Performance evaluation of dimensionality reduction techniques on high-dimensional DNA methylation data.

Area of Science:

• Epigenetics
• Bioinformatics
• Computational Biology

Background:

• DNA methylation (DNAm) is a crucial epigenetic modification.
• DNAm datasets are inherently high-dimensional, posing analytical challenges.
• Effective dimension reduction (DR) is vital for analyzing complex epigenetic data.

Purpose of the Study:

• To evaluate and compare the performance of various DR techniques on DNA methylation datasets.
• To identify optimal DR methods for preserving information and structure in DNAm data.
• To guide the selection of DR methods for epigenome-wide association studies.

Main Methods:

• Leveraged DNAm data from the STRiDE prospective study (258 pregnant women, 862,927 CpG sites).
• Applied multiple DR techniques including PCA, MDS, PLS-DA, ISOMAP, and UMAP.
• Assessed DR performance using Shannon entropy, local-neighborhood metrics (König's measure, Spearman's ρ, trustworthiness/continuity), and global-structure metrics (Kruskal stress, Sammon's stress, residual variance).

Main Results:

• Multidimensional Scaling (MDS) and Principal Component Analysis (PCA) demonstrated superior performance in preserving information and structure.
• Partial Least Squares Discriminant Analysis (PLS-DA) showed competitive results, while ISOMAP yielded moderate outcomes.
• Uniform Manifold Approximation and Projection (UMAP) performed poorly, exhibiting higher entropy and greater structural distortion.

Conclusions:

• PCA and MDS are recommended as the most effective DR techniques for DNA methylation datasets.
• The choice of DR method significantly impacts the analysis of epigenetic data, influencing information and structure preservation.
• Findings provide crucial insights for optimizing bioinformatics pipelines in epigenetics research.