Polygenic Modeling with Bayesian Sparse Linear Mixed Models

Xiang Zhou, Peter Carbonetto, Matthew Stephens

Abstract

Both linear mixed models (LMMs) and sparse regression models are widely used in genetics applications, including, recently, polygenic modeling in genome-wide association studies. These two approaches make very different assumptions, so are expected to perform well in different situations. However, in practice, for a given dataset one typically does not know which assumptions will be more accurate. Motivated by this, we consider a hybrid of the two, which we refer to as a "Bayesian sparse linear mixed model" (BSLMM) that includes both these models as special cases. We address several key computational and statistical issues that arise when applying BSLMM, including appropriate prior specification for the hyper-parameters and a novel Markov chain Monte Carlo algorithm for posterior inference. We apply BSLMM and compare it with other methods for two polygenic modeling applications: estimating the proportion of variance in phenotypes explained (PVE) by available genotypes, and phenotype (or breeding value) prediction. For PVE estimation, we demonstrate that BSLMM combines the advantages of both standard LMMs and sparse regression modeling. For phenotype prediction it considerably outperforms either of the other two methods, as well as several other large-scale regression methods previously suggested for this problem.

The Challenge of Genetic Architecture

In genome-wide association studies (GWAS), researchers typically rely on one of two contrasting models to relate genotypes to phenotypes. Linear Mixed Models (LMMs) assume an "infinitesimal" architecture where every genetic variant contributes a tiny, normally distributed effect. These are excellent for capturing population structure but struggle when a few genes have disproportionately large impacts. Conversely, Sparse Regression (BVSR) assumes a "sparse" architecture where only a small fraction of variants matter. While effective at identifying major genes, BVSR often ignores the "polygenic tail"—the cumulative effect of thousands of tiny contributors.

In practice, the true genetic architecture of a trait is rarely known beforehand. Choosing the wrong model leads to biased estimates of heritability (the proportion of variance explained, or PVE) and poor predictions of individual phenotypes, such as disease risk.

The BSLMM Framework

The Bayesian Sparse Linear Mixed Model (BSLMM) addresses this by treating both LMM and BVSR as special cases within a single hybrid framework. It models phenotypes as a sum of "sparse" effects (a few SNPs with relatively large impacts) and a "random" effect (the collective small impact of all other SNPs).

To make the model user-friendly and computationally feasible, the authors introduced two key innovations:

  1. Interpretable Priors: The model is re-parameterized using two intuitive values: $h$ (the total PVE) and $\rho$ (the proportion of genetic variance explained by sparse effects). By placing uniform priors on these, the model allows the data to determine whether a trait is sparse or polygenic.
  2. Computational Efficiency: To overcome the slow nature of Bayesian MCMC, the authors utilized a linear algebra trick involving a single eigen-decomposition of the relatedness matrix. This reduced the per-iteration cost from $O(n^3)$ to $O(n)$, enabling the analysis of datasets with thousands of individuals and hundreds of thousands of markers.

Performance in PVE Estimation

BSLMM acts as a "natural bridge" between competing models. If a trait is purely polygenic, the model pushes $\rho$ toward 0 (behaving like an LMM); if it is highly sparse, $\rho$ moves toward 1 (behaving like BVSR).

Figure 5: Comparison of BSLMM, LMM, and BVSR in estimating PVE across different genetic architectures. Panels A and C show the Root Mean Square Error (RMSE), while B and D show the actual PVE estimates. BSLMM (purple) consistently provides accurate and low-variance estimates across both sparse and polygenic scenarios.

As shown in the figure above, LMMs are generally robust but can exhibit high variance. Sparse models (BVSR) severely underestimate PVE for highly polygenic traits (Scenario I, right side of panels B and D) because they fail to capture small-effect variants. BSLMM, however, maintains high accuracy across the entire spectrum.

Phenotype Prediction and Real-World Application

While PVE estimation only requires getting the average effect size right, accurate phenotype prediction requires identifying individual effect sizes. BSLMM’s ability to isolate large effects while accounting for the polygenic background makes it a superior predictor.

  • Disease Risk: In the WTCCC 1 study (covering seven common diseases), BSLMM performed as well as or better than both LMM and BVSR.
  • Quantitative Traits: In mouse datasets, BSLMM significantly outperformed standard LMMs and Bayesian Lasso for traits like CD8 and MCH.
  • Human Height: For human height—a classic polygenic trait—BSLMM and LMM gave nearly identical PVE estimates (~0.42), while BVSR fell short (~0.15), confirming BSLMM's robustness.

Conclusion and Limitations

BSLMM provides a powerful, adaptive tool for polygenic modeling that removes the need for researchers to guess a trait's architecture in advance. However, some limitations remain. The current implementation treats case-control status as a quantitative value; while a probit version for binary data exists, it is computationally more expensive. Additionally, the model currently requires the genotype matrix to be stored in memory, placing a practical limit of approximately 10,000 individuals for standard desktop hardware. Future work may focus on scaling these methods to the even larger cohorts now common in modern genomics.

Summary updated

Abstract

Both linear mixed models (LMMs) and sparse regression models are widely used in genetics applications, including, recently, polygenic modeling in genome-wide association studies. These two approaches make very different assumptions, so are expected to perform well in different situations. However, in practice, for a given dataset one typically does not know which assumptions will be more accurate. Motivated by this, we consider a hybrid of the two, which we refer to as a "Bayesian sparse linear mixed model" (BSLMM) that includes both these models as special cases. We address several key computational and statistical issues that arise when applying BSLMM, including appropriate prior specification for the hyper-parameters and a novel Markov chain Monte Carlo algorithm for posterior inference. We apply BSLMM and compare it with other methods for two polygenic modeling applications: estimating the proportion of variance in phenotypes explained (PVE) by available genotypes, and phenotype (or breeding value) prediction. For PVE estimation, we demonstrate that BSLMM combines the advantages of both standard LMMs and sparse regression modeling. For phenotype prediction it considerably outperforms either of the other two methods, as well as several other large-scale regression methods previously suggested for this problem. Software implementing our method is freely available from http://stephenslab.uchicago.edu/software.html.
BibTeX
@inproceedings{zhou2013polygenic,
  author = {Xiang Zhou and Peter Carbonetto and Matthew Stephens},
  title = {{Polygenic Modeling with Bayesian Sparse Linear Mixed Models}},
  booktitle = {PLoS},
  year = {2013},
  doi = {10.1371/journal.pgen.1003264},
}
Figure