pypi rustscenic 0.7.0
v0.7.0

one hour ago

RustScenic v0.7.0: exact TF binning in GRN inference

Released 19 November 2026.

  • GRN inference now gives every distinct TF value its own bin (up to 65,535
    per TF), so candidate split thresholds match scikit-learn's.
    RustScenic 0.6 merged values into at most 255 bins. In a pre-specified
    replication at 20,000 cells, that quantisation made target rankings diverge
    from arboreto's by more than arboreto's own seed-to-seed variation (protocol
    fidelity 0.74 to 0.84 against a 0.90 criterion). Exact binning gives 0.95 to
    0.99 in post hoc refits.
  • Networks therefore change from v0.6. max_bins=255 (API), grn --max-bins 255
    or pipeline(..., grn_max_bins=255) reproduces v0.6 networks bit for bit.
  • New early_stop_mode="paired_oob" applies arboreto's 25-round monitor to
    same-mask out-of-bag improvements, the definition scikit-learn used before
    1.3. The default stopping rule is unchanged ("arboreto").
  • Cost: split search now computes gains only for the bins a node occupies,
    which does not change results. On the bone-marrow replication input (20,000
    cells, 5,000 genes, 1,069 TFs; 10-core laptop, two alternating runs each),
    GRN inference took 284.6 to 285.5 s against 271.7 to 272.0 s for v0.6.1
    (4.9% longer), with the same 0.66 GB peak memory. The binned TF matrix takes
    2 bytes per cell per TF instead of 1, which matters only at large cell and
    TF counts.
  • The audited rustscenic validated recipes pin max_bins=255, so they keep
    reproducing their audited results.

AUCell, cisTarget, topics, ATAC preprocessing, enhancer links and eRegulon
assembly are unchanged. Published workflow benchmarks were measured with the
v0.5 and v0.6 engines.

Install

python -m pip install --upgrade rustscenic

Changelog · Benchmarks and biological results

Don't miss a new rustscenic release

NewReleases is sending notifications on new releases.