RustScenic v0.6.0: atlas workflows and verified activity scoring
Released 9 November 2026.
Built and scaled with the Huang Lab at Mount Sinai. Existing authorship,
contribution records and Apache-2.0 licensing are unchanged.
What changes
- A
rustscenic validatedcommand connects the audited RNA and paired
RNA+ATAC recipes, records settings and reports missing/filtered regulators. - Optional seeded AUCell matches the reference tie-order convention without
densifying expression. Existing unseeded callers retain their outputs. --activity-mode directionalpreserves positive and negative target
activity separately. The inferred networks and historical paired defaults
are unchanged. RNA already retains separate directions.- Updated figures and reproducible evidence cover 21 matched workflow pairs,
real-data case studies and independent scoring confirmation.
Verified benchmarks
| Configured workflow | Real cells | Faster | Lower measured peak memory |
|---|---|---|---|
| RNA+ATAC vs SCENIC+ | 817,740 | 14.4× | 57.1% |
| RNA vs pySCENIC | 1,306,127 | 3.98× | 96.6% |
| Core GRN vs arboreto Dask | 1,306,127 | 3.06× | 78.7% |
These historical measurements use v0.5.0, three seeds and 16 CPUs. Workflow
timings include prepared-input loading through export; the separate core-GRN
clock excludes loading. Selected panels and algorithm differences are in the
methods. The largest paired result uses 5,000 genes,
20,000 regions and 256 TFs. Peak memory is sampled process-tree PSS.
These are historical workflow measurements; the v0.6.0 scorer has separate cost evidence below.
Biological confirmation and cost
Frozen networks trained on 20,000 BMMC cells achieved mean macro-F1 0.7123
versus 0.7113 for SCENIC+ networks on 12,315 external PBMCs from eight donors,
using the same scorer, TF roster and classifier. Separate activity directions
closed the mean deficit on these reused cohorts. Mean recall remained 78.3%
versus 79.8%; equivalence and superior regulatory-edge accuracy are not proven.
The independent 817,740-cell confirmation measured median paired overheads of
2.62% scoring time, 0.29% I/O-inclusive stage time and 4.52% peak PSS versus
union activity. One seed increased peak PSS by 15.94%. Additional activity
columns increase raw array storage by 34.33% and full HDF5 storage by about
27.05%. All 192 donor-fold metrics and numerical controls passed audit.
Complete results · Run the workflow
Install and reproduce
python -m pip install --upgrade rustscenic
rustscenic validated --helpBenchmark evidence and figure source include the measured
workflow versions, settings and individual runs. Full HPC output arrays and
the complete benchmark archive still require durable deposition for the paper.
Historical confirmation builds reused 0.5.1 metadata and are identified by
their recorded hashes. This release has a distinct version. Existing public
analysis defaults are preserved.