A Differentially Private Federated Learning Framework for Secure Genome-Wide Association Studies Across Distributed Clinical Registries

Genome-wide association studies (GWAS) are critical for discovering genetic variants associated with complex diseases, yet their success depends heavily on aggregating large-scale datasets from diverse populations. Centralizing sensitive genomic and clinical data across multiple institutions faces severe regulatory, ethical, and privacy challenges. To address this bottleneck, we present a novel, differentially private federated learning (DP-FL) framework designed for secure, distributed GWAS across heterogeneous clinical registries. Our framework enables collaborative model training—specifically logistic and linear mixed models for genetic association—without transferring raw genomic or phenotypic data. By integrating local differential privacy mechanisms with secure cryptographic aggregation, we provide mathematically provable protection against membership inference and reconstruction attacks, even in the presence of curious or semi-honest central servers. We evaluated our framework using semi-synthetic datasets derived from the 1000 Genomes Project distributed across simulated clinical sites. The results demonstrate that our DP-FL approach achieves statistical power and association mapping accuracy comparable to traditional centralized analysis, while maintaining a robust and configurable privacy budget. This work establishes a scalable, secure paradigm for multi-institutional genomic research, paving the way for collaborative personalized medicine without compromising patient confidentiality.