PPubMed16 May 2020Cited 171×

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

Florian Privé, Keurcien Luu, Michael G B Blum, John J McGrath, Bjarni J Vilhjálmsson, Russell Schwartz

Abstract

In this work, we have compiled different pitfalls that can arise with PCA of genetic data. Then, we have investigated possible solutions to these pitfalls and selected the ones that we found most advantageous, both with respect to properties such as accuracy and robustness, but also computational efficiency and ease of use. We then implemented these solutions in R packages bigsnpr and bigutilsr. The new functions we provide in R package bigsnpr can be directly applied to genotypes stored as PLINK bed/bim/fam files with some missing values. This contrasts with previous releases of package bigsn

A figure from Efficient toolkit implementing best practices for principal component analysis of population genetic data — fig. from the paper

§ The Valyu brief

The paper, in five minutes

Read from the full text via Valyu Contents

Efficient R toolkit (bigsnpr/bigutilsr) provides fast, bias‑corrected PCA for large genetic datasets and best‑practice recommendations.

What’s new

i.Identified common pitfalls in PCA of genetic data and gave best‑practice recommendations
ii.Implemented fast, memory‑efficient PCA, LD‑thinning, outlier detection, and bias‑adjusted projection in the R packages bigsnpr and bigutilsr
iii.Developed an automatic algorithm to remove long‑range LD regions and recover population‑structure PCs
iv.Demonstrated the methods on UK Biobank and 1000 Genomes, showing improved PC selection and unbiased projections

How they did it

The authors derived truncated PCA and related functions (LD thinning, robust Mahalanobis distance, outlier detection, bias‑adjusted projection) and implemented them in the R packages bigsnpr (v1.0.0) and bigutilsr, usable directly on PLINK bed/bim/fam files with some missing values. They applied these tools to the UK Biobank and 1000 Genomes datasets using parallel computation and iterative LD‑region removal to obtain population‑structure PCs and unbiased projections.

What they found

After removing long‑range LD, only 16 PCs capture population structure in the UK Biobank, leading to a recommendation of using 16‑18 PCs for confounding control
Shrinkage bias in simple projection becomes large from PC5 onward; OADP projection reduces shrinkage to ≤1.08 for the first 20 PCs in the UK Biobank (Table 2)
Computing the first 20 PCs on 261 307 variants from 406 545 UK Biobank individuals took 34 min; the full LD‑removal pipeline required ~5 h and 5 iterations
Subsampling 48 942 diverse‑ancestry UK Biobank individuals allowed recovery of ≥40 PCs showing clear population structure, compared to only 16‑18 PCs in the full sample

Caveats

✕bigsnpr requires its own data format and cannot handle missing values without conversion/imputation
✕Simple projection of new individuals is shrinkage‑biased
✕OADP projection does not work for related individuals
✕Outlier detection based on ‘6 SDs from the mean’ assumes a Gaussian distribution

§ The Valyu brief

Reading the full paper and taking notes. This takes a few seconds…

§ Ask this paper

Ask a question about this paper

Valyu reads the full text and answers from what the paper actually says.

§ Related work in other archives

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

The paper, in five minutes

Ask a question about this paper

Similar papers in all four archives

Similar papers in all four archives

The paper, in five minutes