Fast and robust ancestry prediction using principal component analysis.

Area of Science:

• Genetics
• Bioinformatics
• Statistical genomics

Background:

• Population stratification (PS) is a major confounder in genome-wide association studies (GWAS), potentially leading to false-positive associations.
• Principal component analysis (PCA)-based ancestry prediction is commonly used to adjust for PS.
• Existing methods like Simple Projection (SP) can be biased, while Data Augmentation, Decomposition and Procrustes (ADP) transformations are computationally expensive.

Purpose of the Study:

• To develop and propose computationally efficient and unbiased methods for predicting principal component (PC) scores for population stratification adjustment in GWAS.
• To address the limitations of existing methods, namely bias in SP and high computational cost in ADP.

Main Methods:

• Developed bias-adjusted projection (AP) using random matrix theory to estimate and correct SP bias.
• Developed online ADP (OADP) using an efficient online singular value decomposition algorithm to reduce ADP's computational cost.
• Implemented and compared AP and OADP against SP and ADP using extensive simulations and real-world data (UK Biobank).

Main Results:

• AP and OADP were shown to be unbiased in extensive simulation studies.
• OADP and AP demonstrated significant computational speed improvements, being 16-16,000 times faster than ADP.
• Application to UK Biobank data showed AP and OADP required substantially less CPU time (0.82 and 21 hours) compared to the projected ADP time (1628 hours).
• SP exhibited bias in inferring sub-European ancestry, unlike the proposed AP and OADP methods.

Conclusions:

• AP and OADP provide accurate and computationally efficient solutions for population stratification adjustment in large-scale genetic studies.
• These methods overcome the limitations of existing techniques, offering improved performance and reduced computational burden.
• The developed methods are available in the open-source Python software FRAPOSA.