TL;DR
This paper introduces a new Bayesian method (BayesS) to simultaneously estimate heritability, polygenicity, and the relationship between SNP effect size and minor allele frequency (MAF) for complex traits. They found widespread evidence of negative selection acting on human complex traits, particularly reproductive and cardiovascular traits, suggesting that natural selection has shaped the genetic architecture of these traits by favoring variants that are less common.
What Problem Does This Paper Address?
Understanding the genetic basis of complex traits (like height, BMI, or disease susceptibility) is crucial for both human health and evolutionary biology. While Genome-Wide Association Studies (GWAS) have identified many genetic variants associated with these traits, they often explain only a fraction of the heritability, leading to the "missing heritability" problem. Furthermore, inferring the action of natural selection on these traits is challenging because selection often acts subtly across many variants with small effects, rather than through dramatic "sweeps." A key indicator of natural selection is the relationship between the effect size of a genetic variant and its frequency in the population. Specifically, if variants that are detrimental to fitness are kept at low frequencies by negative selection, we'd expect to see a correlation between smaller effect sizes and lower minor allele frequencies (MAF). Existing methods struggle to simultaneously estimate these key genetic architecture parameters and detect selection signals, especially for polygenic traits where selection pressure is distributed across many variants.
Methods
The authors developed a novel Bayesian mixed linear model (BayesS) that addresses these challenges. The core innovation is modeling the joint distribution of SNP effect size and MAF. This is achieved by defining a hierarchical mixture prior for SNP effects. The model simultaneously estimates:
- SNP-based heritability ($h_{SNP}^2$): The proportion of phenotypic variance explained by all SNPs.
- Polygenicity ($\pi$): The proportion of SNPs with non-zero effects.
- Relationship between effect size and MAF (S): This parameter quantifies how effect size changes with MAF. A negative S suggests that low-MAF variants have larger effects (consistent with negative selection), while a positive S suggests the opposite (consistent with positive selection).
The model uses Hamiltonian Monte Carlo (HMC) for sampling the posterior distribution of S and Gibbs sampling for other parameters. To handle large datasets, they also developed a "nested BayesS" approach that partitions the genome into windows, speeding up analysis significantly.
Datasets
The study utilized two major datasets:
- UK Biobank (UKB): Genotype and phenotype data for 126,752 individuals across 28 complex traits (including quantitative traits, diseases, and educational attainment). This dataset is a large-scale resource for studying human complex traits.
- Consortium for the Architecture of Gene Expression (CAGE): Gene expression data for 1,748 individuals, comprising 27,869 gene expression traits (probes). This dataset allowed them to investigate selection on genes influencing gene expression.
Key Insights and Ideas
- Negative Selection is Pervasive: The most striking finding is that for the vast majority of complex traits examined, the relationship between SNP effect size and MAF is negative (S < 0). This strongly suggests that negative selection is a dominant force shaping the genetic architecture of human complex traits, acting to purge deleterious variants.
- Reproductive and Cardiovascular Traits are Most Strongly Selected: These traits showed the strongest signals of negative selection, implying they are more closely linked to fitness than other traits.
- Polygenicity and Pleiotropy: The high degree of polygenicity observed (around 6% of SNPs having non-zero effects on average) highlights the widespread nature of genetic influences on complex traits. This, combined with pervasive selection, underscores the importance of pleiotropy (where one gene affects multiple traits) in how natural selection operates.
- Distinguishing Selection Types: The parameter S is not only useful for detecting selection but also for distinguishing between different modes of selection (e.g., stabilizing vs. directional selection), as demonstrated through forward simulations.
Key Results
- Complex Traits:
- 23 out of 28 complex traits in the UKB showed significant evidence of natural selection (S ≠ 0).
- All significant estimates of S were negative, indicating negative selection.
- Reproductive and cardiovascular traits, followed by height and anthropometric traits, showed the strongest selection signals.
- On average, 6% of SNPs had non-zero effects, explaining 22% of phenotypic variance.
- Gene Expression Traits:
- 30 genes showed significant evidence of negative selection.
- The nested BayesS model was six times faster than the standard BayesS for traits with low polygenicity.
Limitations and Future Work
- Overestimation of Polygenicity: The authors acknowledge that their polygenicity estimate ($\pi$) might be an overestimation of the causal variants, as it reflects the proportion of SNPs with non-zero effects, which can be influenced by linkage disequilibrium (LD) with unobserved causal variants. However, $\pi$ is still useful for comparing polygenicity levels across traits.
- Power with Rare Variants: The power to detect selection signals could be improved by using whole-genome sequencing (WGS) or imputed sequence data, which include a larger number of rare variants. However, this is computationally demanding.
- Nested BayesS Inflation: In simulations, the nested BayesS model showed a slight inflation in estimating positive S values, though this did not affect their conclusions as they found no significant positive S.
- Future Directions: The authors suggest that future GWAS using WGS or imputed data could help discover more rare variants with large effects. They also propose that the identified genes showing evidence of negative selection in gene expression could be targets for further functional studies.
Summary updated