v0.1.1 — Initial archived release

First public release of isp-confound-toolkit, accompanying the methods preprint.

A confound-diagnostic toolkit for judging when foundation-model in-silico
perturbation predictions can be trusted — with Geneformer as a worked
cautionary case.

What's included

• path-D — a machine-precision re-implementation of the per-cell knockout
  response, validated against the official InSilicoPerturber to Pearson
  r = 1.000000 (max abs error 3e-7). Responses are cached and resampled in
  pure NumPy, enabling a controlled coverage-to-estimate analysis.
• Universal-responsiveness control — subtracts the all-perturbation mean so
  hub/high-expression genes don't masquerade as perturbation-specific hits.
• Tokenization-coverage gate — gates in-silico knockouts by a target's real
  top-4096 token coverage.
• Library-size contamination check — catches sequencing-depth artifacts in
  raw-count log2FC.
• De-circularization matrix — separates genuine signal from circular readouts.
• Supporting utilities (isp_confound/): a thin Geneformer wrapper and an
  in-silico knockout engine.

Reproducibility

Curated figures and key CSVs are committed under results/. The large cached
intermediates (path-D caches, DESeq2 outputs, ~465 MB) that reproduce every
figure and increment test in pure NumPy — no GPU — are on Zenodo:
https://doi.org/10.5281/zenodo.20729460

Datasets

Public scPerturb h5ad files (Frangieh et al. 2021; Replogle et al. 2022).
See DATA.md.

License

MIT. If you use this toolkit, please cite it (see CITATION.cff) and Geneformer
(Theodoris et al., Nature 2023) and scPerturb (Peidli et al., Nat Methods 2024).

v0.1.0 — Initial public release

First public release of isp-confound-toolkit, accompanying the methods preprint.

A confound-diagnostic toolkit for judging when foundation-model in-silico
perturbation predictions can be trusted — with Geneformer as a worked
cautionary case.

What's included

• path-D — a machine-precision re-implementation of the per-cell knockout
  response, validated against the official InSilicoPerturber to Pearson
  r = 1.000000 (max abs error 3e-7). Responses are cached and resampled in
  pure NumPy, enabling a controlled coverage-to-estimate analysis.
• Universal-responsiveness control — subtracts the all-perturbation mean so
  hub/high-expression genes don't masquerade as perturbation-specific hits.
• Tokenization-coverage gate — gates in-silico knockouts by a target's real
  top-4096 token coverage.
• Library-size contamination check — catches sequencing-depth artifacts in
  raw-count log2FC.
• De-circularization matrix — separates genuine signal from circular readouts.
• Supporting utilities (isp_confound/): a thin Geneformer wrapper and an
  in-silico knockout engine.

Reproducibility

Curated figures and key CSVs are committed under results/. The large cached
intermediates (path-D caches, DESeq2 outputs, ~465 MB) that reproduce every
figure and increment test in pure NumPy — no GPU — are on Zenodo:
https://doi.org/10.5281/zenodo.20729460

Datasets

Public scPerturb h5ad files (Frangieh et al. 2021; Replogle et al. 2022).
See DATA.md.

License

MIT. If you use this toolkit, please cite it (see CITATION.cff) and Geneformer
(Theodoris et al., Nature 2023) and scPerturb (Peidli et al., Nat Methods 2024).

---

几点说明 / 你可能要调的

1. Tag 用 v0.1.0。比 v1.0.0 更诚实——1.0.0 通常暗示"稳定/完整",而你还在预印本阶段。和 pyproject 的 version = "0.1.0" 也一致。
2. "accompanying the methods preprint" 这句:如果 release 在 bioRxiv 上线之前建,可改成 "accompanying an upcoming methods preprint";上线后可在 release notes 里补 bioRxiv DOI。
   in-silico knockout engine.

Reproducibility

Curated figures and key CSVs are committed under results/. The large cached
intermediates (path-D caches, DESeq2 outputs, ~465 MB) that reproduce every
figure and increment test in pure NumPy — no GPU — are on Zenodo:
https://doi.org/10.5281/zenodo.20729460

Datasets

Public scPerturb h5ad files (Frangieh et al. 2021; Replogle et al. 2022).
See DATA.md.

License

MIT. If you use this toolkit, please cite it (see CITATION.cff) and Geneformer
(Theodoris et al., Nature 2023) and scPerturb (Peidli et al., Nat Methods 2024).