Genetic Risk of Alzheimer's Disease - Sleepless with the Enemy.

The clinical and biological interface between sleep and neurodegeneration is intriguing. A range of sleep disorders occur during in the clinical course of Alzheimer's disease (AD), Parkinson's disease, and other neurodegenerative diseases. For AD, observational studies suggest that sleep disturbance and fragmentation may be risk factors for future disease development, , and a recent meta‐analysis reviewing ~400 observational prospective studies and randomized controlled trials suggested that there was good evidence that sufficient and good‐quality sleep was associated with a reduction in the risk of developing AD. In patients with manifest AD, cell loss is observed in the suprachiasmatic nucleus (ie, the pacemaker for circadian rhythm) and the same patients experience phase delay in their circadian cycle. Further support for a link between sleep deprivation and AD was found in an imaging study using Pittsburgh compound B (PiB) positron emission tomography (PET), which showed an association between self‐reported shorter sleep duration and beta‐amyloid burden in the cortex and precuneus. In keeping with observations made in humans, mouse models of AD revealed that chronic sleep restriction enhances amyloid plaque formation and AD mice display rapid eye movement (REM) sleep deficits. , Whether part of this association arises as a consequence of impairment of the “glymphatic” system, which may result from disrupted sleep, is an area of ongoing research.

An outline of the steps when performing a Mendelian Randomization (MR) Study. (A) SNPs for an exposure of interest are selected from a GWAS of that exposure trait. These independent SNPs can then be used in combination to create an “instrument,” which can be used for causal inference about how the exposure affects an outcome of interest. If individual‐level data is available for genotype, exposure and outcome in the same cohort (such as UK Biobank) then the simplest way to create an instrument is with a genetic risk score (GRS) and this is the approach taken in one‐sample MR. This is often not the case, however, so a more popular method is to use the top SNPs from the summary data of a GWAS of the exposure and look up same SNPs in a GWAS of the outcome, which involves different participants (two‐sample MR). The relationship of SNPs with the exposure, and separately their relationship with the outcome, can be used to calculate the effect of the exposure on the outcome, averaged across all of the SNPs in the instrument. (B) A directed acyclic graph depicts the concept of MR as a form of instrumental variable analysis (dark arrows), along with the core assumptions. SNPs that are strongly associated with the exposure of interest (assumption 1) can be used for causal inference about how the exposure affects the outcome, hypothetically free from confounding and reverse causation. In addition to assumption 1, the instruments should not violate assumptions 2 and 3 (light arrows), that is they should not be associated with confounding factors (this is hard to prove in many settings) and there is no alternative way that the SNP influences the outcome other than via the exposure (ie, there should be no horizontal pleiotropic effects such that SNPs also influence a different exposure trait which then affects the same outcome). (C) Where the assumptions are upheld, the change in outcome for a given change in the exposure can be calculated using the GRS in a one‐sample setting or meta‐analysis of the effects from independent SNPs in a two‐sample setting. An example of the latter is shown in the figure, whereby the SNP‐exposure and SNP‐outcome effects are plotted, and the slope of the black line, which is weighted by inverse variance, is the average causal effect of the exposure on the outcome. In truth, violations of assumption 3 (that there are no pleiotropic effects of exposure SNPs) are common and frequently bias the pooled effect estimate (red line). A whole range of two‐sample MR methods have been created to tackle this situation, including MR‐Egger, median and modal estimates, MR‐PRESSO, and GSMR. AD = Alzheimer's disease; GRS = genetic risk score; GSMR = generalized summary Mendelian randomization; GWAS = genomewide association study; MR = Mendelian randomization; SNP = single nucleotide polymorphism. * The exposure Manhattan plot depicted comes from Ref. 9. [Color figure can be viewed at www.annalsofneurology.org]

The cross‐sectional nature of some of the clinical studies and ‐ with respect to those with prospective designs ‐ a long prodromal phase of AD, make it challenging to know whether changes in sleep patterns or the emergence of sleep disorders are cause or effect of early AD. Even with long follow‐up periods in prospective, observational settings, the possibility of reverse causation remains, and the risk of residual confounding is ever present. Reverse causation arises when the outcome of interest affects the amount of exposure to the risk factor (eg, if one is studying the effect of risk of AD on sleep duration, then an apparent association may also arise if sleep duration has an effect on risk of AD). Potential confounding factors for an association between AD and sleep include changes in physical activity, obesity, sleep apnea, medication use, anxiety, depression, or chronic stress.

Given the limitations that observational studies have when it comes to causal inference, some genetic epidemiological methods have emerged to try to shed insight on association versus causation. Mendelian randomization (MR) is a method that uses genetic variants as so‐called “instrumental variables,” which are present from conception and explain some of the liability towards a given trait (see Figure). Typically, multiple genetic variants in the form of single nucleotide polymorphisms (SNPs) from genomewide association analyses (GWAS) of a trait (such as AD) are used for causal inference about how that trait influences a particular outcome (such as sleep duration). SNPs can be used individually, or their effects pooled using meta‐analysis, or alternatively combined to create a genetic risk score (GRS) for a given exposure trait. In order to be valid “instruments,” the SNPs should conform to several underlying assumptions: (1) strong association with the exposure; (2) lack of association with potential confounders for the exposure and outcome under study; and (3) no alternative pathway by which the SNP affects the outcome other than via the exposure. The latter assumption is that there are no pleiotropic effects of the SNPs that might bias a causal estimate; where pleiotropy is defined as the situation in which a given SNP influences multiple traits/exposures and these in turn (or their downstream effects) influence the outcome. The use of MR as a method for causal inference has grown rapidly in parallel with the emergence of masses of publicly available genetic data and easy access to huge biorepositories, such as the UK Biobank (UKB).

In this edition of the Annals of Neurology, Leng and colleagues used a GRS for AD (AD‐GRS) in participants from the UKB to explore the causal effect of the genetic liability towards AD on sleep duration within an MR framework. First, they showed that the AD‐GRS was associated with sleep duration and also that the apolipoprotein E (ApoE) E4 allele was associated with sleep duration in a dose‐dependent manner. ApoE variants are an important source of common genetic variation pertaining to AD risk and so separate consideration is warranted to ensure that it is not the sole driver of any observed association. The authors went on to show that an increase in the genetic liability toward AD (expressed as a log odds ratio) had the effect of reducing sleep duration by ~2 hours in UKB participants between the age of 55 and 75 years. The resulting inference from this is that changes in sleep duration may occur as an early marker of AD and that mid‐life changes in sleep duration might help identify a higher risk group.

A Lancet Commission report recently updated a list of risk factors for dementia. Sleep duration was not included in this, and on current evidence, it may be more likely to be a prodromal marker of dementia, rather than a causal risk factor. Bidirectional relationships may exist between AD and sleep (as indicated by Leng and colleagues ), and, as mentioned above, there are other studies that suggest associations in the direction of sleep duration affecting AD risk. For several of the other dementia risk factors as well, the direction of effect with dementia (cause or effect) remains uncertain, including hearing loss, obesity, physical activity, and depression. Although MR may offer insight into the direction of a causal effect, there are numerous limitations and the results of MR studies should always be placed in context with the wider experimental literature. One consideration is the type of MR study: a one‐sample MR study involves genotyping (to create the AD‐GRS) and measuring the exposure (here, the manifestation of AD), and the outcome (ie, sleep duration) from the same sample of people. A two‐sample MR study uses summary statistics from two separate GWAS studies comprising the exposure GWAS (ie, AD) from which genomewide significant SNPs are selected as instruments, and the outcome GWAS (ie, sleep duration) where these genetic instruments are applied to assess the causal effect of liability to AD on sleep duration. One‐sample MR, which was the approach used by Leng and colleagues, can give rise to effect estimates consistent with a confounded observational study if the genetic instrument lacks sufficient strength (see assumption 1). As such, replication in an independent setting and use of two‐sample MR (which tends to be more conservative), would add further weight to the finding. As mentioned, there are many potential confounding factors in the relationship between dementia and sleep duration, including but not limited to depression, anxiety, medication use, obesity, sleep apnea, stress, and physical activity. Whereas MR can help mitigate some of the effects of confounding, it will not completely remove those effects. Overall, the results are a positive step toward shedding further light on the emerging links between sleep and neurodegeneration, and future studies should make use of the ever‐expanding complement of methods to explore causality using available genetic and phenotypic data.

Potential Conflicts of Interest

The authors declared no conflict of interest.