Principal component analysis of binary genomics data

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• High-dimensional binary data from genome-wide genetic and epigenetic analyses are increasingly common.
• Classical Principal Component Analysis (PCA) is not directly suitable for the unique mathematical properties of binary data.
• Existing PCA alternatives for binary data are underutilized in bioinformatics.

Purpose of the Study:

• Introduce parametric and nonparametric PCA versions tailored for binary data.
• Evaluate the performance of these methods using simulations and real genomic datasets.
• Provide guidance on selecting appropriate PCA methods for binary genomic data.

Main Methods:

• Exploration of parametric logistic PCA models.
• Application of nonparametric Gifi PCA models.
• Validation using simulated binary data and GDSC1000 cancer genomic data (mutation, CNA, methylation).

Main Results:

• Most evaluated PCA methods successfully identify low-dimensional structures when present in binary data.
• Performance was assessed on component number selection, overfitting, structure recovery, and variable importance.
• Parametric logistic PCA is recommended for data with an assumed probabilistic generating process.

Conclusions:

• The nonparametric Gifi model is suitable when no probabilistic generating process is assumed for binary data.
• Both logistic PCA and Gifi models demonstrate effectiveness in analyzing high-dimensional binary genomic data.
• The study facilitates better exploration of complex genomic alterations using adapted PCA techniques.