Source-linked AI summary
Towards Global Federated Genome-Wide Association Meta-Analysis Using GA4GH TES
Abhijit Chunduru, Matthew Joel, Zilinghan Li, Ravi Madduri
TL;DR
International GWAS needs large, diverse cohorts, but privacy, consent, and data-residency constraints impede centralized genotype pooling. This paper combines APPFL, GA4GH TES, and HiveWatch to execute and monitor federated GWAS meta-analysis while keeping individual-level data local. In a five-site simulation, the approach recovered expected association signals without genotype centralization and showed calibrated test statistics.
Problem
Privacy, consent, and national data-residency constraints limit centralized pooling of individual-level genotypes needed for large, diverse international GWAS.
Method
The pipeline runs local GWAS through APPFL-orchestrated GA4GH TES endpoints, shares aggregate statistics, performs inverse-variance-weighted fixed-effect meta-analysis, and monitors execution with HiveWatch.
Results
14 Type 2 Diabetes and 33 Body Mass Index associations reached p < 5×10−8, with λGC = 1.063 for Type 2 Diabetes and 1.056 for Body Mass Index.
Takeaways & Limitations
The combined federated-learning and GA4GH-standard pipeline provides a deployable infrastructure for privacy-preserving GWAS meta-analysis across international institutions.
Abstract
from arXiv · showhide
Genome-wide association studies (GWAS) gain statistical power from large, ancestrally diverse cohorts, but privacy regulations and data-residency constraints often prevent genomic data from being centrally pooled across institutional or national borders. We present a privacy-preserving federated GWAS meta-analysis pipeline built on the APPFL framework, in which each site computes local GWAS summary statistics and transmits only aggregate results, never individual-level genotypes. Analysis is executed through a global network of Global Alliance for Genomics and Health (GA4GH) Task Execution Service (TES) endpoints, which allows computation to move to the data rather than the reverse. The server performs inverse-variance-weighted fixed-effect meta-analysis and returns aggregated results to all sites, while HiveWatch, our developed geographic observability toolkit, provides real-time monitoring of distributed task execution. In a five-site simulation over 100,000 synthetic individuals and roughly 240,000 variants for Type 2 Diabetes and Body Mass Index, the federated meta-analysis reproduces the association signal expected from a pooled analysis without centralizing any genotype data, showing that standards-based task execution and federated learning enables a practical privacy-preserving infrastructure for international GWAS meta-analysis.
I. INTRODUCTION
Large, diverse GWAS cohorts improve statistical power, but privacy, consent, and data-residency constraints limit centralized genotype pooling. The paper addresses the infrastructure gap with APPFL, GA4GH TES endpoints, and HiveWatch for federated cross-border analysis.
- International GWAS consortia improve power through larger and more ancestrally diverse cohorts, while regulations and residency rules often prohibit exporting individual-level genotypes.
- The pipeline combines APPFL for federation and aggregation, GA4GH TES for interoperable bring-compute-to-data execution, and HiveWatch for real-time monitoring.
- Each site executes analysis locally and returns only summary statistics across an international network of TES endpoints.
A. Federated Meta-Analysis
APPFL runs local GWAS analyses and combines transmitted per-variant aggregates using inverse-variance-weighted fixed-effect meta-analysis. The federated results recover expected association signals without genotype centralization and remain calibrated in the reported simulation.
- Each site computes local GWAS statistics and returns effect sizes, standard errors, allele frequencies, and sample sizes without sharing individual-level genotypes or phenotypes.The pipeline uses ordinary-least-squares GWAS for BMI and a covariate-adjusted score test for Type 2 Diabetes.
- Inverse-variance-weighted fixed-effect aggregation combines site-level estimates into federated Z-statistics and p-values.The method uses per-site weights wi = 1/SEi^2 and assesses genome-wide significance at p < 5 × 10−8.
B. TES Endpoint Network
The pipeline deploys Funnel-based GA4GH TES endpoints in distinct geographic regions and uses their common API to submit, monitor, and retrieve containerized local analyses. Authentication and authorization restrict endpoint access to the trusted federated server.
- Each federated site runs a Funnel-based TES endpoint on cloud compute in a distinct geographic region.
- TES provides a standardized API that lets APPFL orchestrate portable bring-compute-to-data workflows across heterogeneous computing environments.
- Endpoint authentication and authorization allow only the trusted federated server to submit, monitor, and retrieve analysis-task results.
- Individual-level genomic data remain local while APPFL interacts with participating sites through a common TES interface.
C. Geographic Observability with HiveWatch
HiveWatch provides geographic observability for distributed execution across TES endpoints. It records task events, visualizes them on a live map, supports replay, and helps diagnose endpoint-level execution problems.
- HiveWatch records task dispatch, execution, and completion events across distributed endpoints.
- The toolkit visualizes distributed task execution on a live geographic map during runs and supports full replay from logs afterward.
- Endpoint-level visibility makes stalls, stragglers, and failures diagnosable across heterogeneous infrastructures.
III. EVALUATION AND RESULTS
A five-site simulation used synthetic cohorts and traits to evaluate the federated GWAS pipeline. The analysis recovered genome-wide association signals without genotype centralization, with calibrated test statistics for both traits.
- Simulation design: 100,000 synthetic individuals and 239,801 variants were partitioned across five deliberately unequal sites for the evaluation.The site sizes ranged from 6,951 to 40,752 individuals.
- Simulation design: The simulation generated Type 2 Diabetes and Body Mass Index phenotypes from PGS Catalog scores with age and sex covariates.
- Association results: 14 Type 2 Diabetes associations and 33 Body Mass Index associations reached p < 5×10^-8 without genotype centralization.The leading Type 2 Diabetes signals were chr20:43.04 Mb (p = 1.6×10^-34) in HNF4A and chr10:114.75 Mb (p = 1.3×10^-24) in TCF7L2.
- Association results: λGC = 1.063 for Type 2 Diabetes and λGC = 1.056 for Body Mass Index, indicating calibrated test statistics without federation-related inflation.Most variants followed the null diagonal, with true signals departing in the tail.
IV. CONCLUSION
The work combines federated learning with GA4GH standards for privacy-preserving GWAS meta-analysis across international institutions. APPFL, TES, and HiveWatch provide federation, interoperable computation, and operational observability, respectively.
- The pipeline combines FL and GA4GH standards to support deployable, privacy-preserving GWAS meta-analysis across international institutions.
- APPFL provides federated orchestration and exact fixed-effect meta-analysis, TES supports bring-compute-to-data execution, and HiveWatch provides real-time observability.
- The evaluation uses synthetic data, with future work targeting real multi-institution deployments, more sophisticated meta-analysis models, stronger privacy guarantees, and resilient execution.