Architecture

FedGWAS uses Flower to coordinate a multi-client GWAS screening workflow. Each client keeps local genotype data and configuration. The server controls pipeline stages, distributes stage parameters, relays privacy-protected messages, and aggregates intermediate results.

Raw client datasets are not treated as a shared repository artifact. Cross-client operations rely on aggregate statistics, anonymized chunks, and the privacy mechanisms described in Privacy and Masking.

Selected references:

• Beutel et al., Flower: A Friendly Federated Learning Research Framework, arXiv, 2020.
• McMahan et al., Communication-efficient learning of deep networks from decentralized data, AISTATS, 2017.

Repository layout

pipeline/
  client_app.py              # Flower ClientApp entry point
  server_app.py              # Flower ServerApp entry point
  clients/
    base_client.py           # Chunking, anonymization, PLINK calls
    data_loder.py            # YAML loading and Flower state restoration
    local_qc.py              # Local QC and local LR helpers
    iterative_king.py        # Iterative KING chunk sending
    iterative_lr.py          # Iterative LR chunk sending and result mapping
    client_qc_aggregator.py  # Client-side global QC aggregation after decrypting peer shares
    client_to_client.py      # ECDH + AES envelope helpers for server-relayed messages
    c2c_payloads.py          # Typed client-to-client payload encoding
    seed_sync.py             # Global seed persistence and encrypted seed synchronization
    lr_privacy.py            # Tokenized local LR filtering helpers
    logger_manager.py        # Per-client logging
  server/
    strategy_strict.py       # Current federated server strategy
    aggregator_king.py       # Server-side KING aggregation
    aggregator_lr.py         # Server-side LR aggregation
    prg_masking.py           # ECC key exchange helpers used by encrypted relay
  utils/
    monitoring_config.py     # Monitoring settings resolution
    retention_config.py      # Retention settings resolution
    performance/             # Runtime performance monitors and CSV merge helpers
experiments/
  correctness/tiny_even/     # Default tiny two-client validation experiment
  performance/               # Small/medium scalability layouts
  real_world/                # Real genotype experiment layouts
configs/
  config_template.yaml       # Center configuration template
docs/website/
  docusaurus.config.js       # Docusaurus site configuration

Main components

Component	Responsibility
Flower ClientApp	Creates FedLRClient instances for partitions and runs stage-specific client logic.
Flower ServerApp	Creates FederatedGWASStrategy and configures the number of server rounds.
FederatedGWASStrategy	Drives the strict stage machine, forwards encrypted client-to-client payloads, and calls KING/LR aggregators.
DataLoader	Reads center YAML, creates output directories, and restores config from Flower state.
PLINK	Performs dataset filtering, missingness checks, genotype counts, KING, and logistic regression.
Encrypted relay	Uses encryption so seed, QC, relatedness-map, and association-filtering payloads can be relayed by the server without server-side decryption.
Monitoring and retention	Writes per-node metrics and can prune non-essential outputs after successful completion.

Data flow