Human brain aging and Alzheimer’s dementia (AD) impose a huge health and economic burden on modern society. Recent research suggests that cognitive decline begins in middle age (45–65 years old) (Singh-Manoux et al., 2012; Karlamangla et al., 2017). However, there is a substantial heterogeneity in the trajectories of cognitive aging, with some individuals showing minimal or no decline while others of the same age experiencing rapid decline or even dementia. Discovery of sensitive, specific and non-invasive biomarkers is a prerequisite for identifying high-risk individuals, and developing effective strategies that can monitor, delay or prevent the onset of dementia.

A growing body of evidence suggests that brain and cognitive aging are accompanied by extensive metabolic perturbations (Clarke et al., 2018). The altered metabolic profiles can be quantified by metabolomics, a high-throughput biochemical technology that can identify hundreds to thousands of small molecules (metabolites) in biofluids or tissues (Rochfort, 2005). These metabolites represent the endpoints of metabolic processes encompassing the interaction between the internal genome and the external environment (Harrigan and Goodacre, 2012), and thus are closer to disease phenotypes compared to genomic, transcriptomic and proteomic profiles.

While the majority of the existing metabolomics studies in human (Han et al., 2011; Orešič et al., 2011; Björkqvist et al., 2012; Mapstone et al., 2014; Inoue et al., 2015; Casanova et al., 2016; Varma et al., 2018; Huo et al., 2020) have focused on older individuals (aged 65+), several recent studies reported metabolic changes associated with cognitive aging (Bressler et al., 2017; Chouraki et al., 2017; Darst et al., 2018; van der Lee et al., 2018; Proitsi et al., 2018; Tynkkynen et al., 2018) among individuals in midlife (45–65 years). However, most of these studies have examined a small number of metabolites (in general less than 300), leaving the full spectrum of blood metabolome largely unexplored. To date, no consensus metabolites that can predict brain and cognitive aging in midlife have been reported.

Leveraging the comprehensive clinical and cognitive phenotypes in a well-characterized longitudinal twin cohort, the current study aimed to identify metabolites predictive of incident mild cognitive impairment (MCI) or cognitive change using untargeted metabolomics in 30 middle-aged (mean age 57.2 at baseline), male-male twin pairs [15 monozygotic (MZ) pairs, 15 dizygotic (DZ) pairs]. All twins were cognitively normal at baseline (Wave 1, 2003–2007), but one twin in each of these pairs developed incident MCI, while his co-twin remained cognitively normal at the end of 5.5-year follow-up (Wave 2, 2008–2013). Our initial study was followed by replication in an independent cohort of middle age adults (mean age 59.1 at blood draw). Although preliminary, findings of this research are likely to generate novel hypotheses that may lead to the identification of novel metabolic markers predictive of cognitive aging in midlife and provide insights into our understanding of the role of metabolic disturbance in early cognitive aging.

In an untargeted metabolomic analysis including 30 middle-aged twin pairs discordant for progression to MCI, we identified four fast plasma metabolites whose baselines significantly predict cognitive change over time after adjusting for multiple testing and known covariates. Of these, the association between a sphingomyelin (SM [d18:1/20:1, d18:2/20:0]) and the longitudinal change in executive function was replicated in an independent cohort of middle-aged adults. Metabolic network analysis revealed that a reduced metabolic connection could be associated with progression to MCI. The identified metabolites were enriched in several metabolic pathways. Together, our results provide initial evidence that metabolic perturbation affects cognitive aging in midlife and suggest that plasma metabolites may serve as potential biomarkers for early cognitive aging and AD prediction in midlife.

Of the four metabolites predictive of cognitive change identified in twins, higher baseline levels of two sphingomyelins predict a slower decline in executive function, and one of these associations was replicated in WRAP. This finding is in agreement with findings from previous epidemiologic study among older adults (age 65+) (Mielke et al., 2010, 2011; Han et al., 2011; Orešič et al., 2011). The relationship between sphingomyelins and cognitive function was also consistent with previous research demonstrating that a disrupted sphingomyelin pathway is involved in Aβ pathology (Jana and Pahan, 2004), and a reduced level of sphingomyelin in human AD brain than that in cognitively healthy individuals (He et al., 2010). We were unable to replicate the other three metabolites in the WRAP, probably due to the small sample size, the heterogeneous clinical phenotypes across different studies, or other differences such as WRAP included women as well and participants who were enriched with a parental history of AD. The lack of replication could also be attributed to false positives as a result of the small sample size in our discovery cohort.

We also found a higher baseline level of plasma DAG (18:2/20:4) was associated with a slower decrease in short-term memory. While the precise mechanism underlying this association is unclear and awaits further investigation, the association appears to be consistent with previous studies showing that neuronal glycerolipid biosynthesis promotes axon regeneration after optic nerve injuries (Yang et al., 2020), and that diffuse axonal injury causes cognitive impairment including short-term memory (Wallesch et al., 2001). In addition, a higher plasma level of hydroxy-CMPF was associated with a slower decrease in general cognitive ability. The CMPF is a furan fatty acid that inhibits insulin secretion in both human and mouse models (Nagy et al., 2017). A higher level of CMPF has been associated with type 2 diabetes (Prentice et al., 2014) and biological aging (Menni et al., 2013) in human. While these findings are seemingly contradictory to the protective effect of hydroxy-CMPF on cognitive function identified in the current study, it is possible that hydroxylation of CMPF may reduce the plasma concentration of CMPF, which could likely improve neuronal insulin resistance, thereby exerting beneficial effects on cognition.

Our network analysis revealed that loss of metabolites connectivity (the pink module) may be related to cognitive change, and we identified several hub metabolites that differ significantly between MCI converters and non-converters. Although the precise mechanisms are unclear, it is likely that the differences in network connectivity and hub metabolites may explain, at least in part, why some twins progressed to MCI while his co-twin brother did not. For example, we found that 2,2′-methylenebis(6-tert-butyl-p-cresol), ceramide (d16:1/24:1, d18:1/22:1), and gamma-glutamylglutamate were the hub metabolites only in non-converters but not in converters. Ceramides are sphingolipids that play important roles in cellular processes including differentiation, proliferation, and apoptosis (Hannun and Obeid, 2008). Dysregulation of ceramides in the brain has been implied in Alzheimer’s and other neurodegenerative disease (Filippov et al., 2012; Jazvinšćak Jembrek et al., 2015). Gamma-glutamylglutamate acid is a dipeptide consisting of gamma-glutamate and glutamic acid. Glutamic acid is the main excitatory neurotransmitter in the brain, which can be released to neurons and form glutamate; the latter can convert to glutamic acid via astrocytes, which completes the glutamate-glutamine cycle (Daikhin and Yudkoff, 2000). Dysregulation of glutamate signaling has been implicated in neurodegeneration and cognitive behaviors (Vogel et al., 1966; Rahn et al., 2012).

Among the metabolic pathways we identified to be associated with cognitive changes, the aminoacyl-tRNA biosynthesis pathway showed the most significant association with changes in short-term memory. The process of aminoacyl tRNA biosynthesis catalyzes the ligation of amino acids to their cognate tRNAs, and dysregulation of this pathway has been associated with neurodegeneration (Park et al., 2008; Kapur et al., 2017). The observed association of the amino acid (valine, leucine, and isoleucine) pathway with short-term memory is consistent with previous research demonstrating that branched-chain amino acids (BCAAs) were associated with dementia (Tynkkynen et al., 2018) and metabolic conditions such as insulin resistance and diabetes (Roberts et al., 2014), both of which have been linked to cognitive impairment (Ma et al., 2015) and AD (Arnold et al., 2018).

In this study, we identified associations of circulating plasma metabolites with cognitive change over time, independent of known risk factors. The mechanisms underlying the observed associations are unclear but believed to be highly complex. While genome-wide association studies (GWAS) have identified many genetic variants associated with cognitive functions, cognitive impairment and AD in various human populations under different clinical settings (Trampush et al., 2017; Jansen et al., 2019), the identified genetic variants explain only a small proportion of the disease risk. A larger contributor to the phenotypic variation may be due to the complex interactions between environmental and inherited factors. Metabolomics provides a snapshot of hundreds to thousands of small molecules (metabolites) in a biological sample (e.g., plasma) at a given time. The abundance of these metabolites is influenced by both genetic and environmental factors, such as diet, lifestyle, behavior, drug, microbiome, hormones, etc., probably through various epigenetic mechanisms. Metabolomic study of circulating metabolites could facilitate the identification of environmental factors contributing to cognition changes during aging, and provides mechanistic insight into how environmental exposures contribute to cognitive impairment in AD progression. Moreover, we identified that different plasma lipids were associated with different cognitive domains, including general cognitive ability, short term memory, executive memory, but not episodic memory. While the mechanism underlying this differential associations remain to be determined, the lack of association with episodic memory might suggest that the underlying pathological mechanisms are dementia-related but not necessarily specific to AD. However, it is worth noting that the VETSA and WRAP samples are both relatively young. Although prospective studies of AD usually emphasize episodic memory, there is also evidence suggesting that executive function can predict progression to MCI or AD as early or possibly earlier than memory deficits (Clark et al., 2012; Ewers et al., 2012; Gibbons et al., 2012). Executive function deficits, which were associated with metabolites in the present study, are also common in MCI and AD, and executive control functions do affect functioning in other cognitive domains (Buckner, 2004; Baudic et al., 2006). Indeed, there is also evidence for an executive-prominent AD subtype (Mukherjee et al., 2012). Thus, although it remains uncertain, it is possible that the associated deficits that were observed will contribute to AD-related deficits.

Limitations of our study include the small sample size and the inclusion of only Caucasian men in the twin study, the sample differences across studies such as the selection of WRAP biased toward parental history of AD, and the different cognitive measures in our discovery and replication cohorts. The uncertainty of the unknown metabolites represents another limitation. However, our study has several strengths, including a focus on midlife, the well-matched MCI discordant twin pairs (which eliminates confounding by age and sex and minimized confounding by multiple background factors), the inclusion of a replication cohort comprised of middle-aged adults in a similar age range, the prospective design of both discovery and replication cohorts, the comprehensive cognitive phenotypes, and the comprehensive coverage of plasma metabolome by using an untargeted metabolomics approach on the same platform in both cohorts.

In summary, in an untargeted metabolomics analysis of middle-aged adults in two well characterized prospective cohorts, we identified four plasma metabolites that are significantly predictive of cognitive change over time, after adjusting for multiple testing and known clinical covariates. The observed association of a sphingomyelin (SM [d18:1/20:1, d18:2/20:0]) with executive function was replicated in an independent cohort including participants in a similar age range. We also identified metabolic networks and hub metabolites that may be involved in the relationship between metabolic dysregulation and cognitive impairment in midlife. Our results support that metabolic perturbation occurs many years before cognitive impairment, and suggest the possibility of using blood metabolites as non-invasive biomarkers in early prediction of cognitive aging and AD in midlife.

The data analyzed in this study is subject to the following licenses/restrictions: The authors will accept reasonable requests for data access, which will be referred to the Veteran Era Twin Registry where the data are maintained and overseen. Requests to access these datasets should be directed to http://www.vetsatwins.org.

The studies involving human participants were reviewed and approved by twins participating in the VETSA were administered identical protocols at the University of California, San Diego, or Boston University. Participants in the WRAP were administered at the University of Wisconsin. Written consents were obtained from all participants enrolled in both studies. The protocols were approved by Institutional Review Boards at the participating universities. The patients/participants provided their written informed consent to participate in this study.

ZH, BR, and JE performed the statistical analysis. RD, CE, and SJ performed the replication analysis in WRAP. ML, CF, WK, and JZ designed and coordinated the study. ZH and JZ wrote the manuscript. All authors reviewed the manuscript.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. The reviewer LY declared a past co-authorship with several of the authors ZH and JZ to the handling editor.