@apsal/cohortlens-analytics

0
1
flwr new @apsal/cohortlens-analytics

CohortLens Analytics

CohortLens Analytics is a Flower FederatedApp for a bounded pharmacogenomics study using public GDSC2 cancer cell-line data. It asks whether the BRAF p.V600E mutation is associated with response to PLX-4720. Drug response is measured as LN_IC50; a lower value means greater sensitivity in the assay.

This is an emulated federation over public laboratory data. It is not a patient study, a clinical trial, or a recreation of the original GDSC institutions.

How it works

The ServerApp sends one approved, hash-bound analysis contract to three deployment SuperNodes. Each ClientApp:

  1. validates the contract and its time-limited operator approval;
  2. loads only its node-local GDSC snapshot;
  3. fits the fixed OLS model with HC3 robust standard errors; and
  4. returns an allowlisted summary containing the mutation coefficient, uncertainty, sample count, and status.

The ServerApp validates all replies and combines compatible coefficients with inverse-variance fixed-effect meta-analysis. Raw rows, cell-line identifiers, and design matrices are not returned to the coordinator.

flowchart LR
    S["ServerApp<br/>PLX-4720 / BRAF V600E"] --> N1["SuperNode 1<br/>local GDSC2 partition"]
    S --> N2["SuperNode 2<br/>local GDSC2 partition"]
    S --> N3["SuperNode 3<br/>local GDSC2 partition"]
    N1 -->|"effect, HC3 SE, N"| S
    N2 -->|"effect, HC3 SE, N"| S
    N3 -->|"effect, HC3 SE, N"| S

If the released site estimates agree (I² < 0.50), the study finishes. If they disagree and there is enough evidence, a separate CohortLens AgentApp may propose the single predeclared follow-up: repeat the analysis separately for SKCM and non-SKCM cancer lineages. That proposal is not authorization; an operator must approve a new contract before another analytics run.

Data and deployment

The input snapshots are prepared from GDSC2 release 8.5-27Oct23, the Cell Model Passports model list, and its compact driver-mutation summary. Each SuperNode must receive a separate read-only NPZ snapshot and local node configuration values for site-alias, dataset-path, and snapshot-alias. Dataset paths are never placed in the run configuration or FAB.

Preparation and orchestration are provided by the full CohortLens repository. Its README lists the source files, preparation command, SuperNode configuration, contract approval flow, and result collection commands.

Build this app from the repository root:

uv run python scripts/stage_apps.py
uv run flwr build --app apps/cohort-analytics

Run it on a configured deployment federation:

uv run flwr run apps/cohort-analytics supergrid \
  --federation @ACCOUNT/FEDERATION \
  --run-config data/gdsc/prepared/analytics-run.toml --stream

The GDSC workflow requires deployment SuperNodes with prepared local data; a plain Flower simulation does not supply those snapshots.

Scope and limitations

This workflow estimates an association in public, in-vitro cancer cell-line assays. It does not establish a drug effect in people, clinical benefit, causality, or a clinically validated biomarker. The three partitions are virtual laboratories rather than independent replications.

Local rows are not transmitted, but the coordinator does receive site-level summaries. The app does not provide secure aggregation, differential privacy, or protection from dishonest participants. The compact mutation summary also does not justify calling the comparator group definitively BRAF wild type.