Whole genome sequence analyses of brain imaging measures in the Framingham Study.

OBJECTIVE: We sought to identify rare variants influencing brain imaging phenotypes in the Framingham Heart Study by performing whole genome sequence association analyses within the Trans-Omics for Precision Medicine Program.
METHODS: We performed association analyses of cerebral and hippocampal volumes and white matter hyperintensity (WMH) in up to 2,180 individuals by testing the association of rank-normalized residuals from mixed-effect linear regression models adjusted for sex, age, and total intracranial volume with individual variants while accounting for familial relatedness. We conducted gene-based tests for rare variants using (1) a sliding-window approach, (2) a selection of functional exonic variants, or (3) all variants.
RESULTS: We detected new loci in 1p21 for cerebral volume (minor allele frequency [MAF] 0.005, p = 10-8) and in 16q23 for hippocampal volume (MAF 0.05, p = 2.7 × 10-8). Previously identified associations in 12q24 for hippocampal volume (rs7294919, p = 4.4 × 10-4) and in 17q25 for WMH (rs7214628, p = 2.0 × 10-3) were confirmed. Gene-based tests detected associations (p ≤ 2.3 × 10-6) in new loci for cerebral (5q13, 8p12, 9q31, 13q12-q13, 15q24, 17q12, 19q13) and hippocampal volumes (2p12) and WMH (3q13, 4p15) including Alzheimer disease- (UNC5D) and Parkinson disease-associated genes (GBA). Pathway analyses evidenced enrichment of associated genes in immunity, inflammation, and Alzheimer disease and Parkinson disease pathways.
CONCLUSIONS: Whole genome sequence-wide search reveals intriguing new loci associated with brain measures. Replication of novel loci is needed to confirm these findings.

Brain imaging phenotypes (cerebral and hippocampal atrophy or white matter hyperintensity [WMH] burden) are accepted endophenotypes of Alzheimer disease (AD) and vascular brain injury.[1] Identifying genetic loci that influence these measures could lead to the discovery of new biological mechanisms underlying these diseases. Recently, large genome-wide association studies (GWAS) have successfully identified and replicated associations between genetic variants and brain imaging phenotypes.[2-7]

Two regions were consistently reported by association studies (17q25, WMH and 12q24, hippocampal atrophy).[2,4,6,7] The 17q25 region was originally discovered by a GWAS of WMH in stroke-free European individuals from the Cohorts for Heart and Aging Research in Genomic Epidemiology (CHARGE) consortium (n = 9,361) with replication in 2 cohorts (n = 3,024).[7] This locus encompasses 6 genes (WBP2, TRIM65, TRIM47, MRPL38, FBF1, and ACOX1).[7] The 12q24 region was initially reported by a GWAS of dementia-free individuals from the CHARGE consortium (n = 9,232) with replication in 2 samples (n = 2,318) and external validation in the Enhancing Neuroimaging Genetics through Meta-Analysis (ENIGMA) consortium (n = 7,794). These associations implicate genes related to apoptosis (HRK) and ubiquitination (FBXW8).[2]

These prior GWAS have identified common genetic variants (minor allele frequency [MAF] ≥5%) with modest effect sizes that resided mostly in noncoding regions. Besides, the causal variant and the functional basis of associations are unclear at many loci. Thus, more detailed scans of low frequency and rare variants (MAF ≤5%) available from whole genome sequence (WGS) are needed, particularly in GWAS loci and coding regions.

GWAS are performed using single nucleotide variants (SNVs) that are available on commercial genotyping chips or in reference panels. They focus on SNVs that are common in a population (MAF ≥5%). WGS can capture other types of genetic variations such as repeats or indels (insertions/deletions) and have a finer resolution compared to GWAS. Indeed, all genetic variations from an individual can be detected, including rare and individual-specific variations. Thus, WGS association analyses can help to detect new signals that could have been missed by GWAS.

In this study, we sought to identify low-frequency and rare variants associated with brain measures in the Framingham Heart Study by performing WGS association analyses within the Trans-Omics for Precision Medicine Program (TOPMed).

Methods

Description of the Framingham Heart Study

We included participants from the 3 generations of the Framingham Heart Study (FHS). Briefly, the FHS is a prospective, population-based study that began in 1948 to study the determinants of cardiovascular disease. The study has followed participants from the town of Framingham, Massachusetts, whose population was almost entirely white at the beginning of the study. The first generation (original cohort/Gen1) has been followed since 1948 and included 5,209 participants; survivors are still receiving examinations biannualy.[8] The second generation (offspring cohort/Gen2) has been followed since 1971 and comprises 5,124 offspring and spouses of the offspring, including 3,514 biological offspring; they have received examinations once every 4–8 years.[9] The third generation (Gen3) was enrolled in 2002 and included 4,095 children from the largest families of the offspring cohort; they have received 2 examinations 4 years apart and a third examination is currently underway.[10] All cohorts are under active surveillance for cardiovascular events.

Attendees of the 26th Gen1 and 7th Gen2 examinations were invited to participate in brain MRI between March 1999 and June 2005; attendees of the 2nd Gen3 examination were similarly invited to undergo brain MRI from 2009. Among 4,772 individuals with brain MRI data, 96 participants were excluded for stroke or TIA, 73 for dementia, and 102 because of other neurologic conditions such as multiple sclerosis, meningiomas, primary or metastatic brain tumors, or significant head trauma. The remaining 4,501 participants constitute our sample for this study.

Standard protocol approvals, registrations, and patient consents

All participants provided written informed consent. This study was approved by the institutional review board of the Boston University Medical Center.

MRI protocol and phenotyping

Brain MRI acquisition measures and image processing methods have been described in detail elsewhere.[11] Briefly, participants were imaged on a 1T (1999–2005) or 1.5T (after 2005) Magnetom scanner (Siemens Medical, Erlangen, Germany). We used 3D T1-weighted coronal spoiled gradient-recalled echo acquisition scans for all participants. In addition, we used T2 fluid-attenuated inversion recovery (FLAIR) sequences for scans acquired after 2005. All MRI were transferred to the University of California–Davis Medical Center for centralized reading. We used QUANTA 6.2 (Ultra 5 workstation, Sun Microsystems, Santa Clara, CA) for image analyses and interpretation was performed blinded to participants' demographic and clinical characteristics in random order.

The semiautomated segmentation protocols for quantifying total intracranial volume (TCV), total cerebral brain volume (TCBV), hippocampal volume (HPV), and WMH have been described elsewhere.[12] Briefly, TCV was determined by outlining the intracranial vault lying above the tentorium.[13] To accurately distinguish CSF from brain matter, we used a semiautomated analysis of MRI pixel distributions for CSF, gray matter, and white matter.[14] HPV was computed by a semiautomatic multiatlas hippocampal segmentation algorithm.[15] For analyses of WMH, we included the subsample of participants who were imaged on a 1.5T scanner and had FLAIR sequences, and used a semiautomated procedure[13] with high inter-rater reliability[16] for the segmentation and quantification of WMH on FLAIR. TCBV is expressed relative to TCV to account for differences in head size. No substantial bias was observed when compared participants with and without MRI.[11]

Whole genome sequencing

WGS was part of the National Heart, Lung and Blood Institute's TOPMed (nhlbiwgs.org) that serves as an initial step for a larger Precision Medicine Initiative. TOPMed phase I consisted of 11 different studies covering a range of heart, lung, blood, and sleep disorders phenotypes and a total of ∼20,000 samples. Phase I began in October 2014 and completion of sequencing production was done in February 2016. The FHS is part of this phase I with 4,148 sequenced individuals (372 Gen1, 2,194 Gen2, and 1,582 Gen3). The samples were sequenced at >×30 depth of coverage at the Broad Institute of the Massachusetts Institute of Technology and Harvard. Individual genetic variations across the genome were identified in a joint calling of all samples performed by the TOPMed Informatics Resource Center (University of Michigan). Centralized read mapping and genotype calling, along with variant quality metrics and filtering of variants and samples that failed to meet these quality metrics, was also completed by the TOPMed Informatics Research Center. Phenotype harmonization, data management, quality control (QC) to ensure correct sample identification, and general study coordination were provided by the TOPMed Data Coordinating Center. Methods for WGS data acquisition and QC are described in a document that is publicly available on the Database of Genotype and Phenotype (dbGaP) website (ncbi.nlm.nih.gov/projects/gap/cgi-bin/GetPdf.cgi?id=phd006969.1). We excluded variants with a read depth less than 10.

Statistical analysis

We used mixed-effect linear regression models to examine associations between WGS SNVs and brain MRI measures. We adjusted models for age, sex, and TCV for WMH, age, sex, age2, and TCV for HPV and age, sex, and age2 for TCBV. Brain volumes and WMH are related to TCV, and vary with age and between men and women. We created the residuals of the original phenotypes adjusted for covariates and inverse normal transformed them. Then, we tested the association between rank-normalized residuals from these models and individual SNVs while taking into account familial relatedness using an empirical kinship matrix based on analysis of actual genotype similarities between participants. We filtered the results including only variants with an allele count of 10 or greater and used a threshold of p = 1.5 × 10−8 to consider an SNV association as genome-wide significant.[17]

In further analyses, we performed conditional analyses on the previous reported associations in 2 GWAS loci: 12q24 for HPV and 17q25 for WMH.

Within each gene, we also conducted gene-based tests to test the association between rare variants (MAF ≤1%) and MRI measures using 2 different multimarker methods (i.e., sequence kernel association test [SKAT] and burden test [T1]).[18] These methods aggregate individual score test statistics of all rare genetic variations in a gene or a region. SKAT tests are robust to the presence of rare variants with either risk-increasing or risk-decreasing effects. In the T1 test, the summed aggregate effect is considered and only rare variants (MAF <1%) are included to assign a score to each gene or region, and hence this test is most powerful when all variants are either increasing or decreasing risk.

In order to improve our power to detect associations, we used different filters to select only SNVs in a gene that were most likely to affect the phenotype as functional exonic SNVs (missense or loss of function) or all rare SNVs. We also used a sliding-window approach[19] in which the gene length was divided into discrete regions or “windows” and gene window-based test was performed within each region; this approach improves the power to detect effects if only one region within the gene is affecting the phenotype. We used a window size of 4 kb with an overlap of 2 kb between adjacent windows and we filtered the results according to a burden count of 10 or greater. We used a threshold of p = 2.5 × 10−6 to declare a gene association as genome-wide significant (Bonferroni correction for number of genes tested, 0.05/20,000). For the sliding-window approach, we used a more stringent threshold of p = 1.8 × 10−7 to declare a gene association as genome-wide significant to account for overlapping windows (correction for the number of tests performed in each gene, 0.05/279,713). We performed functional annotations of the SNVs with the publicly available variant function predicting software Annovar.[20] All association analyses were carried out using EPACTS 3.2.6 software (University of Michigan) with the EMMAX test. Finally, we used the Meta-Analysis Gene-set Enrichment of Variant Associations (MAGENTA) method to explore pathway-based associations using single-SNV association results. MAGENTA implements a gene set enrichment analysis-based approach, as previously described.[21] In this method, a gene score is calculated for each gene based on the SNV with the lower association p value in a 110 kb upstream, 40 kb downstream window around the gene. This score is corrected for confounding factors (gene size, SNV density, or linkage disequilibrium [LD] between SNVs in the gene) and each gene is ranked on its score. In each pathway, an empirical p value is calculated corresponding to the observed number of genes with a rank above a given significance threshold (95th or 75th percentiles of all gene scores) compared to 106 randomly permuted pathways of the same size.

In total, 10,992 pathways from the Gene Ontology, PANTHER, KEGG, Ingenuity, Reactome, and Biocarta databases were tested for enrichment of associations with each brain MRI phenotype.

Results

We included a total of 2180, 2,170, and 1,667 individuals from FHS in the WGS association analyses of TCBV, HPV, and WMH, respectively (table 1). Most participants were women (∼55%) and the mean age (SD) of individuals was 61.8 years (13.6) for TCBV and HPV and 59.7 years (13.3) for WMH.

Table 1

Main characteristics of the individuals included in the whole genome sequence association analyses of total cerebral brain volume (TCBV), hippocampal volume (HPV), and white matter hyperintensity (WMH)

SNV associations

QQ plots and Manhattan plots for the WGS association analyses of TCBV, HPV, and WMH are presented in figures e-1 and e-2 (links.lww.com/WNL/A32). WGS association analyses revealed at the genome-wide level a new locus in 1p21 for TCBV located between the RWDD3 and PTBP2 genes (top SNV rs181221422, MAF = 0.005, p = 10−8, table 2). A second new locus in 16q23 for HPV within LOC102724084 (top SNV rs9921114, MAF = 0.05, p = 2.7 × 10−8, table 2) was close to the genome-wide level. We confirmed previously identified associations in 12q24 for HPV (rs7294919, MAF = 0.11, p = 4.4 × 10−4) and in 17q25 for WMH (rs7214628, MAF = 0.19, p = 2.0 × 10−3) (table e-1, links.lww.com/WNL/A33). Our best 12q24 HPV association was rs7132910 (MAF = 0.15, p = 3.4 × 10−5) in modest LD with the previously reported association (r2 = 0.49, D' = 0.94 with rs7294919). Conditional analysis revealed that the association observed with rs7132910 was distinct from the one observed with rs7294919 (p = 1.2 × 10−5). The SNP rs7132910 lies at 1.4 kb from the 5′ region of HRK gene and it is an expression quantitative trait locus (eQTL) for HRK and FBXW8 genes in blood.[22] Strong promoter histone marks are described in brain tissues.[22] Our best 17q25 WMH association was rs9889965 (MAF = 0.15, p = 1.2 × 10−6) in modest LD with the previously reported association (r2 = 0.68, D' = 0.88 with rs7214628). Conditional analysis revealed that the association observed with rs9889965 was distinct from the one observed with rs7214628 (p = 3.0 × 10−5). The SNP rs9889965 lies at 113 bp from the 5′ region of TRIM47 gene and it is an eQTL for TRIM65 gene in nerve tibial tissue and TRIM47 in brain, skin, and blood tissues.[22] Strong active transcription starting site histone marks are described in brain tissues.[22] Single-SNV results in 12q24 and 17q25 regions are presented in table e-1 and regional plots are presented in figure 1.

Table 2

Single nucleotide variants (SNVs) in main loci associated with total cerebral brain volume (TCBV) and hippocampal volume (HPV) in single-variant analyses at p ≤ 5 × 10−8

Figure 1

Regional association plots

Regional association plots of the 12q24 hippocampal volume (HPV) (A) and the 17q25 white matter hyperintensity (WMH) (B) regions using Locuszoom software. Single nucleotide variants (SNVs) are plotted with their −log10 (p values) on the left y-axis as a function of the genomic position on the x-axis. Estimated recombination rates on the right y-axis, taken from 1000 Genomes with European panel, are plotted to reflect the local linkage disequilibrium structure around the top associated SNV in purple and correlated proxies, according to a blue to red scale from r2 = 0 to 1.

Gene-based tests

The main gene-based test results are presented in table 3 and regional plots for each gene are provided in figure e-3 (links.lww.com/WNL/A32). Using a sliding-window approach within genes, we detected genome-wide or suggestive associations in new loci for TCBV with the SKAT test (FCHO2-5q13, UNC5D-8p12, CCDC33-15q24) and with the burden test (C9orf84-9q31, UBL3-13q12-q13, SYT3-19q13), for HPV with the SKAT test (LRRTM4-2p12), and for WMH with the burden test (ALCAM-3q13, GBA3-4p15). Using all SNVs within genes, we detected SOCS7 (17q12) significantly associated with TCBV using the SKAT test. No significant results were found when selecting functional exonic SNVs only.

Table 3

Top genes associated with total cerebral brain volume (TCBV), hippocampal volume (HPV), and white matter hyperintensity (WMH) in gene-based tests at p ≤ 2.5 × 10−6

Pathway analyses

MAGENTA analyses revealed interesting pathways associated with brain MRI phenotypes (table e-2, links.lww.com/WNL/A33). Some of them are linked to immunity or inflammation (B lymphocyte pathway, interleukin-4 and interleukin-6 signaling, antigen presentation pathway) or to AD pathology such as the presenilin pathway. These analyses also confirmed the importance of the ubiquitin proteasome pathway that was found associated with WMH (p = 4 × 10−4, false discovery rate = 0.04 for the 75th percentile enrichment cutoff).

Discussion

We investigated low-frequency or rare SNVs influencing brain MRI phenotypes by performing WGS association analyses in the FHS within the TOPMed.

Using single-variant analyses, we identified rare or low-frequency variants in 1p21 (TCBV) and 16q23 (HPV). In 1p21, SNVs were located at 159 kb from PTBP2 (polypyrimidine tract binding protein 2), a brain-specific homologue of PTBP1. Both genes regulate differentiation of neural precursor cells and promote the proliferation and migration of glioma cell lines.[23] In 16q23, the SNVs were located within an ncRNA gene LOC102724084.

Using gene-based tests, we identified 10 potential new genes associated with brain MRI phenotypes. Several of these genes are particularly relevant for brain-related diseases and some are also strongly expressed in the brain.

UNC5D (unc-5 netrin receptor D), related to TCBV, encodes a UNC-5 netrin receptor that plays a role in the regulation of axon guidance and it is strongly expressed in the developing sensory areas of the neocortex in mice[24] and in human brain.[25] Interestingly, a gene of the same family, UNC5C, was reported associated with AD.[26] Moreover, UNC5C genotypes have been found associated with the middle temporal volume and may alter the atrophy of AD regions such as hippocampus and precuneus.[27] The expression of C9orf84 (chromosome 9 open reading frame 84), related to TCBV, was found to be upregulated in the hippocampus of individuals with major depression.[28] UBL3 (ubiquitin like 3) belongs to the ubiquitin pathway that is implicated in the pathogenesis of neurodegenerative disorders.[29,30] UBL3 is strongly expressed in the brain.[25] SYT3 (synaptotagmin 3) is highly expressed in cortex, frontal cortex, anterior cingulate cortex, hippocampus, cerebellum, and the cerebral hemispheres. It belongs to a family of brain-specific proteins present on the membrane of synapse vesicles that play a role in the secretion of neurotransmitter.[31] LRRTM4 (leucine-rich repeat transmembrane neuronal 4) belongs to the leucine-rich repeat proteins that are important regulators of synapse development and function.[32] LRRTM4 was found to regulate excitatory synapse formation in cultured hippocampal neurons.[32] LRRTM4 is strongly expressed in the brain.[25] ALCAM (activated leukocyte cell adhesion molecule or CD166) is a ligand for CD6 and regulates leukocyte extravasation in the inflamed CNS. ALCAM is expressed on human CNS microvascular endothelium, particularly during neuroinflammatory processes, and plays a key role in the recruitment and migration of leukocytes into the brain.[33] GBA3 (glucosylceramidase β3) has been associated with Gaucher disease[34] and GBA genes have been associated with late-onset Parkinson and Lewy body dementia.[35-37] Furthermore, genetic variants near GBA3 have been found associated with white matter lesion progression.[38] SOCS7 (suppressor of cytokine signaling 7) belongs to a family of proteins that play a role in preventing inflammation in the brain. SOCS7 levels were found to be increased in AD human brains.[39] SOCS7 is noticeably expressed in murine brain and mice with disrupted SOCS7 gene showed defects in CSF homeostasis.[40]

The main results were unchanged when including individuals with dementia in the analyses. The significances of ALCAM (WMH), UBL3, and SYRT3 (TCBV) were even improved, whereas LRRTM4 (HPV) was no longer significant, suggesting that this gene may have a role early in the pathogenesis but not in patients with dementia.

The strengths of this study are the population-based sample, the use of quantitative MRI techniques, as well as sequenced data and the fact that we focused on endophenotypes that have substantial variance in our population. This study also has several limitations. First we performed our analyses in a single study. However, the access to datasets with brain MRI phenotypes and WGS data is limited and, to this date, FHS was the only TOPMed phase 1 study with sequenced data and brain MRI measures available. By fine-mapping GWAS regions, we confirmed the association of SNVs that were found associated with the imaging endophenotypes in other studies. A replication or validation in an independent cohort is necessary to confirm the new associations (variants or genes) identified in this study. We were also limited by our sample sizes to discover rare variants (2,180, 2,170, and 1,667 individuals analyzed for TCBV, HPV, and WMH, respectively). Finally, the predominantly European origin of our sample limits the generalization of these results to other ethnic groups.

This FHS WGS-wide search for brain MRI measures reveals rare variants in intriguing new loci associated with brain volumes. Replication of these results in the phase 2 TOPMed studies, whose sequencing is ongoing, is planned to confirm these findings. Further investigation of these loci, such as biological experiments, functional studies, or animal models, will have the potential to validate our findings.

Web resources

TOPMed websites

nhlbi.nih.gov/research/resources/nhlbi-precision-medicine-initiative/topmed

nhlbiwgs.org/

Methods for WGS data acquisition and QC in the FHS are described in a document that is publicly available on dbGaP

ncbi.nlm.nih.gov/projects/gap/cgi-bin/GetPdf.cgi?id=phd006969.1