Generalized genetic association study with samples of related individuals

Zeny Feng, William Wong, Xin Gao, Flávio S. SchenkelThe Annals of Applied Statistics 2011 · DOI · arXiv

Abstract

Genetic association studies are critical for identifying genetic factors linked to complex traits. However, traditional methods often struggle with population stratification and the complexities of related individuals. This paper introduces a novel Generalized Quasi-Likelihood Score (GQLS) test designed for both quantitative and binary traits. By employing a logistic regression model that treats allele frequencies as the response and the trait as a covariate, the GQLS test allows the distribution of trait values to remain unspecified. Simulation studies demonstrate that this approach provides superior statistical power compared to the Family-Based Association Test (FBAT) while maintaining rigorous control over Type I error rates. The method's efficacy is further validated through applications to Holstein cattle breeding data and human genetic data from the Collaborative Study of the Genetics of Alcoholism (COGA).

Introduction

In the search for Single Nucleotide Polymorphisms (SNPs) that influence complex traits—ranging from agricultural productivity to human disease susceptibility—researchers must navigate the challenge of population stratification. When study samples include related individuals, such as families or inbred populations, standard association tests frequently yield false positives. These errors occur because the tests mistake shared genealogical history for true genetic association.

Existing solutions, such as the Family-Based Association Test (FBAT), are robust but often inefficient. FBAT typically decomposes complex pedigrees into smaller nuclear families, discarding valuable information from the broader pedigree structure. Furthermore, many existing methods are specialized for either binary or quantitative traits, lacking a unified framework. The Generalized Quasi-Likelihood Score (GQLS) test addresses these limitations by providing a flexible, powerful, and distribution-independent approach to genetic association in related samples.

Methodology

The "Flipped" Regression Model

The core innovation of the GQLS test is its departure from the standard regression paradigm. In traditional models, the trait is the response variable and the genotype is the covariate. This requires researchers to make specific assumptions about the trait's distribution (e.g., normality), which are often violated in real-world datasets.

In contrast, the GQLS model treats the allele frequency as the response and the trait as the covariate. By using a logistic link function to model the expected allele frequency, the distribution of the trait can remain entirely unspecified. This "flipped" perspective makes the test inherently robust to skewed or non-normal trait data.

Accounting for Relatedness

To account for the inherent correlation between related individuals, the GQLS test incorporates a comprehensive correlation matrix ($\rho$). This matrix is built using kinship and inbreeding coefficients, allowing the model to utilize the exact covariance structure of the entire pedigree. Unlike methods that simplify pedigrees, GQLS treats the entire family tree as a single unit of information, capturing "remote" relationships that are often ignored.

Scalability and Robustness

The framework also allows for the combination of results across multiple subpopulations ($W_{all}$) by summing individual GQLS statistics. This feature ensures the method remains robust against broader population stratification while maximizing the use of available data.

Experimental Results

Simulation Studies

The researchers tested the GQLS method against simulated six-generation pedigrees and smaller family units. The results indicated two primary advantages:

  1. Superior Power: GQLS consistently outperformed FBAT, particularly in scenarios involving binary traits or small sample sizes.
  2. Error Control: The method strictly maintained Type I error rates at desired levels. In contrast, standard trend tests that ignore relatedness showed significant inflation of false-positive signals.

Real-World Applications

Holstein Cattle Data: The method was applied to a dataset of 821 proven bulls in a complex "granddaughter design." Analyzing the Estimated Breeding Value (EBV) for milk yield, GQLS identified 36 significant regions. Notably, it confirmed the presence of the DGAT1 gene on chromosome 14, a well-documented regulator of milk composition, demonstrating the test's ability to recover known biological signals in highly inbred populations.

COGA Study: In an analysis of 1,614 individuals from 143 families regarding alcohol dependence (a binary trait), GQLS identified significant SNPs on chromosomes 2, 6, and 7. These findings aligned with previously reported linkage regions while also highlighting new candidate SNPs for further investigation.

Discussion and Conclusion

The GQLS test represents a significant advancement in genetic association methodology. By utilizing all available family members and avoiding restrictive distributional assumptions, it offers a more information-efficient and powerful alternative to traditional family-based tests.

One current limitation is the requirement for a known pedigree structure. If genealogical records are inaccurate, the kinship matrix may be compromised. Future work may involve integrating genome-wide data to infer relationships directly, further increasing the method's utility in studies where pedigree records are missing or incomplete. Overall, the GQLS framework provides a robust and versatile tool for uncovering the genetic architecture of complex traits in related populations.

Summary updated

Abstract

Genetic association study is an essential step to discover genetic factors that are associated with a complex trait of interest. In this paper we present a novel generalized quasi-likelihood score (GQLS) test that is suitable for a study with either a quantitative trait or a binary trait. We use a logistic regression model to link the phenotypic value of the trait to the distribution of allelic frequencies. In our model, the allele frequencies are treated as a response and the trait is treated as a covariate that allows us to leave the distribution of the trait values unspecified. Simulation studies indicate that our method is generally more powerful in comparison with the family-based association test (FBAT) and controls the type I error at the desired levels. We apply our method to analyze data on Holstein cattle for an estimated breeding value phenotype, and to analyze data from the Collaborative Study of the Genetics of Alcoholism for alcohol dependence. The results show a good portion of significant SNPs and regions consistent with previous reports in the literature, and also reveal new significant SNPs and regions that are associated with the complex trait of interest.
BibTeX
@article{feng2011generalized,
  author = {Zeny Feng and William Wong and Xin Gao and Flávio S. Schenkel},
  title = {{Generalized genetic association study with samples of related individuals}},
  journal = {The Annals of Applied Statistics},
  year = {2011},
  doi = {10.1214/11-aoas465},
}
Figure