Genome-wide physical activity interactions in adiposity ― A meta-analysis of 200,452 adults

1In recent decades, we have witnessed a global obesity epidemic that may be driven by changes in lifestyle such as easier access to energy-dense foods and decreased physical activity (PA). However, not everyone becomes obese in obesogenic environments. Twin studies suggest that changes in body weight in response to lifestyle interventions are in part determined by a person’s genetic constitution. Nevertheless, the genes that are sensitive to environmental influences remain largely unknown.

2Previous studies suggest that genetic susceptibility to obesity, assessed by a genetic risk score for BMI, may be attenuated by PA. A large-scale meta-analysis of the FTO obesity locus in 218,166 adults showed that being physically active attenuates the BMI-increasing effect of this locus by ~30%. While these findings suggest that FTO, and potentially other previously established BMI loci, may interact with PA, it has been hypothesized that loci showing the strongest main effect associations in genome-wide association studies (GWAS) may be the least sensitive to environmental and lifestyle influences, and may therefore not make the best candidates for interactions. Yet no genome-wide search for novel loci exhibiting SNP×PA interaction has been performed. A genome-wide meta-analysis of genotype-dependent phenotypic variance of BMI, a marker of sensitivity to environmental exposures, in ~170,000 participants identified FTO, but did not show robust evidence of environmental sensitivity for other loci. Recent genome-wide meta-analyses of adiposity traits in >320,000 individuals uncovered loci interacting with age and sex, but also suggested that very large sample sizes are required for interaction studies to be successful.

3Here, we report results from a large-scale genome-wide meta-analysis of SNP×PA interactions in adiposity in up to 200,452 adults. As part of these interaction analyses, we also examine whether adjusting for PA or jointly testing for SNP’s main effect and interaction with PA may identify novel adiposity loci.

4We performed meta-analyses of results from 60 studies, including up to 180,423 adults of European descent and 20,029 adults of other ancestries to assess interactions between ~2.5 million genotyped or HapMap-imputed SNPs and PA on BMI and BMI-adjusted waist circumference (WCadjBMI) and waist-hip ratio (WHRadjBMI) (S1–S5 Tables). Similar to a previous meta-analysis of the interaction between FTO and PA, we standardized PA by categorizing it into a dichotomous variable where on average ~23% of participants were categorized as inactive and ~77% as physically active (see Methods and S6 Table). On average, inactive individuals had 0.99 kg/m2 higher BMI, 3.46 cm higher WC, and 0.018 higher WHR than active individuals (S4 and S5 Tables).

5Each study first performed genome-wide association analyses for each SNP’s effect on BMI in the inactive and active groups separately. Corresponding summary statistics from each cohort were subsequently meta-analyzed, and the SNP×PA interaction effect was estimated by calculating the difference in the SNP’s effect between the inactive and active groups. To identify sex-specific SNP×PA interactions, we performed the meta-analyses separately in men and women, as well as in the combined sample. In addition, we carried out meta-analyses in European-ancestry studies only and in European and other-ancestry studies combined.

6We used two approaches to identify loci whose effects are modified by PA. In the first approach, we searched for genome-wide significant SNP×PA interaction effects (PINT<5x10-8). As shown in Fig 1, this approach yielded the highest power to identify cross-over interaction effects where the SNP’s effect is directionally opposite between the inactive and active groups. However, this approach has low power to identify interaction effects where the SNP’s effect is directionally concordant between the inactive and active groups (Fig 1). We identified a genome-wide significant interaction between rs986732 in cadherin 12 (CDH12) and PA on BMI in European-ancestry studies (betaINT = -0.076 SD/allele, PINT = 3.1x10-8, n = 134,767) (S7 Table). The interaction effect was directionally consistent but did not replicate in an independent sample of 31,097 individuals (betaINT = -0.019 SD/allele, PINT = 0.52), and the pooled association P value for the discovery and replication stages combined did not reach genome-wide significance (NTOTAL = 165,864; PINT-TOTAL = 3x10-7) (S1 Fig). No loci showed genome-wide significant interactions with PA on WCadjBMI or WHRadjBMI. CDH12 encodes an integral membrane protein mediating calcium-dependent cell-cell adhesion in the brain, where it may play a role in neurogenesis. While CDH12 rs4701252 and rs268972 SNPs have shown suggestive associations with waist circumference (P = 2x10-6) and BMI (P = 5x10-5) in previous GWAS, the SNPs are not in LD with rs986732 (r2<0.1).

7In our second approach, we tested interaction for loci showing a genome-wide significant main effect on BMI, WCadjBMI or WHRadjBMI (S7–S12 Tables). We adjusted the significance threshold for SNP×PA interaction by Bonferroni correction (P = 0.05/number of SNPs tested). As shown in Fig 1, this approach enhanced our power to identify interaction effects where there is a difference in the magnitude of the SNP’s effect between inactive and active groups when the SNP’s effect is directionally concordant between the groups. We identified a significant SNP×PA interaction of the FTO rs9941349 SNP on BMI in the meta-analysis of European-ancestry individuals; the BMI-increasing effect was 33% smaller in active individuals (betaACTIVE = 0.072 SD/allele) than in inactive individuals (betaINACTIVE = 0.106 SD/allele, PINT = 4x10-5). The rs9941349 SNP is in strong LD (r2 = 0.87) with FTO rs9939609 for which interaction with PA has been previously established in a meta-analysis of 218,166 adults. We identified no loci interacting with PA for WCadjBMI or WHRadjBMI.

8In a previously published meta-analysis, the FTO locus showed a geographic difference for the interaction effect where the interaction was more pronounced in studies from North America than in those from Europe. To test for geographic differences in the present study, we performed additional meta-analyses for the FTO rs9941349 SNP, stratified by geographic origin (North America vs. Europe). While the interaction effect was more pronounced in studies from North America (betaINT = 0.052 SD/allele, P = 5x10-4, N = 63,896) than in those from Europe (betaINT = 0.028 SD/allele, P = 0.006, N = 109,806), we did not find a statistically significant difference between the regions (P = 0.14).

9We tested whether the variance explained by ~1.1 million common variants (MAF≥1%) differed between the inactive and active groups for BMI, WCadjBMI, and WHRadjBMI. In the physically active individuals, the variants explained ~20% less of variance in BMI than in inactive individuals (12.4% vs. 15.7%, respectively; Pdifference = 0.046), suggesting that PA may reduce the impact of genetic predisposition to adiposity overall. There was no significant difference in the variance explained between active and inactive groups for WCadjBMI (8.6% for active, 9.3% for inactive; Pdifference = 0.70) or WHRadjBMI (6.9% for active, 8.0% for inactive; Pdifference = 0.59).

10To further investigate differences in explained variance between the inactive and active groups, we calculated variance explained by subsets of SNPs selected based on significance thresholds (ranging from P = 5x10-8 to P = 0.05) of PA-adjusted SNP association with BMI, WCadjBMI or WHRadjBMI (S13 Table). We found 17–26% smaller explained variance for BMI in the active group than in the inactive group at all P value thresholds (S13 Table).

11Physical activity contributes to variation in BMI, WCadjBMI, and WHRadjBMI, hence, adjusting for PA as a covariate may enhance power to identify novel adiposity loci. To that extent, each study performed genome-wide analyses for association with BMI, WCadjBMI, and WHRadjBMI while adjusting for PA. Subsequently, we performed meta-analyses of the study-specific results. We discovered 10 genome-wide significant loci (2 for BMI, 1 for WCadjBMI, 7 for WHRadjBMI) that have not been reported in previous GWAS of adiposity traits (Table 1, S2–S4 Figs).

12To establish whether additionally accounting for SNP×PA interactions would identify novel loci, we calculated the joint significance of PA-adjusted SNP main effect and SNP×PA interaction using the method of Aschard et al. As illustrated in Fig 1, the joint test enhanced our power to identify loci where the SNP shows simultaneously a main effect and an interaction effect. We identified a novel BMI locus near ELAVL2 in men (PJOINT = 4x10-8), which also showed suggestive evidence of interaction with PA (PINT = 9x10-4); the effect of the BMI-increasing allele was attenuated by 71% in active as compared to inactive individuals (betaINACTIVE = 0.087 SD/allele, betaACTIVE = 0.025 SD/allele) (Table 1, S2–S4 Figs).

13To evaluate the effect of PA adjustment on the results for the 11 novel loci, we performed a look-up in published GIANT consortium meta-analyses for BMI, WCadjBMI, and WHRadjBMI that did not adjust for PA (S22 Table). All 11 loci showed a consistent direction of effect between the present PA-adjusted and the previously published PA-unadjusted results, but the PA-unadjusted associations were less pronounced despite up to 40% greater sample size, suggesting that adjustment for PA may have increased our power to identify these loci.

14The biological relevance of putative candidate genes in the novel loci, based on our thorough searches of the literature, GWAS catalog look-ups, and analyses of eQTL enrichment and overlap with functional regulatory elements, are described in Tables 2 and 3. As the novel loci were identified in a PA-adjusted model, where adjusting for PA may have contributed to their identification, we examined whether the lead SNPs in these loci are associated with the level of PA. More specifically, we performed look-ups in GWAS analyses for the levels of moderate-to-vigorous intensity leisure-time PA (n = 80,035), TV-viewing time (n = 28,752), and sedentary behavior at work (n = 59,381) or during transportation (n = 15,152) [personal communication with Marcel den Hoed, Marilyn Cornelis, and Ruth Loos]. However, we did not find significant associations when correcting for the number of loci that were examined (P>0.005) (S16 Table).

15In addition to uncovering 11 novel adiposity loci, our PA-adjusted GWAS and the joint test of SNP main effect and SNP×PA interaction confirmed 148 genome-wide significant loci (50 for BMI, 58 for WCadjBMI, 40 for WHRadjBMI) that have been established in previous main effect GWAS for adiposity traits (S7–S12 Tables, S4 Fig). The lead SNPs in eight of the previously established loci (5 for BMI, 3 for WCadjBMI), however, showed no LD or only weak LD (r2<0.3) with the published lead SNP, suggesting they could represent novel secondary signals in known loci (S17 Table). To test whether these eight signals are independent of the previously published signals, we performed conditional analyses. Three of the eight SNPs we examined, in/near NDUFS4, MEF2C-AS1 and CPA1, were associated with WCadjBMI with P<5x10-8 in our PA-adjusted GWAS even after conditioning on the published lead SNP, hence representing novel secondary signals in these loci (S17 Table).

16Epigenetic variation may underlie gene-environment interactions observed in epidemiological studies and PA has been shown to induce marked epigenetic changes in the genome. We examined whether the BMI or WHRadjBMI loci reaching P<1x10-5 for interaction with PA (13 loci for BMI, 5 for WHRadjBMI) show overall enrichment with chromatin states in adipose, brain and muscle tissues available from the Roadmap Epigenomics Consortium. However, we did not find significant enrichment (S18 and S19 Tables), which may be due to the limited number of identified loci. The lack of significant findings may also be due to the assessment of chromatin states in the basal state, which may not reflect the dynamic changes that occur when cells are perturbed by PA.

17We also tested whether the loci reaching P<5x10-8 in our PA-adjusted GWAS of BMI or WHRadjBMI show enrichment with chromatin states and found significant enrichment of the BMI loci with enhancer, weak transcription, and polycomb-repressive elements in several brain cell lines, and with enhancer elements in three muscle cell lines (S20 and S21 Tables). We also found significant enrichment of the WHRadjBMI loci with enhancer elements in three adipose and six muscle cell lines, with active transcription start sites in two adipose cell lines, and with polycomb-repressive elements in seven brain cell lines. The enrichment of our PA-adjusted main effect results with chromatin annotations in skeletal muscle in particular, the tissue most affected by PA, could highlight regulatory mechanisms that may be influenced by PA.

18In this genome-wide meta-analysis of more than 200,000 adults, we do not find evidence of interaction with PA for loci other than the established FTO locus. However, when adjusting for PA or jointly testing for SNP main effect and interaction with PA, we identify 11 novel adiposity loci, suggesting that accounting for PA or other environmental factors that contribute to variation in adiposity may increase power for gene discovery.

19Our results suggest that if SNP×PA interaction effects for common variants exist, they are unlikely to be of greater magnitude than observed for FTO, the BMI-increasing effect of which is attenuated by ~30% in physically active individuals. The fact that common SNPs explain less of the BMI variance among physically active compared to inactive individuals indicates that further interactions may exist, but larger meta-analyses, more accurate and precise measurement of PA, and/or improved analytical methods will be required to identify them. We found no difference between inactive and active individuals in variance explained by common SNPs in aggregate for WCadjBMI or WHRadjBMI, and no loci interacted with PA on WCadjBMI or WHRadjBMI. Therefore, PA may not modify genetic influences as strongly for body fat distribution as for overall adiposity. Furthermore, while differences in variance explained by common variants may be due to genetic effects being modified by PA, it is important to note that heritability can change in the absence of changes in genetic effects, if environmental variation differs between the inactive and active groups. Therefore, the lower BMI variance explained in the active group could be partly due to a potentially greater environmental variation in this group.

20While we replicated the previously observed interaction between FTO and PA, it remains unclear what biological mechanisms underlie the attenuation in FTO’s effect in physically active individuals, and whether the interaction is due to PA or due to confounding by other environmental exposures. While some studies suggest that FTO may interact with diet, a recent meta-analysis of 177,330 individuals did not find interaction between FTO and dietary intakes of total energy, protein, carbohydrate or fat. The obesity-associated FTO variants are located in a super-enhancer region and have been associated with DNA methylation levels, suggesting that this region may be sensitive to epigenetic effects that could mediate the interaction between FTO and PA.

21In genome-wide analyses for SNP main effects adjusting for PA, or when testing for the joint significance of SNP main effect and SNPxPA interaction, we identify 11 novel adiposity loci, even though our sample size was up to 40% smaller than in the largest published main effect meta-analyses. Our findings suggest that accounting for PA may facilitate the discovery of novel adiposity loci. Similarly, accounting for other environmental factors that contribute to variation in adiposity could lead to the discovery of additional loci.

22In the present meta-analyses, statistical power to identify SNPxPA interactions may have been limited due to challenges relating to the measurement and statistical modeling of PA. Of the 60 participating studies, 56 assessed PA by self-report while 4 used wearable PA monitors. Measurement error and bias inherent in self-report estimates of PA can attenuate effect sizes for SNP×PA interaction effects towards the null. Measurement using PA monitors provides more consistent results, but the monitors are not able to cover all types of activities and the measurement covers a limited time span compared to questionnaires. As sample size requirements increase nonlinearly when effect sizes decrease, any factor that leads to a deflation in the observed interaction effect estimates may make their detection very difficult, even when very large population samples are available for analysis. Finally, because of the wide differences in PA assessment tools used among the participating studies, we treated PA as a dichotomous variable, harmonizing PA into inactive and active individuals. Considerable loss of power is anticipated when a continuous PA variable is dichotomized. Our power could be enhanced by using a continuous PA variable if a few larger studies with equivalent, quantitative PA measurements were available.

23In summary, while our results suggest that adjusting for PA or other environmental factors that contribute to variation in adiposity may increase power for gene discovery, we do not find evidence of SNP×PA interaction effects stronger than that observed for FTO. While other SNP×PA interaction effects on adiposity are likely to exist, combining many small studies with varying characteristics and PA assessment tools may be inefficient for identifying such effects. Access to large cohorts with quantitative, equivalent PA variables, measured with relatively high accuracy and precision, may be necessary to uncover novel SNP×PA interactions.

24All studies were conducted according to the Declaration of Helsinki. The studies were approved by the local ethical review boards and all study participants provided written informed consent for the collection of samples and subsequent analyses.

25We examined three anthropometric traits related to overall adiposity (BMI) or body fat distribution (WCadjBMI and WHRadjBMI) that were available from a large number of studies. Before the association analyses, we calculated sex-specific residuals by adjusting for age, age2, BMI (for WCadjBMI and WHRadjBMI traits only), and other necessary study-specific covariates, such as genotype-derived principal components. Subsequently, we normalized the distributions of sex-specific trait residuals using inverse normal transformation.

26Physical activity was assessed and quantified in various ways in the participating studies of the meta-analysis (S1 and S6 Tables). Aiming to amass as large a sample size as possible, we harmonized PA by categorizing it into a simple dichotomous variable—physically inactive vs. active—that could be derived in a relatively consistent way in all participating studies, and that would be consistent with previous findings on gene-physical activity interactions and the relationship between activity levels and health outcomes. In studies with categorical PA data, individuals were defined inactive if they reported having a sedentary occupation and being sedentary during transport and leisure-time (<1 h of moderate intensity leisure-time or commuting PA per week). All other individuals were defined physically active. Previous studies in large-scale individual cohorts have demonstrated that the interaction between FTO, or a BMI-increasing genetic risk score, with physical activity, is most pronounced approximately at this activity level. In studies with continuous PA data, PA variables were standardized by defining individuals belonging to the lowest sex- and age-adjusted quintile of PA levels as inactive, and all other individuals as active. The study-specific coding of the dichotomous PA variable in each study is described in S6 Table.

27We included 42 studies with genome-wide data, 10 studies with Metabochip data, and eight studies with both genome-wide and Metabochip data. If both genome-wide and Metabochip data were available for the same individual, we only included the genome-wide data (S1 Table). Studies with genome-wide genotyped data used either Affymetrix or Illumina arrays (S2 Table). Following study-specific quality control measures, the genotype data were imputed using the HapMap phase II reference panel (S2 Table). Studies with Metabochip data used the custom Illumina HumanCardio-Metabo BeadChip containing ~195K SNPs designed to support large-scale follow-up of known associations with metabolic and cardiovascular traits. Each study ran autosomal SNP association analyses with BMI, WCadjBMI and WHRadjBMI across their array of genetic data using the following linear regression models in men and women separately: 1) active individuals only; 2) inactive individuals only; and 3) active and inactive individuals combined, adjusting for the PA stratum. In studies that included families or closely related individuals, regression coefficients were estimated using a variance component model that modeled relatedness in men and women combined, with sex as a covariate, in addition to the sex-specific analyses. The additive genetic effect for each SNP and phenotype association was estimated using linear regression. For studies with a case-control design (S1 Table), cases and controls were analyzed separately.

28All study-specific files for the three regression models listed above were processed through a standardized quality control protocol using the EasyQC software. The study-specific quality control measures included checks on file completeness, range of test statistics, allele frequencies, trait transformation, population stratification, and filtering out of low quality data. Checks on file completeness included screening for missing alleles, effect estimates, allele frequencies, and other missing data. Checks on range of test statistics included screening for invalid statistics such as P-values >1 or <0, negative standard errors, or SNPs with low minor allele count (MAC, calculated as MAF*N, where MAF is the minor allele frequency and N is the sample size) and where SNPs with MAC<5 in the inactive or the active group were removed. The correctness of trait transformation to inverse normal was examined by plotting 2/median of the standard error with the square root of the sample size. Population stratification was examined by calculating the study specific genomic control inflation factor (λGC). If a study had λGC>1.1, the study analyst was contacted and asked to revise the analyses by adjusting for principal components. The allele frequencies in each study were examined for strand issues and miscoded alleles by plotting effect allele frequencies against the corresponding allele frequencies from the HapMap2 reference panel. Finally, low quality data were filtered out by removing monomorphic SNPs, imputed SNPs with poor imputation quality (r2_hat <0.3 in MACH, observed/expected dosage variance <0.3 in BIMBAM, proper_info <0.4 in IMPUTE), and genotyped SNPs with a low call-rate (<95%) or that were out of Hardy-Weinberg equilibrium (P<10−6).

29Beta-coefficients and standard errors were combined by an inverse-variance weighted fixed effect method, implemented using the METAL software. We performed meta-analyses for each of the three models (active, inactive, active + inactive adjusted for PA) in men only, in women only, and in men and women combined. Study-specific GWAS results were corrected for genomic control using all SNPs. Study-specific Metabochip results as well as the meta-analysis results for GWAS and Metabochip combined were corrected for genomic control using 4,425 SNPs included on the Metabochip for replication of associations with QT-interval, a phenotype not correlated with BMI, WCadjBMI or WHRadjBMI, after pruning of SNPs within 500 kb of an anthropometry replication SNP. We excluded SNPs that 1) were not available in at least half of the maximum sample size in each stratum; 2) had a heterogeneity I2 >75%, or 3) were missing chromosomal and base position annotation in dbSNP.

30To identify SNP×PA interactions, we used the EasyStrata R package to test for the difference in meta-analyzed beta-coefficients between the active and inactive groups for the association of each SNP with BMI, WCadjBMI and WHRadjBMI. Easystrata tests for differences in effect estimates between the active and inactive strata by subtracting one beta from the other (βactive−βinactive,) and dividing by the overall standard error of the difference as follows: where r is the Spearman rank correlation coefficient between βactive and βinactive for all genome-wide SNPs. The joint significance of the SNP main and SNP×PA interaction effects was estimated using the method by Aschard et al. which is a joint test for genetic main effects and gene-environment interaction effects where gene-environment interaction is calculated as the difference in effect estimates between two exposure strata, accounting for 2 degrees of freedom.

31Approximate conditional analyses were conducted using GCTA version 1.24. In the analyses for SNPs identified in our meta-analyses of European-ancestry individuals only, LD correlations between SNPs were estimated using a reference sample comprised of European-ancestry participants of the Atherosclerosis Risk in Communities (ARIC) study. In the analyses for SNPs identified in our meta-analyses of all ancestries combined, the reference sample comprised 93% of European-ancestry individuals and 6% of African ancestry participants from ARIC, as well as 1% of CHB and JPT samples from the HapMap2 panel, to approximate the ancestry mixture in our all ancestry meta-analyses. To test if our identified SNPs were independent secondary signals that fell within 1 Mbp of a previously established signal, we used the GCTA—cojo-cond command to condition our lead SNPs on each previously established SNP in the same locus.

32The replication analysis for the CDH12 locus included participants from the EPIC-Norfolk (NINACTIVE = 4,755, NACTIVE = 11,526) and Fenland studies (NINACTIVE = 1,213, NACTIVE = 4,817), and from the random subcohort of the EPIC-InterAct Consortium (NINACTIVE = 2,154, NACTIVE = 6,632). PA stratum-specific estimates of the association of CDH12 with BMI were assessed and meta-analyzed by fixed effects meta-analyses, and the differences between the PA-strata were determined as described above.

33To identify associations of the novel BMI, WCadjBMI or WHRadjBMI loci with other complex traits in published GWAS, we extracted previously reported GWAS associations within 500 kb and r2>0.6 with any of the lead SNPs, from the GWAS Catalog of the National Human Genome Research Institute and European Bioinformatics Institute (S14 Table).

34We examined the cis-associations of the novel BMI, WCadjBMI or WHRadjBMI loci with the expression of nearby genes from various tissues by performing a look-up in a library of >100 published expression datasets, as described previously by Zhang et al. In addition, we examined cis-associations using gene expression data derived from fasting peripheral whole blood in the Framingham Heart Study (n = 5,206), adjusting for PA, age, age2, sex and cohort. For each novel locus, we evaluated the association of all transcripts ±1 Mb from the lead SNP. To minimize the potential for false positives, we only considered associations where our lead SNP or its proxy (r2>0.8) was either the peak SNP associated with the expression of a gene transcript in the region, or in strong LD (r2>0.8) with the peak SNP.

35We used the Uncovering Enrichment Through Simulation method to combine the genetic association data with the Roadmap Epigenomics Project segmentation data. First, 10,000 sets of random SNPs were selected among HapMap2 SNPs with a MAF >0.05 that matched the original input SNPs based on proximity to a transcription start site and the number of LD partners (r2>0.8 in individuals of European ancestry in the 1000 Genomes Project). The LD partners were combined with their original lead SNPs to create 10,000 sets of matched random SNPs and their respective LD partners. These sets were intersected with the 15-state ChromHMM data from the Roadmap Epigenomics Project and resultant co-localizations were collapsed from total SNPs down to loci, which were then used to calculate an empirical P value when comparing the original SNPs to the random sets. We examined the enrichment for all loci reaching P<10−5 for SNP×PA interaction combined, and for all loci reaching P<5x10-8 in the PA-adjusted SNP main effect model combined. In addition, we examined the variant-specific overlap with regulatory elements for each of the index SNPs of the novel BMI, WCadjBMI and WHRadjBMI loci and variants in strong LD (r2>0.8).

36We compared variance explained for BMI, WCadjBMI and WHRadjBMI between the active and inactive groups using two approaches. First, we used a method previously reported by Kutalik et al, and selected subsets of SNPs based on varying P value thresholds (ranging from 5x10-8 to 0.05) from the SNP main effect model adjusted for PA. Each subset of SNPs was clumped into independent regions using a physical distance criterion of <500kb, and the most significant lead SNP within the respective region was selected. For each lead SNP, the explained variance was calculated as: in the active and inactive groups separately, where N is the sample size and P is the P value for SNP main effect in active or inactive strata. Finally, the variance explained by each subset of SNPs in the active and inactive strata was estimated by summing up the variance explained by the SNPs.

37Second, we applied the LD Score regression tool developed by Bulik-Sullivan et al to quantify the proportion of inflation due to polygenicity (heritability) rather than confounding (cryptic relatedness or population stratification) using meta-analysis summary results. LD Score regression leverages LD between causal and index variants to distinguish true signals by regressing meta-analysis summary results on an ‘LD Score’, i.e. the cumulative genetic variation that an index SNP tags. To obtain heritability estimates by PA strata, we regressed our summary results from the genome-wide meta-analyses of BMI, WCadjBMI and WHRadjBMI, stratified by PA status (active and inactive), on pre-calculated LD Scores available in HapMap3 reference samples of up to 1,061,094 variants with MAF≥1% and N>10th percentile of the total sample size.