Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261014T113716Z-figure1-overview-training-corpus/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/figure1_revision_20261014/README.md

@@ -0,0 +1,29 @@
+# Figure 1 revision — 2026-10-14
+
+**Current master: 155 slides.** Figure 1a is replaced at page 7; the exact training-corpus Figure 1b is inserted at page 8. The full 41.4M source-corpus view remains on page 9 as support. All 154 parent slide IDs and all 76 original IDs remain represented. Previous deck files are preserved.
+
+[Editable PPTX](../../PBISC-Diffusion_expanded_master_20261014_figure1.pptx) · [PDF](../../PBISC-Diffusion_expanded_master_20261014_figure1.pdf) · [Page map](page_map.csv)
+
+## Figure 1a: experimental overview
+
+The supplied panel d inspired organic flat cell shapes, soft group envelopes, thin dark arrows and white space. The diagram is newly drawn as editable vectors; none of the reference paper's scientific objects/results were copied. It shows same-sample pseudobulk/member-cell pretraining, paired measured-bulk adaptation and new-bulk generation with Gaussian noise. The PBISC blue method palette is retained; gold marks trainable adapters. Network glyphs are schematic, not literal architecture.
+
+[PNG](../../figures/figure1_revision_20261014/overview/experimental_overview.png) · [Vector PDF](../../figures/figure1_revision_20261014/overview/experimental_overview.pdf) · [Editable SVG](../../figures/figure1_revision_20261014/overview/experimental_overview.svg)
+
+K recipe size and generated-cell count are distinct. Held-out scRNA is evaluation truth, not an inference input. Generic rare-population endpoints do not claim completed source-cluster recovery. Figure 6 remains TODO.
+
+## Figure 1b: actual training corpus
+
+- 33,141,380 unique training source cells; 22 datasets; 5,232 samples; 2,572 corrected biological donor identities.
+- Healthy: 29,304,952 cells (88.4241%); disease / at-risk: 3,836,428 (11.5759%).
+- Condition proportions use cell-level groups filtered to the registered train split.
+- Training cell types were independently aggregated from matched metadata across 432 shards; source row alignment and all-source cell-type totals replay exactly. All 99,769 train overlay cells are retained as Unlabelled.
+- ACPA+ risk is not confirmed RA. Healthy donor status includes stimulation, vaccination and CMV-seropositive contexts. Source-cell counts are not recipe-weighted exposure. The internal 27-donor cross-split caveat is unchanged.
+
+[PNG](../../figures/figure1_revision_20261014/train_corpus/training_corpus_composition.png) · [Vector PDF](../../figures/figure1_revision_20261014/train_corpus/training_corpus_composition.pdf) · [Editable SVG](../../figures/figure1_revision_20261014/train_corpus/training_corpus_composition.svg)
+
+[Training numeric verification](train_corpus/VERIFICATION.json) · [Source roster](train_corpus/source_hashes.json) · [Deck verification](qa/VERIFICATION.json)
+
+CPU-only metadata aggregation, plotting and export. No new training, generation, annotation model or GPU work. No raw expression arrays read. Old source figures and old decks remain unchanged. All parent figures except the explicitly replaced overview have identical embedded bytes.
+
+Build: `tools/build_figure1_master_20261014.py`; export/verify: `data/figure1_revision_20261014/render.py`. Figure sources: `tools/figure1_overview_20261014.py` and `tools/figure1_train_corpus_20261014.py`.

paper-releases/pbisc-diffusion/updates/20261014T113716Z-figure1-overview-training-corpus/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/figure1_revision_20261014/overview/result.json

@@ -0,0 +1,53 @@
+{
+  "schema": "pbisc.figure1-overview.20261014.v1",
+  "slides": [
+    {
+      "kind": "content",
+      "title": "Experimental overview",
+      "section": "Figure 1",
+      "deck_group": "Figure 1",
+      "section_label": "Fig. 1a",
+      "layout": "full",
+      "keywords": [],
+      "figures": [
+        "figures/figure1_revision_20261014/overview/experimental_overview.png"
+      ],
+      "footnote": "Schematic. K defines the pseudobulk recipe; N_gen defines generated-cell count. Held-out scRNA is used only for evaluation.",
+      "footnote_full": "Original schematic redrawn with editable organic vector circles, muted fills, thin dark arrows and open white space, inspired by the supplied reference panel d. No scientific objects or numeric results were copied from that paper. 33,141,380 is the verified unique training-cell universe. A same-sample K-cell recipe supplies pseudobulk conditioning to each member target cell. Pretrained weights initialize LoRA adaptation on measured bulk\u2013scRNA pairs. New measured bulk and independent Gaussian-noise draws generate cells, without supplied cell-type quotas. Cell colours and model-node glyphs are schematic, not counts, measured populations or a literal architecture. The main model uses no cell-type/disease input labels; Figure2 oracle/prediction-hint ablations are separate. Generated annotations are post hoc. Conditioning recipe K and generated-cell number N_gen are distinct. Held-out scRNA is evaluation truth, never shown as an inference input. Rare-population endpoints are evaluation objectives, not a claim of completed source-cluster recovery; Figure6 bulk-only published-finding reproduction remains TODO.",
+      "sources": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/tools/figure1_overview_20261014.py; /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/runs/f01_6k_1x_seed20260627/ALLCELL_DATA_CONTRACT.json; /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/data/figure_assignment_20261013/AUTHOR_SCOPE.md; /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_color_palette_20261013.md; /home/sj_server_1/PBISC/paper-release-20260928/manuscript/scripts/build_figure1_architecture_nature.py"
+    }
+  ],
+  "reference_style": {
+    "path": "/home/sj_server_1/.codex/attachments/15a605fb-2cca-4162-9f87-354b7c57cac6/codex-clipboard-98e511c6-7c97-416f-9054-2e6db758e44e.png",
+    "interpretation": "panel d: organic flat circles, minimal black arrows, white background; existing method palette retained"
+  },
+  "assets": [
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/figure1_revision_20261014/overview/experimental_overview.png",
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/figure1_revision_20261014/overview/experimental_overview.svg",
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/figure1_revision_20261014/overview/experimental_overview.pdf"
+  ],
+  "sources": [
+    {
+      "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/tools/figure1_overview_20261014.py",
+      "sha256": "84969620e40879957b701e8663a5af1b3d125f2085386b0343be2d7ea285b999"
+    },
+    {
+      "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/runs/f01_6k_1x_seed20260627/ALLCELL_DATA_CONTRACT.json",
+      "sha256": "461b46107f75984000747877a11823ec15042b5c2104b4b979729dcee10ee30f"
+    },
+    {
+      "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/data/figure_assignment_20261013/AUTHOR_SCOPE.md",
+      "sha256": "791d9336d4ccf94a888a509808db041e5fda590ee2e15361db8ff36ad2244358"
+    },
+    {
+      "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_color_palette_20261013.md",
+      "sha256": "feff43bfd27e54b39f94f5170b1b75a76cd378985d869b612075f592ae429958"
+    },
+    {
+      "path": "/home/sj_server_1/PBISC/paper-release-20260928/manuscript/scripts/build_figure1_architecture_nature.py",
+      "sha256": "0b462c7f63613c92d208e022622269bea9b2972fe0ddf6fe737f8ae2087a8334"
+    }
+  ],
+  "new_training": false,
+  "new_inference": false
+}

paper-releases/pbisc-diffusion/updates/20261014T113716Z-figure1-overview-training-corpus/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/figure1_revision_20261014/page_map.csv

@@ -0,0 +1,156 @@
+page,parent_154_page,original_76_page,group,panel,title
+1,1,1,Introduction,Introduction,PBISC-Diffusion: manuscript evidence
+2,2,2,Introduction,Introduction,Figures 1-6: the expanded evidence map
+3,3,,Introduction,Introduction,How this expanded deck is organized
+4,4,3,Introduction,Current scope,Manuscript map: the six figures
+5,5,4,Figure 1,Figure 1,Figure 1. Overview and corpus
+6,6,5,Figure 1,Figure 1 plan,"Figure 1: overview, corpus and denoising"
+7,7,7,Figure 1,Fig. 1a,Experimental overview
+8,,,Figure 1,Fig. 1b,Pretraining data: 33.14 million training cells
+9,8,,Figure 1,Figure 1 support,"Full source corpus: training, validation and test"
+10,9,,Figure 1,Figure 1 support,From 41.40M source cells to 33.14M training cells
+11,10,,Figure 1,Figure 1 support,Recipe K and training exposure are different counts
+12,11,13,Figure 1,Fig. 1c,Corpus scale and generation workflow
+13,12,,Figure 1,Fig. 1d,From Gaussian noise to generated cell states
+14,13,15,Figure 2,Figure 2,Figure 2. SLE model ablations
+15,14,16,Figure 2,Figure 2 plan,Figure 2: architecture and gene context
+16,15,,Figure 2,Fig. 2a,Where context width and trunk width change
+17,16,,Figure 2,Fig. 2b,Predicted and oracle hints: one trained denoiser
+18,17,10,Figure 2,Fig. 2c,Architecture: full and late validation progression
+19,18,11,Figure 2,Fig. 2d,Width & context ablation: benchmark
+20,19,,Figure 2,Fig. 2 details,Architecture arms: dimensions and endpoints
+21,20,12,Figure 2,Fig. 2e,Gene context: 2k / 3k / 6k / 15k
+22,21,,Figure 2,Fig. 2f,Gene panels: progression relative to each endpoint
+23,22,,Figure 2,Fig. 2 details,Gene-panel arms: exact sizes and training provenance
+24,23,,Figure 2,Fig. 2g,All ablation arms: parameters and training time
+25,24,,Figure 2,Fig. 2h,Real-cell repeat: a sampling reference
+26,25,,Figure 2,Fig. 2i,Native-pseudobulk model comparisons
+27,26,28,Figure 3,Figure 3,Figure 3. Real-bulk adaptation
+28,27,29,Figure 3,Figure 3 plan,Figure 3: the paired-adaptation story
+29,28,17,Figure 3,Fig. 3a,Bulk and pooled cells: different expression values
+30,29,,Figure 3,Fig. 3a detail,Paired calibration reduces same-donor expression error
+31,30,19,Figure 3,Fig. 3b detail,Consequence: cells off-manifold
+32,31,,Figure 3,Fig. 3 · Adaptation,Adaptation recovers the observed cell mixture
+33,32,,Figure 3,Fig. 3 · Adaptation,Adaptation: mixture error and expression agreement
+34,33,,Figure 3,Fig. 3b,Healthy paired data: full FT and LoRA
+35,34,,Figure 3,Fig. 3c,Full FT and LoRA ranks: quality and parameters
+36,35,,Figure 3,Fig. 3 · Adaptation,What does “all pairs” actually mean?
+37,36,,Figure 3,Fig. 3d,Healthy pair budget: correlation in all cell types
+38,37,,Figure 3,Fig. 3d detail,Healthy pair budget: error in all cell types
+39,38,30,Figure 3,Fig. 3e,"Five pairs, no extra validation: LoRA vs scSemi"
+40,39,,Figure 3,Fig. 3f,Composition: error and donor-level wins
+41,40,35,Figure 4,Figure 4,Figure 4. Disease signal preservation
+42,41,36,Figure 4,Figure 4 plan,Figure 4: disease signal preservation
+43,42,38,Figure 4,Fig. 4a,Disease signals with five training donors
+44,43,39,Figure 4,Fig. 4b,5-fold CV: disease signal
+45,44,40,Figure 4,Fig. 4c,Disease genes at every list length (CAT)
+46,45,41,Figure 4,Fig. 4d,"Cell-type deconvolution, same pairs: disease signal"
+47,46,,Figure 4,Fig. 4 · Disease integrity,Disease gene rankings: complete CV comparison
+48,47,,Figure 4,Fig. 4 · Disease integrity,Disease integrity depends on cohort and endpoint
+49,48,,Figure 4,Fig. 4 · Disease integrity,Five-pair Sepsis: metrics across all support draws
+50,49,47,Figure 5,Figure 5,Figure 5. Rare-population identity
+51,50,48,Figure 5,Figure 5 plan,Figure 5: paired rare-population identity
+52,51,,Figure 5,Fig. 5 · Source-defined populations,Figure 5 targets begin with the source paper
+53,52,,Figure 5,Fig. 5 · Source-defined populations,Ye sepsis: published markers in paired scRNA
+54,53,,Figure 5,Fig. 5 · Source-defined populations,Transcriptional programs and rare-population identity
+55,54,53,Figure 6,Figure 6,Figure 6. Published-finding reproduction
+56,55,54,Figure 6,Figure 6 plan,Figure 6: published-finding reproduction
+57,56,,Figure 6,Figure 6,Figure 6: what the reproduction must establish
+58,57,56,Supplementary,Supplementary,Supplementary evidence
+59,58,57,Supplementary,Supplementary,Supplementary evidence map
+60,59,9,Supp S01,Supp. S01,Corpus reference: assay and source composition
+61,60,6,Supp S01,Supp. S01,Backbone: SDD → PBISC-Diffusion
+62,61,8,Supp S01,Supp. S01,Bulk conditioning
+63,62,14,Supp S01,Supp. S01,Denoiser: joint value + support diffusion
+64,63,20,Supp S01,Supp. S01,Fine-tuning with paired bulk + sc
+65,64,21,Supp S01,Supp. S01,Evaluation design
+66,65,22,Supp S01,Supp. S01,How we score generated cells
+67,66,,Supp S02,Supp. S02,Does the model follow its bulk conditioning?
+68,67,42,Supp S02,Supp. S02,SLE benchmark: PBISC vs scSemiProfiler
+69,68,43,Supp S02,Supp. S02,SLE benchmark: per cell type
+70,69,,Supp S02,Supp. S02,Disjoint-cell controls test the bulk-broadcast baseline
+71,70,,Supp S02,Supp. S02,SLE: what is added beyond the bulk broadcast?
+72,71,18,Supp S03,Supp. S03,Same shift in three paired cohorts
+73,72,,Supp S03,Supp. S03,Input corrections on one real-reference UMAP
+74,73,,Supp S03,Supp. S03,Where generated-cell density differs from real cells
+75,74,,Supp S03,Supp. S03,Input calibration changes cell-type composition
+76,75,,Supp S03,Supp. S03,"Same noise, different input: matched-cell shifts"
+77,76,,Supp S03,Supp. S03,Matched noise separates cell state from cell identity
+78,77,,Supp S03,Supp. S03,Input changes can switch the generated cell label
+79,78,,Supp S03,Supp. S03,Nearest real cells: distance and label agreement
+80,79,,Supp S03,Supp. S03,Composition agreement does not ensure state agreement
+81,80,,Supp S03,Supp. S03,Gene-length effects in the bulk input
+82,81,,Supp S03,Supp. S03,Gene-length effects within label-stable cells
+83,82,,Supp S03,Supp. S03,Raw and TPM shifts persist within cell labels
+84,83,,Supp S03,Supp. S03,Reference calibration leaves specific gene-state shifts
+85,84,,Supp S03,Supp. S03,Donor trait axes and donor identification differ
+86,85,,Supp S03,Supp. S03,Reference choice changes different quality endpoints
+87,86,,Supp S04,Supp. S04,Why the historical and strict protocols differ
+88,87,,Supp S04,Supp. S04,Technical repeats and follow-up are different controls
+89,88,,Supp S04,Supp. S04,Annotation agreement and difficult categories
+90,89,,Supp S04,Supp. S04,Depth matching changes precision and coverage
+91,90,,Supp S04,Supp. S04,Detection and library size are separate diagnostics
+92,91,23,Supp S05,Supp. S05,Healthy cohort: LoRA cells cover the real ones
+93,92,24,Supp S05,Supp. S05,Healthy cohort: realism beyond precision/coverage
+94,93,25,Supp S05,Supp. S05,Healthy cohort: full vs LoRA vs condition encoder
+95,94,32,Supp S05,Supp. S05,Representative-dependent composition
+96,95,64,Supp S05,Supp. S05,Full vs LoRA vs cond.-encoder-only
+97,96,26,Supp S06,Supp. S06,"Pair-budget summary: composition, coverage and steps"
+98,97,27,Supp S06,Supp. S06,Healthy cohort: LoRA rank and K
+99,98,31,Supp S06,Supp. S06,Healthy paired-data learning curves
+100,99,33,Supp S06,Supp. S06,"Time: days for scSemiProfiler, minutes for LoRA"
+101,100,34,Supp S06,Supp. S06,5 donors: time to apply
+102,101,65,Supp S06,Supp. S06,"Pairs needed: composition, coverage, training steps"
+103,102,66,Supp S06,Supp. S06,"LoRA rank: r = 4, 16, 64"
+104,103,67,Supp S06,Supp. S06,K diversification: no gain
+105,104,37,Supp S07,Supp. S07,Five paired cohorts
+106,105,,Supp S07,Supp. S07,Sepsis: exact-five reconstruction across five draws
+107,106,,Supp S07,Supp. S07,Sepsis: realism beyond precision and coverage
+108,107,,Supp S07,Supp. S07,Imperial: first completed exact-five comparison
+109,108,58,Supp S07,Supp. S07,5 donors: LoRA vs scSemiProfiler — cells
+110,109,59,Supp S07,Supp. S07,5 donors: realism beyond precision/coverage
+111,110,60,Supp S07,Supp. S07,"5 donors: precision/coverage at k = 3, 5, 10"
+112,111,61,Supp S07,Supp. S07,5-fold: realism beyond precision/coverage
+113,112,62,Supp S07,Supp. S07,Same training pairs: LoRA vs scSemiProfiler
+114,113,63,Supp S07,Supp. S07,scSemiProfiler and the representative-copy baseline
+115,114,,Supp S08,Supp. S08,Sepsis: patient-aware disease-gene agreement
+116,115,,Supp S08,Supp. S08,Imperial: patient-aware disease-gene agreement
+117,116,44,Supp S08,Supp. S08,Sepsis: fine-tuning variants
+118,117,45,Supp S08,Supp. S08,Pairs needed: donor and disease signal
+119,118,46,Supp S08,Supp. S08,"Cell-type deconvolution, same pairs: composition"
+120,119,,Supp S08,Supp. S08,Deconvolution: expression coverage limits the comparison
+121,120,,Supp S08,Supp. S08,"K=4,000 recipe: paired reconstruction quality"
+122,121,,Supp S08,Supp. S08,"K=4,000 recipe: disease signal does not improve uniformly"
+123,122,,Supp S08,Supp. S08,"K=4,000 recipe: calibration gains are selective"
+124,123,,Supp S09,Supp. S09,Influenza: response order versus response genes
+125,124,,Supp S09,Supp. S09,COVID: calibration on a shared real-cell map
+126,125,,Supp S09,Supp. S09,COVID monocytes: direction after depth matching
+127,126,,Supp S09,Supp. S09,"COVID monocytes: direction, magnitude and null"
+128,127,,Supp S09,Supp. S09,"COVID CD14 monocytes: direction, magnitude and null"
+129,128,,Supp S09,Supp. S09,COVID: biological endpoints give a mixed verdict
+130,129,,Supp S10,Supp. S10,Rare-state frequencies in observed cells
+131,130,,Supp S10,Supp. S10,Rare-state misses: rules 1–8
+132,131,,Supp S10,Supp. S10,Rare-state misses: rules 9–15
+133,132,,Supp S10,Supp. S10,Rare-state fidelity: zero-aware reanalysis
+134,133,,Supp S10,Supp. S10,Sepsis rare states across paired-data budgets
+135,134,,Supp S10,Supp. S10,Sepsis rare-state direction across support budgets
+136,135,,Supp S11,Supp. S11,"Rare-state screen in DRESS, Kawasaki and SLE"
+137,136,52,Supp S11,Supp. S11,Consequence: disease direction flips
+138,137,,Supp S11,Supp. S11,C1Q: single-gene detection versus co-detection
+139,138,,Supp S11,Supp. S11,"Behçet: K=4,000 changes direction more than identity"
+140,139,,Supp S11,Supp. S11,Behcet: one-gene and multi-gene C1Q rules
+141,140,68,Supp S12,Supp. S12,Pairs from other cohorts do not transfer
+142,141,69,Supp S12,Supp. S12,Longer training on other cohorts: gap stays
+143,142,70,Supp S12,Supp. S12,"No pretraining: same cohort, from scratch"
+144,143,71,Supp S12,Supp. S12,"No pretraining: other cohorts, from scratch"
+145,144,72,Supp S12,Supp. S12,Rare states: from scratch vs pretrained
+146,145,74,Supp S13,Supp. S13,Appendix: scSemiProfiler without output threshold
+147,146,,Supp S14,Supp. S14,Bulk-only exploration: calibration populations differ
+148,147,73,Historical archive,Historical archive,Historical scoring and plans
+149,148,49,Historical archive,Historical only,Rare-state fractions: original analysis
+150,149,50,Historical archive,Historical only,Rare-state ranking: original analysis
+151,150,51,Historical archive,Historical only,Rare states: original five-fold analysis
+152,151,55,Historical archive,Historical archive,Bulk-only cohort plan: original snapshot
+153,152,,Historical archive,Historical archive,Historical bulk-only search: no confirmed new finding
+154,153,75,Discussion,Discussion,Discussion and outlook
+155,154,76,Discussion,Discussion,Next: other tissues from a few dozen pairs