The genetic risk of gestational diabetes in South Asian women

South Asian women are at increased risk of developing gestational diabetes mellitus (GDM). Few studies have investigated the genetic contributions to GDM risk. We investigated the association of a type 2 diabetes (T2D) polygenic risk score (PRS), on its own, and with GDM risk factors, on GDM-related traits using data from two birth cohorts in which South Asian women were enrolled during pregnancy. 837 and 4372 pregnant South Asian women from the SouTh Asian BiRth CohorT (START) and Born in Bradford (BiB) cohort studies underwent a 75-g glucose tolerance test. PRSs were derived using genome-wide association study results from an independent multi-ethnic study (~18% South Asians). Associations with fasting plasma glucose (FPG); 2 hr post-load glucose (2hG); area under the curve glucose; and GDM were tested using linear and logistic regressions. The population attributable fraction (PAF) of the PRS was calculated. Every 1 SD increase in the PRS was associated with a 0.085 mmol/L increase in FPG ([95% confidence interval, CI=0.07–0.10], p=2.85×10−20); 0.21 mmol/L increase in 2hG ([95% CI=0.16–0.26], p=5.49×10−16); and a 45% increase in the risk of GDM ([95% CI=32–60%], p=2.27×10−14), independent of parental history of diabetes and other GDM risk factors. PRS tertile 3 accounted for 12.5% of the population’s GDM alone, and 21.7% when combined with family history. A few weak PRS and GDM risk factors interactions modulating FPG and GDM were observed. Taken together, these results show that a T2D PRS and family history of diabetes are strongly and independently associated with multiple GDM-related traits in women of South Asian descent, an effect that could be modulated by other environmental factors.


Introduction
Gestational diabetes mellitus (GDM) is defined as hyperglycemia first diagnosed during pregnancy. This abnormal increase in blood glucose levels is associated with an increased risk of adverse health outcomes for both mother and their fetus/child during pregnancy, and later in life (Farrar et al., 2016). It is estimated that 1% to >30% of live births are affected by GDM worldwide. This prevalence has been shown to vary widely depending on the participants ethnicity, countries/regions, and on the diagnostic criteria used (Archambault et al., 2014;McIntyre et al., 2019). South Asian women (whose ancestry derives from the Indian subcontinent) have a twofold increased odds of developing GDM, compared to white European women (Anand et al., 2016;Cosson et al., 2014;Farrar et al., 2015;McIntyre et al., 2019). The reasons for this disproportionate risk have not been fully characterized.
Gestational diabetes is a complex disorder influenced by multiple genetic and environmental factors such as maternal age, ethnicity, obesity, poor diet quality, and family history of diabetes (Anand et al., 2017;Hedderson et al., 2011;Solomon et al., 1997). Most genetic and environmental GDM risk factors are shared with type 2 diabetes (T2D; Sattar and Greer, 2002;Zhang and Ning, 2011) another condition that is thought to be very closely related to GDM. For example, women with GDM have a higher probability of having at least one parent with T2D, compared to those with normal gestational glycemia (Jang et al., 1998). Furthermore, women with a GDM history have a tenfold higher risk of subsequently being diagnosed with T2D compared to those without a history of GDM (Vounzoulaki et al., 2020). In terms of genetic architecture, both candidate gene and genome-wide association studies (GWASs) demonstrated a considerable overlap between GDM and T2D (Hayes et al., 2013;Kwak et al., 2012;Pervjakova et al., 2022). Finally, T2D polygenic risk scores (PRSs) have also been associated with GDM risk (Lamri et al., 2020;Pervjakova et al., 2022).
It has been demonstrated that environmental exposures such as diet and/or physical activity may modulate the effect of T2D loci (such as TCF7L2, PPARG, and CDKAL1) on the risk of T2D (Dietrich et al., 2019). Nevertheless, only a handful of studies have investigated genetic×environmental interactions on GDM (Chen et al., 2019;Grotenfelt et al., 2016;Popova et al., 2017), and to date, no study has tested the interaction between a genome-wide PRS with other GDM risk factors, on the risk of GDM.
The aims of this investigation were to: (i) test the association of a T2D PRS, generated from an external multi-ethnic GWAS (~18% South Asians), with GDM and related traits (fasting plasma glucose [FPG], 2 hr post-load glucose (2hG), and area under the curve glucose [AUCg] levels) in pregnant South Asian women from the SouTh Asian biRth cohorT (START) and the Born in Bradford (BiB) studies; (ii) To estimate the population attributable fraction (PAF) of the PRS on GDM; and (iii) To determine whether the effect of the PRS is modulated by other GDM risk factors including age, BMI, diet quality, birth country, education, and parity.

Results
The proportion of women classified with GDM using the IADPSG criteria was 25% and 11.2% in START and BiB, respectively, which was lower than the proportion using the South Asian-specific definition of 36.2% and 22.9%, respectively. Notably the proportion of women with GDM was higher in START compared to BiB irrespective of the classification method used.
The proportion of women of Indian origin in START and BiB was 71.8% and 5.1%, while the proportion of Pakistani women was 23.4% and 94.3%, respectively. The proportion of participants born in the Indian sub-continent was higher in START (88.6%) than in BiB (55.6%), and the average number of years spent in Canada or the United Kingdom among these participants was lower in START compared to BiB (6.6 vs. 9.7 years, respectively). The proportions of primiparous women (40.9% vs. 31.7%) and women with one prior pregnancy (42.4% vs.26.9%) were higher in START than in BiB. Conversely, participants with two or more prior pregnancies were more frequent in BiB than START (41.4% vs. 16.6%, respectively). The proportion of vegetarian participants was higher in START than in BiB (36.4% vs. 1.3%). Finally, the proportion of participants with a post-secondary degree/diploma or higher was greater in START than BiB (84.0% vs. 29.0%).
The standardized PRS ranged between -3.23 and 3.12 in START as compared to -3.51 and 4.16 in BiB. The full list of genetic variants included in the PRS as well as their characteristics are shown in Supplementary file 1a. Table 1 shows the baseline characteristics of the South Asian women from the START and BiB stratified by GDM case versus non GDM (IADPSG criteria). As expected, women with GDM had a higher mean fasting, 2hG and AUCg levels than non-GDM participants. Participants with GDM were older, had a higher BMI, and were more likely to report a family history of diabetes compared to women without GDM, in both studies. The overall diet quality was lower in participants with GDM compared to non-GDM participants in START (data not available in BiB). Of note, the average difference in BMI between GDM cases and controls was higher in BiB than in START (3.0 and 1.9, respectively) ( Table 1). Women with GDM had a higher mean PRS compared to women without GDM. Similarly, women with GDM were more likely to have PRS categorized in tertile 2 or 3, compared to tertile 1 ( Table 1).

Genetic risk and GDM-related traits in univariate models
The continuous PRS was associated with FPG, 2hG, and AUCg in START and BiB in univariate models. Every 1 SD increase in the PRS was associated with a 0.09 mmol/L increase in FPG (95% confidence interval [CI]=0.07-0.10), 0.23 mmol/L increase in 2hG (95% CI=0.18-0.28), and a 0.17 unit increase in AUCg z-scores (0.14-0.20) in the meta-analyzed results (Supplementary file 1b).
The PRS was also associated with the risk of GDM IADPSG in univariate models whereby a 1 SD increase in PRS was associated with a 47% increase in risk of GDM after meta-analysis (95% CI=35-60%). A similar association is observed using the South Asian-specific definition of GDM, with moderate between-study heterogeneity observed (Supplementary file 1b).

Multivariable models of GDM risk factors and GDM-related traits
The continuous PRS was strongly and independently associated with FPG, 2hG, and AUCg levels in a multivariable model adjusted for age, BMI, parity, parental history of diabetes, region of birth (South Asia vs. other), education level, and diet quality (available in START only), and the first five PCs ( Table 2). For example, every 1 SD increase in the PRS was associated with a 0.08 mmol/L increase in FPG, and 0.21 mmol/L increase in 2hG levels ( Table 2). The continuous PRS was also associated with a higher risk of GDM in a model with similar adjustments, whereby every 1 SD increase in the PRS was associated with a 45% increase in the risk of GDM IADPSG ( Table 2). Similar association results for GDM using the South Asian-specific criteria were observed and are shown in Supplementary file 1c.
When testing tertiles of PRS with similar covariates, our results show that participants in the second and third PRS tertiles have a 37% and 119% increase in the risk of GDM IADPSG compared to participants in tertile 1, respectively (Supplementary file 1d). Higher PRS tertiles were also associated with higher FPG, 2hG, and AUCg levels (Supplementary file 1d). The effect sizes associated with tertiles 2 were higher in START than BiB across multiple GDM-related traits (2hG, AUCg, and GDM; Supplementary file 1d).

Population attributable fraction and detection rate
In a model adjusted for maternal age, BMI, education, birth in South Asia (yes/no), parental history of diabetes, and diet quality (in START only), the PRS tertile 3 accounted for 12.5% of the population's total GDM IADPSG cases overall, and was higher in START than in BiB ( Table 3). The combined effect of PRS and parental history of diabetes on GDM accounted for ~21.7% of the population's GDM cases in the two studies combined ( Table 3).
The detection rate associated with the top versus lower PRS tertile was equal to 10% for a 5% false positive rate.

Interactions between the PRS and GDM risk factors on GDM
No consistent interactions were observed between the PRS and maternal age; parity; or education level modulating FPG, 2hG, AUCg, or GDM in START or BiB (Table 4 and Supplementary file 1e).
A couple of nominally significant interactions modulating the continuous trait of FPG were observed in START were not confirmed in BiB and vice versa. These included the PRS×BMI and the PRS×birth in South Asia (yes/no) interactions (START P interaction =0.01 and 0.04, respectively), yet non-significant in BiB (P interaction PRS×BMI =0.05 and P interaction PRS×birth in South Asia =0.07), with different effect sizes and opposing direction of effect between the two studies (Supplementary file 1f), resulting in non-significant metaanalysis of these effects (P interaction PRS×BMI =0.42 and P interaction PRS×birth in South Asia =0.26, respectively). Another interaction between the PRS and BMI modulating the risk of GDM was observed in BiB (P interaction =0.03), but not in START (P interaction =0.15; Table 4). Given that the overall direction of effect was similar in the two studies, this interaction remained significant after meta-analysis (P interaction =0.01). Nevetheless, this result in START could be a false negative given the study's smaller sample size (with a power to   detect a similar interaction to BiB of 9.9%). Subgroup analysis shows that the impact of a higher PRS on the risk of GDM was stronger in participants in lower BMI categories (Supplementary file 1f, Figure 1). Finally, a PRS×diet quality interaction on FPG was detected in START (P interaction =0.002; Table 4), whereby the effect of the PRS appeared to be stronger in participants with a low diet quality (Beta=0.17 [95% CI=0.10-0.24]) than in participants with a medium or high diet quality (Beta=0.05 [95% CI=0.00-0.09]) (Supplementary file 1f and Figure 2). Our analysis shows that we have 90% power to detect such an interaction. The overall diet quality score was not available in BiB; hence, this interaction could not be tested for replication.

Discussion
We demonstrate that a T2D PRS, based on an independent and multi-ethnic GWAS meta-analysis (with ~18% South Asian participants), is strongly associated with GDM and related glucose traits among South Asian pregnant women settled in Canada and the United Kingdom. This association is independent of other known GDM risk factors, including age, BMI, parental history of diabetes, and birth country. The PRS highest tertile accounted for 12.5% of the PAF of GDM. Consistent with a recent trans-ethnicity GWAS of GDM, and these results support the hypothesis that GDM and T2D are part of the same underlying pathology (Pervjakova et al., 2022). Family history of T2D is often used as a surrogate marker of the genetic risk of T2D. Our results show that the addition of the PRS to the multivariate models does not nullify the impact of parental history on GDM and vice versa. This suggests that the PRS and family history of diabetes both partially convey independent information. This partial independence could be explained by the fact that the PRS does not entirely capture the genetic association signals with GDM. On the other hand, family history reflects not only genetic similarity, but also shared non-genetic lifestyle factors.
By deriving a T2D PRS and showing its significant association with the risk of GDM, we confirm that the two diseases share a substantial proportion of their genetic background. In their recent publication, Pervjakova et al., 2022 also describe strong genetic similarities between the two traits by comparing the association and effect size of T2D variants to their effect on GDM. This convergence of observations using two different approaches (testing a PRS in our case versus independent loci in Pervjakova et al.) solidifies the hypothesis of a common genetic background between T2D and    GDM. It is however important to note that, although BiB's South Asian mothers were included in both analysis, they represented ~1.2% of the total sample size in Pervjakova et al., which suggests that our congruent conclusions are unlikely to have been driven by the sample overlap between the two studies.
Overall, the evidence for modulation of the PRS's effect on GDM-related traits by other GDM risk factors was weak. Most interactions tested were not significant in both studies. This absence of significance should however be treated with caution since our power analysis suggests that, given our sample size, we are only able to detect strong interaction effects. Two marginal PRS×BMI and PRS×-South Asia born interactions on FPG were observed, these were close to significance in both studies but did not replicate definitively, both in terms effect sizes and direction of effect, which precludes a power issue, and suggests differences in the effect of these environmental factors between the two studies, or possibly false positive results. Furthermore, these interactions would not pass multiple testing corrections if applied. Two potentially stronger PRS×diet quality, and PRS×BMI interactions modulating FPG and GDM were observed in START, and BiB respectively. However, since it was not possible to replicate these interactions (i.e., no comparable diet data available in BiB, and low power in START), future investigations are required in order to validate these observations. If confirmed, these interactions may help identify a subpopulation who will benefit the most from a targeted intervention for the prevention of GDM. Given the transient nature of GDM, another important research question would be the identification of women at greater risk of developing T2D after developing GDM, and how the genetic risk modulates this progression. This could be done by testing the interactions between a GDM/T2D PRS and T2D status in women with prior GDM. This could reveal whether women with prior GDM and a high genetic risk are more likely to develop T2D than women with prior GDM and a low genetic risk. Finally, given the low sensitivity of the PRS themselves, future studies should focus on deriving and estimating the predictive value of a composite score which combines the GDM/T2D PRS, family history of diabetes, prior GDM status, and diet quality score in order to improve the identification of women at higher risk of developing T2D.
The overall clinical implications of our findings should be carefully considered. At present, the use of laboratory-derived genetic information in the clinical setting remains expensive and is not implemented for complex diseases like GDM or T2D. Furthermore, our results show that, despite a strong association, the PRS has a low discriminatory value (detection rate of 10% for a 5% false positive rate) regarding GDM cases. This is in line with the observations of Wald and Old, 2019 stating that most polygenic scores of complex traits derived to date would perform poorly as a screening tests in a clinical setting.
Our study has been considerably strengthened by the use of a PRS optimized for a large population of South Asians from two independent cohorts, as well as by the fact that GDM status was determined using objective OGTT measures. Nevertheless, there are some limitations to our analysis that should be considered: (i) the weights attributed to the genetic variants included in the PRS are derived from a T2D study. Overall, evidence points to a strong correlation between top variants from T2D and GDM GWASs. However, variants at some common loci (e.g., MTNR1B) might have significantly different effect size depending on the phenotype studied (Pervjakova et al., 2022). In addition, variants in at least one locus (HKDC1) have been strongly associated to GDM but not T2D (Pervjakova et al., 2022). More GDM-specific loci, or loci with a different magnitude of effect between GDM and T2D might be identified from future, larger studies. These observations suggest that future PRSs based on a GDM GWAS may have more power to detect gene×environment interactions. (ii) Second, some differences in measurements exist between START and BiB studies, including the timing of weight measurements, and the number of data points included in the calculation of AUCg. However, since data were standardized in both studies, we do not expect that AUCg measurements differences had a major impact on the results. (iii) Finally, the comparison of genetic data between START and BiB revealed the existence of slight genetic heterogeneity, both between and within the samples of these two cohorts. It is our assumption that these differences can be explained by the difference of sample size (START being smaller than BiB), as well as by historical differences in migration patterns from South Asia to Canada and the United Kingdom. For example, most START participants were first-generation migrants from India, whereas the majority of South Asians in BiB are descendants of Pakistani migrants who settled in the United Kingdom for several generations. In order to account for this genetic heterogeneity, we derived our T2D PRS by combining samples from the two studies. This PRS should be more generalizable to other South Asian studies. Another measure implemented to reduce the effect of population stratification was the adjustment for the PC axes in our analysis. Given the absence of heterogeneity in our FPG, 2hG, or GDM IADPSG PC adjusted models, we consider that population stratification effects have been accounted for.

Conclusion
A T2D-derived PRS is strongly associated with the risk of GDM in pregnant women of South Asian descent, independent of parental history of diabetes, and other GDM risk factors.

Study design and participants
START is a prospective cohort study designed to evaluate the environmental and genetic determinants of cardio-metabolic traits among South Asian women and their offspring living in Canada (Anand et al., 2013). In brief, 1012 South Asian pregnant women, aged between 18 and 40 years old, were recruited during their second trimester of pregnancy from the Peel Region (Ontario, Canada) through physician referrals between 2011 and 2015. All START participants provided informed consent, and BiB is a prospective, longitudinal family cohort study designed to investigate the causes of illness, and develop interventions to improve health in a deprived multi-ethnic population in Bradford, England, UK (Wright et al., 2013). Between 2007 and2011, 12,453 women of various ethnic backgrounds (~46% South Asian origin) were recruited between their 24th and 28th week of pregnancy. Detailed information on socio-economic characteristics, ethnicity, family history, environmental, and physical risk factors has been collected (Farrar et al., 2015;Wright et al., 2013)

Measurements and questionnaires SouTh Asian BiRth CohorT
A detailed description of the maternal measurements has been published previously (Anand et al., 2017). Briefly, weight and height were measured using standard procedures, and information about pre-pregnancy weight, family, and personal medical history was collected using questionnaires. Parental history of diabetes was derived from baseline questionnaires and categorized as neither parent, or either one, or both parents had a history of diabetes. Birth country, number of years spent in Canada, and education-related variables were self-reported. Participants' highest level of education was coded as a five-category ordinal variable as: 1-less than high school; 2-high school completed; 3-Diploma or certificate from trade, technical or vocational school; 4-Bachelor's or undergraduate degree, or teacher's college; and 5-Master's, Doctorate or professional degree. A binary 'born in South Asia' variable was categorized as participants born in South Asia (India, Pakistan, Sri Lanka, or Bangladesh versus participants were born in any other country). A validated ethnic-specific food frequency questionnaire (FFQ) was used to collect dietary information (Kelemen et al., 2003). The following steps were implemented in order to calculate the diet quality of each participant: (i) for each of the following four food groups (green leafy vegetables; raw vegetables; other cooked vegetables; and fruits), 1 point was given for consuming ≥the study population median (vs. 0 points if intake <population median); (ii) for each of the following two food groups (fried foods/fast food/snacks; and meat/poultry), 1 point was given for consuming <the study population median (vs. 0 points if intake ≥population median); (iii) the points attributed to each of the six food groups mentioned above were summed in order to derive a continuous food score (ranging from 0 to 6 points), which was subsequently divided into three categories (Low diet quality -if food score=1 or 2; Medium diet quality -if food score=3 or 4; and High diet quality if food score=5 or 6). (iv) A binary diet quality variable used in our analysis was coded as follows (Low diet quality -if food score=1 or 2; medium or high quality -if food score≥3) (Anand et al., 2017).

Born in Bradford
Maternal height was measured during the recruitment visit (24-28th weeks of pregnancy) using standard procedures. In the absence of pre-pregnancy weight data, weight from the first antenatal clinic visit (average 12 weeks of pregnancy) was used to calculate BMI. Ethnicity of participants and years spent in the United Kingdom were self-reported at recruitment through an interview administered questionnaire; missing ethnicity data were backfilled from primary care data when available. The South Asian ethnicity of all participants included in this analysis was validated using genetic data. Parental history of diabetes and 'born in South Asia' variables were derived from the baseline questionnaire data and coded as in START. Since only a very small proportion of BiB's participants completed an FFQ that included information about fruits and vegetables intake, the diet quality score could not be derived in BiB. Data regarding the participant's highest educational qualification were equalized (using UK standards) and recoded into the following categories: 1-less than 5 General Certificate of Secondary Education (GCSE) equivalent; 2-5 GCSE equivalent; 3-A-level equivalent; and 4higher than A-level. Data for unclassifiable foreign degrees were considered as missing.

Outcomes
Study participants without prior T2D were invited to undertake a 75-g oral glucose tolerance test (OGTT) in both START and BiB, and FPG, and 2hG levels were measured (1 hr post-load glucose was measured in START only). AUCg was calculated using the FPG and 2hG glucose levels in BiB, and using the FPG, 1 hr post-load glucose, and 2hG levels in START (Anand et al., 2017). Given the difference in the number of data points included in the calculation of AUC between the two studies and the skewness of the distributions, values were log-transformed, winsorized, and standardized in each study before analysis. Gestational diabetes status of women without pre-existing T2D was primarily defined based on OGTT results in both studies using the International Association of Diabetes and Pregnancy Study Group (IADPSG) GDM criteria (FPG≥5.1 mmol/L or higher, or a 1hG≥10.8 or a 2hG≥8.5 mmol/L or higher) (Metzger et al., 2010). Our secondary outcome was GDM using BiB's South Asian specific definition (FPG of 5.2 mmol/L or higher, or a 2hG of 7.2 mmol/L or higher) (Farrar et al., 2015), which will be referred to as the South Asian-specific definition hereafter. Self-reported GDM status or data from the birth chart were used to determine GDM's status if OGTT measures were unavailable (N=65 and 31 in START and BiB, respectively). Women with preexisting diabetes at baseline were not included in this analysis. Pre-pregnancy diabetes status was determined using maternal self-reported data (about diabetes diagnosis, diabetes medication, and/ or insulin intake prior to pregnancy) in START. In BiB, information on pre-pregnancy diabetes was backfilled from electronic medical records.
In order to keep a single pregnancy (and a single GDM status) per mother in BiB, only pregnancies with no missing data for GDM were included. For mothers with available data at multiple pregnancies at this stage, pregnancies with no missing data across all covariates (age, BMI, family history, birth country, parity, and education level) were prioritized. Next, only pregnancies with the least amount of missing data across all covariates were kept. The following two additional filtering approaches were then applied for mothers with multiple pregnancies remaining: (i) if GDM was not diagnosed at any of the pregnancies, phenotype data at the latest available time point was kept (i.e., keep older GDM controls) and (ii) if GDM was diagnosed during any of the pregnancies included in the study, the earliest time point where GDM was diagnosed was kept (i.e., keep younger GDM cases). DNA extraction, genotyping, imputation, and filtering SouTh Asian BiRth CohorT DNA was extracted and genotyped for 867 mothers using the Illumina Human CoreExome-24 and Infinium CoreExome-24 arrays (Illumina, San Diego, CA). About 837 samples passed standard quality control procedures (Anderson et al., 2010). Genotype data was handeled using PLINK v1.90b6.8 (Chang et al., 2015) . Genotypes were phased and imputed using SHAPEIT v2.12 (Delaneau et al., 2014), and IMPUTE v2.3.2 (Howie et al., 2009), respectively, using the 1000 Genomes (phase 3) data as a reference panel (Auton et al., 2015). Variants with an info score <0.7 were removed from analysis. In total, 837 START participants with both genotypes and available GDM status, FPG, 1hG, and/or 2hG levels were included in the analysis.

Deriving the PRS
Given the absence of publicly available South Asian-specific T2D or GDM GWAS data at the time of the analysis, weights were derived from the DIAGRAM's 2014 multi-ethnic T2D GWAS meta-analysis, which included over 18% of South Asians (~63% European and 19% other ethnic backgrounds) (Mahajan et al., 2014). A grid search approach was used to identify the optimal parameters (17 p values tested, ranging from 5×10 -8 to 1 with 0.1 increase; 4 heritability values tested: 0.023, 0.06, 0.08, and 0.12). START and BiB genotypes were pooled. About 70% of the samples' data were used for training and 30% for validation (random sampling stratified by study) in order to minimize the impact of population stratification. The PRS was derived using LDpred2 (Privé et al., 2020). The best PRS (i.e., that maximized the AUC) was characterized by a p value≤0.0014 and an h 2 =0.08 (N SNVs =6492). The PRS was standardized (mean=0, standard deviation=1) in both studies before analysis.

Statistical analysis Regression models
The statistical analysis was conducted using R (v3.6.3) (R core Team, 2016). Linear regression models were used to test the association between the PRS and FPG, 2hG and AUCg. PRS and GDM associations were tested using logistic regression. Both univariate and multivariate models were constructed with adjustment for GDM risk factors (age, BMI, parity, birth in South Asia [yes vs. no]), education level, and diet quality (in START only) and the first five PCs (in order to minimize the effect of population stratification). Interactions between the PRS and each risk factor was also tested. Interaction plots were produced using the interactions R package (v1.2.0.9000) (Long, 2021).

Population attributable fractions
The estimated PAFs and their corresponding standard errors were calculated using the AF R package (v.0.1.5) (Dahlqwist and Sjolander, 2019). To this end, continuous variables were recoded into categorical variables: age was divided into two categories ([29-31, 32-43] vs. 19-28); BMI was stratified into a two categories variable using South Asian obesity cutoff points suggested by Gray et al., 2011 (<23 vs. ≥23); the PRS was divided into two categories (tertiles 1+2 versus tertile 3); parity was divided into two categories (primiparity versus 1 pregnancy or more); education level was divided into two categories (completed high school or lower versus higher degree, diploma, or certificate in START; and A-level equivalent or lower versus higher than A-level in BiB).

Detection and false positive rates
Detection rate (sensitivity) and false positive rate (1-specificity) for the OR of association of PRS tertile 3 versus 1 was estimated using the risk-screening converter tool developed by Wald and Morris, 2011. Power analysis for interactions Power to detect interactions was estimated using the InteractionPoweR R package (v0.1.1) (Baranger et al., 2022). Monte-carlo simulation was used using 10.000 simulations and an alpha of 0.05.