Ultra-high dimensional variable selection with application to normative aging study: DNA methylation and metabolic syndrome

Grace Yoon, Yinan Zheng, Zhou Zhang, Haixiang Zhang, Tao Gao, Brian Joyce, Wei Zhang, Weihua Guan, Andrea A. Baccarelli, Wenxin Jiang, Joel Schwartz, Pantel Vokonas, Lifang Hou, Lei LiuBMC Bioinformatics 2017 · DOI

TL;DR

This paper tackles the challenge of identifying key DNA methylation markers associated with metabolic syndrome from an ultra-high dimensional dataset. They introduce an enhanced variable selection method (ISIS with elastic net) that significantly improves accuracy and reduces false positives compared to standard approaches, leading to the identification of four promising CpG markers linked to biologically relevant genes.

What Problem Does This Paper Address?

Metabolic syndrome is a growing global health concern, and understanding its biological underpinnings is crucial for prevention and treatment. DNA methylation, an epigenetic mechanism that regulates gene expression, is a promising area of research for uncovering these links. However, analyzing DNA methylation data presents a significant challenge: it's "ultra-high dimensional" (many more potential markers than samples) and the markers are often highly correlated. Traditional statistical methods struggle with this complexity, leading to unreliable results and a high rate of false positives. This paper aims to develop and apply a robust method to navigate this high-dimensional, correlated data landscape and pinpoint specific DNA methylation markers associated with metabolic syndrome.

Methods

The core of this paper lies in their proposed methodology: an Iterative Sure Independence Screening (ISIS) method combined with an elastic net penalty.

  • ISIS: This is a two-step approach. First, a "screening" step uses a simplified measure (like marginal correlation) to quickly identify a manageable subset of potentially relevant variables from the vast initial set. This is crucial for reducing computational burden and noise. The "iterative" aspect means this screening process can be repeated to refine the selection.
  • Elastic Net: This is a penalized regression method that combines the strengths of L1 (LASSO) and L2 (ridge) regularization. LASSO is known for its ability to perform variable selection (setting some coefficients to zero), while ridge regression is good at handling correlated predictors by shrinking their coefficients. The elastic net offers a balance, providing sparsity like LASSO while being more stable with correlated features.
  • Binomial Model: They use a binomial model to predict the metabolic syndrome index, which is treated as an ordinal variable (0 to 5, representing the number of satisfied conditions).

The novelty lies in the synergistic application of ISIS for initial ultra-high dimensional screening followed by elastic net for refined variable selection, specifically tailored to the characteristics of DNA methylation data. They also incorporate bootstrap resampling to assess the stability and reliability of their selected markers.

Datasets

The primary dataset used is the Normative Aging Study (NAS) data.

  • NAS Data: This longitudinal cohort study provided DNA methylation levels from 484,548 CpG sites across 659 subjects. The data was collected from buffy coat samples and processed using the Illumina 450K BeadChip array.
  • Simulation Data: To rigorously evaluate their method, they also generated synthetic data that mimicked the correlation structure of the real NAS methylation data, allowing for controlled testing of their approach.

Key Insights and Ideas

  • Screening is Key for Ultra-High Dimensionality: The paper strongly demonstrates that an initial screening step (like ISIS) is essential for improving the performance of variable selection methods (like elastic net) when dealing with ultra-high dimensional data. It effectively filters out noise and reduces the burden on the subsequent, more computationally intensive, selection process.
  • Reducing False Positives: The ISIS approach significantly reduces the number of false positive markers identified compared to using elastic net alone. This is critical for generating reliable biological insights and avoiding wasted effort on spurious associations.
  • Biologically Relevant Findings: The identified CpG markers are not just statistically significant; they are mapped to genes (ABCG1 and DHCR24) that have known roles in lipid metabolism, cholesterol synthesis, and metabolic pathways, lending strong biological plausibility to their findings.
  • Robustness of the Method: The results are consistent across different choices of the screening parameter 'd' and across multiple bootstrap samples, indicating the robustness of their ISIS-elastic net approach.

Key Results

  • Identification of Four Key CpGs: The proposed method identified four CpG sites that were strongly associated with the metabolic syndrome index in the NAS data.
  • Improved AUC: Using 5-fold cross-validation, their ISIS-elastic net method achieved higher AUC (Area Under the Curve) values for predicting metabolic syndrome (both binary and multiclass definitions) compared to the elastic net method alone, even with a significantly smaller number of selected variables. For instance, the average multiclass AUC was 0.6441 for ISIS (d=25) versus 0.6219 for elastic net only.
  • Reduced False Positives in Simulation: In simulations, the ISIS method correctly identified true significant coefficients with high frequency while substantially reducing the selection of redundant (false positive) markers compared to elastic net alone.

Limitations and Future Work

  • Study Population: The current analysis is limited to white male subjects in the NAS study.
  • Validation: The authors suggest validating their findings in other cohorts, including the Coronary Artery Risk Development in Young Adults Study (CARDIA), and examining diverse populations.
  • Longitudinal Data: The NAS data includes repeated measurements over time. The paper notes that their method could be extended to longitudinal data analysis.
  • Mediation Analysis: They express interest in using mediation analysis to determine if DNA methylation mediates the relationship between interventions (like diet or exercise) and health outcomes.
  • Adaptive Tuning: The paper suggests that an adaptive choice of the tuning parameter 'c' for the elastic net, guided by cross-validation, could lead to improved performance, especially for smaller sample sizes.

Summary updated

BibTeX
@article{yoon2017ultra,
  author = {Grace Yoon and Yinan Zheng and Zhou Zhang and Haixiang Zhang and Tao Gao and Brian Joyce and Wei Zhang and Weihua Guan and Andrea A. Baccarelli and Wenxin Jiang and Joel Schwartz and Pantel Vokonas and Lifang Hou and Lei Liu},
  title = {{Ultra-high dimensional variable selection with application to normative aging study: DNA methylation and metabolic syndrome}},
  journal = {BMC Bioinformatics},
  year = {2017},
  doi = {10.1186/s12859-017-1568-1},
}
Figure