Epigenetic gestational age acceleration: a prospective cohort study investigating associations with familial, sociodemographic and birth characteristics

Preterm birth (≤ 37 weeks gestation) is associated with numerous health consequences such as increased mortality [1], hypertension [2, 3], insulin resistance [4] and respiratory problems in later life [5, 6]. Indeed, each additional gestational day at birth is associated with improved medical and neuropsychological outcomes in childhood [7]. Gestational age (GA) at delivery is typically determined via early obstetric ultrasound or last menstrual period (LMP), with ultrasound methods considered the more reliable [8], ‘gold standard’ procedure.

Recently, DNA methylation (DNAm) has been used to predict GA at delivery [9, 10]. This method builds on work that used DNAm to predict chronological age [11] and subsequent work, showing that the differences between predicted and chronological age are associated with disease outcomes. Those with DNAm-predicted ages that exceeded their chronological ages (age acceleration, AA) have a higher risk of cancer incidence [12–14], Alzheimer’s disease [15] and mortality [15–20]. Of note, the term AA is commonly used in the literature to describe both positive and negative differences (i.e. predicted ages above or below chronological ages), which could be misleading, but we use the term to be consistent with previous literature.

In a similar way, the gestational epigenetic clocks developed by Bohlin et al. [10] and Knight et al. [9] can be compared with actual GA to determine gestational age acceleration (GAA). There is little existing research on the association of potential predictors of GAA or the potential outcomes of GAA.

The aim of this study was to apply a previously published model for predicting GA from DNAm [10] and use this model to estimate GAA in order to (i) explore potential predictors of GAA by assessing the association of a broad range of socioeconomic variables and parental characteristics with GAA and (ii) explore potential outcomes of GAA by assessing associations of GAA with delivery and postnatal factors, using data from the Accessible Resource for Integrated Epigenomic Studies (ARIES) project, a subsample of child-mother pairs from the Avon Longitudinal Study of Parents and Children (ALSPAC).

Our analyses indicated that male sex, higher maternal pre-pregnancy BMI and vaginal delivery are associated with higher GAA. Our results also indicated that higher GAA was associated with birth size (birthweight, birth length and head circumference). There was no clear evidence of any associations of GAA with parental education, relationship status, smoking, alcohol consumption, depression or age, nor with housing tenure, financial difficulties, parity, pregnancy complications or APGAR scores.

Unlike in AA research, where AA has been associated with negative outcomes such as all-cause mortality and Alzheimer’s disease [15–20], it is currently unclear whether accelerated GA at birth is beneficial or detrimental to a fetus or newborn. From previous research using ARIES data, we know sex, birthweight, caesarean section and maternal BMI have also been associated with AA [22]. A recent study has also shown associations of GAA with birth size and sex [23]. However, in their main analysis, associations with birth size were in the opposite direction to our analysis when the raw GAA-GA difference was used as the outcome, i.e. higher GAA was associated with smaller size at birth. When calculating GAA using the residuals from a regression of DNAm-predicted GA on reported GA, as GAA was calculated in this study, there was no clear association between GAA and size at birth. In further contrast, we did not replicate their associations with maternal age, APGAR scores (at 1 min) and pregnancy complications (pre-eclampsia). The discrepancies in the results could be related to the focus on raw differences rather than residuals in the main analysis of Girchenko et al. [23]. We did not estimate GAA as the difference between DNAm-predicted GA and reported GA because the confounding effect of GA is not accounted for in this approach, whereas the residual-based approach ensures GAA is uncorrelated with GA.

Another explanation for the discrepancies between our findings and the findings of Girchenko et al. [23] is their use of the Knight et al. [9] GA prediction model rather than the Bohlin et al. [10] model applied in this study. We have previously noted several key methodological differences in the derivation of the Knight et al. [9] and the Bohlin et al. [10] models that influence the accuracy of the GA prediction in this cohort [21], such as the inclusion of preterm infants in the test set of the Knight et al. model [9] which is inappropriate for a data set with few pre-term births (as in this study). Additionally, the number of CpGs (148) included in the Knight et al. [9] training model was close to the sample size of the training set (207) and the model was then tested with a much larger sample size which may have resulted in an overfitting of the model. In contrast, the number of CpGs (96) in the Bohlin et al. [10] training model was much lower than the sample size of the training set (1068). Thus, the Bohlin et al. [10] model provided us with the best estimate of GAA for this cohort [21], which was reflected in the stronger correlation found between reported GA in the ARIES cohort and the model predictions of GA compared to the Knight et al. [9] model. The Bohlin et al. [10] model for prediction of GA from DNAm performed well in our data, adding support to the notion that DNAm could potentially be used as a marker of GA in data sets where GA has not been measured.

In this study, we were able to apply the epigenetic clock created by Bohlin et al. [10] in ARIES data. Despite the original model being trained only on a Norwegian cohort, the methods of Bohlin et al. [10] transferred to a UK cohort with considerable accuracy (r = 0.65). Additionally, the use of ARIES and ALSPAC data (a large and rich source of longitudinal data for children and their families) allowed us to assess the associations of GAA with a wide range of socioeconomic, parental and perinatal factors. The longitudinal nature of ARIES also enabled comparisons between GAA and AA at multiple ages.

Although the ARIES sub-sample is a more affluent sample of the full ALSPAC cohort, our results were robust to adjustment for SES. This is in line with some evidence that such differences are unlikely to severely bias association studies [24–27]. To account for missing data, we used multiple imputation, which maximised our statistical power. The results were consistent when using complete-case observed data.

Following Gervin et al.’s [28] methods, the regression analyses were adjusted for cell type proportions, even though GA is associated with variation in cell type proportions and this adjustment could therefore potentially bias the results. However, comparing the data before and after adjustment, there were no substantial differences in the results. Additionally, only weak associations were found between GA and cell type composition in this study (Additional file 1: Figure S1). Another potential issue with adjusting for cell type composition arises from deriving cord blood references from full-term births only; cord blood cell counts may be inaccurate for cord blood methylation profiles of preterm infants. However, a small proportion of the participants in our study were pre-term (3%) so it is unlikely that this will have substantially impacted the results.

Interestingly, there is little overlap between probes used in the Bohlin et al. [10] model of GA predication and the Horvath [11] model of age prediction, with only one CpG site (cg08965235 in the latent-transforming growth factor beta-binding protein 3 gene) overlapping between the models. In contrast to the Horvath [11] model, which compares accurate measures of chronological age and epigenetic predictions, the Bohlin et al. [10] model compares potentially inaccurate LMP/ultrasound estimates to epigenetic predictions. Consequently, inaccurate GA estimates may have impacted upon the GAA estimates in this research, especially as the majority of the GA estimates in ALSPAC are derived from LMP, since ultrasound estimation of GA was not common at the time of recruitment. Additionally, the GA predictions using the Bohlin et al. [10] model were more accurate using ultrasound methods rather than LMP. This may mean that our estimates may not be as accurate as if ultrasound GAs had been used and this could explain the discrepancy in accuracy between the Bohlin et al. [10] test data and the ARIES data, as detailed by Simpkin and colleagues [21].

Our results suggest that higher maternal BMI is strongly associated with higher GAA and that higher GAA is strongly associated with larger size at birth (birthweight, birth length and head circumference). In addition, we found weaker associations of sex and delivery method with GAA. Our results may indicate that having a BMI over 25 kg/m² is associated with more rapid development of the fetus in utero. The implications of GAA for postnatal growth, development and health warrant further investigation.