Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Effect of apolipoprotein E polymorphism on cognition and brain in the Cambridge Centre for Ageing and Neuroscience cohort:

Effect of apolipoprotein E polymorphism on cognition and brain in the Cambridge Centre for Ageing... Polymorphisms in the apolipoprotein E (APOE) gene have been associated with individual differences in cognition, brain structure and brain function. For example, the ε4 allele has been associated with cognitive and brain impairment in old age and increased risk of dementia, while the ε2 allele has been claimed to be neuroprotective. According to the ‘antagonistic pleiotropy’ hypothesis, these polymorphisms have different effects across the lifespan, with ε4, for example, postulated to confer benefits on cognitive and brain functions earlier in life. In this stage 2 of the Registered Report – https:// osf.io/bufc4, we report the results from the cognitive and brain measures in the Cambridge Centre for Ageing and Neuroscience cohort (www.cam-can. org). We investigated the antagonistic pleiotropy hypothesis by testing for allele-by-age interactions in approximately 600 people across the adult lifespan (18–88 years), on six outcome variables related to cognition, brain structure and brain function (namely, fluid intelligence, verbal memory, hippocampal grey-matter volume, mean diffusion within white matter and resting-state connectivity measured by both functional magnetic resonance imaging and magnetoencephalography). We found no evidence to support the antagonistic pleiotropy hypothesis. Indeed, Bayes factors supported the null hypothesis in all cases, except for the (linear) interaction between age and possession of the ε4 allele on fluid intelligence, for which the evidence for faster decline in older ages was ambiguous. Overall, these pre-registered analyses question the antagonistic pleiotropy of APOE polymorphisms, at least in healthy adults. Keywords Cognition, apolipoprotein E, lifespan, brain, ageing Received: 13 December 2019; accepted: 27 August 2020 Introduction MRC Cognition & Brain Sciences Unit, University of Cambridge, Cambridge, UK Apolipoprotein E (APOE) is a protein that plays an important Department of Psychiatry, University of Cambridge, Cambridge, UK role in lipid metabolism (including cholesterols) and has been Department of Psychiatry, Warneford Hospital, University of Oxford, implicated in synaptogenesis, repair of injured nerve tissue and Oxford, UK the modulation of beta-amyloid plaques and neurofibrillary tan- Wellcome Centre for Integrative Neuroimaging, University of Oxford, gles that characterise Alzheimer’s disease (AD) (for review, see Oxford, UK Belloy et al., 2019; Rocchi et al., 2003). The gene coding for Department of Clinical Neurosciences, University of Cambridge, APOE is located on chromosome 19 and is polymorphic in the Cambridge, UK general population. The three most common alleles are ε2, ε3 and Institute of Health and Wellbeing, University of Glasgow, Glasgow, UK ε4, with approximate allele frequencies of 6%, 78% and 15% in Institute of Cognitive Neuroscience, University College London, London, UK healthy Caucasian Europeans (Eisenberg et al., 2010). Possession Language and Genetics Department, Max Planck Institute for of the ε4 allele has been associated with poorer cognitive abilities Psycholinguistics, Nijmegen, The Netherlands and more rapid longitudinal decline in healthy older people, par- Donders Institute for Brain, Cognition and Behaviour, Radboud ticularly in episodic memory (e.g. Jack et al., 2015; Jochemsen University, Nijmegen, The Netherlands et al., 2012; Jorm et al., 2007; Lyall et al., 2016; Marioni et al., 2016; Mondadori et al., 2007; Schiepers et al., 2012; Schultz Corresponding author: et al., 2008; Shin et al., 2014; see Wisdom et al., 2011 for a meta- Richard N. Henson, MRC Cognition & Brain Sciences Unit, University of analysis). It has also been associated with a three-to-fourfold Cambridge, 15 Chaucer Road, Cambridge CB2 7EF, UK. increase in the risk of late onset AD in a gene dose-dependent Email: rik.henson@mrc-cbu.cam.ac.uk Creative Commons CC BY: This article is distributed under the terms of the Creative Commons Attribution 4.0 License (https://creativecommons.org/licenses/by/4.0/) which permits any use, reproduction and distribution of the work without further permission provided the original work is attributed as specified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/open-access-at-sage). 2 Brain and Neuroscience Advances Figure 1. Illustration of potential interactions between APOE status Figure 2. Classical power analyses for (any) polynomial term of GLM and age on cognitive and/or brain health. for (1) separate, categorical analysis of ε3ε3 versus ε2+ε4− (green According to the ‘antagonistic pleiotropy’ hypothesis, ε4-carriers (red line) line), (2) separate, categorical analysis of ε4+ε2− versus ε3ε3 (red experience a benefit in early life but a detriment in later life, relative to non- line) and (3) linear, parametric ‘dose’ effect of ε2−/ε4+ load (blue carriers (ε3ε3 group, blue line). The characteristics of ε2-carriers (green line) are less clear, but generally thought to benefit later in life. Note that every person dotted line). The vertical lines and numbers refer to prior effect sizes has two alleles, so ‘ε4-carrier’ (indicated by ‘ε4*’ in legend) means someone with in literature, as described in text. ε4ε4 or ε3ε4 (but excluding ε4ε2 – see text for more details). manner (Farrer et al., 1997) and with an earlier age at onset by lifespan cohort of healthy adults uniformly distributed from 18 to nearly 6 years on average for ε4 carriers (Blacker et al., 1997). 88 years, which allows us to overcome many of the limitations of Indeed, some have argued that APOE has no influence on cogni- former studies of the effects of ε4 at different ages. tion in mid- or late-life beyond increasing risk for AD, such that Another strand of the literature has examined the possible effects found on cognition reflect the decades of a pre-sympto- protective effects of the ε2 allele, which has been reported to matic period of AD pathology (e.g. Vemuri et al., 2010; see also reduce risk of AD (Farrer et al., 1997). The relationship of these O’Donoghue et al., 2018). putative protective effects with age is less well studied (partly There have also been claims that the ε4 allele confers cogni- due to its rarer frequency), but one suggestion is that ε2 is associ- tive benefits earlier in life, that is, individuals aged 5–35 years ated with more robust neurodevelopment in early life (relative to (e.g. Puttonen et al., 2003; Rusted et al., 2013; Yu et al., 2000, ε3), while another is that it is associated with lower levels of though, see Ihle et al., 2012 for a meta-analysis of studies that neurodegeneration (relative to ε4) in later life (Suri et al., 2013). found no consistent effect of ε4 status on cognitive tests in this CamCAN’s lifespan cohort allows novel tests for these interac- age-group). These claims led to the ‘antagonistic pleiotropy’ tive or additive effects of the ε2 allele with age (green line in hypothesis of Han and Bondi (2008), whereby the APOE ε4 Figure 1). allele is hypothesised to be advantageous in early life but disad- The mean effect size of ε4-carrier versus non-carrier status on vantageous in later life (and potentially neutral in middle-age, cognition is typically small when averaging across ages. For consistent with the lack of association with cognition in a meta- example, the meta-analysis of Wisdom et al. (2011) showed a analysis of middle-aged people by Lancaster et al., 2017). The Cohen effect size (d) across studies of d = −0.14 on episodic antagonistic pleiotropy hypothesis is illustrated in Figure 1 (cf. memory and d = −0.05 on global cognition, averaged across ages red versus blue lines). from 20 to 90 years (where d = 0.20 is considered ‘small’). This Cognitive benefits in early life (while still fertile), or impair- effect size would be difficult to detect with the approximately ments that only arise later in life, might explain in evolutionary 600 participants in the CamCAN imaging sample who have valid terms why the ε4 allele persists in the population at a relatively cognitive, neuroimaging and genetic data, including APOE status high frequency (Fullerton et al., 2000). However, most studies of (see Figure 2); a sample that is considerably smaller than several ε4 have examined young, middle-aged or older samples sepa- recent large studies (e.g. Lyall et al., 2019a; Marioni et al., 2016; rately. Moreover, studies of older participants are often biased by Shin et al., 2014). However, if the antagonistic pleiotropy hypoth- the over representation of super-healthy individuals who are esis is correct, then the effect size depends on age, such that the motivated to respond to adverts or join volunteer panels, as effect size for the interaction between APOE and age could be opposed to population-based recruitments. bigger than the effect size averaging over age (see Figure 1 and To examine the effect of APOE across the adult lifespan, we section ‘Methods’). Consistent with this view, the meta-analysis used the ‘imaging’ sample of Cambridge Centre for Ageing and of Wisdom et al. (2011) also found a significant linear effect of Neuroscience (CamCAN; www.cam-can.org) to test for a con- age, with the detrimental effect of ε4 increasing with age. tinuous interaction between age and APOE status on a variety of Similarly, more recent studies (e.g. Lyall et al., 2019b; Marioni cognitive and brain measures. CamCAN is a population-derived, et al., 2016; Shin et al., 2014) have generally reported stronger Henson et al. 3 detrimental effects of ε4 in older than younger groups. Moreover, samples of 73 and 203 adults, respectively, found similar effects in a large (genome-wide) study of multiple cohorts, Davies et al. of ε4- and ε2-carriers (relative to ε3ε3) on various DTI metrics, (2015) found that rs10119 – a single-nucleotide polymorphism including increases in mean diffusivity (MD) and decreases in (SNP) neighbouring the APOE locus – showed a negative corre- fractional anisotropy (FA), where both of these changes are reflec- lation of R = −0.424 between the average cohort age (from 55 to tive of poorer WM health. Similar MD increases and FA decreases 80 years) and the effect size for this SNP on general cognitive have been reported in AD (Stebbins and Murphy, 2009). However, ability. This corresponds to an interaction effect size of d = −0.936 the parallel rather than opposite effect of ε4 and ε2, coupled with (Figure 2). Furthermore, the power to detect such an interaction the fact that neither study found any interactions with age, led between age and APOE status is likely to benefit from the wider both groups to suggest that the APOE variants reflected neurode- age range in CamCAN than is typically available in the cohorts velopmental differences, rather than any changes related to AD. A that have been tested so far (such as the UK Biobank), which study of N ~ 650 older individuals (aged ~73 years) found reduced have tended to focus on middle and older ages. FA in ε4-carriers (but not ε2-carriers) in two of 15 major white- In addition to cognition, the APOE variants have been associ- matter tracts, but not in the first principal component across tracts ated with differences in brain structure and function, as measured (Lyall et al., 2014). A more recent and much larger study from the with magnetic resonance imaging (MRI) or magnetoencephalog- UK Biobank (Lyall et al., 2019b) found neither effect of ε4 nor raphy (MEG). Indeed, it is possible that such allele-specific interaction with age, on the first principal component of MD or FA effects might be seen earlier in the lifespan than they can be seen across 22 WM tracts. However, as the authors noted, the restricted in cognition (Siebner et al., 2009). This could be because neuro- age range of mostly 50–70 years may have limited their ability to imaging measures are more sensitive than behavioural measures detect age effects. CamCAN’s larger range of ages make it better at detecting subclinical abnormalities. Another possibility for the suited to investigate the effects of ε4 and ε2 and their interactions lower sensitivity of cognitive measures (versus brain measures) is with age. that young ε4-carriers show compensatory changes (e.g. increased Several studies have assessed functional connectivity between functional activity or connectivity) that maintain cognitive perfor- brain regions, acquiring, for example, blood oxygenation level mance comparable to non-carriers; however, at some point in dependent (BOLD)-weighted MRI while people rest. There has older age, these compensatory mechanisms fail and a deleterious been a focus on the default mode network (DMN), which includes effect on cognition is unmasked. A related idea is that the increased hippocampus as well as medial parietal and medial frontal regions. activity in younger ε4-carriers ‘may cause neurophysiological Using independent component analysis (ICA), for example, changes that lead to earlier age-related decline in brain function’ Westlye et al. (2011) reported increased expression of DMN con- (Buckner et al., 2008; Filippini et al., 2009a). Both of these nectivity for N = 33 ε4-carriers versus N = 62 non-carriers (aged hypotheses predict an interaction between APOE variant and age. 50–80 years). Based on the similarity of the DMN and the regions Indeed, if one were to find a greater interaction between APOE disrupted in AD, it has been suggested that enhanced metabolism and age on brain measures than on cognitive measures, this would in the DMN may provide regional conditions that are conducive support the idea that brain function compensates for the detrimen- to amyloid deposition (Buckner et al., 2008). However, tal effects of APOE on brain structure, compensating in youth but Trachtenberg et al. (2012) found similar effects of ε4 and ε2 on no longer being able to compensate in old age. functional connectivity in a young group and argued that the Many neuroimaging studies of APOE variants have focused effects of APOE do not relate only to AD risk, but rather to the on volumetric differences in structural (e.g. T1-weighted) MRI putative role of APOE in neurodevelopment (see also Deary et al., images, particularly atrophy of grey matter in regions of the 2002; O’Donoghue et al., 2018). Moreover, other results are medial temporal lobe, such as the hippocampus, which occurs divergent, with both increased and decreased functional connec- early in AD and is associated with memory problems. tivity reported for APOE variants (e.g. Damoiseaux et al., 2012; Unfortunately, the results have been mixed, with most studies Filippini et al., 2009a; Fleisher et al., 2009; Machulda et al., 2011; not finding a mean effect of APOE, ε4 or ε2, particularly in hip- Pietzuch et al., 2019; Sheline et al., 2010 for review). Indeed, pocampus (e.g. Bunce et al., 2012; Filippini et al., 2009b; Habes Pietzuch et al. (2019) suggested that this divergence ‘may relate to et al., 2016; Jack et al., 2015; Lyall et al., 2019b; Machulda the age of the sample groups’ (see also Damoiseaux et al., 2012 et al., 2011; Westlye et al., 2011; Taylor et al., 2015; see Table 1 for potential moderating effects of age and sex). Indeed, in a in Lyall et al., 2013 and Fouquet et al., 2014 for review). cross-sectional study, Shu et al. (2016) reported that, while both Nonetheless, despite the converging evidence against any effect ε2- and ε4-carriers had decreased DMN connectivity compared to of APOE, we tested main effects on hippocampal volume and ε3 homozygotes, they showed opposite effects of age. the interaction with age because of their historical and theoreti- Interpretation of changes in functional connectivity measured cal relevance. by functional MRI (fMRI) is non-trivial, since the measures can Other MRI studies have examined effects of APOE on white- be affected by vascular as well as neural factors (Geerligs et al., matter integrity, for example, the number of white-matter hyper- 2017; Tsvetanov et al., 2016). MEG provides a more direct meas- intensities (WMHI) in MRI images (e.g. Lyall et al., 2019b). ure of neural activity, albeit with worse spatial resolution than While CamCAN does not have MRI data suitable for estimating fMRI. To our knowledge, Koelewijn et al. (2019) is the only WMHI in the full sample, it does possess diffusion-weighted MEG study to examine resting-state functional connectivity in images (DWIs) that can be used to estimate diffusion tensor imag- APOE variants with an appreciable number of N = 159 healthy ing (DTI) metrics. Given APOE’s established role in cholesterol adults (N = 159; 51 ε4-carriers versus 108 non-carriers aged 18– transport, it is conceivable that different alleles may modulate 65 years, though mostly young; see Cuesta et al., 2015, for a age-related myelin degradation, which would affect such DTI smaller study). These authors reported that they could classify ε4 metrics. Two studies (Heise et al., 2011; Westlye et al., 2012) with status with an accuracy of 63.5%, based on the strongest 4 Brain and Neuroscience Advances connections between brain regions. Though the CamCAN data CBU in Cambridge and Max Planck Institute for Psycholinguistics were recorded with eyes closed rather than open and on a differ- in Nijmegen, respectively), and the link between them only de- ent MEG system, we replicated their analysis pipelines, as closely blinded when the Stage 1 version of this study was accepted and as possible, for the CamCAN data. restricted to APOE allelic status. In summary, we tested the antagonistic pleiotropy hypothesis, with regard to both ε4- and ε2-carriers, on a number of cognitive and brain measures that are available in the CamCAN cohort; Choice of dependent variables and models specifically, those measures that have been linked to APOE status In order to minimise the number of statistical tests, we restricted in previous studies, empirically and/or theoretically. We used a ourselves to six dependent variables – two cognitive and four quadratic expansion of age (see section ‘Methods’) to test for brain measures – which we deemed to have the strongest prior interactions between APOE status and age. In the case of signifi- evidence for associations with APOE allelic status (e.g. episodic cant non-linear effects (e.g. inverted U-shaped fits as in Figure memory) or are most theoretically relevant (e.g. hippocampal 1), we planned to estimate the age of the vertex (peak scores) for volume). These measures also have prior effect sizes published in each allele group. Since it was possible that we would identify no the literature with which to estimate power (see section ‘Power significant quadratic component (if, for example, the beneficial calculation’ below). effects of ε4 or ε2 are only seen below the age of 18 years, that is, For the cognitive data, we focused on two abilities that have during early development), such that any interaction between age shown the most consistent associations with APOE status in the and APOE status could reflect just an accelerated decline in old literature: (1) fluid intelligence and (2) episodic memory (Davies age, we also compared linear slopes. et al., 2015; O’Donoghue et al., 2018; Wisdom et al., 2011). For the former, we used the Cattell test (first principal component across four sub-tests) and for the latter we used the WAIS logical Methods memory test of verbal memory (first principal component across Data selection immediate recall, delayed recall and delayed recognition) – see Shafto et al. (2014) for more details. None of the CamCAN participants had a diagnosis of dementia For grey matter, we focused on (3) hippocampal volume, as or mild-cognitive impairment at recruitment; all reported them- extracted following application of FreeSurfer 6.0 (https://surfer. selves to be in good cognitive health, and all scored above con- nmr.mgh.harvard.edu/fswiki/DownloadAndInstall) to 1 mm-iso- ventional cut-offs for dementia on the mini-mental state tropic T1-weight magnetization prepared rapid gradient echo examination (MMSE) and Addenbrooke’s cognitive examina- (MPRAGE) scans (see https://camcan-archive.mrc-cbu.cam. tion – revised (ACE-R) screening tests (see Shafto et al., 2014). ac.uk/dataaccess/pdfs/CAMCAN700_MR_params.pdf for fur- This study was conducted in compliance with the Helsinki ther details of MRI scans). To correct for inter-individual differ- Declaration and was approved by the local ethics committee, ences in head size, hippocampal volumes were adjusted for total Cambridgeshire 2 Research Ethics Committee (reference: 10/ intracranial volume (TIV). H0308/50). The cognitive, MRI and MEG data from CamCAN For white matter, we used (4) the participant loadings of the are available on request from https://camcan-archive.mrc-cbu. first principal component of MD values across major white-matter cam.ac.uk/dataaccess/. The cognitive data were already scored tracts defined by the John Hopkins Atlas (averaged across hemi- (available on above website) and the MRI data had already been sphere, as in de Mooij et al., 2018), derived from 2-mm-isotropic preprocessed (see Taylor et al., 2017). DWI data preprocessed according to Taylor et al. (2017). DNA was prepared from saliva samples, which underwent For resting-state fMRI (rsfMRI) connectivity, we focused on genome-wide genotyping using the Illumina Infinium (5) mean connectivity within the DMN, following the optimised ‘OmniExpressExome’ SNP-chip. This chip covers >960,000 pre-processing pipeline described in Geerligs et al. (2017). SNP markers spread through the genome, capturing a large pro- Finally, for the resting-state MEG (rsMEG) connectivity, we portion of common variation. The common SNPs rs7412 and followed the procedure of Koelewijn et al. (2019) and examined rs429358 were used to determine APOE ε2, ε3 and ε4 allelic sta- (6) the participant loadings of the first principal component of tus of the participants. Full raw genotype data were first filtered connection strengths in source space. For further details of the in GenomeStudio according to standard procedures (Guo et al., preprocessing of the imaging measures and minor deviations 2014). Additional quality control checks were performed in from the Stage 1 report, see Supplemental Table 1. −6 PLINK (removal of SNPs for which Hardy Weinberg p < 1 × 10 , missingness > 0.05, minor allele frequency < 0.05; removal of individuals with total SNP missingness > 0.05 or where multidi- Statistical models mensional scaling indicated non-European origin). After quality For all six dependent variables, we modelled the effects of age by control, the data set included 675,373 directly genotyped SNPs. a second-order polynomial expansion, implemented in a general Genotype data were imputed using the Haplotype Reference linear model (GLM). A standard quadratic expansion of age is Consortium version 1.1 panel in the Michigan Imputation Server given by (https://imputationserver.sph.umich.edu). Genotypes of rs7412 were extracted from the raw genotype data, while genotypes for 2 yx =+ ββ + β x 01 2 rs429358 were derived by imputation. Prior studies have shown that this imputation procedure has high accuracy for determining where y is the dependent variable (e.g. cognitive score), x is APOE allelic status (Radmanesh et al., 2014). The phenotypic (mean-corrected) age and β are the polynomial parameters to 02 − and genotypic data were held in separate laboratories (MRC be estimated. Note that there is an equivalent parabolic form Henson et al. 5 The ‘apoe_power_simulations.m’ MATLAB script for these   yx =+ ββ () − β 01 2 analyses is provided in the website: https://osf.io/ehs9n/. Using in which 2 is the vertex (age of maximal performance as shown a Bonferroni-corrected alpha value of 0.05/6 (given the six out- in Figure 1), which can be estimated as come variables above – no correction was made for the three contrasts of APOE groups, because these contrasts share sub- ββ =− / 2β 21 2 sets of the same data, nor for the number of polynomial effects, since the main interest was in non-zero terms that were relevant All regressors of interest in the GLM were Z-scored. The GLMs to the antagonistic pleiotropy hypothesis), the cohort provides were fit in the MATLAB function ‘apoe_getdata.m’ and equiva- >80% power for effect sizes between 0.20 and 0.44 (i.e. small lently in the R script ‘apoe_lm_brms.R’, available on https://osf. to medium effects), depending on the comparison and model io/ehs9n/. Outliers on any of the six phenotypic variables were (Figure 2; note that although we use ‘power’ to refer to a single defined by residuals that were 1.5 times the interquartile range, study, strictly it refers to the expected outcome over a series of after adjusting for polynomial effects of age. studies). The following effect sizes from the literature are In the ‘categorical’ GLMs, the effects of age were modelled shown in Figure 2: separately for three sub-groups (Table 1): ε2-carriers without ε4, or henceforth the ‘ε2+ε4−’ group (i.e. ε2ε2, ε2ε3); ε4-carriers 1. The effect size of d = −0.14 of ε4 on episodic memory without ε2, or henceforth the ‘ε4+ε2−’ group (i.e. ε3ε4, ε4ε4) when averaging across all adult ages (Wisdom et al., and the ‘ε3ε3’ reference group who do not carry either ε2 or ε4 2011), for which our power is only 17% for ε4-carriers (we ignored any ε2ε4 cases, which were less than 3% of our sam- versus non-carriers or 41% if there is a parametric effect ple, as is common in the field). of ε2−/ε4+ load. Two comparisons of pairs of groups were planned: (1) ε3ε3 2. More important for the present hypothesis about age- versus ε2+ε4− and (2) ε4+ε2− versus ε3ε3. The mean and linear dependence of ε4 effects, the recent meta-analysis of age terms were tested separately as one-tailed t-contrasts, where Davies et al., 2015 found an effect size of d = −0.936 for the tail of the test depended on the direction of the expected effect the linear effect of age for the rs10119 region associated on the specific phenotypic variable (e.g. a detrimental effect of ε4 with APOE on general cognitive ability, for which the would produce lower Cattell scores of fluid intelligence, but present power is close to 100% in all cases. higher values of MD scores of white-matter integrity). For scores 3. For hippocampal volume, Taylor et al. (2015) reported like Cattell, where larger values mean better performance, one an effect size for ε4 of d = −0.56, for which this study has would expect a negative difference for both planned comparisons, a power of close to 100% (and 95% for ε2 if the effect is that is, the ε3ε3 group minus the ε2+ε4− group (since ε2 is comparable in size). hypothesised to be beneficial) and the ε4+ε2− group minus the 4. For MD from DTI, Westlye et al. (2012) reported an MD ε3ε3 group (since ε4 is hypothesised to be detrimental). This effect size for ε4 of d = +0.77, for which the present direction applies to both the mean and the linear slope of the poly- power is close to 100% in all cases. nomial effect of age (e.g. the slope of the age effect should be 5. For rsfMRI, Westlye et al. (2011) reported an effect size more negative for the ε3ε3 group than the ε2+ε4− group and for for ε4 of d = +1.23 in right hippocampus/amygdala (part the ε2+ε4− group than the ε3ε3 group). The quadratic term was of DMN), for which we have ample power (off the scale predicted, on the basis of the antagonistic pleiotropy hypothesis, in Figure 2). Note that, however, the effect sizes for this to be negative for all groups (i.e. an inverted U-shape), but can be functional connectivity effect and for the DTI effect in combined with the mean and linear terms to estimate the vertex the above Westlye et al.’s paper may be biased upwards (inflexion point) of a parabolic fit (Figure 1), as detailed above, because both were selected after thresholding voxels that which was predicted to be earlier for ε4-carriers than non-carriers. showed a basic effect of APOE status. The above categorical analyses allowed the effects of ε4 and ε2 to 6. For rsMEG, Koelewijn et al. (2019) reported an area- differ qualitatively. under-curve of the receiver-operating characteristic A second ‘parametric’, or gene-dose, GLM modelled a linear (ROC) of 63.5% for classification of ε4 status. This cor- effect of a decreasing number of ε2 alleles and increasing num- responds to a Cohen’s d of +0.49 (Salgado, 2018), for bers of ε4 alleles, which is potentially a more sensitive model if which this study has a power of close to 100%. the two alleles have the same quantitative (and additive) effects. The latter is consistent with evidence for a load effect for ε4 at least (e.g. Wisdom et al., 2011), including the cumulative Bayes factors increased risk for AD (Farrer et al., 1997). For the same tests on the six dependent variables, we also calculated the Bayes factors (BFs) for the null versus alternative hypothesis for Power calculation the APOE-by-age interaction terms. For this, we used the Savage– Dickey ratio, after logspline interpolation of the posterior sampling We simulated statistical power for various effect sizes and distribution, generated from 100,000 Markov chain Monte Carlo N = 608 participants. The number of 608 was an estimate of the (MCMC) iterations of the ‘brm’ function from the ‘brms’ R pack- number of CamCAN individuals who have valid cognitive and age (Bürkner, 2017, 2018) using the same linear model as for the genetic data; an estimate made before the genetic and pheno- above classical statistics. This is provided in the R script ‘apoe_lm_ typic data were combined (see section ‘Results’ for final num- brms.R’ on https://osf.io/ehs9n/. We imposed unit normal priors on bers for each phenotypic variable). To simulate random genetic all of the (Z-scored) polynomial terms for the age-by-APOE status sampling, multiple random draws were simulated from the interactions. Note that this is a deviation from Stage 1 of this study, expected population frequencies of each allele combination. 6 Brain and Neuroscience Advances Table 1. Number of outliers and valid data points for each phenotypic variable for each of the three genetic groups: ε2+ε4−, ε3ε3 and ε4+ε2−. Count N initial N outliers (excluded) N final Phenotypic variable ε2+ε4− ε4+ε2− Total ε2+ε4− ε4+ε2− Total ε3ε3 ε3ε3 Fluid intelligence 560 0 2 1 3 74 331 152 557 Episodic memory 592 0 1 1 2 78 355 157 590 Hippocampal volume 546 1 4 7 12 70 322 142 534 White-matter MD 525 7 26 12 45 63 286 131 480 rsfMRI in DMN 548 0 6 3 9 71 322 146 539 rsMEG PCA 514 2 7 7 16 67 295 136 498 MD: mean diffusivity; rsfMRI: resting-state functional magnetic resonance imaging; DMN: default mode network; rsMEG: resting-state magnetoencephalography; PCA: principal component analysis. which stated that we would use Jeffreys-Zellner-Siow (JZS) priors. This is because the sampling of a Cauchy prior on the mean (with scale 1/√2), as required by the JZS approach, produced unstable posterior distributions. Nonetheless, simulations show that the BFs are very similar for both priors (see https://jaquent.github.io/post/ comparing-different-methods-to-calculate-bayes-factors-for-a-sim- ple-model/). Covariates of no interest Covariates of no interest included: (1) sex (which, though bal- anced in CamCAN sample, has been shown to modulate APOE effects in some studies, e.g. Damoiseaux et al., 2012): (2) educa- tion level (since this tends to increase with year of birth in cross- sectional data and is known to correlate with cognitive measures later in life), (3) an estimate of socio-economic status (SES) and (4) a summary measure of cardiovascular health (first principal component of CamCAN’s measures of blood pressure and elec- Figure 3. Number of CamCAN participants (out of 610) with each APOE trocardiogram (ECG); see Tsvetanov et al., 2015), since this has genotype (filled bars) and number expected from White Europeans also been shown to be affected by APOE (e.g. Lyall et al., 2016; (empty bars). Oberlin et al., 2015). Separate models were fit with versus with- out covariates because education, cardiovascular health and age covary positively and therefore cannot be disentangled with Results cross-sectional data. Note that, while the Stage 1 report stated that genetic ancestry would also be a covariate, because it became There were a total of 651 genetic samples with matching pheno- clear after molecular genetic analysis that the majority of the typic IDs, of which three were excluded for low quality and sample had European white ancestry, we excluded the small three were excluded because they were related to others in the number of individuals with a different ancestry from the analyses sample. Of the remaining 645, principal component analyses of (see start of section ‘Results’), so it was no longer necessary to genome-wide SNP data showed that only 35 had non-European include ancestry as a covariate. ancestry, which we also excluded so as to optimise the homoge- SES was estimated by total family income, ranked into five neity of the sample, leaving a total sample of 610 participants. levels (<£18k, <£31k, <£52k, <£100k and >£100k per The frequency of APOE alleles closely matched that expected annum) and had 27 missing values. Education was also ranked from the frequency within healthy Europeans (Eisenberg et al., into five levels (no qualifications at age 16, practical qualifica- 2010), as seen in Figure 3. There were 78 in the ε2+ε4− group, tions at age 16, academic qualifications at age 16, qualifications 159 in the ε4+ε2− group and 357 in the ε3ε3 group (plus 16 at age 18 and university degree or higher), with 47 missing val- ε2ε4 carriers, who were not analysed further, as explained in ues. Cardiovascular health was the first principal component of section ‘Methods’). mean heart rate and heart rate variability (after low and high- The numbers of participants with data for each phenotypic pass filtering of the ECG recorded during the MEG session), as variable are shown in the leftmost column of Table 1. From these, well as systolic and diastolic blood pressure, after excluding outliers that were 1.5 times the interquartile range, after adjusting 169 missing values, respectively. All missing values were for linear and quadratic effects of age, were removed, producing replaced by the predictions of a quadratic fit of age to remaining the final numbers in the rightmost column of Table 1. The white- values, and the results Z-scored for each variable. matter MD measure contained the highest number of outliers, Henson et al. 7 Figure 4. Scatter plots for each phenotypic variable against age, grouped by APOE status (colour). Outliers are indicated by crosses (but were excluded from analyses). The solid lines show second-order polynomial fits of age on remaining points for each APOE group. most likely because diffusion-weighted MRI data are well-known ε4+ε2− versus ε3ε3; the third group of columns refers to (3) the to be sensitive to noise, such as that caused by head motion. parametric contrast of ε2−/ε4+ dose, with a linear increase across Scatter plots of the six phenotypic variables against age are ε2ε2, ε2ε3/ε3ε2, ε3ε3, ε3ε4/ε4ε3 and ε4ε4. The sign of these shown in Figure 4. Additional information regarding definition three contrasts matches the sign of the expected, prior effect size of these variables is provided in Supplemental Table 1. All six for ε4-carriers versus non-carriers reported in section ‘Power cal- variables showed significant effects of age (see Supplemental culation’ (i.e. the direction of effect was predicted to be negative Table 2). The fits come from a second-order polynomial expan- for first three phenotypic variables and positive for last three phe- sion of age to each APOE group. There was little apparent evi- notypic variables). Within each group, the parameter estimates for dence for antagonistic pleiotropy (i.e. different effects of age for the interactions between APOE status and zeroth-, first- and sec- each group), though the Cattell test of fluid intelligence showed ond-order polynomial effects of age are shown (where the zeroth- some suggestion of ε4-carriers performing worse in later life and order effect is equivalent to the main effect of APOE status). better in early life (i.e. for red line relative to blue and green lines The models fit the data well for the first four phenotypes in top left panel of Figure 4). Although the quadratic component (explaining 14%–70% of the variance). The fit was not so good was significant for three variables, it was never large relative to for fMRI and MEG (explaining only approximately 5% of the the linear component, and more importantly, it never differed sig- variance, though still significant), which we attribute to these nificantly between APOE groups (see the next section), meaning data being noisier. Only one of the three polynomial effects for that there was little value in comparing the peaks across APOE any of the three contrasts for any of the six phenotypic variables groups (i.e. age of maximum or minimum value). survived our pre-specified, one-tailed alpha value of 0.05/6. This was the zeroth-order effect on resting-state fMRI connectivity in the default mode network, where connectivity was higher on average for the ε4+ε2− group than ε3ε3 group (cf. red and blue Planned comparisons using classical lines in middle bottom panel of Figure 4). This is consistent with (frequentist) statistics Westlye et al. (2011), but because this main effect of the ε4 allele Results of the critical interactions between APOE and age are did not interact with either of the linear or quadratic age terms, it shown in Table 2. These come from the same linear model as provides no support for the antagonistic pleiotropy hypothesis. shown in Figure 4 (without any covariates). In the three groups of The only phenotypic variable to show any suggestion of the columns, the first two groups refer to the two planned categorical predicted interaction between APOE status and age was the contrasts across APOE groups of (1) ε3ε3 versus ε2+ε4−, (2) Cattell measure of fluid intelligence (surviving p < 0.05 but not 8 Brain and Neuroscience Advances Table 2. GLM results for each phenotypic variable (row). Contrast ε3ε3 group versus ε2+ε4− group ε4+ε2− group versus ε3ε3 group Parametric (dose) effect of ε2−ε4+ Poly. order Fit 0 1 2 Fit 0 1 2 Fit 0 1 2 2 2 2 Fluid intelligence R = 0.455 0.017 0.077 0.072 R = 0.500 −0.133 −0.228 0.151 R = 0.490 −0.192 −0.235 0.153 df = 399 (0.122) (0.158) (0.158) df = 477 (0.107) (0.115) (0.115) df = 551 (0.102) (0.102) (0.102) T = 0.139 T = 0.489 T = 0.458 T = 1.246 T = 1.982 T = 1.310 T = 1.881 T = 2.313 T = 1.507 2 2 2 Episodic memory R = 0.136 0.025 0.134 0.072 R = 0.145 −0.102 −0.251 0.023 R = 0.145 −0.138 −0.223 0.026 df = 427 (0.244) (0.319) (0.319) df = 506 (0.230) (0.249) (0.249) df = 584 (0.213) (0.213) (0.213) T = 0.102 T = 0.423 T = 0.225 T = 0.443 T = 1.006 T = 0.926 T = 0.648 T = 1.046 T = 0.121 2 2 2 Hippocampal volume R = 0.406 −22.12 −2.378 −19.86 R = 0.414 27.78 −3.697 2.579 R = 0.411 −7.161 −11.89 −9.45 df = 386 (14.81) (19.40) (19.40) df = 458 (13.68) (14.85) (14.85) df = 528 (12.69) (12.70) (12.71) T = 1.494 T = 0.123 T = 1.023 T = 2.031 T = 0.249 T = 0.174 T = 0.564 T = 0.936 T = 0.743 2 2 2 MD of WM tracts R = 0.710 8.54e-6 4.34e-6 7.42e-6 R = 0.701 −9.97e-6 −5.25e-6 −4.37e-6 R = 0.698 3.90e-6 4.74e-6 1.43e-6 df = 343 (5.56e-6) (7.26e-6) (7.26e-6) df = 411 (5.16e-6) (5.56e-6) (5.56e-6) df = 474 (4.82e-6) (4.84e-6) (4.81e-6) T = 1.536 T = 0.597 T = 1.023 T = 1.93 T = 0.945 T = 0.786 T = 0.809 T = 0.981 T = 0.298 2 2 2 DMN rsfMRI con- R = 0.042 1.28e-4 −3.82e-3 −2.35e-4 R = 0.060 7.66e-3 −1.38e-4 2.83e-3 R = 0.046 4.58e-3 −1.72e-3 −2.78e-4 nectivity df = 387 (2.84e-3) (3.70e-3) (3.70e-3) df = 462 (2.76e-3) (2.98e-3) (2.98e-3) df = 533 (2.56e-3) (2.56e-3) (2.56e-3) T = 0.045 T = 1.032 T = 0.634 T = 2.774 T = 0.046 T = 0.951 T = 1.77 T = 0.671 T = 0.914 2 2 2 rsMEG connectivity R = 0.051 0.059 0.047 −0.083 R = 0.042 −0.046 −0.033 0.005 R = 0.034 0.006 −0.007 −0.031 df = 356 (0.044) (0.056) (0.056) df = 425 (0.042) (0.044) (0.044) df = 492 (0.039) (0.039) (0.039) T = 1.362 T = 0.841 T = 1.486 T = 1.098 T = 0.721 T = 0.116 T = 0.150 T = 0.169 T = 0.793 MD: mean diffusivity; WM: white matter; DMN: default mode network; rsfMRI: resting state functional magnetic resonance imaging; rsMEG: resting state magnetoencephalography; Poly.: polynomial; df: degrees of freedom; R : adjusted R-squared of full model. The three groups of columns refer to the planned contrasts across APOE groups. Within each group, the first column gives the overall model fit, and the next three columns give the parameter estimates, their associated standard error (in brackets) and unsigned T-statistic for the interaction between the APOE contrast and the three polynomial expansions of age: zeroth (constant), first (linear) and second (quadratic), where zeroth-order term is equivalent to main effect of APOE contrast (see Supplemental Table 2 for parameters for main effects of Age). Effects with p < 0.05 are shown in bold, but note that only one survived the pre-specified Bonferroni correction for six multiple, one- tailed comparisons (for which|T| > 2.40 for the minimum number of df’s here; see text), where direction of effect was predicted to be negative for first three phenotypic variables and positive for last three phenotypic variables (see text). These results are without covariates; see Supplemental Table 3 for corresponding results with covariates. Henson et al. 9 Table 3. Bayes factors (BFs) for various hypotheses about the first-order (linear) effect of age being zero (BF01) for each contrast and each phenotypic variable, given various means for a unit normal prior. Contrast ε3ε3 group versus ε2+ε4− group ε4+ε2− group versus ε3ε3 group Parametric (dose) effect of ε2−/ε4+ BF01 Lin = 0 Lin < 0 Lin = P Lin = 0 Lin < 0 Lin = P Lin = 0 Lin < 0 Lin = P Fluid intelligence 18.63 29.79 29.42 4.08 2.09 5.99 2.30 1.16 3.31 Episodic memory 15.69 23.30 15.82 13.56 8.05 13.72 13.56 8.05 13.72 Hippocampal volume 19.40 17.71 22.65 25.11 20.95 29.20 19.43 11.79 22.15 BF01 Lin = 0 Lin > 0 Lin = P Lin = 0 Lin > 0 Lin = P Lin = 0 Lin > 0 Lin = P MD of WM tracts 22.21 15.35 29.31 22.17 63.92 30.52 24.36 14.58 32.18 DMN rsfMRI connectivity 9.10 29.84 21.08 20.66 21.45 43.90 18.91 27.59 41.65 rsMEG connectivity 10.59 6.62 11.65 15.13 32.11 17.38 22.26 25.64 25.28 BF: Bayes factor; MD: mean diffusivity; WM: white matter; DMN: default mode network; rsfMRI: resting state functional magnetic resonance imaging; rsMEG: resting state magnetoencephalography. The column headings for each analysis are explained in the text. The value of P in the column ‘Lin = P’ was determined from the literature cited in section ‘Power calculation’, that is, P = [−0.936, −0.14, −0.56, +0.77, +1.23, +0.49] for the six phenotypic variables. The direction of the one-tailed test in the column (i.e. ‘Lin > 0’ or ‘Lin < 0’) was determined by the sign of P (i.e. negative for the first three phenotypic variables and positive for the remaining three). correction for multiple comparisons), for which the slope was for the linear interaction between age and APOE status being more negative for the ε4+ε2− group than ε3ε3 group (cf. red and greater or less than zero ( ‘Lin > 0’ or ‘Lin < 0’), given a prior of blue lines in top left panel of Figure 4), and was negatively greater or less than zero, where the greater/lesser direction depends related to increasing ε2−/ε4+ dose in the parametric model. on the analysis (i.e. ε2 being protective would predict a more posi- The only other effect with p < 0.05 (uncorrected) was the tive (less negative) slope than the reference group, and ε4 being a zeroth-order effect for the ε4+ε2− group on hippocampal vol- risk factor would predict a more negative slope than the reference ume, which tended to be larger on average than in the ε3ε3 group. group) and (3) BFs for the linear interaction between age and However, we suspect that this is a false positive since it was in APOE status being zero given the prior expectation of an effect the opposite direction to our predictions, and so we do not con- size (P) equal to that from the literature that was used to power this sider it further. study ( ‘Lin = P’), as listed in section ‘Power calculation’. In all cases, the BFs provided ‘substantial’ (BF > 3) or ‘strong’ (BF > 10) evidence (https://en.wikipedia.org/wiki/Bayes_factor) Adjusting for covariates for no interaction between APOE status and age on any of the phenotypic variables, regardless of whether their prior expecta- The above linear model was repeated with five additional covari- tion was equal to zero or equal to the effect size from the litera- ates: male/female sex, education level, SES and cardiovascular ture, with the exception of fluid intelligence: While there was health. The inclusion of these covariates did not reveal any new strong evidence for no interaction between the ε2 allele and age significant effects of APOE (Supplemental Table 3). The greater on fluid intelligence, this was not true for the predicted direction DMN fMRI connectivity for the ε4+ε2− group relative to the of interaction between ε4 allele and age (or of overall ε4 load), ε3ε3 group continued to survive correction, and the parametric where the BFs were around 2, that is, ambiguous. effect of dose on the linear age slope for fluid intelligence contin- ued to survive p < 0.05 uncorrected. The linear interaction between age and ε4+ε2− group and ε3ε3 group on fluid intelli- General discussion gence no longer survived p < 0.05 after adjusting for covariates. In short, there was no evidence for the antagonistic pleiotropy This study provided no support for, and mostly evidence against, hypothesis using classical null-hypothesis testing. Since this the ‘antagonistic pleiotropy’ hypothesis (Han and Bondi, 2008), could reflect false negatives, given our relatively small sample whereby the ε4 variant of APOE is proposed to be advantageous for genetic effects, we calculated BFs for the null versus alternate in early life but disadvantageous in later life. The study also pro- hypotheses. vided evidence against the related hypothesis that the ε2 variant of APOE is advantageous, particularly in later life (Suri et al., 2013). While our sample was relatively small for genetic analysis BFs of cognitive and neural phenotypes, it was sufficiently powered Since the dominant effect of age on phenotypic variables was for classical statistics when based on previous effect sizes linear (see Supplemental Table 1), and linear effects were reported, and moreover, furnished BFs that provided evidence in reported in the prior literature (see section ‘Power calculation’), favour of the null hypothesis that these APOE variants do not we only report here the BFs for the first-order polynomial term. interact (linearly) with age in their putative effects on brain struc- The three columns in Table 3 show, for each APOE analysis: ture or cognition. (1) BFs for the linear interaction between Age and APOE status The only effect of APOE status that survived our a priori cor- being zero, given a prior expectation of zero ( ‘Lin = 0’), (2) BFs rection for the six phenotypic variables tested was the main effect 10 Brain and Neuroscience Advances of ε4-carriers (more precisely, our ε4+ε2− group versus our ε3ε3 is a cross-sectional sample, which may be confounded with effects reference group) on the mean fMRI functional connectivity of birth-year and is unable to examine true ageing within individu- within the default mode network. This reflected higher connec- als. Fourth, we cannot discount a role for antagonistic pleiotropy for tivity for the ε4-carriers, consistent with Westlye et al. (2012). phenotypic measures that we did not examine in this study such as More importantly, there was no evidence that this effect inter- other measures of cognition or the brain. acted with age, and therefore even if it reflects a true genetic effect on brain functional connectivity, it is not evidence for the Acknowledgements antagonistic pleiotropy hypothesis. The authors thank Kamen Tsvetanov for providing the cardiovascular While all phenotypic variables showed strong associations summary measures and Alex Quent for help with the Bayes Factor with age, there was only one suggestion that such age associa- estimation. tions depended on APOE status, namely, a steeper, linear age- related decline in fluid intelligence in the ε4+ε2− group than Declaration of conflicting interests ε3ε3 group. This effect did not survive correction for multiple The authors declare that there is no conflict of interest. comparisons. Moreover, the BF for this linear effect was ambigu- ous, so this effect needs replication before providing support for the antagonistic pleiotropy hypothesis. Funding None of the analyses showed any evidence of differences for The author(s) disclosed receipt of the following financial support for the ε2-carriers (i.e. when comparing our ε4+ε2− group versus our research, authorship and/or publication of this article: The Cambridge ε3ε3 reference group). While the number of ε2-carriers was Centre for Ageing and Neuroscience (CamCAN) research was supported lower than ε4-carriers, the lack of interaction between ε2 and age by the Biotechnology and Biological Sciences Research Council (Grant is unlikely to simply reflect low power, because BFs provided No. BB/H008217/1). This project has also received funding from the European Union’s Horizon 2020 research and innovation programme evidence (BFs > 9) for the interaction being zero. (‘LifeBrain’, Grant Agreement No. 732592). R.N.H. was supported by The classical power estimates may have been over-estimated the Medical Research Council (SUAG/010 RG91365). S.S. was sup- because the effect sizes reported in the literature are likely to be ported by an Alzheimer’s Society Junior Research Fellowship (Grant inflated owing to publication bias and/or ‘winner’s curse’. Reference No. 441). E.K. was supported by the LifeBrain grant. R.A.K. Indeed, the BFs for the null hypothesis were highest when the was supported by the UK Medical Research Council (SUAG/014 prior mean was based on a published effect size. Nonetheless, the RG91365). J.B.R. was supported by the Wellcome Trust (103838) and the BF still provided substantial to strong evidence for no effect even Medical Research Council (SUAG/051 RG91365). D.C. was funded by when the prior mean was zero. the Wellcome Trust and Alzheimer’s Research UK. E.E. and S.E.F. were While this study had lower power than many previous stud- supported by the Max Planck Society. ies of APOE on cognition, it is nonetheless larger than many prior APOE studies using neuroimaging measures, particularly ORCID iD measures of functional connectivity using fMRI or MEG, which Richard N. Henson https://orcid.org/0000-0002-0712-2639 often use small and/or biased samples of the population. Relative to these studies, this study gains power by virtue of the wide and Supplemental material near-uniform age range from 18 to 88 years, in a population- derived sample. Nevertheless, it is important to note that, by Supplemental material for this article is available online. only considering adults, we cannot address a version of the antagonistic pleiotropy hypothesis in which the interactions References between APOE status and age only occur during cognitive and Belloy ME, Napolioni V and Greicius MD (2019) A quarter century of brain development, that is, in individuals under 18 years. This APOE and Alzheimer’s disease: Progress to date and the path for- issue could be examined in larger cognitive and neuroimaging ward. Neuron 101(5): 820–838. lifespan cohorts like the European LifeBrain consortium Blacker D, Haines JL, Rodes L, et al. (1997) ApoE-4 and age at onset (Walhovd et al., 2018). of Alzheimer’s disease: The NIMH genetics initiative. Neurology Finally, there are other caveats associated with our study. First, 48(1): 139–147. there may be other types of bias in the selection of adults used in Buckner RL, Andrews-Hanna JR and Schacter DL (2008) The brain’s this study, particularly those towards the older end of our age range, default network: Anatomy, function, and relevance to disease. for example, owing to survival bias and exclusion of those with Annals of the New York Academy of Sciences 1124(1): 1–38. evidence of cognitive decline. Indeed, it is possible that the effects Bunce D, Anstey KJ, Cherbuin N, et al. (2012) APOE genotype and ento- of APOE variants are only detectable in individuals with AD pathol- rhinal cortex volume in non-demented community-dwelling adults in midlife and early old age. Journal of Alzheimer’s Disease 30(4): ogy (e.g. Vemuri et al., 2010) and there were few such people in our 935–942. cohort, even those in the pre-symptomatic phase of AD. There will Bürkner PC (2017) brms: An R package for Bayesian multilevel models also be a bias induced by the exclusion of those who were unable to using Stan. Journal of Statistical Software. Epub ahead of print 29 undergo an MRI scan, for example, because of cardiovascular prob- August. DOI: 10.18637/jss.v080.i01. lems that require a pacemaker or stent, but which might also be Bürkner PC (2018) Advanced Bayesian multilevel modeling with the R related to APOE allelic status and cognitive/brain health. Second, package brms. The R Journal 80(1): 1–28. our sample only included White Europeans (Caucasians), whereas Cuesta P, Garcés P, Castellanos NP, et al. (2015) Influence of the APOE APOE variants may have stronger effects in other genetic groups, ε4 allele and mild cognitive impairment diagnosis in the disruption and we did not have any direct measures of neuropathology or take of the MEG resting state functional connectivity in sources space. into account potential effects of a range of medications. Third, this Journal of Alzheimer’s Disease 44(2): 493–505. Henson et al. 11 Damoiseaux JS, Seeley WW, Zhou J, et al. (2012) Gender modulates the Koelewijn L, Lancaster TM, Linden D, et al. (2019) Oscillatory hyperac- APOE ε4 effect in healthy older adults: Convergent evidence from tivity and hyperconnectivity in young APOE-ɛ4 carriers and hypo- functional brain connectivity and spinal fluid tau levels. The Journal connectivity in Alzheimer’s disease. Elife 8: e36011. of Neuroscience 32(24): 8254–8262. Lancaster C, Tabet N and Rusted J (2017) The elusive nature of Davies G, Armstrong N, Bis JC, et al. (2015) Genetic contributions to APOE ε4 in mid-adulthood: Understanding the cognitive profile. variation in general cognitive function: A meta-analysis of genome- Journal of the International Neuropsychological Society 23(3): wide association studies in the CHARGE consortium (N = 53 949). 239–253. Molecular Psychiatry 20(2): 183–192. Lyall DM, Celis-Morales C, Lyall LM, et al. (2019a) Assessing for inter- Deary IJ, Whiteman MC, Pattie A, et al. (2002) Ageing: Cognitive action between APOE ε4, sex, and lifestyle on cognitive abilities. change and the APOE ε4 allele. Nature 418(6901): 932. Neurology 92(23): e2691–e2698. de Mooij SMM, Henson RNA, Waldorp LJ, et al. (2018) Age differen- Lyall DM, Cox SR, Lyall LM, et al. (2019b) Association between APOE tiation within gray matter, white matter, and between memory and e4 and white matter hyperintensity volume, but not total brain vol- white matter in an adult life span cohort. The Journal of Neurosci- ume or white matter integrity. Brain Imaging and Behavior. Epub ence 38(25): 5826–5836. ahead of print 22 March. DOI: 10.1007/s11682-019-00069-9. Eisenberg DTA, Kuzawa CW and Hayes MG (2010) Worldwide allele Lyall DM, Harris SE, Bastin ME, et al. (2014) Alzheimer’s disease sus- frequencies of the human apolipoprotein E gene: Climate, local ceptibility genes APOE and TOMM40, and brain white matter integ- adaptations, and evolutionary history. American Journal of Physical rity in the Lothian Birth Cohort 1936. Neurobiology of Aging 35(6): Anthropology 143(1): 100–111. 1513e25–1513e33. Farrer LA, Cupples LA, Haines JL, et al. (1997) Effects of age, sex, and Lyall DM, Royle NA, Harris SE, et al. (2013) Alzheimer’s disease sus- ethnicity on the association between apolipoprotein E genotype and ceptibility genes APOE and TOMM40, and hippocampal volumes in Alzheimer’s disease. JAMA 278(16): 1349–1356. the Lothian Birth Cohort 1936. PLoS One 8(11): e80513. Filippini N, MacIntosh BJ, Hough MG, et al. (2009a) Distinct patterns Lyall DM, Ward J, Ritchie SJ, et al. (2016) Alzheimer’s disease genetic of brain activity in young carriers of the APOE-epsilon4 allele. Pnas risk factor APOE e4 and cognitive abilities in 111,739 UK Biobank 106(17): 7209–7214. participants. Age and Ageing 45(4): 511–517. Filippini N, Rao A, Wetten S, et al. (2009b) Anatomically-distinct genetic Machulda MM, Jones DT, Vemuri P, et al. (2011) Effect of APOE ε4 associations of APOE e{open}4 allele load with regional cortical status on intrinsic network connectivity in cognitively normal elderly atrophy in Alzheimer’s disease. Neuroimage 44(3): 724–728. subjects. Archives of Neurology 68(9): 1131–1136. Fleisher AS, Sherzai A, Taylor C, et al. (2009) Resting-state BOLD Marioni RE, Campbell A, Scotland G, et al. (2016) Differential effects networks versus task-associated functional MRI for distinguishing of the APOE e4 allele on different domains of cognitive ability Alzheimer’s disease risk groups. Neuroimage 47(4): 1678–1690. across the life-course. European Journal of Human Genetics 24(6): Fouquet M, Besson FL, Gonneaud J, et al. (2014) Imaging brain effects 919–923. of APOE4 in cognitively normal individuals across the lifespan. Mondadori CRA, De Quervain DJF, Buchmann A, et al. (2007) Better Neuropsychology Review 24(3): 290–299. memory and neural efficiency in young apolipoprotein E ε4 carriers. Fullerton SM, Clark AG, Weiss KM, et al. (2000) Apolipoprotein E vari- Cerebral Cortex 17(8): 1934–1947. ation at the sequence haplotype level: Implications for the origin and Oberlin LE, Manuck SB, Gianaros PJ, et al. (2015) Blood pressure inter- maintenance of a major human polymorphism. American Journal of acts with APOE ε4 to predict memory performance in a midlife sam- Human Genetics 67(4): 881–900. ple. Neuropsychology 29(5): 693–702. Geerligs L, Tsvetanov KA, Cam-CAN, et al. (2017) Challenges in O’Donoghue MC, Murphy SE, Zamboni G, et al. (2018) APOE genotype measuring individual differences in functional connectivity using and cognition in healthy individuals at risk of Alzheimer’s disease: fMRI: The case of healthy aging. Human Brain Mapping 38(8): A review. Cortex 104(July): 103–123. 4125–4156. Pietzuch M, King AE, Ward DD, et al. (2019) The influence of genetic Guo Y, He J, Zhao S, et al. (2014) Illumina human exome genotyp- factors and cognitive reserve on structural and functional resting- ing array clustering and quality control. Nature Protocols 9(11): state brain networks in aging and Alzheimer’s disease. Frontiers in 2643–2662. Aging Neuroscience 11(March): 30. Habes M, Toledo JB, Resnick SM, et al. (2016) Relationship between Puttonen S, Elovainio M, Kivimäki M, et al. (2003) The combined effects APOE genotype and structural MRI measures throughout adulthood of apolipoprotein E polymorphism and low-density lipoprotein choles- in the study of health in Pomerania population-based cohort. Ameri- terol on cognitive performance in young adults. Neuropsychobiology can Journal of Neuroradiology 37(9): 1636–1642. 48(1): 35–40. Han SD and Bondi MW (2008) Revision of the apolipoprotein E com- Radmanesh F, Devan WJ, Anderson CD, et al. (2014) Accuracy of impu- pensatory mechanism recruitment hypothesis. Alzheimer’s and tation to infer unobserved APOE epsilon alleles in genome-wide Dementia 4(4): 251–254. genotyping data. European Journal of Human Genetics 22(10): Heise V, Filippini N, Ebmeier KP, et al. (2011) The APOE ɛ4 allele 1239–1242. modulates brain white matter integrity in healthy adults. Molecular Rocchi A, Pellegrini S, Siciliano G, et al. (2003) Causative and suscepti- Psychiatry 16(9): 908–916. bility genes for Alzheimer’s disease: A review. Brain Research Bul- Ihle A, Bunce D and Kliegel M (2012) APOE ε4 and cognitive function letin 61(1): 1–24. in early life: A meta-analysis. Neuropsychology 26(3): 267–277. Rusted JM, Evans SL, King SL, et al. (2013) APOE e4 poly- morphism Jack CR, Wiste HJ, Weigand SD, et al. (2015) Age, sex, and APOE ε4 in young adults is associated with improved attention and indexed by effects on memory, brain structure, and β-amyloid across the adult distinct neural signatures. NeuroImage 65: 364–373. life Span. JAMA Neurology 72(5): 511–519. Salgado JF (2018) Transforming the area under the normal curve (AUC) Jochemsen HM, Muller M, Van der Graaf Y, et al. (2012) APOE ε4 dif- into Cohen’s d, Pearson’s r , odds-ratio, and natural log odds-ratio: pb ferentially influences change in memory performance depending on Two conversion tables. European Journal of Psychology Applied to age. The SMART-MR study. Neurobiology of Aging 33(4): 832. Legal Context 10(1): 35–47. e15–832.e22. Schiepers OJG, Harris SE, Gow AJ, et al. (2012) APOE E4 status pre- Jorm AF, Mather KA, Butterworth P, et al. (2007) APOE genotype and dicts age-related cognitive decline in the ninth decade: Longitudinal cognitive functioning in a large age-stratified population sample. follow-up of the Lothian Birth Cohort 1921. Molecular Psychiatry Neuropsychology 21(1): 1–8. 17(3): 315–324. 12 Brain and Neuroscience Advances Schultz MR, Lyons MJ, Franz CE, et al. (2008) Apolipoprotein E geno- and functional MRI, MEG, and cognitive data from a cross-sectional type and memory in the sixth decade of life. Neurology 70(19 PART adult lifespan sample. Neuroimage 144(Pt B): 262–269. 2): 1771–1777. Trachtenberg AJ, Filippini N, Ebmeier KP, et al. (2012) The effects of Shafto M, a Tyler LK, Dixon M, et al. (2014) The Cambridge Centre APOE on the functional architecture of the resting brain. Neuroim- for Ageing and Neuroscience (Cam-CAN) study protocol: A cross- age 59(1): 565–572. sectional, lifespan, multidisciplinary examination of healthy cogni- Tsvetanov KA, Henson RNA, Tyler LK, et al. (2015) The effect of age- tive ageing. BMC Neurology 14(1): 204. ing on fMRI: Correction for the confounding effects of vascular reac- Sheline YI, Morris JC, Snyder AZ, et al. (2010) APOE4 allele disrupts resting tivity evaluated by joint fMRI and MEG in 335 adults. Human Brain state fMRI connectivity in the absence of amyloid plaques or decreased Mapping 36(6): 2248–2269. CSF Aβ42. The Journal of Neuroscience 30(50): 17035–17040. Tsvetanov KA, Henson RNA, Tyler LK, et al. (2016) Extrinsic and Shin MH, Kweon SS, Choi JS, et al. (2014) The effect of an APOE poly- intrinsic brain network connectivity maintains cognition across the morphism on cognitive function depends on age. Journal of Neurol- lifespan despite accelerated decay of regional brain activation. Jour- ogy 261(1): 66–72. nal of Neuroscience 36(11): 3115–3126. Shu H, Shi Y, Chen G, et al. (2016) Opposite neural trajectories of apo- Vemuri P, Wiste HJ, Weigand SD, et al. (2010) Effect of apolipopro- lipoprotein E ε4 and ε2 alleles with aging associated with different tein E on biomarkers of amyloid load and neuronal pathology in risks of Alzheimer’s disease. Cerebral Cortex 26(4): 1421–1429. Alzheimer’s disease 67(3): 308–316. Siebner HR, Callicott JH, Sommer T, et al. (2009) From the genome to Walhovd KB, Fjell AM, Westerhausen R, et al. (2018) Healthy minds the phenome and back: Linking genes with human brain function and 0–100 years: Optimising the use of European brain imaging cohorts structure using genetically informed neuroimaging. Neuroscience (‘Lifebrain’). European Psychiatry 47(Jan): 76–87. 164(1): 1–6. Westlye ET, Lundervold A, Rootwelt H, et al. (2011) Increased hip- Stebbins GT and Murphy CM (2009) Diffusion tensor imaging in pocampal default mode synchronization during rest in middle-aged Alzheimer’s disease and mild cognitive impairment. Behavioural and elderly APOE ε4 carriers: Relationships with memory perfor- Neurology 21(1–2): 39–49. mance. The Journal of Neuroscience 31(21): 7775–7783. Suri S, Heise V, Trachtenberg AJ, et al. (2013) The forgotten APOE Westlye LT, Reinvang I, Rootwelt H, et al. (2012) Effects of APOE allele: A review of the evidence and suggested mechanisms for on brain white matter microstructure in healthy adults. Neurology the protective effect of APOE e2. Neuroscience and Biobehavioral 79(19): 1961–1969. Reviews 37(10): 2878–2886. Wisdom NM, Callahan JL and Hawkins KA (2011) The effects of apoli- Taylor JL, Scanlon BK, Farrell M, et al. (2015) APOE-epsilon4 and poprotein E on non-impaired cognitive functioning: A meta-analysis. aging of medial temporal lobe gray matter in healthy adults older Neurobiology of Aging 32(1): 63–74. than 50 years. Neurobiology of Aging 35(11): 2479–2485. Yu YW, Lin CH, Chen SP, et al. (2000) Intelligence and event-related Taylor JR, Williams N, Cusack R, et al. (2017) The Cambridge Centre for potentials for young female human volunteer apolipoprotein E epsi- Ageing and Neuroscience (Cam-CAN) data repository: Structural lon4 and non-epsilon4 carriers. Neuroscience Letters 294(3): 179–181. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Brain and Neuroscience Advances SAGE

Effect of apolipoprotein E polymorphism on cognition and brain in the Cambridge Centre for Ageing and Neuroscience cohort:

Loading next page...
 
/lp/sage/effect-of-apolipoprotein-e-polymorphism-on-cognition-and-brain-in-the-KInWz0McCM
Publisher
SAGE
Copyright
Copyright © 2022 by SAGE Publications Ltd and British Neuroscience Association, unless otherwise noted. Manuscript content on this site is licensed under Creative Commons Licenses
ISSN
2398-2128
eISSN
2398-2128
DOI
10.1177/2398212820961704
Publisher site
See Article on Publisher Site

Abstract

Polymorphisms in the apolipoprotein E (APOE) gene have been associated with individual differences in cognition, brain structure and brain function. For example, the ε4 allele has been associated with cognitive and brain impairment in old age and increased risk of dementia, while the ε2 allele has been claimed to be neuroprotective. According to the ‘antagonistic pleiotropy’ hypothesis, these polymorphisms have different effects across the lifespan, with ε4, for example, postulated to confer benefits on cognitive and brain functions earlier in life. In this stage 2 of the Registered Report – https:// osf.io/bufc4, we report the results from the cognitive and brain measures in the Cambridge Centre for Ageing and Neuroscience cohort (www.cam-can. org). We investigated the antagonistic pleiotropy hypothesis by testing for allele-by-age interactions in approximately 600 people across the adult lifespan (18–88 years), on six outcome variables related to cognition, brain structure and brain function (namely, fluid intelligence, verbal memory, hippocampal grey-matter volume, mean diffusion within white matter and resting-state connectivity measured by both functional magnetic resonance imaging and magnetoencephalography). We found no evidence to support the antagonistic pleiotropy hypothesis. Indeed, Bayes factors supported the null hypothesis in all cases, except for the (linear) interaction between age and possession of the ε4 allele on fluid intelligence, for which the evidence for faster decline in older ages was ambiguous. Overall, these pre-registered analyses question the antagonistic pleiotropy of APOE polymorphisms, at least in healthy adults. Keywords Cognition, apolipoprotein E, lifespan, brain, ageing Received: 13 December 2019; accepted: 27 August 2020 Introduction MRC Cognition & Brain Sciences Unit, University of Cambridge, Cambridge, UK Apolipoprotein E (APOE) is a protein that plays an important Department of Psychiatry, University of Cambridge, Cambridge, UK role in lipid metabolism (including cholesterols) and has been Department of Psychiatry, Warneford Hospital, University of Oxford, implicated in synaptogenesis, repair of injured nerve tissue and Oxford, UK the modulation of beta-amyloid plaques and neurofibrillary tan- Wellcome Centre for Integrative Neuroimaging, University of Oxford, gles that characterise Alzheimer’s disease (AD) (for review, see Oxford, UK Belloy et al., 2019; Rocchi et al., 2003). The gene coding for Department of Clinical Neurosciences, University of Cambridge, APOE is located on chromosome 19 and is polymorphic in the Cambridge, UK general population. The three most common alleles are ε2, ε3 and Institute of Health and Wellbeing, University of Glasgow, Glasgow, UK ε4, with approximate allele frequencies of 6%, 78% and 15% in Institute of Cognitive Neuroscience, University College London, London, UK healthy Caucasian Europeans (Eisenberg et al., 2010). Possession Language and Genetics Department, Max Planck Institute for of the ε4 allele has been associated with poorer cognitive abilities Psycholinguistics, Nijmegen, The Netherlands and more rapid longitudinal decline in healthy older people, par- Donders Institute for Brain, Cognition and Behaviour, Radboud ticularly in episodic memory (e.g. Jack et al., 2015; Jochemsen University, Nijmegen, The Netherlands et al., 2012; Jorm et al., 2007; Lyall et al., 2016; Marioni et al., 2016; Mondadori et al., 2007; Schiepers et al., 2012; Schultz Corresponding author: et al., 2008; Shin et al., 2014; see Wisdom et al., 2011 for a meta- Richard N. Henson, MRC Cognition & Brain Sciences Unit, University of analysis). It has also been associated with a three-to-fourfold Cambridge, 15 Chaucer Road, Cambridge CB2 7EF, UK. increase in the risk of late onset AD in a gene dose-dependent Email: rik.henson@mrc-cbu.cam.ac.uk Creative Commons CC BY: This article is distributed under the terms of the Creative Commons Attribution 4.0 License (https://creativecommons.org/licenses/by/4.0/) which permits any use, reproduction and distribution of the work without further permission provided the original work is attributed as specified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/open-access-at-sage). 2 Brain and Neuroscience Advances Figure 1. Illustration of potential interactions between APOE status Figure 2. Classical power analyses for (any) polynomial term of GLM and age on cognitive and/or brain health. for (1) separate, categorical analysis of ε3ε3 versus ε2+ε4− (green According to the ‘antagonistic pleiotropy’ hypothesis, ε4-carriers (red line) line), (2) separate, categorical analysis of ε4+ε2− versus ε3ε3 (red experience a benefit in early life but a detriment in later life, relative to non- line) and (3) linear, parametric ‘dose’ effect of ε2−/ε4+ load (blue carriers (ε3ε3 group, blue line). The characteristics of ε2-carriers (green line) are less clear, but generally thought to benefit later in life. Note that every person dotted line). The vertical lines and numbers refer to prior effect sizes has two alleles, so ‘ε4-carrier’ (indicated by ‘ε4*’ in legend) means someone with in literature, as described in text. ε4ε4 or ε3ε4 (but excluding ε4ε2 – see text for more details). manner (Farrer et al., 1997) and with an earlier age at onset by lifespan cohort of healthy adults uniformly distributed from 18 to nearly 6 years on average for ε4 carriers (Blacker et al., 1997). 88 years, which allows us to overcome many of the limitations of Indeed, some have argued that APOE has no influence on cogni- former studies of the effects of ε4 at different ages. tion in mid- or late-life beyond increasing risk for AD, such that Another strand of the literature has examined the possible effects found on cognition reflect the decades of a pre-sympto- protective effects of the ε2 allele, which has been reported to matic period of AD pathology (e.g. Vemuri et al., 2010; see also reduce risk of AD (Farrer et al., 1997). The relationship of these O’Donoghue et al., 2018). putative protective effects with age is less well studied (partly There have also been claims that the ε4 allele confers cogni- due to its rarer frequency), but one suggestion is that ε2 is associ- tive benefits earlier in life, that is, individuals aged 5–35 years ated with more robust neurodevelopment in early life (relative to (e.g. Puttonen et al., 2003; Rusted et al., 2013; Yu et al., 2000, ε3), while another is that it is associated with lower levels of though, see Ihle et al., 2012 for a meta-analysis of studies that neurodegeneration (relative to ε4) in later life (Suri et al., 2013). found no consistent effect of ε4 status on cognitive tests in this CamCAN’s lifespan cohort allows novel tests for these interac- age-group). These claims led to the ‘antagonistic pleiotropy’ tive or additive effects of the ε2 allele with age (green line in hypothesis of Han and Bondi (2008), whereby the APOE ε4 Figure 1). allele is hypothesised to be advantageous in early life but disad- The mean effect size of ε4-carrier versus non-carrier status on vantageous in later life (and potentially neutral in middle-age, cognition is typically small when averaging across ages. For consistent with the lack of association with cognition in a meta- example, the meta-analysis of Wisdom et al. (2011) showed a analysis of middle-aged people by Lancaster et al., 2017). The Cohen effect size (d) across studies of d = −0.14 on episodic antagonistic pleiotropy hypothesis is illustrated in Figure 1 (cf. memory and d = −0.05 on global cognition, averaged across ages red versus blue lines). from 20 to 90 years (where d = 0.20 is considered ‘small’). This Cognitive benefits in early life (while still fertile), or impair- effect size would be difficult to detect with the approximately ments that only arise later in life, might explain in evolutionary 600 participants in the CamCAN imaging sample who have valid terms why the ε4 allele persists in the population at a relatively cognitive, neuroimaging and genetic data, including APOE status high frequency (Fullerton et al., 2000). However, most studies of (see Figure 2); a sample that is considerably smaller than several ε4 have examined young, middle-aged or older samples sepa- recent large studies (e.g. Lyall et al., 2019a; Marioni et al., 2016; rately. Moreover, studies of older participants are often biased by Shin et al., 2014). However, if the antagonistic pleiotropy hypoth- the over representation of super-healthy individuals who are esis is correct, then the effect size depends on age, such that the motivated to respond to adverts or join volunteer panels, as effect size for the interaction between APOE and age could be opposed to population-based recruitments. bigger than the effect size averaging over age (see Figure 1 and To examine the effect of APOE across the adult lifespan, we section ‘Methods’). Consistent with this view, the meta-analysis used the ‘imaging’ sample of Cambridge Centre for Ageing and of Wisdom et al. (2011) also found a significant linear effect of Neuroscience (CamCAN; www.cam-can.org) to test for a con- age, with the detrimental effect of ε4 increasing with age. tinuous interaction between age and APOE status on a variety of Similarly, more recent studies (e.g. Lyall et al., 2019b; Marioni cognitive and brain measures. CamCAN is a population-derived, et al., 2016; Shin et al., 2014) have generally reported stronger Henson et al. 3 detrimental effects of ε4 in older than younger groups. Moreover, samples of 73 and 203 adults, respectively, found similar effects in a large (genome-wide) study of multiple cohorts, Davies et al. of ε4- and ε2-carriers (relative to ε3ε3) on various DTI metrics, (2015) found that rs10119 – a single-nucleotide polymorphism including increases in mean diffusivity (MD) and decreases in (SNP) neighbouring the APOE locus – showed a negative corre- fractional anisotropy (FA), where both of these changes are reflec- lation of R = −0.424 between the average cohort age (from 55 to tive of poorer WM health. Similar MD increases and FA decreases 80 years) and the effect size for this SNP on general cognitive have been reported in AD (Stebbins and Murphy, 2009). However, ability. This corresponds to an interaction effect size of d = −0.936 the parallel rather than opposite effect of ε4 and ε2, coupled with (Figure 2). Furthermore, the power to detect such an interaction the fact that neither study found any interactions with age, led between age and APOE status is likely to benefit from the wider both groups to suggest that the APOE variants reflected neurode- age range in CamCAN than is typically available in the cohorts velopmental differences, rather than any changes related to AD. A that have been tested so far (such as the UK Biobank), which study of N ~ 650 older individuals (aged ~73 years) found reduced have tended to focus on middle and older ages. FA in ε4-carriers (but not ε2-carriers) in two of 15 major white- In addition to cognition, the APOE variants have been associ- matter tracts, but not in the first principal component across tracts ated with differences in brain structure and function, as measured (Lyall et al., 2014). A more recent and much larger study from the with magnetic resonance imaging (MRI) or magnetoencephalog- UK Biobank (Lyall et al., 2019b) found neither effect of ε4 nor raphy (MEG). Indeed, it is possible that such allele-specific interaction with age, on the first principal component of MD or FA effects might be seen earlier in the lifespan than they can be seen across 22 WM tracts. However, as the authors noted, the restricted in cognition (Siebner et al., 2009). This could be because neuro- age range of mostly 50–70 years may have limited their ability to imaging measures are more sensitive than behavioural measures detect age effects. CamCAN’s larger range of ages make it better at detecting subclinical abnormalities. Another possibility for the suited to investigate the effects of ε4 and ε2 and their interactions lower sensitivity of cognitive measures (versus brain measures) is with age. that young ε4-carriers show compensatory changes (e.g. increased Several studies have assessed functional connectivity between functional activity or connectivity) that maintain cognitive perfor- brain regions, acquiring, for example, blood oxygenation level mance comparable to non-carriers; however, at some point in dependent (BOLD)-weighted MRI while people rest. There has older age, these compensatory mechanisms fail and a deleterious been a focus on the default mode network (DMN), which includes effect on cognition is unmasked. A related idea is that the increased hippocampus as well as medial parietal and medial frontal regions. activity in younger ε4-carriers ‘may cause neurophysiological Using independent component analysis (ICA), for example, changes that lead to earlier age-related decline in brain function’ Westlye et al. (2011) reported increased expression of DMN con- (Buckner et al., 2008; Filippini et al., 2009a). Both of these nectivity for N = 33 ε4-carriers versus N = 62 non-carriers (aged hypotheses predict an interaction between APOE variant and age. 50–80 years). Based on the similarity of the DMN and the regions Indeed, if one were to find a greater interaction between APOE disrupted in AD, it has been suggested that enhanced metabolism and age on brain measures than on cognitive measures, this would in the DMN may provide regional conditions that are conducive support the idea that brain function compensates for the detrimen- to amyloid deposition (Buckner et al., 2008). However, tal effects of APOE on brain structure, compensating in youth but Trachtenberg et al. (2012) found similar effects of ε4 and ε2 on no longer being able to compensate in old age. functional connectivity in a young group and argued that the Many neuroimaging studies of APOE variants have focused effects of APOE do not relate only to AD risk, but rather to the on volumetric differences in structural (e.g. T1-weighted) MRI putative role of APOE in neurodevelopment (see also Deary et al., images, particularly atrophy of grey matter in regions of the 2002; O’Donoghue et al., 2018). Moreover, other results are medial temporal lobe, such as the hippocampus, which occurs divergent, with both increased and decreased functional connec- early in AD and is associated with memory problems. tivity reported for APOE variants (e.g. Damoiseaux et al., 2012; Unfortunately, the results have been mixed, with most studies Filippini et al., 2009a; Fleisher et al., 2009; Machulda et al., 2011; not finding a mean effect of APOE, ε4 or ε2, particularly in hip- Pietzuch et al., 2019; Sheline et al., 2010 for review). Indeed, pocampus (e.g. Bunce et al., 2012; Filippini et al., 2009b; Habes Pietzuch et al. (2019) suggested that this divergence ‘may relate to et al., 2016; Jack et al., 2015; Lyall et al., 2019b; Machulda the age of the sample groups’ (see also Damoiseaux et al., 2012 et al., 2011; Westlye et al., 2011; Taylor et al., 2015; see Table 1 for potential moderating effects of age and sex). Indeed, in a in Lyall et al., 2013 and Fouquet et al., 2014 for review). cross-sectional study, Shu et al. (2016) reported that, while both Nonetheless, despite the converging evidence against any effect ε2- and ε4-carriers had decreased DMN connectivity compared to of APOE, we tested main effects on hippocampal volume and ε3 homozygotes, they showed opposite effects of age. the interaction with age because of their historical and theoreti- Interpretation of changes in functional connectivity measured cal relevance. by functional MRI (fMRI) is non-trivial, since the measures can Other MRI studies have examined effects of APOE on white- be affected by vascular as well as neural factors (Geerligs et al., matter integrity, for example, the number of white-matter hyper- 2017; Tsvetanov et al., 2016). MEG provides a more direct meas- intensities (WMHI) in MRI images (e.g. Lyall et al., 2019b). ure of neural activity, albeit with worse spatial resolution than While CamCAN does not have MRI data suitable for estimating fMRI. To our knowledge, Koelewijn et al. (2019) is the only WMHI in the full sample, it does possess diffusion-weighted MEG study to examine resting-state functional connectivity in images (DWIs) that can be used to estimate diffusion tensor imag- APOE variants with an appreciable number of N = 159 healthy ing (DTI) metrics. Given APOE’s established role in cholesterol adults (N = 159; 51 ε4-carriers versus 108 non-carriers aged 18– transport, it is conceivable that different alleles may modulate 65 years, though mostly young; see Cuesta et al., 2015, for a age-related myelin degradation, which would affect such DTI smaller study). These authors reported that they could classify ε4 metrics. Two studies (Heise et al., 2011; Westlye et al., 2012) with status with an accuracy of 63.5%, based on the strongest 4 Brain and Neuroscience Advances connections between brain regions. Though the CamCAN data CBU in Cambridge and Max Planck Institute for Psycholinguistics were recorded with eyes closed rather than open and on a differ- in Nijmegen, respectively), and the link between them only de- ent MEG system, we replicated their analysis pipelines, as closely blinded when the Stage 1 version of this study was accepted and as possible, for the CamCAN data. restricted to APOE allelic status. In summary, we tested the antagonistic pleiotropy hypothesis, with regard to both ε4- and ε2-carriers, on a number of cognitive and brain measures that are available in the CamCAN cohort; Choice of dependent variables and models specifically, those measures that have been linked to APOE status In order to minimise the number of statistical tests, we restricted in previous studies, empirically and/or theoretically. We used a ourselves to six dependent variables – two cognitive and four quadratic expansion of age (see section ‘Methods’) to test for brain measures – which we deemed to have the strongest prior interactions between APOE status and age. In the case of signifi- evidence for associations with APOE allelic status (e.g. episodic cant non-linear effects (e.g. inverted U-shaped fits as in Figure memory) or are most theoretically relevant (e.g. hippocampal 1), we planned to estimate the age of the vertex (peak scores) for volume). These measures also have prior effect sizes published in each allele group. Since it was possible that we would identify no the literature with which to estimate power (see section ‘Power significant quadratic component (if, for example, the beneficial calculation’ below). effects of ε4 or ε2 are only seen below the age of 18 years, that is, For the cognitive data, we focused on two abilities that have during early development), such that any interaction between age shown the most consistent associations with APOE status in the and APOE status could reflect just an accelerated decline in old literature: (1) fluid intelligence and (2) episodic memory (Davies age, we also compared linear slopes. et al., 2015; O’Donoghue et al., 2018; Wisdom et al., 2011). For the former, we used the Cattell test (first principal component across four sub-tests) and for the latter we used the WAIS logical Methods memory test of verbal memory (first principal component across Data selection immediate recall, delayed recall and delayed recognition) – see Shafto et al. (2014) for more details. None of the CamCAN participants had a diagnosis of dementia For grey matter, we focused on (3) hippocampal volume, as or mild-cognitive impairment at recruitment; all reported them- extracted following application of FreeSurfer 6.0 (https://surfer. selves to be in good cognitive health, and all scored above con- nmr.mgh.harvard.edu/fswiki/DownloadAndInstall) to 1 mm-iso- ventional cut-offs for dementia on the mini-mental state tropic T1-weight magnetization prepared rapid gradient echo examination (MMSE) and Addenbrooke’s cognitive examina- (MPRAGE) scans (see https://camcan-archive.mrc-cbu.cam. tion – revised (ACE-R) screening tests (see Shafto et al., 2014). ac.uk/dataaccess/pdfs/CAMCAN700_MR_params.pdf for fur- This study was conducted in compliance with the Helsinki ther details of MRI scans). To correct for inter-individual differ- Declaration and was approved by the local ethics committee, ences in head size, hippocampal volumes were adjusted for total Cambridgeshire 2 Research Ethics Committee (reference: 10/ intracranial volume (TIV). H0308/50). The cognitive, MRI and MEG data from CamCAN For white matter, we used (4) the participant loadings of the are available on request from https://camcan-archive.mrc-cbu. first principal component of MD values across major white-matter cam.ac.uk/dataaccess/. The cognitive data were already scored tracts defined by the John Hopkins Atlas (averaged across hemi- (available on above website) and the MRI data had already been sphere, as in de Mooij et al., 2018), derived from 2-mm-isotropic preprocessed (see Taylor et al., 2017). DWI data preprocessed according to Taylor et al. (2017). DNA was prepared from saliva samples, which underwent For resting-state fMRI (rsfMRI) connectivity, we focused on genome-wide genotyping using the Illumina Infinium (5) mean connectivity within the DMN, following the optimised ‘OmniExpressExome’ SNP-chip. This chip covers >960,000 pre-processing pipeline described in Geerligs et al. (2017). SNP markers spread through the genome, capturing a large pro- Finally, for the resting-state MEG (rsMEG) connectivity, we portion of common variation. The common SNPs rs7412 and followed the procedure of Koelewijn et al. (2019) and examined rs429358 were used to determine APOE ε2, ε3 and ε4 allelic sta- (6) the participant loadings of the first principal component of tus of the participants. Full raw genotype data were first filtered connection strengths in source space. For further details of the in GenomeStudio according to standard procedures (Guo et al., preprocessing of the imaging measures and minor deviations 2014). Additional quality control checks were performed in from the Stage 1 report, see Supplemental Table 1. −6 PLINK (removal of SNPs for which Hardy Weinberg p < 1 × 10 , missingness > 0.05, minor allele frequency < 0.05; removal of individuals with total SNP missingness > 0.05 or where multidi- Statistical models mensional scaling indicated non-European origin). After quality For all six dependent variables, we modelled the effects of age by control, the data set included 675,373 directly genotyped SNPs. a second-order polynomial expansion, implemented in a general Genotype data were imputed using the Haplotype Reference linear model (GLM). A standard quadratic expansion of age is Consortium version 1.1 panel in the Michigan Imputation Server given by (https://imputationserver.sph.umich.edu). Genotypes of rs7412 were extracted from the raw genotype data, while genotypes for 2 yx =+ ββ + β x 01 2 rs429358 were derived by imputation. Prior studies have shown that this imputation procedure has high accuracy for determining where y is the dependent variable (e.g. cognitive score), x is APOE allelic status (Radmanesh et al., 2014). The phenotypic (mean-corrected) age and β are the polynomial parameters to 02 − and genotypic data were held in separate laboratories (MRC be estimated. Note that there is an equivalent parabolic form Henson et al. 5 The ‘apoe_power_simulations.m’ MATLAB script for these   yx =+ ββ () − β 01 2 analyses is provided in the website: https://osf.io/ehs9n/. Using in which 2 is the vertex (age of maximal performance as shown a Bonferroni-corrected alpha value of 0.05/6 (given the six out- in Figure 1), which can be estimated as come variables above – no correction was made for the three contrasts of APOE groups, because these contrasts share sub- ββ =− / 2β 21 2 sets of the same data, nor for the number of polynomial effects, since the main interest was in non-zero terms that were relevant All regressors of interest in the GLM were Z-scored. The GLMs to the antagonistic pleiotropy hypothesis), the cohort provides were fit in the MATLAB function ‘apoe_getdata.m’ and equiva- >80% power for effect sizes between 0.20 and 0.44 (i.e. small lently in the R script ‘apoe_lm_brms.R’, available on https://osf. to medium effects), depending on the comparison and model io/ehs9n/. Outliers on any of the six phenotypic variables were (Figure 2; note that although we use ‘power’ to refer to a single defined by residuals that were 1.5 times the interquartile range, study, strictly it refers to the expected outcome over a series of after adjusting for polynomial effects of age. studies). The following effect sizes from the literature are In the ‘categorical’ GLMs, the effects of age were modelled shown in Figure 2: separately for three sub-groups (Table 1): ε2-carriers without ε4, or henceforth the ‘ε2+ε4−’ group (i.e. ε2ε2, ε2ε3); ε4-carriers 1. The effect size of d = −0.14 of ε4 on episodic memory without ε2, or henceforth the ‘ε4+ε2−’ group (i.e. ε3ε4, ε4ε4) when averaging across all adult ages (Wisdom et al., and the ‘ε3ε3’ reference group who do not carry either ε2 or ε4 2011), for which our power is only 17% for ε4-carriers (we ignored any ε2ε4 cases, which were less than 3% of our sam- versus non-carriers or 41% if there is a parametric effect ple, as is common in the field). of ε2−/ε4+ load. Two comparisons of pairs of groups were planned: (1) ε3ε3 2. More important for the present hypothesis about age- versus ε2+ε4− and (2) ε4+ε2− versus ε3ε3. The mean and linear dependence of ε4 effects, the recent meta-analysis of age terms were tested separately as one-tailed t-contrasts, where Davies et al., 2015 found an effect size of d = −0.936 for the tail of the test depended on the direction of the expected effect the linear effect of age for the rs10119 region associated on the specific phenotypic variable (e.g. a detrimental effect of ε4 with APOE on general cognitive ability, for which the would produce lower Cattell scores of fluid intelligence, but present power is close to 100% in all cases. higher values of MD scores of white-matter integrity). For scores 3. For hippocampal volume, Taylor et al. (2015) reported like Cattell, where larger values mean better performance, one an effect size for ε4 of d = −0.56, for which this study has would expect a negative difference for both planned comparisons, a power of close to 100% (and 95% for ε2 if the effect is that is, the ε3ε3 group minus the ε2+ε4− group (since ε2 is comparable in size). hypothesised to be beneficial) and the ε4+ε2− group minus the 4. For MD from DTI, Westlye et al. (2012) reported an MD ε3ε3 group (since ε4 is hypothesised to be detrimental). This effect size for ε4 of d = +0.77, for which the present direction applies to both the mean and the linear slope of the poly- power is close to 100% in all cases. nomial effect of age (e.g. the slope of the age effect should be 5. For rsfMRI, Westlye et al. (2011) reported an effect size more negative for the ε3ε3 group than the ε2+ε4− group and for for ε4 of d = +1.23 in right hippocampus/amygdala (part the ε2+ε4− group than the ε3ε3 group). The quadratic term was of DMN), for which we have ample power (off the scale predicted, on the basis of the antagonistic pleiotropy hypothesis, in Figure 2). Note that, however, the effect sizes for this to be negative for all groups (i.e. an inverted U-shape), but can be functional connectivity effect and for the DTI effect in combined with the mean and linear terms to estimate the vertex the above Westlye et al.’s paper may be biased upwards (inflexion point) of a parabolic fit (Figure 1), as detailed above, because both were selected after thresholding voxels that which was predicted to be earlier for ε4-carriers than non-carriers. showed a basic effect of APOE status. The above categorical analyses allowed the effects of ε4 and ε2 to 6. For rsMEG, Koelewijn et al. (2019) reported an area- differ qualitatively. under-curve of the receiver-operating characteristic A second ‘parametric’, or gene-dose, GLM modelled a linear (ROC) of 63.5% for classification of ε4 status. This cor- effect of a decreasing number of ε2 alleles and increasing num- responds to a Cohen’s d of +0.49 (Salgado, 2018), for bers of ε4 alleles, which is potentially a more sensitive model if which this study has a power of close to 100%. the two alleles have the same quantitative (and additive) effects. The latter is consistent with evidence for a load effect for ε4 at least (e.g. Wisdom et al., 2011), including the cumulative Bayes factors increased risk for AD (Farrer et al., 1997). For the same tests on the six dependent variables, we also calculated the Bayes factors (BFs) for the null versus alternative hypothesis for Power calculation the APOE-by-age interaction terms. For this, we used the Savage– Dickey ratio, after logspline interpolation of the posterior sampling We simulated statistical power for various effect sizes and distribution, generated from 100,000 Markov chain Monte Carlo N = 608 participants. The number of 608 was an estimate of the (MCMC) iterations of the ‘brm’ function from the ‘brms’ R pack- number of CamCAN individuals who have valid cognitive and age (Bürkner, 2017, 2018) using the same linear model as for the genetic data; an estimate made before the genetic and pheno- above classical statistics. This is provided in the R script ‘apoe_lm_ typic data were combined (see section ‘Results’ for final num- brms.R’ on https://osf.io/ehs9n/. We imposed unit normal priors on bers for each phenotypic variable). To simulate random genetic all of the (Z-scored) polynomial terms for the age-by-APOE status sampling, multiple random draws were simulated from the interactions. Note that this is a deviation from Stage 1 of this study, expected population frequencies of each allele combination. 6 Brain and Neuroscience Advances Table 1. Number of outliers and valid data points for each phenotypic variable for each of the three genetic groups: ε2+ε4−, ε3ε3 and ε4+ε2−. Count N initial N outliers (excluded) N final Phenotypic variable ε2+ε4− ε4+ε2− Total ε2+ε4− ε4+ε2− Total ε3ε3 ε3ε3 Fluid intelligence 560 0 2 1 3 74 331 152 557 Episodic memory 592 0 1 1 2 78 355 157 590 Hippocampal volume 546 1 4 7 12 70 322 142 534 White-matter MD 525 7 26 12 45 63 286 131 480 rsfMRI in DMN 548 0 6 3 9 71 322 146 539 rsMEG PCA 514 2 7 7 16 67 295 136 498 MD: mean diffusivity; rsfMRI: resting-state functional magnetic resonance imaging; DMN: default mode network; rsMEG: resting-state magnetoencephalography; PCA: principal component analysis. which stated that we would use Jeffreys-Zellner-Siow (JZS) priors. This is because the sampling of a Cauchy prior on the mean (with scale 1/√2), as required by the JZS approach, produced unstable posterior distributions. Nonetheless, simulations show that the BFs are very similar for both priors (see https://jaquent.github.io/post/ comparing-different-methods-to-calculate-bayes-factors-for-a-sim- ple-model/). Covariates of no interest Covariates of no interest included: (1) sex (which, though bal- anced in CamCAN sample, has been shown to modulate APOE effects in some studies, e.g. Damoiseaux et al., 2012): (2) educa- tion level (since this tends to increase with year of birth in cross- sectional data and is known to correlate with cognitive measures later in life), (3) an estimate of socio-economic status (SES) and (4) a summary measure of cardiovascular health (first principal component of CamCAN’s measures of blood pressure and elec- Figure 3. Number of CamCAN participants (out of 610) with each APOE trocardiogram (ECG); see Tsvetanov et al., 2015), since this has genotype (filled bars) and number expected from White Europeans also been shown to be affected by APOE (e.g. Lyall et al., 2016; (empty bars). Oberlin et al., 2015). Separate models were fit with versus with- out covariates because education, cardiovascular health and age covary positively and therefore cannot be disentangled with Results cross-sectional data. Note that, while the Stage 1 report stated that genetic ancestry would also be a covariate, because it became There were a total of 651 genetic samples with matching pheno- clear after molecular genetic analysis that the majority of the typic IDs, of which three were excluded for low quality and sample had European white ancestry, we excluded the small three were excluded because they were related to others in the number of individuals with a different ancestry from the analyses sample. Of the remaining 645, principal component analyses of (see start of section ‘Results’), so it was no longer necessary to genome-wide SNP data showed that only 35 had non-European include ancestry as a covariate. ancestry, which we also excluded so as to optimise the homoge- SES was estimated by total family income, ranked into five neity of the sample, leaving a total sample of 610 participants. levels (<£18k, <£31k, <£52k, <£100k and >£100k per The frequency of APOE alleles closely matched that expected annum) and had 27 missing values. Education was also ranked from the frequency within healthy Europeans (Eisenberg et al., into five levels (no qualifications at age 16, practical qualifica- 2010), as seen in Figure 3. There were 78 in the ε2+ε4− group, tions at age 16, academic qualifications at age 16, qualifications 159 in the ε4+ε2− group and 357 in the ε3ε3 group (plus 16 at age 18 and university degree or higher), with 47 missing val- ε2ε4 carriers, who were not analysed further, as explained in ues. Cardiovascular health was the first principal component of section ‘Methods’). mean heart rate and heart rate variability (after low and high- The numbers of participants with data for each phenotypic pass filtering of the ECG recorded during the MEG session), as variable are shown in the leftmost column of Table 1. From these, well as systolic and diastolic blood pressure, after excluding outliers that were 1.5 times the interquartile range, after adjusting 169 missing values, respectively. All missing values were for linear and quadratic effects of age, were removed, producing replaced by the predictions of a quadratic fit of age to remaining the final numbers in the rightmost column of Table 1. The white- values, and the results Z-scored for each variable. matter MD measure contained the highest number of outliers, Henson et al. 7 Figure 4. Scatter plots for each phenotypic variable against age, grouped by APOE status (colour). Outliers are indicated by crosses (but were excluded from analyses). The solid lines show second-order polynomial fits of age on remaining points for each APOE group. most likely because diffusion-weighted MRI data are well-known ε4+ε2− versus ε3ε3; the third group of columns refers to (3) the to be sensitive to noise, such as that caused by head motion. parametric contrast of ε2−/ε4+ dose, with a linear increase across Scatter plots of the six phenotypic variables against age are ε2ε2, ε2ε3/ε3ε2, ε3ε3, ε3ε4/ε4ε3 and ε4ε4. The sign of these shown in Figure 4. Additional information regarding definition three contrasts matches the sign of the expected, prior effect size of these variables is provided in Supplemental Table 1. All six for ε4-carriers versus non-carriers reported in section ‘Power cal- variables showed significant effects of age (see Supplemental culation’ (i.e. the direction of effect was predicted to be negative Table 2). The fits come from a second-order polynomial expan- for first three phenotypic variables and positive for last three phe- sion of age to each APOE group. There was little apparent evi- notypic variables). Within each group, the parameter estimates for dence for antagonistic pleiotropy (i.e. different effects of age for the interactions between APOE status and zeroth-, first- and sec- each group), though the Cattell test of fluid intelligence showed ond-order polynomial effects of age are shown (where the zeroth- some suggestion of ε4-carriers performing worse in later life and order effect is equivalent to the main effect of APOE status). better in early life (i.e. for red line relative to blue and green lines The models fit the data well for the first four phenotypes in top left panel of Figure 4). Although the quadratic component (explaining 14%–70% of the variance). The fit was not so good was significant for three variables, it was never large relative to for fMRI and MEG (explaining only approximately 5% of the the linear component, and more importantly, it never differed sig- variance, though still significant), which we attribute to these nificantly between APOE groups (see the next section), meaning data being noisier. Only one of the three polynomial effects for that there was little value in comparing the peaks across APOE any of the three contrasts for any of the six phenotypic variables groups (i.e. age of maximum or minimum value). survived our pre-specified, one-tailed alpha value of 0.05/6. This was the zeroth-order effect on resting-state fMRI connectivity in the default mode network, where connectivity was higher on average for the ε4+ε2− group than ε3ε3 group (cf. red and blue Planned comparisons using classical lines in middle bottom panel of Figure 4). This is consistent with (frequentist) statistics Westlye et al. (2011), but because this main effect of the ε4 allele Results of the critical interactions between APOE and age are did not interact with either of the linear or quadratic age terms, it shown in Table 2. These come from the same linear model as provides no support for the antagonistic pleiotropy hypothesis. shown in Figure 4 (without any covariates). In the three groups of The only phenotypic variable to show any suggestion of the columns, the first two groups refer to the two planned categorical predicted interaction between APOE status and age was the contrasts across APOE groups of (1) ε3ε3 versus ε2+ε4−, (2) Cattell measure of fluid intelligence (surviving p < 0.05 but not 8 Brain and Neuroscience Advances Table 2. GLM results for each phenotypic variable (row). Contrast ε3ε3 group versus ε2+ε4− group ε4+ε2− group versus ε3ε3 group Parametric (dose) effect of ε2−ε4+ Poly. order Fit 0 1 2 Fit 0 1 2 Fit 0 1 2 2 2 2 Fluid intelligence R = 0.455 0.017 0.077 0.072 R = 0.500 −0.133 −0.228 0.151 R = 0.490 −0.192 −0.235 0.153 df = 399 (0.122) (0.158) (0.158) df = 477 (0.107) (0.115) (0.115) df = 551 (0.102) (0.102) (0.102) T = 0.139 T = 0.489 T = 0.458 T = 1.246 T = 1.982 T = 1.310 T = 1.881 T = 2.313 T = 1.507 2 2 2 Episodic memory R = 0.136 0.025 0.134 0.072 R = 0.145 −0.102 −0.251 0.023 R = 0.145 −0.138 −0.223 0.026 df = 427 (0.244) (0.319) (0.319) df = 506 (0.230) (0.249) (0.249) df = 584 (0.213) (0.213) (0.213) T = 0.102 T = 0.423 T = 0.225 T = 0.443 T = 1.006 T = 0.926 T = 0.648 T = 1.046 T = 0.121 2 2 2 Hippocampal volume R = 0.406 −22.12 −2.378 −19.86 R = 0.414 27.78 −3.697 2.579 R = 0.411 −7.161 −11.89 −9.45 df = 386 (14.81) (19.40) (19.40) df = 458 (13.68) (14.85) (14.85) df = 528 (12.69) (12.70) (12.71) T = 1.494 T = 0.123 T = 1.023 T = 2.031 T = 0.249 T = 0.174 T = 0.564 T = 0.936 T = 0.743 2 2 2 MD of WM tracts R = 0.710 8.54e-6 4.34e-6 7.42e-6 R = 0.701 −9.97e-6 −5.25e-6 −4.37e-6 R = 0.698 3.90e-6 4.74e-6 1.43e-6 df = 343 (5.56e-6) (7.26e-6) (7.26e-6) df = 411 (5.16e-6) (5.56e-6) (5.56e-6) df = 474 (4.82e-6) (4.84e-6) (4.81e-6) T = 1.536 T = 0.597 T = 1.023 T = 1.93 T = 0.945 T = 0.786 T = 0.809 T = 0.981 T = 0.298 2 2 2 DMN rsfMRI con- R = 0.042 1.28e-4 −3.82e-3 −2.35e-4 R = 0.060 7.66e-3 −1.38e-4 2.83e-3 R = 0.046 4.58e-3 −1.72e-3 −2.78e-4 nectivity df = 387 (2.84e-3) (3.70e-3) (3.70e-3) df = 462 (2.76e-3) (2.98e-3) (2.98e-3) df = 533 (2.56e-3) (2.56e-3) (2.56e-3) T = 0.045 T = 1.032 T = 0.634 T = 2.774 T = 0.046 T = 0.951 T = 1.77 T = 0.671 T = 0.914 2 2 2 rsMEG connectivity R = 0.051 0.059 0.047 −0.083 R = 0.042 −0.046 −0.033 0.005 R = 0.034 0.006 −0.007 −0.031 df = 356 (0.044) (0.056) (0.056) df = 425 (0.042) (0.044) (0.044) df = 492 (0.039) (0.039) (0.039) T = 1.362 T = 0.841 T = 1.486 T = 1.098 T = 0.721 T = 0.116 T = 0.150 T = 0.169 T = 0.793 MD: mean diffusivity; WM: white matter; DMN: default mode network; rsfMRI: resting state functional magnetic resonance imaging; rsMEG: resting state magnetoencephalography; Poly.: polynomial; df: degrees of freedom; R : adjusted R-squared of full model. The three groups of columns refer to the planned contrasts across APOE groups. Within each group, the first column gives the overall model fit, and the next three columns give the parameter estimates, their associated standard error (in brackets) and unsigned T-statistic for the interaction between the APOE contrast and the three polynomial expansions of age: zeroth (constant), first (linear) and second (quadratic), where zeroth-order term is equivalent to main effect of APOE contrast (see Supplemental Table 2 for parameters for main effects of Age). Effects with p < 0.05 are shown in bold, but note that only one survived the pre-specified Bonferroni correction for six multiple, one- tailed comparisons (for which|T| > 2.40 for the minimum number of df’s here; see text), where direction of effect was predicted to be negative for first three phenotypic variables and positive for last three phenotypic variables (see text). These results are without covariates; see Supplemental Table 3 for corresponding results with covariates. Henson et al. 9 Table 3. Bayes factors (BFs) for various hypotheses about the first-order (linear) effect of age being zero (BF01) for each contrast and each phenotypic variable, given various means for a unit normal prior. Contrast ε3ε3 group versus ε2+ε4− group ε4+ε2− group versus ε3ε3 group Parametric (dose) effect of ε2−/ε4+ BF01 Lin = 0 Lin < 0 Lin = P Lin = 0 Lin < 0 Lin = P Lin = 0 Lin < 0 Lin = P Fluid intelligence 18.63 29.79 29.42 4.08 2.09 5.99 2.30 1.16 3.31 Episodic memory 15.69 23.30 15.82 13.56 8.05 13.72 13.56 8.05 13.72 Hippocampal volume 19.40 17.71 22.65 25.11 20.95 29.20 19.43 11.79 22.15 BF01 Lin = 0 Lin > 0 Lin = P Lin = 0 Lin > 0 Lin = P Lin = 0 Lin > 0 Lin = P MD of WM tracts 22.21 15.35 29.31 22.17 63.92 30.52 24.36 14.58 32.18 DMN rsfMRI connectivity 9.10 29.84 21.08 20.66 21.45 43.90 18.91 27.59 41.65 rsMEG connectivity 10.59 6.62 11.65 15.13 32.11 17.38 22.26 25.64 25.28 BF: Bayes factor; MD: mean diffusivity; WM: white matter; DMN: default mode network; rsfMRI: resting state functional magnetic resonance imaging; rsMEG: resting state magnetoencephalography. The column headings for each analysis are explained in the text. The value of P in the column ‘Lin = P’ was determined from the literature cited in section ‘Power calculation’, that is, P = [−0.936, −0.14, −0.56, +0.77, +1.23, +0.49] for the six phenotypic variables. The direction of the one-tailed test in the column (i.e. ‘Lin > 0’ or ‘Lin < 0’) was determined by the sign of P (i.e. negative for the first three phenotypic variables and positive for the remaining three). correction for multiple comparisons), for which the slope was for the linear interaction between age and APOE status being more negative for the ε4+ε2− group than ε3ε3 group (cf. red and greater or less than zero ( ‘Lin > 0’ or ‘Lin < 0’), given a prior of blue lines in top left panel of Figure 4), and was negatively greater or less than zero, where the greater/lesser direction depends related to increasing ε2−/ε4+ dose in the parametric model. on the analysis (i.e. ε2 being protective would predict a more posi- The only other effect with p < 0.05 (uncorrected) was the tive (less negative) slope than the reference group, and ε4 being a zeroth-order effect for the ε4+ε2− group on hippocampal vol- risk factor would predict a more negative slope than the reference ume, which tended to be larger on average than in the ε3ε3 group. group) and (3) BFs for the linear interaction between age and However, we suspect that this is a false positive since it was in APOE status being zero given the prior expectation of an effect the opposite direction to our predictions, and so we do not con- size (P) equal to that from the literature that was used to power this sider it further. study ( ‘Lin = P’), as listed in section ‘Power calculation’. In all cases, the BFs provided ‘substantial’ (BF > 3) or ‘strong’ (BF > 10) evidence (https://en.wikipedia.org/wiki/Bayes_factor) Adjusting for covariates for no interaction between APOE status and age on any of the phenotypic variables, regardless of whether their prior expecta- The above linear model was repeated with five additional covari- tion was equal to zero or equal to the effect size from the litera- ates: male/female sex, education level, SES and cardiovascular ture, with the exception of fluid intelligence: While there was health. The inclusion of these covariates did not reveal any new strong evidence for no interaction between the ε2 allele and age significant effects of APOE (Supplemental Table 3). The greater on fluid intelligence, this was not true for the predicted direction DMN fMRI connectivity for the ε4+ε2− group relative to the of interaction between ε4 allele and age (or of overall ε4 load), ε3ε3 group continued to survive correction, and the parametric where the BFs were around 2, that is, ambiguous. effect of dose on the linear age slope for fluid intelligence contin- ued to survive p < 0.05 uncorrected. The linear interaction between age and ε4+ε2− group and ε3ε3 group on fluid intelli- General discussion gence no longer survived p < 0.05 after adjusting for covariates. In short, there was no evidence for the antagonistic pleiotropy This study provided no support for, and mostly evidence against, hypothesis using classical null-hypothesis testing. Since this the ‘antagonistic pleiotropy’ hypothesis (Han and Bondi, 2008), could reflect false negatives, given our relatively small sample whereby the ε4 variant of APOE is proposed to be advantageous for genetic effects, we calculated BFs for the null versus alternate in early life but disadvantageous in later life. The study also pro- hypotheses. vided evidence against the related hypothesis that the ε2 variant of APOE is advantageous, particularly in later life (Suri et al., 2013). While our sample was relatively small for genetic analysis BFs of cognitive and neural phenotypes, it was sufficiently powered Since the dominant effect of age on phenotypic variables was for classical statistics when based on previous effect sizes linear (see Supplemental Table 1), and linear effects were reported, and moreover, furnished BFs that provided evidence in reported in the prior literature (see section ‘Power calculation’), favour of the null hypothesis that these APOE variants do not we only report here the BFs for the first-order polynomial term. interact (linearly) with age in their putative effects on brain struc- The three columns in Table 3 show, for each APOE analysis: ture or cognition. (1) BFs for the linear interaction between Age and APOE status The only effect of APOE status that survived our a priori cor- being zero, given a prior expectation of zero ( ‘Lin = 0’), (2) BFs rection for the six phenotypic variables tested was the main effect 10 Brain and Neuroscience Advances of ε4-carriers (more precisely, our ε4+ε2− group versus our ε3ε3 is a cross-sectional sample, which may be confounded with effects reference group) on the mean fMRI functional connectivity of birth-year and is unable to examine true ageing within individu- within the default mode network. This reflected higher connec- als. Fourth, we cannot discount a role for antagonistic pleiotropy for tivity for the ε4-carriers, consistent with Westlye et al. (2012). phenotypic measures that we did not examine in this study such as More importantly, there was no evidence that this effect inter- other measures of cognition or the brain. acted with age, and therefore even if it reflects a true genetic effect on brain functional connectivity, it is not evidence for the Acknowledgements antagonistic pleiotropy hypothesis. The authors thank Kamen Tsvetanov for providing the cardiovascular While all phenotypic variables showed strong associations summary measures and Alex Quent for help with the Bayes Factor with age, there was only one suggestion that such age associa- estimation. tions depended on APOE status, namely, a steeper, linear age- related decline in fluid intelligence in the ε4+ε2− group than Declaration of conflicting interests ε3ε3 group. This effect did not survive correction for multiple The authors declare that there is no conflict of interest. comparisons. Moreover, the BF for this linear effect was ambigu- ous, so this effect needs replication before providing support for the antagonistic pleiotropy hypothesis. Funding None of the analyses showed any evidence of differences for The author(s) disclosed receipt of the following financial support for the ε2-carriers (i.e. when comparing our ε4+ε2− group versus our research, authorship and/or publication of this article: The Cambridge ε3ε3 reference group). While the number of ε2-carriers was Centre for Ageing and Neuroscience (CamCAN) research was supported lower than ε4-carriers, the lack of interaction between ε2 and age by the Biotechnology and Biological Sciences Research Council (Grant is unlikely to simply reflect low power, because BFs provided No. BB/H008217/1). This project has also received funding from the European Union’s Horizon 2020 research and innovation programme evidence (BFs > 9) for the interaction being zero. (‘LifeBrain’, Grant Agreement No. 732592). R.N.H. was supported by The classical power estimates may have been over-estimated the Medical Research Council (SUAG/010 RG91365). S.S. was sup- because the effect sizes reported in the literature are likely to be ported by an Alzheimer’s Society Junior Research Fellowship (Grant inflated owing to publication bias and/or ‘winner’s curse’. Reference No. 441). E.K. was supported by the LifeBrain grant. R.A.K. Indeed, the BFs for the null hypothesis were highest when the was supported by the UK Medical Research Council (SUAG/014 prior mean was based on a published effect size. Nonetheless, the RG91365). J.B.R. was supported by the Wellcome Trust (103838) and the BF still provided substantial to strong evidence for no effect even Medical Research Council (SUAG/051 RG91365). D.C. was funded by when the prior mean was zero. the Wellcome Trust and Alzheimer’s Research UK. E.E. and S.E.F. were While this study had lower power than many previous stud- supported by the Max Planck Society. ies of APOE on cognition, it is nonetheless larger than many prior APOE studies using neuroimaging measures, particularly ORCID iD measures of functional connectivity using fMRI or MEG, which Richard N. Henson https://orcid.org/0000-0002-0712-2639 often use small and/or biased samples of the population. Relative to these studies, this study gains power by virtue of the wide and Supplemental material near-uniform age range from 18 to 88 years, in a population- derived sample. Nevertheless, it is important to note that, by Supplemental material for this article is available online. only considering adults, we cannot address a version of the antagonistic pleiotropy hypothesis in which the interactions References between APOE status and age only occur during cognitive and Belloy ME, Napolioni V and Greicius MD (2019) A quarter century of brain development, that is, in individuals under 18 years. This APOE and Alzheimer’s disease: Progress to date and the path for- issue could be examined in larger cognitive and neuroimaging ward. Neuron 101(5): 820–838. lifespan cohorts like the European LifeBrain consortium Blacker D, Haines JL, Rodes L, et al. (1997) ApoE-4 and age at onset (Walhovd et al., 2018). of Alzheimer’s disease: The NIMH genetics initiative. Neurology Finally, there are other caveats associated with our study. First, 48(1): 139–147. there may be other types of bias in the selection of adults used in Buckner RL, Andrews-Hanna JR and Schacter DL (2008) The brain’s this study, particularly those towards the older end of our age range, default network: Anatomy, function, and relevance to disease. for example, owing to survival bias and exclusion of those with Annals of the New York Academy of Sciences 1124(1): 1–38. evidence of cognitive decline. Indeed, it is possible that the effects Bunce D, Anstey KJ, Cherbuin N, et al. (2012) APOE genotype and ento- of APOE variants are only detectable in individuals with AD pathol- rhinal cortex volume in non-demented community-dwelling adults in midlife and early old age. Journal of Alzheimer’s Disease 30(4): ogy (e.g. Vemuri et al., 2010) and there were few such people in our 935–942. cohort, even those in the pre-symptomatic phase of AD. There will Bürkner PC (2017) brms: An R package for Bayesian multilevel models also be a bias induced by the exclusion of those who were unable to using Stan. Journal of Statistical Software. Epub ahead of print 29 undergo an MRI scan, for example, because of cardiovascular prob- August. DOI: 10.18637/jss.v080.i01. lems that require a pacemaker or stent, but which might also be Bürkner PC (2018) Advanced Bayesian multilevel modeling with the R related to APOE allelic status and cognitive/brain health. Second, package brms. The R Journal 80(1): 1–28. our sample only included White Europeans (Caucasians), whereas Cuesta P, Garcés P, Castellanos NP, et al. (2015) Influence of the APOE APOE variants may have stronger effects in other genetic groups, ε4 allele and mild cognitive impairment diagnosis in the disruption and we did not have any direct measures of neuropathology or take of the MEG resting state functional connectivity in sources space. into account potential effects of a range of medications. Third, this Journal of Alzheimer’s Disease 44(2): 493–505. Henson et al. 11 Damoiseaux JS, Seeley WW, Zhou J, et al. (2012) Gender modulates the Koelewijn L, Lancaster TM, Linden D, et al. (2019) Oscillatory hyperac- APOE ε4 effect in healthy older adults: Convergent evidence from tivity and hyperconnectivity in young APOE-ɛ4 carriers and hypo- functional brain connectivity and spinal fluid tau levels. The Journal connectivity in Alzheimer’s disease. Elife 8: e36011. of Neuroscience 32(24): 8254–8262. Lancaster C, Tabet N and Rusted J (2017) The elusive nature of Davies G, Armstrong N, Bis JC, et al. (2015) Genetic contributions to APOE ε4 in mid-adulthood: Understanding the cognitive profile. variation in general cognitive function: A meta-analysis of genome- Journal of the International Neuropsychological Society 23(3): wide association studies in the CHARGE consortium (N = 53 949). 239–253. Molecular Psychiatry 20(2): 183–192. Lyall DM, Celis-Morales C, Lyall LM, et al. (2019a) Assessing for inter- Deary IJ, Whiteman MC, Pattie A, et al. (2002) Ageing: Cognitive action between APOE ε4, sex, and lifestyle on cognitive abilities. change and the APOE ε4 allele. Nature 418(6901): 932. Neurology 92(23): e2691–e2698. de Mooij SMM, Henson RNA, Waldorp LJ, et al. (2018) Age differen- Lyall DM, Cox SR, Lyall LM, et al. (2019b) Association between APOE tiation within gray matter, white matter, and between memory and e4 and white matter hyperintensity volume, but not total brain vol- white matter in an adult life span cohort. The Journal of Neurosci- ume or white matter integrity. Brain Imaging and Behavior. Epub ence 38(25): 5826–5836. ahead of print 22 March. DOI: 10.1007/s11682-019-00069-9. Eisenberg DTA, Kuzawa CW and Hayes MG (2010) Worldwide allele Lyall DM, Harris SE, Bastin ME, et al. (2014) Alzheimer’s disease sus- frequencies of the human apolipoprotein E gene: Climate, local ceptibility genes APOE and TOMM40, and brain white matter integ- adaptations, and evolutionary history. American Journal of Physical rity in the Lothian Birth Cohort 1936. Neurobiology of Aging 35(6): Anthropology 143(1): 100–111. 1513e25–1513e33. Farrer LA, Cupples LA, Haines JL, et al. (1997) Effects of age, sex, and Lyall DM, Royle NA, Harris SE, et al. (2013) Alzheimer’s disease sus- ethnicity on the association between apolipoprotein E genotype and ceptibility genes APOE and TOMM40, and hippocampal volumes in Alzheimer’s disease. JAMA 278(16): 1349–1356. the Lothian Birth Cohort 1936. PLoS One 8(11): e80513. Filippini N, MacIntosh BJ, Hough MG, et al. (2009a) Distinct patterns Lyall DM, Ward J, Ritchie SJ, et al. (2016) Alzheimer’s disease genetic of brain activity in young carriers of the APOE-epsilon4 allele. Pnas risk factor APOE e4 and cognitive abilities in 111,739 UK Biobank 106(17): 7209–7214. participants. Age and Ageing 45(4): 511–517. Filippini N, Rao A, Wetten S, et al. (2009b) Anatomically-distinct genetic Machulda MM, Jones DT, Vemuri P, et al. (2011) Effect of APOE ε4 associations of APOE e{open}4 allele load with regional cortical status on intrinsic network connectivity in cognitively normal elderly atrophy in Alzheimer’s disease. Neuroimage 44(3): 724–728. subjects. Archives of Neurology 68(9): 1131–1136. Fleisher AS, Sherzai A, Taylor C, et al. (2009) Resting-state BOLD Marioni RE, Campbell A, Scotland G, et al. (2016) Differential effects networks versus task-associated functional MRI for distinguishing of the APOE e4 allele on different domains of cognitive ability Alzheimer’s disease risk groups. Neuroimage 47(4): 1678–1690. across the life-course. European Journal of Human Genetics 24(6): Fouquet M, Besson FL, Gonneaud J, et al. (2014) Imaging brain effects 919–923. of APOE4 in cognitively normal individuals across the lifespan. Mondadori CRA, De Quervain DJF, Buchmann A, et al. (2007) Better Neuropsychology Review 24(3): 290–299. memory and neural efficiency in young apolipoprotein E ε4 carriers. Fullerton SM, Clark AG, Weiss KM, et al. (2000) Apolipoprotein E vari- Cerebral Cortex 17(8): 1934–1947. ation at the sequence haplotype level: Implications for the origin and Oberlin LE, Manuck SB, Gianaros PJ, et al. (2015) Blood pressure inter- maintenance of a major human polymorphism. American Journal of acts with APOE ε4 to predict memory performance in a midlife sam- Human Genetics 67(4): 881–900. ple. Neuropsychology 29(5): 693–702. Geerligs L, Tsvetanov KA, Cam-CAN, et al. (2017) Challenges in O’Donoghue MC, Murphy SE, Zamboni G, et al. (2018) APOE genotype measuring individual differences in functional connectivity using and cognition in healthy individuals at risk of Alzheimer’s disease: fMRI: The case of healthy aging. Human Brain Mapping 38(8): A review. Cortex 104(July): 103–123. 4125–4156. Pietzuch M, King AE, Ward DD, et al. (2019) The influence of genetic Guo Y, He J, Zhao S, et al. (2014) Illumina human exome genotyp- factors and cognitive reserve on structural and functional resting- ing array clustering and quality control. Nature Protocols 9(11): state brain networks in aging and Alzheimer’s disease. Frontiers in 2643–2662. Aging Neuroscience 11(March): 30. Habes M, Toledo JB, Resnick SM, et al. (2016) Relationship between Puttonen S, Elovainio M, Kivimäki M, et al. (2003) The combined effects APOE genotype and structural MRI measures throughout adulthood of apolipoprotein E polymorphism and low-density lipoprotein choles- in the study of health in Pomerania population-based cohort. Ameri- terol on cognitive performance in young adults. Neuropsychobiology can Journal of Neuroradiology 37(9): 1636–1642. 48(1): 35–40. Han SD and Bondi MW (2008) Revision of the apolipoprotein E com- Radmanesh F, Devan WJ, Anderson CD, et al. (2014) Accuracy of impu- pensatory mechanism recruitment hypothesis. Alzheimer’s and tation to infer unobserved APOE epsilon alleles in genome-wide Dementia 4(4): 251–254. genotyping data. European Journal of Human Genetics 22(10): Heise V, Filippini N, Ebmeier KP, et al. (2011) The APOE ɛ4 allele 1239–1242. modulates brain white matter integrity in healthy adults. Molecular Rocchi A, Pellegrini S, Siciliano G, et al. (2003) Causative and suscepti- Psychiatry 16(9): 908–916. bility genes for Alzheimer’s disease: A review. Brain Research Bul- Ihle A, Bunce D and Kliegel M (2012) APOE ε4 and cognitive function letin 61(1): 1–24. in early life: A meta-analysis. Neuropsychology 26(3): 267–277. Rusted JM, Evans SL, King SL, et al. (2013) APOE e4 poly- morphism Jack CR, Wiste HJ, Weigand SD, et al. (2015) Age, sex, and APOE ε4 in young adults is associated with improved attention and indexed by effects on memory, brain structure, and β-amyloid across the adult distinct neural signatures. NeuroImage 65: 364–373. life Span. JAMA Neurology 72(5): 511–519. Salgado JF (2018) Transforming the area under the normal curve (AUC) Jochemsen HM, Muller M, Van der Graaf Y, et al. (2012) APOE ε4 dif- into Cohen’s d, Pearson’s r , odds-ratio, and natural log odds-ratio: pb ferentially influences change in memory performance depending on Two conversion tables. European Journal of Psychology Applied to age. The SMART-MR study. Neurobiology of Aging 33(4): 832. Legal Context 10(1): 35–47. e15–832.e22. Schiepers OJG, Harris SE, Gow AJ, et al. (2012) APOE E4 status pre- Jorm AF, Mather KA, Butterworth P, et al. (2007) APOE genotype and dicts age-related cognitive decline in the ninth decade: Longitudinal cognitive functioning in a large age-stratified population sample. follow-up of the Lothian Birth Cohort 1921. Molecular Psychiatry Neuropsychology 21(1): 1–8. 17(3): 315–324. 12 Brain and Neuroscience Advances Schultz MR, Lyons MJ, Franz CE, et al. (2008) Apolipoprotein E geno- and functional MRI, MEG, and cognitive data from a cross-sectional type and memory in the sixth decade of life. Neurology 70(19 PART adult lifespan sample. Neuroimage 144(Pt B): 262–269. 2): 1771–1777. Trachtenberg AJ, Filippini N, Ebmeier KP, et al. (2012) The effects of Shafto M, a Tyler LK, Dixon M, et al. (2014) The Cambridge Centre APOE on the functional architecture of the resting brain. Neuroim- for Ageing and Neuroscience (Cam-CAN) study protocol: A cross- age 59(1): 565–572. sectional, lifespan, multidisciplinary examination of healthy cogni- Tsvetanov KA, Henson RNA, Tyler LK, et al. (2015) The effect of age- tive ageing. BMC Neurology 14(1): 204. ing on fMRI: Correction for the confounding effects of vascular reac- Sheline YI, Morris JC, Snyder AZ, et al. (2010) APOE4 allele disrupts resting tivity evaluated by joint fMRI and MEG in 335 adults. Human Brain state fMRI connectivity in the absence of amyloid plaques or decreased Mapping 36(6): 2248–2269. CSF Aβ42. The Journal of Neuroscience 30(50): 17035–17040. Tsvetanov KA, Henson RNA, Tyler LK, et al. (2016) Extrinsic and Shin MH, Kweon SS, Choi JS, et al. (2014) The effect of an APOE poly- intrinsic brain network connectivity maintains cognition across the morphism on cognitive function depends on age. Journal of Neurol- lifespan despite accelerated decay of regional brain activation. Jour- ogy 261(1): 66–72. nal of Neuroscience 36(11): 3115–3126. Shu H, Shi Y, Chen G, et al. (2016) Opposite neural trajectories of apo- Vemuri P, Wiste HJ, Weigand SD, et al. (2010) Effect of apolipopro- lipoprotein E ε4 and ε2 alleles with aging associated with different tein E on biomarkers of amyloid load and neuronal pathology in risks of Alzheimer’s disease. Cerebral Cortex 26(4): 1421–1429. Alzheimer’s disease 67(3): 308–316. Siebner HR, Callicott JH, Sommer T, et al. (2009) From the genome to Walhovd KB, Fjell AM, Westerhausen R, et al. (2018) Healthy minds the phenome and back: Linking genes with human brain function and 0–100 years: Optimising the use of European brain imaging cohorts structure using genetically informed neuroimaging. Neuroscience (‘Lifebrain’). European Psychiatry 47(Jan): 76–87. 164(1): 1–6. Westlye ET, Lundervold A, Rootwelt H, et al. (2011) Increased hip- Stebbins GT and Murphy CM (2009) Diffusion tensor imaging in pocampal default mode synchronization during rest in middle-aged Alzheimer’s disease and mild cognitive impairment. Behavioural and elderly APOE ε4 carriers: Relationships with memory perfor- Neurology 21(1–2): 39–49. mance. The Journal of Neuroscience 31(21): 7775–7783. Suri S, Heise V, Trachtenberg AJ, et al. (2013) The forgotten APOE Westlye LT, Reinvang I, Rootwelt H, et al. (2012) Effects of APOE allele: A review of the evidence and suggested mechanisms for on brain white matter microstructure in healthy adults. Neurology the protective effect of APOE e2. Neuroscience and Biobehavioral 79(19): 1961–1969. Reviews 37(10): 2878–2886. Wisdom NM, Callahan JL and Hawkins KA (2011) The effects of apoli- Taylor JL, Scanlon BK, Farrell M, et al. (2015) APOE-epsilon4 and poprotein E on non-impaired cognitive functioning: A meta-analysis. aging of medial temporal lobe gray matter in healthy adults older Neurobiology of Aging 32(1): 63–74. than 50 years. Neurobiology of Aging 35(11): 2479–2485. Yu YW, Lin CH, Chen SP, et al. (2000) Intelligence and event-related Taylor JR, Williams N, Cusack R, et al. (2017) The Cambridge Centre for potentials for young female human volunteer apolipoprotein E epsi- Ageing and Neuroscience (Cam-CAN) data repository: Structural lon4 and non-epsilon4 carriers. Neuroscience Letters 294(3): 179–181.

Journal

Brain and Neuroscience AdvancesSAGE

Published: Oct 7, 2020

Keywords: Cognition; apolipoprotein E; lifespan; brain; ageing

References