Study population
The study was based on 535 individuals from the Swedish Adoption/Twin Study of Aging (SATSA)(11), a sub-study of the population-based Swedish Twin Registry (STR)(12). Cognitive abilities covering processing speed, verbal ability, spatial ability, episodic memory, and working memory were tested during up to 10 in-person testing occasions. A measure of general cognitive ability was created based on all domains. Individuals diagnosed with dementia were censored from the time of diagnosis and onwards. The mean number of cognitive assessments was 5.4 assessments (SD=2.3, range 1-10) over an average of 12.2 years (SD=7.3, range 0-24). Blood samples were collected from the third in-person testing occasion and onwards, and DNA methylation measured from the first available blood sample was used in this study. The sample consisted of 313 (58.5%) women and 222 (41.5%) men, with a mean age of 64.8 (SD=8.3, range 48-94) years at first participation and 68.2 (SD=9.5, range 48-94) years at first blood sample. At the time of blood sampling, 95 individuals were current smokers (17.8%). The sample included 238 (82 monozygotic, 156 dizygotic) complete twin pairs.
EWAS of empirical Bayes estimates for level and change in cognitive abilities
In the first step of analyses, we performed an EWAS to identify epigenome-wide significant (threshold pre-defined at p<2.4x10-7(13)) and suggestive (threshold pre-defined at p<10-5(9)) signals.
As longitudinal models are computationally intense and hence not ideal for the EWAS setting, we first obtained EB estimates by applying linear and quadratic latent growth curve models(14) to each cognitive domain. Thus, individual measures of cognitive level at the intercept age and of the linear and quadratic change across time were obtained and used as separate outcomes in epigenome-wide analyses. To obtain more precise EB estimates, cognitive information across all in-person testing occasions was used, regardless of when methylation was measured. A quadratic model best fit the data for all domains except working memory, where the linear model showed the best fit. Intercept age was set at 65 for all domains except verbal ability where the intercept age 70 best fit the data (based on previous work(15)).
The epigenome-wide analyses were then modelled in linear regressions, with DNA methylation at each CpG site as the exposure and the EB estimates as separate outcomes. The models were adjusted for sex, age and smoking at time of blood sample, methylation array, and relatedness among the twins. Estimates for linear and quadratic slopes were scaled to represent 10-year change.
Significant findings from the EWAS of DNA methylation and EB estimates of level and change in cognitive abilities are presented in Table 1, and suggestive findings in Additional file 1. In total, 5 CpG sites reached epigenome-wide significance, all with level of cognitive ability at the intercept age65: cg18064256 (PPP1R13L) with lower level of processing speed and spatial ability; cg04549090 (NRXN3) with higher level of spatial ability; cg08011941 (ENTPD8) and cg25651129 (-) with higher level and cg09988380 (POGZ) with lower level of working memory.
Another 131 suggestive associations were identified (Additional file 1). Of note is that cg18064256 also showed a suggestive association with general cognitive ability at age 65 and 10-year linear change in processing speed, and cg04549090 with level of general cognitive ability at age 65. Another 11 CpG sites showed suggestive associations with more than one cognitive domain or growth feature (Additional file 1). All sites with a significant or suggestive p-value were carried forward to follow-up analyses of the respective cognitive domain.
Between-within models of DNA methylation and empirical Bayes estimates for level and change in cognitive abilities
We applied between-within models(16), where the between-pair estimate represents the average effect in the population while the within pair estimate represents the effect after adjusting for genetic and other familial factors shared within the twin pair. As in the epigenome-wide analyses, DNA methylation at each CpG site was modelled as the exposure and EB estimates for cognitive level and change as the outcome, and the models adjusted for sex, age and smoking at time of blood sample, and methylation array.
Table 1: Significant epigenome-wide associations of DNA methylation and level and change in cognitive abilities
Cognitive domain
and CpG site
|
|
|
Total sample
n=535 individuals
|
Between-pair effect n=297 pairs
|
Within-pair effect
n=238 complete pairs
|
Gene
|
Position1
|
Beta
|
SE
|
P-value
|
Beta
|
SE
|
P-value
|
Beta
|
SE
|
P-value
|
Processing speed (intercept)
|
|
|
|
|
|
|
|
|
|
|
|
cg18064256
|
PPP1R13L
|
19:45905621
|
-1.77
|
0.31
|
1.55e-08
|
-2.22
|
0.51
|
2.20e-05
|
-1.14
|
0.34
|
9.42e-04
|
Spatial ability (intercept)
|
|
|
|
|
|
|
|
|
|
|
|
cg04549090
|
NRXN3
|
14:79033036
|
1.95
|
0.36
|
1.23e-07
|
2.34
|
0.55
|
2.95e-05
|
1.02
|
0.37
|
6.55e-03
|
cg18064256
|
PPP1R13L
|
19:45905621
|
-2.01
|
0.38
|
1.67e-07
|
-2.15
|
0.56
|
1.45e-04
|
-1.74
|
0.38
|
5.99e-06
|
Working memory (intercept)
|
|
|
|
|
|
|
|
|
|
|
|
cg09988380
|
POGZ
|
1:151431765
|
-2.11
|
0.40
|
1.88e-07
|
-2.64
|
0.61
|
2.57e-05
|
-1.04
|
0.48
|
0.03
|
cg25651129
|
-
|
8:11474056
|
1.89
|
0.35
|
1.34e-07
|
2.41
|
0.64
|
2.05e-04
|
1.21
|
0.44
|
6.28e-03
|
cg08011941
|
ENTPD8
|
9:140333139
|
2.00
|
0.35
|
2.67e-08
|
2.49
|
0.62
|
7.93e-05
|
1.01
|
0.47
|
0.03
|
Note. Significant (p<2.4x10-7) associations from epigenome-wide analyses of DNA methylation and level and change in processing speed, verbal and spatial ability, episodic and working memory, and general cognitive ability in the total sample, followed by results between- and within twin pairs. Empirical Bayes estimates for level of cognitive ability at the intercept age (age 70 for verbal ability, 65 for all other domains) as well as 10-year linear and quadratic change were modelled as separate outcomes. Linear regression was applied to the total sample, and between-within models to compare estimates between and within twin pairs. All models were adjusted for age, sex, smoking, methylation array, and number of testing waves with cognitive measures. 1 Genome Reference Consortium Human Build 37 (GRCh37)
Results from between-within models are presented in Table 1 for significant associations from the epigenome-wide analyses, and in Additional file 1 for suggestive associations. All the significant associations presented were substantially reduced with, on average, halved beta values for the association between methylation on cognitive abilities within twin pairs compared to between pairs.
Latent growth curve models of DNA methylation and level and change in cognitive abilities
Latent growth curve models with age in decades as the time scale were fitted simultaneously with identified methylation sites to evaluate the trajectory features of cognitive abilities during late-life, using cognitive data from the time of methylation measurement and onwards. The intercept term here represents the level of cognitive ability at the intercept age (70 years for verbal ability, 65 years for all other domains), while the linear term represents the instantaneous linear rate of change at the intercept age, and the quadratic term the acceleration of change across age. As in the epigenome-wide analyses, the models were adjusted for sex, age and smoking at time of blood sample, methylation array, and relatedness among the twins. To evaluate the significance of the effect of methylation on level and change taken together, a likelihood ratio test was performed, comparing the model fit of the full model to that of a null-model with only covariates and no methylation included. Standardized mean differences (Cohen’s d equivalents) in cognitive abilities by 1 SD higher DNA methylation were calculated for the intercept level and for change over 10 years from the intercept age (see methods section)(17).
Growth features for each cognitive domain from a null model (without DNA methylation predictors) are presented in Additional file 2. The intercept level ranged from 51.0 to 54.7, the linear slope from -0.5 to -3.4, and the quadratic slope from -0.56 to -1.4. The effects of DNA methylation on the intercept level, 10-year linear change, and 10-year quadratic change in cognitive abilities are presented in Table 2 (significant associations in the epigenome-wide analyses) and Additional file 3 (suggestive associations in the epigenome-wide analyses). Figure 1 visualizes the estimated growth trajectories with one SD higher methylation for the significant sites, alongside the estimated trajectories from the corresponding null model.
One SD higher methylation level in cg18064256 (PPP1R13L) was associated with lower levels of processing speed and spatial ability at age 65, with a steeper linear decrease at age 65, but slightly less accelerating decrease. One SD higher methylation in cg04549090 (NRXN3) was associated with higher levels of spatial ability at age 65, a less steep rate of linear change at the same age, followed by a more accelerating decline. cg09988380 was associated with lower level of working memory at age 65 but a less steep linear rate of change. cg25651129 and cg08011941 were associated with higher level of working memory at age 65 but with a steeper linear decline.
The standardized effect sizes for the associations between DNA methylation and intercept cognitive level ranged between 0.14-0.18 (Table 2), thus considered of small magnitude(18). The associations between DNA methylation and 10-year change from age 65 to 75 were of very small magnitude (0.01-0.09), but would reach larger magnitudes when cumulating over decades.
Characterization of the CpG sites
To characterize longitudinal change in methylation at the five significant CpG sites and to identify meQTLs, we extracted results from a study by Wang and colleagues (19), which studied longitudinal change in methylation levels during aging and cis-meQTLs (within 1 million base pairs) in the SATSA sample. None of the five sites were significantly associated with age in the study by Wang et al., nor was any evidence of cis-meQTLs driving methylation identified.
Table 2: The association between DNA methylation and longitudinal trajectories of cognitive abilities
Cognitive domain and CpG site
|
CpG on intercept
|
CpG on linear change
|
CpG on quadratic change
|
LRT CpG
|
Cohen’s d equivalent
|
Beta
|
SE
|
P-value
|
Beta
|
SE
|
P-value
|
Beta
|
SE
|
P-value
|
P-value
|
Intercept
|
Change 65-75
|
Processing speed
|
|
|
|
|
|
|
|
|
|
|
|
|
cg18064256
|
-1.35
|
0.32
|
2.90e-05
|
-0.38
|
0.26
|
0.15
|
0.03
|
0.15
|
0.85
|
1.01e-05
|
0.14
|
0.04
|
Spatial ability
|
|
|
|
|
|
|
|
|
|
|
|
|
cg04549090
|
1.37
|
0.35
|
1.11e-04
|
0.24
|
0.22
|
0.27
|
-0.15
|
0.16
|
0.34
|
5.25e-04
|
0.14
|
0.01
|
cg18064256
|
-1.71
|
0.35
|
1.41e-06
|
-0.35
|
0.23
|
0.13
|
0.10
|
0.17
|
0.54
|
8.53e-07
|
0.17
|
0.03
|
Working memory
|
|
|
|
|
|
|
|
|
|
|
|
|
cg09988380
|
-1.61
|
0.37
|
1.48e-05
|
0.30
|
0.21
|
0.16
|
--
|
--
|
--
|
9.93e-05
|
0.16
|
0.03
|
cg25651129
|
1.83
|
0.35
|
2.80e-07
|
-0.85
|
0.21
|
4.03e-05
|
--
|
--
|
--
|
2.63e-07
|
0.18
|
0.09
|
cg08011941
|
1.41
|
0.37
|
1.38e-04
|
-0.11
|
0.22
|
0.61
|
--
|
--
|
--
|
4.22e-04
|
0.14
|
0.01
|
Note. Mean cognitive level, 10-year linear change, and 10-year quadratic change in cognitive abilities in relation to DNA methylation at sites significant in EWAS analyses. Beta values, standard errors, and p-values were obtained from latent growth-curve models, with age (in decades) as the underlying time scale. Age was centered at 65 for all domains. The models were further adjusted for sex, smoking, and methylation array. The model fit was compared to a null model not including DNA methylation to assess the significance of the effect of DNA methylation on cognitive level and change. Standardized mean differences (Cohen’s d equivalents) by 1 standard deviation higher DNA methylation at respective site was calculated for the intercept level and for 10-year change in cognitive abilities.
SE: standard error; LRT: likelihood ratio test
We also performed lookup in the online mQTL database(6) (filtering on middle-age individuals) to identify cis- and trans-meQTLs driving methylation at the significant CpG sites. We here identified two potential (not meeting a strict p<10-14 significance level(6)) trans-meQTLs: chrX:118976619:I is associated with methylation levels at cg04549090 (p=4.77x10-08) and rs144382559 on chromosome 10 with cg08011941 (p=9.53x10-09). To study whether these two meQTLs were associated with methylation level in the SATSA sample, the two SNPs were extracted from genotype data and modelled in linear regression models as predictors of DNA methylation at the relevant site, and of the EB estimates for the relevant cognitive domain. The SNP on the X-chromosome was modelled separately for men and women. Neither of the SNPs were associated with methylation levels in this sample (chrX:118976619:I with cg04549090, β=-0.02, p=0.94 in women, β=0.11, p=0.57 in men; rs144382559 with cg08011941, β=-0.08, p=0.76), nor with cognitive level at age 65 (chrX:118976619:I with spatial ability, β=26.98, p=0.60 in women, β=-16.31, p=0.52 in men; rs144382559 with working memory, β=-1.87, p=0.33).
To investigate whether DNA methylation in blood leukocytes is correlated with that in brain cells, we performed lookup in IMAGE-CpG(20), an online tool to compare methylation levels in blood and brain from live human tissues, and the Blood Brain DNA Methylation Comparison Tool(21), where methylation levels can be compared in blood and four different brain regions (prefrontal cortex, entorhinal cortex, superior temporal gyrus, and cerebellum) from post-mortem samples. According to the Blood Brain DNA Methylation Comparison Tool(21), blood methylation levels of cg18064256 showed a moderate correlation with levels in the entorhinal cortex (r=0.40, p=5.4x10-4) and the superior temporal gyrus (r=0.32, p=5.7x10-3). None of the other CpG sites showed significant correlations between blood and brain methylation levels in either online tool.
To investigate expression of the genes across tissues, we performed additional lookup in the Human Protein Atlas(22) (available from http://www.proteinatlas.org). PPP1R13L (cg18064256) and POGZ (cg09988380) are both expressed in several tissues, including brain and blood where both show low brain region and blood cell type specificity. NRXN3 (cg04549090) is primarily expressed in the brain, with low region specificity, and blood where it is primarily expressed in basophils. ENTPD8 (cg08011941) is mainly expressed in the intestines and is generally not expressed in brain or blood cells.