A Genome-Wide Association Study of Dementia Using the Electronic Medical Record

Abstract

Dementia is characterized as a decline in cognitive function, including memory, language and problem-solving abilities. In this paper, we conducted a Genome-Wide Association Study (GWAS) using data from the electronic Medical Records and Genomics (eMERGE) network. This study has two aims, (1) to investigate the genetic mechanism of dementia and (2) to discuss multiple p-value thresholds used to address multiple testing issues. Using the genome-wide significant threshold ($p\le 5\times {10}^{-8}$), we identified four SNPs. Controlling the False Positive Rate (FDR) level below $0.05$ leads to one extra SNP. Five SNPs that we found are also supported by QQ-plot comparing observed p-values with expected p-values. All these five SNPs belong to the TOMM40 gene on chromosome 19. Other published studies independently validate the relationship between TOMM40 and dementia. Some published studies use a relaxed threshold ($p\le 1\times {10}^{-5}$) to discover SNPs when the statistical power is insufficient. This relaxed threshold is more powerful but cannot properly control false positives in multiple testing. We identified 13 SNPs using this threshold, which led to the discovery of extra genes (such as ATP10A-DT and PTPRM). Other published studies reported these genes as related to brain development or neuro-development, indicating these genes are potential novel genes for dementia. Those novel potential loci and genes may help identify targets for developing new therapies. However, we suggest using them with caution since they are discovered without proper false positive control.

1. Introduction

Dementia is a term used to describe a decline in cognitive function, including memory, language and problem-solving abilities [1]. Dementia affects millions among the ageing population, and the probability of having a form of the condition increases with age [2]. Approximately one-third of people aged 85 or older are likely to have dementia. It is also a leading cause of disability and death among older adults and a significant burden on public health [3].

The cause of dementia is complex. Both genetic and environmental factors can influence dementia [4]. Scientists believe that genetic factors account for about 60% of the risk of developing dementia. Many genetic risk factors for dementia have been identified in previous research, including specific genetic mutations [5,6]. Some studies have shown an association between the risk genes and dementia, such as the APOE gene and Alzheimer’s disease (the most common form of dementia), as demonstrated by Strittmatter et al. [7] in 1993. Cervantes et al. [8] proposed that TOMM40, as an APOE cluster gene, is a risk of Alzheimer’s disease. Liu et al. [9] investigated the association of the SORL1 gene expression with Alzheimer’s disease. However, the genetic basis of most cases of dementia is not fully comprehended, leading to ongoing research in this area.

Genome-Wide Association Study (GWAS) is a powerful tool for identifying genetic variants associated with specific traits or diseases. GWAS can identify genetic variants that are more common in populations with specific traits or diseases than in the general population by analyzing patterns of variation in large numbers of individual genomes [10]. Most GWAS studies test the association between disease outcome and each individual SNP, one by one. These tests lead to numerous p-values, while multiple testing adjustment is critical to control false positives in many tests. The two most popular multiple-testing approaches for GWAS are the genome-wide significant threshold ($p\le 5\times {10}^{-8}$) [11] and false discovery rate (FDR) [12]. Many GWAS studies suffer from a lack of power issue, which has led to the development and applications of many novel approaches. Such methods often have various limitations or require additional assumptions. For example, Xu et al. [13] proposed a pseudo-supervised machine learning approach for discovering SNPs, which utilizes information on the SNP-SNP relationship to achieve better power while properly controlling FDR. However, this method cannot handle covariates, which largely limits its applications when covariates play an essential role in affecting disease outcomes and must be adjusted in genomic data analysis. Alternatively, some studies reported SNPs without properly controlling false positives. Authors used subjectively decided p-value thresholds in these studies ($p\le 1\times {10}^{-5}$) [14,15]. Such an approach is useful since it enables researchers to discover potentially interesting signals when the sample size is insufficient to achieve decent statistical power while controlling false positives properly. However, this approach should be used with extreme caution and its results should be properly interpreted to highlight the fact that there is no proper FDR control.

In this study, we have two aims. Firstly, we aim to provide insights into the biological mechanism underlying the development of dementia and help identify potential targets for developing new therapies. Secondly, we discuss three different p-value thresholds used to address multiple testing problems, which can help to discover SNPs when the statistical power is insufficient. To achieve the goals, we conducted a GWAS on the data of a published study from the electronic Medical Records and Genomics (eMERGE) network. The eMERGE is a consortium of biobanks linked to electronic medical record data and funded by NHGRI for genomic research in electronic medical record (EMR) systems [16]. The eMERGE actively tracks dementia investigations and many participants have lived to advanced ages under continuous observation, which provides data to support our study. We used the genomic data obtained from the consent group of Disease-Specific (Dementia). In our GWAS study, we apply and compare three discovering SNP approaches to investigate the genetic variants associated with dementia. Based on the difference in the results of these three approaches, we discuss their strength and weakness.

The rest of this paper is organized as below. Section 2 introduces the database, patient cohort, quality control and statistical analysis. Section 3 describes the results of covariates in the GWAS model and exploratory GWAS of dementia. Section 4 presents our discussion and conclusions.

2. Materials and Methods

2.1. Database

The Electronic Medical Records and Genomics (eMERGE) Network (https://emerge-network.org/, accessed on 11 December 2022) is a consortium of ten participating sites to develop, disseminate and apply methods that combine DNA biorepositories with EMR systems for large-scale, high-throughput genetic studies [17]. Using electronic phenotyping methods, the consortium used DNA samples from all participating sites to explore the genetic determinants of over forty phenotypes, including dementia. We used the data of the consent group of Disease-Specific (Dementia).

2.2. Patient Cohort

We selected participants with five or more visits. Diseases with any of the following conditions were defined by this study as a case group for dementia. They were senile dementia, presenile dementia, senile dementia delusional (paranoid) features, senile dementia with depressive features, senile dementia with delirium or confusion, arteriosclerotic dementia, vascular dementia with delirium, vascular dementia with delusions, dementia due to alcohol, dementia due to drugs, Alzheimer’s disease, pick’s disease of the brain, other frontotemporal dementia, dementia with Lewy bodies and dementia with Parkinsonism.

2.3. Quality Control

We performed a series of QC steps on genotype data provided by 1433 participants (578 cases and 855 controls before quality control). Missing genotypes introduce bias and reduce the analysis’s efficacy. Thus, we screened out variants with more than 20% missing individuals and excluded individuals with more than 20% missing variants. Moreover, we filtered the variants and individuals using a more stringent threshold (0.02; > 2%); 4,569,604 variants and 34 individuals were removed. Furthermore, SNPs are those variations with a minor allele frequency (MAF) greater than 1% [18]. In this paper, we used 0.01 as the MAF threshold and removed 27,994,441 variants. The variants that passed the threshold were SNPs. We also filtered out SNPs with HWE $p\le 1\times {10}^{-10}$, which excluded 26631 SNPs. These filters were applied separately to each of these 22 chromosomes to remove poorly performing samples and variants/SNPs using tools implemented in PLINK [19].

After samples and marker quality control, the total genotyping rate is 0.994328. There are 5,449,492 variants and 1399 participants (559 cases and 840 controls) pass filters and QC.

2.4. Statistical Method

SNPs were tested for association with dementia in PLINK using logistic regression analysis that assumed an additive genetic model. We summarized the patient characteristics in the disease group and the control group to select the significant variables ($p<0.05$). Meanwhile, population structure could cause confounding in GWAS, which may produce spurious associations if we do not properly process it. We addressed this problem by including principal components (PCs) as covariates. In our GWAS model, we simultaneously input the significant variables and the top 10 PCs as covariates.

We selected three different discovering SNP approaches to identify the significant SNPs: (1) Genome-wide significant threshold ($p\le 5\times {10}^{-8}$), which was a popular criteria. (2) Benjamini–Hochberg method was applied to adjusting the false discovery rate (FDR). We set the cut-off of FDR adjusted p-value as $0.05$. (3) We use a relaxed threshold, $p\le 1\times {10}^{-5}$, to discover more SNPs but sacrifice proper false positive control.

3. Results

3.1. Important Non-Genomics Factors Affecting Risk of Dementia

To identify important covariates to be adjusted in GWAS analysis, we investigated the relationship between dementia disease status and some selected patient characteristics. The characteristics of the case-control groups are shown in

Table 1. Demographic and clinical characteristics of the patients (full analysis population).

Characteristic	Case (N = 578)	Control (N = 855)	p-Value
Gender (%)			$1.26\times {10}^{-3}$
Female	346 (59.86)	438 (51.22)
Male	232 (40.14)	417 (48.77)
Birth year (%)			<$2\times {10}^{-16}$
Birth in 1900–1919	451 (78.03)	486 (56.84)
Birth in 1920–1929	126 (21.80)	352 (41.16)
Birth in 1930–1939	1 (0.17)	17 (1.99)
BMI (Mean ± SD)	25.51 (±4.62)	26.43 (±4.73)	$7.11\times {10}^{-4}$
Smoke Status (%)			0.46
Never	72 (43.11)	223 (37.99)
Current	9 (5.39)	39 (6.64)
Past	86 (54.50)	325 (55.37)
STC (Mean ± SD)	232.41 (±42.37)	229.47 (±43.10)	0.26
H-PSA (Mean ± SD)	4.17 (±6.18)	3.63 (±4.43)	0.53
Race (%)			0.34
Black or African American	31 (5.37)	32 (3.74)
Unknown	11 (1.90)	10 (1.17)
American Indian or Alaska Native	2 (0.35)	1 (0.12)
Asian	11 (1.90)	19 (2.22)
White	523 (90.48)	793 (92.70)

BMI = Body Mass Index; H-PSA = Highest record of PS (prostate specific antigen); STC = Serum total cholesterol.

. The proportion of females is 59.86% in dementia patients, which is higher than males (40.14%). However, it is nearly the same proportion in the control group (51.22% and 48.77%). The proportion of patients with advanced age in the case group is 78.03%, higher than those born in 1920–1929 and 1930–1939. We can see the mean value and standard error values of BMI. The mean $(\pm SD)$ BMI of the case group is 25.51 $(\pm 4.62)$, which is lower than that of the control group $\left(26.43\right(\pm 4.73\left)\right)$. These three variables, the birth year, gender and BMI, have statistically significant differences between the case and control groups ($p\le 0.05$). In addition to the significant variables, we also considered smoke status, STC (serum total cholesterol), H-PSA (highest record of prostate-specific antigen) and race variables. These four variables were not significant.

Birth year, gender and BMI show a strong relationship with dementia, which are visualized in Figure 1. There were no participants born in 1930–1939 in the case group of dementia in the final filtered data and the proportion of those born in 1900–1919 was higher than in 1920–1929. The birth year and gender variables passed Pearson’s Chi-squared test ($p=4.50\times {10}^{-12}$, $p=1.94\times {10}^{-05}$) between the case and control groups. The patients had a lower median BMI (24.9) than the control group (25.9). Meanwhile, the BMI variable passed the Wilcoxon test ($p=2.35\times {10}^{-4}$). Based on this exploratory data analysis result, we decided to include birth year, gender and BMI, as well as the top 10 PCs as covariates in the logistic regression of GWAS analysis.

3.2. GWAS of Dementia

We performed a GWAS analysis of our collected data after merging and filtering with the non-genomic data. The ability of GWAS to identify genetic associations depends on the overall quality of the data. To avoid false negative and false positive associations, we performed quality control procedures on the data to explore true genetic associations [20]. First, we filtered out the variants with missing individuals of more than 20%, which resulted in 533,207 variants being removed. Secondly, when we set the threshold of missing sample rate to 20%, no samples were deleted. Then, we used a more stringent threshold filtering (2%) to filter the variants and individuals. A total of 4,036,397 variants and 34 individuals were removed. Furthermore, 27,994,441 variants were removed due to a minor allele less than the threshold (0.01). Meanwhile, 5,476,123 variants were maintained as SNPs. Finally, 26,631 SNPs were removed due to the Hardy–Weinberg exact test (HWE $p\le 1\times {10}^{-10}$). Therefore, after samples and marker quality control, the total genotyping rate was 0.994328. A total of 5,449,492 SNPs and 1399 participants passed the filters and QC. Among the remaining phenotypes, 559 were cases and 840 were controls. We performed a GWAS of dementia on the genome data after QC to identify significant SNPs exploring novel treatments.

We fit logistic regression to test the association between dementia disease status and every individual SNP. The birth year, gender, BMI and the top 10 PCs were used as covariates in logistic regression models. The p-value of the SNP coefficient in each logistic regression represents whether an SNP significantly affects the risk of dementia. These logistic regression models lead to 5,449,492 p-values for all SNPs that passed QC. The negative-log transferred p-values are visualized in the Manhattan plot Figure 2. Using the genome-wide significant threshold ($p\le 5\times {10}^{-8}$), we identified four significant SNPs (rs11556505, $p=3.536\times {10}^{-11}$; rs2075650, $p=4.394\times {10}^{-11}$; rs34404554, $p=8.368\times {10}^{-11}$; rs71352238, $p=1.229\times {10}^{-11}$). Those four SNPs were represented as green dots in Figure 2. The number of the discovered SNPs was improved when we used another popular decision rule, FDR, which we identified one more SNP, rs34095326 ($FDR=4.860\times {10}^{-3}<0.05$). This SNP was represented using a red dot as in Figure 2. All these five SNPs are located on chromosome 19 and belong to the same gene, TOMM40. The relationship between TOMM40 and dementia is reported in other studies, such as [21], which serve as independent evidence of our findings.

In our study, we found that the p-values largely adhered to the expected p-values until the deviation of the five SNPs at the right-hand side tail, as shown in Figure 3. The expected p-values are calculated by assuming no SNPs are associated with the risk of dementia. Hence, these five SNPs’ deviation indicates a strong signal of associations and a low probability of false positive results. We confirmed that the five points at the tail above the diagonal are identical to the five SNPs identified by the FDR threshold. This provides independent support to find five significant SNPs (i.e., FDR rule) instead of four SNPs (i.e., the genome-wide significant threshold rule). Given our evidence, we conclude five SNPs are significantly associated with dementia with properly controlled false positives.

This study is not very well powered, given its sample size and the number of tests to be conducted. So, we decided to explore more potential signal (non-significant) SNPs or genes, sacrificing proper false positive control. Using another threshold $p\le 1\times {10}^{-5}$ (used by other published studies [14,15]) as the decision rule, we identify13 SNPs. These SNPs are located on 5 genes, including TOMM40, ATP10A-DT, PTPRM, MED21 and two undefined genes (details can be found in

Table 2. Detailed information of significant SNPs with $p\le 1\times {10}^{-05}$(sorted by p-value from smallest to largest). We marked the genes associated with dementia using (*) and a series of novel genes related to brain development or neurodevelopment using (-).

	SNPID	CHR	p-Value	FDR	BP	GENE
1	rs11556505	19	$3.536\times {10}^{-11}$	$1.852\times {10}^{-06}$	45396144	TOMM40 *
2	rs2075650	19	$4.394\times {10}^{-11}$	$1.852\times {10}^{-06}$	45,395,619	TOMM40 *
3	rs34404554	19	$8.368\times {10}^{-11}$	$2.352\times {10}^{-06}$	45,395,909	TOMM40 *
4	rs71352238	19	$1.229\times {10}^{-10}$	$2.590\times {10}^{-06}$	45,394,336	TOMM40 *
5	rs34095326	19	$2.882\times {10}^{-07}$	$4.860\times {10}^{-03}$	45,395,844	TOMM40 *
6	rs72689267	15	$2.899\times {10}^{-06}$	$4.182\times {10}^{-01}$	26,117,761	ATP10A-DT -
7	rs668168	18	$6.111\times {10}^{-06}$	$5.216\times {10}^{-01}$	8,392,719	PTPRM -
8	rs144822097	12	$7.455\times {10}^{-06}$	$9.956\times {10}^{-01}$	27,183,821	MED21
9	rs670305	18	$8.009\times {10}^{-06}$	$5.216\times {10}^{-01}$	8,392,750	PTPRM -
10	rs7178765	15	$8.222\times {10}^{-06}$	$5.929\times {10}^{-01}$	26,121,173	ATP10A-DT -
11	rs4785108	16	$9.097\times {10}^{-06}$	$4.489\times {10}^{-01}$	60,446,014	LOC101927605
12	rs9888985	16	$9.795\times {10}^{-06}$	$4.489\times {10}^{-01}$	60,427,440	LOC101927605
13	rs10458022	6	$9.910\times {10}^{-06}$	$9.999\times {10}^{-01}$	58,308,335	LOC101927293

CHR = chromosome; BP = base–pair position; * it has been confirmed that this gene is associated with dementia; ^- it has been confirmed that this gene is related to brain development or neurodevelopmental disorders.

). We represent the extra SNPs identified using threshold $p\le 1\times {10}^{-5}$ using blue dots in Figure 2.

Among the genes identified using a relaxed p-value threshold, two genes were confirmed related to brain development and neurodevelopment. In a genetic study involving intellectual disability, autism and psychosis, ATP10A was identified as a gene that may affect neurodevelopmental disorders [22]. Therefore, we investigated these genes to explore their association with dementia. We found that PTPRM was a crucial gene involved in the formation of synapses regulated by zinc ions, which was related to the transmission of information in the brain [23]. The details of those genes are shown in

Table 3. The function of the novel significant genes. ATP10A-DT and PTPRM are related to brain development and neurodevelopment. The third column shows the reference that supports the gene function.

	Gene	Function	Reference
1	MED21	an enzyme in humans	[24]
2	ATP10A-DT	it can affect neurodevelopmental disorders	[22]
3	PTPRM	it involved in the formation of synapses regulated by zinc ions, which is related to the transmission of information in the brain.	[23]

.

Note that we need to highlight that the new SNPs or genes discovered with relaxed threshold ${10}^{-5}$ should be used with extreme caution since false positives are not properly controlled. We highly suggest using validation studies to confirm such relationships before using these SNPs and genes for critical decisions.

4. Discussions and Conclusions

We explored the biological mechanism underlying the development of dementia by conducting a GWAS of dementia and discussed three different p-value thresholds used to address multiple testing problems.

We investigated the relationship between dementia and patients’ characteristics and revealed three significant factors, birth year, gender and BMI, which are confirmed in the literature. For example, recent studies have suggested that ageing and gender are risk factors for dementia [25,26] and our study provides further validation of the hypothesis. The relationship between BMI and dementia is controversial in the literature. Our analysis found a lower BMI increases the probability of dementia in a cohort of patients with a normal BMI range (mean 25.51 and standard deviation 4.62). This result does not align with our intuition but is supported by other studies (e.g., [27]). Furthermore, we obtained the top 10 PCs to address the confound due to population structure. We input the three significant variables and the top 10 PCs into the GWAS model as covariates.

We applied and compared three discovering SNP approaches to investigate the genetic variants associated with dementia. Based on the difference in the results of these three approaches, we discuss their strength and weakness. In our analysis, we found that the genome-wide significant threshold $p\le 5\times {10}^{-8}$ is the least powerful approach, which discovered four SNPs on the TOMM40 gene located in Chromosome 19. The FDR adjustment approach is more powerful while keeping false positives of the study under control, which can discover one more significant SNP in the same gene. Using the threshold of $p\le 1\times {10}^{-5}$, we can obtain 13 SNPs on a few uncharacterized locations on chromosome 2 and on several other genes. Among these genes, only MED21 is not discussed in related literature, which might be a false hit or a novel discovery. All other genes were reported to be associated with brain development or neuro-development, supported by the literature [22,23]. We believe these neuro-development genes are likely to be real dementia-related genes since neurological and neuropsychiatric matters are the primary causes of dementia [28].

Based on the different results of three different discovering SNP approaches in this study, we suggest using FDR adjustment if the false positives need to be properly controlled. However, GWAS studies often suffer from the issue of lack of power caused by the curse of dimensionality [29]. Investigators often cannot afford to study enough samples to address the multiple-testing issues when a huge number of tests need to be conducted in GWAS. In this situation, using a relaxed threshold, such as ${10}^{-5}$, could help us find more potential signals. However, this approach needs to be used with extreme caution because it cannot properly control false positives in the study. When reporting these SNPs, authors should highlight the fact of no proper false positive control. We strongly suggest conducting validation studies to validate the SNPs discovered with a relaxed threshold if these SNPs are used to make important decisions.

In summary, we investigated three decision rules to discover significant SNPs for GWAS analysis and discussed their strengths and weaknesses. Our analysis results confirmed the significant associations of variants in TOMM40 with dementia. We discovered potential novel dementia SNPs, which were reported to be associated with brain development or neuro-development and novel SNPs with no related literature discussion. These findings reveal the genetic mechanism of dementia and may provide opportunities for identifying novel dementia treatment.