HELIX (https://www.projecthelix.eu/) is a collaborative project across six established and ongoing longitudinal population-based birth cohort studies in six European countries: EDEN - Étude des Déterminants pré et postnatals du développement et de la santé de l’Enfant (France); Rhea - The Rhea Mother–Child Study in Crete (Greece); KANC - Kaunas Cohort (Lithuania), MoBa - Norwegian Mother, Father and Child Cohort Study (Norway), INMA - INfancia y Medio Ambiente (Spain), and BiB - Born in Bradford (United Kingdom) (Maitre et al., 2018). In summary, HELIX aims to investigate the effects of the early-life exposome on child health, identify determinants and exposure patterns, understand molecular mechanisms, and assess the role of genetic background. All the participants in the study signed an ethical consent, and the study was approved by the ethical committees of each study area.

PRS were calculated in the 1,155 European ancestry children from the HELIX project with genetic data, and the GxE interaction analyses were performed in 1,086 children with genetic, phenotypic, and exposure data ( Supplementary Figure S1 ).

Maternal tobacco smoking was self-reported by the mother during different trimesters of pregnancy, from which we created two variables: any and sustained maternal smoking. Any maternal smoking was categorized as “smokers” (mothers who smoked at any point during pregnancy) and “non-smokers” (mothers who did not smoke at all during pregnancy). Sustained maternal smoking was categorized as “non-smokers,” “non-sustained smokers” (mothers who smoked only during the first trimester), and “sustained smokers” (mothers who smoked until the third trimester).

Exposure to second-hand smoke in childhood was assessed through a harmonized questionnaire administered to the parents at the age of ∼8 years. Children were classified as “exposed” if they were exposed to second-hand smoke at home or in other indoor spaces and as “unexposed” if they were not exposed at all.

BW (in grams) was collected as part of the study protocol of each cohort and harmonized in the context of the European Study of Cohort for Air Pollution Effects (ESCAPE) project (Agier et al., 2021). Whenever possible, gestation duration (in weeks) was defined as the interval between the start of the last menstruation and delivery; when the date of the last menstruation was missing, ultrasound-based estimates were used; when both measures were missing, obstetrician estimates were used. BW was converted in gestational age- and sex-standardized z-scores using the INTERGROWTH-21st reference curves (Villar et al., 2014). Six individuals had gestational age >300 days, which is outside the reference curves, and thus, for them, we could not calculate zBW.

At the age of ∼8 years, following a harmonized protocol across cohorts, anthropometry was assessed (Vrijheid et al., 2020). Children were asked to be in light clothing and without shoes, and then, height to the nearest 0.1 cm was measured with a stadiometer, and weight (in kg) was measured with a digital weight scale. BMI (in kg/m²) was calculated and converted into age- and sex-standardized z-scores (zBMI) using the international World Health Organization (WHO) reference curves (De Onis and Lobstein, 2010). Waist circumference (WC, in cm) was measured in a standing position, at the high point of the iliac crest at the end of a gentle expiration, using a metric tape and recorded in duplicate (Seca 201, Seca Corporation). Bioelectric impedance readings were performed with the Bodystat 1500 (Bodystat, Douglas, Isle of Man) equipment after 5 min of lying down. The proportion of fat mass (FM) was calculated using published age- and race-specific equations validated for use in children (Clasey et al., 2011). Using the distribution of the full study population combining all HELIX cohorts, we calculated age- and sex-standardized z-scores for waist circumference (zWC) and proportion of fat mass (zFM).

Child peripheral blood DNA samples were collected at the age of ∼8 years. Genome-wide genotyping was performed using the Infinium Global Screening Array (GSA) MD version 1 (Illumina) at the Human Genomics Facility, Erasmus MC (HuGe-F). Genotype calling was performed using the GenTrain2.0 algorithm based on a custom cluster file implemented in the GenomeStudio software and annotation with the GSAMD-24v1-0_20,011,747_A4 manifest.

Sample quality control was performed using the PLINK program (Purcell et al., 2007). The following filtering was applied for sample quality control: sample call rate <97% (n = 43), sex inconsistencies (n = 8), heterozygosity (>3 standard deviations) (n = 0), relatedness (sharing more than 18.5% of alleles) (n = 10), and duplicated samples (n = 19). The Peddy program was used to predict the ancestry from GWAS data and contrasted with self-reported ethnicity (Pedersen and Quinlan, 2017). Discordant samples were filtered out (n = 12). The following filtering was applied for the variant quality control: variant call rate <95% (n = 4,046), variants in non-canonical pseudoautosomal region (PAR) (n = 47), variants with minor allele frequency (MAF) < 1% (n = 178,017), and variants not in Hardy–Weinberg equilibrium (HWE), thus with a p-value <1E-06 (n = 913).

Genome-wide genotype imputation was performed with the Imputation Michigan server using the Haplotype Reference Consortium (HRC) cosmopolitan panel, version r1.1 2016. Before imputation, PLINK GWAS data were converted into VCF format, and variants were aligned with the reference genome. The phasing of the haplotypes was performed with Eagle v2.4 and the imputation with minimac4. Chromosome X was imputed, including PAR and non-PAR regions. In total, we retrieved 40,405,505 variants after imputation. Several quality control filters were applied to the imputed dataset: 1) imputation accuracy (R ²) <0.9, 2) MAF <1% and 3) HWE p-value <1E-06, giving rise to a final post-imputation dataset consisting of 1,304 samples and 6,143,757 variants (human genome build GRCh37/hg19 and plus strand). The first 20 principal components were computed from the GWAS data of European ancestry children using PLINK with the LD clumping option.

PRSs for BW, BMI, WC, and FM were computed for 1,155 children of the European ancestry using the PRSice v2 program and the imputed and quality controlled genetic data (Choi and O’Reilly, 2019). Summarized results of reference GWAS (base data) were retrieved from the Early Growth Genetics (EGG) consortium and from the PanUK Biobank—European population (Supplementary Table S1). PRSs were calculated operating with the average score method, which computes the score as a sum of the summary statistics for the effective allele multiplied by the number of effective alleles observed, divided by the number of alleles included in the PRS. SNP clumping was set at r ² > 0.1, and the rest of the arguments were left as default in PRSice v2. For each trait, we calculated 10 PRSs based on ten different p-value thresholds (Pt) of the base GWAS for SNP inclusion: Pt = {0.00000005, 0.000005, 0.0001, 0.001, 0.01, 0.05, 0.1, 0.2, 0.5, and 1}. For BW, we computed two sets of 10 PRSs: one using the EGG base GWAS (which includes around 25% of the HELIX children) (PRS-EGG) and the other using the PanUK Biobank base GWAS (with no overlap of individuals) (PRS-PanUK).

The optimal PRS for each trait among the 10 PRSs calculated with different p-value thresholds (Pt) was selected using PRSice v2. In particular, linear regression models were computed between the phenotypic trait and each PRS adjusting for the 10 first GWAS principal components (PCs). The PRS with the highest model-fit (R ²) was considered the best and used in downstream analyses. To control for overfitting due to parameter optimization (i.e., 10 Pt tested), the association between the best PRS and the phenotypic trait was corrected by performing 10,000 permutations of the phenotype and calculating an empirical p-value.

Finally, the best PRSs were centered and scaled by subtracting the mean and divided by the SD. They were analyzed as continuous variables and categorized into three groups: PRS-low (<25% percentile), PRS-mid (from 25 to 75% percentile), and PRS-high (>75% percentile). PRS was constructed based on the weights of the base GWAS; thus, the PRS group representing the highest genetic risk for zBW is the PRS-low, while for the other traits is the PRS-high group.

Frequencies (for categorical variables) and means and SDs (for continuous variables) were calculated. The cross-correlation of the PRSs and phenotypic traits was calculated with Pearson’s correlation coefficients and represented in heatmaps with the corrplot R package (Wei et al., 2017).

To test the association of the PRS and maternal smoking during pregnancy with phenotypic traits in the selected 1,086 children of the study, we fitted several linear regression models adjusted for covariates:

Phenotypic trait = PRS-continuous + any maternal smoking during pregnancy + covariates

Covariates for zBW were sex, gestational age, 10 first GWAS PCs, and maternal education. Covariates for zBMI-related traits were 10 first GWAS PCs, maternal education (in three levels), and second-hand smoke in childhood. Maternal education, which is associated with socioeconomic status and maternal age, was included as a covariate as it is a potential confounder for maternal smoking during pregnancy.

Effect sizes are expressed as the change in the z-score of the phenotypic trait (where each unit represents one standard deviation from the mean considering age and sex), by maternal smoking status (non-smokers as reference), by PRS group (PRS-low as reference), or by one standard deviation of the PRS (when treated in continuous). The effect modification of the PRSs on the association between maternal smoking and the phenotypes was tested by including an interaction term between these variables in the linear regression model:

Phenotypic trait = PRS-continuous + any maternal smoking during pregnancy + PRS-continuous * any maternal smoking during pregnancy + covariates

Partial F-tests were used to determine statistically significant differences between models with and without the interaction term (global interaction p-value). For statistically significant GxE interactions, models stratified by the PRS group were run.

Finally, we performed a series of sensitivity analyses to explore the effect on zBW of excluding cohorts included in the EGG-based GWAS summary statistics (INMA and MoBa) and of excluding non-preterm children. We also tested the effect on zBMI-related traits by not adjusting the models for childhood second-hand smoke and by excluding children unexposed to second-hand smoke.

All the analyses were conducted in R (version 4.0.3), and scripts can be found in a GitHub repository (https://github.com/georginafp/analysis_GxE).

The description of the main variables of the HELIX children is shown in Table 1. All the children were of European ancestry and were distributed across six ongoing European birth cohorts. In the study sample, 45.8% were females, the mean gestational age was 39.6 ± 1.6 weeks (4.9% preterm children), the mean age at the HELIX visit was 8.1 ± 1.5 years, and 52.5% were born to mothers with high educational attainment. Mean (standard deviation—SD) of BW and BMI were 3,392.9 (490.8) grams and 17.03 (2.63) kg/m², respectively. For the analyses, phenotypic traits were sex- and age-standardized using international reference curves (zBW and zBMI) or own distributions (zWC and zFM). In total, 16.9% of the mothers reported having smoked at some point during pregnancy, and 9.7% were classified as sustained smokers as they smoked during the whole pregnancy period. Childhood exposure to second-hand smoke at home or other places was 36.8%. The correlation between pregnancy and childhood exposure was ∼0.4 and has been described elsewhere in detail (Vives-Usano et al., 2020). The description of the main variables by cohort is shown in Supplementary Table S2.

Ten PRSs were calculated per phenotypic trait for the 1,155 HELIX children. The distribution of the PRSs, the total phenotypic variation explained (R ²), and the associations of the best PRS in quantiles are shown in Supplementary Figure S2. The best PRS for each trait was selected for downstream analyses. They contained 3,316, 18,563, 60,993, 9,203, and 62,011 SNPs and explained 4.9, 2.5, 4.7, 4.9, and 3.3%, respectively, for zBW-EGG, zBW-PanUK, zBMI, zWC and zFM. After correcting through permutations, empirical p-values for the association of the best PRSs with phenotypic traits were 9.99E-05. The cross-correlation of the PRSs and phenotypic traits is shown in Figure 1. zBMI-related traits were correlated among them (r > 0.72) but not with zBW (r < 0.13). A similar pattern was observed for the PRSs but with a lower strength (r > 0.63, among PRSs for BMI-related traits; r = 0.48 among PRSs for zBW (EGG vs. PanUK); r <|0.06|, between PRSs for zBW and PRSs for BMI-related traits). The correlation coefficient between the traits and their PRSs ranged from 0.18 (zFM) to 0.29 (zBW).

The mean and SD of each trait, by the PRS group and by maternal smoking status during pregnancy, are shown in Supplementary Table S3. Children in the PRS-high group had higher phenotypic values than children in the PRS-mid group, which had higher values than children in the PRS-low group. Moreover, children exposed to maternal smoking presented lower zBW and higher values of the zBMI-related traits. To test the association of maternal smoking (any/sustained) and of the PRSs (continuous/three groups) with phenotypic traits, we fitted linear regression models adjusted for covariates.

In the model including the continuous PRS, any maternal smoking in pregnancy was at least marginally associated with lower child zBW and higher zBMI, zWC, and zFM (p-values <5.64E-02) ( Table 2 ). The associations were only observed among children born from sustained smokers (p-values <2.99E-02) and not among children born of non-sustained smokers (p-values >0.402). In particular, children exposed to sustained smoking during pregnancy had 0.500 (standard error, SE = 0.100) lower zBW (adjusted for the PRS-PanUK), 0.321 (0.125) higher zBMI, 0.215 (0.099) higher zFM, and 0.280 (0.010) higher zWC than unexposed children. The effects of sustained maternal smoking on the phenotypic traits in the models including the PRS in three groups were similar ( Table 2 ).

In the model including sustained maternal smoking, the best PRSs (treated in continuous) were positively associated with the respective phenotypic traits (p-values <5.55E-08) (Table 3). The same pattern was observed when PRSs were categorized into three groups: low (<25% percentile), intermediate (25–75% percentile), and high (>75% percentile). Children in the PRS-high group of each trait had 0.634 (0.081) higher zBW (PRS-EGG), 0.538 (0.082) higher zBW (PRS-PanUK), 0.743 (0.103) higher zBMI, 0.451 (0.071) higher zFM, and 0.524 (0.080) higher zWC than children in the PRS-low group. Effect sizes remained similar when the models were adjusted for any maternal smoking during the pregnancy instead of sustained maternal smoking. Being above the 75th percentile of the PRS compared to being below the 25th percentile had a stronger effect (in absolute terms) than being exposed to maternal smoking during the whole pregnancy: 0.538 (PRS-PanUK) vs. −0.428 for zBW and 0.743 vs. 0.309 for BMI.

Then, we tested the interaction between maternal smoking during pregnancy and the best PRS for each trait in the association with the phenotype. We observed a GxE interaction for zBW between maternal smoking during pregnancy and the PRS-EGG treated in three groups (p-values interaction = 0.052 and 0.01, for any and sustained smoking, respectively) (Supplementary Table S4 and 5). Maternal smoking during the whole pregnancy was associated with decreased zBW in children who had a certain genetic predisposition to low zBW (-0.399 (0.173) and -0.655 (0.157) for children in the PRS-low and in the PRS-mid groups, respectively). In contrast, the effect of sustained smoking was negligible in children within the PRS-high group (-0.097 (0.195)) (Table 4). However, this interaction was not observed when we analyzed the PRS-PanUK (p-value interaction = 0.096 and 0.341, for any and sustained smoking, respectively) (Table 4 , Supplementary Table S4 and 5). Finally, there was no evidence for a GxE interaction for zBMI-related traits (p-values for interaction >0.237) (Supplementary Table S4 and 5).

We, then, performed a series of sensitivity analyses. First, we repeated the analyses of sustained maternal smoking in relation to zBW without INMA and MoBa cohorts which are present in the EGG-based GWAS summary statistics (Supplementary Table S6), and the GxE interaction was still observed with a similar magnitude of the effect. The R ² values explained by the PRS-EGG in the subsets of children without overlap were 4.8% (without INMA) and 4.2% (without INMA and MoBa). Second, we repeated the analyses of sustained maternal smoking in relation to zBW without preterm children, and the results did not change substantially (Supplementary Table S7). Third, the models of zBMI-related traits were analyzed again without adjustment for childhood second-hand smoke or restricting the analysis to children unexposed to second-hand smoke (Supplementary Table S8). Effect sizes did not change much in both cases; however, the associations of sustained maternal smoking during pregnancy were not significant any more in the restricted population, likely to the reduced sample size (n = 672).