Longitudinal multi-omics study reveals common etiology underlying association between plasma proteome and BMI trajectories in adolescent and young adult twins

FinnTwin12 is a longitudinal cohort based on a population of Finnish twins born between 1983 and 1987 aimed at investigating adolescent behavioral development and health habits [30, 31]. Participants were identified from the Central Finnish Population Register and completed four questionnaires (response rate range: 85–90%) at approximate ages 12, 14, 17, and 22. A subset of twins, referred to as intensive subset, was studied more intensively, with additional psychiatric interviews and questionnaires starting at the wave corresponding to age 14. Besides health, lifestyle, and psychological indicators, questionnaires collected up to four self-reported weight and height measures for these participants across the four waves, from which BMI values (kg.m⁻²) were calculated. At the 22-year assessment wave, 786 twins from the intensive subset participated in in-person assessments and provided venous blood plasma samples after overnight fasting. We assessed the validity of self-reports of weight and height at the 22-year assessment wave versus in-person measurements in a sample of 756 of these participants (Additional file 2). Self-reported and measured values correlated strongly for weight (Pearson correlation: r = 0.98) and height (r = 0.99). From the blood samples, extensive omic data were derived, including proteomics, metabolomics, and genotyping to generate the PRS data used in the current study. Of these participants, 651 had complete longitudinal anthropometric measurements over the 10 years of follow-up and constituted the final FinnTwin12 sample (Table 1). The sample included 250 monozygotic twins (106 complete pairs) and 401 dizygotic twins (164 complete pairs). Height and weight across the waves can be found in the supplementary material (Additional file 3: Table S1).

The Netherlands Twin Register (NTR) is a population-based cohort which includes twins and multiples from the Netherlands [32]. Adult participants of the NTR have completed a survey on health and behavior every 2–3 years since 1991. In contrast to the FinnTwin12 cohort, where data collection is based on age, the NTR data collection in adult twins is independent of age; adult participants are invited to fill out a survey every ~ 3 years. The blood samples for the omics experiments described in this study were collected in participants of the NTR biobank project (2004–2008) [33]. Venous blood samples were drawn between 7 and 10 AM after overnight fasting, and fertile women had their blood sample drawn in their pill-free week or on day 2–4 of their menstrual cycle. Information on BMI was obtained during the home visit for blood collection as part of the NTR Biobank project. Additionally, we calculated BMI from self-reported height and weight from NTR Survey 5 (2000), Survey 6 (2002), and Survey 7 (2004).

Array-based transcriptomics data were available for 3369 individuals, of which we selected 965 participants for whom blood sampling was performed before age 30 (i.e., age at blood sampling < 30 years old). Of these participants, we retained those for whom BMI was available at the time of blood sampling as well as at least one other BMI measure at Survey 5, Survey 6, or Survey 7. The final NTR sample comprised 665 participants with 2 to 4 BMI measurements over a mean follow-up period of ~ 6 years (Table 1). The sample included 359 monozygotic twins (141 complete pairs) and 306 dizygotic twins (103 complete pairs).

Proteins from the plasma samples of FinnTwin12 participants were subjected to Liquid Chromatography-Electrospray Ionization-Mass Spectrometry (LC–ESI–MS/MS) as described previously [34]. First, proteins were precipitated with acetone and subjected to in-solution digestion according to the standard protocol of the Turku Proteomics Facility (Turku Proteomics Facility, Turku, Finland). After digestion, peptides were desalted using a 96-well Sep-Pak C18 plate (Waters), evaporated to dryness, and stored at − 20 °C. A commercial kit (High Select™ Top14 Abundant Protein Depletion Mini Spin Columns, cat. Number: A36370, Thermo Scientific) was used to deplete the 14 most abundant proteins from plasma before the proteomic analysis. Samples were first analyzed by independent data acquisition LC–MS/MS using a Q Exactive HF mass spectrometer. Data were further analyzed using Spectronaut software and included local normalization of the data [35], as described elsewhere [34]. Raw matrix counts were log₂-transformed and kit-depleted proteins were removed, including human serum albumin (HSA), albumin, IgG, IgA, IgM, IgD, IgE, kappa and lambda light chains, alpha-1 acid glycoprotein, alpha-1 antitrypsin, alpha-2 macroglobulin, apolipoprotein A1, fibrinogen, haptoglobin, and transferrin. None of the participants had scores greater or less than 5 standard deviations from the mean on the first two principal components (PCs) derived from principal component analysis (PCA), indicating none of the participants were identified as outliers. We excluded proteins with > 10% missing values. Following the identification of 4 batches, imputation of missing values was performed (proportion of missing values in dataset: 0.86%) using the lowest observed value per batch. Corrections for batch effects were performed with Combat [36]. The final proteomic dataset comprised 439 proteins and protein abundances were scaled, such that one unit corresponded to one standard deviation (sd) with zero mean.

Details of preprocessing can be found in Jansen et al. [37]. In short, heparinized whole blood was transferred into PAXgene Blood RNA tubes (Qiagen, Valencia, Florida, USA) within 20 min of sampling and stored at − 30 °C. Total RNA was extracted using the PAXgene Blood RNA MDx kit protocol in 96-well format with the BioRobot Universal System (Qiagen, Valencia, Florida, USA). RNA quality and quantity was assessed by Caliper AMS90 with HT DNA5K/RNA LabChips [37‐39].

Samples were randomly assigned to plates and co-twins were randomized across plates to avoid bias in family correlation estimates. For cDNA synthesis, 50 ng of RNA was reverse transcribed and amplified in a plate format on a Biomek FX liquid handling robot (Beckman Coulter, Brea, California, USA) using Ovation Pico WTA reagents per the manufacturer’s protocol (NuGEN, San Carlos, California, USA). Products purified from single-primer isothermal amplification were then fragmented and labeled with biotin using Encore Biotin Module (NuGEN). Prior to hybridization, the labeled cDNA was analyzed using electrophoresis to verify the appropriate size distribution (Caliper AMS90, HT DNA 5 K/RNA LabChip). Samples were hybridized to Affymetrix U219 array plates, and array hybridization, washing, staining, and scanning were carried out per the manufacturer’s protocol (GeneTitan, Affymetrix, Santa Clara, California, USA).

Gene expression data were required to pass standard Affymetrix quality control metrics (Affymetrix expression console). Probes were removed when their location was uncertain or if their location intersected a polymorphic single-nucleotide polymorphism (SNP). Expression values were obtained using robust multi-array average normalization implemented in Affymetrix Power Tools (v 1.12.0). We excluded samples that displayed an average Pearson correlation below 0.8 with the probe set expression values of other samples and samples with incorrect sex chromosome expression. In the analyses, we retained only probes for genes encoding proteins associated with BMI or BMI changes in FinnTwin12. We attempted to match protein-coding gene names; however, we failed to retrieve three protein-coding genes (AMY2A, C4A, and CFHR1) identified in FinnTwin12 from the NTR transcriptomic data. We focused on log₂-transformed gene-level expression in our analyses, where we obtained gene-level expression by averaging probe-level expression per gene, for each gene associated with protein-coding genes. The same analyses were also performed for log₂ probe expression as supplementary analyses.

Metabolites were quantified from EDTA plasma samples using high-throughput proton nuclear magnetic resonance spectroscopy (¹H-NMR) on a laboratory setup that combines a Bruker AVANCE III 500 MHz and a Bruker AVANCE III HD 600 MHz spectrometer (Nightingale Health Ltd, Helsinki, Finland) [40, 41]. In short, this method provides simultaneous quantification of routine lipids, lipoprotein subclass profiling with lipid concentrations within 14 subclasses, fatty acid composition, and various low molecular weight metabolites, including amino acids, ketone bodies, and glycolysis-related metabolites in molar concentration units.

Of the 651 participants from the FinnTwin12 sample, 638 had metabolomic data available of which the preprocessing has been described elsewhere [42]. The data included 114 metabolites, of which 13 had missing values. None of the metabolites had more than 10% missing values. Since the distribution of missing values cannot be assumed to be random, we imputed all missing values with the lowest observed value of each metabolic biomarker. Besides LDL cholesterol as derived from the NMR platform, LDL cholesterol was quantified using the Friedewald equation [43]. Similar to the proteomic data, no outliers were detected on the first two PCs. The metabolite values were log₂-transformed and scaled so that one unit corresponded to a change of one sd, with a mean of zero.

In the NTR, apart from six non-overlapping metabolites (Additional file 3: Table S2), we only retained metabolites that were significantly associated with BMI trajectories in FinnTwin12. Description for the NTR metabolomics (NMR platform) has been supplied in detail elsewhere [44]. Missing values represented less than 3% of the total data points, and no metabolite had more than 20% missing values, thus imputation was performed by the lowest observed value for each metabolite. Data were log₂-transformed, and a technical variable indicating batch was used as a covariate in all relevant analyses.

PRSs were calculated and analyzed to test whether genetic susceptibility could explain the association of protein or metabolite levels with BMI trajectories. In FinnTwin12, genotype data for most participants (645 of 651) were available and genotyping was performed using the Illumina Human CNV370-Duo chip. Processing details are available elsewhere [45]. We calculated three PRSs describing the genetic susceptibility to BMI [46], waist to hip ratio adjusted with BMI (WHR) [47], and CAD [48]. After correcting the PRSs for population stratification [49], by regressing out the top ten genetic PCs, the PRSs were scaled to a mean of zero and unit variance.

Of the 665 NTR participants, 576 were genotyped across multiple platforms (Additional file 4) [46, 49‐57]. For genotyped NTR participants, we calculated the PRS of BMI [46], using the infinitesimal prior (LDpred-inf) in Ldpred [57]. Detailed information on genotyping and PRS calculation in NTR participants is included in the supplementary material (Additional file 4). The PRS was first corrected for population stratification [49] and for technical batch (i.e., platform) by regressing out the top ten genetic PCs and platform indicator, respectively, and then the PRS was scaled to have a mean of zero and unit variance.

We assessed the goodness of fit of the linear and logarithmic schemes using the Comparative Fit Index (CFI), the Tucker-Lewis Index (TLI), and the Root Mean Square Error of Approximation (RMSEA). A model with RMSEA less than 0.06, CFI greater than 0.95, and TLI greater than 0.95 was considered a good fit [59, 60]. Whenever the linear modeling did not enable good model fit, the logarithmic scheme was preferred when it did enable a good model fit.

We performed the second sensitivity analysis in NTR. In NTR, surveys were collected regardless of the age of the participants; thus, each wave comprised all participants that were 18 years or older at the time of the survey. Because LGCM corrected for within-survey age differences, we used a second approach that did not correct for wave-level age differences, which were larger in NTR than in FinnTwin12 (Table 1). Instead of considering reference ages by which ages are corrected, as in the main analyses, we took each individual’s actual age at baseline as the first time point and then modeled changes in BMI from this individual’s baseline age. We then quantified the associations between these changes in BMI with gene expression.

To quantify associations between the plasma proteomics data and BMI trajectories, we used mixed-effects models. Mixed-effects modeling involves using two types of effects, fixed and random, the former being constant across individuals and the latter varying between individuals. We tested associations between proteins and BMI at blood sampling and slope of BMI (i.e., BMI changes) in FinnTwin12. We also tested associations between proteins and intercept of BMI (i.e., baseline BMI) as the intercept of BMI was that of young adolescents, whereas the BMI at blood sampling was that of young adults. BMI at blood sampling, changes in BMI, and intercept of BMI were used as dependent variables, and omic variables (e.g., proteins) were used as independent variables. We included age at blood sampling, age squared at blood sampling, sex, and interactions between sex and age and age squared as covariates. BMI intercept was also included as a covariate to correct for baseline BMI differences when testing associations between proteins and changes in BMI. Use of quadratic age and interactions with sex as covariates aimed to ensure adjustment for potential nonlinear relationships with outcomes and combined age–sex confounding on associations, respectively. The model included two random effects whose main function was to correct for clustering in the data due to family relatedness: a binary indicator for zygosity and the family identifier. We tested the association between the coefficients of the fixed effects and each protein using a t-test with Satterthwaite approximation [64] and Bonferroni correction. We considered an association to be significant if the Bonferroni-corrected p-value was less than 0.05. Associations between BMI trajectories and protein levels were therefore significant if the nominal p-value was less than 0.05/439 = 1.1 × 10⁻⁴. The same modeling was used for identification of metabolites associated with BMI trajectories (Bonferroni: nominal p-value < 0.05/114 = 4.4 × 10⁻⁴) as well as to quantify associations between PRSs and proteins (Bonferroni: nominal p-value < 0.05/439 = 1.1 × 10⁻⁴).

We repeated the same modeling in NTR, replacing proteomic variables, being modeled as independent variables, with probe- or gene-level expression of protein-coding genes identified in FinnTwin12, or with metabolites. We added a technical covariate to correct for batches in the modeling between metabolites and BMI trajectories. Multiple correction was applied to the number of tests performed, with the associations of BMI and changes in BMI with omic variables (e.g., metabolite, gene) studied independently. We did not examine the associations between the omics layers and the BMI intercept, as BMI at baseline and BMI at blood sampling reflected both BMI at adulthood. The association between PRS of BMI and gene expression at the probe and gene level was also quantified.

We applied classical twin models (CTMs) to quantify the genetic and environmental contributions to protein abundances, and to decompose the associations between protein and BMI trajectories into genetic and environmental correlations. CTMs hinge on the comparison between MZ and DZ twins [65, 66], since MZ twins share approximately 100% of the genomic sequence of their DNA, while DZ twins share, on average, only 50% of their segregating genes. By comparing the magnitude of correlations in DZ twin pairs (rDZ) to those in MZ twin pairs (rMZ), we can infer the presence of genetic factors. In univariate CTM, the total variance (V) of a trait can be decomposed into several components: additive genetic (A), non-additive or dominant genetic (D), shared or common environmental (C), and nonshared or unique environmental (E). In classical twin modeling of pairs reared together, simultaneous quantification of the C and D components is not possible. A ratio rDZ/rMZ above or below 0.5 is therefore an indication that C or D, respectively, can be modeled. Since the quantification of D requires considerable statistical power, we have only estimated A and E, with C where appropriate. Heritability (h²) is defined as A/V, and the standardized coefficients e² and c² denote the part of the variance V explained by E and C, respectively [67, 68].

We first fitted saturated models to obtain estimates of the twin correlations by maximum likelihood (ML), including sex and age at blood sampling as covariates. Analyses were conducted in R software using the OpenMx package, version 2.20.7 [69‐72]. In univariate settings, proteins for which rMZ/rDZ < 0.5 were fitted by AE models. Proteins for which rMZ/rDZ > 0.5 were fitted with ACE models, which were compared with AE, CE, and E only. Negative variance estimates were allowed [73]. The choice for the best model for these proteins was based on Akaike Information Criterion (AIC). Model comparison by AIC is available in the supplementary material (Additional file 3: Table S3).

In bivariate models, we decomposed phenotypically significant covariances between proteins and BMI trajectories. Since twin correlations indicated an AE model for BMI at blood sampling and changes in BMI in the univariate framework, we decomposed their phenotypic correlations with proteins into genetic (rA) and nonshared environmental (rE) correlations in the bivariate framework. The bivariate twin analysis between changes in BMI and protein abundance included the BMI intercept as a covariate in addition to sex and age. Proteins modeled by CE rather than AE (i.e., having h² = 0) were excluded from the bivariate modeling. Cross-twin cross-trait correlations can be found in the Supplementary Document (Additional file 3: Table S4, Table S5).

A correlation network was constructed to provide a multi-omics framework of the mechanisms underlying BMI change in FinnTwin12, using the significant findings from the prior analyses. The network included the 14 proteins that were associated with the change in BMI and the 3 PRSs. Metabolites associated with the change in BMI were also included, but the set of 48 highly correlated lipoproteins (out of 53 metabolites) were reduced by PCA into 3 principal components (LIPO.PC1, LIPO.PC2, and LIPO.PC3), explaining 87% of the variance. This allowed for better visual representation of the connections between omics. Each of the 25 final entries was linearly regressed on BMI slope, BMI intercept, and BMI at blood sampling, as well as simple and quadratic age at blood sampling, sex, age by sex interaction, and quadratic age by sex interaction. Correlations between these 25 residual variables were used to construct the multi-omics network. We colored the edges according to the significance of the correlation assessed by Pearson’s correlation (i.e., nominal p-value < 0.05 or Bonferroni-corrected p-value < 0.05) and the direction of the correlation (i.e., positive or negative). In NTR, we used a similar methodology to correlate metabolomics residual variables (3 lipoprotein PCs and 5 low molecular weight molecules) with gene expression (N = 14).

We examined associations between plasma protein abundance and BMI trajectories (i.e., BMI at blood sampling and slope of BMI) from mixed-effects models using scaled BMI trajectory markers (Fig. 1). Proteins were defined as independent variables along with covariates in mixe-effects modeling. A total of 66 proteins out of 439 proteins showed a significant association with BMI at blood sampling, of which 42 had a negative and 24 had a positive association with BMI (Table 2). Proteoglycan 4 had the strongest association with BMI (coefficient: 0.46; p = 2.1E − 34); an increase of one sd in the abundance of proteoglycan 4 was associated with an increase of almost a half of sd in BMI (Table 2).

A total of 14 protein abundances were associated with changes in BMI, independent of baseline BMI levels (Table 3). These proteins included complement factors and proteoglycan 4, with proteoglycan 4 having the strongest association with adolescent BMI changes (estimate: 0.27; p = 4.3E − 10). Ten proteins had a positive association with changes in BMI and 4 had negative associations. Thirteen out of 14 proteins were also correlated with BMI at blood sampling.

Similarly, we investigated associations between proteins and BMI intercept (i.e., baseline BMI) (Additional file 3: Table S6), of which 11 significant associations were detected. Of these, 9 were also associated with BMI at blood sampling. These findings suggest that the proteome in young adults contains proteins associated with both childhood and adulthood BMI. Proteoglycan 4, complement factor H, C-reactive protein, and secreted phosphoprotein 24 were associated with BMI at blood sample collection, BMI intercept, and BMI slope.

Associations between proteins with triponderal mass index (TMI) at baseline (~ 12 years old) and slope of TMI can be found in the supplementary material (Additional file 3: Table S7). In total, we found 10 and 22 proteins associated with baseline TMI and TMI changes, respectively. Seven of the ten proteins associated with baseline TMI were also associated with baseline BMI. All proteins associated with BMI changes were also associated with TMI changes. This shows that BMI and TMI associations with the plasma proteome tend to have similar proteins involved.

Univariate twin modeling was applied to quantify the genetic and environmental sources of variance in the 66 proteins identified as associated with either BMI at blood sampling and/or the slope of BMI (Fig. 2A). According to the ratio between rMZ and rDZ (Fig. 2B), we obtained 13 CE models, 1 ACE model, and 52 AE models. On average, genetic and nonshared environmental factors explained 35% (range: 0–78) and 59% (range: 22–89) of the variance in protein abundances, respectively (Fig. 2C). Thirteen proteins showed no evidence for genetic effects (h² = 0), and all of these had non-zero shared environmental components that accounted for an average of 28% (range: 19–41) of the variance, with nonshared environmental factors accounting for 72% (range: 59–81) of the variance in the abundances of these proteins (Fig. 2A). Heritability of BMI at blood sampling was 72% [95% CI: 62,78], and the nonshared environment accounted for the remaining (28%) of the variance in BMI at blood sampling. Heritability of the slope of BMI was 63% [95% CI: 51,73] while the nonshared environment explained the remaining (37%) variance in BMI changes. All univariate estimates and their 95% CIs are available in the supplementary material (Additional file 3: Table S3, Table S8).

The decomposition of significant phenotypic covariation between plasma protein abundances (h² > 0) with BMI and BMI changes were performed with bivariate genetic twin modeling. For all associations between proteins and BMI at blood sampling or BMI change, we applied bivariate AE models. Of the 53 BMI-protein correlations tested, we observed 43 significant genetic correlations and 12 significant nonshared environmental correlations (Fig. 3A). For 8 protein–BMI associations, we observed both significant genetic and nonshared environmental correlations. Overall, genetic correlations ranged from − 0.50 to 0.50 (absolute mean: rA = 0.29) and environmental correlations from − 0.33 to 0.44 (absolute mean: rE = 0.14). Seven proteins had genetic correlations with BMI greater than 0.4 in absolute value, and proteoglycan 4 had the highest nonshared environmental correlation with BMI among all proteins tested (rE = 0.44 [95% CI 0.28,0.57]). The full set of genetic and nonshared environmental correlations is available in the supplemental material (Additional file 3: Table S9).

Similarly, the significant associations between proteins and changes in BMI were decomposed into genetic and nonshared environmental correlations for 11 of the 14 proteins associated with changes in BMI (Fig. 3B). We observed 7 significant genetic correlations and 3 significant nonshared environmental correlations (Additional file 3: Table S10). Because BMI change during adolescence predicted adult BMI independently of baseline BMI (linear regression: R² = 45%), we examined whether the association between BMI at blood sampling and BMI change might confound the observed correlations between proteins and BMI. After adjusting the slope of BMI by BMI at blood sampling before bivariate twin modeling (Fig. 3B), only apolipoprotein A-IV remained genetically correlated with the slope of BMI (rA = 0.20 [95% CI 0.02, 0.38]), while all other correlations were no longer significant (Additional file 3: Table S11). These results suggest that the genetic and environmental factors influencing the association between changes in BMI during adolescence and protein levels in adulthood may reflect associations with adult BMI.