FedFM: Federated Cross-Ancestry Fine-Mapping 

Background

A genome-wide association study (GWAS) points to a region of the genome; fine-mapping narrows it to the few variants likely to be causal, reported as a credible set. That inference depends on the correlation structure between nearby variants - linkage disequilibrium - which cannot be reconstructed from the per-variant summary statistics that meta-analysis shares. Fine-mapping is therefore one of the few common genomic analyses that data-sharing restrictions genuinely block, and it is most valuable exactly where those restrictions bite hardest: across ancestrally diverse cohorts, where differences in linkage disequilibrium between populations sharpen the credible set.

This project federates it without approximation. Each site computes a fixed set of aggregate quantities over its own genotypes and phenotypes, on its own hardware, and returns only those; a coordinator combines them and fits the multi-ancestry fine-mapping model SuSiEx. The combined fit is mathematically identical to one computed on pooled individual-level data, so federating costs nothing in statistical power.

Four GA4GH standards make such a run governable rather than merely possible: Data Use Ontology (DUO) for whether a site's terms permit the study, Data Repository Service (DRS) for which data were analysed, Tool Registry Service (TRS) for which tool version analysed them, and Task Execution Service (TES) for dispatching it.

Ongoing Work

The software runs today across Argonne National Laboratory (US), Covenant University (Nigeria) and Mohamed bin Zayed University of Artificial Intelligence (UAE) - three continents, mixed schedulers - on synthetic cohorts, so the entire pipeline is publicly reviewable.

Current priorities: a v1.0 release that any third party can run from a generated partner bundle (November 2026); the full three-site production run, with the federated and pooled fits checked for equality at every locus (December 2026); one site joining over a TES service instead of a Globus Compute endpoint; and submitting the project's DUO consent-code and DRS bundle-identifier conventions for work stream review (February 2027).

How to participate

We are looking for sites holding genomic data they cannot share, and for anyone operating a TES service willing to serve as an execution backend. A participating site installs two Python packages, authorises one identity, and runs a bundle generated with its configuration already resolved - no VPN, no inbound network access, and no data leaves the institution. Contributors to the method and reviewers of the DUO and DRS profiles are equally welcome.

References

appfl-bio-suite - implementation, documentation and partner setup guides: https://github.com/mfjoel01/appfl-bio-suite

APPFL - the federated learning framework it builds on: https://github.com/APPFL/APPFL (documentation: http://appfl.rtfd.io/)

Yuan K, Longchamps RJ, Pardinas AF, et al. Fine-mapping across diverse ancestries drives the discovery of putative causal variants underlying human complex traits and diseases. Nature Genetics 56, 1841-1850 (2024). https://doi.org/10.1038/s41588-024-01870-z

Wharrie S, et al. HAPNEST: efficient, large-scale generation and evaluation of synthetic datasets for genotypes and phenotypes. Bioinformatics 39(9) (2023). https://doi.org/10.1093/bioinformatics/btad535