Skip to main content

Association Screening

This module implements the federated association screening stage: local logistic-regression filtering with privacy-preserving tokens, then federated case–control logistic regression.

Local logistic-regression filtering​

During local_lr, each client runs:

plink --bfile <plink_prefix> --logistic --out <log_dir>/local_lr_temp

parse_insignificant_snps reads the resulting .assoc.logistic file and selects SNPs with P >= local_lr_threshold when configured, falling back to p_threshold. The client sends privacy-preserving (salted) tokens for those SNPs instead of raw SNP IDs.

The server forwards token arrays back to clients. During local_lr_filter_response, clients compute the intersection locally, resolve local SNP IDs, and remove that intersection from local data with exclude_snps.

Federated logistic regression​

After filtering, the workflow enters init_chunks_lr and iterative_lr:

  1. Clients repartition filtered data, usually by SNP.
  2. Clients send one encoded PLINK chunk per stage call.
  3. run_server_lr reconstructs and merges chunks.
  4. The server runs plink --logistic.
  5. The server parses .assoc.logistic and returns SNP p-values.
  6. Clients use persisted SNP map files to map anonymous SNP IDs back to local IDs when possible.
  7. Clients classify final p-values with global_lr_threshold when configured.

Current scope​

The current implementation focuses on logistic regression for binary phenotypes. Continuous phenotypes, richer covariate handling, population stratification, or alternate association tests should be added through the PLINK command construction, configuration schema, and result parsers.