Federated Generalized Linear Mixed Models for Collaborative Genome-wide Association Studies

TL;DR

This paper introduces dMEGA, a novel federated method for genome-wide association studies that preserves privacy, accounts for population structure, and efficiently estimates genetic associations across multiple sites.

Contribution

dMEGA is a new federated generalized linear mixed model approach that enables privacy-preserving, flexible, and efficient association testing in multi-site genomic studies.

Findings

Accurately detects genetic associations in simulated datasets.

Demonstrates efficiency and robustness on real genomic data.

Effectively accounts for population stratification and disease etiology.

Abstract

As the sequencing costs are decreasing, there is great incentive to perform large scale association studies to increase power of detecting new variants. Federated association testing among different institutions is a viable solution for increasing sample sizes by sharing the intermediate testing statistics that are aggregated by a central server. There are, however, standing challenges to performing federated association testing. Association tests are known to be confounded by numerous factors such as population stratification, which can be especially important in multiancestral studies and in admixed populations among different sites. Furthermore, disease etiology should be considered via flexible models to avoid biases in the significance of the genetic effect. A rising challenge for performing large scale association studies is the privacy of participants and related ethical concerns of stigmatization and marginalization. Here, we present dMEGA, a flexible and efficient method for performing federated generalized linear mixed model based association testing among multiple sites while underlying genotype and phenotype data are not explicitly shared. dMEGA first utilizes a reference projection to estimate population-based covariates without sharing genotype dataset among sites. Next, dMEGA uses Laplacian approximation for the parameter likelihoods and decomposes parameter estimation into efficient local-gradient updates among sites. We use simulated and real datasets to demonstrate the accuracy and efficiency of dMEGA. Overall, dMEGA's formulation is flexible to integrate fixed and random effects in a federated setting.