Linear probe

[lagillenwater]

Summary

Adds the harness to evaluate whether Stack-generated treated transcriptomes predict
PDTO drug response (AUC), with published methods and baselines to compare against.

What's here

• Stack gene panel + embedding input (stack_panel.py, build_stack_panel.py,
  prep_stack_input.py) — map CoderData genes onto Stack's 15,012-gene vocabulary and
  write organoid/cell-line cohorts as pseudo-cells for embedding/generation.
• Viability signatures (signatures.py, fetch_hallmark.py,
  data/static/hallmark_signatures.gmt) — curated death/proliferation sets and a loader
  for the MSigDB Hallmark sets (p53/apoptosis up, E2F/G2-M down), scored with a
  random-gene-set negative control + within-drug label-permutation null.
• L1000 builders + viability-bridge gate (l1000.py, validate_bridge_readout.py,
  bridge_generated_to_auc.py, build_l1000_context.py) — one shared code path that
  (a) pairs real L1000 treated−DMSO deltas with GDSC2 AUC and (b) turns Stack-generated organoid deltas into AUC
  predictions. The .gctx is read in chunks (bounded memory); cmapPy is lazy-imported.
• Modular viability adapters (adapters.py, score_viability_adapters.py) — three
  published transcriptome→viability methods, selectable (--methods, default all), each
  carrying its citation:
  • hallmark — single-sample signature scoring (Liberzon et al. 2015; Barbie et al. 2009)
  • szalai — L2 linear regression on the perturbation signature (Szalai et al., NAR 2019)
  • xgboost — elastic-net gene selection + gradient-boosted trees (Lu, Chen & Qin,
    BMC Bioinformatics 2021)
    Supervised adapters train on real L1000Δ→GDSC2 AUC and transfer to the generated
    organoid deltas (each cohort z-scored to cross the platform gap).
• Baselines (baselines_soragni.py, predict_expression_baselines.py, transfer
  scripts) — drug-mean floor, GDSC2→Soragni transfer, and expression-prediction
  baselines, scored identically (global + interaction rho).

Notes

• Green: 89 pytest, ruff, and pyright (strict on src/tests).
• Adds the xgboost dependency (needs OpenMP/libomp locally).