Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261013T094815Z-all-celltypes-final-palette-142/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/celltypes_palette_master_20261013/README.md

@@ -0,0 +1,45 @@
+# All-cell-type pair budgets and approved method palette
+
+This version combines two user requests: expand former page 31 beyond NK fractions, then apply `PBISC-Diffusion_color_palette_20261013.md` throughout the working deck.
+
+## Page changes
+
+- **31:** donor fraction Pearson correlation for T, B, Monocyte, NK, DC and Other.
+- **32:** fraction MAE in percentage points for the same six cell types, methods, budgets and donors.
+- **84:** the former global composition/coverage/NK/selected-step summary is retained as supplementary context.
+- Total: **142 slides**, retaining each of the 140 parent pages once plus two new cell-type panels. Original 76-slide IDs remain represented once.
+
+The new panels reuse the historical healthy ImMicro pooled five-fold predictions: 107 unique donors, 11 method/budget settings, six types (7,062 donor/type/arm rows; 66 summaries). Pearson is calculated across the pooled donors, not averaged across folds. The MAE gives each donor equal weight.
+
+## Protocol and interpretation
+
+The budget is **n training + seven validation pairs** per fold. Calibration uses the same n+7 pairs. The all point uses 78–79 training pairs plus validation. Raw zero-shot at zero uses no pairs. These curves are separate from the genuine total-five-pair comparison on pages 33–34.
+
+All six types remain visible, including less favourable results. The old NK curve is reproduced to numerical precision. At n=5, raw→LoRA MAE is T 19.238→6.071 pp, B 2.684→1.640, Monocyte 17.088→3.437, NK 3.274→4.453, DC 0.846→1.090 and Other 2.358→0.454. Correlation and absolute error answer different questions; e.g. Monocyte correlation decreases 0.837→0.813 despite the lower error. No universal superiority or significance claim is added.
+
+Pearson axes are shared; MAE panels explicitly use different y-axis ranges by cell type. Undefined correlations would remain NA rather than zero; none of the 66 observed correlations is undefined. No new model training, generation or calibration fitting was performed.
+
+## Approved palette
+
+- PBISC LoRA: `#168FE5`.
+- PBISC paired/calibrated: `#59A8DF`; raw/zero-shot: `#C3E2F4`.
+- scSemiProfiler: `#83C6A8`.
+- BayesPrism: `#D9B23F`; CIBERSORTx: `#F0D77D`; other comparison methods use yellow-family shades.
+- Real/reference traces remain neutral. Biological category colors, cohort colors and quantitative heatmap scales retain their original meaning.
+
+Full FT uses a distinct dark PBISC blue (`#0D6EA8`) consistently, while calibrated conditions retain the approved medium blue.
+
+The palette conversion is figure-specific. Shared RGB values, rasterized method points, and literal colour captions require source-specific handling; a global image recolour is not used. Original scientific assets and previous decks are preserved.
+
+## Sources and reproduction
+
+- `../pair_budget_celltypes_20261013/result.json`: source hashes and numerical verification.
+- `../pair_budget_celltypes_20261013/summary.csv`: all 66 summaries.
+- `../pair_budget_celltypes_20261013/donor_fractions_long.csv`: source-bound donor fractions.
+- `../final_palette_20261013/result.json`: explicit semantic asset map and preservation evidence.
+- `page_map.csv`, `manifest.json`, `qa/VERIFICATION.json`: final deck navigation, source bindings and QA.
+- `../../tools/build_celltypes_palette_master_20261013.py`: build from saved plots and palette assets; no experiment launch.
+
+## Palette preservation evidence
+
+The semantic map covers 121 assets: all 119 parent figure assets and the two new cell-type plots. Vector changes preserve scientific paint geometry; selective source-native replays verify the same values and non-caption text layout. Literal colour captions were updated to neutral wording. The one PNG-only COVID method-arrow figure uses source-defined method masks, preserving all 6,760,419 non-method pixels and the entire alpha channel exactly; no projection was refitted.

paper-releases/pbisc-diffusion/updates/20261013T094815Z-all-celltypes-final-palette-142/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/celltypes_palette_master_20261013/page_map.csv

@@ -0,0 +1,143 @@
+current_page,parent_140_page,original_131_page,original_76_page,group,panel,title
+1,1,1,1,Introduction,Introduction,PBISC-Diffusion: manuscript evidence
+2,2,2,2,Introduction,Introduction,Figures 1-6: the expanded evidence map
+3,3,3,,Introduction,Introduction,How this expanded deck is organized
+4,4,4,3,Introduction,Current scope,Manuscript map: the six figures
+5,5,6,4,Figure 1,Figure 1,Figure 1. Overview and corpus
+6,6,7,5,Figure 1,Figure 1 plan,"Figure 1: overview, corpus and denoising"
+7,7,9,7,Figure 1,Fig. 1a,Experimental overview
+8,8,11,9,Figure 1,Fig. 1b,Pretraining corpus
+9,9,16,13,Figure 1,Fig. 1c,Corpus scale and generation workflow
+10,10,,,Figure 1,Fig. 1d,From Gaussian noise to generated cell states
+11,11,19,15,Figure 2,Figure 2,Figure 2. SLE model ablations
+12,12,20,16,Figure 2,Figure 2 plan,Figure 2: architecture and gene context
+13,13,,,Figure 2,Fig. 2a,Where context width and trunk width change
+14,14,,,Figure 2,Fig. 2b,Predicted and oracle hints: one trained denoiser
+15,15,12,10,Figure 2,Fig. 2c,Architecture: full and late validation progression
+16,16,13,11,Figure 2,Fig. 2d,Width & context ablation: benchmark
+17,17,,,Figure 2,Fig. 2 details,Architecture arms: dimensions and endpoints
+18,18,14,12,Figure 2,Fig. 2e,Gene context: 2k / 3k / 6k / 15k
+19,19,,,Figure 2,Fig. 2f,Gene panels: progression relative to each endpoint
+20,20,,,Figure 2,Fig. 2 details,Gene-panel arms: exact sizes and training provenance
+21,21,,,Figure 2,Fig. 2g,All ablation arms: parameters and training time
+22,22,,,Figure 2,Fig. 2h,Real-cell repeat: a sampling reference
+23,23,15,,Figure 2,Fig. 2i,Native-pseudobulk model comparisons
+24,24,52,28,Figure 3,Figure 3,Figure 3. Real-bulk adaptation
+25,25,53,29,Figure 3,Figure 3 plan,Figure 3: the paired-adaptation story
+26,26,21,17,Figure 3,Fig. 3a,Bulk and pooled cells: different expression values
+27,27,,,Figure 3,Fig. 3a detail,Paired calibration reduces same-donor expression error
+28,28,23,19,Figure 3,Fig. 3b detail,Consequence: cells off-manifold
+29,29,5,,Figure 3,Fig. 3b,Healthy paired data: full FT and LoRA
+30,30,51,,Figure 3,Fig. 3c,Full FT and LoRA ranks: quality and parameters
+31,,,,Figure 3,Fig. 3d,Healthy pair budget: correlation in all cell types
+32,,,,Figure 3,Fig. 3d detail,Healthy pair budget: error in all cell types
+33,32,54,30,Figure 3,Fig. 3e,"Five pairs, no extra validation: LoRA vs scSemi"
+34,33,55,,Figure 3,Fig. 3f,Composition: error and donor-level wins
+35,34,60,35,Figure 4,Figure 4,Figure 4. Disease signal preservation
+36,35,61,36,Figure 4,Figure 4 plan,Figure 4: disease signal preservation
+37,36,63,38,Figure 4,Fig. 4a,Disease signals with five training donors
+38,37,67,39,Figure 4,Fig. 4b,5-fold CV: disease signal
+39,38,68,40,Figure 4,Fig. 4c,Disease genes at every list length (CAT)
+40,39,71,41,Figure 4,Fig. 4d,"Cell-type deconvolution, same pairs: disease signal"
+41,40,86,47,Figure 5,Figure 5,Figure 5. Rare-population identity
+42,41,87,48,Figure 5,Figure 5 plan,Figure 5: paired rare-population identity
+43,42,102,53,Figure 6,Figure 6,Figure 6. Published-finding reproduction
+44,43,103,54,Figure 6,Figure 6 plan,Figure 6: published-finding reproduction
+45,44,107,,Figure 6,Figure 6,Figure 6: what the reproduction must establish
+46,45,108,56,Supplementary,Supplementary,Supplementary evidence
+47,46,109,57,Supplementary,Supplementary,Supplementary evidence map
+48,47,8,6,Supp S01,Supp. S01,Backbone: SDD → PBISC-Diffusion
+49,48,10,8,Supp S01,Supp. S01,Bulk conditioning
+50,49,17,14,Supp S01,Supp. S01,Denoiser: joint value + support diffusion
+51,50,38,20,Supp S01,Supp. S01,Fine-tuning with paired bulk + sc
+52,51,39,21,Supp S01,Supp. S01,Evaluation design
+53,52,41,22,Supp S01,Supp. S01,How we score generated cells
+54,53,18,,Supp S02,Supp. S02,Does the model follow its bulk conditioning?
+55,54,72,42,Supp S02,Supp. S02,SLE benchmark: PBISC vs scSemiProfiler
+56,55,73,43,Supp S02,Supp. S02,SLE benchmark: per cell type
+57,56,74,,Supp S02,Supp. S02,Disjoint-cell controls test the bulk-broadcast baseline
+58,57,75,,Supp S02,Supp. S02,SLE: what is added beyond the bulk broadcast?
+59,58,22,18,Supp S03,Supp. S03,Same shift in three paired cohorts
+60,59,24,,Supp S03,Supp. S03,Input corrections on one real-reference UMAP
+61,60,25,,Supp S03,Supp. S03,Where generated-cell density differs from real cells
+62,61,26,,Supp S03,Supp. S03,Input calibration changes cell-type composition
+63,62,27,,Supp S03,Supp. S03,"Same noise, different input: matched-cell shifts"
+64,63,28,,Supp S03,Supp. S03,Matched noise separates cell state from cell identity
+65,64,29,,Supp S03,Supp. S03,Input changes can switch the generated cell label
+66,65,30,,Supp S03,Supp. S03,Nearest real cells: distance and label agreement
+67,66,31,,Supp S03,Supp. S03,Composition agreement does not ensure state agreement
+68,67,32,,Supp S03,Supp. S03,Gene-length effects in the bulk input
+69,68,33,,Supp S03,Supp. S03,Gene-length effects within label-stable cells
+70,69,34,,Supp S03,Supp. S03,Raw and TPM shifts persist within cell labels
+71,70,35,,Supp S03,Supp. S03,Reference calibration leaves specific gene-state shifts
+72,71,36,,Supp S03,Supp. S03,Donor trait axes and donor identification differ
+73,72,37,,Supp S03,Supp. S03,Reference choice changes different quality endpoints
+74,73,40,,Supp S04,Supp. S04,Why the historical and strict protocols differ
+75,74,42,,Supp S04,Supp. S04,Technical repeats and follow-up are different controls
+76,75,43,,Supp S04,Supp. S04,Annotation agreement and difficult categories
+77,76,46,,Supp S04,Supp. S04,Depth matching changes precision and coverage
+78,77,47,,Supp S04,Supp. S04,Detection and library size are separate diagnostics
+79,78,44,23,Supp S05,Supp. S05,Healthy cohort: LoRA cells cover the real ones
+80,79,45,24,Supp S05,Supp. S05,Healthy cohort: realism beyond precision/coverage
+81,80,48,25,Supp S05,Supp. S05,Healthy cohort: full vs LoRA vs condition encoder
+82,81,57,32,Supp S05,Supp. S05,Representative-dependent composition
+83,82,116,64,Supp S05,Supp. S05,Full vs LoRA vs cond.-encoder-only
+84,31,49,26,Supp S06,Supp. S06,"Pair-budget summary: composition, coverage and steps"
+85,83,50,27,Supp S06,Supp. S06,Healthy cohort: LoRA rank and K
+86,84,56,31,Supp S06,Supp. S06,Healthy paired-data learning curves
+87,85,58,33,Supp S06,Supp. S06,"Time: days for scSemiProfiler, minutes for LoRA"
+88,86,59,34,Supp S06,Supp. S06,5 donors: time to apply
+89,87,117,65,Supp S06,Supp. S06,"Pairs needed: composition, coverage, training steps"
+90,88,118,66,Supp S06,Supp. S06,"LoRA rank: r = 4, 16, 64"
+91,89,119,67,Supp S06,Supp. S06,K diversification: no gain
+92,90,62,37,Supp S07,Supp. S07,Five paired cohorts
+93,91,64,,Supp S07,Supp. S07,Sepsis: exact-five reconstruction across five draws
+94,92,65,,Supp S07,Supp. S07,Sepsis: realism beyond precision and coverage
+95,93,66,,Supp S07,Supp. S07,Imperial: first completed exact-five comparison
+96,94,110,58,Supp S07,Supp. S07,5 donors: LoRA vs scSemiProfiler — cells
+97,95,111,59,Supp S07,Supp. S07,5 donors: realism beyond precision/coverage
+98,96,112,60,Supp S07,Supp. S07,"5 donors: precision/coverage at k = 3, 5, 10"
+99,97,113,61,Supp S07,Supp. S07,5-fold: realism beyond precision/coverage
+100,98,114,62,Supp S07,Supp. S07,Same training pairs: LoRA vs scSemiProfiler
+101,99,115,63,Supp S07,Supp. S07,scSemiProfiler and the representative-copy baseline
+102,100,69,,Supp S08,Supp. S08,Sepsis: patient-aware disease-gene agreement
+103,101,70,,Supp S08,Supp. S08,Imperial: patient-aware disease-gene agreement
+104,102,82,44,Supp S08,Supp. S08,Sepsis: fine-tuning variants
+105,103,83,45,Supp S08,Supp. S08,Pairs needed: donor and disease signal
+106,104,84,46,Supp S08,Supp. S08,"Cell-type deconvolution, same pairs: composition"
+107,105,85,,Supp S08,Supp. S08,Deconvolution: expression coverage limits the comparison
+108,106,120,,Supp S08,Supp. S08,"K=4,000 recipe: paired reconstruction quality"
+109,107,121,,Supp S08,Supp. S08,"K=4,000 recipe: disease signal does not improve uniformly"
+110,108,122,,Supp S08,Supp. S08,"K=4,000 recipe: calibration gains are selective"
+111,109,76,,Supp S09,Supp. S09,Influenza: response order versus response genes
+112,110,77,,Supp S09,Supp. S09,COVID: calibration on a shared real-cell map
+113,111,78,,Supp S09,Supp. S09,COVID monocytes: direction after depth matching
+114,112,79,,Supp S09,Supp. S09,"COVID monocytes: direction, magnitude and null"
+115,113,80,,Supp S09,Supp. S09,"COVID CD14 monocytes: direction, magnitude and null"
+116,114,81,,Supp S09,Supp. S09,COVID: biological endpoints give a mixed verdict
+117,115,89,,Supp S10,Supp. S10,Rare-state frequencies in observed cells
+118,116,91,,Supp S10,Supp. S10,Rare-state misses: rules 1–8
+119,117,92,,Supp S10,Supp. S10,Rare-state misses: rules 9–15
+120,118,93,,Supp S10,Supp. S10,Rare-state fidelity: zero-aware reanalysis
+121,119,95,,Supp S10,Supp. S10,Sepsis rare states across paired-data budgets
+122,120,96,,Supp S10,Supp. S10,Sepsis rare-state direction across support budgets
+123,121,97,,Supp S11,Supp. S11,"Rare-state screen in DRESS, Kawasaki and SLE"
+124,122,98,52,Supp S11,Supp. S11,Consequence: disease direction flips
+125,123,99,,Supp S11,Supp. S11,C1Q: single-gene detection versus co-detection
+126,124,100,,Supp S11,Supp. S11,"Behçet: K=4,000 changes direction more than identity"
+127,125,101,,Supp S11,Supp. S11,Behcet: one-gene and multi-gene C1Q rules
+128,126,123,68,Supp S12,Supp. S12,Pairs from other cohorts do not transfer
+129,127,124,69,Supp S12,Supp. S12,Longer training on other cohorts: gap stays
+130,128,125,70,Supp S12,Supp. S12,"No pretraining: same cohort, from scratch"
+131,129,126,71,Supp S12,Supp. S12,"No pretraining: other cohorts, from scratch"
+132,130,127,72,Supp S12,Supp. S12,Rare states: from scratch vs pretrained
+133,131,129,74,Supp S13,Supp. S13,Appendix: scSemiProfiler without output threshold
+134,132,106,,Supp S14,Supp. S14,Bulk-only exploration: calibration populations differ
+135,133,128,73,Historical archive,Historical archive,Historical scoring and plans
+136,134,88,49,Historical archive,Historical only,Rare-state fractions: original analysis
+137,135,90,50,Historical archive,Historical only,Rare-state ranking: original analysis
+138,136,94,51,Historical archive,Historical only,Rare states: original five-fold analysis
+139,137,104,55,Historical archive,Historical archive,Bulk-only cohort plan: original snapshot
+140,138,105,,Historical archive,Historical archive,Historical bulk-only search: no confirmed new finding
+141,139,130,75,Discussion,Discussion,Discussion and outlook
+142,140,131,76,Discussion,Discussion,Next: other tissues from a few dozen pairs