BEATRICE: Bayesian fine-mapping from summary data using deep variational inference.

Area of Science:

• Genetics
• Computational Biology
• Statistical Genetics

Background:

• Identifying causal variants from genome-wide association studies (GWAS) is difficult due to variant sparsity and high linkage disequilibrium.
• Existing methods struggle to accurately pinpoint causal variants in complex genetic regions.

Purpose of the Study:

• Introduce BEATRICE, a novel framework for identifying putative causal variants from GWAS summary statistics.
• Develop a robust method to address the challenges of variant sparsity and correlation in fine-mapping.

Main Methods:

• Utilize a hierarchical Bayesian model with a binary concrete prior on causal variants.
• Employ a variational algorithm minimizing KL divergence for causal configuration inference.
• Integrate a deep neural network as an inference machine for parameter estimation.
• Implement a stochastic optimization procedure to compute posterior inclusion probabilities and credible sets.

Main Results:

• BEATRICE demonstrates superior coverage with comparable power and set sizes compared to state-of-the-art methods.
• Performance gains of BEATRICE increase with a higher number of causal variants.
• BEATRICE successfully identified the APOE ε2 allele, a known Alzheimer's disease risk factor, which was missed by baseline methods.

Conclusions:

• BEATRICE offers a powerful and accurate approach for fine-mapping causal variants in GWAS.
• The framework's ability to identify key variants like APOE ε2 highlights its clinical relevance for diseases like Alzheimer's.