Accelerometer-measured physical activity and cognitive functioning: A Mendelian Randomization study

Physical activity and cognitive functioning are strongly intertwined. However, the causal relationships underlying this association are still unclear. Physical activity can enhance brain functions, but healthy cognition may also promote engagement in physical activity. Here, we assessed the bidirectional relationships between physical activity and general cognitive functioning using Latent Heritable Confounder Mendelian Randomization (LHC-MR). Association data were drawn from two large-scale genome-wide association studies (UK Biobank and COGENT) on accelerometer-measured moderate, vigorous, and average physical activity (N = 91,084) and cognitive functioning (N = 257,841). After Bonferroni correction, we observed significant LHC-MR associations suggesting that increased fraction of both moderate (b = 0.32, CI 95% = [0.17,0.47], P = 2.89e-05) and vigorous physical activity (b = 0.22, CI 95% = [0.06,0.37], P = 0.007) lead to increased cognitive functioning. In contrast, we found no evidence of a causal effect of average physical activity on cognitive functioning, and no evidence of a reverse causal effect (cognitive functioning on any physical activity measures). These findings provide new evidence supporting a beneficial role of moderate and vigorous physical activity (MVPA) on cognitive functioning.


Abstract
Physical activity and cognitive functioning are strongly intertwined. However, the causal relationships underlying this association are still unclear. Physical activity can enhance brain functions, but healthy cognition may also promote engagement in physical activity. Here, we assessed the bidirectional relationships between physical activity and general cognitive functioning using Latent Heritable Confounder Mendelian Randomization (LHC-MR).
Association data were drawn from two large-scale genome-wide association studies (UK These findings provide new evidence supporting a beneficial role of moderate and vigorous physical activity (MVPA) on cognitive functioning.

Significance Statement
Whether the observed correlation between physical activity and cognitive functioning reflects causal relationships in either direction is still unclear. To investigate the potential causal relationships underlying this association, we conducted recently-improved geneticallyinformed analyses with two large-scale datasets including accelerometer-measured physical activity. In line with theoretical models and previous experimental work explaining the mechanisms underlying the association between physical activity and cognitive functioning, our results revealed a one-way association: Higher levels of moderate and vigorous physical activity can potentially cause higher cognitive functioning. These findings suggest that physical activity plays a fundamental role in general cognitive functioning, and that health policies and interventions promoting moderate or vigorous physical activity are relevant to improve cognitive performance and to delay its decline.
For example, greater cognitive functioning may be required to counteract the innate tendency for effort minimization, and thereby influence a person's ability to engage in physically active behaviors (36)(37)(38)(39)(40)(41). Of note, these mechanisms are not mutually exclusive and could therefore lead to bidirectionally reinforcing relationships (i.e., positive feedback loop) between physical activity and cognitive functioning (42).
Although previous findings point to a potential mutually beneficial interplay between physical activity and cognitive functioning, these findings mainly stem from observational designs and analytical methods that cannot fully rule out the influence of social, behavioral, and genetic confounders or uni-directional causation (42). Accordingly, evidence for a one-or two-way association between physical activity and cognitive functioning could be considered weak.
Mendelian Randomization (MR) is a statistical approach for causal inference that can overcome this weakness of traditional observational studies. Specifically, MR uses genetic variants that are randomly distributed in a population as instruments to reduce the risk of confounding or reverse causation (43,44). MR-based effect estimates rely on three main assumptions (45) stating that genetic instruments i) are strongly associated with the exposure (relevance assumption), ii) are independent of confounding factors of the exposure-outcome relationship (independence assumption), and iii) are not associated to the outcome conditional on the exposure and potential confounders (exclusion restriction assumption assume that at least half of or the most "frequent" genetic instruments are valid/non-pleiotropic. However, despite the extensions and their assumptions, these methods still suffer from two major limitations. First, they only use a subset of markers as instruments (genome-wide significant markers), which often dilutes the true relationship between traits. Second, they ignore the presence of a potential latent heritable confounder of the exposure-outcome relationship (e.g., body mass index, educational attainement, level of physical activitx at work, or material deprivation).
The present study applies the Latent Heritable Confounder MR (LHC-MR) method (46), which addresses the aforementioned limitations, to simultaneously estimate the bidirectional causal effects between physical activity and cognitive functioning, while accounting for possible heritable confounders of their relationship. Unlike standard MR, LHC-MR accounts for sample overlap from genome-wide genetic instruments, thereby allowing the exposure and the outcome to originate either from overlapping datasets or the same dataset. As such, LHC-MR is an innovative method that exploits the full geneome-wide architecture of the exposure trait to maintain sufficient power when few genome-wide significant instruments are available, as is the case for physical activity. In the current study, LHC-MR maintained a reasonable level of power to accurately estimate the bidirectional relationships, while standard MR methods were likely underpowered (46).
Here, the causal estimates were modelled based on summary statistics from large-scale GWAS of accelerometer-measured physical activity (47) and general cognitive functioning (48,49).
Since it has been suggested that the intensity of physical activity may be an important consideration, with moderate intensity having greater beneficial effects than vigorous intensity (50)(51)(52)(53)(54), we assessed whether the causal effect estimates on cognitive functioning were dependent on physical activity intensity (i.e., moderate vs. vigorous vs. average).

Results
Three measures derived from accelerometer wear were used as a proxy for physical activity: average, moderate, and vigorous physical activity. These three measures were used in LHC-MR to investigate the possible bidirectional causal effects between them and cognitive functioning. The model tested was adjusted for age, sex, genotyping chip, first ten genomic principal components (PC), center, and season (month) of wearing accelerometer. The Bonferroni correction was used to control for familywise error rates, yielding an α = 0.05 / (2 directions × 3 tests) = 0.008. As was found with average physical activity, there was no evidence for the presence of a heritable confounder. Standard MR methods yielded non-significant causal estimates in both directions ( Table 2).  Table 2).

Sensitivity analyses
We

Discussion
This study used a genetically informed method that provides evidence of putative causal relations to investigate the bidirectional associations between accelerometer-based physical activity and general cognitive functioning. Drawing on large-scale GWAS, we found evidence for potential causal effects, suggesting that higher levels of moderate and vigorous physical activity lead to increased cognitive functioning. In the opposite direction, we did not observe evidence of a causal effect of cognitive functioning on physical activity. Hence, our study suggests a favorable effect of moderate and vigorous physical activity on cognitive functioning, but does not provide evidence that increased cognitive functioning promotes engagement in more physical activity.
Previous reviews and meta-analyses of observational studies showed a beneficial effect of physical activity on cognitive functioning (13,16,17,34). However, the evidence arising from intervention studies was inconclusive (18,19,(21)(22)(23)55). It has been argued that these inconsistencies may primarily be attributed to the design-specific tools used to assess physical activity (21 Yet, while the assumptions of LHC-MR may not hold, the assumptions of the other five methods are known not to hold because of insufficient genome-wide significant instrument. To the best of our knowledge, our study is the first to investigate the potential causal relationship between physical activity and cognitive functioning using a genetically informed method. We are aware of only two other, non-genetic studies that examined the potential bidirectional associations between physical activity and cognitive functioning (8,20). In contrast to the present study, those two studies observed a positive influence of cognitive functioning on physical activity. At least two factors can explain the differences in the results observed. First, both those studies are based on longitudinal assessment (Granger causality) of the two traits, while our approach is based on a genetically instrumented causal inference technique (LHC-MR). Second, these studies draw on self-reported physical activity rather than accelerometer-measured physical activity, which may have biased the observed associations between cognitive functioning and physical activity.
Our results obtained with recently-improved genetically-informed analyses (LHC-MR) highlight the potential critical role of physical activity, specifically of moderate and vigorous intensity, on cognitive functioning. However, it should be noted that the estimated effect of moderate physical activity on cognitive functioning was about 1.5 times stronger in magnitude than the effect of vigorous physical activity. This result is consistent with previous literature that have suggested that moderate physical activity yields higher benefits of cognitive functioning compared to vigorous physical activity (50)(51)(52)(53)(54). However, to the best of our knowledge, this study is the first to assess and compare the causal relationships of moderate and vigorous physical activity with cognitive functioning with a genetically informed method based on large-scale datasets. Several mechanisms can explain how physical activity of varying intensities may enhance brain functioning. Physical activity promotes neurogenesis, gliogenesis, neuronal excitability, angiogenesis, cortical thickness, and growth factor production (19,(28)(29)(30)(31)(32)(33)(34). Moreover, the particular benefits of moderate physical activity could be explained by differences in the quantity of hormones released in the blood. For example, one study observed an inverted U relationships between physical activity intensity and endocannabinoids, with vigorous intensities reducing the concentrations in peripheral endocannabinoids compared with moderate intensities (56). Consequently, vigorous physical activity may be less effective in enhancing cognitive functions than moderate physical activity.
Another potential explanation is the stress response associated with vigorous physical activity yielding a large cortisol release that can have a detrimental effect on aspects of cognitive functioning, such as memory (57,58).
The LHC-MR method revealed two causal relations that are consistent with each other.
Importantly, these findings are consistent with theoretical and experimental work explaining the mechanisms underlying the association between the physical activity and cognitive functioning (13)(14)(15)(16)(17)(18)(19)(28)(29)(30)(31)(32)(33)(34)(35). Results obtained with both the LHC and standard MR methods showed no evidence of an effect of average physical activity on cognitive functioning. This finding can likely be explained by physical activities of low intensity (i.e., < 100 mg) that are part of the average physical activity, which further suggests that physical activity should be of moderate-to-vigorous intensity to benefit cognitive functioning. Yet, this result contrasts with previous studies showing that total volume of physical activity was associated with cardiovascular disease (59) and all causes of mortality (60). These discrepancies may be due to potential differences in the types of physical activity performed by the participants, or by the fact that cardiovascular disease and mortaility could be more responsive to light physical This difference could be explained in at least two ways. Firstly, previous studies examining the positive effect of cognitive functions on physical activity relied on self-reported physical activity, which can bias the observed associations (8,24,27). Secondly, our study relied on general cognitive functioning, whereas previous results highlight the specific importance of inhibition resources that may be required to counteract an innate tendency for effort minimization (27,(36)(37)(38)(39)41). Therefore, future studies should investigate the specific relationships between motor inhibition and physical activity when such data is available. Our findings provide preliminary support for a unidirectional relation whereby higher levels of moderate and vigorous physical activity lead to improved cognitive functioning. These results underline the essential role of moderate and vigorous physical activity in maintaining or improving general cognitive functioning. Therefore, health policies and interventions that promote moderate and vigorous physical activity are relevant to improve cognitive functioning or to delay its decline.

Data sources and instruments
This study used de-identified GWAS summary statistics from original studies that were approved by relevant ethics committees. The current study was approved by the Ethics

Physical activity
Accelerometer-measured physical activity was assessed based on summary statistics from a recent GWAS (47) analyzing accelerometer-based physical activity data from the UK Biobank.
In the UK Biobank, about 100,000 participants wore a wrist-worn triaxial accelerometer (Axivity AX3) that was set up to record data for seven days. Individuals with less than 3 days (72 h) of data or not having data in each 1-hour period of the 24-h cycle or for whom the accelerometer could not be calibrated were excluded. Data for non-wear segments, defined as consecutive stationary episodes ≥ 60 min where all three axes had a standard deviation < 13 mg, were imputed. The details of data collection and processing can be found elsewhere (64).
We examined three measures derived from the three to seven days of accelerometer wear: the average acceleration in milli-gravities (mg), the fraction of accelerations > 100 mg and < 425 mg to estimate moderate physical activity (65), and fraction of accelerations ≥ 425 mg to estimate vigorous physical activity (65) (Manhattan plots and Q-Q plots can be found in Supplementary Material 1 and 2). The GWAS for average physical activity (nmax = 91,084) identified 2 independent genome-wide significant SNPs (P < 5e-09), with a SNP-based heritability of ~ 14%.
As for the other two physical activity measures, the fractions of accelerations corresponding to moderate and vigorous physical activity were obtained by running new GWAS on the decomposed acceleration data from UK Biobank using the BGENIE software (66). The phenotype for moderate physical activity was limited to acceleration magnitudes ranging from 100 to < 425 mg, whereas vigorous physical activity was limited to acceleration magnitudes ranging from 425 to 2,000 mg. These acceleration fractions were adjusted for age, sex, and the first 40 PC, and the analyzed individuals were restricted to unrelated white-British. The two datasets of average physical activity summary statistics, alongside the moderate and vigorous physical activity summary statistics, were used in LHC-MR to investigate the possible bidirectional effect that exists between these physical activity traits and cognitive functioning.

General cognitive functioning
General cognitive functioning was assessed based on summary statistics from a recent GWAS combining cognitive and genetic data from the UK Biobank and the COGENT consortium (N = 257,841) (48). The phenotypes of these cohorts are well-suited to meta-analysis because their pairwise genetic correlation has been shown to be high (49). In the UK Biobank (nmax = 222,543) participants were asked to complete 13 multiple-choice questions that assessed verbal and numerical reasoning. The verbal and numerical reasoning score was based on the number of questions answered correctly within a two-minute time limit. Each respondent took the test up to four times. The phenotype consists of the mean of the standardized score across the measurement occasions for a given participant. In the COGENT consortium (nmax = 35,298), general cognitive function is statistically derived from a principal components analysis of individual scores on a neuropsychological test battery (67). The phenotype estimates overall cognitive functioning and is relatively invariant to the battery used and specific cognitive abilities assessed (68,69). These COGENT data used to assess general cognitive functioning were also used in another GWAS study (49). The GWAS identified 226 independent genomewide significant SNPs, with a SNP-based heritability of ~20%.

Statistical analysis
MR is an epidemiological method in which the randomized inheritance of genetic variation is considered as a natural experiment to estimate the potential causal effect of a modifiable risk factor or exposure on health-related outcomes in an observational design (43,44). MR draws on the assumption that genetic variants, because they are randomly allocated at conception, are less associated with other risk factors that may be confounders of the exposure and the outcome, and are immune to reverse causality since diseases or health-related outcomes have no reverse effect on genetic variants. Consequently, these genetic variants can be used as instrumental variables, potentially making MR less vulnerable to confounding or reverse causation than conventional approaches in observational studies (43,44).   Notes. Causal estimates from 5 standard Mendelian Randomization (MR) methods on alternating exposure and outcome traits. For both moderate and vigorous physical activity as exposure, the cutoff was decreased to 6.33e-5 because of the low number of genome wide significant single nucleotide polymorphisms (SNPs) to use as instruments. Corrected α = 0.008.  genotyping chip, first ten genomic principal components (PC), center, and season (month) of wearing accelerometer. * = significant effect after Boneferroni correction (i.e., p-value < .008).

Physical activity
Causal effect estimate * *