CASCADE

CASCADE predicts whether a CRISPR hit observed in one lab will replicate in another — calibrated probability, honest uncertainty, before you spend a year chasing it.

CASCADE is a federated CRISPR-screen meta-analysis engine plus the Replication Oracle: given a screen hit (gene effect in context A) it returns a calibrated P(replicate in context B) with a conformal interval and an honest abstention when the context pair falls outside observed support. The provably-correct core is inverse-variance random-effects meta-analysis with provenance/QC weighting; the Oracle is a deliberately shallow head (logistic → isotonic → Mondrian conformal → kNN support gate) over real cross-lab labels.

---

Quick start

pip install cascade-oracle

from cascade.oracle import ReplicationOracle
oracle, meta = ReplicationOracle.load("artifacts/oracle_v0.pkl")   # trained_on_real_data: true
pred = oracle.predict_pair(pair)        # pair = a ReplicationPair (gene effect in context A)
print(pred.p_replicate, pred.lower, pred.upper, pred.abstained)

cascade meta '{"studies":[{"beta":-1.2,"variance":0.04},{"beta":-0.9,"variance":0.06}]}'
cascade benchmark --leaderboard      # score any model on the frozen Replication Benchmark v1
cascade mcp-serve                    # the splicr-mcp authority server (stdio)

---

Benchmark results

Frozen Replication Benchmark v1 (gene + cell-line + study-disjoint test set, real DepMap 26Q1 + Sanger + BioGRID-ORCS + Replogle). AUROC, all methods scored by the same harness:

Method	overall	cross_lab	cross_context	cross_study	cross_cell_type
mean (prior)	0.500	0.500	0.500	0.500	0.500
ridge	0.758	0.562	0.599	0.511	0.612
WBC (Billmann 2023)	0.596	0.431	0.459	0.646	0.470
MAIC (IC meta-analysis)	0.905	0.919	0.912	0.738	0.545
CASCADE Oracle	0.849	0.862	0.781	0.527	0.618

Honest reading (this is the point, not a footnote). A simple information-content meta-analysis (MAIC) — which is essentially an essentiality detector — beats the CASCADE Oracle on raw AUROC for most fitness strata. Much of "replication" in fitness screens reduces to "is this a common-essential gene," which a one-line IC score captures. CASCADE's distinct, defensible value is: (1) calibrated probabilities + honest abstention (MAIC/WBC emit uncalibrated raw scores); (2) it is the only method that conditions on the context pair and wins on the hardest stratum, cross-cell-type (0.618); (3) the replication-prediction framing with distribution-free conformal coverage. CASCADE beats WBC on cross_lab (0.862 vs 0.431) and beats ridge overall (0.849 vs 0.758).

The §HONESTY scientific gate (Oracle vs additive B4 + group-prior B5, 5 seeds): LOCO-cell-line PASS (Oracle 0.849 vs 0.71, +0.14 AUROC, p≈0, coverage 0.92); LOSO-study FAIL (0.41, reported plainly — see limitations).

---

⚠️ Limitations (read these — see docs/known_issues.md)

• LOSO-study AUROC ≈ 0.41 (worse than chance): cross-study replication over BioGRID-ORCS is near-unpredictable — heterogeneous phenotype readouts and per-study thresholds. Root-cause analysis in results/loso_failure_analysis.json.
• cross_study stratum AUROC 0.527 (near chance) for the same reason; label noise documented.
• High abstention (~40%): the Oracle declines outside the observed support manifold rather than guessing.
• cross_cell_type n = 213: real signal (p≈0.002) but modest and ceiling-limited — Replogle gives only two cell types (K562↔RPE1).
• Common-essentials ceiling: AUROC on fitness strata is inflated by easy pan-essential genes; a simple IC baseline (MAIC) exploits this and outperforms CASCADE there.
• Federation is a 2-party same-machine pilot (FSCP), not production.

---

Reproduce all results

pip install cascade-oracle
cascade benchmark --leaderboard     # reproduces the benchmark table from committed result JSONs
# Full regeneration from real data (requires the pinned sources, see Data):
python scripts/compute_wbc_baseline.py
python scripts/compute_maic_baseline.py
python scripts/run_ablation.py
python scripts/run_loso_analysis.py
python scripts/run_biological_interpretation.py
cascade phase3 --seeds 0 1 2 3 4    # frozen-benchmark evidence pack
cascade gate   --seeds 0 1 2 3 4    # the scientific gate
# Expected: cross-lab AUROC 0.862; runtime a few minutes on CPU per step.

Every result JSON carries trained_on_real_data: true + a provenance hash; scripts/ci_realness.sh enforces no fabricated data in product paths.

---

Data (all real, all pinned)

Every source is recorded in data/PROVENANCE.json with url + sha256 + release + license; nothing is substituted with synthetic data. Raw files are not committed (large; re-fetchable) — see data/acquire.py.

Source	Release	License
Broad DepMap (Chronos gene effect)	Public 26Q1	CC BY 4.0
Sanger Project Score (Chronos)	v2	CC BY 4.0
BioGRID-ORCS	2.0.18 (homo sapiens)	MIT / ORCS terms
scPerturb / Replogle 2022 (Perturb-seq)	Zenodo 13350497	CC BY 4.0
DepMap omics + STRING v12 + GO	26Q1 / v12 / current	CC BY 4.0

---

Citation

@software{cascade2026,
  title  = {CASCADE: a federated CRISPR-screen replication oracle},
  author = {Satani, Sahaj},
  year   = {2026},
  url    = {https://github.com/ssatanis/CASCADE},
  note   = {bioRxiv preprint TODO:DOI}
}

Baselines compared: WBC — Billmann et al. 2023, Cell Systems (PMID:37201508); MAIC — Baillie lab (github.com/baillielab/maic).

License

Apache-2.0 — see LICENSE.