SniffCell v0.9.8a1

5hmC-aware SniffCell alpha

This prerelease adds separate 5mC and 5hmC marker discovery and read scoring,
while retaining the original modifiedC atlas path. It is published from
5hmC_compatitible_atlas for early testing.

Read the benchmark blog and detailed implementation update
for bar plots, control definitions, and the compatibility checks.

Install

pip install 'sniffcell==0.9.8a1'
sniffcell --version

The package version is 0.9.8a1; the Git tag is v0.9.8a1. This is an alpha,
not a stable release. Normal stable installs remain on v0.9.7. The container
is published under ghcr.io/fu-yilei/sniffcell:0.9.8a1 when its build finishes;
this prerelease does not move Docker latest.

Atlas and input behavior

Download the four-donor GRCh38 brain atlas:
489,972 scoring rows, including the original 152,961-row paper backbone,
194,662 added 5mC markers, and 142,349 added 5hmC markers. The atlas is a
separate download, not bundled in the Python wheel.

With --bam-modification auto, each marker uses its declared channel.
Separate m/h BAM calls supply both assays. With an m-only BAM, h-specific
markers have no evidence; missing h is not treated as measured zero 5hmC.
An m tag alone does not distinguish a pure-5mC model from an older
modifiedC-like model. Existing unlabeled catalogs retain modifiedC behavior.
MDB-based discovery additionally needs the MDB dependency and source inputs;
using the precomputed catalog for deconvolution does not require rebuilding MDB.

Held-out evidence

On the same 46,373,872 bcontrol1 FANS control reads:

Atlas	Precision	Recall	Correct assignments
Original paper backbone	87.01%	18.33%	8,502,346
Integrated dual	88.64%	31.03%	14,391,267

The benchmark atlas excludes bcontrol1 from ONT discovery and reference means.
The distributed atlas includes all four donors and is a distinct artifact.
These results are a single-donor proof, not an independent validation of the
four-donor build or an estimate of the contribution of 5hmC alone. Other folds,
bulk checks, and single-channel ablations remain deferred.

Validation and compatibility

The main-branch tissue atlases, CLI update, and packaging changes are merged.
The merged source passed 273 tests plus 3 subtests. FGF14 historical-code
compatibility passed, and the post-merge backbone and dual runs reproduced
all eight benchmark output tables exactly. Genome-wide old-code parity has
not been established. The original paper backbone fields in the distributed
catalog are preserved exactly.

Full Changelog: v0.9.7...v0.9.8a1