Reconstruction-oriented methods

Data, transformations, validation and inference

Data sources and roles

Exact source files, hashes, inclusion tables and provenance paths are in the download ledger. Where a public accession or licence is unavailable, the release preserves that gap instead of inventing metadata.

Response construction

1. Use the already frozen 2,000-coordinate response array; the public package cannot verify where or when those HVGs were selected because the source H5AD and selection log are absent.
2. Aggregate cells to one target-level perturbation response; cells are observational support and the perturbation target is the inferential unit.
3. Eligibility is exactly 2,160 non-targeting-excluded targets with at least 20 cells, fixed before fold assignment. Fewer than 20 cells is the frozen low-support exclusion threshold for this benchmark.
4. Partition every realization into five non-overlapping 432-target outer folds, balanced by outcome-blind cell-count quintile and external-feature availability.
5. Fit imputation, standardization and α selection on the applicable training data only.
6. Compare every out-of-fold prediction with zero change and the outer-training response mean.

Feature construction

The safe model uses 16 DMD/context or curated priority values and 16 paired missingness indicators. “Response-independent” refers to the audited build path; acquisition chronology is not independently proven. DepMap and dataset-derived response summaries are excluded.

Historical-name note. Frozen feature identifiers that contain older terms such as rescue or priority are retained byte-for-byte for reproducibility. Their current semantics are counteralignment or historical input provenance; they are not therapeutic-rescue labels or current ranks.

Full provenance remains available in the feature table and leakage ledger.

Cross-validation and regularisation

• 20 prespecified realizations; each uses five non-overlapping balanced folds of 432 targets.
• Five-fold inner CV selects α from an expanded grid 0.01, 0.1, 1, 10, 100, 1000, 10000, 100000, 1e6, 1e7, 1e8 using training data only, plus an explicit train-mean-only comparator.
• For standardized design matrix X and 2,000-output response matrix Y, each fit minimizes ||Y − 1β₀ᵀ − XΒ||²F + α||Β||²F; the intercept is not penalized.
• Across 100 outer folds, α=1,000 was selected 26 times and α=10,000 was selected 74 times; 0/100 folds selected the finite-grid upper boundary. The former uniform α=1,000 boundary claim is withdrawn.
• Fold assignments are sensitivity realizations within one dataset, not biological replications.

Estimands and uncertainty

RMSE is calculated for each target over its 2,000-dimensional response vector. Raw cosine compares the predicted and observed target vectors. Residual cosine first subtracts the applicable outer-training mean response from both vectors. Primary contrasts are paired target-level differences versus the declared baseline.

The joint hierarchical bootstrap resamples realizations first and target modules second (10,000 replicates; seed 20260737). The public layer does not currently expose the target-module membership/count table required to rerun that exact second stage; the reported interval is auditable but not independently rebuildable from zero until that object is released.

The strict 16-outcome family requires paired RMSE support, multiplicity control and raw-direction support. All 16 rows—including model, split, endpoint, estimate, interval, adjusted P when defined and gate—are exposed on the Benchmark page and in typed JSON. The legacy 16/16 result is historical only.

Software and exact reconstruction

Commands, code paths, source hashes, environment records and expected outputs are distributed through the two release layers. Start with the release manifest, frozen protocol, model card and schema 1.1 dictionary. Package versions not present in a locked source object are explicitly not asserted here.