Genetic variation underlying cognition and its relation with neurological outcomes and brain imaging

Cognitive performance and daily functioning

As a methodological validation, we looked whether a cross-sectional relation was present between the polygenic score (PGS) and cognition. Indeed, an increase in the PGS was significantly associated with a higher general cognitive performance (‘G-factor’) (β=0.08, p=1.2x10^-13), as well as with individual cognitive tests (Figure 2A). The PGS was also significantly associated with more years of education (β=0.29, p=4.3x10^-7). No associations with daily functioning were found. Adjusting for years of education caused an attenuation of the associations, yet only the associations for Stroop 1 became non-significant. Nearly all associations attenuated after removing variants associated with the developmental component of cognition (Figure 2A, Supplementary Figure S1). To explore the developmental component further, we created a PGS of the same 170 genetic variants using the weights for educational attainment. This PGS showed similar associations with all cognitive tests. No individual variant was significantly associated with any of the outcomes. All results for the cross-sectional analyses of cognition, daily functioning and education are shown in Supplementary Table S3.

Figure 2. Association of genetic variants for general cognitive function with (decline in) cognition and daily functioning, and educational attainment. Association between genetic variants and cognitive performance and daily functioning at one point in time, as well as years of education, adjusted for age and sex with and without adjustment for years of education (A), and change in cognitive performance and daily functioning over time (B), additionally adjusted for baseline measurement and time between baseline and follow-up measurement. Three polygenic scores are presented: a cognition polygenic score including all independent lead variants (N=170); a cognition polygenic score only including variants with a p>0.05 for the association with the developmental component of cognition, i.e. educational attainment and intracranial volume (N=36); and an educational attainment polygenic score, which contains the lead genetic variants for cognitive performance (N=170) but uses the weights for educational attainment. Larger blocks indicate higher t-values. Higher scores indicate better performance, except for the Stroop test, the Basic Activities of Daily Living and Instrumental Activities of Daily Living. Significance levels are indicated by asterisks: *p<0.05, nominally significant; **p<0.0038 (A) or p<0.0040 (B), adjusted for the number of independent traits as calculated through 10,000 permutations; ***p<2.2x10^-5 (A; 0.0038/170) or p<2.4x10^-5 (B; 0.0040/170), additionally adjusted for the number of genetic variants.

Figure 2B shows that a higher PGS was associated with less cognitive decline (β_ΔG-factor=0.03, p=1.3x10^-3), although this association became non-significant after adjusting for years of education (β_ΔG-factor=0.02, p=1.6x10^-2). A higher PGS was also associated with less decline in basic activities of daily living (BADL), although this was not significant after correcting for multiple comparisons (β=-0.02, p=4.4x10^-2). Removing genetic variants associated with educational attainment and intracranial volume did not show a substantial enrichment of the effects. In contrast, the PGS for educational attainment even showed a slightly stronger association with cognitive decline than the cognition PGS (β_ΔG-factor=0.03, p=8.3x10^-4) (Figure 2B, Supplementary Figure S2). In the single-variant analysis, no variant reached statistical significance. MMSE measurements were available in a larger sample (N=9,369) with up to six measurements and a maximum follow-up of 25.2 years. In this sample, we observed a modest but significant relation between the PGS and yearly Mini-Mental State Examination (MMSE) change using linear mixed models (β=3.5x10^-3, p=4.3x10^‑4). Supplementary Table S4 contains the complete results for the longitudinal analyses for cognition and daily functioning.

Clinical outcomes

No significant association was found between the PGS and any of the clinical outcomes (Figure 3). Out of all 170 individual lead variants, none was significantly associated with the risk of one of dementia, parkinsonism or stroke. An increased risk for dementia was found after excluding variants associated with the developmental component at a p>0.05 threshold (hazard ratio 1.06, p=0.040), although this did not survive correction for multiple testing (Supplementary Figure S3). The PGS for educational attainment was not related to any of the neurological outcomes. Full results for the analyses of clinical outcomes are presented in Supplementary Table S5.

Figure 3. Polygenic scores for general cognitive function and disease-free probability for dementia, parkinsonism and stroke. Kaplan-Meier curves presenting the association between low (i.e. below the median) and high (above the median) polygenic scores and the disease-free probability over time for dementia, parkinsonism, and stroke. Abbreviations: polygenic score (PGS).

Brain imaging markers

We found that a higher PGS was significantly related to a larger intracranial volume (β=0.05, p=7.5x10^-4), but not with the other volumetric measures or with global white matter microstructural integrity (Figure 4). At a nominal significance level, a higher PGS was associated with a higher fractional anisotropy (FA) in the medial lemniscus, and a lower mean diffusivity (MD) in the inferior-fronto-occipital fasciculus and the posterior thalamic radiation (minimal p=2.2x10^-2), but this did not survive correction for multiple testing (Figure 5). Removing genetic variants associated with the developmental component of cognition did not show a pattern of enrichment of the associations. The associations between the educational attainment PGS with the brain imaging markers were comparable to those of the cognition PGS (Figure 4, Supplementary Figure S4-S5). No individual variant reached the significance threshold for the association with any of the brain imaging markers after multiple comparisons correction. Full results of the brain imaging analyses can be found in Supplementary Table S6, Supplementary Table S7

Figure 4. Association between genetic variants for general cognitive function and global brain imaging markers. Association between genetic variants for general cognitive function and both volumetric and global white matter microstructural integrity markers. For the volumetric outcomes, analyses were adjusted for age and sex, and additionally for intracranial volume if the outcome was not intracranial volume. For the microstructural integrity outcomes, analyses were adjusted for age, sex, white matter and white matter lesion volume. Three polygenic scores are presented: a cognition polygenic score including all independent lead variants (N=170); a cognition polygenic score only including variants with a p>0.05 for the association with the developmental component of cognition, i.e. educational attainment and intracranial volume (N=36); and an educational attainment polygenic score, which contains the lead genetic variants for cognitive performance (N=170) but uses the weights for educational attainment. Also, the five top genetic variants for the association with these brain imaging markers are presented. Positive associations depicted in blue correspond to a larger volume or a better white matter microstructural integrity. Larger blocks indicate higher t-values. Significance levels are indicated by asterisks: *p<0.05, nominally significant; **p<0.0101, adjusted for the number of independent traits as calculated through 10,000 permutations. No association was significant after additional adjustments for the number of genetic variants tested (p<5.9x10^-5; 0.0101/170).

Figure 5. Association of polygenic scores for cognition and tract-specific diffusion-MRI measures. Nominally significant tracts are color-coded: dark-blue – inferior-fronto-occipital fasciculus; green – medial lemniscus; yellow – posterior thalamic radiation. Non-significant tracts are colored in light-blue.

Study population

This study was conducted within the Rotterdam Study, an ongoing population-based cohort study in the Netherlands with the aim to investigate causes and determinants of diseases in the elderly [28]. This cohort was initiated in 1990 and extended in 2000 and 2006, with a total of 14,926 participants aged 45 years and older who undergo examinations every three to four years. Assessment of dementia, parkinsonism and stroke has been performed since the start of the study. In 2002, an extensive cognitive test battery was added to the core protocol. MRI scanning was implemented in the study protocol from 2005 onwards [29]. Out of 14,926 subjects, genotyping was successfully performed in 11,496 participants. Figure 1 gives an overview of the selection of participants for the different analyses, presented in a flowchart. According to the Population Study Act Rotterdam Study, the Ministry of Health, Welfare and Sports of the Netherlands has given approval for the Rotterdam Study. All participants have given written informed consent [28].

Outcome selection

Outcomes were selected based on their link with either development or neurodegeneration. Educational attainment and intracranial volume (ICV) are established markers of cognitive and brain reserve and can therefore be used to study the developmental component of cognition [30,31]. On the other hand, dementia, parkinsonism, and stroke are clinical outcomes related to accelerated cognitive decline and neurodegeneration [5,14–16]. Furthermore, daily functioning, global brain tissue volumes, and diffusion-MRI measurements have been associated with impaired cognition in the elderly [4,16,23], and can be used as a marker for neurodegeneration.

Genotyping

The Illumina 550K, 550K duo and 610 quad arrays were used for genotyping. Samples with a call rate below 97.5% were removed, as well as gender mismatches, excess autosomal heterozygosity, duplicates or family relations, ethnic outliers, variants with call rates lower than 95.0%, failing missingness test, Hardy-Weinberg equilibrium p-value smaller than 10^-6 and allele frequencies smaller than 1%. Genotypes were imputed using MaCH/minimac software to the 1000 Genomes phase I version 3 reference panel.

Polygenic scores

We calculated a PGS using the lead genetic variants with their corresponding effect sizes for general cognitive function [3]. Genetic variants that were not available in the reference panel and variants with an r²<0.30 were excluded (N=7 and N=1, respectively). For the remaining genetic variants (N=170), the allele dosage was multiplied by the reported effect estimate (Supplementary Table S1). Subsequently, the weighted effects of all variants were added up and the resulting PGSs were standardized into Z-scores.

Since we aimed to differentiate the developmental component of general cognitive function from degenerative effects, we calculated additional PGSs where variants associated with educational attainment and intracranial volume were removed at multiple p-value thresholds. For each variant, we used the lowest p-value threshold for either educational attainment or intracranial volume. The p-values were extracted from the summary statistics of a GWAS on educational attainment performed in a discovery sample of 766,345 individuals [27], and a GWAS on intracranial volume performed in a discovery sample of 26,577 individuals [32]. The different p-value thresholds for the association with educational attainment and intracranial volume, with the corresponding number of variants that remained, as well as the explained variance of the G-factor in our dataset are shown in Supplementary Table S2. When applying the strictest p-value threshold for the exclusion of developmental variants (p>0.05), 36 genetic variants remained.

Additionally, to explore the developmental component of the lead genetic variants for general cognitive function further, we created PGSs of the same 170 variants using the weights of the educational attainment GWAS (Supplementary Table S1).

Cognitive test battery

For the cognition and daily functioning analyses, only participants who had two measurements for at least one of the tests underlying these outcomes were included. The MMSE was assessed as a measure of global cognitive function, and was collected since the initiation of the Rotterdam Study. From 2002 onwards, cognitive function was additionally assessed using multiple cognitive tests: the 15-word verbal learning test (15-WLT), the Stroop test (consisting of reading, color naming and interference tasks, error-adjusted scores), the letter-digit substitution task (LDST), the verbal fluency test (using animal categories) and the Purdue pegboard (PPB) test for the left hand, right hand and both hands [2,33–36]. A measure of general cognitive function (‘G-factor’) was obtained through principal component analysis on the delayed recall score of the 15-WLT, Stroop interference test, LDST, verbal fluency task and the PPB test, as described previously [2]. The G-factor explained 53.4% and 51.9% of the variance in cognitive test scores in our population at baseline and follow-up visit, respectively. Z-scores were calculated in order to make comparable test results.

Self-reported years of education was used as a measure of educational attainment.

Assessment of daily functioning

Two components of daily functioning were assessed: BADL and instrumental activities of daily living (IADL). The Dutch version of the Stanford Health Assessment Questionnaire was used to measure BADL [37], and IADL was measured using the Dutch version of the IADL scale [38]. To prevent selective loss of data, IADL items scored as non-applicable were imputed using the variables age, sex, BADL scores and all other available IADL items. Both BADL and IADL scores were standardized into Z-scores.

Assessment of clinical outcomes

The assessment of dementia, parkinsonism (including Parkinson’s disease) and stroke have previously been described in detail [39–41]. In summary, history of these clinical outcomes was assessed during the baseline interview. Participants were screened at baseline and subsequent center visits for dementia with the MMSE and the Geriatric Mental Schedule organic level, and for signs of parkinsonism. Participants with a positive screening were further examined and were evaluated by a panel led by an experienced neurologist who made the definitive diagnosis. After enrollment, participants were continuously monitored for dementia, parkinsonism and stroke through automated linkage of the study database with files from general practitioners. Follow-up for parkinsonism (including Parkinson’s disease) was available until the 1^st January 2015 and for dementia and stroke until the 1^st January 2016.

MRI acquisition and processing

We performed a multi-sequence brain MRI scan on a 1.5 tesla research dedicated MRI scanner (GE Signa Excite). Imaging details are provided elsewhere [29]. In short, the scan protocol included a T1-weighted image, a T2-weighted fluid-attenuated inversion recovery (FLAIR) sequence, a proton density weighted image and a spin echo echo planar diffusion weighted image for the diffusion-MRI. A multimodal algorithm was used based on T1-weighted, T2-weighted and FLAIR images to segment voxels into grey matter, white matter, white matter lesion volume, cerebrospinal fluid and background tissue using a k-nearest-neighbor-algorithm trained on six manually labelled atlases [42,43]. We estimated supratentorial intracranial volume by summing total grey and white matter volume and cerebrospinal fluid [42].

For the diffusion-MRI, three volumes were performed without diffusion weighting of which the average was used (b-value=0 s/mm², maximum b-value was 1000 s/mm²). Diffusion tensors were computed using ExploreDTI to obtain FA and MD in normal-appearing white matter voxels. We segmented fifteen white matter tracts using probabilistic tractography and atlas-based masking [44]. Tracts were grouped based on anatomic location or presumed function into brain stem tracts, projection tracts, association tracts, limbic system tracts and callosal tracts. Tract-specific FA and MD but also white matter volumes and white matter lesion volumes in specific tracts were obtained as previously described [44]. In general, a lower FA and a higher MD are indicative of lower microstructural white matter integrity.

Data analysis

Linear regression models were used to assess the associations between the PGS and cognition, daily functioning, volumetric brain outcomes and white matter microstructural integrity. For the cross-sectional analysis of cognition and daily functioning we analyzed the first measurement. We additionally assessed the association with change in MMSE using linear mixed models with an interaction between the PGS and time. We used a random intercept and slope for time and included participants with a minimum of two MMSE measurements. Cox proportional hazard models were used to study the association between the PGS and the incidence of dementia, parkinsonism, and stroke. The proportional hazards and linearity assumptions were met. All models were adjusted for age and sex. Models assessing cognition and daily functioning were performed with and without adjustment for educational attainment. Longitudinal cognition and daily functional analyses were adjusted for time between baseline and follow-up visit, and additionally for baseline measurements in the linear regression analyses. Volumetric brain outcomes were adjusted for intracranial volume when the outcome was not intracranial volume, and additionally for white matter and white matter lesion volume in the analyses for white matter microstructural integrity. The abovementioned analyses were repeated for all genetic variants separately.

Since outcomes for the different analyses may be correlated, we used permutation testing in order to assess the number of independent outcomes for each subsection. Based on this information, we defined the multiple testing p-value thresholds for the different analyses, namely p<0.0038 for the cross-sectional and p<0.0040 for the longitudinal analyses of cognitive performance and daily functioning; p<0.0101 for the volumetric and global diffusion-MRI brain measures, and p<0.0022 for the tract-specific diffusion-MRI analyses; and p<0.0129 for the clinical outcomes. For the analyses of the genetic variants separately, we additionally used the Bonferroni correction for multiple testing, using the formula k/170 with k representing the p-value threshold as obtained by permutation testing. Analyses were performed using the IBM SPSS Statistics 21 and R 3.4.0 software.