Physical activity and sedentary behavior; mechanistic insights and role in disease prevention


 Even though physical activity and sedentary behavior are moderately heritable, little is known about the mechanisms that influence these traits. Here, we combine data for up to 674,980 individuals from 51 studies in a trans-ancestry meta-analysis of genome-wide association studies for self-reported moderate-to-vigorous intensity physical activity during leisure time (MVPA); leisure screen time (LST); sedentary commuting; and sedentary behavior at work. We identify 99 loci that associate with at least one trait. Loci associated with LST are enriched for genes whose expression in skeletal muscle is altered by resistance training. Molecular dynamics simulations suggest that the Glu to Ala substitution encoded by rs2229456 (ACTN3) – associated with more MVPA – disrupts salt bridge interactions and makes the alpha actinin 3 filaments more flexible. In isolated type IIA muscle fibers, the Ala-encoding allele is associated with lower maximal force and power during an isometric contraction, suggesting protection from exercise-induced muscle damage. Finally, Mendelian Randomization analyses show that the causal effect of LST on BMI is 2-3 times larger than the effect of body mass index (BMI) on LST, and that beneficial effects of LST and MVPA on several risk factors and diseases are mediated or confounded by BMI. Taken together, our results provide mechanistic insights into the regulation of MVPA and into the role of LST and MVPA in disease prevention. These insights may facilitate the development of tailored physical activity interventions.


Introduction
Low levels of physical activity (PA) have a major effect on disease burden and an estimated >5 million deaths per year might be prevented by ensuring adequate levels 1 . Despite efforts to increase PA levels in the population 2 , an estimated 28% of the world's population is insufficiently active, and the prevalence of physical inactivity in high-income countries rose from 31.6% in 2001 to 36.8% in 2016 3 . Trends of decreasing PA levels over time coincide with increases in the time spent sedentary 4 , which may pose an independent risk for public health 5,6 .
Physical activity and sedentary behavior are affected by public policy and social support, as well as by cultural, environmental and individual factors 7 . Factors like socio-economic status, built environment, and media all influence PA at a population level 7 . In parallel, innate biological factors (e.g. age, sex hormones, preexisting medical conditions, epigenetics and genetics) also explain a moderate proportion of the interindividual variability in PA and sedentary behavior.
Heritability estimates (h 2 ) range from 31% to 71% in large twin studies 8,9 . Pinpointing the genetic factors that influence daily PA will improve our understanding of this complex behavior, and may 1) facilitate unbiased causal inference; 2) help identify vulnerable subpopulations; and 3) fuel the design of tailored interventions to effectively promote PA. A mechanistic understanding of PA at a molecular level may even allow its beneficial effects to be attained through pharmacological intervention 10 .
Genome-wide association studies (GWAS) have identified thousands of loci associated with cardiometabolic risk factors and diseases 11 . However, similar efforts for PA have been sparse and have had limited success. This likely reflects the comparatively small sample size of these efforts 12 , along with heterogeneous assessments of PA across studies. More recently, GWAS using data from UK Biobank identified three loci associated with self-reported PA (n~377,000 individuals) and five with accelerometry-assessed PA (n~91,000) 13,14 . Hence, on the assumption that PA is a highly polygenic trait, many common variants influencing PA undoubtedly remain to be identified. 4 Here, we combine data from up to 674,980 individuals from 51 studies in a transancestry meta-analysis of GWAS for moderate-to-vigorous intensity PA during leisure time (MVPA); leisure screen time (LST); sedentary commuting; and sedentary behavior at work. This yields 104 independent association signals in 99 loci, implicating brain and muscle, amongst others organs. Follow-up analyses improve our understanding of the molecular basis of leisure time PA and sedentary behavior and their role in disease prevention.

Genome-wide analyses yield 99 loci associated with physical activity and sedentary traits
In our primary meta-analysis of European-ancestry men and women combined (Supp Tables 1-2), we identify 91 loci that are associated (P<5×10 -9 ) with at least one of four self-reported traits, one representing MVPA and three reflecting sedentary behaviors (Supp Table 3 independent SNPs in 88 loci (35 not previously reported 13,15 ) -are associated with LST, explaining 2.75% of its variance. We also identify 11 loci for MVPA (six not previously reported 13,15,16 , four overlap with LST) and four loci for sedentary behavior at work (all previously reported 13,15 , Supp Table 3). No loci are identified for sedentary commuting.
SNP-heritability estimates range from 8% for MVPA to 16% for LST (Supp Table 5 Genetic correlations of self-reported LST and MVPA with objective, accelerometryassessed daily PA traits in UK Biobank range from 0.14 to 0.44 (Figure 1b). Importantly, five of the eight loci previously identified for objectively assessed daily PA in UK Biobank data 13,14 show directionally consistent associations (P<0.05) with self-reported LST and/or MVPA (Supp Table 6). Vice versa, 39 LST-and four MVPA-associated loci show directionally consistent associations (P<0.05) with at least one objectively-assessed PA trait (using accelerometry) in UK Biobank (Supp Table 7). In line with this, each additional LST-decreasing and MVPAincreasing allele in unweighted genetic predisposition scores of the 88 LST-and ten MVPA-6 associated loci, respectively, are associated with higher objectively assessed daily PA levels in UK Biobank (P< 5×10 -5 for LST; P<5×10 -3 for MVPA, Supp Table 7). Taken together, these results suggest that in spite of their limitations, self-reported, intensity-and domain-specific PA traits can be meaningful proxies for daily PA in large-scale genetic association studies.
As external validation, we use the European-ancestry summary statistics of LST and MVPA to construct polygenic scores (PGSs), and examine their associations with MVPA in 8,195 BioMe participants of European (n=2,765), African (n=2,224) and Hispanic (n=3,206) ancestry. In general, a higher PGS for MVPA is associated with higher odds of engaging in more than 20 mins/week of MVPA, and a higher PGS for LST with lower odds of engaging in MVPA. Individuals at the highest decile of the PGS for LST are 26% less likely to spend more than 20 mins per week on MVPA than individuals at deciles 4 to 6 (OR [95% CI] = 0.74 [0.55-0.99]) (Figure 3).

Shared genetic architecture between physical activity, adiposity, and other traits
Using LD score regression implemented in the LD-Hub 17 , we observe significant (P<4.6×10  Table   9). In line with moderate genetic correlations, 11 of the 99 self-reported PA loci have previously been associated with obesity-related traits [18][19][20][21][22][23][24] . In addition, PGSs for higher MVPA and lower LST are associated with lower BMI in up to 23,723 participants from the BioMe Biobank (Supp Table 8), and a phenome-wide association study in 8,959 BioMe European ancestry samples shows a negative association between the PGS for MVPA and morbid obesity (P=1.1×10

7
Besides adiposity, higher PA levels are also genetically correlated with a more favorable cardiometabolic status, including lower triglyceride, total cholesterol, fasting glucose, and fasting insulin levels; and lower odds of type 2 diabetes and coronary artery disease; as well as better mental health outcomes; a lower risk of lung cancer; and with longevity (Figure 4, Supp Figure   2).

Causal inference between physical activity, adiposity, and disease outcomes
To assess directions of causality between PA and BMI, we next perform two-sample Mendelian Randomization (MR) analyses using multiple MR methods that utilize genome-wide full summary results or genome-wide significant loci (Supp Table 10 consistently show that LST and BMI causally affect each other, with the causal effect (per 1 SD unit increase in each trait) of higher LST on higher BMI being 2-3-fold larger than the effect of BMI on LST ( Figure 5A, Table 1a). Results are similar for bi-directional causal inference tests using body fat percentage instead of BMI (Table 1b). However, CAUSE cannot distinguish a model of causality from horizontal pleiotropy for body fat percentage and LST (Table 1b). We do not observe evidence for causal effects between MVPA and adiposity in either direction. CAUSE also illustrates a causal effect of higher LST on higher recalled adiposity and height in childhood (  (Figure 5B-C, Supp Table 11).

Enrichment of altered gene expression in skeletal muscle following resistance training
While behavior is mainly influenced by signals from the brain, in the case of PA, characteristics of skeletal muscle can play a facilitating or restricting role 29 . Therefore, we next examine if genes in LST-and MVPA-associated loci are enriched for altered mRNA expression in skeletal muscle following an acute bout of exercise or a period of training or inactivity 30 (Online Methods). A mild enrichment for transcripts with an altered expression in skeletal muscle after resistance training is observed for genes nearest to lead SNPs in LST-associated loci (P=0.02). Table 12). Of the ten genes driving the enrichment, PDE10A may play a critical role in regulating cAMP and cGMP levels in the striatum -a brain region that harbors the central reward system and is important for physical activity regulation 31 -and in regulating striatum output 32 ; ILF3 and NECTIN2 -near APOE -influence the host response to viral infections 33,34 ; EXOC4 plays a role in insulin-stimulated glucose uptake in skeletal muscle 35 ; and IMMP2L influences transport of proteins across the inner mitochondrial membrane 36 (Supp

Enrichment for genes involved in visual information processing and the reward system
To further improve the understanding of the biological factors that influence PA, we perform a tissue enrichment analysis using DEPICT 37 . LST-and MVPA-associated loci (P<1×10 −5 ) are most significantly enriched for genes expressed in the retina, visual cortex, occipital lobe, and cerebral cortex. This suggests that: 1) possibly subtle differences in the ability to receive, integrate and process visual information influence the likelihood to engage in MVPA; 2) MVPA improves the expression of genes that play a role in such visual processes in these tissues; and/or 3) MVPA can slow down age-related perceptual and cognitive decline 38 . The LSTassociated loci yield similar tissue enrichment results, with retina being the most significantly enriched tissue. Interestingly, enrichment for genes expressed in retina was also observed in the High Runner mouse model 39 . Areas related to the reward system (e.g. hippocampus, limbic system) and to memory and navigation (e.g. entorhinal cortex, parahippocampal gyrus, temporal lobe and limbic system) are also enriched in both LST-and MVPA-associated loci (Supp Figure 6, Supp Table 13).
We next use CELLECT 40 to identify enriched cell types using single cell RNA sequencing data from the Tabula Muris and mouse brain projects 41 . In Tabula

Prioritized candidate genes point to cell signaling, endocytosis, and myopathy
To explore mechanisms by which the identified loci may influence LST and MVPA, we next pinpoint genes in GWAS identified loci: 1) that contribute to enriched tissues or are identified by DEPICT's gene prioritization algorithm (Supp Tables 13,15); 2) whose expression in brain, blood, and/or skeletal muscle is anticipated to mediate the association between locus and outcome based on Summary-based MR 43 (Supp Table 16 Table 17) 45 ; 4) that show chromatin-chromatin interactions with credible variants in central nervous system cell types (such genes may be further from lead SNPs, Supp Table  17); 5) associated with PA in GWAS in humans and mice (see below) and located <100 Kb of the lead variant in either humans or mice (Supp Tables 18-19); and 6) driving enrichment of altered expression in skeletal muscle following resistance exercise training (Supp Table 12).
Integrating results across approaches yields 32 candidate genes in eight MVPA-associated loci;

Physical activity loci under selection point to signal transduction and wound healing
As much higher PA levels were required to ascertain sufficient nutrition in times of hunting and gathering and pre-mechanical farming as compared with today's Westernized societies 46 , a higher capacity to be physically active may have been selected for. To explore this, we examine if MVPA and LST association signals overlap with regions identified in three genome-wide selection screens [47][48][49] . Here, we show that 22 genes located <100kb of lead SNPs in three MVPA and/or LST-associated loci are located in three of 412 regions under selection in the past 50,000 years 47 (Supp Table 22). The protein-coding genes nearest the lead SNPs (<10kb) -DNM3, MST1R and FOXP1 -are also prioritized by other approaches (Supp Tables 20-22) and play a role in cell signaling and wound healing, amongst others (Supp Box 1). We next identify genes located <10kb of 15 loci under selection in the past 10,000 years -based on results from an ancient DNA scan 48 -and <100kb of PA association signals. This yields one additional gene (GRM5) that harbors a GWAS lead SNP for LST (rs1391954, Supp Table 22).
GRM5 encodes a metabotropic glutamate receptor that activates phospholipase C 50 ; another key player in signal transduction 51 , inflammation and wound healing 52 , amongst other processes. No lead SNPs for LST or MVPA are located within 1Mb of five loci under very recent selection 49 . In summary, we show that four loci selected for in the past 10-50,000 years are associated with leisure time PA and sedentary behavior today.

Overlap with genetics of voluntary wheel-running behavior in mice unveils a new human transcript
Many of the biological factors influencing PA levels are likely shared across species 53 .
Identifying loci that are associated with PA across multiple species may help prioritize candidate genes in such loci, and shed light on the mechanisms by which overlapping loci influence PA.
To this end, we compare our findings with loci identified in a GWAS for spontaneous PA in 100 inbred mouse strains, performed using the Hybrid Mouse Diversity Panel (HMDP) 54 (Supp Table 18). Nine genes in two LST-associated loci are also located within ±1Mbp of two lead SNPs for distance run and average running speed in mice (P<4.1×10 −6 ) (Supp Table 19). Of the eight genes that overlap across humans and mice in one of these two loci, TESC -highly expressed in the striatum -harbors the intronic lead SNP rs2173650 in humans (Supp Box 1).
In the mouse however, a gene without an established orthologue in humans -the lncRNA 4930413E15Rik -is considered likely causal for high voluntary wheel running behavior in mice selectively bred for 61 generations 39 . Using single cell RNA-sequencing data from GTEx 55 , we show that a sequence 1.4 Mb from rs2173650 with high conservation to the mouse 4930413E15Rik is expressed in several human reproductive tissues (Supp Figure 8).

Enrichment for associations with previously reported candidate genes
Candidate gene studies in humans have aimed to identify and characterize the role of genes in exercise (i.e., PA behavior) and fitness (i.e., PA ability) for decades. We next examine if variants in genes that have been linked to or associated with exercise and fitness show evidence of associations with self-reported LST and MVPA 12,56-60 . Of the 58 previously The C allele in rs1625595 ~300kb upstream of ACTN3 is significantly associated with higher MVPA (P=1.9×10 -11 ) as well as with higher ACTN3 expression in skeletal muscle (GTEx, P=6.6×10 -5 ). alpha-actinin-3 (ACTN3) forms a structural component of the muscle's Z-disc that is exclusively expressed in type IIA and IIX muscle fibers 64 Table   24).

The ACTN3 spectrin repeat region shows greater flexibility with 635A
Given the striking finding of MVPA and LST being associated with the ACTN3 missense variant rs2229456, but not with the ACTN3-truncating variant rs1815739, we next examine if rs2229456 (encoding E635A) has functional consequences for ACTN3's mechanistic properties at the molecular level. The results of computer-based (steered) molecular dynamics simulations and umbrella sampling (see Online Methods for details) show that the ancestral E635-encoding allele product facilitates salt bridge interactions at residue 635 with surrounding residues (e.g. R638 and Q639, Figure 6A 635A shows a more linear force versus distance relationship, with greater variance in the potential of mean force ( Figure 6D). Taken together, these results indicate that the ACTN3 635A dimer -associated with higher MVPA -exhibits greater flexibility than the E635 dimer.

Maximal force and fiber power lower in single IIA muscle fibers with ACTN3 635A
We next examine if a higher predicted ACTN3 dimer flexibility in the presence of 635A has functional consequences in isolated human skeletal muscle fibers. To this end, we compare functional readouts in 298 isolated type I and IIA fibers from Vastus Lateralis biopsies obtained from eight healthy, young, untrained male participants before and after an eccentric exercise bout 65,66 . Results from a 15,000 iteration Markov chain Monte Carlo model show that within four individuals homozygous for the R577-encoding allele, stable maximal force -with fibers submerged in activating solution -and fiber power during isotonic load clamps are lower in 32±7 fibers (mean±SD) from three E635A heterozygous individuals than in 46 fibers from an individual that is homozygous for the E635-encoding allele ( Figure 6E, Online methods).
Strikingly, both outcomes are similar for fibers from R577 homozygous individuals carrying one 635A-encoding allele and 39±6 fibers from four individuals homozygous for the 577X-encoding allele, who are anticipated to have ACTN2 incorporated in their Z-disc, instead of ACTN3.
Associations are most striking after an eccentric exercise intervention, and are -as expectedmore pronounced in type IIA than in type I fibers (Supp Figure 11). We do not observe evidence that differences at a single fiber level extrapolate to whole-body exercise performance in data from 266 de novo genotyped healthy young men 67 (data not shown), which may reflect the relatively low statistical power we have to detect such associations when compared with the

Discussion
By doubling the sample size compared with earlier GWAS, we identify 104 independent association signals in 99 loci, including 42 newly identified loci, for self-reported traits reflecting MVPA during leisure time and sedentary behavior. Around half of these also show evidence of directionally consistent associations with objectively assessed PA traits. Genetic correlations and two-sample MR analyses show that lower LST results in lower adiposity. Protective causal effects of higher PA acting through lowering BMI are observed for longevity and odds of type-2 diabetes, attention deficit hyperactivity disorder and depression. Tissue and cell type enrichment analyses suggest a role for visual information processing and the reward system in MVPA and LST, including enrichment for dopaminergic neurons. Loci associated with LST are enriched for genes whose expression in skeletal muscle is altered by resistance training. In silico annotation using a range of approaches helps prioritize 263 unique candidate genes across 68 LST-and seven MVPA-associated loci. Of these, five genes are located in three loci that have been under selection in the last 10-50,000 years and suggest a role for cell signaling and wound healing. Moreover, the 29 candidate genes flagged by >1 prioritization approach point to pathways related to cell signaling, endocytosis, and myopathy. Two LST-associated loci are also associated with voluntary wheel-running behavior in mice. One of these loci harbors a sequence with high conservation to a presumed causal lncRNA in mice for which we identify a previously unknown transcript in humans that is expressed in reproductive tissues. Finally, results from molecular dynamics simulations, umbrella sampling, and single fiber experiments suggest that a missense variant (rs2229456 encoding E635A in ACTN3) likely increases MVPA at least in part by reducing susceptibility to exercise-induce muscle damage.
Recent MR studies reported causal protective effects of objectively assessed PA on depression, colorectal cancer, and breast cancer 68,69 , but did not examine mediation by BMI.
The MR study for cancer concluded that a 1 SD increase in device-measured PA was associated with lower odds of breast (OR=0.51) and colorectal cancer (OR=0.66) 68 . Both the direction and size of the effect are consistent with our univariable MR results. Furthermore, a causal effect of objectively assessed, but not self-reported PA (i.e., MVPA) on depression has been reported 69 . Our MR results for LST on depression show that while the PA trait matters, the self-reported nature of it is inconsequential. According to an earlier study, TV viewing has an attenuated effect but still causes coronary artery disease when adjusting for BMI 15 . The discrepancy with our results -suggesting mediation or confounding by BMI -highlights the importance of including variants associated with both PA and BMI in multivariable MR analysis, to prevent loss of precision and potentially even biased estimates 28 .
Aiming to improve the understanding of the molecular basis of PA, we perform a range of largely complementary approaches to identify candidate genes through which the association signals are anticipated to act. Strikingly, of the 263 unique genes prioritized across 68 LST-and seven MVPA-associated loci, only seven genes are prioritized by two approaches when using traditional cut-offs within each approach. This illustrates the complexity of in silico gene prioritization for complex behaviors, especially when proof-of-concept genes are sparse and a gold standard approach for prioritization is nonexistent. When combining results from multiple approaches, applying more lenient criteria in individual approaches is justifiable. First, because the odds that a gene with an FDR of (e.g.) 0.20 in two methodologically independent approaches represents a false positive is low (i.e. 0.04); and secondly because in silico gene In conclusion, our results shed light on genetic variants and molecular mechanisms that influence physical activity and sedentary behavior in daily life. As would be expected for complex behaviors that involve both motivation and physical ability, these mechanisms occur in multiple organs and organ systems. We also provide evidence that the causal effect of sedentary behavior on adiposity is 2-3 times larger than vice versa, and support the important public health message that a physically active lifestyle mitigates the risk of multiple diseases in major part through an effect on BMI.

Samples and study design
We conducted the largest meta-analyses for physical activity (PA) traits to date, including results from up to 674,980 individuals (including nearly half-a-million from the UK Biobank) to identify genetic loci associated with PA and sedentary behavior across different ancestries. We first examined genome-wide, ancestry-and sex-stratified associations in 51 studies with questionnaire-based data on: 1) moderate-to-vigorous intensity PA during leisure time (MVPA); 2) leisure screen time (LST); 3) sedentary commuting behavior; and/or 4) sedentary behavior at work, using study-specific, tailored analysis plans (Supp Table 2). Next, we performed ancestry-specific, inverse-variance weighted fixed effects meta-analyses of summary statistics for each of the four self-reported traits (Figure 1a), including data from up to 674,980 individuals of European (93.9%), African (1.8%), East Asian (0.9%), South Asian (1.4%), and Hispanic (2.0%) ancestry (Supp Table 1 Each study obtained informed consent from participants and approval from the appropriate institutional review boards.

Self-reported physical activity traits
We defined four self-reported PA traits: MVPA (more than 20 mins/week or not, i.e., the median in most studies); LST (hours/day); sedentary commuting behavior (driving a car vs. using other modes in those employed and commuting 73 ) and sedentary behavior at work (mostly sitting and no heavy lifting vs. other in those employed 74 ).
The self-reported outcomes are domain-and intensity-specific PA and sedentary traits that, unlike accelerometry-based outcomes, are subject to misclassification and bias by recall and awareness of the beneficial effects of PA, amongst others. Furthermore, different studies used different questionnaires to capture PA, and so we defined cohort-specific traits that make optimal use of the available data, whilst striving for consistency across studies (Supp Table 2).
As a result, and based on the zero-inflated negative binomial nature of the distribution of MVPA in most studies, we had to analyze MVPA as a dichotomous outcome, which had a negative impact on statistical power.
Besides trait definitions, the average age per cohort ranged from early adulthood to oldage (17-74 years old). The power to detect genetic factors that influence PA was thus likely compromised by misclassification of physically active and inactive individuals, and heterogeneity by the inclusion of older age groups in the meta-analysis, as the heritability of PA decreases with increasing age 75 Table 1.

Genotyping, imputation and quality control
Detailed information about the genotyping platform and quality control measures for each study are presented in Supp Table 2. Quality control following study level analyses was conducted using standard procedures 76 .

GWAS and meta-analyses
GWAS were performed within each study in a sex-and ancestry-specific manner. Additive genetic models accounting for family relatedness (where appropriate) were adjusted for age, age 2 , principal components (PCs) reflecting population structure and additional study-specific covariates as presented in Supp Table 2. For all outcomes, we examined associations with and without adjusting for BMI. To avoid drawing conclusions that are driven by collider bias 77 , we did

21
To identify genome-wide significant loci, we defined a distance criterion of ±1 Mb surrounding each genome-wide significant peak (P<5×10 −9 ). We extracted previously reported genome-wide significant associations within 1 Mb of any index variants we identified from the NHGRI-EBI GWAS Catalog 11 and PhenoScanner V2 81 . A locus is considered previously reported if any variant we extracted at that locus was in LD (r 2 >0.1) with a lead variant that has been associated with objectively assessed or self-reported PA traits previously. To identify PAassociated loci that were previously associated with obesity-related traits, we performed a look up for each lead variant (and their proxies with LD r 2 >0.2) in the GWAS catalog and PhenoScanner V2.

SNP-based heritability estimation
To estimate the heritability explained by genotyped SNPs for each PA trait, we used BOLT-REML variance components analysis 82 , a Monte Carlo average information restricted maximum likelihood algorithm implemented in the BOLT-LMM v2.3.3 software. Like in most GWAS for complex traits, the SNP heritability (up to 16%) was lower than the heritability estimates from twin studies (31 -71%) 8,9 . A recent study using whole genome sequencing data showed that additionally including rare variants can fully recover the 'missing' heritability 83 . This warrants further, much larger studies in which the association of PA traits with rare variants is examined.
Although we performed a trans-ancestry meta-analysis, data from relatively few individuals of Non-European ancestries were available to us, and our functional follow-up analyses were conducted based on the European ancestry results. Studies with data from more individuals of non-European ancestry will no doubt further increase the understanding of PA etiology.

Joint and conditional analyses
To identify additional independent signals in associated loci, we performed approximate joint and conditional SNP association analyses in each locus, using GCTA 84 . Any lead SNPs 22 identified in known long-range high-LD regions 85 were treated as a single large locus in the GCTA analysis. We used unrelated European ancestry participants from the UK Biobank as the reference sample to acquire conditional P-values for association.

Phenome-wide association study with physical activity polygenic scores
To assess the out-of-sample predictive power of the variants associated with self-reported PA, We used logistic regression to separately model each PheWAS trait as a function of the two PGSs, adjusting for age, age 2 , sex and the top ten PCs. Interpretation of results was restricted 23 to outcomes with more than ten cases. Multiple testing thresholds for statistical significance were set to P<4.8×10 −5 (0.05/1,039).

Genetic correlations
To explore a possibly shared genetic architecture, we next estimated genetic correlations of the four self-reported traits examined in this study and five accelerometry-assessed PA traits assessed in UK Biobank 14 with relevant complex traits and diseases based on established associations at the trait level using LD score regression implemented in the LD-Hub web resource 17 . To define significance, we applied a Bonferroni correction for the 108 selected phenotypes available on LD-Hub (P<4.6×10 −4 ). Supp Table 9 shows the complete set of pairwise genetic correlations of the four self-reported PA traits with relevant complex traits and diseases. Next, we prioritized traits and diseases showing evidence of genetic overlap (significantly associated with at least one of the PA traits). These can be divided into six categories: lifestyle traits, anthropometric traits, psychiatric diseases, other diseases (cardiometabolic diseases and cancer), biomarkers and others (Figure 4). Using objectively assessed PA traits (i.e. using accelerometry) instead of self-reported traits yielded similar results (Supp Figure 1).

Two sample Mendelian Randomization
We performed Mendelian Randomization (MR) analyses to disentangle the causality between LST and MVPA on the one hand and BMI on the other hand. Since UK Biobank provided the majority (>75%) of the samples in the meta-analyses, we selected summary statistics of variants for relevant traits and diseases using data from the largest publicly available GWAS without data from UK Biobank participants on the MR-Base platform 87 (Supp Table 10), aiming to minimize bias due to sample overlap in the two sample MR analysis 88 . Genetic instrumental variables for each of the traits and diseases consisted of genome-wide significant (P<5×10 -8 ) index SNPs. Index SNPs were LD clumped (r 2 >0.001 within a 10 Mb window) to remove any 24 correlated variants. In the multivariable MR that evaluates the independent effects of each risk factor, the genetic instrumental variables from two risk factors were combined. For both LST and MVPA, independent loci associated with PA and BMI were used as instrumental variables.
We followed several steps to evaluate potential causality. As

Enrichment for genes with altered expression in skeletal muscle after an intervention
A high degree of physical fitness and a strong adaptive response to exercise interventions facilitate a physically active lifestyle. To identify plausible candidate genes in GWAS-identified loci, we examined enrichment for transcripts whose expression in skeletal muscle was changed after an acute bout of aerobic exercise, aerobic training, an acute bout of resistance exercise, resistance training, and inactivity 30 . We excluded individuals with pre-existing conditions such as chronic kidney disease, chronic obstructive pulmonary disease, frailty, metabolic syndromes, and obesity. We also excluded athletes because in this subgroup, transcripts with differential expression in response to such interventions are likely not representative for the general population 92 . Enrichment was examined for genes nearest to, or within 1 Mb of lead variants for LST-and MVPA-associated loci. We used FDR<0.01 as the threshold for altered expression after intervention. A sensitivity analysis with a series of different FDR cut-offs (0.001 to 0.5) showed that results were robust.

Annotation using DEPICT, CELLECT, SMR, FINEMAP and 3D chromatin interactions
We used DEPICT 37 to identify enriched gene sets and tissues, as well as to prioritize candidate genes in the identified loci, using variants with P<1×10 −5 in the primary meta-analysis of European ancestry men and women combined as input. We also used CELLECT 40 to identify enriched cell types for PA, by combining MVPA and LST GWAS summary statistics with single cell RNA sequencing data. We sought to further refine the set of prioritized candidate genes To identify variants in GWAS-identified loci with a high posterior probability of being causal, we used LST and MVPA summary statistics as input for FINEMAP 44 . We used default parameters and selected a maximum of 10 putative causal variants per locus. The output variants identified as credible were mapped to genes using tissue-specific HiC chromatin conformation capture data 98 . We integrated all HiC data in the brain (dorsolateral prefrontal cortex, hippocampus, neural progenitor cell, and adult and fetal cortex) available on FUMA v1.3.5, using the same approach. Genes in GWAS-identified loci containing FINEMAP-identified credible coding variants with a CADD score > 12.33 were also prioritized.

Mouse experiments
Females from 100 genetically distinct strains from the Hybrid Mouse Diversity Panel (HMDP) 99 were purchased from Jackson Laboratories (University of Tennessee Health Science Center).
They arrived at UCLA at 5 to 8 weeks of age and were housed 1-4 weeks until wheel testing. All mice were ~3 months old at the start of the experimental protocol, and were randomized into  55 to study expression that might be present at low levels in specific human tissues.

Enrichment for previously reported candidate genes
We next conducted a literature review of previously reported genes with evidence of a role in exercise (i.e., PA behavior) and fitness (i.e., PA ability) and identified 58 such candidate genes

Molecular dynamics simulation for E635A
Alpha-actinin is a structural member of vertebrate muscle Z-discs, and primarily functions to cross-link neighboring actin filaments of opposite polarity from adjacent sarcomeres. This binding can occur over a range of angles from 60 to 120°, creating a tetragonal lattice with a lattice spacing of 19 to 25 nm [100][101][102] . In addition to its interaction with actin, alpha-actinin binds and anchors titin to the Z-disc 103 . The alpha-actinin homodimer is formed from two antiparallel subunits composed of an N-terminal actin-binding domain and a C-terminal calmodulin homology domain (CAM), separated by four spectrin-like repeats. Each repeat consists of a triple α-helix coiled-coil ( Figure 6A). Alpha-actinin 3 (ACTN3) at 901 amino acids in length is one of four isoforms of alpha -actinin and is exclusively found in human type-II (also known as fast-twitch) skeletal muscle fibers. The naturally occurring truncating mutation R557X in ACTN3 has a potential impact on injury risk during exercise, increased muscle-damage following eccentric training and increased flexibility for 557X homozygotes 63 .
As no structure for human ACTN3 has yet been experimentally determined, we constructed a homology model of the E635 variant monomeric filament using the fully annotated protein (Uniprot ID Q08043) using Phyre2 104 , with the 635A variant mutated in silico. For each variant, the spectrin repeats of the ACTN3 monomer were aligned with the crystal structure of the rod domain of alpha -actinin (PDB ID 1HCI), to give the dimeric form of ACTN3. Molecular dynamics (MD) system preparation and simulation was conducted with GROMACS 2020.1 105 .
The MD topology was created with GROMACS pdb2gmx using the ACTN3 dimer model and parameterized with the CHARMM36 all-atom force field 106 . The ACTN3 dimer was placed in a rectangular simulation box with a 1.0 nm buffer between the protein and the box extent, with periodic boundary conditions in all three spatial axes. The system was solvated with TIP3P water molecules and using GROMACS genion, random solvent molecules were replaced with K + and Clto a concentration of 150 mM with additional K + ions added to provide an allele product, associated with higher MVPA.

Steered molecular dynamics and Umbrella sampling for E635A
We which the spectrin repeats were initially aligned. Suitable frames from each steered molecular dynamics simulation were selected that differed by no more than 0.2 nm from 0 to -3 nm (a contraction of the dimer by 3 nm) and were used as the starting topology for a series of 10 ns umbrella sampling simulations. Analysis of the umbrella sampling simulations was conducted using g_wham to yield the potential of mean force versus reaction coordinate for each variant.

31
When a compressive force was applied between the center of mass of the two actinbinding domains, the force required to compress the two actin-binding domains by 0.2 nm was lower for 635A compared with E635 (~28 v. 55 kJ/mol/nm). Furthermore, the force-to-distance relationship to a compressive distance of -1.2 nm -where the two respective forces converge (67 kJ/mol/nm) -was notably more linear for 635A than for E635 (Supp Figure 10). Greater variability is also seen for 635A in the force versus distance relationship among triplicate steered molecular dynamics simulations. To explore this further, we used umbrella sampling to examine the change in potential of mean force (free energy surface) over the reaction coordinate corresponding to the compression of the ACTN3 dimer.
Umbrella sampling of the ACTN3 dimer variants showed that 635A reaches an energy minimum at a distance between actin-binding domains that is 0.35 nm shorter than E635. Initial