Placental epigenetic clocks derived from crowdsourcing: Implications for the study of accelerated aging in obstetrics

Summary

Epigenetic gestational age acceleration has been implicated in obstetric syndromes including preeclampsia, yet robust conclusions require accurate and unbiased epigenetic age models. Herein, we curated 1,842 public placental methylomes and organized a DREAM challenge to develop models of gestational age. Participants were blinded to the test data that we generated from 384 placentas encompassing normal and complicated pregnancies. Models developed during and post-challenge compared favorably to existing models in terms of accuracy, yet they were better calibrated throughout gestation and indicated that reports of accelerated epigenetic aging in preterm preeclampsia were likely due to modeling artifacts. The models show that accelerated aging is associated with a decrease in birthweight percentiles in male neonates delivered at term. By contrast, preterm accelerated aging was protective against delivery of a small-for-gestational-age neonate regardless of fetal sex. This work informs our understanding of the fetal sex-dimorphic role of the placenta epigenome in obstetrics.

Graphical abstract

Highlights

• •Public placenta methylomes curated and used to crowdsource epigenetic clock models
• •New data in 384 normal and complicated pregnancies were generated to validate models
• •Robust clocks reassess the association between epigenetic age acceleration and preeclampsia
• •The link between placental aging and birthweight is modified by fetal sex and gestational age

Epigenetics; Pregnancy

Introduction

The placenta is a unique organ that functions as the fetal lung, gut, kidney, and liver, provides a physical barrier to protect the fetus from infection, and acts as a key immunological interface for maternal-fetal interactions.¹ Placental complications, particularly failure of deep placentation and vascular remodeling, are considered primary contributors to the great obstetrical syndromes such as pre-eclampsia, preterm birth, and fetal growth restriction.^2,3 Abnormal placental aging, characterized by cellular senescence, mitochondrial dysfunction, histopathological alterations, and shifts in epigenetic aging patterns, has also been linked to these syndromes.^{4,5,6,7,8,9,10,11} The placenta has a distinct hypomethylated and highly variable epigenome, particularly in promoters and repetitive elements, reflecting its specialized functions and sensitivity to environmental influences.¹² Placental DNA methylation profiles capture environmental exposures and reflect the organ’s normal development, the effects of physiological factors (e.g., fetal sex¹³), and pathological processes.¹⁴ Thus, assessing the epigenetic age of the placenta, based on DNA methylation data, offers a promising approach to understanding its biological aging and potentially linking accelerated or delayed aging to the onset and severity of obstetric disease.⁵

Biological age, as measured by epigenetic clocks of blood and other tissues, can be influenced by environmental factors, including smoking, obesity, sleep patterns, diet, exercise, stress, and diseases like cancer and diabetes.^{15,16,17,18,19,20,21} Estimating placental biological age through epigenetic data remains challenging, partly due to difficulties in obtaining placental samples that span most of gestation age spectrum, particularly from uncomplicated pregnancies. Nevertheless, such an endeavor is important to identify potential associations between accelerated placental aging and pregnancy outcomes.

Two notable efforts have demonstrated the potential of using DNA methylation profiles to develop placental epigenetic clocks for predicting gestational age (GA). Mayne et al. developed the first placental clock using the Illumina 27K CpG methylation array, achieving a mean absolute error (MAE) of 2.6 weeks in an independent dataset.^5,22 Subsequently, Lee et al. introduced three advanced models using the more comprehensive Illumina 450K array CpGs: robust placental clock (RPC), control placental clock (CPC), and refined robust placental clock (RRPC).²² Among these, the RPC, which was trained on samples from both normal and complicated pregnancies, was the most accurate, with an MAE of 0.96 weeks in their test dataset.²² The CPC model, trained on samples from relatively normal pregnancies, also outperformed Mayne et al.’s clock (MAE 1.02 vs. 2.6 weeks), likely due to the broader methylome coverage of the 450K array and increased sample size. Despite its accuracy, the RPC overestimates GA in preterm gestations, suggesting the need for further refinement in predicting biological age of preterm placentas. Additionally, evidence indicates that fetal sex, maternal complications, and chemical exposures to environmental pollutants (e.g., arsenic) impact epigenetic aging of the placenta.^{5,8,23,24,25,26,27,28}

Deviations from normal epigenetic aging in the placenta have been linked to adverse pregnancy outcomes such as preeclampsia and fetal growth restriction, suggesting that accelerated or delayed placental aging may serve as a biomarker for these conditions.^5,8,12,24 However, the findings have been inconclusive to date, partly due to the use of different epigenetic clocks and the lack of reproducibility. Developing a more accurate placental clock, particularly for preterm samples, is therefore important for advancing our understanding of placental aging and its relevance to pregnancy complications leading to preterm delivery.

To build new epigenetic clocks of the placenta, herein we curated a comprehensive dataset of DNA methylation profiles from tissue samples covering a range of gestational ages and obstetrical conditions and leveraged the dialogue for reverse engineering assessments and methods (DREAM) crowdsourcing framework to involve the computational biology community in this effort through two sub-challenges. In sub-challenge 1, we evaluated models predicting GA using CpG probes available on the Illumina 450K array. In sub-challenge 2, participants developed models using the broader methylome coverage available on the Illumina EPIC (850K) array, which was not previously described in this context. This challenge, therefore, relied on a larger sample size and provided deeper coverage of the GA span. Model evaluation was performed on a large, independent, blinded test dataset with diverse clinical outcomes that we have generated. This approach ensured a robust development process by preventing data overfitting, as participants had no access to the test set. The challenge attracted participants with expertise in obstetrics, aging, epigenetics, and deep learning. Insights from the challenge and prior research also informed the development of a robust post-challenge placental clock (PCPC). By developing a more accurate placental epigenetic clock and profiling the placentas of women with normal and complicated pregnancies, this study aimed to provide a new clinical tool together with a deeper understanding of the link between placental aging and pregnancy-related complications that impact long-term health outcomes in the offspring.

Results

Development of placental epigenetics clocks

Placental clock DREAM challenge

We curated DNA methylation profiles of 1,842 placental samples from 18 studies archived in the Gene Expression Omnibus (GEO) and ArrayExpress databases (Table S1) corresponding to a wide range of gestational ages (Figure 1). Data were generated using either the Illumina 450K (N = 930) or 850K (N = 912) methylation arrays. Including data obtained using the 850K platform allowed for a larger sample size and enhanced CpG coverage, facilitating a more comprehensive analysis of methylation patterns. The cohorts included samples from normal term pregnancies as well as those with pregnancy complications including preterm birth and preeclampsia. For the Placenta Clock DREAM Challenge, one fraction of the publicly available data (n = 100 samples profiled on the 850K platform) was designated as the leaderboard set to provide feedback to participants during model development. The remaining 1,742 samples (450K, N = 930; 850K, N = 812) were allocated to the training dataset (Figure 2). The challenge was divided into two sub-challenges (SC1 and SC2): the first used only the CpG sites shared between both methylation platforms, while the second allowed for the additional probes available only on the 850K EPIC array. An independent test dataset of 384 methylation profiles was generated at Wayne State University/Pregnancy Research Branch (NICHD) and was reserved for final evaluation of predictive models, thereby avoiding overfitting. The test dataset included samples from women with diverse clinical outcomes: normal term deliveries (≥37 weeks, N = 82), small-for-gestational-age (SGA) neonates delivered at term (N = 45), term preeclampsia (N = 43), preterm preeclampsia (N = 42), spontaneous preterm labor and delivery (sPTD, N = 91), preterm pre-labor rupture of membranes (PPROM, N = 76), and SGA neonates delivered preterm (N = 5; Figure 1). Ultrasounds were performed before 22 weeks for 383 of the 384 women in the test set, ensuring optimal gestational dating.²⁹ The clinical characteristics of the study population used as the test set are provided in Table 1.

Figure 1.

Distribution of gestational age by dataset and phenotypic group

(A) Gestational age distribution across 1,842 publicly available placental samples, grouped by dataset IDs from various studies used in the Placental Clock DREAM Challenge. Each dataset is represented by a unique color, and gestational age is plotted on the x axis.

(B) Gestational age distribution for 384 test samples from the test dataset, grouped by phenotypic categories: term controls, term small for gestational age (SGA), term preeclampsia (PE), preterm PE, preterm labor (PTL), preterm SGA, and preterm pre-labor rupture of membranes (PPROM).

Figure 2.

Overview of the placental clock DREAM challenge

The placental clock DREAM challenge aimed to develop models for predicting gestational age using DNA methylation profiles from placental samples. We curated 1,842 placenta methylomes analyzed with HumanMethylation450K (n = 930) and HumanMethylation850K (n = 912) arrays from public repositories. Of these, 1,742 samples were used as the training dataset, while 100 samples profiled on the 850K platform were specifically set aside as leaderboard data to provide real-time feedback during model development. A blinded test dataset of 384 samples was generated using the HumanMethylation850K array for final evaluation. Participants submitted dockerized models that were assessed on a virtual Linux server. Performance metrics, including root mean squared error (RMSE), mean absolute error (MAE), and Pearson correlation coefficient, were calculated to rank teams.

See also STAR Methods, DREAM challenge.

Table 1.

Demographic and clinical characteristics of the women included in the study

	Control (n = 82)	PTL (n = 91)	PPROM (n = 76)	Preterm PE (n = 42)	Preterm SGA (n = 5)	Term PE (n = 43)	Term SGA (n = 45)	p
Age (years)	23 (21–26.8)	23 (21–27)	25 (21–30)	24 (21.2–27)	29 (21–29)	25 (22–28)	23 (20–25)	0.311
African American	77/82 (93.9%)	83/91 (91.2%)	74/76 (97.4%)	42/42 (100%)	5/5 (100%)	42/43 (97.7%)	42/45 (93.3%)	0.323
Nulliparous	16/82 (19.5%)	20/90 (22.2%)	26/76 (34.2%)	17/42 (40.5%)	4/5 (80%)	15/43 (34.9%)	18/45 (40%)	0.007
Smoking	16/82 (19.5%)	20/91 (22%)	15/76 (19.7%)	6/42 (14.3%)	2/5 (40%)	10/43 (23.3%)	8/45 (17.8%)	0.816
Drugs	23/82 (28%)	22/91 (24.2%)	19/76 (25%)	13/42 (31%)	1/5 (20%)	8/43 (18.6%)	13/45 (28.9%)	0.883
BMI (kg/m2)	29 (24.2–33.5)	27.2 (22.8–31.8)	28.6 (23–33.2)	30.2 (27.1–36.3)	26.3 (21.3–26.8)	31.2 (25.8–39)	27.4 (23.4–34.7)	0.041
Gestational age (weeks)	39.1 (38.6–39.6)	35.3 (33.3–36.3)	34.7 (31.1–36)	35.1 (33.2–36.6)	36.1 (33.7–36.7)	38.7 (38–39.3)	39 (38.6–39.6)	<0.001
Birth weight (g)	3245 (3056–3474)	2365 (1890–2730)	2262.5 (1499–2545)	2212.5 (1639–2506)	1595 (1350–2075)	3111 (2902–3440)	2605 (2415–2735)	<0.001
Birth weight percentile	39.2 (29.9–59.2)	35.8 (23.8–46.5)	32 (19.3–43.2)	16 (7.7–34)	1.2 (1–3)	39.4 (18.2–54.9)	5.1 (2.2–8.4)	<0.001
Fetal sex (female)	41/82 (50%)	42/91 (46.2%)	36/76 (47.4%)	28/42 (66.7%)	0/5 (0%)	15/43(34.9%)	28/45 (62.2%)	0.007

Continuous variables, summarized as median (interquartile range), were compared using one-way analysis of variance. Categorical variables, presented as number (%), were analyzed using Fisher’s exact tests.

Baseline performance of current placental clocks

To establish a baseline for the models developed in the DREAM challenge, we evaluated the performance of three existing placental clocks on the test dataset: the RPC, the CPC, and the RRPC.²² The RPC achieved the best performance, with a root-mean-square error (RMSE) of 1.36 weeks and a MAE of 1.08 weeks. The CPC, which is based on “control” samples, showed the lowest accuracy, with an RMSE of 1.7 weeks (MAE = 1.34 weeks), while the RRPC, built from uncomplicated term pregnancies, demonstrated an intermediate accuracy (RMSE = 1.56 weeks, MAE = 1.2 weeks; Figure S1). An important observation is that all three existing models predicted higher epigenetic age than the chronological GA of preterm placentas regardless of the clinical phenotype (Figure S1). This suggests that either accelerated placental aging affects multiple pathways leading to spontaneous or indicated preterm birth, or that existing models are biased and do not necessarily capture underlying biology in this GA range.

Top-performing models generated in the placental clock DREAM challenge

We received 49 model submissions from seven independent teams spanning expertise in data science, computational biology, epigenetics, aging, and obstetrics. The ranking and performance metrics of the seven teams on the test set are presented in Table 2. The top-ranking model (Team 1) was based on methylation features from 450K arrays, achieving an RMSE of 1.24 weeks and a MAE of 0.99 weeks (Figure 3A). The RMSE and MAE for team 2 were 1.44 and 1.15 weeks, respectively, and for team 3 were 1.68 and 1.24 weeks, respectively (Figures 3B–3C). Although the relative ranking of these three teams was the same as based on the leaderboard set results, the model with the highest accuracy on the leaderboard set was ranked 4th based on results on the test set (Figure S2). This likely reflects the overfitting of the leaderboard set and highlights the need for a test set inaccessible to model developers. For the top two performing models in Figure 3, the parity line (predicted = actual) passes through the center of the scatterplot irrespective of the clinical phenotypes, suggesting good calibration across the entire GA span even though data modelers did not have access to the test data or to accuracy estimates of their models on the test set.

Table 2.

Final rankings and performance metrics of top teams in the placental clock DREAM challenge

Team	Name	RMSE (weeks)	MAE (weeks)	Pearson’s correlation	Sub-challenge
1	Herdiantri Sufriyana (@herdiants)	1.24	0.99	0.95	SC1
2	Team - ANTS	1.44	1.15	0.95	SC2
3	Team - CS-BBK	1.68	1.24	0.94	SC1
4	Orsolya Pipek (@opipek)	2.51	2.21	0.94	SC1
5	Team - Metformin-121	2.72	2.12	0.82	SC2
6	Daniel Su (@danielsu)	2.80	2.29	0.78	SC1
7	Tanay Panja (@tpanja)	3.01	2.28	0.80	SC1

This table presents the final rankings of participants in the placental clock DREAM challenge after evaluation on the test dataset. Teams are ranked based on their root mean squared error (RMSE) in predicting gestational age, with lower RMSE indicating better performance. The table also includes mean absolute error (MAE), correlation between predicted and actual gestational age, and the sub-challenge (SC1 or SC2) corresponding to the submitted model.

Figure 3.

Performance of top three teams in the placental clock DREAM challenge

(A–C) Scatterplots of chronological gestational age (GA) versus predicted GA by team 1 (A), team 2 (B), and team 3 (C). Each point is colored by phenotypic group: preterm labor (PTL), preterm pre-labor rupture of membranes (PPROM), preterm preeclampsia (preterm PE), term PE, term small for gestational age (SGA), and controls. Performance metrics (correlation, root mean squared error [RMSE] and mean absolute error [MAE]) are displayed for each team based on all 384 samples shown in Table 1.

To assess the robustness of the team rankings, we evaluated the models on 1,000 bootstrap resamples of the test data and calculated the Bayes factor as the ratio between the number of instances in which a model was better than the next-ranked model and the number of times when the reverse was true. The results of this bootstrap analysis (most Bayes factors >3) suggest that the ranking of the seven participating teams was robust to fluctuations in the test set (Figure 4).

Figure 4.

Robustness of team rankings in the placental clock DREAM challenge

The violin plots display the distribution of team performance (root mean squared error, RMSE).

across 1,000 bootstrap resampling iterations of the test set. Bayes factor evaluates the relative likelihood of a model (k) outperforming the next-ranked model (k + 1), providing a measure of ranking robustness.

See also STAR Methods, DREAM challenge.

Team 1 employed a multi-stage approach and fit a separate elastic net regression model for each clinical phenotype, after which they weighed the estimate from each model with the probability of belonging to a given phenotype. While the clinical phenotype was partly available for the training set, it was imputed for the test set. Their final models incorporated 10,433 CpG sites. Team 2 selected predictor CpGs based on evidence of correlation with GA on the training set and prior knowledge from epigenome-wide association studies (EWAS). The features were combined in a penalized regression model for GA prediction. The final model included 279 predictor CpGs. Team 3 used a purely data-driven approach based on a large neural network model using all available features. Their approach utilized contrastive learning with noise-based data augmentation to extract latent features, followed by fine-tuning to predict GA.³⁰

“Wisdom of crowds” placental clock

It is customary in DREAM challenges to evaluate the “wisdom of crowds” (WOC) by aggregating the predictions from several or all models generated in the challenge. Since the top three teams based on the test set accuracy also ranked 2nd to 4th on the leaderboard set, generating models that were substantially more accurate than all the remaining teams (>1.5 weeks decrease in RMSE, Figure 4), we created the WOC model as the averaged predictions of teams 1, 2, and 3. The scatterplot comparing the WOC predicted versus reported gestational ages is presented in Figure 5A. The ensemble model demonstrated improved performance, achieving an RMSE of 1.18 weeks and a MAE of 0.92 weeks. While this performance estimate may be slightly optimistic, as the top three teams were selected based on accuracy on the same test set, it suggests that prediction error can be further reduced by considering alternative models.

Figure 5.

Performance of post-challenge ensemble models

(A and B) Scatterplots show chronological gestational age (GA) on the x axis versus predicted GA by the “wisdom of crowds” placental clock (A) and the automated machine learning solution, AutoGluon (B), respectively. Points are color-coded by phenotypic groups: preterm labor (PTL), preterm pre-labor rupture of membranes (PPROM), preterm preeclampsia (preterm PE), term PE, term small for gestational age (SGA), and controls. Performance metrics (correlation, root mean squared error [RMSE] and mean absolute error [MAE]) are displayed for each model based on all 384 samples shown in Table 1.

Automated machine learning solution (Autogluon)

When completing a data challenge, a recurring question is whether the participating teams had sufficient skill and resources to generate close to optimal models to address the research question asked in the challenge. Given the size of the dataset, it is possible that participants encountered limitations of computing time and computer memory resources for fitting more complex models. To rule out this scenario, we employed Amazon’s AutoGluon, an automated machine learning (AutoML) tool developed to democratize machine learning.³¹ AutoGluon automatically trains and tunes multiple types of models, including random forests, neural networks, gradient boosting machines, and others, and then stacks them to form a powerful ensemble. This approach ensures that the best-performing models are selected without requiring manual intervention or expert knowledge. AutoGluon has consistently outperformed other AutoML frameworks across a wide range of datasets, as shown in a benchmark study involving 104 independent datasets, where it delivered top-tier results within a few hours of computation.³² When applied to SC1 training data, the Autogluon model achieved an accuracy similar to that of team 2 (RMSE = 1.48 weeks, MAE = 1.19 weeks, Figure 5B).

Post challenge placental clock

The performance of the existing placental clocks, and those derived during and after the challenge (WOC and AutoGluon), provide an understanding of the achievable accuracy when predicting GA from DNA methylation data. To facilitate broader scientific interpretation and ease of sharing with the community, we developed a PCPC. This clock utilized elastic net regression, a widely used approach in both previous placental clocks and this challenge. It was trained on all 1,842 publicly available samples (1,742 training + 100 leaderboard) using 450K array CpGs. The PCPC model included 503 CpG probes and its performance is presented in Figure 6. The model achieved an RMSE of 1.3 weeks, an MAE of 1.04 weeks, and a correlation of 0.96 between predicted and actual GA, comparable to the best-performing challenge model and the RPC. Like the top models derived during the challenge, the PCPC model does not systematically over-predict the age of placentas delivered preterm.

Figure 6.

Performance of post challenge placental clock

Scatterplot shows chronological gestational age (GA) on the x axis versus predicted GA by the post challenge placental clock. Points are color-coded by phenotypic groups: preterm labor (PTL), preterm pre-labor rupture of membranes (PPROM), preterm preeclampsia (preterm PE), term PE, term small for gestational age (SGA), and controls. Performance metrics (correlation, root mean squared error [RMSE] and mean absolute error [MAE]) are displayed based on all 384 samples shown in Table 1.

Across all models evaluated in this study, including the existing RPC clock (MAE = 1.08 weeks), the top-performing DREAM model (0.99 weeks), the wisdom-of-crowds ensemble (0.92 weeks), the PCPC (1.04 weeks), and AutoGluon (1.19 weeks), the apparent ceiling in prediction accuracy of approximately 1.0 week MAE likely reflects a practical limit on the achievable accuracy of placental epigenetic clocks, due to inherent uncertainty in the true GA and noise in the methylation data. To gain further insight into the regulatory characteristics of genomic targets of the probes included in the PCPC, we conducted an overrepresentation analysis in various genomic contexts and functional annotations. First, we examined the distribution of these probes across CpG island, shore, and open sea regions. This analysis revealed significant enrichment in CpG shores (OR = 1.3, q < 0.01) and open sea regions (OR = 1.2, q < 0.05). Next, we assessed whether the PCPC CpG probes were overrepresented in key regulatory elements, including enhancers, promoters, dual-function (promoter and enhancer) regions, and non-coding RNAs. The analysis showed enrichment in dual-function regions (OR = 1.8, q < 0.001), enhancers (OR = 1.6, q < 0.01), and enhancer RNAs (eRNAs) (OR = 1.7, q < 0.001). Finally, the gene ontology analysis (Table S2) of genes mapped to these probes identified biological processes related to embryonic skeletal system development (OR = 4.8, q < 0.001), skeletal system morphogenesis (OR = 3.9, q < 0.001), and inner ear morphogenesis (OR = 3.9, q < 0.05). This enrichment pattern suggests that the loci represented in the PCPC model may play a role in growth and development.

Biological implications of epigenetic clocks of the placenta

Epigenetic age acceleration

Epigenetic age acceleration, defined as the difference between predicted epigenetic age and chronological age, has been proposed as a method to assess abnormal aging in the placenta and other tissues. Such analyses could shed light on the relationship between epigenetic age and environmental factors, such as lifestyle choices and disease.^8,23,24,33 Figures 7, S3, and S4 display the epigenetic age acceleration values across different phenotypes in the test dataset, including PTL, PPROM, PE, SGA, and controls, using various epigenetic clocks. A typical pattern among the existing clocks, CPC (Figure 7A), RPC (Figure S3A), and RRPC (Figure S3B), was that placentas from preterm deliveries consistently showed significantly higher predicted gestational ages compared to their actual gestational ages, regardless of the clinical phenotype. The overestimation of GA for preterm neonates was most pronounced with the CPC (Figure 7A), which was trained on mostly uncomplicated pregnancies, assuming that this was an ideal choice to study abnormal aging in obstetrical diseases.²²

Figure 7.

Epigenetic age acceleration across clinical phenotypes

(A and B) Boxplots comparing epigenetic age acceleration, defined as the difference between epigenetic age and chronological gestational age, for six clinical phenotypes using the control placental clock (A) and post challenge placental clock (B). Mean acceleration values and one-sample t test p values are noted for each of the six phenotypes: preterm labor (PTL), preterm pre-labor rupture of membranes (PPROM), preterm preeclampsia (preterm PE), term PE, term small for gestational age (SGA), and controls.

See also STAR Methods, Quantification and statistical analysis.

Unlike with the existing clocks, the predictions from team 1’s model (Figure S4A) exhibited a reduced bias in preterm samples, with an overprediction of GA by only 0.7 weeks compared to the almost 2-week overprediction by CPC (Figure 7A). While a slight trend of overestimation remained, it is more likely an artifact. On the other hand, the predictions from the PCPC (Figure 7B) and team 2 (Figure S4B) did not show any bias for preterm samples. However, these models exhibited a negative bias for term pregnancies, where the predicted epigenetic age was consistently lower than the actual GA. A similar pattern of slower increases in epigenetic age with chronological age has been observed in other tissues.³⁴ This trend may also suggest that the CpGs included in the PCPC are capturing aspects of the placental aging spectrum more closely associated with fetal growth rate rather than the broader biological aging process.

These findings have substantial implications for assessing the relationship between accelerated placental aging and obstetrical disease. Specifically, the observed biases in existing clocks may have yielded spurious associations between accelerated aging and pregnancy complications. Improved models, such as those developed in the DREAM challenge, may provide a more accurate framework for investigating these associations in future studies.

Association between epigenetic age acceleration and pregnancy complications at term

Using the PCPC model, we assessed the associations of eGA acceleration with pregnancy complications at term, specifically PE (Table S3) and SGA (Table S4), by fitting logistic regression models. The models included the adverse pregnancy outcome as the dependent variable and eGA acceleration as the predictor, adjusting for relevant maternal characteristics and fetal sex. Considering prior evidence of heterogeneity in the relationship between eGA acceleration and pregnancy outcomes/risk factors,^8,23,26,28 interaction terms were included to allow for significant interactions between eGA and covariates. Term preeclampsia was not associated with eGA acceleration (OR = 0.96, p = 0.75) after adjusting for maternal age, obesity, nulliparity, smoking, fetal sex, spontaneous labor, and drug use. However, we observed a fetal sex-specific association between eGA acceleration and SGA. Specifically, a one-week increase in eGA acceleration was not significantly associated with SGA among male neonates (OR = 1.73, p = 0.057), but significantly associated with lower odds of SGA among female neonates (OR = 0.59, p = 0.045). This remained true when defining SGA status based on a customized birthweight standard³⁵ that accounts for maternal height, weight, parity, ethnicity, and fetal sex (OR = 1.4, p = 0.21 for males and OR = 0.58, p = 0.049, for females). The analysis based on customized SGA definition is more suitable in term gestations when fetal size is more strongly influenced by maternal characteristics and fetal sex, and a sizable proportion of SGA neonates may be constitutionally small rather than growth-restricted.^36,37

Alternatively, analyzing the birth weight percentile as a function of eGA acceleration in a linear regression analysis revealed that a one-week increase in eGA acceleration was associated with a 9.6-unit decrease in birth weight percentiles for male neonates (p < 0.001), with no significant association observed for female neonates (Figure 8A, Table S5). Similar trends in the relationship between birth weight percentiles and eGA acceleration were observed when using the CPC (Figure S4A), RPC (Figure S5A), and models from team 1 and team 2 (Figure S5A). The positive association between SGA as a binary outcome and eGA acceleration in male neonates was also observed when using CPC and RPC and team 2 models. These findings highlight a notable sex-specific relationship between eGA acceleration and fetal growth at term.

Figure 8.

Correlation between epigenetic age acceleration and birth weight percentiles for the test set

(A and B) Scatterplots showing the relationship between birth weight percentiles on y axis and epigenetic age acceleration as determined by the post challenge placenta clock on the x axis in term (A) and preterm (B) samples.

Association between epigenetic age acceleration and pregnancy complications in preterm gestation

Similarly to the analysis in term placentas, we evaluated the association between eGA acceleration and adverse pregnancy outcomes observed preterm (GA<37), namely preterm PE (Table S6), preterm SGA (Table S7), PPROM (Table S8), and PTL (Table S9). Among these outcomes, a significant association was found only for preterm SGA. Specifically, a one-week increase in eGA acceleration was significantly associated with reduced odds of SGA preterm (OR = 0.68, p = 0.035), suggesting that preterm placentas with more advanced epigenetic age may be protective against SGA. Analyzing the birth weight percentile as a function of eGA acceleration in a linear regression analysis revealed that a one-week increase in eGA acceleration was associated with a 5-unit increase in birth weight percentile among males (p = 0.007, Figure 8B, Table S10). While the increase was less pronounced in female than in male neonates, the interaction between sex and eGA acceleration was not statistically significant (p = 0.067, Figure 8B). Determinations of age acceleration using other models showed similar trends, consistent with the PCPC results. The increase in birth weight percentiles with eGA acceleration was significant for males when using CPC (Figure S5B), RPC (Figure S6B), and the model from team 2 (Figure S8B), with no significant relationship observed among females. In contrast, the model from team 1 showed positive relationship between eGA acceleration and birth weight percentiles, regardless of fetal sex (Figure S7B). Decreased odds of SGA with increasing eGA acceleration, regardless of fetal sex was observed when using other placenta clocks. These findings highlight the SGA-protective role of accelerated placental aging in preterm gestation.

Discussion

Pregnancy complications such as preterm birth and associated adverse outcomes are prevalent in the US, especially among ethnic minorities.^38,39,40,41 These complications impose a substantial economic burden on society and are difficult to predict owing to heterogeneity in underlying causes.^{42,43,44,45,46} The placenta has long been recognized as a critical factor in the development of great obstetrical syndromes, with defective deep placentation identified as a key contributor to many complications.² Abnormal placental aging has been proposed as a contributing factor to the genesis of these syndromes.^{4,5,6,7,8,9,10,11} Epigenetic modifications, which regulate gene expression and are sensitive to environmental stressors, provide a promising framework for investigating the placental contributions to obstetrical syndromes.^{47,48,49,50,51} Among these modifications, DNA methylation is the most widely studied. Unlike other tissues, the placenta is characterized by extensive hypomethylation and high variability, reflecting its specialized role in supporting fetal growth and its dynamic response to environmental influences.^12,14

Epigenetic clocks, which estimate the biological age of tissues based on the methylation status of specific genomic regions, have evolved significantly since their introduction in 2011.^52,53,54 First-generation clocks predicted chronological age with remarkable accuracy across various tissues.^52,55,56,57 Second-generation clocks, such as PhenoAge and GrimAge, incorporated health-related outcomes to estimate biological age as a marker of health status.^58,59,60 Most recently, third-generation clocks, including DunedinPoAm, have leveraged longitudinal data to estimate the rate of aging, enabling more refined and personalized assessments.⁶¹ This generation also includes universal clocks applicable to multiple species, such as the pan-mammalian epigenetic clock.⁶² A unifying feature of these clocks is that departures between predicted epigenetic age and chronological age obtained with these models correlate with aging-related conditions, albeit with varying correlation strengths. A typical application of epigenetic clocks has been to test the association of departures between predicted and actual age and lifestyle factors, such as smoking, exercise, and obesity, hence, demonstrating their ability to serve as a proxy for biological age.^{15,16,17,18,19,20,21}

In the context of placental biology, epigenetic clocks offer a unique opportunity, as the purported placental origin unites the “great obstetrical complications.” However, estimating placental biological age through epigenetic data are challenging, primarily due to the difficulty of obtaining samples from uncomplicated pregnancies at early/preterm gestational ages. Preterm placentas are associated with some underlying pathology, and longitudinal sampling is not feasible. Consequently, previous placental clocks have either relied on term samples from uncomplicated pregnancies or included preterm samples from pregnancies with complications.^5,22

Of the existing placental clocks, the CPC is particularly suited for studying placental aging in the context of obstetrical complications like prematurity, as it was developed using “control” placentas with likely fewer pathologies.²² However, applying the CPC to our dataset consistently overpredicted GA in preterm placentas. This suggests that most preterm placentas undergo accelerated aging irrespective of whether preterm birth is spontaneous or medically indicated. However, previous studies using placental staining reported both accelerated and delayed villous maturation defects within spontaneous PTL cases, and cases of delayed villous maturation were more common than accelerated villous maturation.^10,11 Therefore, these subgroups of PTL should represent two extremes on the placental aging spectrum, indicating that a shift in the biological age (toward accelerated aging) observed with CPC could represent a modeling artifact.

Here, we hypothesized that most placentas, whether collected at term or preterm, predominantly follow a typical aging trajectory, with abnormally aged placentas representing only a small subset. A clock built on a larger dataset, encompassing term and preterm samples, could provide robust GA estimates independent of study design. Supporting this, the overprediction of GA in preterm placentas was less pronounced with the RPC, which was trained on 1,102 placentas from pregnancies with and without complications, though some overprediction persisted.²²

To advance placental clock development, we curated an even larger dataset of publicly available DNA methylation profiles and leveraged the DREAM challenge framework, which we have previously leveraged for prediction of preterm birth using maternal transcriptomics, proteomics, and microbiome data.^63,64 Compared to Lee et al.,²² our study included significantly more placental samples analyzed on the 850K platform, increasing the sample size by 1.7-fold. A key aspect of our approach was that model development was decoupled from model assessment since the test data were not available in any form to participants in the challenge, and hence overfitting of the test set was avoided. The data challenge attracted diverse modeling approaches, including the use of clinical metadata to enhance predictions, data-driven feature selection, and neural networks that learned novel feature spaces. The best-performing models achieved an RMSE of 1.24 weeks and a MAE of 0.99 weeks for GA prediction, surpassing the performance of the robust placental clock (RPC; RMSE = 1.36 weeks, MAE = 1.08 weeks). More importantly, these models showed good calibration across gestational ages, offering better predictive accuracy for both term and preterm samples.

The use of auto machine learning software (AutoGluon) led to a model that achieved an RMSE of 1.48 weeks (MAE = 1.19 weeks), an accuracy similar to that of the second-ranked team. This result suggests that the additional clinical insight used by team 1 and additional information derived by team 2 from a meta-analysis were at least as important as systematically exploring and combining hundreds of different types of models, as done automatically by AutoGluon. The finding that domain-specific knowledge leads to more accurate models is consistent with results from other data challenges.⁶⁵

Given the complexity of the approach taken by team 1, which makes sharing and use of the model somewhat difficult, and since 100 methylomes were left out from model training to be used as the leaderboard set, we developed a PCPC that would address these limitations while being inspired by the approaches of the top teams. While based on a simple linear regression just as RPC, the PCPC model achieved a slightly higher accuracy and, importantly, did not systematically overpredict the age of preterm placentas. PCPC predicted lower epigenetic age for term placentas than chronological age, a phenomenon consistent with the Horvath clock⁵⁶ that underestimates epigenetic age in older individuals, suggesting a plateauing of biological aging.^34,66,67 This pattern aligns with placental growth dynamics, where rapid early growth transitions to maintenance to meet fetal metabolic demands, followed by functional decline characterized by oxidative stress, inflammation, and senescence as GA advances.^68,69,70

Our functional analysis of the methylation sites at the basis of the PCPC model showed enrichment of CpG shores, open seas, and enhancers, genomic features crucial for regulating tissue-specific gene expression and development. Gene ontology analysis further linked these CpGs to biological processes associated with development. These findings are consistent with prior pan-tissue and pan-mammalian epigenetic clocks, which similarly reported enrichment of developmental pathways, emphasizing the importance of these processes in predicting chronological aging.^56,62 We then examined the associations between epigenetic age acceleration and obstetrical complications, including PE, SGA, PPROM, and PTL, using the PCPC model. The only significant associations of accelerated aging were observed herein with SGA status or birthweight percentiles. In preterm gestation, an increase in eGA acceleration was correlated with higher birth weight percentiles and decreased odds of delivering SGA neonates. The impact of eGA acceleration was less pronounced or absent in female neonates, although this sex-specific effect modification did not reach statistical significance. Similar findings were obtained with clocks generated before or during the DREAM challenge. These observations, alongside the established link between preterm placentas and various pathologies, imply that accelerated epigenetic aging in premature neonates may be an adaptive response to chronic intrauterine stressors such as nutrient depravation or suboptimal resource allocation. Such adaptations might protect against clinically overt small-for-gestational-age diagnosis, highlighting the placenta’s compensatory role in supporting fetal development under adverse conditions.⁷¹ The connections previously established between maternal risk factors such as dyslipidemia,²³ depression during early gestation,²⁸ or smoking²⁶ and accelerated placental aging suggests that these factors may act as stressors that prompt the placenta to expedite aging, potentially maintaining fetal development in challenging maternal environments. Additionally, placental eGA acceleration was linked to shorter NICU stays among extremely preterm black infants,²⁶ highlighting the complex relationship between maternal health, placental adaptations, and neonatal outcomes.

By contrast, eGA acceleration in term gestations was associated with decreased birth weight percentiles in male neonates, suggesting accelerated epigenetic aging may have different implications for male than female fetal growth. This latter finding is consistent with observations by Tekola-Ayele et al., who reported an association between epigenetic age acceleration and reduced fetal growth in males but not females.⁸ These findings are supported by results based on age predictions from all clocks evaluated herein.

This study stands out primarily due to its large sample size, incorporating more 850K arrays to increase the dataset by 1.7 times compared to the previous study by Lee et al.,²² providing a robust basis for developing models. Unlike in our previous DREAM Challenge in which GA was predicted from transcriptomics data and participants could see the accuracy of their models on the test set for up to 5 consecutive submissions,⁶³ herein the test set accuracy was inaccessible to challenge participants. This avoided the situation in which the model with the best overall accuracy fitted better the more numerous term samples and over-predicted the preterm samples. The study also allowed us to assess the implications of using different clocks on the association between epigenetic age acceleration and pregnancy complications in an underserved and high-risk population of mostly African American patients. A further strength is derived from leveraging the DREAM challenge framework, which brought together a diverse computational and genomics expertise, thus enriching the study with diverse backgrounds and innovative approaches. Finally, the development of a PCPC, which can be easily shared and used by the community, provides the basis for further reproducible research.

Limitations of the study

Despite the significant contributions of this study, several limitations must be acknowledged. First, the improvement in overall accuracy to predict placenta age from methylation data were rather modest compared to existing models, yet the reduction in bias for preterm placentas was substantial. Ideally, epigenetic clocks should be developed using longitudinal samples from normal pregnancies to establish baseline trajectories, which could then be utilized to identify deviations associated with pregnancy complications. However, longitudinal sampling of placentas from normal pregnancies is practically and ethically unfeasible, limiting our ability to model typical gestational methylation patterns fully. Moreover, the ultimate objective of identifying non-invasive biomarkers for adverse pregnancy outcomes requires tools that can predict risks early and non-invasively. Placentas cannot be sampled early and non-invasively, underscoring the need for the development of blood-based biomarkers that can reflect deviations in epigenetic aging as determined by placental clocks. This represents a significant area for future research, aiming to bridge the gap between current capabilities and clinical needs for early and non-invasive detection of pregnancy complications. Another limitation is that our study relied exclusively on array-based DNA methylation data, primarily due to the availability of large, publicly accessible datasets.^72,73 While these platforms remain widely used, they cover only a subset of the methylome, and sequencing-based platforms offer more comprehensive coverage that may provide additional insights into placental aging.^74,75 However, using such platforms to develop epigenetic clocks will require the generation and sharing of large-scale placental sequencing datasets.

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Adi L. Tarca (atarca@med.wayne.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

Data

The large training set we curated can be downloaded from https://www.synapse.org/Synapse:syn59834055 and it is described in detail here: https://www.synapse.org/Synapse:syn59520082/wiki/628529. For the subset of patients in the test dataset who were asked and agreed to broad data sharing, the methylation profiles were deposited in the Gene Expression Omnibus (GEO: GSE287219).

Code

The description and the code used to train the models of the top 3 teams are available at: https://www.synapse.org/Synapse:syn62407322/wiki/629409, https://www.synapse.org/Synapse:syn61846522/wiki/629109, and https://www.synapse.org/Synapse:syn61964146/wiki/629445, for Team 1, 2, and 3, respectively. The Team 1 Placenta Clock model can be used using the rplec package in R and also available at: https://github.com/herdiantrisufriyana/rplec. The Post Challenge Placenta Clock model was implemented as an R package and available at: https://github.com/dw1227/PCPCmodel.

Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Consortia

The DREAM Placenta Clock Challenge Consortium is listed below, and author affiliations are available in Table S11.

Chih-Han Huang, Tsai-Min Chen, Hsuan-Kai Wang, Jhih-Yu Chen, Edward S.C. Shih, Chih-Hsun Wu, Wei-Quan Fang, Sz-Hau Chen, Kuei-Lin Huang, Tanay Panja, Orsolya Anna Pipek, Sheng-Yao Su, Victor Tarca, and Oladejo Ahmodu.

Acknowledgments

This research was supported in part by the Pregnancy Research Branch (formerly Perinatology Research Branch) of the Eunice Kennedy Shriver National Institute of Child Health and Human Development, National Institutes of Health, U.S. Department of Health, and Human Services (NICHD/NIH/DHHS) under contract HHSN275201300006C. A.L.T. was supported by the Wayne State University Perinatal Initiative in Maternal, Perinatal and Child Health. N.G.-L. was supported by a grant from the National Institute of Allergy and Infectious Diseases (NIAID)/NIH (RAI184481A), and the Next Gen Pregnancy Initiative, Burroughs Wellcome Fund (1263500). R.R. has contributed to this work as part of his official duties as an employee of the US Federal Government. H.S. was supported by the Postdoctoral Accompanies Research Project from the National Science and Technology Council in Taiwan (NSTC113-2811-E-A49A-003). E.C.-Y.S. was supported by a grant from the National Science and Technology Council in Taiwan (NSTC113-2221-E-A49-193-MY3). M.S. was supported by the UCSF March of Dimes Prematurity Research Center. N.A. was supported by the NIH (R35GM138353), the March of Dimes, and the Alfred E. Mann Foundation.

Author contributions

A.L.T., R.R., N.G.-L., T.C., and F.T.-A. designed the DNA methylation study that generated the test data. A.L.T., G.B., B.D., C.N., N.G.-L., J.A., N.A., G.S., and M.S. and D.B. designed and organized the challenge. H.S., T.P., I.A., E.C.-Y.S., S.H., A.v.B., and C.W. developed and submitted placenta clock models. The DREAM Preterm Birth Prediction Challenge Consortium participated in the challenge or contributed to testing the model submission platform. A.L.T. and G.B. developed tools to evaluate participant submissions, and performed the post-challenge analyses. A.L.T., G.B., R.R., and F.T.-A. interpreted the results. G.B. and A.L.T. drafted the manuscript. All authors reviewed, edited, and approved the final manuscript.

Declaration of interests

All authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Biological samples
Human placenta samples	Perinatology Research Branch, an intramural program of the Eunice Kennedy Shriver NICHD, NIH, DHHS, Wayne State University (Detroit, MI, USA), and the Detroit Medical Center (Detroit, MI, USA)	N/A
Critical commercial assays
Quick-DNA™ MagBead Plus Kit	Zymo Research	D4081
Infinium MethylationEPIC BeadChip Kit	Illumina	WG-317-1003
Deposited data
Raw methylation data for training set	Multiple studies76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91	www.synapse.org/syn59834055
Raw and preprocessed new methylation data	This paper	GEO: GSE287219
Software and algorithms
Minfi	Aryee et al.92	https://www.bioconductor.org/packages/release/bioc/html/minfi.htm
CpG impUtation Ensemble	Li et al.93	https://github.com/GangLiTarheel/CUE
ChAMP	Tian et al.94	https://www.bioconductor.org/packages/release/bioc/html/ChAMP.html
AutoGluon	Erickson et al.31	https://auto.gluon.ai/stable/index.html
Glmnet	Friedman et al.95	https://cran.r-project.org/web/packages/glmnet/index.html
Planet	Yuan et al.79	https://www.bioconductor.org/packages/release/bioc/html/planet.html
BulkCentileCalc_Global_v8.0.6.2	Gardosi et al.35	https://www.gestation.net/cc/about.htm
Other
Resource website for the DREAM Placental Clock Challenge, including data, software code, and vignettes	This paper	https://www.synapse.org/Synapse:syn59520082

Experimental model and study participant details

Test data generation and processing

Study design, sample collection

The test set data was obtained from a study designed to compare the DNA methylome of placental tissues among various obstetrical syndromes. Samples were obtained from term controls (n = 82), preterm prelabor rupture of membranes (PPROM) (n = 76), preterm and term preeclampsia (PE) (n = 85), preterm labor (PTL) (n = 91), and pregnancies with small-for-gestational-age (SGA) fetuses (n = 50). Patients were enrolled in a study conducted at Wayne State University/Detroit Medical Center / Pregnancy Research Branch (NICHD). Participants provided written informed consent, and the collection and use of the samples and clinical data for research purposes were approved by the Institutional Review Boards of Wayne State University and the Eunice Kennedy Shriver National Institute of Child Health and Human Development, National Institutes of Health, U.S. Department of Health and Human Services.

Clinical definitions

Gestational age (GA) was determined by the self-reported last menstrual period and confirmed by ultrasound. If there was a discrepancy between the menstrual dating and the ultrasound findings that supports redating, GA was determined by the ultrasound biometry.⁹⁶ Preeclampsia was defined as the new onset of elevated blood pressure in the presence of proteinuria after 20 weeks of gestation, following the diagnostic criteria of the American College of Obstetricians and Gynecologists.⁹⁷ Preterm labor was defined as the spontaneous onset of labor, characterized by regular uterine contractions and cervical changes,⁹⁸ that occurs between 20 0/7 weeks of gestation and 36 6/7 weeks of gestation. Preterm prelabor rupture of membranes was diagnosed with a sterile speculum examination that documented pooling of amniotic fluid in the vagina, along with a positive nitrazine test and/or positive ferning tests when necessary.⁹⁹ A small-for-gestational-age (SGA) neonate was defined as having a birthweight percentile less than the 10th percentile for their specific gestational age at delivery by using the standard of Alexander et al.¹⁰⁰

DNA extraction

Genomic DNA was extracted from tissues using the Quick-DNA™ MagBead Plus Kit (Zymo Research, Irvine, CA) automated on the epMotion® 5075 liquid handler (Eppendorf, Enfield, CT) for high throughput sample isolation. Briefly, solid tissues were digested with Proteinase K in DNA Elution Buffer and Biofluid & Solid Tissue Buffer. The samples were centrifuged at ≥10,000 x g with a microcentrifuge for one minute to pellet the debris. The supernatants were transferred to new tubes. The samples were then mixed with Quick-DNA™ MagBinding Buffer and MagBinding Beads. After binding DNA in samples, the beads were pre-washed with Pre-wash buffer and washed with g-DNA Wash Buffer three times. Finally, the gDNA was eluted with DNA Elution Buffer. Purified DNA was quantified using a Qubit 3.0 fluorometer with a Qubit dsDNA BR assay kit (Life Technologies, Carlsbad, CA), according to the manufacturer’s protocol. Purified DNA was then stored at −20°C for the next application.

DNA methylation assays

Samples were sent to the University of Michigan Epigenomics Core for quantitation using the Qubit High Sensitivity dsDNA assay and quality assessment with the TapeStation Genomic DNA kit. For each sample, 250 ng of DNA was bisulfite converted using Zymo’s EZ DNA Methylation kit, following the manufacturer’s incubation parameters specific for Illumina methylation arrays. The cleaned-up samples were then sent to the UM Advanced Genomics Core for hybridization to the Infinium MethylationEPIC BeadChip array, as well as for washing and scanning, according to the manufacturer’s instructions (Illumina EPIC Datasheet). For the subset of patients who were asked and agreed to broad data sharing, the methylation profiles were deposited in the Gene Expression Omnibus (identifier GEO: GSE287219).

Method details

Public data acquisition and processing

Data aggregation

We curated publicly available DNA methylation data from the Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo/) and ArrayExpress (https://www.ebi.ac.uk/biostudies/arrayexpress). Eighteen placental tissue DNA methylation datasets were identified and had gestational age information available.^{76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91} Omics data from nine of the 18 studies were obtained by using the Illumina HumanMethylation 450K BeadChip (450K) platform, and data from the other nine were obtained by using Illumina MethylationEPIC BeadChip (850K) arrays.

Methylation data processing

The raw idat files for each study were downloaded using the R/Bioconductor package, GEOquery,¹⁰¹ and beta values were extracted using the R/Bioconductor package, minfi.⁹² For datasets without available idat files, matrices containing raw methylation (red) and unmethylation (green) signals were downloaded manually and transformed into beta values with minfi. Technical replicates and samples for which gestational age could not be ascertained were excluded. The final dataset comprised 1,842 placental samples: 930 samples were analyzed using the 450K array, and 912 samples were analyzed using the 850K array. The distribution of gestational ages of placentas derived from each study are shown in Figure 1.

Data imputation

For the 930 samples profiled using the HumanMethylation450 BeadChip arrays, we imputed methylation beta values for the 413,000 probes unique to the 850K arrays. The imputation was performed using the CpG impUtation Ensemble (CUE)⁹³ for 339000 probes. The remaining probe data were imputed using the default k-nearest neighbor method implemented in the R/Bioconductor package ChAMP.⁹⁴

DREAM challenge

DREAM challenge organization

The Placental Clock DREAM Challenge was structured as two sub-challenges. In sub-challenge 1 (SC1), participants were tasked with developing a model using only the CpG probes shared between the 450K and 850K arrays to predict gestational age. The aim of sub-challenge 2 (SC2) was to utilize data from the 850k arrays for gestational age prediction. Participants had the option to use the provided imputed data or apply their own imputation methods. Alternatively, they could choose to utilize only data from samples obtained on the 850k arrays.

The challenge had two rounds: leaderboard and final evaluation (Figure 2). The teams were provided with training data composed of publicly available DNA methylation profiles of 1,742 placenta samples. To gain access to the data, participants were required to comply with a data use agreement, restricting use of the data outside of the Challenge and providing guidelines on ethical participation in the Challenge. Participants built their models, dockerized their environment, and submitted their models through the Synapse platform (https://www.synapse.org/Synapse:syn59520082).

In the leaderboard round, models were evaluated on the leaderboard data composed of the publicly available DNA methylation profiles from 100 placenta samples, randomly selected from the publicly available placental DNA methylation datasets generated using the Illumina 850K array. Information on pathology status was not used as a selection criterion, but the final leaderboard set included a mix of normal and complicated pregnancies. Teams were limited to 10 total submissions with the last submission selected as the final submission. Public leaderboards with correlation, RMSE, and MAE were updated immediately after submission. This helped teams to gain immediate feedback on their algorithms and modify them accordingly.

In the final evaluation round, the final submission was evaluated the n = 384 test set generated at Wayne State University / Detroit Medical Center / Pregnancy Research Branch (NICHD)to determine the team’s ranking in each sub-challenge. After the close of the Challenge, models were evaluated in additional post-challenge analyses.

Most Challenge components were supported by the Synapse platform (http://www.synapse.org), including documentation of datasets and tasks, access to the data, submission of models, leaderboard, and the discussion forum. Model evaluations on the leaderboard set and the private WSU test set were performed on a virtual Linux server hosted using the Google GCP platform, which automatically updated the challenge website with RMSE scores obtained on the leaderboard dataset.

Participant engagement

Information about the challenge was disseminated through the DREAM Challenges website (https://dreamchallenges.org) and through social media platforms. Since the test data was not available to participants, they were asked to submit their models in Docker environments to be evaluated by the organizers. The Docker environments contained the necessary instructions to recreate the operating system, software infrastructure, and model parameters for each sub-challenge so that they could be applied to any test set. The organizers provided examples of instructions to build the Docker containers and train models using R, Python, and Julia statistical software. The organizers remained actively engaged with participants through the challenge forum to aid in troubleshooting model submission throughout the challenge.

Approaches of the top 3 performing teams

Team 1 (H.S, E.C.-Y.S.): The best-performing model from Team 1 was for SC1 (450K array CpGs). They preprocessed the data by filtering probes that fell near single nucleotide polymorphisms (SNPs), aligned to multiple locations, targeted X and Y chromosomes, or were not reliably detected in all samples (p-value < 0.01). Beta values were then normalized using beta-mixture quantile normalization.¹⁰² A multistage predictive modeling strategy was developed to estimate GA from placental DNA methylation data. First, for patients included in the training set, the Team 1 extracted all medical conditions included in the public metadata, such as preterm birth, preeclampsia, growth restriction, etc. For individual training dataset and the test dataset where such metadata was not available, the risk that the patient would be affected by each such condition was calculated from DNA methylation scores using random forest models. Then, for each condition associated with the GA at collection (relative to the term control group), an elastic net regression model was trained using beta values of differentially methylated (q ≤0.05) CpG sites with gestational age to predict gestational age. A baseline GA model, trained on samples without any medical conditions, was first used to predict GA for all samples. GA residuals from the baseline model were then predicted for condition-specific subsets using separate models. These residuals, weighted by each condition probabilities, were combined to adjust GA predictions of the baseline GA model. Further details are available at https://www.synapse.org/Synapse:syn62407322/wiki/629409.

Team 2 (T.P., S.H., A.V.B.): The best-performing model from Team 2 was for SC2 (850K array CpGs). Team 2 exclusively used the 812 training samples analyzed on the 850K platform and employed the same preprocessing and normalization steps as Team 1. CpG features were selected by calculating Spearman’s correlation between normalized methylation values and GA, retaining the top 2,000 positively and 2,000 negatively correlated CpG probes. Additionally, 10,827 CpG probes with age-association identified in a prior epigenome-wide association analysis²² were included, resulting in a total of 12,435 probes for the model. Elastic net regression was applied with log-transformed GA as the response variable and selected CpG probes as predictors. Regularization parameters were optimized via 30-fold cross-validation, with alpha set to 0.1 and lambda tuned to address multicollinearity and high-dimensionality in the data. Further details are available at https://www.synapse.org/Synapse:syn61846522/wiki/629109.

Team 3 (I.A., C.W.): CpG feature filtering was performed by Team 3 as described above for Team 1 and 2. The remaining 346,407 probes were used to train the model without any normalisation. Unlike Team 1 and Team 2, no prior knowledge or metadata was used by Team 3. Gestational age prediction was performed using multi-layer perceptron neural networks, also called deep learning. Such models create a projection of the original methylation feature space and use those to regress GA values.³⁰ Several model tuning steps were based on the accuracy estimates provided by the organizers on the leaderboard set. Further details are available at https://www.synapse.org/Synapse:syn61964146/wiki/629445.

Post-DREAM challenge clocks

Wisdom of Crowds placenta Clock: To capitalize on the collective accuracy of top-performing models, we defined the Wisdom of Crowds clock as the average of the predictions from the top three teams in the final rankings. This ensemble approach leveraged the complementary strengths of the best models to improve prediction accuracy.

AutoGluon Clock: To create a baseline model for the DREAM Challenge, we used an automated machine learning (autoML) framework called AutoGluon. Developed by Amazon to democratize ML applications, AutoGluon is a Python-based software. Model training was performed using the data for SC1 utilizing four days of computational time and “best performance” as model fitting options.

Post-Challenge Placental Clock: A Post-Challenge Placental Clock (PCPC) model was also developed using elastic net regression, implemented with the glmnet⁹⁵ package in R. This standard algorithm uses a combination of ridge and lasso regression models, with hyperparameters being fine-tuned via 10-fold cross-validation to balance predictive accuracy and model complexity. The PCPC was trained on all 1,842 publicly available samples (1,742 training + 100 leaderboard), using 450K array CpGs. Data preprocessing involved filtering out probes near single nucleotide polymorphisms (SNPs), those aligned to multiple locations, targeted X and Y chromosomes, or were not reliably detected across all samples (p-value < 0.01). Normalization was conducted using the beta-mixture quantile normalization method¹⁰² implemented with the ChAMP package⁹⁴ in R. This PCPC model aimed to build on insights from the challenge while providing a robust and simple linear regression equation to calculate placental age from methylation beta values of specific CpGs. An R package implementing the PCPC model is available at https://github.com/dw1227/PCPCmodel/.

Quantification and statistical analysis

Assessment of submissions

Performance metrics, including Root Mean Squared Error (RMSE), mean absolute error (MAE), and Pearson correlation coefficient, were calculated on the leaderboard dataset and final test set. For teams who submitted a model for each sub-challenge, the one with lowest RMSE was retained in the final tram ranking.

Robustness analysis of team ranks

The robustness of the final team ranking was evaluated through bootstrap resampling. This approach involved generating 1,000 bootstrap samples of the test and recalculating the RMSE and ranking of each team. The consistency of model rankings across these resamples was assessed by calculating the Bayes factor as the number of times a team’s RMSE outperformed the next-ranked team divided by the number of times the reverse was true.

Analysis of the demographic data

Continuous variables were compared across multiple groups using one-way analysis of variance (ANOVA). Proportions were compared using Fisher’s exact test for contingency tables. A p-value of less than 0.05 was considered statistically significant.

Functional analysis of the CpGs in post challenge placental clock

The CpGs on the EPIC array were mapped to CpG islands, shores, open seas, and regulatory elements such as promoters, enhancers, and long non-coding RNAs using a custom annotation resource that has more annotated probes compared to the default Illumina annotation.¹⁰³ These mappings also included genes whose expression could be regulated by CpG methylation status. Enrichment analysis for the regulatory elements and gene ontology biological processes (MSigDB database, “C5” collection) associated with probes included in the PCPC was performed using a hypergeometric test in R using all CpGs shared between 450k and 850K arrays as the background. p-Values were adjusted using the Benjamini-Hochberg procedure¹⁰⁴ to control the false discovery rate, with a q-value < 0.05 deemed statistically significant.

Analysis of the epigenetic age acceleration in relation to obstetrical disease

The epigenetic ages based on the placental clocks proposed by Lee et al.²² were calculated using the R/Bioconductor package planet⁷⁹ following data normalization.¹⁰² Other epigenetic age estimates, such as those from models developed during the DREAM challenge, were calculated using models as submitted by participants in the challenge. Epigenetic age acceleration was calculated as the difference between the epigenetic age and chronological gestational age. To assess whether epigenetic age acceleration deviated significantly from zero, a one-sample t-test was performed. Associations between eGA acceleration and binary adverse pregnancy outcomes, including PE, PPROM, PTL, and SGA were assessed using logistic regression models. The relationship between eGA acceleration and birth weight percentile (continuous) was examined through linear regression models. Both types of models included the pregnancy outcome as the dependent variable and eGA acceleration as the predictor while also adjusting for relevant maternal characteristics and fetal sex. Models were fit separately for term and preterm samples. To account for potential heterogeneity in the associations, interaction terms between eGA acceleration and covariates were included when significant. A p-value of less than 0.05 was considered statistically significant. All statistical analyses were performed using R statistical software.

Published: July 23, 2025