TL;DR
This paper introduces BEAVR, a new statistical framework for estimating how many genetic variants contribute to complex traits within specific genomic regions. By accurately mapping this "regional polygenicity," the study reveals that complex traits are highly polygenic across the genome, with heritability strongly linked to the number of causal variants in a region.
What Problem Does This Paper Address?
Understanding the genetic architecture of complex traits (like height, BMI, or blood pressure) is crucial for predicting disease risk and developing targeted therapies. A key aspect of this architecture is "polygenicity" – the number of genetic variants (SNPs) that influence a trait. While genome-wide polygenicity has been studied, how this polygenicity is distributed across the genome (regional polygenicity) has been less clear. This lack of regional resolution makes it difficult to pinpoint specific genomic areas that contribute most to trait variation. A major challenge in estimating polygenicity is the correlation between nearby SNPs due to linkage disequilibrium (LD), which complicates statistical modeling and can lead to intractable computations.
Methods
The authors developed a novel statistical framework called Bayesian Estimation of Variants in a Region (BEAVR). BEAVR uses a Bayesian approach with a spike-and-slab prior to model the probability of a SNP having a non-zero effect. Key features include:
- Regional Focus: BEAVR is designed to estimate polygenicity within specific genomic regions (e.g., 6-Mb windows), rather than just genome-wide.
- LD Accounting: It explicitly accounts for linkage disequilibrium (LD) within each region using an in-sample LD matrix.
- Efficient MCMC: To overcome computational challenges, BEAVR employs an optimized Markov Chain Monte Carlo (MCMC) algorithm that leverages the expected sparsity of causal variants (i.e., only a small fraction of SNPs are truly causal). This significantly speeds up computation, reducing complexity from O(M³) to O(K,M,) where K, is the number of causal SNPs.
- Input: BEAVR takes marginal effect sizes from GWAS, in-sample LD information, and estimated regional heritability as input.
Datasets
- Simulated Data: The authors generated synthetic GWAS data to test BEAVR's performance under various genetic architectures, sample sizes, and heritability levels.
- UK Biobank: Real-world data from the UK Biobank was used for analysis. This included genotype data from approximately 290,641 unrelated individuals of White British ancestry. Five complex traits were analyzed: Body Mass Index (BMI), Height, Waist-hip ratio, Diastolic blood pressure, and Systolic blood pressure.
- 1000 Genomes Project: Used as an external reference panel for LD in some simulation scenarios.
Key Insights and Ideas
- Regional Polygenicity is Highly Variable: The study reveals substantial variation in how polygenic different genomic regions are for the same trait.
- Heritability is Driven by Causal SNP Count: The findings strongly support the hypothesis that regional heritability is primarily explained by the number of causal SNPs in that region, more so than just the total number of SNPs.
- LD Matters for Accurate Estimation: BEAVR's ability to accurately estimate polygenicity, especially in scenarios with high polygenicity, is attributed to its explicit modeling of LD. Methods that ignore or simplify LD can exhibit significant bias.
- "Omnigenic" Model Support: The widespread polygenicity across many regions aligns with the "omnigenic" model, suggesting that most genes, even those not directly related to a trait's core function, can have small effects.
Key Results
- Unbiased Estimates: BEAVR provides relatively unbiased estimates of regional polygenicity across a wide range of simulated genetic architectures, with a relative bias < 2%.
- GENESIS Bias: In contrast, the GENESIS method showed severe downward bias (over 70%) in high polygenicity settings.
- Ubiquitous Polygenicity: For the analyzed UK Biobank traits, over one-third of 6-Mb genomic regions harbored at least one causal variant. For height, this proportion reached nearly 80%.
- Significant Variation: On average, 48.9% of regions contained at least 5 causal SNPs, and 5.44% contained at least 50 causal SNPs.
- Heritability-Polygenicity Link: A strong linear relationship was found between regional heritability and the number of causal SNPs (R² = 0.278 on average across traits), outperforming the relationship with the total number of SNPs.
- Computational Speedup: The optimized BEAVR algorithm achieved a 12-fold improvement in run-time for regions with 5,000 SNPs compared to a naive implementation.
Limitations and Future Work
- Gaussian Assumption: BEAVR assumes causal effect sizes follow a Gaussian distribution. While robust to some deviations, extreme deviations or very small effect sizes can lead to downward bias. The authors suggest exploring other distributions (e.g., mixtures of Gaussians) for future work.
- LD Matrix Block Structure: Genome-wide polygenicity estimates assume a block-structured LD matrix.
- Array Data Limitations: Analyses were limited to genotyped SNPs, meaning true causal SNPs not present on the array were not accounted for. Future work could investigate analyses on imputed data.
- Reference Panel LD: Using out-of-sample LD reference panels can lead to inaccurate estimates if the panel is not ancestrally homogeneous with the target population. Future work could explore more sophisticated LD matching strategies.
Summary updated