Gene regulation contributes to explain the impact of early life socioeconomic disadvantage on adult inflammatory levels in two European cohort studies

Individuals growing up during childhood in a socioeconomically disadvantaged family experience a higher rate of inflammation-related diseases later in life. Little is known about the mechanisms linking early life experiences to the functioning of the immune system decades later. Here we explore the relationship across social-to-biological layers of early life social exposures on levels of adulthood inflammation (C-reactive protein) and the mediating role of gene regulatory mechanisms, epigenetic and transcriptomic profiling from blood, in 2,329 individuals from two European cohort studies. Consistently across both studies, we find transcriptional activity explains a substantive proportion (up to 78%) of the estimated effect of early life disadvantaged social exposures on levels of adulthood inflammation. Furthermore, we show that mechanisms other than DNA methylation potentially regulate those transcriptional fingerprints. These results further our understanding of social-to-biological transitions by pinpointing the role of pro-inflammatory genes regulation that cannot fully be explained by differential DNA methylation.


INTRODUCTION
Inflammatory processes have been suggested as a potential pathophysiological mechanism underlying the development of several major chronic diseases, including cardiovascular, metabolic and psychotic disorders 1,2,3 . Systemic inflammation is also one of the most studied pathways in the context of the social-to-biological transition 4,5,6,7,8 . Although early life disadvantaged socioeconomic conditions have been consistently shown to lead to elevated levels of inflammation in adulthood 9 , the underlying biological processes through which those socioeconomic exposures are embodied remain unclear.
Various studies have suggested a dysregulation of pro-inflammatory genes transcription in leukocytes as a potential biological mechanism for the embodiment of socioeconomic disadvantage 10,11,12,13,14 . This line of reaserch rests on the hypothesis that socioeconomic disadvantage activates the sympathetic nervous system associated to body's flight or fight response to stressors that eventually leave a leukocyte transcriptional fingerprint. Similar findings have been reported in experiments conducted with macaques 15,16 . Other studies have focused on the epigenetic regulators of transcriptional activity in leukocytes, in particular the DNA methylation levels at specific CpG sites 17,15,18 . These research endeavors were conducted without directly relating DNA methylation to gene transcription levels, whose functional relationship does not always hold 19,20 .
To date, the contribution of pro-inflammatory genes regulation to socioeconomic inequalities in inflammatory levels has not been investigated empirically, although frequently discussed theoretically 10,17 . To explore the chain of biological mechanisms linking early life socioeconomic disadvantage to adult inflammatory levels via DNA methylation and gene transcription in leukocytes, we used data from a Finnish (1,623 participants) and a Swiss (706 participants) population-based cohort study, and evaluated the consistency of findings across these two populations. First, we calculated the portion of the effect of early life socioeconomic disadvantage on adult heightened inflammatory levels explained by transcriptional level of pro-inflammatory genes either inferred by All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint this version posted April 6, 2020. . https://doi.org/10.1101/2020.04.03.20050872 doi: medRxiv preprint applying 2-sample Mendelian randomization 21 on the whole transcriptome (2sMR) or belonging to the conserved transcriptional response to adversity (CTRA) model 12 . We measured levels of systemic inflammation through circulating C-reactive protein (CRP), as it is a sensitive marker of inflammation and elevated amounts have been associated with increased risk for several diseases 3 . Second, we established the most likely functional relationship between cis (based on their genomic location) epigenetic and transcriptional changes induced by early life socioeconomic disadvantage via Bayesian networks model selection 22 .
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint this version posted April 6, 2020.

Pro-inflammatory genes selection via 2sMR
We used summary statistics of the associations between genetic instruments (Single Nucleotide Polymorphisms, SNPs) and gene transcription in leukocytes or CRP in blood obtained via large-scale GWAS conducted by the eQTLGen consortium 23 and the CHARGE consortium 3 , respectively (see Methods). There were 10,701 genes for which 2sMR provided a test of the causal association between gene transcription and CRP levels. There was an average of 39 semi-independent instruments (cis eQTLs in a linkage disequilibrium r 2 ≤0.25, see Methods) per each gene and the average F-statistic was 97. Finally, 81 genes were significant at q≤0.05 (see Table S1 and Methods). Of note, one of them, NFKB1, was as well one of the CTRA genes. Functional enrichment analysis revealed enriched inflammatory bowel disease, influenza A and prion disease pathways.

Characteristics of populations
As reported in Table 1, on average Finnish participants (YFS study) were 42-year old (age ranging between 34 and 49 years), while Swiss participants (SKIPOGH study) were 53-year old (age ranging between 25 and 88 years). Women represented 46% of the YFS population and 52% of the SKIPOGH population. CRP levels in SKIPOGH were distributed differently than in YFS, as participants in the upper quintile had CRP≥3 mg/L in SKIPOGH while had CRP>2.2 mg/L in YFS (see Table 1). Individuals with high/low parental occupational position were 18%/26% in YFS and 25%/28% in SKIPOGH.

Mediation by genes transcription
Individuals with low parental occupational position (potentially counter to fact) had an increase in (the geometric mean of) adulthood CRP of 22.6% [95% confidence intervals (CI): 2.7%, 46.4%] for YFS and 65.1% [95% CI: 32.7%,106.9%] for SKIPOGH compared to having (potentially counter to fact) high parental occupational position (total effect, see Figure 1). The point estimate of the geometric mean of CRP for individuals with low parental occupational position was 0.75 mg/L in YFS and 0.52 mg/L in All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint this version posted April 6, 2020.  50.3%] in SKIPOGH. The partial redundancy between the two gene sets was supported by correlation patterns between the principal components and between the transcription levels (see Figures S1 and S2).
Overall, in both cohorts the magnitude of effects was in the same direction and that of indirect effects was similar, with estimates of 17.2% and 13.9% indicating that pro-inflammatory genes transcription mediates a substantive portion of the estimated total effect of parental occupational position on adult CRP.

Regulation of genes transcription by local DNA methylation
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint this version posted April 6, 2020. We applied Bayesian network scoring to 4,076 CpGs located nearby (<1500 bp) the 2sMR-implicated and CTRA genes (see Table S2), for five causal structures (Group A to E, see Figure S5 and Methods). Table 2, for most CpGs (Group A: N=3,612 (88.6%) and N=3,253 (79.8%) in YFS and SKIPOGH, respectively), there was no substantive evidence (logarithm of Bayes factor larger than Overall, a non-empty set of CpGs was present across the two cohort studies for Groups A, C, D and E. Furthermore, this consistency at pattern level did correspond to a replication at the level of single site for 2,894 CpGs belonging to Group A only (see Table S2).

As described in
Compared to CpGs in Group A, sites in either Group C or D (N=41 and N=53 in CRYFS and SKIPOGH, respectively) had similar distribution in CpG island position (island or shore or shelf vs open sea, χ 2 =0.1 and 0.2, P value=0.76 and 0.48, %, YFS and SKIPOGH, respectively) and increased TSS or UTR annotation (56% vs 44% and 64% vs 44%, χ 2 =2.8 and 7.9, P value=0.15 and 0.005, YFS and SKIPOGH, respectively).

Sensitivity analyses
Total and indirect effect estimates were in the same direction and of similar (for YFS) and higher (for SKIPOGH) magnitude to those reported in main analysis when excluding participants with high values All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint this version posted April 6, 2020.  Figure S3). By varying the amount of variance explained by principal components summarizing transcription levels of the genes' sets, indirect effect estimates slightly increased when variance moved from 30% to 50% and remained stable when variance moved from 50% to 60% (see Table S3). This finding supports the fact that indirect effects reported in main analyses do not underestimate the amount of mediation by proinflammatory genes regulation. In SKIPOGH, when using a dichotomous CRP (≥3 mg/L) as outcome, the estimated indirect effect was substantive (see Table S4), backing the results presented in main analyses with an imputed continuous CRP.
When parental education was considered as the exposure, the magnitude of indirect effects was in the same direction and slightly lower than that reported in main analysis (see Figure S4). Furthermore, the Bayesian scoring procedure supported a pattern of models similar to the one found for parental occupational position, but no consistent evidence for models whereby parental education drives transcription which then affects DNA methylation levels (see Table S5). When adding a group of negative control DAGs to the groups described in Figure S5, the Bayesian scoring procedure correctly did not assign that group to any of the CpGs.
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

DISCUSSION
We found that both Finnish and Swiss adults having experienced disadvantaged socioeconomic conditions early in life displayed heightened levels of systemic inflammation compared to their more advantaged counterparts, and that a substantive portion of this effect was transmitted by shifts in leukocytes pro-inflammatory genes transcription. Via 2sMR, we identified a new set of proinflammatory genes putatively influenced by socioeconomic conditions in early life. Finally, we showed a consistent pattern of inter-individual variation in local DNA methylation being either unrelated to or driven by inter-individual variation in transcription.
To our knowledge, this is the first study examining regulatory mechanisms linking early life socioeconomic conditions to inflammatory levels in adulthood by integrating epigenetic and transcriptomic profiling from blood in two European populations. The observed socioeconomic inequalities in CRP (that is the estimated total effect) were in line with a previous multi-cohort study (N=13,078) from four European countries 6 , whereby the pooled association between paternal occupational position and CRP resulted in an increased (geometric mean of) CRP of 20.9% [95% CI: 11.6%, 31.0%]. In our study, some differences across the two cohorts were observed in the estimated magnitude of total and indirect effects. As populations were sampled using different designs, were from different European countries and had a different age distribution, these characteristics may have influenced those inter-cohort differences. Remarkably, in both populations those effects were in the same direction and the magnitudes of mediation were similar. Indeed, the joint indirect effect estimates were 17.2% and 13.9% (2sMR+CTRA genes in YFS and SKIPOGH, respectively), indicating a substantive mediation. However, the magnitude of the indirect effects can be an overestimate since unmeasured confounding may have increased the association between genes transcription and heightened inflammation. Furthermore, another limitation of our study is that CRP was not measured through a high-sensitivity assay in SKIPOGH. However, main and sensitivity analysis provided consistent results, and two thirds of SKIPOGH participants had CRP measured with high precision at All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint this version posted April 6, 2020. . https://doi.org/10.1101/2020.04.03.20050872 doi: medRxiv preprint 2 mg/L (coefficient of variation<5%). Overall, our finding of a mediating role of pro-inflammatory genes regulation is supported by previous studies in humans and primates. Those research endeavours supported transcriptional fingerprint of decreased glucocorticoid and increased proinflammatory signalling as key mechanism in the embodiment of socioeconomic disadvantage 10,15 .
Our transcriptome-wide analysis also identified 81 pro-inflammatory genes via 2sMR, a method that is protected from confounding and reverse causation biases. By using summary statistics from two large data sets we increased statistical power, and by performing inference via three 2sMR methods based on orthogonal assumptions we strengthened the reliability of inference. Nevertheless, we cannot exclude bias in our 2sMR analysis due to selection of instruments in the same data set where we conducted inference 25 , which leads to underestimation of causal effect sizes. Furthermore, we applied a stringent, rather conservative threshold for genes' selection, and we could not apply 2sMR to all human genes, therefore we cannot exclude the existence of other undetected pro-inflammatory genes.
We showed in both Finnish and Swiss individuals that early in life socioeconomic conditions may regulate gene transcription in leukocytes without involving local DNA methylation changes, or with local DNA methylation playing the role of regulated event. This finding is in line with recent experimental observations obtained via a novel sequencing technology 19 , Bayesian network analyses of ancestry-related differences in DNA methylation/gene transcription in cord blood 35 and whole blood monocytes 36 , and a Mendelian randomization study 24 . As pointed out therein 24 , the inferred direction of causality between DNA methylation and gene transcription can be biased because their measurement is noisy. Nevertheless, our findings are based on interpreting only models with substantial statistical evidence (logarithm of Bayesian factor larger than one). At the same time, that threshold could have biased our findings, as we could not assign any of the investigated causal structures to about 10% of the CpGs. Our results do not exclude the possibility that at some cis CpGs methylation levels do drive gene transcription and they point to the presence of other regulating All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint this version posted April 6, 2020. All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint this version posted April 6, 2020.

Study populations
We used individual-participant data from two multi-centre population-based studies: the Cardiovascular Risk in Young Finns Study (YFS) 26  (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint this version posted April 6, 2020.

Causal models and measures
The causal structures posited in our study are represented in the directed acyclic graphs (DAGs) displayed in Figure 2. The DAG in figure 2A evaluates the portion of the effect of early life socioeconomic conditions (exposure) on adult inflammatory levels (outcome) transmitted by intracellular processes regulating transcription of genes in immune cells (mediators). This "indirect" effect captures all potential pre-and post-transcriptional regulating mechanisms including DNA methylation, histone modifications, transcription factors binding and microRNAs 30,31 . These processes mutually regulate in both physiological and diseases conditions. In the second part of our study, we focused on local DNA methylation as regulating mechanism (mediator) of a gene' transcription (outcome) according to the causal structure in Figure 2B. We chose DNA methylation as it is the most widely studied epigenetic mechanism in the field of social epidemiology 32 and as it represents a potential therapeutic target 33 . Since the relationship between levels of local DNA methylation and genes transcription can be bidirectional 34,35,36 or absent 19 , for each pro-inflammatory gene and DNA methylation site we did not posit any specific causal structure between gene transcrition and DNA methylation measurements. Instead, we investigated the most likely one among all possible causal structures (see Figure S5 and section entitled Bayesian network model selection).
Parental occupational position defined as the highest household occupation was used as an indicator of socioeconomic conditions in early life (main analysis). In YFS, the occupational position (manual, lower-grade non-manual and higher-grade non-manual, farmer) of the head of the household was assessed in 1980 (study's baseline), while SKIPOGH participants reported at the study's follow-up the profession of their father when they were children. Paternal occupational position is a commonly used indicator of socioeconomic conditions in early life 37 and is closely related to parental occupational All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint this version posted April 6, 2020. . position in the Swiss context until the early 1980s (see SI Appendix). In a sensitivity analysis, we adopted a measure of parental education as an alternative proxy of socioeconomic conditions in early life (see SI Appendix).
Occupational position was categorized according to the European SocioEconomic Classification (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint this version posted April 6, 2020. Age, sex, centre of blood collection were included as potential confounders.
In main analyses, we excluded individuals with unknown parental occupational position (N=27 [1.6%] in YFS and N=15 [2%] in SKIPOGH), father not working for health reasons (N=1 in SKIPOGH) and missing CRP (N=1 in SKIPOGH) for a total of 1,623 YFS and 706 SKIPOGH participants (mediation analysis). Parental occupational position (or parental education) was modelled as ordered variable, hypothesizing a dose-response relationship 41 . CRP was modelled as a continuous outcome and transformed via the natural logarithm to symmetrize its distribution. Finally, parental occupational position, DNA methylation and transcriptomic data were available in 1,367 YFS and 661 SKIPOGH participants (Bayesian network scoring analysis).

Pro-inflammatory genes selection
We selected pro-inflammatory genes based on two strategies independently of the study populations. First, we selected genes based on a transcriptome-wide causal inference analysis via 2sample Mendelian Randomization (2sMR) aiming at establishing leukocytes transcripts (exposure) driving levels of CRP in blood (outcome) 21 . 2sMR is potentially robust with respect to unmeasured confounders and exploits publicly available summary statistics of large-scale GWAS 42 . In our study, we used summary statistics of the associations between about 10M genetic instruments (Single Nucleotide Polymorphisms, SNPs) and gene transcription in leukocytes or CRP levels in blood obtained via GWAS conducted by the eQTLGen consortium 23 (N=31,684) and the CHARGE consortium 3 (N=148,164), respectively (see SI Appendix).
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint this version posted April 6, 2020. . https://doi.org/10.1101/2020.04.03.20050872 doi: medRxiv preprint Critical to 2sMR are assumptions about validity of genetic variants as instrumental variables 43 . To satisfy the assumption that chosen variants are predictive of the exposure, we selected only eQTLs (P value<1.9*10 -5 corresponding to a false discovery rate <0.05) 23 and -to reduce bias due to reverse causation -discarded those with a strong association to CRP (P value<1.9*10 -5 ). Remaining assumptions were addressed by using cis eQTLs (likely direct effect of SNP on nearby gene transcription) and applying three complementary 2sMR methods 43 (see SI Appendix). All 2sMR methods were run with a minimum of 10 instruments. We estimated F-statistics to assess instruments strengths (F-stat>10 is suggestive of strong instruments). We computed the q value 44 from the P values estimated via the three 2sMR methods and declared a gene transcription to be driving CRP levels when q≤0.05 in all three 2sMR methods.
Alternatively, we defined our set of pro-inflammatory genes as those having been identified by the conserved transcriptional response to adversity (CTRA) model 12 . The CTRA is composed by 53 genes characterized by up/down-regulation depending on whether their immune function is inflammationor antiviral/antibody-related, respectively 14 . This gene set has been associated to both early life and adult socioeconomic conditions 12,14 .

Functional enrichment
We run gene ontology (GO) and pathways enrichment analysis of the list of genes selected via 2sMR with g:Profiler 45 . The list of genes tested via 2sMR was submitted as the background gene list and multiple testing correction was performed with the g:SCS algorithm that takes into account the structure of functionally annotated gene sets 45 . GOs or pathways were declared as significantly enriched when g:SCS adjusted P values were <0.05.

Mediation approach
We adopted a counterfactual mediation framework based on natural effect models 46 to estimate the portion of the effect of early life socioeconomic conditions on adulthood inflammation transmitted simultaneously or en-bloc by transcription levels of multiple genes in leukocytes. Importantly, the All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint this version posted April 6, 2020. . https://doi.org/10.1101/2020.04.03.20050872 doi: medRxiv preprint identification of these joint indirect effects is not dependent on the true causal order of the mediators 46,47 . This property is useful in our case since the interrelations between the transcription of different genes cannot be determined unequivocally. The magnitude of total and joint indirect effects corresponded to exponentiated parameters of natural effects models and is interpretable as percentage of change in the geometric mean of CRP. The proportion mediated (PM) by a bloc of genes was derived by the ratio of the joint indirect and total effect parameters. See SI Appendix for detailed information on interpretation and estimation of natural effect models' parameters.
We applied principal component analysis (PCA) 48 to the selected pro-inflammatory gene sets for estimating a representation of genes' transcription. In our mediation model, using principal components of genes' transcription instead of measured transcription levels has the advantage of increasing the reliability of joint indirect effect estimates since it allows to both reduce the number of mediators and the impact of measurement errors in the mediators 49 . Indeed, PCA is an established dimensionality reduction method allowing the data to be described using a small number of uncorrelated variables (the principal components, PCs) while retaining as much information (variance in the original data) as possible. Furthermore, given the ordering of PCs according to decreasing values of explained variance, PCA is also a noise reduction method when retaining only the principal components with highest explained variance 48 .
In main analyses, we selected the number of principal components (PCs) in order to explain about 50% of transcription variance. Confidence intervals (95%) of total and joint indirect effects were estimated through percentiles from 5,000 bootstrap draws (with replacement).

Local DNA methylation
For each of the pro-inflammatory genes selected either via 2sMR or the CTRA model, we considered DNA methylation sites (CpGs) within 1,500 bp of the annotated gene location. So defined cis CpGs should ensure that their potential effect on transcription is direct and not through other genes. All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Bayesian network model selection
We applied Bayesian network scoring 22 to select the most likely causal pathway between local DNA methylation and gene transcription that are affected by early life socioeconomic conditions. For each CpG we evaluated the DAG structures represented in Figure S1. For each DAG, a posterior probability was estimated through the Bayesian Gaussian score and uniform priors across the eleven DAGs (bnlearn R package 50 ). Finally, to alleviate the potential brittleness of inference from a large number of causal structures, we assessed posterior probabilities of 5 groups of DAGs, by specifying a partition of the DAGs 51 . Namely, group A encapsulates models where paternal occupational position does not influence DNA methylation nor transcription levels; group B encapsulates models where paternal occupational position causes changes in DNA methylation and transcription levels, and DNA methylation also causes changes in transcription; group C encapsulates models where paternal occupational position drives transcription levels, but methylation and transcription levels are not related; group D encapsulates models where paternal occupational position drives transcription levels that in turn influences methylation levels; group E encapsulates models where paternal occupational position influences DNA methylation but not transcription levels (see Figure S1). To select the group of DAGs with substantial evidence, we computed a Bayes factor (BF) from the two highest posterior probabilities and choose the group with the natural logarithm of BF larger than 1. In other words, a group of DAGs is selected only if its posterior probability is at least roughly three times larger than any other group's posterior probability 52 .

Sensitivity analyses
We investigated the stability of analyses upon removing participants with CRP>10mg/L, which may reflect short term immune responses. We computed indirect effects when summarizing transcription levels with 30% and 60% of explained variance by principal components to assess whether the effect All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint this version posted April 6, 2020. . https://doi.org/10.1101/2020.04.03.20050872 doi: medRxiv preprint sizes reported in main analyses underestimate the amount of mediation. In SKIPOGH, we run mediation analysis with a binary outcome (CRP≥3 mg/L) as the so defined heightened inflammation was independent of the imputation procedure.
We ran mediation and Bayesian scoring analyses when parental education was used as an indicator of early life socioeconomic conditions. Finally, we repeated the model selection via Bayesian scoring when adding a group of negative control DAGs, namely models where uncoupled DNA methylation and gene transcription drive (both or either one) paternal occupational position.

DATA AVAILABILITY
Summary statistics for eQTLs were downloaded from https://eqtlgen.org/. Summary statistics for CRP are available upon request to the CHARGE consortium.
YFS and SKIPOGH individual data are available from the PIs of the study upon reasonable request. Figure 1. Size of effects and proportion mediated (PM) estimated by mediation analysis in each population. CRP change (%) and 95% confidence intervals (CI) are reported for the total effect of low (vs high) parental occupational position on CRP levels in adulthood, and for indirect effects through transcription levels of proinflammatory genes en-bloc. Joint indirect effects via 2sMR, CTRA or 2sMR+CTRA genes were estimated by summarizing transcription levels with principal components corresponding to about 50% of the transcription variance in each genes' set.

FIGURES
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint this version posted April 6, 2020. . https://doi.org/10.1101/2020.04.03.20050872 doi: medRxiv preprint Figure 2. Causal structures posited in our study. For the sake of simplicity, we do not draw confounders (age, sex, centre of blood collection). Early life socioeconomic conditions represent the exposure. A) Inflammatory levels in adulthood represents the outcome, and transcription of pro-inflammatory leukocytes genes in adulthood the investigated intermediate mechanism. B) Transcription level of a pro-inflammatory gene in leukocytes represents the outcome and DNA methylation levels at a specific cis CpG in leukocytes the mediating mechanism.
All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint this version posted April 6, 2020. information about parental occupational position, transcriptome and CRP. Categorical characteristics are summarized through their absolute and relative (%) prevalence. Age is summarized through its mean and range. CRP in YFS is summarized with median and lower to upper quintiles range, while in SKIPOGH is binarized as 3mg/L is the highest left truncation value (see Methods).

Characteristics
YFS SKIPOGH N 1,623 706 Table 2. Cardinality of the most likely group of model structures across the five tested competing groups (see Figure S5) for a total of 4,076 CpG sites in up to N=1,367 YFS and N=661 SKIPOGH participants. The most likely group of causal structures is selected when its posterior probability is at least roughly three times larger than any other group's posterior probability. None corresponds to CpGs for which none of the five tested groups reached winning evidence or models could not be estimated as the transcription level was not available in a few genes (37 and 33 CpGs in YFS and SKIPOGH, respectively). All rights reserved. No reuse allowed without permission.
(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.