SGM scBoundary Supplementary Software for Frozen Selective-Prediction Evaluation
Description
This repository contains the article-focused supplementary software and audit materials supporting the manuscript “Lower Error Need Not Preserve Biological Fidelity: A Frozen Selective-Prediction Evaluation of Single-Cell Perturbations.” The package includes frozen protocols and configurations, scGen Article Gate I analysis code, validation tests, result receipts, figure-generation scripts and outputs, environment records, A/B-machine provenance, SHA256 manifests, reproduction instructions, and a BSD-3-Clause license. The frozen Article Gate I decision is NO-GO and remains unchanged. The repository does not include H5AD source data, Tahoe-100M, McFarland data, the separate SGM-DEGAWARE branch, OP3 full counts, or de_test.h5ad. Public source-data access is documented in the manuscript and package. The included materials support audit and reproducibility without authorising Gate II or additional scientific computation.
Files
Steps to reproduce
1. Download SGM_Supplementary_Software_v1.2.zip and verify its SHA256: 60155a2becef611db728fccf9669ff6f9fb208701dc63a45a0f85b245b6dcc62 2. Extract the archive into a new directory. Begin by reviewing: 04_RESULTS/manuscript/SGM_AB_Computational_Provenance_v1.2.md 00_PROTOCOL/SCGEN_ARTICLE_GATE_I_v0.1.yaml 04_RESULTS/gate0b_input/sciplex3_article_gate_i_receipt.json 3. Create the Python environment: conda create --name sgm-scgen python=3.9.18 conda activate sgm-scgen python -m pip install -r 02_ANALYSIS/requirements-article.txt 4. From the extracted archive root, verify package integrity and run the tests: python verify_package_integrity.py python -m unittest discover -s 02_ANALYSIS/tests -v 5. If the pinned R environment is available, run: Rscript 02_ANALYSIS/scripts/verify_r_environment.R Rscript 02_ANALYSIS/scripts/gate_i_limma_voom.R --self-test 6. Rebuild the manuscript figures from the archived result tables: python figure_generation/generate_corrected_manuscript_figures.py python figure_generation/generate_corrected_manuscript_figures_svg.py Full independent recomputation requires separately obtaining the public sci-Plex3 input from GEO accession GSE139944 or scPerturb v1.4 and matching the archived byte, MD5, and SHA256 records. Large H5AD files are not redistributed. The frozen pipeline phases are prepare, predict, freeze, and reveal. Article Gate I remains a binding NO-GO; frozen thresholds must not be relaxed, and locked OP3 outcomes must not be accessed as a rescue analysis.