mm10_41celltypes.msdref: the first mouse deconvolution reference
Wanding Zhou
43b7a27
A pure addition -- every v10 file ships unchanged, nothing withdrawn.
41 Liu 2021 mouse brain major types over the 6,765,325 CpGs that could ever
separate some class subset (30.9% of mm10), packed as uint16 M/U from
pseudobulks of all 54,447 cells with no fold held back. Until now every
.msdref was human and mouse shipped only classifiers plus the wg decoder.
It is the first version-3 artifact: it carries a validation confusion
trailer, which `deconv --group-threshold` reads to report near-collinear
classes under one joined label (CA3/DG-po) rather than splitting mass between
columns the data cannot separate. The trailer was measured on a
fold-held-out twin, since an all-cells build has no held-out cells to measure
with. A binary older than that format refuses the file rather than misreading
it, so the registry pin moves only with a YAME release that can read it.
Validated before staging: all 41 pools self-deconvolve to 100.0%; on P21
spatial DMT mouse brain (2,235 in-tissue pixels)…
bank classifiers: hg38_celltype_{full,lite}, mm10_brain_{full,lite}
Wanding Zhou
61e38dd
Pooled-bank cell-type classifiers (one booster, no routing) from the
20260904 bank-full recipe. _full carries a calibrated resolver for every
class pair; _lite only for pairs no LCA binstring separates. Human 63
classes, mouse 41 brain classes.
hg38_celltype is now a violation-framework model with no fitted parameters: it reads each class's expected methylation state off the MRMP binstrings and calls the class the query contradicts least, weighting patterns by sqrt(n_cpg). Confidence is the margin to the runner-up, which collapses when no class fits.
External test 0.842 / macro-F1 0.853 at >=100k covered CpGs. Colorectal cancer single cells call 100% gastrointestinal from liver and lymph metastases alike, and the sink that sent globally demethylated tumours to Acinar.(Pancreas) drops from 30.0% of TCGA calls to 2.3%.
Classifier bundles are renamed .ubjx -> .clfx: .ubjx named a payload format rather than a role and had already stopped being accurate (hg38_sex carried a plain-text linear spec, no UBJSON). Bundles are detected by MSBNDL1 magic, never by extension, so .ubjx stays accepted. The previous filenames remain available at --revision v5.
mm10_celltype.ubjx is unchanged, pending a Liu 2021 rebuild.
