Investigating the relationships between unfavourable habitual sleep and metabolomic traits: evidence from multi-cohort multivariable regression and Mendelian randomization analyses

Several systematic reviews and large biobank studies have reported associations of self-reported insomnia symptoms, short and long sleep duration, and chronotype (i.e. having an evening rather than morning preference) with increased risk of cardiovascular disease, type 2 diabetes, and risk factors for these [1‐9]. The mechanisms underlying these associations are unclear, and it is plausible that specific sleep traits may contribute to the misalignment of various behavioural and internal physiological processes, including aspects of metabolism that causes adverse cardiometabolic health.

There is some evidence of poor sleep quality, shorter sleep duration, and having an evening chronotype being associated with higher triglyceride, total cholesterol and low-density lipoprotein cholesterol (LDL-C) levels, and lower high-density lipoprotein cholesterol (HDL-C) concentrations [10‐12]. However, the extent to which these associations are explained by confounding factors, such as body mass index [11], is unclear. Beyond conventional multivariable-adjusted regression analyses, we have previously demonstrated that sleep duration modifies the associations of genetic variation with triglycerides, LDL-C and HDL-C in a large sleep-gene interaction analysis, suggesting that possible different biological mechanisms underlie the associations of short and long sleep duration with these lipid traits [13]. However, these genetic interaction analyses do not assess causality and, like previous multivariable-adjusted regression analyses, have focused on a limited number of lipid traits.

Mendelian randomization (MR) uses genetic variants that are robustly associated with an exposure as an instrumental variable to obtain unconfounded effects of that exposure on an outcome of interest [14‐16]. Recent MR analyses have suggested a causal effect of insomnia symptoms on coronary heart disease [17] and of short (< 6 h) sleep duration on myocardial infarction risk [18].

The aim of this study was to determine the possible causal effect of sleep traits on metabolomic traits. We compared findings from adjusted multivariable regression (AMV) and MR analysis, to determine the relationships between self-reported insomnia symptoms (usually vs. sometimes/rare/never), total habitual sleep duration (per 1 h longer), and chronotype (evening vs. morning preference) and 113 nuclear magnetic resonance (NMR) metabolomic traits. Cross-sectional AMV was performed with adjustment for age, sex, and BMI in 17,370 individuals from 10 cohorts of mostly Europeans. Two-sample MR used summary results from genome-wide association studies (GWAS) of different sleep traits in 1,331,010 (insomnia) [19], 446,118 (sleep duration) [20], and 651,295 (chronotype) [21] European adults and summary results from four GWAS of 113 circulating metabolomic measures from NMR in 38,618 European adults. In secondary analyses, we explored effects of short (< 7 vs. 7- < 9 h) and long (≥ 9 vs. 7- < 9 h) sleep duration on the metabolomic traits. We highlight results that were consistent across both methods, as the different key sources of bias of the two methods (e.g. residual confounding in AMV and unbalanced horizontal pleiotropy in MR, respectively) mean that, where there is consistency, this is more likely to reflect a causal effect [22].

Full results of all AMV and MR analysis, including MR sensitivity analysis results, are presented in Additional file 1: Tables S7 to S16.

With the present multi-cohort effort, we intended to identify the potential biochemical mechanisms linking sleep to cardiometabolic disease risk. We found consistent evidence with both AMV and MR that usually (vs. sometimes, rarely or never) experiencing insomnia symptoms cause lower concentrations of citrate, total very large HDL particles and phospholipids in very large HDL particles and higher concentrations of glycoprotein acetyls. There was little consistency between AMV and MR results for total habitual sleep duration across all metabolomic traits, though a longer total sleep duration was associated with higher concentrations of creatinine in both methods. For chronotype, whilst having an evening preference was associated with higher isoleucine concentrations at our multiple-testing threshold in the AMV analyses, MR analyses did not support causality. Chronotype did not pass multiple testing with any other metabolites. Therefore, our findings do not support the notion that sleep traits have widespread effects on the investigated metabolomic traits. Nevertheless, they suggest that insomnia symptoms may influence cardiometabolic disease (as previously shown in MR [17]) through increased inflammation and also result in lower citrate levels.

The lack of a more widespread impact of sleep traits on multiple metabolomic traits is in contrast with some experimental sleep studies, although direct comparisons are not possible. For example, targeted and untargeted mass spectrometry measurements performed in frequently sampled blood (every 2 h) from 12 healthy men revealed that 109 out of 171 metabolites exhibited a circadian rhythm [51]. Furthermore, in controlled experimental conditions, this circadian variation was maintained for 78 out of these 109 metabolites over a 24-h period of total sleep deprivation. For 27 metabolites, including some lipids (13 glycerophospholipids and 3 sphingolipids), as well as tryptophan, serotonin, taurine, and 8 acylcarnitines, marked acute increases in concentrations were observed during 24 h of sleep deprivation compared with the 24 h of habitual sleep [51]. Importantly, the MR analyses assessed long-term (lifelong), rather than acute, effects of a predisposition for unfavourable quality or quantity of sleep on metabolic disturbances, which could explain the generally stronger effects in the total sleep duration MR analyses.

Glycoprotein acetyls, which we identified as a novel trait potentially influenced by insomnia symptoms, are elevated in response to infection and inflammation. C-reactive protein (CRP) is the most widely recognized marker of acute and chronic inflammation in epidemiological studies. Whilst observational studies have shown that higher circulating CRP is associated with increased cardiovascular disease risk, MR studies suggest this is not a causal relationship [52, 53]. Glycoprotein acetyls have emerged as a potentially better measure of cumulative inflammation than CRP, since glycoprotein acetyls increase late in the inflammatory process and levels are relatively stable within individuals over many years [54, 55]. In AMV analyses in prospective cohorts, glycoprotein acetyls were positively associated with cardiovascular diseases and type 2 diabetes, independently of established risk factors and CRP [55]. If these associations are shown to be causal, then it is possible that cumulative chronic inflammation, as measured by glycoprotein acetyls, mediates the effect of insomnia on coronary heart disease identified in MR analyses [17]. However, we acknowledge that our results for the effect of insomnia on glycoprotein acetyls require replication in independent and larger studies and testing in ancestries other than Europeans.

The inverse association of insomnia symptoms with citrate in both AMV and MR analyses is novel. A recent narrative review highlighted the physiological control of plasma citrate concentrations in health and disease [56]. One possible mechanism through which insomnia might influence citrate is via the relationship of insomnia with night-time eating (which is also accompanied with higher night physical activity) [57], which would result in higher TCA cycle activity and consequently lower plasma citrate concentrations. However, despite a plausible role, there is a paucity of clinical and epidemiological studies of the effect of citrate levels on disease outcomes [56]. Citrate is converted to Acetyl-CoA by the enzyme ATP citrate lyase (ACLY). This action is on the path to cholesterol biosynthesis up stream of HMGCR, the enzyme that is the target of statins [58]. Both MR and RCT evidence show ACLY inhibition reduce LDLc levels and proportionately coronary heart disease risk by a similar amount to statins [59‐61]. However, this provides only indirect evidence for a role of citrate on cardiovascular risk and it is notable that we found no strong evidence in this study of an effect of insomnia on LDLc. Therefore, the meaning of a possible effect of insomnia on citrate levels, and whether it mediates any effect of insomnia on cardiovascular disease risk is hard to discern. Whether our findings for citrate replicate would also be important to clarify.

We found evidence for associations of experiencing insomnia symptoms with higher concentrations of very large total HDL particles and phospholipids in very large HDL particles. MR and randomized controlled trials suggest that circulating HDL cholesterol is not causally related to cardiovascular disease [62‐64]. The amount of cholesterol carried in HDL particles increases with increasing particle size and emerging evidence highlights the importance of considering size, structure, and composition of lipoprotein particles when exploring their effects on cardiovascular disease [65]. In AMV analyses, inverse associations of very large, large, medium, and small HDL particles with cardiovascular disease have been observed, but these attenuated to the null with adjustment for lipids used by clinicians [42]. Thus, the relevance of possible insomnia effects on very large HDL particle concentrations, and specifically phospholipids in these particles, is unclear and require additional studies.

We found evidence in both AMV and MR analyses of a possible association of longer total sleep duration with higher creatine concentrations, a biomarker used to estimate kidney function. Established cardiovascular risk factors, such as high blood pressure and type 2 diabetes, are associated with higher creatinine concentrations [66]. Findings from multivariable regression suggest that the association of kidney function with cardiovascular disease largely reflects confounding and/or reverse causality [67]. Thus, our observations possibly suggest that longer sleep duration is an additional risk factor for chronic kidney disease rather than cardiovascular diseases, though we acknowledge MR sensitivity analyses did not support a causal effect. It is also possible longer sleep duration results in higher creatine concentrations via dehydration, though we might then have expected similar effects on more of the other metabolite concentrations. We also found a novel association of longer total sleep duration with the branched-chain amino acid isoleucine in MR analyses, though this association was not observed in the AMV analyses. This raises the possibility of masking (negative) confounding in the AMV analyses, though it would be surprising for this to specifically affect this one branch chain amino acid. It is also possible that the MR analyses are biased by unbalanced pleiotropy, although the MR-Egger intercept being very close to zero would argue against that. Higher concentrations of branched-chain amino acids, including isoleucine, are associated with increased risk of cardiovascular disease [42], though this has not been explored in MR studies. MR analyses supports a causal effect of the branched-chain amino acids on type 2 diabetes [68], and our results suggest that longer total sleep duration may mediate some of this effect. Although the mechanism of action how sleep induces higher isoleucine concentrations, speculatively, this might be the result of protein degradation required for gluconeogenesis. More research is required to further elaborate on this hypothesis.

Key strengths of our study are its novelty and the comparison of results from the largest AMV study of sleep traits with multiple circulating metabolomic measures [22] with equivalent results from MR. We harmonized questionnaire-based sleep data across all contributing studies and the NMR metabolomic platform was consistent across studies in both the AMV and MR analyses. We were able to increase the power of our two-sample MR analyses by combining unpublished summary-level GWAS results from three cohorts (total N = 13,693) with those of the largest published GWAS of the same NMR platform (N = 24,925) to date [35]. Two-sample MR assumes that the two samples are from the same underlying population and independent of each other. Given all GWAS were undertaken in adults of European ancestry and the lack of overlap in studies contributing to the metabolite GWAS with any of the sleep trait GWAS, we are confident this assumption is largely met. Most observed differences in mean metabolomic concentrations were close to the null, and in general (true) null results are less subject to bias than non-null results [69].

Important limitations include the lack of statistical power, particularly to explore possible non-linear associations for sleep duration. The platform misses a high proportion of currently quantifiable metabolites in human serum/plasma, including markers of energy balance, microbiota metabolism, vitamins, co-factors, and xenobiotics, that may be influenced by sleep traits [51]. Still, the NMR platform used in the analyses covers considerably more of the lipidome than conventional clinical chemistry measures (total cholesterol, LDL-C, HDL-C, and triglycerides) that have previously been explored and in addition includes amino acids, glycolysis metabolites, ketone bodies, and an inflammatory marker. Whilst we adjusted for age, sex, and BMI, the results obtained in multivariable-adjusted regression may be exaggerated by residual confounding from unobserved confounders such as socioeconomic position, smoking, and physical activity. As the AMV results were cross-sectional, it is also possible that variation in metabolomic traits influences sleep patterns, and some of the multivariable regression results not verified in MR are due to reverse causality. In addition, we restricted the analyses to cohorts containing mostly European participants (one cohort contributing to AMV meta-analysis, HELIUS, included non-European participants, whereas all MR analyses were restricted to Europeans). This reduces the potential for population stratification to bias our MR analyses, but hampers generalization of our findings to other ancestry groups. In addition, the cohorts contributing in the AMV meta-analysis vary in participant characteristics, in particular by age. In the cohorts used in the AMV analyses only, 2.4% reported taking medication to aid sleep. This very small proportion means these are very unlikely to have introduced any bias into our analyses. However, it is known that many prescribed, and over the counter medications, can influence sleep, and in our study, as in others exploring sleep, we were not able to do a detailed assessment of all medications. The MR results which reflect a potential lifelong genetic tendency should be less influenced by medication use. Furthermore, the use of questionnaire-based data on sleep traits might have increased measurement error. As people do not know the concentrations of their circulating metabolites or genetic variants related to those, such error is likely to be random and would therefore be expected in both analyses to bias towards the null. Accelerometer-based sleep measures could be useful to further explore the effects we have studied, but previous observational and genetic studies suggest only moderate agreement between questionnaire-based and accelerometer-based sleep duration [21, 70], and it is unclear whether the two are measuring the same construct. The MR results may have been influenced by weak instrument bias, which, if present, would be expected to bias results towards the null. The very large F-statistics for our main analyses (2537 to 13,967), and even for our secondary analyses of short and long duration (208 and 646, respectively), suggest that weak instrument bias is unlikely to have a major impact. Sensitivity analyses exploring possible bias due to directional horizontal pleiotropy were mostly consistent with the main IVW findings, though MR-Egger estimates were imprecise as expected with this method which is statistically less efficient than the main IVW method.

Taken together, our findings do not suggest widespread metabolic disruption caused by sleep traits. However, the evidence for possible effects of insomnia symptoms on glycoprotein acetyls and citrate and longer total sleep duration on creatinine and isoleucine might explain some of the effects, found in MR analyses, of these sleep traits on cardiometabolic diseases. These warrant further investigation.