Guidelines on optimizing DNA methylation reference panels for cell-type deconvolution | JoVE Visualize

Area of Science:

Epigenetics
Bioinformatics
Computational Biology

Background:

Accurate cell-type deconvolution is essential for interpreting Epigenome-Wide Association Studies (EWAS).
Estimating cell-type fractions typically relies on DNA methylation (DNAm) reference panels derived from sorted or single-cell data.
Two primary methods exist for constructing these reference panels: machine learning and effect size/cell-type specificity optimization.

Purpose of the Study:

To compare the performance of machine learning versus effect size optimization for building DNA methylation reference panels.
To identify optimal strategies for constructing DNA methylation reference panels for improved cell-type deconvolution.
To provide guidelines for future DNA methylation reference panel development.

Main Methods:

Demonstrated the superiority of an effect size and cell-type specificity optimization approach over standard machine learning for reference panel construction.
Evaluated panel performance on independent datasets, highlighting overfitting issues with machine learning due to small sorted sample sizes.
Compared DNA methylation reference panels built from hypomethylated versus hypermethylated markers for adult blood cell types.

Main Results:

The effect size optimization method demonstrated superior performance and reduced overfitting compared to machine learning approaches.
DNA methylation reference panels built using cell-type specific hypomethylated markers resulted in improved estimation of cell-type fractions in adult blood.
Standard machine learning models tended to overfit and underperform on independent data when trained on limited sorted samples.

Conclusions:

The effect size optimization approach is preferable for building DNA methylation reference panels due to its robustness against overfitting.
Utilizing hypomethylated markers in adult blood panels enhances the accuracy of cell-type fraction estimation.
These findings offer critical guidelines for developing more accurate DNA methylation reference panels and include an optimized panel-building function in the EpiDISH R-package.