Modeling dependency structures in 450k DNA methylation data

MinLinMo: a minimalist approach to variable selection and linear model prediction

Generating prediction models from high dimensional data often result in large models with many predictors. Causal inference for such models can therefore be difficult or even impossible in practice. The stand-alone software package MinLinMo emphasizes small linear prediction models over highest possible predictability with a particular focus on including variables correlated with the outcome, minimal memory usage and speed. MinLinMo is demonstrated on large epigenetic datasets with prediction models for chronological age, gestational age, and birth weight comprising, respectively, 15, 14 and 10 predictors. The parsimonious MinLinMo models perform comparably to established prediction models requiring hundreds of predictors.

Stability selection enhances feature selection and enables accurate prediction of gestational age using only five DNA methylation sites

BACKGROUND

DNA methylation (DNAm) is robustly associated with chronological age in children and adults, and gestational age (GA) in newborns. This property has enabled the development of several epigenetic clocks that can accurately predict chronological age and GA. However, the lack of overlap in predictive CpGs across different epigenetic clocks remains elusive. Our main aim was therefore to identify and characterize CpGs that are stably predictive of GA.

RESULTS

We applied a statistical approach called 'stability selection' to DNAm data from 2138 newborns in the Norwegian Mother, Father, and Child Cohort study. Stability selection combines subsampling with variable selection to restrict the number of false discoveries in the set of selected variables. Twenty-four CpGs were identified as being stably predictive of GA. Intriguingly, only up to 10% of the CpGs in previous GA clocks were found to be stably selected. Based on these results, we used generalized additive model regression to develop a new GA clock consisting of only five CpGs, which showed a similar predictive performance as previous GA clocks (R2 = 0.674, median absolute deviation = 4.4 days). These CpGs were in or near genes and regulatory regions involved in immune responses, metabolism, and developmental processes. Furthermore, accounting for nonlinear associations improved prediction performance in preterm newborns.

CONCLUSION

We present a methodological framework for feature selection that is broadly applicable to any trait that can be predicted from DNAm data. We demonstrate its utility by identifying CpGs that are highly predictive of GA and present a new and highly performant GA clock based on only five CpGs that is more amenable to a clinical setting.