zhouwanding commited on
Commit
0aabcd7
·
verified ·
1 Parent(s): 909bbe2

README: hg38_62celltypes.msdref is the primary deconvolution reference

Browse files
Files changed (1) hide show
  1. README.md +18 -8
README.md CHANGED
@@ -52,7 +52,7 @@ upscale input + truth) live in the GitHub repo
52
 
53
  | file | task | framework | labels / notes |
54
  |------|------|-----------|----------------|
55
- | `hg38_wg.updecx` | whole-genome CpG upscaling | `UPDEC2` | all 29,401,795 hg38 CpGs; v5 primary upscaler (2.8 GB). Its decoder upgrade is deferred to v8. |
56
  | `hg38_10k1.updecx` | CpG upscaling (single block) | MLP decoder | block 10k1 (10,000 CpGs); small demo model |
57
  | `hg38_celltype.clfx` | cell-type annotation (human) | xgboost, routing tree | 33 human cell types; 11-node tree + 3 satellites; cross-cohort accuracy **0.903** — see below |
58
  | `mm10_celltype_brain.clfx` | cell-type annotation (mouse **brain**) | xgboost, routing tree | 41 mouse-brain cell types; 4-node tree + 134 satellites; **class-balanced accuracy 0.969 ± 0.001** at native coverage (10-fold CV). Brain only — a non-brain cell is forced onto the nearest brain type |
@@ -66,15 +66,25 @@ unmarked bundle, and rejects the retired flat `xgboost` mark. Upscale decoders (
66
  Deconvolution takes a standalone `.msdref`, not a bundle:
67
 
68
  ```
69
- methscope deconv mixture.cg hg38_33celltypes.msdref -o props.tsv
70
  ```
71
 
72
- `hg38_33celltypes.msdref` is 33 human cell types pooled from donor-resolved
73
- normal WGBS, packed as uint16 M/U over the 7.9 M CpGs that could ever separate
74
- some class subset (26.9% of hg38). It replaces `hg38_65celltypes.refx`, which
75
- was withdrawn — the current `methscope` can neither read nor produce that
76
- format. The solver rebuilds its pattern set per query from the CpGs that query
77
- actually measured, so no pattern budget is baked in.
 
 
 
 
 
 
 
 
 
 
78
 
79
  ## The routing-tree classifiers
80
 
 
52
 
53
  | file | task | framework | labels / notes |
54
  |------|------|-----------|----------------|
55
+ | `hg38_wg.updecx` | whole-genome CpG upscaling | `UPDEC2` | all 29,401,795 hg38 CpGs; v5 primary upscaler (2.8 GB). Its decoder upgrade is deferred to v9. |
56
  | `hg38_10k1.updecx` | CpG upscaling (single block) | MLP decoder | block 10k1 (10,000 CpGs); small demo model |
57
  | `hg38_celltype.clfx` | cell-type annotation (human) | xgboost, routing tree | 33 human cell types; 11-node tree + 3 satellites; cross-cohort accuracy **0.903** — see below |
58
  | `mm10_celltype_brain.clfx` | cell-type annotation (mouse **brain**) | xgboost, routing tree | 41 mouse-brain cell types; 4-node tree + 134 satellites; **class-balanced accuracy 0.969 ± 0.001** at native coverage (10-fold CV). Brain only — a non-brain cell is forced onto the nearest brain type |
 
66
  Deconvolution takes a standalone `.msdref`, not a bundle:
67
 
68
  ```
69
+ methscope deconv hg38_62celltypes.msdref mixture.cg -o props.tsv
70
  ```
71
 
72
+ `hg38_62celltypes.msdref` (v8) is the primary deconvolution reference: 60 cell
73
+ types from the Zhou 2025 single-cell atlas pseudobulks (early-embryonic SCT/VCT
74
+ excluded) plus Loyfer hepatocytes and BLUEPRINT granulocytes, packed as uint16
75
+ M/U over the 11.8 M CpGs that could ever separate some class subset (40.1% of
76
+ hg38). The single-cell backbone matters: against the 33-class reference, whose
77
+ immune compartment is BLUEPRINT bulk, single-cell-derived immune mixtures
78
+ misread badly (a 70/30 macrophage/monocyte mixture returned 0.14/0.56), while
79
+ the 62-class reference reads the same mixtures to within a few points of truth
80
+ and resolves T-cell sub-types (Tmem/Tnaive × CD4/CD8).
81
+
82
+ `hg38_33celltypes.msdref` is the earlier bulk-pooled reference: 33 human cell
83
+ types from donor-resolved normal WGBS (7.9 M kept CpGs, 26.9% of hg38). It
84
+ replaces `hg38_65celltypes.refx`, which was withdrawn — the current `methscope`
85
+ can neither read nor produce that format. For both references the solver
86
+ rebuilds its pattern set per query from the CpGs that query actually measured,
87
+ so no pattern budget is baked in.
88
 
89
  ## The routing-tree classifiers
90