Federated Polygenic Risk Score

Performs genome-wide association studies across multiple biobanks without pooling individual-level data. Each site computes local summary statistics, which are then securely aggregated.

• Eliminates the need for a central data warehouse
• Enables inclusion of historically underrepresented populations
• Dramatically increases statistical power through larger virtual cohorts

Employs cryptographic secure multi-party computation (SMPC) to combine local model updates. The central server computes the weighted sum of gradient vectors without ever inspecting an individual institution's contribution.

• Protects against gradient leakage attacks
• Ensures no single point of data vulnerability
• Compatible with Federated Averaging (FedAvg) optimization

Integrates formal differential privacy (DP) bounds into the aggregation process. Calibrated statistical noise is injected into model updates, providing a mathematically provable guarantee against membership inference.

• Configurable epsilon budget for privacy-utility trade-offs
• Prevents reconstruction of individual genotypes
• Meets GDPR and HIPAA compliance requirements for cross-border studies

Architected for a cross-silo federated learning environment with a small number of reliable institutional participants—such as national biobanks or hospital networks—each possessing substantial local compute and curated genomic datasets.

• Assumes stateful clients with persistent availability
• Supports complex, multi-iteration training workflows
• Leverages high-bandwidth institutional connections for efficient communication

Addresses the critical preprocessing challenge of non-IID data across sites. A standardized pipeline maps disparate electronic health record (EHR) disease definitions to a common ontology before federated computation begins.

• Uses GA4GH standards for interoperability
• Resolves semantic mismatches in clinical coding
• Ensures a consistent analytical cohort across all participating nodes

Minimizes bandwidth bottlenecks by applying gradient sparsification and quantization to model updates before transmission. Only the most significant weight deltas are communicated, reducing payload size by orders of magnitude.

• Reduces per-round communication cost by 100-1000x
• Maintains model convergence guarantees
• Critical for scaling to millions of genetic variants

A mathematical framework that provides provable privacy guarantees for federated PRS models. By injecting calibrated noise into the aggregated effect size estimates, differential privacy ensures that the presence or absence of any single individual's genome in the training data is statistically indistinguishable.

• Privacy Budget (ε): A tunable parameter controlling the privacy-utility tradeoff; lower epsilon means stronger privacy
• Gaussian Mechanism: Adds noise proportional to the sensitivity of the PRS weight update function
• Local vs. Central DP: Local DP adds noise on each client before aggregation; central DP adds noise at the server after secure aggregation

The specific federated topology used for genomic PRS computation, involving a small number of reliable institutional participants such as national biobanks, hospital networks, or pharmaceutical research consortia.

• High Trust Environment: Unlike cross-device FL, silos are authenticated and audited entities
• Substantial Local Compute: Each silo possesses GPU clusters capable of training large PRS models locally
• Regulatory Compliance: Designed to satisfy GDPR, HIPAA, and local biobank governance requirements through data residency

The critical preprocessing step that standardizes disease definitions and clinical measurements across disparate electronic health record systems before federated PRS training can begin. Without harmonization, a model trained on inconsistent phenotype labels produces meaningless risk scores.

• Common Data Models: Mapping local EHR codes to standardized ontologies like OMOP or ICD-10
• Algorithmic Phenotyping: Using rule-based or ML-driven algorithms to identify cases and controls consistently
• Cohort Heterogeneity: Addressing differences in disease prevalence, ascertainment bias, and population stratification across sites

The precursor computation to federated PRS. A Federated Genome-Wide Association Study computes per-variant effect sizes and p-values across distributed biobanks without pooling individual-level data, providing the summary statistics needed to select and weight variants for the polygenic score.

• Meta-Analysis Approaches: Fixed-effects and random-effects models combine per-site association statistics
• Invariant Transformations: Only aggregate statistics like allele frequencies and effect estimates are shared
• PRS Derivation: The resulting federated summary statistics directly inform which SNPs are included in the risk score model