SMetABF: A rapid algorithm for Bayesian GWAS meta-analysis with a large number of studies included

Jianle Sun, Ruiqi Lyu, Luojia Deng, Qianwen Li, Yang Zhao, Yue ZhangPLOS Computational Biology 2022 · DOI

Introduction

Genome-wide association studies (GWAS) are fundamental to identifying the genetic underpinnings of complex diseases. However, individual studies often suffer from limited statistical power or inconsistencies due to population stratification. Meta-analysis addresses these issues by pooling data across multiple studies. The Bayesian framework, specifically the MetABF method, is considered the gold standard for this task as it calculates an Approximate Bayes Factor (ABF) to evaluate genetic associations.

However, MetABF faces a significant computational hurdle: to find the optimal association, researchers must test every possible subset of studies. For $n$ studies, there are $2^n$ possible subsets. As the number of studies grows, this "exponential explosion" makes exhaustive search computationally impossible. SMetABF is a new, high-speed algorithm designed to overcome this limitation.

SMetABF optimizes the subset selection process by replacing exhaustive searches with Shotgun Stochastic Search (SSS). While standard Markov chain Monte Carlo (MCMC) methods explore the search space incrementally, SSS explores the entire "neighborhood" of a current subset in each iteration—simultaneously adding, deleting, or replacing elements.

This approach allows SSS to identify high-probability subsets much faster and more reliably than traditional MCMC. The framework builds on the existing MetABF model, calculating the ABF by assuming effect sizes follow a multivariate normal distribution, which accounts for both correlation and heterogeneity between different studies.

SMetABF (SSS) performance compared to exhaustive search (EXH) and MCMC across varying odds ratios. SSS (green) consistently tracks the exhaustive search results, while MCMC (orange) often fails to reach the global maximum.

Performance and Accuracy

The researchers validated SMetABF using synthetic datasets (up to $N=25$ studies) to compare its performance against exhaustive search and MCMC. The results demonstrated that SMetABF is both highly accurate and significantly faster.

In simulations, the SSS curve almost perfectly coincided with the exhaustive search curve—the theoretical maximum—whereas MCMC frequently fell short. Furthermore, SSS reached the maximum ABF in significantly fewer iterations; for instance, SSS with 500–1,000 iterations outperformed MCMC with 10,000–20,000 iterations.

Comparison of mean logABF (accuracy) and computation time (speed) across different prior configurations. SSS (green) matches the accuracy of exhaustive search (EXH, dark blue) while maintaining a speed comparable to or better than MCMC (orange).

Real-World Application: Parkinson’s and Autoimmunity

To demonstrate its utility, the authors applied SMetABF to 59 distinct datasets, including 29 studies on Parkinson’s Disease (PD) and 30 on autoimmune disorders (such as Multiple Sclerosis and Systemic Sclerosis).

Key biological insights included:

  • The PD-Autoimmune Connection: A mixed meta-analysis revealed that certain loci, such as BTNL2, are strongly associated with both PD and autoimmune disorders, supporting the theory that Parkinson’s may have an underlying autoimmune component.
  • Locus-Specific Behavior: The study identified "pure" PD risk factors like SNCA and MAPT, while others like KANSL1 showed associations across both disease categories, suggesting distinct biological pathways for different risk factors.

Conclusion

SMetABF provides a robust and efficient solution for large-scale Bayesian GWAS meta-analysis. By significantly reducing computation time without sacrificing accuracy, it enables the analysis of massive, multi-study datasets that were previously inaccessible. The authors have released the method as an R package (GWASmeta) and a Shiny web tool to facilitate its use in the broader biomedical research community. Future work may expand SMetABF to handle three-dimensional arrays, allowing for the simultaneous analysis of multiple genetic loci across multiple traits and studies.

Summary updated

BibTeX
@article{sun2022smetabf,
  author = {Jianle Sun and Ruiqi Lyu and Luojia Deng and Qianwen Li and Yang Zhao and Yue Zhang},
  title = {{SMetABF: A rapid algorithm for Bayesian GWAS meta-analysis with a large number of studies included}},
  journal = {PLOS Computational Biology},
  year = {2022},
  doi = {10.1371/journal.pcbi.1009948},
}
Figure