Penalized Principal Component Analysis Using Smoothing.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Principal Component Analysis (PCA) is standard for genomic data dimensionality reduction and population stratification correction.
• Traditional PCA may lack sparseness and efficiency for complex genomic datasets.
• Penalized Eigenvalue Problem (PEP) offers an optimization-based approach to compute sparse eigenvectors.

Purpose of the Study:

• To extend the Penalized Eigenvalue Problem (PEP) by incorporating smoothing to the L1 penalty.
• To enable efficient computation of analytical gradients for faster optimization.
• To demonstrate the utility of smoothed penalized eigenvectors in genomic data analysis.

Main Methods:

• Developed a smoothed L1 penalty for the PEP, allowing analytical gradient computation.
• Extended PEP to compute higher-order eigenvectors using Singular Value Decomposition (SVD) principles.
• Conducted four experimental studies, including analysis of the 1000 Genomes Project dataset, polygenic risk score computation, and clustering.

Main Results:

• The smoothed PEP demonstrated increased numerical stability and produced meaningful eigenvectors on the 1000 Genomes dataset.
• Replacing standard penalized eigenvectors with smoothed versions improved prediction accuracy in polygenic risk scores.
• Smoothed PEP enhanced the discernibility of clusters in clustering applications and showed competitive performance against state-of-the-art sparse PCA algorithms.

Conclusions:

• Smoothed PEP provides a numerically stable and efficient method for computing sparse eigenvectors in genomic data.
• The proposed method offers practical advantages in downstream applications such as risk prediction and data clustering.
• Smoothed PEP represents a valuable advancement in sparse dimensionality reduction techniques for bioinformatics.