TL;DR
This paper introduces LDER-GE, a novel statistical method that significantly improves the accuracy of estimating the contribution of gene-environment (GE) interactions to human complex traits. By leveraging full linkage disequilibrium (LD) information from summary statistics, LDER-GE offers a more powerful and efficient way to uncover these crucial genetic influences compared to existing methods.
What Problem Does This Paper Address?
Understanding the genetic basis of complex human traits, like height or susceptibility to diseases, is a major challenge. While we know genes play a role, a significant portion of the heritable variation remains unexplained – the so-called "missing heritability." Gene-environment (GE) interactions, where the effect of a gene depends on environmental factors (or vice versa), are thought to contribute to this missing heritability. However, detecting and quantifying these GE interactions is difficult because their effects are often small and require sophisticated statistical methods. Existing methods that use genome-wide association study (GWAS) summary statistics often rely on simplified LD information, limiting their statistical power and accuracy. This paper aims to bridge this gap by developing a method that uses comprehensive LD information to more precisely estimate the variance explained by GE interactions.
Methods
The core innovation is the LDER-GE (Linkage-Disequilibrium Eigenvalue Regression for Gene-Environment interactions) method. It builds upon the principles of the original LDER framework, which uses eigen-decomposition of the LD matrix to capture full LD information. LDER-GE transforms the standard GWAS Z-score summary statistics using this full LD matrix. This transformation consolidates association information and, crucially, allows for a more statistically efficient estimation of the GE interaction variance component. Unlike previous LD score regression (LDSC)-based methods that primarily use the diagonal of the squared LD matrix, LDER-GE utilizes the entire LD matrix, leading to improved precision. The method employs a method of moments estimation approach and uses delete-block-wise jackknife for standard error estimation.
Datasets
The study utilized the UK Biobank (UKBB) dataset, a large-scale biomedical database.
- For simulations and reference panel construction, a subset of UKBB data (N=276,050 individuals, M=396,330 variants) was used.
- For real-data analysis, a larger UKBB dataset (N=307,259 unrelated European-ancestry individuals, M=966,766 variants) was used.
- The 1000 Genomes project dataset was used to construct an out-of-sample LD reference panel for simulations.
Key Insights and Ideas
- Full LD is Crucial: The paper strongly advocates for the use of the complete LD matrix, not just diagonal elements, for accurate estimation of GE interaction variance. This is a key takeaway that extends previous findings on heritability estimation to GE interactions.
- GE Interactions are Substantial: The real-data analysis revealed that GE interactions contribute a significant portion to phenotypic variance, sometimes approaching the magnitude of narrow-sense heritability, particularly for traits like Type-II Diabetes and Coronary Artery Disease.
- Sex and BMI are Key Environmental Modifiers: The analysis highlighted sex and Body Mass Index (BMI) as particularly important environmental covariates that significantly modify genetic effects across a range of health-related traits.
- LDER-GE Outperforms Existing Methods: The simulations and real-data results consistently show LDER-GE to be more statistically efficient and powerful than LDSC-based methods.
Key Results
- Statistical Efficiency: LDER-GE demonstrated an average of 23% higher statistical efficiency compared to LDSC-based methods in simulations, equivalent to approximately a 51% increase in sample size.
- Detection Power: In the UK Biobank analysis, LDER-GE identified 34 significant GE interaction pairs, a 48% increase compared to the 23 pairs identified by the LDSC-based method.
- Type-I Error Control: LDER-GE effectively controls type-I error rates, producing unbiased results across various simulation scenarios.
- Standard Error Reduction: The standard error of LDER-GE estimates was, on average, 21% lower than that of the LDSC-based method, leading to narrower confidence intervals and higher statistical power.
- Aggregated GE Variance: LDER-GE successfully estimated aggregated multi-GE interaction variance, revealing substantial contributions for several phenotypes, with 13 out of 22 phenotypes showing contributions greater than 10% of narrow-sense heritability.
Limitations and Future Work
- Polygenic Assumption: LDER-GE, like LDSC, assumes polygenic GE effects. Violations of this assumption can lead to underestimation of variance components.
- LD Reference Panel Quality: The accuracy of LD estimation is crucial, especially for variants with low minor allele frequencies (MAFs). The paper recommends using high-quality reference panels and variants with MAF > 0.05.
- Covariate Correlation: The current model does not explicitly differentiate between GE covariate correlations, which could introduce estimation bias. Future work could explore methods to address this.
- Liability-Scale Transformation Bias: For binary phenotypes, the liability-scale transformation can introduce bias, particularly for rare diseases with substantial GE interaction variance. The paper notes that while the bias is present, it's not always substantial, but warrants caution.
- Future Directions: The authors suggest incorporating variant functional annotation or allele frequency information to enhance estimation and explore how LDER-GE's eigen-decomposition technique could be used to aggregate full LD information for methods like SNP-level GxE analysis.
Summary updated