Evidence for the placenta-brain axis: multi-omic kernel aggregation predicts intellectual and social impairment in children born extremely preterm

Abstract

Background: Children born extremely preterm are at heightened risk for intellectual and social impairment, including Autism Spectrum Disorder (ASD). There is increasing evidence for a key role of the placenta in prenatal developmental programming, suggesting that the placenta may, in part, contribute to origins of neurodevelopmental outcomes.

Methods: We examined associations between placental transcriptomic and epigenomic profiles and assessed their ability to predict intellectual and social impairment at age 10 years in 379 children from the Extremely Low Gestational Age Newborn (ELGAN) cohort. Assessment of intellectual ability (IQ) and social function was completed with the Differential Ability Scales-II and Social Responsiveness Scale (SRS), respectively. Examining IQ and SRS allows for studying ASD risk beyond the diagnostic criteria, as IQ and SRS are continuous measures strongly correlated with ASD. Genome-wide mRNA, CpG methylation and miRNA were assayeds with the Illumina Hiseq 2500, HTG EdgeSeq miRNA Whole Transcriptome Assay, and Illumina EPIC/850 K array, respectively. We conducted genome-wide differential analyses of placental mRNA, miRNA, and CpG methylation data. These molecular features were then integrated for a predictive analysis of IQ and SRS outcomes using kernel aggregation regression. We lastly examined associations between ASD and the multi-omic-predicted component of IQ and SRS.

Results: Genes with important roles in neurodevelopment and placental tissue organization were associated with intellectual and social impairment. Kernel aggregations of placental multi-omics strongly predicted intellectual and social function, explaining approximately 8% and 12% of variance in SRS and IQ scores via cross-validation, respectively. Predicted in-sample SRS and IQ showed significant positive and negative associations with ASD case-control status.

Limitations: The ELGAN cohort comprises children born pre-term, and generalization may be affected by unmeasured confounders associated with low gestational age. We conducted external validation of predictive models, though the sample size (N = 49) and the scope of the available out-sample placental dataset are limited. Further validation of the models is merited.

Conclusions: Aggregating information from biomarkers within and among molecular data types improves prediction of complex traits like social and intellectual ability in children born extremely preterm, suggesting that traits within the placenta-brain axis may be omnigenic.

Keywords: Differential expression analysis; Epigenome-wide association; Multi-omic aggregation; Placental gene regulation; Prenatal neurodevelopmental programming; Social and cognitive impairment.

Fig. 1

Scheme for kernel aggregation and prediction models. (1) Design matrices for CpG sites, mRNAs, and miRNAs are aggregated to form a linear or Gaussian kernel matrix that measures the similarity of samples. (2) Clinical variables are regressed out of the outcomes IQ and SRS and from the omic kernels to limit influence from these variables. (3) Using 50-fold Monte Carlo cross-validation on 75–25% training-test splits, we train prediction models with the kernel matrices for IQ and SRS in the training set and predict in the test sets. Prediction is assessed in every fold with adjusted $R^2$ and averaged for an overall prediction metric

Fig. 2

Associations between SRS, IQ, and ASD and with clinical variables. a Scatter plot of SRS (X-axis) and IQ (Y-axis) colored by ASD case (orange) and control (blue) status. b Boxplots of SRS and IQ across ASD case–control status. P value from a two-sample Mann–Whitney test is provided. c Caterpillar plot of multivariable linear regression parameters of IQ and SRS using clinical variables. Points give the regression parameter estimates with error bars showing the 95% FDR-adjusted confidence intervals [48]. The null value of 0 is provided for reference with the dotted line

Fig. 3

In-sample predictive performance of kernel models. a Adjusted mean $R^2$ (Y-axis) of best kernel models over various numbers of the top biomarkers (X-axis) in the CpG (dark blue), miRNA (orange), and mRNA (light blue) omics over 50 Monte Carlo folds. The X-axis scale is logarithmic. b Bar plots of adjusted mean $R^2$ (Y-axis) for optimally tuned kernel predictive models using all combinations of omics (X-axis) over 50 Monte Carlo folds. The error bar gives a spread of one standard deviation around the mean adjusted $R^2$

Fig. 4

Association of ASD case/control status with predicted SRS and IQ. a Box-plots of in-sample predicted IQ (left) and SRS (right) over ASD case/control in ELGAN over tenfold cross-validation. b Box-plots of out-sample predicted IQ (left) and SRS (right) over ASD case/control in MARBLES external validation dataset. $P$-values presented as from a Mann–Whitney test of differences across the ASD case/control groups