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 data set 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.
Introduction
In Genome-Wide Association Studies (GWAS), researchers aim to predict physical traits (phenotypes) based on hundreds of thousands of genetic markers (SNPs). Historically, this has been approached through two distinct statistical lenses. The first, Linear Mixed Models (LMMs), assumes an "infinitesimal" architecture where every marker has a tiny, normally distributed effect. The second, Sparse Regression, assumes that only a small fraction of markers have any effect at all.
In reality, complex traits like height or cholesterol levels rarely fit perfectly into one of these categories; they often present a mixture of many tiny effects and a few larger ones. The Bayesian Sparse Linear Mixed Model (BSLMM) addresses this by providing a hybrid framework that adaptively learns a trait's specific genetic architecture from the data, rather than requiring researchers to choose a model a priori.
The BSLMM Framework
BSLMM treats the effect of each genetic marker as a mixture of two normal distributions:
- Polygenic Background: A distribution capturing the small effects of all markers.
- Extra Effects: A distribution capturing the sparse, larger signals that stand out from the background.
Adaptive Priors and Computational Efficiency
Unlike previous hybrid models that required manual tuning of sensitive parameters, BSLMM uses a Bayesian framework to estimate hyper-parameters—such as the proportion of markers with large effects—directly from the dataset.
To make this computationally feasible for modern GWAS, the authors implemented a linear algebra optimization. While standard fitting of these models typically requires matrix inversions that scale cubically ($n^3$) with the number of individuals, BSLMM utilizes an eigen-decomposition trick. This reduces the cost of each Markov chain Monte Carlo (MCMC) iteration to be linear ($n$) relative to the sample size, allowing the model to run on standard desktop hardware with thousands of individuals and hundreds of thousands of markers.
Key Metrics: PVE and PGE
BSLMM introduces a nuanced way to decompose heritability through two primary metrics:
- PVE (Proportion of Variance Explained): The total variance explained by all genotypes.
- PGE (Proportion of Genetic variance explained by "large" effects): This metric indicates what fraction of the total genetic variance is attributable to markers with larger-than-average effects.
This decomposition allows biologists to quantify how much of a trait's "missing heritability" is hidden in the polygenic background versus identifiable large-effect loci.
Empirical Results
The researchers validated BSLMM across simulated data, human GWAS data (height and blood lipids), WTCCC disease data, and mouse datasets.
Phenotype Prediction
In the WTCCC disease data, BSLMM performed as well as or better than both LMMs and sparse regression across seven common diseases. In heterogeneous stock mouse data, BSLMM consistently outperformed five other state-of-the-art regression methods, including the Bayesian Lasso. The model proved particularly robust; while sparse models failed on polygenic traits and LMMs struggled with sparse architectures, BSLMM remained stable and accurate across all scenarios.
Heritability Estimation
When applied to human height in an Australian cohort, BSLMM estimated a PVE of 0.41. This is highly consistent with previous LMM-based studies, but BSLMM provided the additional insight that only approximately 12% of that genetic variance (PGE) was derived from markers with large effects, confirming height's highly polygenic nature.
Discussion and Limitations
While BSLMM represents a significant step forward, some challenges remain. The current MCMC approach, though optimized, may still face scaling issues with "Biobank-scale" datasets involving hundreds of thousands of individuals. Additionally, while the authors developed a "probit" version for binary disease traits, they found that treating disease status as a quantitative variable was computationally more efficient and yielded similar prediction accuracy. Future work may also expand the model to include non-additive effects, such as gene-gene or gene-environment interactions.
BSLMM is implemented as part of the GEMMA software package.
Summary updated