 2018. https://doi.org/10.1038/s41588-018-0183-z.

  36. Lambert SA, Abraham G, Inouye M. Towards clinical utility of polygenic risk scores. Hum Mol Genet. 2019;28:R133–42.

    Article  CAS  Google Scholar 

  37. Janssens ACJW. Validity of polygenic risk scores: are we measuring what we think we are? Hum Mol Genet. 2019. https://doi.org/10.1093/hmg/ddz205.

  38. Curtis D. Polygenic risk score for schizophrenia is more strongly associated with ancestry than with schizophrenia. Psychiatr Genet. 2018. https://doi.org/10.1097/YPG.0000000000000206.

  39. Tahmasebi H, Trajcevski K, Higgins V, Adeli K. Influence of ethnicity on population reference values for biochemical markers. Crit Rev Clin Lab Sci. 2018. https://doi.org/10.1080/10408363.2018.1476455.

  40. Casey JA, Schwartz BS, Stewart WF, Adler NE. Using electronic health records for population health research: a review of methods and applications. Annu Rev Public Health. 2016. https://doi.org/10.1146/annurev-publhealth-032315-021353.

  41. Dennis J, Sealock JM, et al. (2020): Clinical laboratory test-wide association scan of polygenic scores identifies biomarkers of complex disease. GWAS Catalog. https://www.ebi.ac.uk/gwas/. Accessed 20 Nov 2020.

  42. Straub P, Dennis J, Sealock JM, et al (2020): BitBucket. https://bitbucket.org/straubp_vandy/quality_labs/.

    Google Scholar 

  43. Sealock JM, Dennis J, Straub P, et al (2020): BitBucket. https://bitbucket.org/juliasealock/labwas/.

    Google Scholar 

Download references

Acknowledgements

The authors thank the Vanderbilt University Medical Center Biobank and Mass General Brigham Biobank for providing genomic and health information data.

Funding

JKD is supported by the Canadian Institutes of Health Research (award MFE- 142936). JMS is supported by NIH/NIGMS training grant 5T32GM080178-12. JDM is supported by AHA grant 16FTF30130005 and NIH GM130791-01. GC is supported by NIH UL1TR000427 and 1U24CA242637-01. LKD is supported by NIH 1R01MH118233-01, 5U54MD010722-04, 5R01MH113362-03, and 1R56MH120736-01. JWS is supported by NIH R01 MH118233. DMR is supported by 5R01MH113362, R01MH11629, R01MH120736, and R01MH118233.

This project was conducted in part using the resources of the Advanced Computing Center for Research and Education at Vanderbilt University, Nashville, TN. The datasets used for this project were obtained from Vanderbilt University Medical Center’s Synthetic Derivative, which is supported by numerous sources: institutional funding, private agencies, and federal grants. These include the NIH funded Shared Instrumentation Grant S10RR025141 and CTSA grants UL1TR002243, UL1TR000445, and UL1RR024975. The project described was supported by the National Center for Research Resources, Grant UL1 RR024975-01, and is now at the National Center for Advancing Translational Sciences, Grant 2 UL1 TR000445-06. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.

Author information

Author notes

  1. Jessica K. Dennis and Julia M. Sealock contributed equally to this work.

Authors and Affiliations

  1. Division of Genetic Medicine, Department of Medicine, Vanderbilt University Medical Center, Nashville, TN, 37232, USA

    Jessica K. Dennis, Julia M. Sealock, Peter Straub, Donald Hucks, Ky’Era Actkins, Annika Faucon, Slavina B. Goleva, Maria Niarchou, Kritika Singh, Theodore Morley, Douglas M. Ruderfer & Lea K. Davis

  2. Vanderbilt Genetics Institute, Vanderbilt University Medical Center, Nashville, TN, 37232, USA

    Jessica K. Dennis, Julia M. Sealock, Peter Straub, Donald Hucks, Ky’Era Actkins, Annika Faucon, Slavina B. Goleva, Maria Niarchou, Kritika Singh, Theodore Morley, Douglas M. Ruderfer & Lea K. Davis

  3. Department of Medical Genetics, University of British Columbia, Vancouver, BC, V5Z 4H4, Canada

    Jessica K. Dennis

  4. Psychiatric & Neurodevelopmental Genetics Unit, Center for Genomic Medicine, Massachusetts General Hospital, Boston, MA, 02114, USA

    Younga H. Lee, Yen-Chen Anne Feng, Tian Ge & Jordan W. Smoller

  5. Department of Psychiatry, Harvard Medical School, Boston, MA, 02115, USA

    Younga H. Lee, Tian Ge & Jordan W. Smoller

  6. Stanley Center for Psychiatric Research, Broad Institute of Harvard and MIT, Cambridge, MA, 02142, USA

    Younga H. Lee, Yen-Chen Anne Feng, Tian Ge & Jordan W. Smoller

  7. Department of Microbiology, Immunology, and Physiology, Meharry Medical College, Nashville, TN, 37232, USA

    Ky’Era Actkins

  8. Analytic and Translational Genetics Unit, Center for Genomic Medicine, Massachusetts General Hospital, Boston, MA, 02114, USA

    Yen-Chen Anne Feng

  9. Department of Molecular Physiology and Biophysics, Vanderbilt University Medical Center, Nashville, TN, 37232, USA

    Slavina B. Goleva & Lea K. Davis

  10. Department of Psychiatry and Behavioral Sciences, Vanderbilt University Medical Center, Nashville, TN, 37232, USA

    Douglas M. Ruderfer & Lea K. Davis

  11. Departments of Medicine and Biomedical Informatics, Vanderbilt University Medical Center, Nashville, TN, 37232, USA

    Douglas M. Ruderfer, Jonathan D. Mosley & Lea K. Davis

  12. Department of Biostatistics and Medical Informatics, University of Wisconsin-Madison, Madison, WI, 53706, USA

    Guanhua Chen

  13. Division of Genetic Medicine, Department of Medicine, Vanderbilt Genetics Institute, Vanderbilt University, 511-A Light Hall, 2215 Garland Ave, Nashville, TN, 37232, USA

    Lea K. Davis

Authors
  1. Jessica K. Dennis
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Julia M. Sealock
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. Peter Straub
    View author publications

    You can also search for this author in PubMed Google Scholar

  4. Younga H. Lee
    View author publications

    You can also search for this author in PubMed Google Scholar

  5. Donald Hucks
    View author publications

    You can also search for this author in PubMed Google Scholar

  6. Ky’Era Actkins
    View author publications

    You can also search for this author in PubMed Google Scholar

  7. Annika Faucon
    View author publications

    You can also search for this author in PubMed Google Scholar

  8. Yen-Chen Anne Feng
    View author publications

    You can also search for this author in PubMed Google Scholar

  9. Tian Ge
    View author publications

    You can also search for this author in PubMed Google Scholar

  10. Slavina B. Goleva
    View author publications

    You can also search for this author in PubMed Google Scholar

  11. Maria Niarchou
    View author publications

    You can also search for this author in PubMed Google Scholar

  12. Kritika Singh
    View author publications

    You can also search for this author in PubMed Google Scholar

  13. Theodore Morley
    View author publications

    You can also search for this author in PubMed Google Scholar

  14. Jordan W. Smoller
    View author publications

    You can also search for this author in PubMed Google Scholar

  15. Douglas M. Ruderfer
    View author publications

    You can also search for this author in PubMed Google Scholar

  16. Jonathan D. Mosley
    View author publications

    You can also search for this author in PubMed Google Scholar

  17. Guanhua Chen
    View author publications

    You can also search for this author in PubMed Google Scholar

  18. Lea K. Davis
    View author publications

    You can also search for this author in PubMed Google Scholar

Contributions

JKD, JMS, GC, and LKD conceived and designed the study. All authors contributed to the acquisition, analysis, or interpretation of data. JKD, JMS, and LKD drafted the manuscript. All authors contributed to revision of the manuscript. All authors read and approved the final manuscript.

Corresponding author

Correspondence to Lea K. Davis.

Ethics declarations

Ethics approval and consent to participate

BioVU Consent form is provided to patients in the outpatient clinic environments at VUMC. The consent states policies on data sharing and privacy and, upon consent, makes any blood leftover from clinical care eligible for BioVU banking. The VUMC Institutional Review Board oversees BioVU and approved this project. All data included in this study was de-identified and unlinked to any identifying information. This study was reviewed by the Vanderbilt University Medical Center IRB (IRB#172020) and designated as non-human subjects research. The research was conducted in accordance with the principles of the Declaration of Helsinki.

Consent for publication

Not applicable

Competing interests

JWS is an unpaid member of the Bipolar/Depression Research Community Advisory Panel of 23andMe, is a member of the Leon Levy Foundation Neuroscience Advisory Board, and received an honorarium for an internal seminar at Biogen, Inc. He is PI of a collaborative study of the genetics of depression and bipolar disorder sponsored by 23andMe for which 23andMe provides analysis time as in-kind support but no payments. The remaining authors declare that they have no competing interests.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Additional file 1.

Contains supplementary methods information on genotyping and quality control, definition of coronary artery disease and lipids lowering mediation, and polygenic scoring methods.

Additional file 2:

Contains all supplementary tables. Table S1. Descriptive statistics generated by the QualityLab pipeline. Table S2. Lipid-altering medications abstracted from free text in clinical notes. Table S3. Characteristics of patients with clean lab and genotyping data. Table S4. Estimates of SNP-based heritability and sample sizes for all labs passing QualityLab using REML. Table S5. Estimates of SNP-based heritability for blood lipid levels across study populations and estimation methods. Table S6. LabWAS of HDLPGS in the European sample in BioVU. Table S7. LabWAS of HDLPGS in the African sample in BioVU. Table S8. LabWAS of LDLPGS in the European sample in BioVU. Table S9. LabWAS of LDLPGS in the African sample in BioVU. Table S10. LabWAS of TGPGS in the European sample in BioVU. Table S11. LabWAS of TGPGS in the African sample in BioVU. Table S12. LabWAS of CADPGS in the European sample in BioVU. Table S13. LabWAS of CADPGS in the European sample in BioVU controlling for CAD diagnosis. Table S14. LabWAS of CADPGS in the African sample in BioVU. Table S15. LabWAS of CADPGS in the African sample in BioVU controlling for CAD diagnosis. Table S16. LabWAS of CAD diagnosis in the European sample in BioVU. Table S17. LabWAS of HDLPGS in the European sample in MGB. Table S18. LabWAS of LDLPGS in the European sample in MGB. Table S19. LabWAS of TGPGS in the European sample in MGB. Table S20. LabWAS of CADPGS in the European sample in MGB. Table S21. LabWAS of CADPGS in the European sample in MGB controlling for CAD diagnosis. Table S22. The GWAS Catalog study accession numbers for summary statistics from GWAS of each lab trait.

Additional file 3:

Contains all supplementary figures. Figure S1. QualityLab data visualizations. Figure S2. Manhattan and QQ plots from GWAS of BioVU lipids. Figure S3.. Predictive abilities of different polygenic scoring methods. Figure S4. Lipids levels by genetic ancestry. Figure S5. Histogram of BioVU lab heritability estimates. Figure S6. LDL genetic correlation sensitivity analyses. Figure S7. LabWAS Manhattan of CAD diagnosis. Figure S8. LabWAS of Lipids PGS in MGBB. Figure S9. LabWAS of CAD PGS in MGBB.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Dennis, J.K., Sealock, J.M., Straub, P. et al. Clinical laboratory test-wide association scan of polygenic scores identifies biomarkers of complex disease. Genome Med 13, 6 (2021). https://doi.org/10.1186/s13073-020-00820-8

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s13073-020-00820-8

Keywords

Genome Medicine

ISSN: 1756-994X

Contact us