TL;DR
This paper introduces a novel two-variance component mixed model that simultaneously leverages identity-by-state (IBS) sharing and approximate pedigree information from genetic markers to improve genetic prediction accuracy. The model demonstrates significant gains in prediction accuracy and also corrects for inflated test statistics in association studies involving related individuals.
What Problem Does This Paper Address?
Predicting an individual's genetic predisposition to traits or diseases is a cornerstone of modern genetics. Traditional methods, like Best Linear Unbiased Prediction (BLUP), often rely on either identity-by-state (IBS) sharing derived from genetic markers or information from family pedigrees. While both approaches have proven useful, they typically operate independently. This paper addresses the gap in methods that can effectively combine these two sources of genetic information – IBS sharing and pedigree structure – to achieve more accurate genetic predictions, particularly in family-based studies where relatedness is a significant factor. Furthermore, existing mixed model association tests can produce unreliable results (inflated statistics) when dealing with related individuals, a problem this paper also aims to solve.
Methods
The core innovation is a two-variance component mixed model. This model extends the standard BLUP framework by incorporating two distinct sources of genetic covariance:
- IBS Sharing: This is captured by the standard Genetic Relationship Matrix (GRM) computed directly from observed genetic markers. This component accounts for the heritability explained by typed SNPs.
- Approximate Pedigree Structure: This component is modeled using a thresholded GRM. Here, pairwise genetic relationships below a certain threshold are set to zero. The idea is to capture stronger, more direct relatedness patterns akin to a pedigree, thereby accounting for heritability explained by untyped SNPs.
The model uses these two components to estimate genetic effects and make predictions. For association testing, a similar two-variance component framework is employed, which is shown to correct for inflated test statistics observed in standard mixed models when dealing with related individuals.
Datasets
The study utilizes two real-world human genetic datasets:
- CARe (Candidate-gene Association Resource): This dataset comprises African-American individuals from multiple cohorts (ARIC, CARDIA, CFS, JHS, MESA) with high-quality genotype data. Phenotypes analyzed include BMI, height, LDL, and HDL.
- FHS (Framingham Heart Study): This dataset includes SHARe samples with high-quality genotype data and phenotypes for BMI and height.
Both CARe and FHS are family studies, meaning they contain many related individuals, making them ideal for testing methods that account for relatedness. The authors also conducted extensive simulations using genotypes from these datasets and simulated genotypes to evaluate the model's performance under various scenarios.
Key Insights and Ideas
- Synergy of IBS and Pedigree: The most significant insight is that combining IBS sharing (from typed SNPs) and approximate pedigree structure (from untyped SNPs) provides a synergistic benefit, leading to better genetic prediction than using either source alone. This suggests that these two forms of genetic information capture complementary aspects of an individual's genetic architecture.
- Correcting Association Inflation: The standard mixed model association tests, widely used in GWAS, can be unreliable in family studies due to inflated test statistics. The two-variance component model effectively addresses this inflation, providing more trustworthy results for association studies in related populations.
- Future-Proofing for Larger Datasets: Simulations indicate that the relative gains in prediction accuracy achieved by the two-variance component model will likely persist and even increase as sample sizes grow, making it a valuable tool for future large-scale genetic studies.
Key Results
- Improved Prediction Accuracy: In simulations and real data analyses, the two-variance component model consistently outperformed standard BLUP and BLUP using only the thresholded GRM.
- Real data analyses showed relative improvements in prediction accuracy of up to 20% for quantitative phenotypes.
- Simulations projected that these gains would hold for larger sample sizes.
- Association Test Calibration: The two-variance component model significantly reduced the inflation of test statistics in association studies with related individuals, bringing them closer to the expected null distribution.
- In simulations, inflation of up to 11% was observed with standard MLM, while the two-variance component model nearly eliminated this inflation.
- Real data analyses showed up to a 1.05-fold inflation with standard MLM, which was also mitigated by the new model.
- Accurate Heritability Estimation: The model accurately estimated the genetic components of heritability (h²g and h² - h²g), whereas standard models provided intermediate estimates.
Limitations and Future Work
- Applicability to Related Individuals: The method is primarily designed for and most effective in datasets with related individuals. Its utility in completely unrelated populations might be limited.
- Computational Complexity: Implementing the two-variance component model, especially for prediction, can be computationally intensive, requiring O(N²) memory and O(N³) time per REML iteration. The authors suggest that future work could focus on optimizing these computations using faster algorithms and relationship inference techniques.
- Case-Control Ascertainment: The current approach does not explicitly address case-control ascertainment, which could be an avenue for future research by integrating techniques from other methods.
- Non-infinitesimal Models: The authors suggest exploring the integration of their two-variance component framework with non-infinitesimal models (which account for a few large-effect loci) for even greater prediction accuracy.
- Extending to Other Data Types: The authors note that their approach is applicable to datasets with related individuals and available genotypes, and that large-scale human datasets of this type are increasingly common. They also mention parallel work in cattle showing similar benefits, though with some breed-specific nuances.
Summary updated