Ancestry informative markers for fine-scale individual assignment to worldwide populations

Peristera Paschou, Joshua P. Lewis, Asif Javed, Petros DrineasJournal of Medical Genetics 2010 · DOI

TL;DR

This paper introduces a hierarchical decision tree approach for fine-scale human ancestry inference using genetic markers. They demonstrate that a small, carefully selected set of Ancestry Informative Markers (AIMs) can achieve nearly perfect accuracy in assigning individuals to their populations of origin, significantly reducing the number of markers needed compared to whole-genome analysis.

What Problem Does This Paper Address?

Understanding human population structure and individual ancestry is crucial for fields like evolutionary biology, medical genetics, and forensics. While whole-genome analysis can reveal fine-scale genetic differences, it's computationally expensive and requires a vast number of genetic markers (SNPs). The challenge is to identify a minimal set of "ancestry informative markers" (AIMs) that can accurately predict an individual's ancestry at a fine-grained level, mimicking the resolution of whole-genome data but with significantly reduced cost and complexity. Previous work had identified AIMs, but a systematic evaluation of their performance across diverse global populations and a method for reducing redundancy within these marker sets were lacking.

Methods

The authors developed a hierarchical decision tree framework to decompose the complex problem of worldwide population structure into manageable steps. At each node of the tree, they propose a panel of AIMs to differentiate between populations. Two primary methods were used to select AIMs:

  • PCA-correlated SNPs (PCAIMs): These are SNPs that strongly correlate with the top Principal Components (PCs) derived from Principal Component Analysis (PCA) of the entire dataset. PCA is a standard technique for dimensionality reduction and revealing population structure.
  • Informativeness for Assignment (In) metric: This is a supervised metric that directly quantifies how informative a marker is for assigning individuals to predefined populations, based on allele frequencies and self-reported ancestry.

To address redundancy in the selected AIMs (often due to linkage disequilibrium), they employed a clustering-based redundancy removal algorithm. For ancestry prediction, they used a k-Nearest Neighbors (k-NN) classification approach, specifically 5-NN.

Datasets

The primary dataset used was the Human Genome Diversity Panel (HGDP), which contains genotype data for approximately 650,000 SNPs from 1043 individuals across 51 populations, categorized into seven broad geographic regions.

A second dataset, the HapMap Phase 3 database, was used for validation, including populations not well-represented in the HGDP, such as admixed populations and those with limited reference data.

Key Insights and Ideas

  • Hierarchical approach is powerful: Decomposing the global population structure into a decision tree significantly simplifies the problem of ancestry inference and allows for tailored marker selection at different levels of resolution.
  • Small panels of AIMs are highly effective: Contrary to the intuition that whole-genome data is necessary for fine-scale ancestry, a small, judiciously selected set of AIMs can achieve near-perfect accuracy.
  • Redundancy removal is crucial: Clustering-based methods effectively reduce the number of AIMs needed without sacrificing accuracy, making the panels more interpretable and efficient.
  • PCAIMs are robust: PCA-correlated SNPs (PCAIMs) proved to be a reliable and efficient method for selecting AIMs, performing well across various populations and levels of the decision tree.
  • Supervised metrics can overfit: While "In" metrics showed good performance in initial validation, they tended to overfit the training data, leading to poorer generalization in cross-validation compared to PCAIMs.

Key Results

  • High Accuracy with Reduced Markers: In most cases, ancestry inference accuracy using small AIM panels (as few as 50-300 SNPs) was comparable to using the full 650,000 SNPs, often exceeding 98% accuracy for broad geographic regions.
  • European Homogeneity Challenge: European populations, due to their genetic homogeneity, presented the greatest challenge for fine-scale differentiation, even with large marker sets.
  • HapMap Validation: The developed AIM panels performed well on HapMap populations, accurately assigning individuals to their closest geographic neighbors, even for admixed or poorly represented groups.
  • Panel Sizes Matter: The study explored three panel sizes (P1, P2, P3), demonstrating that even the smallest panels (P1) provided high accuracy, with larger panels offering marginal improvements in some cases.

Limitations and Future Work

  • European Differentiation: The authors acknowledge that distinguishing between closely related European populations (e.g., French, Orcadian, Italian) remains difficult even with small SNP panels, suggesting larger panels or more advanced methods might be needed.
  • Small Sample Sizes in HGDP: The HGDP dataset, while comprehensive, has small sample sizes for some populations, which may limit the representation of intra-population variation and thus affect classification accuracy.
  • Admixed Populations: For highly admixed populations like some Mexican Americans, classification can be challenging, and a confidence metric was introduced to classify such individuals as "unknown" when they lie far from reference populations.
  • Future Work: The authors suggest exploring more advanced classification methodologies and distance metrics. They also highlight the potential for using these AIMs in evolutionary genetics to study relationships between extant populations, in medical genetics for disease risk assessment, and in forensic and commercial ancestry testing, emphasizing the need for clear communication of limitations to clients. The extensibility of their methods to larger and more geographically dense datasets is also noted.

Summary updated

Abstract

The proposed methods and ancestry informative marker panels, in combination with the increasingly more comprehensive databases of human genetic variation, open new horizons in a variety of fields, ranging from the study of human evolution and population history, to medical genetics and forensics.
BibTeX
@article{paschou2010ancestry,
  author = {Peristera Paschou and Joshua P. Lewis and Asif Javed and Petros Drineas},
  title = {{Ancestry informative markers for fine-scale individual assignment to worldwide populations}},
  journal = {Journal of Medical Genetics},
  year = {2010},
  doi = {10.1136/jmg.2010.078212},
}
Figure