Federated Learning in Genomic Research

Abstract

Genomic research requires large, diverse cohorts for statistically powered discovery, yet privacy regulations and institutional data-governance policies prevent centralisation of sensitive genetic data, and federated learning that trains models across distributed datasets without sharing raw data offers a privacy-preserving alternative, yet adoption into genomic research varies across application domains and federation architectures. We evaluated 216 federated learning programmes for genomic research across centres in Switzerland, Austria, and France between 2018 and 2022, spanning five application categories: federated genome-wide association studies, federated variant calling and quality control, federated polygenic risk score training, federated single-cell analysis, and federated clinical genomics classifiers. A Federated Genomic Learning Index (FGLI) was constructed from five sub-scores -- model accuracy relative to centralised training, privacy guarantee strength, communication efficiency, heterogeneity robustness across sites, and regulatory compliance -- with weights from regression against sustained adoption into multi-site genomic pipelines. FGLI correlated with adoption at r = +0.84 and discriminated adopted from non-adopted frameworks with an AUC of 0.884. Federated GWAS scored highest (mean FGLI 0.826), while federated single-cell analysis trailed at 0.600. Only 35.6 percent exceeded the 0.75 threshold. Model accuracy carried the largest weight (beta = +0.278), followed by privacy guarantee strength (beta = +0.230).