Unfolding of hidden white blood cell count phenotypes for gene discovery using latent class mixed modeling

Abstract

Resting-state white blood cell (WBC) count is a marker of inflammation and immune system health. There is evidence that WBC count is not fixed over time and there is heterogeneity in WBC trajectory that is associated with morbidity and mortality. Latent class mixed modeling (LCMM) is a method that can identify unobserved heterogeneity in longitudinal data and attempts to classify individuals into groups based on a linear model of repeated measurements. We applied LCMM to repeated white blood cell (WBC) count measures derived from electronic medical records of participants of the National Human Genetics Research Institute (NHRGI) electronic MEdical Record and GEnomics (eMERGE) network study, revealing two WBC count trajectory phenotypes. Advancing these phenotypes to GWAS, we found genetic associations between trajectory class membership and regions on chromosome 1p34.3 and chromosome 11q13.4. The chromosome 1 region contains CSF3R, which encodes the granulocyte colony stimulating factor receptor. This protein is a major factor in neutrophil stimulation and proliferation. The association on chromosome 11 contain genes RNF169 and XRRA1; both involved in the regulation of double strand break DNA repair.

Introduction

White blood cell count (WBC) count is a marker of systemic inflammation and immune system health. WBC count varies acutely in response to infection and other environmental exposures. However, resting-state WBC count -- the WBC level when the immune system is neither challenged nor suppressed -- may be an indicator of chronic disease risk. Elevated resting WBC count has been associated with metabolic syndrome^1–4, cardiovascular disease ^5,6 and mortality^7–11. This may reflect excess inflammation as evidenced by WBC count, or leukocytes may contribute directly to disease¹² .

While WBC count is impacted by modifiable factors such as smoking^13–15 and body composition^16–18, resting-state WBC count is also influenced by ancestry, and has been found to be partly under genetic control, with heritability estimated at around 40%¹⁹. Individuals with African ancestry, on average, have lower WBC count compared to individuals with European ancestry, attributed to lower neutrophil count^20,21. Among those with African ancestry, total WBC and neutrophil count has been associated with SNP rs2814778 in the ACKR1/DARC gene, via admixture mapping^22,23. This association was replicated in several genome-wide association studies, including our own, and a meta-analysis^24–27.

There is evidence that resting-state WBC count is not fixed over time. Longitudinal analysis has shown a U-shaped pattern in WBC counts over the lifespan, dipping around age 60 and then increasing⁹. Similarly, cross-sectional data has shown higher WBC count in individuals older than 65 years old¹⁰. Heterogeneity in WBC count trajectory also exists and some trajectories are associated with morbidity and mortality⁸. Because WBC count is also influenced by adiposity, changes in resting-state WBC count may reflect age-related change in body composition. However, in a mouse model, different strains exhibited different WBC count trajectories, indicating these trajectories may be under genetic control²⁸.

Deep phenotyping aims to increase the granularity of a phenotype in hopes that a more precise phenotype will increase the power of a genome-wide association study (GWAS) and lead to larger effect size estimates²⁹. Extending a phenotype over time by harnessing the information contained in longitudinal data instead of simple aggregation is one strategy to deepen phenotype^30,31. Different trajectories of WBC count over the lifespan may be a fruitful deep phenotype to use in GWAS.

Trajectory heterogeneity may be difficult to discern in large, observational datasets using standard statistical methods. Trajectory analysis is a method that can identify unobserved heterogeneity in longitudinal data and attempts to classify individuals into groups based on a linear model of repeated measurements over time³². As such, this method is particularly suited to the type of data gathered in the electronic medical records (EMR) which contains information about multiple traits, gathered repeatedly over time. Trajectory analysis, applied to EMR data, has been used to characterize and identify risk factors for multimorbidity³³, depression³⁴, dementia-related cognitive decline³⁵, and adverse birth weight outcomes³⁶. Trajectory-based phenotypes have been shown to be heritable, used in candidate gene studies, linkage analysis and GWAS, and have been associated with genetic risk scores for a number of complex traits (e.g. systolic blood pressure, BMI, schizophrenia, alcohol use and smoking behavior, ADHD and Autism)^37–44.

Here we applied a trajectory analysis, using latent class mixed modeling (LCMM)⁴⁵, to longitudinal WBC count data obtained from the EMR from the electronic MEdical Record and GEnomics (eMERGE) Network study. We then conducted a GWAS and identified genetic variants associated with the trajectory classes derived in the deep phenotyping step.

Results

Resting-state WBC count data was identified for 14 018 participants. LCMM requires a minimum number of repeated measurements to appropriately model trajectory (here a minimum of three data points for a quadratic model). In our sample, 4 762 participants were excluded due to insufficient data. Excluded participants were younger than included participants (56.6 vs 64.1 year, respectively). There was also a higher proportion of participants of genetically determined African Ancestry (AA) among those excluded. A higher proportion of participants from the Vanderbilt University site and a lower proportion of participants from the Marshfield Clinic site were excluded.

LCMM selection

We evaluated model fit based on Bayesian Information Criteria (BIC), average posterior probability of class membership ≥ 70%, and minor class sample size ≥ 10%. A summary of the models fit are presented in Table 1. Details for all models tested are available in the Supplementary Materials. Based on these criteria, we determined that the two-class solution was the best fitting model tested.

Table 1.

LCMM fit statistics.

Model	Link	Classes	BIC	Entropy	Average Posterior Probability				Sample Size (%)
1a	linear	1	339342.22	--	--				9742 (100)
1b	beta	1	301289.38	--	--	--			9742 (100)
2	beta	2	299820.82	0.27	0.73	0.75			3349 (35)	6292 (65)
3	beta	3	299145.93	0.44	0.75	0.71	0.70		209 (2)	6725 (70)	2702 (28)
4	beta	4	298777.82	0.52	0.77	0.73	0.65	0.70	151 (1)	7639 (78)	1595 (16)	357 (4)

Participants were assigned to a trajectory class based on the class for which they had a higher posterior probability of membership, given their data and the model fit. Figure 1A shows the mean predicted trajectory based on the LCMM for each class. Class 1 is modeled by the equation $-0.10137\ast age\_at\_event+0.00043\ast age\_at\_even{t}^{\mathrm{2⬚}}{}^{}$. The equation for the Class 2 trajectory is $-4.00651-0.07115\ast age\_at\_event+0.00082\ast age\_at\_even{t}^2$. Figure 1B shows the mean observed trajectory and 95% confidence interval for classed participant data. Class 2 was the major trajectory identified, representing 65% of the participants, showed a stable resting-state WBC count trajectory and then increased after age 60. The Class 1 WBC count trajectory decreased steadily across the lifespan and accounted for 35% of sample participants. The trajectories cross at about age 70 and the 95% confidence intervals overlap from ages 68 to 72. The average posterior probability of Class 1 and Class 2 membership was moderately high at 73% and 75%, respectively, but the entropy (a measure of confidence, bounded by 0 and 1) of classification was low (0.27).

Figure 1.

Predicted mean class-specific trajectory (A) and observed mean class-specific trajectory and 95% confidence interval (B) of WBC count by age from the LCMM.

The median number of observations, age-at-event, years of follow-up, WBC counts, and BMI were similar in magnitude between classes (Table 2). Likewise, the distributions of males, those in genetically determined ancestry (GDA) groups, and study site for each class were comparable. The differences between Class 1 and Class 2 demographics were statistically significant, but this may reflect the large sample size rather than meaningful differences in cohort makeup.

Table 2.

Descriptive characteristics of each trajectory class.

	Class 1 (N=3349)	Class 2 (N=6292)	p-value
	MEDIAN [IQR]
OBSERVATIONS	8 [5–12]	7 [4–10]	< 0.0001
AGE AT EVENT	62.5 [52.3–71.8]	64.6 [55.1–72.9]	< 0.0001
NUMBER YEARS FOLLOW-UP	11.7 [6.5–17.5]	10.9 [5.9–15.9]	< 0.0001
WBC COUNT	7.4 [6.2–8.6]	6.2 [5.4–7.1]	< 0.0001
BMI	28.9 [25.1–33.3]	27.9 [24.9–31.3]	< 0.0001
	%
MALE	41	45	< 0.0001
GENETICALLY DETERMINED ANCESTRY
EUROPEAN	88	91
AFRICAN	11	8
ASIAN	1	1	0.0002
SITE
GROUP HEALTH	21	22
MARSHFIELD	37	31
MAYO	12	23
NORTHWESTERN	11	7
VANDERBILT	19	16	<0.0001

GWAS

The results of the joint and GDA group stratified GWAS analyses comparing Class 1 to Class 2 trajectory phenotype are summarized in Figure 2. The Q-Q plots and lambda values of 0.9926, 1.0216 and 0.9888 indicate good control for population stratification in the Joint, AA and European Ancestry (EA) groups, respectively (Figure 2). The Manhattan plots show three regions of interest, with p-values less than 10⁻⁷, for genetic associations with our trajectory phenotype classes: 1q23.2, 1p34.3, and 9q33.1 (Figure 2).

Figure 2.

Manhattan and Q-Q plots summarizing the results of the WBC count trajectory phenotype GWAS, joint and GDA stratified analyses. A) Manhattan plot for the joint analysis. B) Q-Q plot for the joint analysis. C) Manhattan plot for the AA analysis. D) Q-Q plot for the AA analysis. E) Manhattan plot for the EA analysis. F) Q-Q plot for the EA analysis.

The strongest region associated with trajectory class membership we identified was on chromosome 1q23.3 in the joint and AA group analyses with the lead SNP rs2814778 (p-value = 9.83 × 10⁻⁹, joint analysis; 9.56 × 10⁻⁹, AA group analysis). In the AA group, the T allele of SNP rs2814778 was associated with a two-fold higher risk of having the Class 1 trajectory phenotype (OR: 2.23 95% CI: 1.23–4.07). The T allele is the minor allele among individuals of AA, with an allele frequency of 0.21 in our study. By contrast, the T allele is nearly fixed in the EA group, with an allele frequency of 0.996. SNP rs2814778 is located in the first exon of the ACKR1/DARC gene, and together with rs12075, polymorphisms at these loci determine the Duffy blood group. Among individuals of AA, homozygotes for the C allele of rs2814778 have the Duffy-Null phenotype, which is exhibits a strong association with low neutrophil count²³.

We identified an additional significant region on chromosome 1p34.3 in the joint analysis (p-value = 3.54 × 10⁻⁸, lead SNP rs12094900). This SNP is in moderate to high linkage disequilibrium with several other SNPs in a region that contains the genes MRPS15, OSCP1, and CSF3R, (Figure 3). The MRPS15 gene encodes a mitochondrial ribosomal protein. OSCP1, also known as NOR1, is a tumor-suppressor gene associated with nasopharyngeal cancer. CSF3R encodes the receptor for the granulocyte colony-stimulating factor (G-CSF) cytokine. This cytokine-receptor complex stimulates the creation of granulocytes and activates neutrophils.

Figure 3.

Regional association plot for chromosome 1p34.3, joint analysis.

In the AA group analysis, two SNPs on chromosome 9q33.1 we associated with trajectory class membership at the genome-wide threshold (3.40 × 10⁻⁹, lead SNP rs55736771). This SNP is located the third intron of the ASTN2 gene. The protein encoded by this gene is expressed in the brain and has been associated with age of Alzheimer’s disease onset, schizophrenia, and neurodevelopmental disorders in males^46–48.

Given that rs2814778 polymorphism in the ACKR1/DARC gene has a strong effect on resting-state WBC count among individuals of AA and, in our analysis, Class 1 trajectory members had a significantly higher median WBC count compared to Class 2 members, we were concerned that median WBC count was confounding an association between rs2814778 and the trajectory phenotype. To assess potential confounding, we ran an additional set of GWAS analyses, this time adjusting for median WBC count. A comparison of our minimally- and fully-adjusted models is presented in Table 3.

Table 3.

Comparison of significant GWAS results before and after adjusting for median WBC count.

CHR	SNP	Minor allele	Closest gene		OR	P	OR*	P*
Joint
1	rs2814778	C	ACKR1	exon	0.50	9.83E-09	0.68	0.01131
1	rs12094900	G	CSF3R	intergenic	0.83	3.54E-08	0.80	3.62E-07
3	rs75135222	C	DCP1A	intron	1.46	5.53E-07	1.46	9.59E-05
11	rs143804085	G	RNF169	intron	1.31	3.97E-05	1.55	1.10E-07
11	rs117711119	A	XRRA1	intron	1.30	6.52E-05	1.55	1.86E-07
11	rs117258637	G	SPCS2	intron	1.29	9.74E-05	1.54	2.29E-07
11	rs11606575	G	CEP164	intron	0.86	2.11E-04	0.76	8.30E-07
AA
1	rs2814778	T	ACKR1	exon	2.15	9.59E-09	1.26	0.2174
1	rs4657616	G	POP3	intron	2.69	1.24E-07	1.59	0.06193
1	rs856068	C	IFI16	intron	1.97	2.94E-07	1.40	0.05618
1	rs2501339	G	VSIG8	intron	2.17	6.35E-07	1.76	0.006375
6	rs7761344	A	SASH1	intron	1.42	7.89E-04	2.12	5.60E-07
9	rs55736771	A	ASTN2	intron	1.89	3.40E-09	1.79	6.80E-05
EA
11	rs79852880	G	XRRA1	intron	1.30	7.65E-05	1.54	3.21E-07
11	rs143804085	G	RNF169	intron	1.29	1.65E-04	1.53	5.42E-07
11	rs117258637	G	SPCS2	intron	1.27	2.86E-04	1.51	7.24E-07

Adjusting for median WBC removed the association signal at rs2814778 in both the joint and AA group analyses and attenuated the association on chromosome 9. The association between trajectory class membership and rs12094900 was attenuated slightly, after adjustment for median WBC count (p=3.62× 10⁻⁷). Interestingly, adjusting for median WBC count revealed a new region of interest for association with trajectory class membership on chromosome 11q13.4 in both the joint and EA group analyses (1.10 × 10⁻⁷, joint analysis lead SNP rs143804085; 3.21 × 10⁻⁷, EA group analysis lead SNP rs79852880). The regional association plot for the joint analysis shows a large number of SNPs in high LD with rs143804085, with an equivalent level of association, across a four megabase region (Figure 4). This region contains six genes; rs143804085 falls within an intron of RNF169, the five remaining genes are: CHRDL2, MIR4696, XRRA1, SPCS2, and NEU3. The ring finger protein specified by RNF169 is involved with the regulation of DNA double strand break repair⁴⁹. Chordin-like 2 protein, encoded by CHRDL2, associates with members of the TGF-ꞵ superfamily and may play a role in myoblast and osteoblast differentiation and maturation⁵⁰. MIR4696 encodes a microRNA, which are involved in post-translational gene expression regulation⁵¹. The XRRA1 gene product is believed to regulate cell response to X-radiation exposure; the lead SNP in the EA group analysis falls within the XRRA1 gene⁵². SPCS2 encodes a subunit of the microsomal signal peptidase complex, which removes signaling peptides from newly-formed proteins as they move to the endoplasmic reticulum⁵³. Neuraminidases 3, the gene product of NEU3, catabolizes gangliosides in the brain, thereby regulating neuronal function⁵⁴.

Figure 4.

Regional association plot for chromosome 11p13.4, joint analysis.

Discussion

The LCMM we used to identify unobserved heterogeneity in these longitudinal data for deep phenotype discovery identified two distinct latent resting-state WBC count trajectories. The major class trajectory, Class 2, was slightly U-shaped (concave up), with the point of inflection at approximately 60 years of age. The predicted trajectory corresponded well to previous reports of WBC count change over the lifespan^9,10. The Class 1 steady-state WBC count trajectory decreased across the lifespan and may indicate individuals with important differences in inflammation and immune health. While the average posterior probabilities of each class indicate distinct trajectories we discovered by the LCMM, the entropy value of 0.27 suggests a degree of imprecision in the clustering of participants into trajectory classes⁵⁵. Here, the low entropy value likely reflects the crossing in the predicted trajectories. Individuals with most of their data in the region of the crossing have posterior probabilities of class membership close to 50% for both classes and entropy nearing zero, driving the entropy of the whole model down, despite the moderately high average posterior probabilities for trajectory class membership.

Though not free from misclassification, using trajectory class membership predicted by LCMM as our phenotype of interest, we identified three regions associated with longitudinal change in WBC level in our GWAS; two on chromosome 1 at 1p34.3 and 1q23.2, and one at chromosome 9q33.1. Our previous work with this cohort identified WBC count quantitative trait loci in the 1q23.2 region at the ACKR1/DARC locus among those of AA and the 17q21.1 region among those of European ancestry²⁶. In this study, we found the T allele of the rs2814778 polymorphism in the ACKR1/DARC was also associated with decreasing WBC count with age (Class 1 phenotype) in the AA group. Individuals that carry the T allele -- all of our participants of EA and 37% of our participants of AA -- express the Duffy antigen on their RBCs. The Duffy antigen is a chemokine receptor and has been found to preferentially bind inflammatory chemokines⁵⁶. Given the proinflammatory nature of the Duffy antigen, we would expect to see a positive association between the T allele of rs2814778 and the Class 2 phenotype, but this is not the case. Rather, because there was a significant difference in median WBC count between trajectory phenotype classes, we believe the association to Duffy region is reflecting the strong QTL there. Indeed, when we adjusted our analysis for potential confounding by median WBC count, the association between trajectory class membership and rs2814778 disappeared.

Similarly, adjusting for median WBC count removed the suggested association at chromosome 9q33.1. The associated SNPs are in ASTN2. The gene appears to be primarily expressed in the brain, prostate, and testis, with a lower expression level in the adrenal gland. Variation in this gene has been associated with neurological disorders^46–48. Though ASTN2 has also been associated with osteoarthritis, that mechanism was attributable to influences on femur shape⁵⁷. There is no obvious mechanism for an association with median WBC count or trajectory class.

The association with region on 1p34.3 and WBC trajectory was slightly attenuated after adjusting for potential confounding by median WBC count, but the signal remained. Of the genes in this region, CSF3R is the most biologically plausible candidate explaining this association. CSF3R encodes the receptor for the granulocyte colony-stimulating factor (G-CSF) cytokine. The gene which encodes this cytokine, CSF3, is in close proximity to the 17q21.1 locus, which we found to be associated with WBC count in a previous study of this cohort²⁶. G-CSF, working with its receptor expressed on the surface of hematopoietic progenitor cells and neutrophilic granulocytes, stimulates granulopoiesis and activates neutrophils. Deficiency in G-CSF is associated with severe neutropenia, and G-CSF therapy is the major treatment for neutropenia, regardless of cause^58,59.

Adjusting for median WBC count revealed an additional area of interest on chromosome 11q13.4, with several SNPs in high LD across a four megabase region, just above the genome-wide significance threshold. Of the six genes in this region, two (RNF169 and XRRA1) are involved in DNA repair and cell cycle arrest and are the most biologically plausible in relation to WBC count trajectory, as bone marrow, which gives rise to the WBCs, is the most rapidly replacing tissue in the body. In response to DNA damage, RNF169 protein negatively regulates the ubiquitin-dependent signaling cascade for double strand break repair, turning off the DNA damage signal and promoting mitosis after cell recovery⁶⁰. RNF169 is also overexpressed in peripheral blood mononuclear cells, but not granulocytes (www.genecards.org)⁶¹. Similarly, XRRA1 appears to regulate the cell cycle in response to X-radiation. It is expressed in normal tissue, including WBCs, as well as cancer cells⁵². This LD region is associated with decreasing resting-state WBC count over time and polymorphisms within maybe contributing to this phenotype by decreasing mitosis or increasing apoptosis rates in response to cell damage with age.

While we successfully revealed two latent WBC count trajectory phenotypes in our longitudinal EMR-derived data and found novel genetic associations with these trajectories, it is difficult to determine the biological or clinical relevance of these trajectory phenotypes with the available demographic and diagnosis codes in our dataset, given their complex relationships.

We used very little a priori biological knowledge to inform model construction. Incorporating longitudinal measures of BMI and smoking status may improve the precision of discovered trajectory phenotypes and make for easier interpretation of clinical relevance. However, age matched BMI measurements were available for less than half of participants with longitudinal WBC count data and LCMM with multiple imputed datasets is computationally intensive. Smoking status of eMERGE participants was not available. Care should be taken when incorporating longitudinal EHR covariate data into analyses that require “complete” data as this requirement can bias the individuals available for the study⁶². Indeed, requiring a minimum number of longitudinal WBC count data points preferentially excluded younger eMERGE participants and participants with AA.

To further the goals of precision medicine, more precise phenotyping of complex disease is needed. The pattern of change over time can be controlled by loci not apparent when considering cross-sectional data. Though it has limitations, we have shown that trajectory analysis with LCMM is a useful tool to integrate longitudinal measures for deep phenotype construction. This method was also a fruitful way to partition phenotypic heterogeneity, with respect to gene discovery.

Materials and Methods

Participants

Resting-state WBC count data was identified from the EMR of 14 018 participants in the electronic Medical Records and Genomics (eMERGE) Network. Currently, the eMERGE Network is a consortium of twelve U.S. cohorts linked to EMR data for conducting large-scale, high- throughput genetic research⁶³. Our study used a subset of participating sites include the following: (1) Kaiser Permanente Washington (formerly Group Health Cooperative) and University of Washington partnership, Seattle,WA; (2) Marshfield Clinic, Marshfield, WI; (3) Mayo Clinic, Rochester, MN; (4) Northwestern University, Evanston, IL; and (5) Vanderbilt University, Nashville, TN⁶⁴. Participant informed consent was obtained by the recruiting eMERGE site. The study was approved by each site’s internal Institutional Review Board.

The WBC data extraction algorithm used has been previously published²⁶. Generally, we excluded participants and visits where the participant’s diagnosis or medication use may have perturbed WBC count outside of resting- state. To further exclude abnormal WBC counts, we excluded visits where the WBC count was outside two standard deviations of the median, within subject. Additionally, identifying a trajectory-based deep phenotype through LCMM requires that participants have one more WBC count measure than the order of magnitude of the model fitted. Because WBC count appears to be slightly U-shaped over the lifespan, we chose to fit a quadratic model and therefore participants with fewer than three WBC count visits were excluded. A summary of subject- and visit-level exclusion criteria are listed in Table 4. After exclusions, 9 742 participants were available for trajectory model analysis.

Table 4.

Summary of subject- and visit-level exclusions for EMR WBC count data.

Subject-level exclusion criteria
Any indication at any time of HIV
Dialysis at any time
<3 WBC count records
Visit-level exclusion criteria
Inpatient or emergency visit
Splenectomy record prior to lab
Prior diagnosis of myelodysplastic syndrome
Medications with minor impacts on WBC (aspirin at high doses)
Strongly immune affecting medications (oral or IV steroids chemotherapeutic agents such as methotrexate)
Indication of concurrent ‘‘active chemotherapy’’ regimen 6 months prior to 3 months after index visit
Prior indication of Alzheimer’s disease
Blood dyscrasia (leukemia, myeloma, bone marrow failure, aplastic anemia, etc.)
‘‘Active infection’’ in prior or subsequent 30 days
Other acute and chronic infections
Within subject outlier WBC count

Deep phenotype

LCMM was fitted using the R package lcmm^45,65. This method uses a modified Marquardt iterative algorithm and maximum likelihood theory to estimate the LCMM. The Bayesian Information Criterion (BIC), posterior probability of class membership, and percentage of class membership were used to evaluate model fit. When comparing models, that with the lower BIC is the preferred model. An average posterior probability of class membership of ≥ 70% and ≥ 10% of the population assigned to the minor class are also indications that distinct trajectories are being modelled^33,66. We followed the model fitting procedure suggested by Andruff and colleagues, first fitting the mixed model specifying one latent class and then adding classes until the model fit criteria are satisfied⁶⁷. In all models, age at event (i.e. WBC count draw) was our time varying covariate, fit with a linear and quadratic fixed effect and a linear random effect. The lcmm package calculates the posterior probability of class membership for each participant in each class using Bayes Theorem as the probability of class membership given the participant’s data and the model fit⁴⁵. Participants were then assigned membership in a class based on the class for which they had the highest posterior probability of class membership. Participant class membership was used as the phenotype in subsequent analyses. To further characterize the discrimination of the latent classes, we calculated the entropy of each model using the participant posterior probability of class membership, as suggested by van de Schoot and colleagues⁶⁸. Entropy, E_k, was calculated as: $E_k=1-\left({\sum }_i^{⬚}{\sum }_k^{⬚}{\widehat{P}}_{ik}ln{\widehat{P}}_{ik}\right)/\left(IlnK\right)$, for i individuals in k classes. E_k is 0 when the posterior probabilities for all participants are equal, indicating no separation of classes, and 1 when the classes are discrete partitions⁵⁵.

Statistics

We described the characteristics of each class, comparing several covariates. Continuous values were compared using the Kruskal-Wallis rank sum test and categorical values using the chi-squared test. Median age-at-event, WBC count and BMI were first calculated within subject and then the median was calculated for each class, i.e. the grand median. All analyses were performed in R version 3.3.0⁶⁹. R code is available upon request.

Genotyping

As reported previously²⁶, most subjects were genotyped on the Illumina Human660W-Quadv1_A (660 W) genotyping platform. A subset of subjects who were self-reported (Northwestern University) or observer reported (Vanderbilt University) to have African ancestry was genotyped on the Illumina Human1M-Duo (1 M) genotyping platform. Genotyping calls for both platforms were made at CIDR and Broad using BeadStudio version 3.3.7 and Gentrain version 1.0. Both samples and SNPs were assessed for quality and subsequently filtered from the production data, if thresholds were not met⁷⁰. We produced a unified genotype variant dataset for the different genotyping platforms, imputing missing genotype calls using the Michigan Imputation Server with the HRC1.1 haplotype reference set^71–73. Cryptic relatedness was assessed for all sites, and pairs at half-sibling level (kinship coefficient θ= k₁/4 + k₂/2 = 1/8) or higher were randomly broken (by dropping one) before assessing whole-genome association. Subjects identified for filtering at each particular site through the quality control/quality assurance process were subsequently filtered for the entire merged dataset. The eMERGE imputed genotype dataset is available on dbGaP study accession phs001584.v1.p1.

Assigning genetically-determined ancestry (GDA)

Principal components analysis was performed using independent, autosomal SNPs with missing call rates < 5.0% and minor allele frequency > 5.0% across the merged data set of 17 150 unique subjects, as described previously²⁶. We used k-means clustering on the first two principal components, specifying three groups, to assign an individual’s GDA as either European Ancestry (EA), African Ancestry (AA) or Asian Ancestry⁷⁴.

GWAS

GWAS analyses of the WBC count trajectory phenotypes discovered in the phenotyping phase were performed in PLINK⁷⁵. The minor class, Class 1, was set as the “at risk” phenotype. We performed analyses pooling subjects of all genetic ancestries (Joint analysis) as well as analyses stratified by GDA (EA group and AA group only, due to small Asian Ancestry sample size). All analyses were adjusted for sex, median BMI, and principal components 1 and 2, to account for possible confounding by ancestry. Joint and EA group analyses were also adjusted for study site, but this covariate was dropped from the AA group analysis due to collinearity. We assumed an additive genetic model, with SNP genotypes were coded as 0, 1 and 2 copies of the minor allele. We filtered out SNPs with a minor allele frequency of < 0.03. Manhattan and Q-Q plots we made using the GWASTools R package⁷⁶. Regional association plots were generated by LocusZoom⁷⁷.