Efficient toolkit implementing best practices for principal component analysis of population genetic data

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Principal Component Analysis (PCA) is crucial for inferring ancestry and controlling population structure in genetic analyses.
• However, PCA is prone to pitfalls such as capturing Linkage Disequilibrium (LD) instead of population structure, shrinkage bias in projected Principal Components (PCs), outlier detection, and handling uneven population sizes.

Purpose of the Study:

• To explore common pitfalls in PCA of genetic data.
• To present efficient solutions and best practices for addressing these issues.
• To provide user-friendly R package implementations for these solutions.

Main Methods:

• Applied PCA to UK Biobank and 1000 Genomes Project datasets.
• Developed and utilized an algorithm for removing long-range LD regions.
• Investigated shrinkage bias in projected genotypes.
• Implemented solutions in R packages bigsnpr and bigutilsr.

Main Results:

• Identified PCs 19-40 in UK Biobank capturing LD structure, not population structure.
• Recovered 16 PCs capturing only population structure after LD region removal.
• Recommended using 16-18 PCs for UK Biobank population structure analysis.
• Demonstrated significant shrinkage bias in projected genotypes from PC5 onwards in 1000 Genomes data.
• Showcased efficient, unbiased genotype projection using bigsnpr.

Conclusions:

• PCA pitfalls in genetic data analysis are significant but addressable.
• The proposed methods and R package implementations offer efficient solutions for accurate population structure inference.
• Recommendations are provided for best practices in PCA for genetic and other omics data analysis.