Tracing Sub-Structure in the European American Population with PCA-Informative Markers

Peristera Paschou, Petros Drineas, Jamey Lewis, Caroline M. Nievergelt, Deborah A. Nickerson, Joshua D. Smith, Paul M. Ridker, Daniel I. Chasman, Ronald M. Krauss, Elad ZivPLoS 2008 · DOI

Abstract

Genetic structure in the European American population reflects waves of migration and recent gene flow among different populations. This complex structure can introduce bias in genetic association studies. Using Principal Components Analysis (PCA), we analyze the structure of two independent European American datasets (1,521 individuals-307,315 autosomal SNPs). Individual variation lies across a continuum with some individuals showing high degrees of admixture with non-European populations, as demonstrated through joint analysis with HapMap data. The CEPH Europeans only represent a small fraction of the variation encountered in the larger European American datasets we studied. We interpret the first eigenvector of this data as correlated with ancestry, and we apply an algorithm that we have previously described to select PCA-informative markers (PCAIMs) that can reproduce this structure. Importantly, we develop a novel method that can remove redundancy from the selected SNP panels and show that we can effectively remove correlated markers, thus increasing genotyping savings. Only 150-200 PCAIMs suffice to accurately predict fine structure in European American datasets, as identified by PCA. Simulating association studies, we couple our method with a PCA-based stratification correction tool and demonstrate that a small number of PCAIMs can efficiently remove false correlations with almost no loss in power. The structure informative SNPs that we propose are an important resource for genetic association studies of European Americans. Furthermore, our redundancy removal algorithm can be applied on sets of ancestry informative markers selected with any method in order to select the most uncorrelated SNPs, and significantly decreases genotyping costs.

Introduction

In Genome-Wide Association Studies (GWAS), researchers search for genetic variants that appear more frequently in individuals with a specific disease. However, these studies are often confounded by population stratification. If a "case" group happens to have a different ancestral background than the "control" group, any genetic variant that naturally varies by ancestry will appear to be a "disease gene," even if it has no biological link to the illness.

While European Americans are often treated as a genetically homogenous group, they are actually a "mosaic" reflecting centuries of migration. Previous methods to correct for this required either massive datasets (300,000+ markers) or prior knowledge of an individual's specific pedigree. This paper identifies a "middle way": a tiny, highly informative panel of markers that can capture this fine-scale structure without needing to know a person's family history beforehand.

Methods

The authors utilized an unsupervised approach based on Principal Component Analysis (PCA). PCA identifies the primary axes of genetic variation in a dataset; in this case, the first "eigenvector" largely corresponds to a North-to-South European ancestry gradient.

The core technical innovation is a two-step selection process:

  1. PCA-Informative Markers (PCAIMs): A scoring algorithm identifies SNPs that correlate most strongly with the top principal components.
  2. Redundancy Removal (Greedy QR): Standard selection methods often pick SNPs that are physically close together or highly correlated (in Linkage Disequilibrium). To fix this, the authors framed the selection as a Column Subset Selection Problem (CSSP). They applied a "greedy QR" decomposition algorithm to filter out redundant markers, ensuring each selected SNP provides unique information.

Datasets

The study analyzed two independent European American datasets totaling 1,521 individuals and 307,315 autosomal SNPs:

  • CHORI Dataset: 980 individuals from clinical trials (CAP and PRINCE studies).
  • CORIELL Dataset: 541 individuals, including patients with Parkinson’s and healthy controls.
  • HapMap Phase II: Used as a reference for African (YRI), European (CEU), and East Asian (CHB/JPT) populations to identify outliers and orient the ancestry axes.

Key Insights and Results

  • The "Continuum" of Ancestry: European American genetic structure isn't a set of discrete clusters; it’s a continuum. Most variation exists along a single axis (likely North-to-South).
  • Reference Limitations: The standard "CEPH" European reference samples represent only a tiny fraction of the actual variation encountered in the larger European American datasets studied.
  • High Accuracy with Few Markers: Just 150–200 non-redundant PCAIMs were enough to predict an individual's position on the ancestry axis with a correlation coefficient of >0.9 compared to the full 300,000 SNP set.
  • Stratification Correction: In simulated association studies, using these ~200 markers to correct for ancestry effectively eliminated false positives (spurious correlations) with almost no loss in statistical power.
  • Cross-Validation: Markers selected from the CHORI dataset performed equally well when applied to the CORIELL dataset, proving the markers are robust across different cohorts.

Limitations and Future Work

While the North-South axis is the dominant signal, other subtle structures (such as Ashkenazi Jewish ancestry) exist and might require additional principal components to capture fully. Furthermore, the specific SNPs identified are optimized for European Americans; different panels would be needed for other admixed populations (e.g., Hispanic Americans or African Americans).

Future work could involve expanding these panels to account for more "axes" of variation and integrating them into standard low-cost genotyping arrays for clinical research. The redundancy removal algorithm can be applied to ancestry-informative markers selected with any method to significantly decrease genotyping costs.

Summary updated

Abstract

Genetic structure in the European American population reflects waves of migration and recent gene flow among different populations. This complex structure can introduce bias in genetic association studies. Using Principal Components Analysis (PCA), we analyze the structure of two independent European American datasets (1,521 individuals-307,315 autosomal SNPs). Individual variation lies across a continuum with some individuals showing high degrees of admixture with non-European populations, as demonstrated through joint analysis with HapMap data. The CEPH Europeans only represent a small fraction of the variation encountered in the larger European American datasets we studied. We interpret the first eigenvector of this data as correlated with ancestry, and we apply an algorithm that we have previously described to select PCA-informative markers (PCAIMs) that can reproduce this structure. Importantly, we develop a novel method that can remove redundancy from the selected SNP panels and show that we can effectively remove correlated markers, thus increasing genotyping savings. Only 150-200 PCAIMs suffice to accurately predict fine structure in European American datasets, as identified by PCA. Simulating association studies, we couple our method with a PCA-based stratification correction tool and demonstrate that a small number of PCAIMs can efficiently remove false correlations with almost no loss in power. The structure informative SNPs that we propose are an important resource for genetic association studies of European Americans. Furthermore, our redundancy removal algorithm can be applied on sets of ancestry informative markers selected with any method in order to select the most uncorrelated SNPs, and significantly decreases genotyping costs.
BibTeX
@inproceedings{paschou2008tracing,
  author = {Peristera Paschou and Petros Drineas and Jamey Lewis and Caroline M. Nievergelt and Deborah A. Nickerson and Joshua D. Smith and Paul M. Ridker and Daniel I. Chasman and Ronald M. Krauss and Elad Ziv},
  title = {{Tracing Sub-Structure in the European American Population with PCA-Informative Markers}},
  booktitle = {PLoS},
  year = {2008},
  doi = {10.1371/journal.pgen.1000114},
}
Figure