TL;DR
The SuSiE model is a popular tool for genetic fine-mapping, but it can produce unreliable results in small studies. This paper introduces a simple modification to SuSiE that accounts for uncertainty in residual variance, significantly improving its accuracy in small sample settings. This enhancement is crucial for applications like molecular QTL studies in rare cell types where sample sizes are inherently limited.
What Problem Does This Paper Address?
Genetic fine-mapping aims to pinpoint specific genetic variants associated with a particular trait or disease. The SuSiE (Sum of Single Effects) model is a powerful tool for this task, especially when dealing with many correlated genetic variants. However, the authors discovered a critical limitation: when applied to datasets with small sample sizes (fewer than 50 individuals), the original SuSiE model tends to produce "miscalibrated" results. This means that the confidence intervals (called Credible Sets or CSs) it generates are not as reliable as they should be, leading to an increased rate of false positive findings. This issue is particularly problematic in emerging areas of genetic research, such as studying rare cell types or specific tissue samples, where obtaining large datasets is often challenging.
Methods
The core of the paper's contribution lies in a modification to the SuSiE model. The original SuSiE model estimates the residual variance (a measure of unexplained variation in the data) using maximum likelihood. The authors propose a Bayesian approach where this residual variance is integrated out using a diffuse prior. This is achieved by augmenting the "Single Effect Regression" (SER) model, which SuSiE is built upon, with a conjugate inverse-gamma prior on the residual variance. This modified model, termed "SS-SER" (for "Servin and Stephens"), is then integrated into the SuSiE framework. The fitting procedure for SuSiE, known as Iterative Bayesian Stepwise Selection (IBSS), is retained, but it now uses the SS-SER model instead of the default SER. This approach aims to better account for uncertainty in the residual variance, especially in small sample sizes, without sacrificing the computational efficiency of the original SuSiE.
Datasets
The paper uses simulated data for its primary evaluations. These simulations are designed to mimic real-world genetic fine-mapping scenarios, varying sample sizes (N=20 to 100) and heritability (h²=25% to 75%). They also analyze two real-world RNA-seq datasets:
- Aracena et al. [10]: Transcriptional response in primary macrophages after influenza infection (n=35).
- Lopes et al. [11]: eQTLs in microglia, a cell type relevant to brain aging (n=47). For these real datasets, the authors fine-mapped gene expression QTLs by selecting SNPs within a 200kb window of each gene.
Key Insights and Ideas
- Small samples are a big problem for standard SuSiE: The original SuSiE model's calibration breaks down significantly when sample sizes are small, leading to an over-reliance on potentially false positive associations.
- Accounting for residual variance uncertainty is key: The authors demonstrate that explicitly modeling the uncertainty in the residual variance, rather than treating it as a fixed point estimate, is crucial for improving calibration in small sample sizes.
- SS-SER offers a practical solution: The proposed SS-SER modification integrates seamlessly into the existing SuSiE framework, offering a computationally efficient way to achieve better calibration.
- Improved calibration comes at a cost: While the SS-SER model improves calibration, it can lead to slightly larger Credible Sets and a reduction in statistical power. This is a trade-off for increased reliability.
- Real-world data confirms simulation findings: The application of the SS-SER model to real eQTL datasets shows a substantial reduction in the number of reported Credible Sets, particularly in cases where the default SuSiE exhibited signs of miscalibration.
Key Results
- In simulations with small sample sizes (n < 50), the default SuSiE model showed coverage as low as 44.1% (compared to the target 95%) when heritability was 25%.
- The modified SuSiE with SS-SER improved coverage to 82.1% in the same scenario.
- The SS-SER model produced larger Credible Sets (CSs) and lower CS purity compared to the default SER, which is a consequence of increased uncertainty being accounted for.
- On real eQTL data, SuSiE with SS-SER reported significantly fewer unique CSs (e.g., 389 vs. 1020 for one dataset) and showed a marked reduction in miscalibration signs (e.g., fewer genes with multiple CSs).
Limitations and Future Work
The authors acknowledge that even with the SS-SER modification, some miscalibration can persist in certain settings. They suggest that further work is needed to improve performance for small-sample studies. The paper also notes that while their approach improves calibration, there is a trade-off with a slight loss of power and an increase in CS size. Future research could explore more sophisticated methods for handling residual variance uncertainty or investigate alternative approaches to fine-mapping in low-sample scenarios.
Summary updated