Horizontal Federated Learning

All participating nodes must operate on a shared, pre-harmonized feature space. In genomics, this means identical variant calling pipelines, standardized reference genomes (e.g., GRCh38), and consistent annotation formats across institutions.

• Each client's local dataset has the same columns, data types, and encoding
• Enables direct weight averaging without feature translation layers
• Requires rigorous phenotype harmonization before training begins
• Common in multi-hospital GWAS consortia studying the same disease

The defining characteristic of HFL is sample-space partitioning. Each institution contributes different patients or specimens, ensuring no single entity's records exist at multiple nodes. This is the natural topology for genomic biobanks.

• Patient A exists only at Hospital 1; Patient B only at Hospital 2
• Eliminates the need for entity resolution or private record linkage
• Increases statistical power by pooling diverse populations
• Introduces non-IID challenges when demographics or sequencing protocols differ across sites

HFL is the native topology for the FedAvg algorithm, where each client trains a local copy of the model on its own samples, then transmits only weight updates to the aggregation server.

• Local training: each hospital computes gradients on its private genomic data
• Secure aggregation: the server computes a weighted mean of updates
• Weights are proportional to local dataset size to account for sample imbalance
• Converges reliably when local data distributions are approximately similar

HFL is the backbone of privacy-preserving collaborative genomics, enabling multi-institutional studies that would be impossible under centralized data-sharing agreements.

• Federated GWAS: jointly computing variant-trait associations across biobanks
• Polygenic risk score training: building more diverse, generalizable risk models
• Variant pathogenicity classification: training classifiers on rare disease cohorts
• Gene expression prediction: multi-center training of models like Enformer
• Real-world example: the GA4GH community standards explicitly support HFL topologies for cross-institutional genomic queries

HFL provides a baseline privacy improvement by keeping raw sequence data local, but weight updates can still leak information about training samples through model inversion or membership inference attacks.

• Defense layers: differential privacy noise injection on gradient updates
• Secure aggregation prevents the server from inspecting individual contributions
• Trusted execution environments protect computation at each node
• Honest-but-curious server model is the standard threat assumption
• Cryptographic guarantees require combining HFL with SMPC or homomorphic encryption

HFL contrasts sharply with vertical federated learning (VFL), where institutions hold different features about the same set of entities. Understanding this distinction is critical for architectural decisions.

• HFL: same features, different samples → weight averaging works directly
• VFL: different features, same samples → requires entity alignment and split neural networks
• HFL is simpler to implement but requires feature-space standardization
• VFL is common when one hospital has genomic data and another has imaging for the same patients
• Hybrid federated transfer learning addresses scenarios fitting neither pure HFL nor VFL

The foundational optimization algorithm for horizontal federated learning. Each client trains a local model on its own genomic samples, and the central server aggregates these updates by computing a weighted average based on each client's dataset size. This ensures that larger, more statistically significant cohorts have proportionally greater influence on the global model without ever seeing the underlying DNA sequences.

A critical obstacle in genomic consortia where local datasets are not independently and identically distributed. Different hospitals may sequence patients with distinct ethnic backgrounds, disease subtypes, or using different sequencing machines. This distributional skew can cause naive federated averaging to diverge or fail to converge, requiring specialized optimization strategies like FedProx or SCAFFOLD to stabilize training.

A cryptographic protocol that ensures the central server can only compute the sum of encrypted model updates from all participating institutions, without ever being able to inspect any single institution's contribution. This prevents an honest-but-curious server from reconstructing genomic features through gradient leakage attacks, providing a critical second layer of defense beyond raw data isolation.

A mathematical framework that injects calibrated statistical noise into model updates before transmission. By clipping gradient norms and adding Gaussian noise, the system provides a provable (ε, δ)-differential privacy guarantee, ensuring that the final model's parameters are statistically indistinguishable whether or not any single patient's genome was included in the training cohort.

The dominant architectural pattern for genomic federated learning, involving a small number of reliable, stateful institutional clients—such as major research hospitals or national biobanks—each possessing substantial on-premise GPU compute. Unlike cross-device FL with millions of unreliable smartphones, cross-silo genomic FL assumes persistent connectivity and allows for more complex synchronous training protocols.

A primary bottleneck in horizontal FL where transmitting full model weights for a genomic foundation model with hundreds of millions of parameters can saturate institutional network links. Techniques like gradient compression, sparsification, and quantization reduce the bitrate of updates, while local SGD increases the number of local training steps between communication rounds to minimize overhead.