Results

Cohort Characteristics
After sample QC, 1,366 samples from the LBCs (n = 446 from LBC1921 and n = 920 from LBC1936), 752 samples from the LifeLines DEEP cohort, and 403 samples from the BSGS cohort (after we removed one individual from each MZ twin pair) had methylation, phenotype, and genotype data. Cohort characteristics of these samples are provided in Table 2. The LifeLines DEEP participants had a much wider age range (18–81 years) and were on average much younger (mean 45.5 ± 13.3 years) than LBC participants (69.5 ± 0.8 years in LBC1936 and 79.1 ± 0.6 years in LBC1921). The mean age in the BSGS cohort was 14 ± 2.4 years. BMI and height distributions for each cohort are shown in Figures S1 and S2. The BMI and height phenotypes were adjusted for age and sex in each cohort.

Methylome-wide Association Analysis
To create a multi-probe methylation predictor, we first conducted a methylome-wide association analysis. A total of 431,951 and 407,935 CpG probes remained in the LBC and LifeLines DEEP datasets, respectively, after QC and probe filtering. Probes with an association p value < 1.16 × 10−7 in the LBC dataset and a p value < 1.22 × 10−7 in the LifeLines DEEP dataset were considered to be significantly associated after Bonferroni correction for the number of probes tested. After removal of correlated probes, nine CpG probes in the LBC dataset and five probes in the LifeLines DEEP dataset were associated with BMI and were used for generating methylation-profile scores (Table S1). Two probes (cg06500161 and cg11024682) were significantly associated with BMI in both cohorts—cg06500161 is found in an intronic region of ABCG1 (ATP-binding cassette, sub-family G, member 1 [MIM: 603076]), and cg11024682 is intronic to one isoform of SREBF1 (sterol regulatory element binding transcription factor 1 [MIM: 184756]). Both genes are known to be involved in lipid metabolism, but neither has been identified by GWASs to harbor genetic variants that are associated with BMI.
For height, no CpG probes passed the p value threshold in the LBCs, whereas only a single probe passed the threshold in the LifeLines DEEP cohort. Therefore, to generate a height-profile score, we used a less stringent association p value of <0.001 for probe selection. 507 and 949 CpG probes were selected in the LBCs and LifeLines DEEP cohort, respectively. Quantile-quantile plots for each MWAS are shown in Figure S3. We observed inflation in the lambda values—for BMI, lambdas were 1.53 and 1.17 in the LBCs and LifeLines DEEP cohort, respectively, whereas for height, lambdas were 1.12 and 1.36, respectively. Lambdas close to 1 (SD = 0.1) were observed with permutation analysis (performed in both the LBCs and LifeLines DEEP cohort), which indicates that the inflation was due to real signal and not an artifact of our assumption of the null distribution of the test statistic.

Proportion of BMI and Height Variance Explained by Profile Scores in the LBCs and LifeLines DEEP Cohort
Consistent with expectation, all methylation- and genetic-profile scores were correlated with their respective traits in the anticipated direction (Table S2). The methylation-profile scores explained 6.9% and 4.9% (p value < 1 × 10−15 and 7 × 10−10, respectively) of the variation in BMI in the LBCs and LifeLines DEEP cohort, respectively, whereas the genetic-profile scores explained 8.0% and 9.4% (p value < 1 × 10−15), respectively (Figure 1). When both the methylation- and genetic-profile scores were included in an additive model for BMI, each remained independently associated with BMI. The proportion of variance explained by the additive model was 14.0% and 13.6% in the LBCs and LifeLines DEEP cohort, respectively, suggesting a mainly additive effect of the two scores on BMI (Figure 1).
The BMI methylation-profile scores, based on 78 probes selected from an MWAS in the larger Framingham Heart Study (M.M.M., unpublished data) but weighted with effect sizes estimated in the LBCs, explained 7.3% of the variation in BMI in the LifeLines DEEP cohort, whereas a profile score based on the effects estimated in the LifeLines DEEP cohort explained 11% of the variation in the LBCs. As before, the methylation-profile scores showed an additive effect with the genetic-profile scores (Figure S4). Compared to the methylation-profile scores derived from the MWAS in the LBCs or LifeLines DEEP cohort, the larger R2 values for the profile scores based on probes identified in the Framingham cohort suggest that the larger sample size in the latter study provided more power to identify additional CpG probes and hence allowed us to explain a higher proportion of variance in BMI.
The height methylation-profile scores were associated with height and explained 0.31% and 0.76% (p value = 0.02 and 0.01 of the variation in the LBCs and LifeLines DEEP cohort, respectively). The height genetic-profile scores explained 18.5% and 19.8% (p value < 1 × 10−15) of the inter-individual variation in the height phenotype in the LBCs and LifeLines DEEP cohort, respectively (Figure 1). The additive model including both methylation- and genetic-profile scores explained 18.5% and 20.1% of the variation in the height phenotype in the LBCs and LifeLines DEEP cohort, respectively. However, the methylation-profile score showed no independent association in the LBCs (p = 0.16) and remained only marginally associated (p = 0.035) with the height phenotype independently of the genetic-profile score in the LifeLines DEEP cohort.
For BMI, the interaction model explained a slightly larger proportion of variance than did the additive model in the LBCs (15% versus 14%; ANOVA p value = 5 × 10−6) but not in the LifeLines DEEP cohort (Table S3). There was no significant interaction between the genetic- and methylation-profile scores for height in either cohort.

Proportion of BMI Variance Explained in BSGS Adolescent Individuals
The methylation-profile scores derived from the MWAS analysis in the LBC individuals did not explain any variation (adjusted R2 = −0.001) in the sex- and age-adjusted BMI phenotype from the BSGS cohort, whereas that derived from the mostly middle-aged individuals of the LifeLines DEEP study explained 3.6% (p value = 8 × 10−5; Figure 2). Methylation scores based on the CpG probes identified in the larger Framingham MWAS but weighted with effect sizes from the older LBC individuals explained 3.0% of the variation in BMI in adolescent individuals. Based on the same CpG probes but effect sizes derived from the younger, albeit smaller, LifeLines DEEP cohort, the methylation-profile scores explained almost twice (5.4%) the variation in BMI in adolescent individuals (Figure 2).
Given that the proportion of variance explained in a prediction setting is a function of sample sizes of the discovery cohorts, the R2 values from different-sized cohorts are not directly comparable. We therefore compared the ratio of the methylation score R2 to the genetic score R2 to look at the relative contribution of the methylation- and genetic-profile scores to variance in BMI in both BSGS adolescents and older cohorts. As shown in Table S4, in all cases, the methylation-profile scores had a lower contribution to BMI variance in the BSGS cohort than in the other cohorts. The methylation predictor derived from older individuals (probes and weights for the methylation-profile score derived from the LBC MWAS) performed the worst.
A BMI methylation score based on a fixed-effect meta-analysis of the LBC and LifeLines DEEP MWAS results, whereby a Bonferroni correction for 374,629 common probes in the two cohorts (p value < 1.33 × 10−7) was used for selecting probes, performed better than the methylation score based on the LBC MWAS. However, despite the larger sample size, it performed worse than the predictor based on the LifeLines DEEP MWAS: its adjusted R2 was 0.028 (p value = 4.0 × 10−4).

Correcting for Cell Count
In the LBCs, all cell counts except neutrophils were associated with sex- and age-adjusted BMI (p < 0.05), but only monocytes were associated with sex- and age-adjusted height. In contrast, in the LifeLines DEEP cohort, all cell counts were significantly associated with BMI, but not with height. Adjusting for cell count reduced some of the inflation observed in the uncorrected analysis—for BMI, lambdas were 1.28 and 1.00 in the LBCs and LifeLines DEEP cohort, respectively, whereas for height, lambdas were 1.00 and 1.15, respectively. The proportion of variance explained by the methylation scores after cell-count correction is shown in Figure S5. The cell-count-corrected methylation scores based on the MWAS discovery in the LBCs and LifeLines DEEP cohort remained significantly associated with BMI and showed an additive effect, although the proportion of variance explained was substantially less in the LBCs (3.2%). For height, the methylation-profile score was still marginally associated with the sex- and age-adjusted height phenotype in the LifeLines DEEP cohort (adjusted R2 = 0.0041; p value = 0.045), but not in the LBCs.