Epigenome-wide cross-tissue predictive modeling and comparison of cord blood and placental methylation in a birth cohort

Abstract

Aim: We compared predictive modeling approaches to estimate placental methylation using cord blood methylation.

Materials & methods: We performed locus-specific methylation prediction using both linear regression and support vector machine models with 174 matched pairs of 450k arrays.

Results: At most CpG sites, both approaches gave poor predictions in spite of a misleading improvement in array-wide correlation. CpG islands and gene promoters, but not enhancers, were the genomic contexts where the correlation between measured and predicted placental methylation levels achieved higher values. We provide a list of 714 sites where both models achieved an R² ≥0.75.

Conclusion: The present study indicates the need for caution in interpreting cross-tissue predictions. Few methylation sites can be predicted between cord blood and placenta.

Keywords: 450k arrays; DNA methylation; cord blood; epigenetics; methylation prediction; placenta; support vector machine.

Figure 1.. Prediction accuracy of both linear models and support vector machine models was evaluated by using R² and root mean square error between measured and predicted methylation values.

For both models, the distribution of R² (A) and root mean square error (B) in all CpG sites are shown. The filled square and triangle represent the mean. LM: Linear model; RMSE: Root mean square error; SVM: Support vector machine.

Figure 2.. R² and root mean square error distributions stratified by annotation category in both linear models and support vector machine models.

The categories were designated using Illumina annotation. The filled square and triangle represent the mean. LM: Linear regression model; RMSE: Root mean square error; SVM: Support vector machine.

Figure 3.. Scatter plots between measured placental methylation values (y-axis) and both support vector machine-predicted placenta (blue) and cord blood (red) methylation values (x-axis) in three CpG sites where the R² achieves, respectively, high (A) and low (B) values.

Above each plot is the corresponding CpG site’s Illumina ID and the value of the R² between measured and predicted placental methylation values obtained by tenfold cross-validation. SVM: Support vector machine.

Figure 4.. R² between measured and support vector machine-predicted placental methylation levels in terms of the average (A) and the standard deviation (B) of both cord blood and placental methylation values across all subjects.

Each hexagon bin’s color corresponds to the mean of the R² of all the CpG sites contained in that bin and the overlaid red contour lines show the density of the set of analyzed sites. SVM: Support vector machine.