Federated learning for biobank data at the CMU-NVIDIA Hackathon

This article describes ten federated learning projects developed at the CMU-NVIDIA Hackathon for biomedical applications, demonstrating how sensitive biobank data can be analyzed collaboratively while preserving privacy.

• FedGen: Generates synthetic genomic datasets for testing federated genome-wide association studies
• FedPathHarmony: Harmonizes histopathology images across sites using stain normalization techniques
• FedViz: Provides visualization dashboard for auditing biobank data harmonization and readiness
• OmniGenome: Builds federated pangenome graphs without exposing raw sequence data
• Med_SNP_Deconvolution: Enables ancestry classification across sites without centralizing genotype data
• RAIDers: Discovers rare disease subtypes using synthetic cohorts of 8,000+ simulated patients
• OncoLearn: Performs multi-omic cancer subtyping using federated transfer learning
• PRSAggregator: Harmonizes polygenic risk score computation across diverse populations
• MuFFLe: Integrates multimodal data (RNA, clinical, imaging) for cancer prognosis prediction
• FedProFit: Predicts protein fitness across distributed datasets using frozen language models

These projects demonstrate that federated learning enables biobanks to collaborate on sophisticated AI research while maintaining data sovereignty and privacy compliance.