Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261013T005008Z-reuse-audit-results-deck/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/results_review_20261013_provenance.json

@@ -0,0 +1,693 @@
+{
+  "schema": "pbisc.results-review.reuse.v1",
+  "as_of_kst": "2026-10-13 09:45 KST",
+  "slides": [
+    {
+      "slide": 1,
+      "section": "Results review",
+      "title": "PBISC-Diffusion: results available now",
+      "protocol": "2026-10-13 09:45 KST | Existing evidence reused; protocol updates are identified separately.",
+      "old_deck_slide": null,
+      "notes": "",
+      "assets": []
+    },
+    {
+      "slide": 2,
+      "section": "Figure roadmap",
+      "title": "What is already supported?",
+      "protocol": "Figure 6 remains reserved for user-prepared reproduction data.",
+      "old_deck_slide": null,
+      "notes": "",
+      "assets": []
+    },
+    {
+      "slide": 3,
+      "section": "Figure 1",
+      "title": "From pretraining to five-pair adaptation",
+      "protocol": "Complete existing concept: K_cond, depth and N_gen are distinct; corpus labels are descriptive.",
+      "old_deck_slide": null,
+      "notes": "",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paper_figures_strictpairs_20261012/figures/fig1_overview.pdf",
+          "sha256": "bf87a0094962a32355b10d49c602950a01b22b96f0cea39bd56a503af1ad6f07",
+          "clip_y_fraction": [
+            0,
+            0.395
+          ]
+        }
+      ]
+    },
+    {
+      "slide": 4,
+      "section": "Figure 1",
+      "title": "Condition, depth and independent cell draws",
+      "protocol": "Complete existing concept: K_cond, depth and N_gen are distinct; corpus labels are descriptive.",
+      "old_deck_slide": null,
+      "notes": "",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paper_figures_strictpairs_20261012/figures/fig1_overview.pdf",
+          "sha256": "bf87a0094962a32355b10d49c602950a01b22b96f0cea39bd56a503af1ad6f07",
+          "clip_y_fraction": [
+            0.395,
+            0.65
+          ]
+        }
+      ]
+    },
+    {
+      "slide": 5,
+      "section": "Figure 1",
+      "title": "Corpus composition and training exposure",
+      "protocol": "Complete existing concept: K_cond, depth and N_gen are distinct; corpus labels are descriptive.",
+      "old_deck_slide": null,
+      "notes": "",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paper_figures_strictpairs_20261012/figures/fig1_overview.pdf",
+          "sha256": "bf87a0094962a32355b10d49c602950a01b22b96f0cea39bd56a503af1ad6f07",
+          "clip_y_fraction": [
+            0.65,
+            1
+          ]
+        }
+      ]
+    },
+    {
+      "slide": 6,
+      "section": "Figure 2",
+      "title": "Native-pseudobulk model comparisons",
+      "protocol": "Completed held-out rescoring. Architecture 1% floor and gene-panel 0.1% floor are separate sweeps.",
+      "old_deck_slide": null,
+      "notes": "",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paper_figures_strictpairs_20261012/figures/fig2_ablation_and_support.pdf",
+          "sha256": "153c665acdb119caf72ebe0428ec0683a9de345ccdbda77c3a35e09608fe968e",
+          "clip_y_fraction": [
+            0.035,
+            0.35
+          ]
+        }
+      ]
+    },
+    {
+      "slide": 7,
+      "section": "Figure 2",
+      "title": "Gene-panel performance and gene coverage",
+      "protocol": "Existing panel benchmark; 6k has five sampling seeds, other panels one. Do not claim universal 6k superiority.",
+      "old_deck_slide": 12,
+      "notes": "Original deck slide 12\nSLE, 42 donors, 1,006 genes in every panel. 6k: 5 seeds; base seed best of 5 on r and DEG.\nSources: tools/sec03_panel.py; data/sec03_panel_bench.csv, data/sec03_panel_seeds.csv, data/sec03_panel_loss.csv, data/sec03_panel_headline.csv\n",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec03_panel.pdf",
+          "sha256": "ddbfdbab9cfc924ab7f1f6a3f5c59a74c17648166b329986bb8a075c5eda2986",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 8,
+      "section": "Figure 2",
+      "title": "Healthy real bulk: full fine-tuning and LoRA rank",
+      "protocol": "Completed original donor-level five-fold comparison; r16 is a quality/parameter compromise.",
+      "old_deck_slide": null,
+      "notes": "",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paper_figures_strictpairs_20261012/figures/fig2_ablation_and_support.pdf",
+          "sha256": "153c665acdb119caf72ebe0428ec0683a9de345ccdbda77c3a35e09608fe968e",
+          "clip_y_fraction": [
+            0.35,
+            0.68
+          ]
+        }
+      ]
+    },
+    {
+      "slide": 9,
+      "section": "Figure 2",
+      "title": "Existing learning curve: matched data access",
+      "protocol": "Historical curve: n training + 7 validation pairs; early stopping. These are not the new total-n budgets.",
+      "old_deck_slide": 31,
+      "notes": "Original deck slide 31\nImMicro, donor-level 5-fold CV pooled at each pair count: for every fold, scSemiProfiler's representatives are exactly the n pairs LoRA trained on (plus the inner-validation pairs); 'all' (open symbols, right of the dotted line) = every training donor of the fold, i.e. the 5-fold result, reported separately from the 5-pair comparison. Source: data/sec08_pairs_realism.csv, data/sec08_pairs_signal.csv, data/sec09_same_pairs.csv.\nSources: tools/fig23_healthy.py; data/fig23_healthy.csv\nWang et al. scSemiProfiler: advancing large-scale single-cell studies through semi-profiling with deep generative models and active learning. Nat Commun 15:5989 (2024). doi:10.1038/s41467-024-50150-1",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/fig3_healthy_pairs.pdf",
+          "sha256": "37ea09fc9802cb5d799ca5b3db54d9a36f3eec43ba023563952f844b730d8802",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 10,
+      "section": "Figure 2",
+      "title": "Additional pairs improve donor-level information",
+      "protocol": "Historical n + 7 validation design. The fixed-3000-step strict curve is a separate protocol update.",
+      "old_deck_slide": 26,
+      "notes": "Original deck slide 26\nImMicro, donor-level 5-fold CV, pooled scoring (107 donors); x = LoRA training pairs per fold (the last point is every training donor of the fold, 78\u201379). Calibration fitted on exactly the pairs the LoRA point saw. Selected step = early-stopping step on 7 inner-validation pairs (cap 20,000). Values: coverage LoRA 0.84 / 0.79 / 0.76 / 0.78 / 0.78; NK fraction r LoRA 0.20 / 0.40 / 0.56 / 0.70 / 0.70 at 5 / 10 / 20 / 40 / all pairs. Source: data/sec08_pairs_realism.csv, data/sec08_pairs_signal.csv, data/sec08_selected_steps_per_fold.csv.\nSources: tools/fig23_healthy.py; data/fig23_healthy.csv\nKynk\u00e4\u00e4nniemi et al. Improved precision and recall metric for assessing generative models. NeurIPS 2019\nNaeem et al. Reliable fidelity and diversity metrics for generative models. ICML 2020",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/fig2_healthy_pairs.pdf",
+          "sha256": "78a83d2603464842b32553814d1cfa665383b74f7e19438ea03440d6719c0db2",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 11,
+      "section": "Figure 3",
+      "title": "Measured bulk differs from pseudobulk input",
+      "protocol": "Existing paired-input diagnostic; mean-shift removal uses paired pseudobulks and is not a deployment method.",
+      "old_deck_slide": 17,
+      "notes": "Original deck slide 17\nImMicro, 107 healthy donors. Model inputs (log1p CPM, 5,508 genes), not learned ctx. D = median donor-to-donor distance; PCA-30 values are lower bounds (full space 3.19 D, 0.91 shared). Shift removal is a diagnostic that uses paired pseudobulks. Pseudobulk PC1 is sc pool batch (pool R\u00b2 0.96 on PC1).\nSources: tools/sec05_condition_space.py; data/sec05_condition_space_numbers.csv, data/sec05_condition_space_coords.csv\n",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec05_condition_space.pdf",
+          "sha256": "e0701bd49471a731a4ec6489770a7187cda28727c8a176ab681ad9381624f4c2",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 12,
+      "section": "Figure 3",
+      "title": "Healthy donors: completed exact-five comparison",
+      "protocol": "ImMicro: exactly 5 pairs, no extra validation, 3000-step LoRA; five draws, remaining 102 donors per draw.",
+      "old_deck_slide": 30,
+      "notes": "Original deck slide 30\nMain comparison on healthy donors: per draw, LoRA-5 and scSemiProfiler get the same 5 paired donors (scSemiProfiler: as its representatives) and are tested on the other 102 from their real bulk; 5 draws. Zero-shot (no pairs) shown for reference. Dashed: real vs real cells. Values (LoRA-5 / scSemiProfiler-5): coverage 0.792 / 0.036; precision 0.952 / 0.518; comp_l1 0.176 / 0.233; c2st_knn_auc 0.943 / 0.999; mmd2_bw1 0.023 / 0.186; en_all_sq 1.547 / 3.649. Source: data/sec09_fewpair_cells.csv, data/sec09_extra_fewpair.csv.\nSources: tools/fig23_healthy.py; data/fig23_healthy.csv\nWang et al. scSemiProfiler: advancing large-scale single-cell studies through semi-profiling with deep generative models and active learning. Nat Commun 15:5989 (2024). doi:10.1038/s41467-024-50150-1\nLopez-Paz & Oquab. Revisiting classifier two-sample tests. ICLR 2017",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/fig3_healthy_fewpair.pdf",
+          "sha256": "d93b8cc9c49be301038c021c00e6d41af87aae06ddeeebca2fb8092a2fe808f1",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 13,
+      "section": "Figure 3",
+      "title": "Healthy cell-type composition from saved scores",
+      "protocol": "CPU redraw of completed exact-five draws. Each donor is averaged over held-out draws; descriptive, no new p-values.",
+      "old_deck_slide": null,
+      "notes": "",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/review_20261013_healthy_composition.pdf",
+          "sha256": "381c35831e0e03b1ac5f78b176857240bf38b100ec32b227973430bf58fb0123",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 14,
+      "section": "Figure 3",
+      "title": "Representative-dependent composition errors",
+      "protocol": "Existing healthy five-draw diagnostic. This evaluates composition dependence, not literal copying of every cell.",
+      "old_deck_slide": 32,
+      "notes": "Original deck slide 32\nImMicro, 5 donors. scSemiProfiler assigns each test sample one representative; 'copy' uses that representative's observed composition as the prediction. Right: Spearman over test samples between each method's composition error and the copy's error. Source: data/sec09_scsemi_representative.csv (scsemi/analysis/representative_dependence_20261002).\nSources: tools/fig23_healthy.py; data/fig23_healthy.csv\nWang et al. scSemiProfiler: advancing large-scale single-cell studies through semi-profiling with deep generative models and active learning. Nat Commun 15:5989 (2024). doi:10.1038/s41467-024-50150-1",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/fig3_healthy_copy.pdf",
+          "sha256": "644c45e3a49eb7d16b15c4da1736ce5eb94ea8d2dd0213ce820a6d111b215964",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 15,
+      "section": "Figure 3",
+      "title": "Operational cost after supplied pretraining",
+      "protocol": "Existing measurements: adaptation + target generation; pretraining, GPU wait and scoring excluded.",
+      "old_deck_slide": 34,
+      "notes": "Original deck slide 34\nSame 25 draws as the '5 donors' slides (5 cohorts x 5 draws; same 5 training donors and test samples for both methods). GPU time on one GB10, every step alone on the GPU (shared lock). LoRA-5: fine-tune = RUN.json wall_seconds of the 3,000-step LoRA; generate = wall clock of the sampling job in the queue log (model load + DDIM 50, 1,500 cells per test sample). scSemiProfiler-5 (upstream c2aea02, default budget): stage 1 + stage 2 pretraining of every representative (model receipts) + its per-target inference (5 ministages x 150 epochs, export receipt total_target_seconds). Not counted on either side: CPU data preparation, scoring, work lost to interruptions, and PBISC's one-time pretraining (15.9 h; 'Fine-tuning with paired bulk + sc' slide), which every cohort reuses. scSemiProfiler time grows with the number of test samples because it trains one model per sample. Per-draw ratio scSemiProfiler / LoRA: ImMicro 15\u201319\u00d7; Sepsis 26\u201347\u00d7; Imperial 19\u201326\u00d7; Giroux 54\u201363\u00d7; Ji 11\u201320\u00d7 (all 25 draws 11\u201363\u00d7; keyword and figure = mean of each cohort's 5 draws).\nSources: tools/sec09_time_to_apply.py; data/sec09_time_to_apply.csv, data/sec09_time_to_apply_runs.csv; data/sec06_ft_method.csv (pretrain_wall_hours)\nWang et al. scSemiProfiler: advancing large-scale single-cell studies through semi-profiling with deep generative models and active learning. Nat Commun 15:5989 (2024). doi:10.1038/s41467-024-50150-1",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec09_time_to_apply.pdf",
+          "sha256": "614303f5f76d1ce103b93ebb38973c93c2463e20d78055cd430f8fea2d84929a",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 16,
+      "section": "Figure 4",
+      "title": "Paired cohorts and their evaluation units",
+      "protocol": "Existing paired-cohort map. Five donors means 5 pairs in Sepsis, 12-15 Imperial, 12-14 Giroux, 7-9 Ji.",
+      "old_deck_slide": 37,
+      "notes": "Original deck slide 37\nUsable pair = passed the pairing identity check and >= 200 cells (fold >= 0 in data/<cohort>/folds.csv); dropped 15 identity-fail, 11 borderline and 1 low-cell pairs of 430. ImMicro: no identity flag, all used. Imperial, Giroux, Ji: several pairs per donor. Case vs control = samples in scores/<cohort>/pooled/summary.json. Giroux 'moderate' = GEO 'severe' (paper term); its moderate vs mild subjects are 4F/0M vs 0F/5M (scored), so its disease metrics exclude sex-linked genes. Labels: data/<cohort>/QC_REPORT.md; ImMicro: GEO GSE314922/GSE314416.\nSources: tools/sec09_paired_validation.py; data/sec09_cohorts.csv, data/sec09_quoted_numbers.csv, data/sec09_slide_text.json\n",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec09_cohorts.pdf",
+          "sha256": "948538720b7f3be73cab30e3e26819ab9c46021c1d65bab1d112eb0a3afac88d",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 17,
+      "section": "Figure 4",
+      "title": "Disease signal: completed five-donor draws",
+      "protocol": "Existing five-donor protocol; sample counts differ by cohort. Original sample-level estimator; corrected update separate.",
+      "old_deck_slide": 38,
+      "notes": "Original deck slide 38\nMain comparison: 5 donors per draw, LoRA\u20135 vs scSemiProfiler with the same 5 donors as representatives (and the two zero\u2013shot PBISC versions), 5 draws per cohort, every other usable sample tested. logFC r = Pearson r of case\u2013vs\u2013control logFC (mean CP10K) with the observed cells, mean over T, B, monocyte and NK. Partial r = the same after both logFCs are residualised on the input bulk's logFC: the cell\u2013type signal the bulk does not carry. Top\u201350 DEG: LoRA\u20135 ahead in 12/20 draws, tied in 8 (Giroux and Ji: few DEG in the observed cells, mostly 0 for every method). Giroux's 5\u2013donor test sets hold 10\u201312 samples from 4 subjects (2 vs 2), so its disease metrics are noisy; there scSemiProfiler\u20135 is ahead. Giroux: sex\u2013linked genes excluded (severity is confounded with sex). Per cohort, mean over draws (r / partial r / DEG): Sepsis LoRA\u20135 0.35 / 0.30 / 16.4, scSemiProfiler\u20135 0.40 / 0.33 / 5.2; Imperial LoRA\u20135 0.75 / 0.60 / 113.6, scSemiProfiler\u20135 0.56 / 0.43 / 46.8; Giroux LoRA\u20135 0.30 / 0.19 / 0.2, scSemiProfiler\u20135 0.40 / 0.24 / 0.0; Ji LoRA\u20135 0.33 / 0.30 / 0.2, scSemiProfiler\u20135 0.15 / 0.09 / 0.0. Source: data/fig4_disease.csv (scores[_nosex]/<cohort>/fewshot_n5/draw<j>[_scsemi5]).\nSources: tools/fig4_disease.py; data/fig4_disease.csv\nWang et al. scSemiProfiler: advancing large-scale single-cell studies through semi-profiling with deep generative models and active learning. Nat Commun 15:5989 (2024). doi:10.1038/s41467-024-50150-1\nVi\u00f1as Torn\u00e9 et al. Systema: a framework for evaluating genetic perturbation response prediction beyond systematic variation. Nat Biotechnol 44:1050-1059 (2026). doi:10.1038/s41587-025-02777-8",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/fig4_disease_fewpair.pdf",
+          "sha256": "b7c727cf4c76755c4234f07a22f5f314000aa762982d141eb4f6ac60c802bfb6",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 18,
+      "section": "Figure 4",
+      "title": "Disease signal: completed five-fold validation",
+      "protocol": "Original donor-level split and original signal estimator. These results are distinct from the exact-five update.",
+      "old_deck_slide": 39,
+      "notes": "Original deck slide 39\nReported separately from the 5-pair comparison: donor-level 5-fold CV, each sample generated by the fold model that never trained on it, one pooled scoring call per method; scSemiProfiler uses exactly LoRA's training pairs as representatives. Per cohort LoRA vs scSemiProfiler (r / partial r / DEG): Sepsis 0.53 / 0.47 / 23 vs 0.44 / 0.36 / 6; Imperial 0.82 / 0.69 / 143 vs 0.71 / 0.54 / 87; Giroux 0.46 / 0.32 / 5 vs 0.54 / 0.39 / 0; Ji 0.48 / 0.45 / 15 vs 0.27 / 0.23 / 0. Giroux: sex-linked genes excluded. Source: data/fig4_disease.csv (scores[_nosex]/<cohort>/pooled, scsemi_cv_20261002/pooled; sepsis DEG from pooled_deg).\nSources: tools/fig4_disease.py; data/fig4_disease.csv\nWang et al. scSemiProfiler: advancing large-scale single-cell studies through semi-profiling with deep generative models and active learning. Nat Commun 15:5989 (2024). doi:10.1038/s41467-024-50150-1\nVi\u00f1as Torn\u00e9 et al. Systema: a framework for evaluating genetic perturbation response prediction beyond systematic variation. Nat Biotechnol 44:1050-1059 (2026). doi:10.1038/s41587-025-02777-8",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/fig4_disease_fivefold.pdf",
+          "sha256": "003d005069e886f31855cbf455b8f287bf59fe3532ee1cf95826888961689fb7",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 19,
+      "section": "Figure 4",
+      "title": "Disease-gene agreement across list lengths",
+      "protocol": "Existing CAT curves; ranking agreement and chance line. Patient-aware reanalysis reuses the same predictions.",
+      "old_deck_slide": 40,
+      "notes": "Original deck slide 40\nCorrespondence at the top (Irizarry et al. 2005): per cell type, every gene tested on both sides is ranked by |limma-voom moderated t| (case vs control, same donor x type pseudobulks as the DEG panels); CAT(k) = share of the observed top-k genes that are also in the generated top-k with the same logFC sign. Mean over T / B / Mono / NK; 5 donors = mean over 5 draws (scSemiProfiler: its 25 draws), the main comparison; 5-fold = pooled call, shown separately. Chance = k / (2G), G = genes tested on both sides (0.012\u20130.016 at k = 100). No significance threshold, so cohorts where few genes pass FDR (Giroux, Ji) still get a curve. Rank-biased overlap (Webber et al. 2010, p = 0.98) gives the same order: scores_contrast/{fewshot,pooled}_rbo.csv. Giroux: sex-linked genes excluded. Sepsis CAT100 5 donors LoRA 0.092 / scSemi 0.037 / calibration 0.069 / raw 0.028 / chance 0.012; 5-fold LoRA 0.163; Imperial CAT100 5 donors LoRA 0.243 / scSemi 0.133 / calibration 0.198 / raw 0.179 / chance 0.012; 5-fold LoRA 0.270; Giroux CAT100 5 donors LoRA 0.048 / scSemi 0.075 / calibration 0.050 / raw 0.048 / chance 0.016; 5-fold LoRA 0.068; Ji CAT100 5 donors LoRA 0.091 / scSemi 0.030 / calibration 0.021 / raw 0.021 / chance 0.014; 5-fold LoRA 0.152.\nSources: tools/sec09_cat_curves.py; data/sec09_cat_curves.csv\nIrizarry et al. Multiple-laboratory comparison of microarray platforms. Nat Methods 2:345-350 (2005). doi:10.1038/nmeth756\nWebber, Moffat & Zobel. A similarity measure for indefinite rankings. ACM Trans Inf Syst 28(4):20 (2010). doi:10.1145/1852102.1852106\nSquair et al. Confronting false discoveries in single-cell differential expression. Nat Commun 12:5692 (2021). doi:10.1038/s41467-021-25960-2",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec09_cat_curves.pdf",
+          "sha256": "09f3bbd11ca995584b0015800e2afbe02c3fa080efac52851f836f0d956ab8f2",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 20,
+      "section": "Figure 4",
+      "title": "Cell-type expression against deconvolution",
+      "protocol": "Completed BayesPrism/CIBERSORTx comparisons; same five-donor support, method-supported expression metrics.",
+      "old_deck_slide": 41,
+      "notes": "Original deck slide 41\nCell-type deconvolution, not single cells. Case-vs-control per-cell-type logFC (mean CP10K; T, B, Monocyte, NK) from BayesPrism's posterior cell-type expression Z and CIBERSORTx HiRes, scored with the benchmark's own functions; partial r = the same r after both logFCs are residualised on the input bulk's logFC (the cell-type signal the bulk does not carry); top-50 DEG = Welch, summed over types (many ties at 0 in Ji and Giroux). 5-donor draws (the 5-pair comparison), mean over draws with a score. Giroux excludes sex-linked genes, as everywhere in the deck. CIBERSORTx: official Docker images (x86, run under qemu emulation on server_1), single-cell reference mode with S-mode batch correction (up to 400 reference cells per type), fractions in cell units (divided by each type's mean UMI per reference cell). CIBERSORTx HiRes gives per-sample cell-type expression for the genes it imputes per type (others are written as 1 and count as zero here); it imputed no gene of the four contrast types in 1 Ji ChILI and 2 Giroux draws, which therefore have no CIBERSORTx disease score. Source: plans/bayesprism_paired_20261006/results/RESULTS_full.md, RESULTS_cibersortx.md.\nSources: tools/secS_deconvolution.py; data/secS_deconvolution.csv; plans/bayesprism_paired_20261006/results/{RESULTS_full.md, RESULTS_cibersortx.md, TABLE_*_full.csv, TABLE_*_cibersortx.csv}\nChu et al. Cell type and gene expression deconvolution with BayesPrism enables Bayesian integrative analysis across bulk and single-cell RNA sequencing in oncology. Nat Cancer 3:505-517 (2022). doi:10.1038/s43018-022-00356-3\nNewman et al. Determining cell type abundance and expression from bulk tissues with digital cytometry. Nat Biotechnol 37:773-782 (2019). doi:10.1038/s41587-019-0114-2",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/secS_deconv_disease.pdf",
+          "sha256": "31ddaf4d792943114ae52c07a330a3f63d636b8adca9ca311aa6b0449a812f80",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 21,
+      "section": "Figure 4",
+      "title": "Composition against deconvolution",
+      "protocol": "Completed deconvolution outputs already exist. Verified identical Sepsis supports use adoption instead of method reruns.",
+      "old_deck_slide": 46,
+      "notes": "Original deck slide 46\nCell-type deconvolution (cell-type fractions and per-type expression, not single cells) given the same paired donors as PBISC LoRA, on the paired benchmark's own splits, observed truth and scorer: 5-donor draws (reference = the draw's 5 training donors' observed cells, mixture = the test samples' real bulk) and 5-fold CV (reference = each fold's LoRA training + inner-validation pairs; pooled). BayesPrism: recommended gene set (protein-coding minus chrX/Y/M and ribosomal genes, 15-19k), reference states = gate | AIFI L2; its better theta (initial or final) in cell units is shown. CIBERSORTx: official Docker images (x86, run under qemu emulation on server_1), single-cell reference mode with S-mode batch correction (up to 400 reference cells per type), fractions in cell units (divided by each type's mean UMI per reference cell). CIBERSORTx HiRes gives per-sample cell-type expression for the genes it imputes per type (others are written as 1 and count as zero here); it imputed no gene of the four contrast types in 1 Ji ChILI and 2 Giroux draws, which therefore have no CIBERSORTx disease score. Composition L1 over the six-type gate, lower is better. 5-fold: PBISC has the lowest L1 except in Giroux (scSemiProfiler). Pipeline check: on pseudobulk mixtures BayesPrism recovers the observed mRNA shares (e.g. T 0.39 vs 0.38), so the error comes from real bulk. Source: plans/bayesprism_paired_20261006/results/RESULTS_full.md, RESULTS_cibersortx.md.\nSources: tools/secS_deconvolution.py; data/secS_deconvolution.csv; plans/bayesprism_paired_20261006/results/{RESULTS_full.md, RESULTS_cibersortx.md, TABLE_*_full.csv, TABLE_*_cibersortx.csv}\nChu et al. Cell type and gene expression deconvolution with BayesPrism enables Bayesian integrative analysis across bulk and single-cell RNA sequencing in oncology. Nat Cancer 3:505-517 (2022). doi:10.1038/s43018-022-00356-3\nNewman et al. Determining cell type abundance and expression from bulk tissues with digital cytometry. Nat Biotechnol 37:773-782 (2019). doi:10.1038/s41587-019-0114-2\nWang et al. scSemiProfiler: advancing large-scale single-cell studies through semi-profiling with deep generative models and active learning. Nat Commun 15:5989 (2024). doi:10.1038/s41467-024-50150-1",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/secS_deconv_composition.pdf",
+          "sha256": "54a24da70bf18e208b459775586cdc75643d2bd585526c8e92875e415cf811e7",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 22,
+      "section": "Figure 5",
+      "title": "Rare-state frequencies in the observed cells",
+      "protocol": "Saved corrected analysis of original five-donor outputs; median across draws, state share within parent population.",
+      "old_deck_slide": null,
+      "notes": "",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/review_20261013_rare_observed.pdf",
+          "sha256": "1a87f32f42e590c8a96f53b4e0351584528c229678bffddcb149bad4e8237c24",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 23,
+      "section": "Figure 5",
+      "title": "Rare-state fidelity after zero-aware reanalysis",
+      "protocol": "Original five-donor predictions; zero draws retained and actual donor IDs. 59 state/cohort combinations, including controls.",
+      "old_deck_slide": null,
+      "notes": "",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/review_20261013_rare_corrected_comparison.pdf",
+          "sha256": "85d148c2f6923fec18760167b4170a3e0202478cdbec4927f7c4b0ef869280af",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 24,
+      "section": "Figure 5",
+      "title": "Behcet: one-gene and multi-gene C1Q rules",
+      "protocol": "Descriptive unmatched reference includes training donors. Same-three-pair calibration; 2 identity-pass + 1 uncertain pair.",
+      "old_deck_slide": null,
+      "notes": "",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paper_figures_strictpairs_20261012/figures/fig5_behcet_rules_diagnostic.pdf",
+          "sha256": "6d0662a0aad0ab319e0c1ba7c468da41f89d891f970e9f0fabd903b0d2fc45f6",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 25,
+      "section": "Work status",
+      "title": "Remaining work has three different meanings",
+      "protocol": "Existing evidence stays visible; incomplete replacements are not presented as completed new results.",
+      "old_deck_slide": null,
+      "notes": "",
+      "assets": []
+    },
+    {
+      "slide": 26,
+      "section": "Supporting evidence",
+      "title": "Existing control-cell embedding diagnostic",
+      "protocol": "Existing same-noise input diagnostic; not the new LoRA/scSemi OOF embedding. UMAP overlap is not a quantitative score.",
+      "old_deck_slide": 19,
+      "notes": "Original deck slide 19\nImMicro, 107 donors. Final model f01 6k 1\u00d7 E16 (DDIM 50, EMA); each cell starts from the same noise under every input. Precision/coverage from one scoring call. Composition L1: raw 0.455, pseudobulk 0.166, cohort-mean null 0.161. The UMAP is fitted on real cells and overstates overlap.\nSources: tools/sec05_cells_offmanifold.py; data/sec05_cells_metrics.csv, data/sec05_cells_umap_coords.csv\nKynk\u00e4\u00e4nniemi et al. Improved precision and recall metric for assessing generative models. NeurIPS 2019\nNaeem et al. Reliable fidelity and diversity metrics for generative models. ICML 2020",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec05_cells_umap.pdf",
+          "sha256": "137e1d1351e9dc79aa3398590468d53835ef4341fd5d64dc29c5ebb02d2dfaed",
+          "clip_y_fraction": null
+        },
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec05_cells_metrics.pdf",
+          "sha256": "a555f4bced0da2cc6ea25d35c27b0a4e0992c377f40abc356edda85e9e2fed92",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 27,
+      "section": "Supporting evidence",
+      "title": "Generated-cell quality across all cohorts",
+      "protocol": "Completed five-donor comparisons. Multi-visit cohorts used more than five paired samples.",
+      "old_deck_slide": 58,
+      "notes": "Original deck slide 58\nBase f01 6k\u00b71x\u00b7E16 (EMA, DDIM 50) + LoRA r=16, 3,000 steps. LoRA-5 / scSemi-5 = LoRA / scSemiProfiler trained on the same 5 donors (5\u201315 pairs); test = all other usable pairs. Zero-shot = base model without fine-tuning, fed each test pair's real bulk as measured; zero-shot + calibration = the same model fed that bulk after gene-wise calibration, alpha fitted on the same 5 training donors' pairs (both in the same call as LoRA-5). scSemi-5 scored in a separate call, same test set and reference (sha-checked); 25/25 draws. Verdict = better in >= 80% of shared draws. LoRA-5 lower L1 than scSemi-5 in 4/5 (Giroux split 3/5). LoRA-5 > zero-shot + calibration: coverage 5/5, precision 3/5, depth-matched 5/5 and 3/5 cohorts; > zero-shot (real bulk): coverage 5/5, precision 5/5, depth-matched 5/5 and 4/5. Depth-matched P/C vs scSemi-5: same ranking, coverage LoRA-5 0.83\u20130.97 vs scSemi-5 0.15\u20130.28. Composition L1, LoRA-5 minus zero-shot + calibration, per-donor Wilcoxon (gain = p < 0.05): ImMicro -0.03 (p 2e-10), Sepsis +0.02 (p 0.79), Imperial +0.02 (p 0.38), Giroux -0.06 (p 0.38), Ji -0.01 (p 0.70); minus zero-shot (real bulk): ImMicro -0.27 (p 2e-18), Sepsis -0.04 (p 0.15), Imperial -0.35 (p 2e-08), Giroux -0.11 (p 0.04), Ji -0.23 (p 0.02). scANVI labels; P/C = PCA-50, k=10 balls. Scores: scores/<cohort>/fewshot_n5/.\nSources: tools/sec09_paired_validation.py; data/sec09_fewpair_cells.csv, data/sec09_scsemi_status.csv, data/sec09_slide_text.json\nKynk\u00e4\u00e4nniemi et al. Improved precision and recall metric for assessing generative models. NeurIPS 2019\nNaeem et al. Reliable fidelity and diversity metrics for generative models. ICML 2020\nWang et al. scSemiProfiler: advancing large-scale single-cell studies through semi-profiling with deep generative models and active learning. Nat Commun 15:5989 (2024). doi:10.1038/s41467-024-50150-1",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec09_fewpair_cells.pdf",
+          "sha256": "ed9849628f3b1556846bef6359aa09ffd79983541049bc518a74187bac916e21",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 28,
+      "section": "Supporting evidence",
+      "title": "Precision and coverage across neighbourhood sizes",
+      "protocol": "Completed sensitivity at k=3,5,10; no new sensitivity campaign required.",
+      "old_deck_slide": 60,
+      "notes": "Original deck slide 60\nPrecision and coverage recomputed with k = 3, 5, 10 nearest-neighbour balls in the scorer's PCA-50 (k = 10 reproduces the main-slide values exactly). Means over the 5 draws per cohort; error bars = SD; grey = real vs real (R1 vs R2, same samples). Both metrics fall with smaller k for every set including the real control (real coverage at k = 3: 0.92\u20130.99), so compare arms with the real row, not with 1. Expected coverage of a perfect generator (Naeem et al. 2020, eq. 9) at k = 3/5/10: 0.88/0.97/1.00 in every cohort. LoRA-5 first on coverage at every k in 25/25 draws; on precision in 23/21/20 of 25 draws (k = 3/5/10); not first: Sepsis 0/1/2, Imperial 2/2/2, Giroux 0/1/1 draws at k = 3/5/10, beaten by zero-shot + calibration. 5-fold pooled design (not drawn): coverage ranking of the 5 arms across k unchanged; LoRA first at every k in 5/5 cohorts. Scores: scores_extra/<cohort>/fewshot_n5/draw<j>/metrics.csv.\nSources: tools/sec09_extra_metrics.py; data/sec09_extra_ksens.csv; plans/paired_finetune_intradataset_20260927/scores_extra/\nNaeem et al. Reliable fidelity and diversity metrics for generative models. ICML 2020\nKhayatkhoei & AbdAlmageed. Emergent asymmetry of precision and recall for measuring fidelity and diversity of generative models in high dimensions. ICML 2023",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec09_extra_ksens.pdf",
+          "sha256": "a3ef5e661ccfff24e5ee0e25b5a6e33c9c34f6bc3e41aac0d222b319a87413c1",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 29,
+      "section": "Supporting evidence",
+      "title": "Same training pairs: five-fold method comparison",
+      "protocol": "Completed original five-fold LoRA/scSemi comparison; not the new total-n curve.",
+      "old_deck_slide": 62,
+      "notes": "Original deck slide 62\n5-fold CV: for every fold, scSemiProfiler's representatives are exactly LoRA's training + inner-val pairs (the other four folds; checked per fold), test = the fold's samples from their real bulk only. Pooled = one scoring call per method over every usable sample, same donors and reference (sha-checked). scSemiProfiler = upstream c2aea02, default settings, official output. GPU time per fold on one GB10: LoRA = fine-tune + sampling 1,500 cells per test sample; scSemiProfiler = pretraining of every representative + one model per test sample (receipts); data preparation and scoring not counted on either side. scSemiProfiler ahead on composition: Giroux. scSemiProfiler ahead on logFC r: Giroux.  Equal pair budget, ImMicro (n = 5, 10, 20, 40 pairs): scSemiProfiler L1 0.218\u20130.227 and coverage 0.04\u20130.04 at every n; LoRA L1 0.171\u20130.188, coverage 0.76\u20130.84 (drawn on the 'Pairs needed: composition, coverage, training steps' slide). Equal pair budget, Sepsis (n = 5, 10, 20, 40 pairs): scSemiProfiler L1 0.516\u20130.536 and coverage 0.10\u20130.10 at every n; LoRA L1 0.365\u20130.428, coverage 0.77\u20130.82 (drawn on the 'Pairs needed: composition, coverage, training steps' slide). Source: plans/paired_finetune_intradataset_20260927/scores/<cohort>/{pooled,scsemi_cv_20261002/pooled}/ and scsemi/analysis/cv_budget_20261005/.\nSources: tools/sec09_same_pairs.py; data/sec09_same_pairs.csv; plans/paired_finetune_intradataset_20260927/scsemi/analysis/cv_budget_20261005/\n",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec09_same_pairs.pdf",
+          "sha256": "85e1244526ce3967fad550ba47ed9a121fb91bcd3798c8fb42d0ab09d5d989b4",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 30,
+      "section": "Supporting evidence",
+      "title": "Pairs from other cohorts do not replace local pairs",
+      "protocol": "Completed leave-one-cohort-out test at K=1000; target-cohort sc was excluded from training.",
+      "old_deck_slide": 68,
+      "notes": "Original deck slide 68\nEach held-out cohort is scored on all its usable samples against the same frozen reference as its within-cohort benchmark. Zero-shot = base model on real bulk. Calibration from other cohorts = gene-wise alpha (mean pseudobulk CPM / mean bulk CPM) fitted on the training cohorts' pairs, cohort-balanced; 0 target pairs. LoRA r16 / r64 / full fine-tune = trained on the other PBMC cohorts' pairs plus Liu 2021 (training only: its cells are in the pretraining corpus; 44 pairs after excluding 14 hashtag-negative-heavy rows), cohort-balanced batches, inner-validation on 10 % of each training cohort's donors, step-saturation protocol (20k-step blocks, stop when a block lowers the inner-validation loss by < 1 % of the total), value at the last milestone (number in the bar = steps). Ji's bulk is whole blood and is never trained on. Own-cohort LoRA = LoRA r16 on the cohort's own pairs, 5-fold pooled. Giroux's logFC r excludes 196 sex-linked genes. Best on composition among zero-target-pair arms: ImMicro Calibration from other cohorts, Sepsis Calibration from other cohorts, Imperial Full fine-tune, other cohorts, Giroux Calibration from other cohorts, Ji (whole blood) Calibration from other cohorts.\nSources: tools/sec08_cross_cohort.py; data/sec08_cross_cohort.csv, data/sec08_cross_cohort_steps.csv; plans/paired_crosscohort_20261002/ (README, scores/aggregate.csv)\n",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec08_cross_cohort.pdf",
+          "sha256": "cae7ea9560a094324f8a71105524a5d72b5977168c9478b52649c8fd3338445c",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 31,
+      "section": "Supporting evidence",
+      "title": "Longer cross-cohort training leaves a gap",
+      "protocol": "Existing saturation experiment; no additional saturation run introduced here.",
+      "old_deck_slide": 69,
+      "notes": "Original deck slide 69\nSame runs as the previous slide. Each point = the milestone checkpoint at the end of a 20k-step block (or at the early stop), sampled and scored on the held-out cohort. Dashed = calibration from other cohorts; dotted = LoRA on the cohort's own pairs. Training stopped (steps): LoRA r16 ImMicro 28k, Sepsis 26k, Imperial 40k, Giroux 35k, Ji 40k; LoRA r64 ImMicro 39k, Sepsis 59k, Imperial 69k, Giroux 69k, Ji 51k; Full fine-tune ImMicro 65k, Sepsis 108k, Imperial 93k, Giroux 120k, Ji 86k. Saturation steps across cohorts: r16 up to 40k, r64 up to 69k, full up to 120k. Training time to saturation (RUN.json wall seconds, one GB10): LoRA r16 ImMicro 21 min, Sepsis 20 min, Imperial 30 min, Giroux 28 min, Ji 31 min; LoRA r64 ImMicro 27 min, Sepsis 44 min, Imperial 54 min, Giroux 53 min, Ji 41 min; Full fine-tune ImMicro 59 min, Sepsis 93 min, Imperial 82 min, Giroux 107 min, Ji 77 min. Gap = no milestone of any mode reaches the own-cohort LoRA's coverage (true in all five cohorts); full fine-tune composition L1 from the first to the last milestone: ImMicro +0.005, Sepsis +0.087, Imperial -0.052, Giroux -0.078, Ji +0.192 (drift >= 0.05 in Sepsis and Ji; improved by >= 0.05 in Imperial and Giroux).\nSources: tools/sec08_cross_cohort.py; data/sec08_cross_cohort.csv (minutes = RUN.json wall_seconds / 60), data/sec08_cross_cohort_steps.csv; plans/paired_crosscohort_20261002/ (README, scores/aggregate.csv, runs/hoL_<cohort>/K1000/<lora16sat|lora64sat|fullsat>/RUN.json)\n",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec08_cross_cohort_steps.pdf",
+          "sha256": "ac0e1b49613dba0f44c1bfbfbb7c964998ee4ff5782f686fa0f4162f9634da33",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 32,
+      "section": "Supporting evidence",
+      "title": "From-scratch control: current evidence and limits",
+      "protocol": "Completed Sepsis fold0 pilot, not all folds. Pretraining supports speed/few-pair adaptation, not universal superiority.",
+      "old_deck_slide": 70,
+      "notes": "Original deck slide 70\nPilot on one fold: sepsis fold 0, 17 test donors (12 sepsis, 5 controls); folds 1-4 not run, so no claim yet. From scratch = the same denoiser (54.2 M weights) re-initialised at random (tools/ft_sdd_aug.py --init random; seed 20260627) and trained in full on the cohort's own pairs at the pretraining lr 2e-4, batch 256, EMA 0.999, inner-val selection every 1,000 steps, patience 10, cap 100k steps; the checkpoint supplies only the architecture, 6k panel, value scaler and bulk z-score constants. Pretrained + LoRA = f01 6k 1x E16 + LoRA r16 (lr 5e-4). Same training pairs (5 or 58, + 7-8 inner-val), real-bulk conditions K = 1,000, test donors and sampling seed; recipes = K-diversified pseudobulk augmentation. The 58-pair scratch arms were still improving at the 100k cap. Coverage = PCA-50 k = 10 balls (precision, C2ST, MMD, E_N agree; scores_extra/ye2026_sepsis/fold0_scratch_20261005/). Donor composition r = per cell type (T, B, Mono, NK) Pearson r over the 17 donors of generated vs observed fraction, mean of the four. Dashed = zero-shot. Training time = RUN.json wall seconds on one GB10 (LoRA also needs the one-time 15.9 h pretraining, which every cohort reuses). Source: plans/paired_finetune_intradataset_20260927/scores/ye2026_sepsis/fold0_scratch_20261005/ and README 'No-pretraining ablation'.\nSources: tools/sec07_no_pretraining.py; data/sec07_no_pretraining.csv; plans/paired_finetune_intradataset_20260927/scores/ye2026_sepsis/fold0_scratch_20261005/\n",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec07_no_pretraining.pdf",
+          "sha256": "4a46aa5e38a600d2f67e051e165b1ffdc1afb6c564b2352a9d58eb5e5e4f0f7b",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 33,
+      "section": "Supporting evidence",
+      "title": "scSemiProfiler readout sensitivity",
+      "protocol": "Completed modified-readout check; main results use official output. Models were not retrained for the sensitivity.",
+      "old_deck_slide": 74,
+      "notes": "Original deck slide 74\nAppendix, sensitivity only. scSemiProfiler's official output (upstream commit c2aea02) is the decoder's ZINB mean with every value <= 10 normalized units set to 0; the cut is fixed in the code (vaegan.myVAE.sample, bound = 10, and fast_semi: xsemi * (xsemi > 10)), not an option. At our cohorts' depth (about 1,000-2,000 UMI per cell on the panel) it removes most low- and mid-expressed genes. Here the same trained per-target models were read out again without that cut (CPU, no retraining, target seeds as in the run), exported and scored exactly as the official arm (same 25 five-donor draws, donors, reference; Giroux disease metrics without sex-linked genes). A control re-read with the cut reproduces the official output: pseudobulk r >= 0.996 over 1790 targets, coverage within 0.015. Because this changes the method's official read-out, the main slides keep the official output. Bars = mean of 5 draws, dots = draws. Source: plans/paired_finetune_intradataset_20260927/scsemi/analysis/output_variants_20261003/ (README section 'Read-out sensitivity').\nSources: plans/paired_finetune_intradataset_20260927/scsemi/tools/{resample_output_variants,compare_output_variants}.py; tools/secA_scsemi_readout.py; data/secA_scsemi_readout.csv\n",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/secA_scsemi_readout.pdf",
+          "sha256": "8d88d3ba4d9e6bab62c739d03830dd5c90a72a580ea6ac1bca679ab6584ccad2",
+          "clip_y_fraction": null
+        }
+      ]
+    },
+    {
+      "slide": 34,
+      "section": "Supporting evidence",
+      "title": "SLE pseudobulk development benchmark",
+      "protocol": "Completed pseudobulk benchmark; separate from external paired real-bulk validation. PBISC five seeds, scSemi one run.",
+      "old_deck_slide": 42,
+      "notes": "Original deck slide 42\nGSE135779, 42 held-out donors (30 SLE / 12 HC). PBISC f01 6k E16, 5 seeds (mean \u00b1 SD); baselines 1 draw. scSemiProfiler also gets real sc of 14 in-cohort donors (1,500 cells each). Partial r: both logFCs residualised on the input bulk's logFC (the cell-type signal the bulk does not carry), 1,006 genes: PBISC 0.208 \u00b1 0.012, scSemiProfiler 0.167.\nSources: tools/sec04_sle_benchmark.py; data/sec04_sle_benchmark.csv, data/sec04_sle_numbers.csv\nWang et al. scSemiProfiler: advancing large-scale single-cell studies through semi-profiling with deep generative models and active learning. Nat Commun 15:5989 (2024). doi:10.1038/s41467-024-50150-1\nXu et al. Probabilistic harmonization and annotation of single-cell transcriptomics data with deep generative models. Mol Syst Biol 17:e9620 (2021)",
+      "assets": [
+        {
+          "path": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec04_sle_benchmark.pdf",
+          "sha256": "17d1cbaf83f9e5f764ab61e445a8d9d10f5f5434b5abefae5dedf2d27210e292",
+          "clip_y_fraction": null
+        }
+      ]
+    }
+  ],
+  "sources": {
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/data/slides_spec.json": {
+      "sha256": "6b8fa055e20a3379ce2e50ca19dbbe285d81e4e95d2c42597e236945b8b2bc23",
+      "role": "original 76-slide deck specification"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/immicro/fewshot_n5/draw0/per_donor.csv": {
+      "sha256": "6aae52c2d1a30139025aeae7cc2b44cbecb574f8bd486e2bfbadc02c4aa5444b",
+      "role": "existing healthy scores"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/immicro/fewshot_n5/draw0_scsemi5/per_donor.csv": {
+      "sha256": "4ae35b3c03aa54dfb0e8ffbc03847b015c9defc6d857e2eae51472ea0f015355",
+      "role": "existing healthy scores"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/immicro/fewshot_n5/draw1/per_donor.csv": {
+      "sha256": "b988d8608d022898da41c04275110d95f0b28daa145bec10027ed7e9193cd473",
+      "role": "existing healthy scores"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/immicro/fewshot_n5/draw1_scsemi5/per_donor.csv": {
+      "sha256": "29e5d70b1c15ad2d850dd3cc805b32f693b415c75507920f9ca87e2c4c55f7d0",
+      "role": "existing healthy scores"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/immicro/fewshot_n5/draw2/per_donor.csv": {
+      "sha256": "1caf3d114796869aa6a0fb6edbeddc93dfc8efe8dbd7bd332cb9ea12f0d1a79f",
+      "role": "existing healthy scores"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/immicro/fewshot_n5/draw2_scsemi5/per_donor.csv": {
+      "sha256": "0f257a24ef9dde20e4f1ec8dce1bbcc7199a5ec01d8efa37de27f32d5689eef2",
+      "role": "existing healthy scores"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/immicro/fewshot_n5/draw3/per_donor.csv": {
+      "sha256": "a9bf586d886d1530712b53ae01457b20fb2b9018e698d4bfcb3b6c80a883e5d7",
+      "role": "existing healthy scores"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/immicro/fewshot_n5/draw3_scsemi5/per_donor.csv": {
+      "sha256": "9f9d6a9777d160895e4cc102e496cbf385a8713e95e871a77ebb3165779cbe28",
+      "role": "existing healthy scores"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/immicro/fewshot_n5/draw4/per_donor.csv": {
+      "sha256": "c6bd999a3c963f9ef7f5dc8deeb5a7255c5dae709b4d7fdb330df1efba4c8e8f",
+      "role": "existing healthy scores"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/immicro/fewshot_n5/draw4_scsemi5/per_donor.csv": {
+      "sha256": "43ae3ef666fe05f31f023fb8bd0af132478aea67d2f3fb4e56918c4c1b9242fe",
+      "role": "existing healthy scores"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paper_figures_strictpairs_20261012/source_data/rare_all_states.csv": {
+      "sha256": "9dc216b4172ff9e46ac0c498f52049ac1869df742761208cc48597ad3397b36f",
+      "role": "existing corrected rare-state summaries"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paper_figures_strictpairs_20261012/source_data/rare_all_states_aggregate.csv": {
+      "sha256": "5fd9a2a4051ee6443ed20ec26db2499f9bf7b89116e3a373abc1bfacd8737317",
+      "role": "existing zero-aware rare-state metrics"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paper_figures_strictpairs_20261012/figures/fig1_overview.pdf": {
+      "sha256": "bf87a0094962a32355b10d49c602950a01b22b96f0cea39bd56a503af1ad6f07",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paper_figures_strictpairs_20261012/figures/fig2_ablation_and_support.pdf": {
+      "sha256": "153c665acdb119caf72ebe0428ec0683a9de345ccdbda77c3a35e09608fe968e",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec03_panel.pdf": {
+      "sha256": "ddbfdbab9cfc924ab7f1f6a3f5c59a74c17648166b329986bb8a075c5eda2986",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/fig3_healthy_pairs.pdf": {
+      "sha256": "37ea09fc9802cb5d799ca5b3db54d9a36f3eec43ba023563952f844b730d8802",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/fig2_healthy_pairs.pdf": {
+      "sha256": "78a83d2603464842b32553814d1cfa665383b74f7e19438ea03440d6719c0db2",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec05_condition_space.pdf": {
+      "sha256": "e0701bd49471a731a4ec6489770a7187cda28727c8a176ab681ad9381624f4c2",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/fig3_healthy_fewpair.pdf": {
+      "sha256": "d93b8cc9c49be301038c021c00e6d41af87aae06ddeeebca2fb8092a2fe808f1",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/review_20261013_healthy_composition.pdf": {
+      "sha256": "381c35831e0e03b1ac5f78b176857240bf38b100ec32b227973430bf58fb0123",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/fig3_healthy_copy.pdf": {
+      "sha256": "644c45e3a49eb7d16b15c4da1736ce5eb94ea8d2dd0213ce820a6d111b215964",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec09_time_to_apply.pdf": {
+      "sha256": "614303f5f76d1ce103b93ebb38973c93c2463e20d78055cd430f8fea2d84929a",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec09_cohorts.pdf": {
+      "sha256": "948538720b7f3be73cab30e3e26819ab9c46021c1d65bab1d112eb0a3afac88d",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/fig4_disease_fewpair.pdf": {
+      "sha256": "b7c727cf4c76755c4234f07a22f5f314000aa762982d141eb4f6ac60c802bfb6",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/fig4_disease_fivefold.pdf": {
+      "sha256": "003d005069e886f31855cbf455b8f287bf59fe3532ee1cf95826888961689fb7",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec09_cat_curves.pdf": {
+      "sha256": "09f3bbd11ca995584b0015800e2afbe02c3fa080efac52851f836f0d956ab8f2",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/secS_deconv_disease.pdf": {
+      "sha256": "31ddaf4d792943114ae52c07a330a3f63d636b8adca9ca311aa6b0449a812f80",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/secS_deconv_composition.pdf": {
+      "sha256": "54a24da70bf18e208b459775586cdc75643d2bd585526c8e92875e415cf811e7",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/review_20261013_rare_observed.pdf": {
+      "sha256": "1a87f32f42e590c8a96f53b4e0351584528c229678bffddcb149bad4e8237c24",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/review_20261013_rare_corrected_comparison.pdf": {
+      "sha256": "85d148c2f6923fec18760167b4170a3e0202478cdbec4927f7c4b0ef869280af",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paper_figures_strictpairs_20261012/figures/fig5_behcet_rules_diagnostic.pdf": {
+      "sha256": "6d0662a0aad0ab319e0c1ba7c468da41f89d891f970e9f0fabd903b0d2fc45f6",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec05_cells_umap.pdf": {
+      "sha256": "137e1d1351e9dc79aa3398590468d53835ef4341fd5d64dc29c5ebb02d2dfaed",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec05_cells_metrics.pdf": {
+      "sha256": "a555f4bced0da2cc6ea25d35c27b0a4e0992c377f40abc356edda85e9e2fed92",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec09_fewpair_cells.pdf": {
+      "sha256": "ed9849628f3b1556846bef6359aa09ffd79983541049bc518a74187bac916e21",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec09_extra_ksens.pdf": {
+      "sha256": "a3ef5e661ccfff24e5ee0e25b5a6e33c9c34f6bc3e41aac0d222b319a87413c1",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec09_same_pairs.pdf": {
+      "sha256": "85e1244526ce3967fad550ba47ed9a121fb91bcd3798c8fb42d0ab09d5d989b4",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec08_cross_cohort.pdf": {
+      "sha256": "cae7ea9560a094324f8a71105524a5d72b5977168c9478b52649c8fd3338445c",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec08_cross_cohort_steps.pdf": {
+      "sha256": "ac0e1b49613dba0f44c1bfbfbb7c964998ee4ff5782f686fa0f4162f9634da33",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec07_no_pretraining.pdf": {
+      "sha256": "4a46aa5e38a600d2f67e051e165b1ffdc1afb6c564b2352a9d58eb5e5e4f0f7b",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/secA_scsemi_readout.pdf": {
+      "sha256": "8d88d3ba4d9e6bab62c739d03830dd5c90a72a580ea6ac1bca679ab6584ccad2",
+      "role": "figure"
+    },
+    "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/sec04_sle_benchmark.pdf": {
+      "sha256": "17d1cbaf83f9e5f764ab61e445a8d9d10f5f5434b5abefae5dedf2d27210e292",
+      "role": "figure"
+    }
+  },
+  "new_training": false,
+  "new_generation": false,
+  "cpu_work": "saved-table redraws only",
+  "pdf_sha256": "3035f7e8a4194702a2a64e28b6cfc6080b649f4cf242a1734ca2064aef16c652",
+  "pptx_sha256": "84c6a9862df82288fd952da748f1eacf2ff1be29822799ec8119a57529d0d0c4"
+}

paper-releases/pbisc-diffusion/updates/20261013T005008Z-reuse-audit-results-deck/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/results_review_20261013_slide_map.csv

@@ -0,0 +1,35 @@
+slide,section,title,protocol,old_deck_slide
+1,Results review,PBISC-Diffusion: results available now,2026-10-13 09:45 KST | Existing evidence reused; protocol updates are identified separately.,
+2,Figure roadmap,What is already supported?,Figure 6 remains reserved for user-prepared reproduction data.,
+3,Figure 1,From pretraining to five-pair adaptation,"Complete existing concept: K_cond, depth and N_gen are distinct; corpus labels are descriptive.",
+4,Figure 1,"Condition, depth and independent cell draws","Complete existing concept: K_cond, depth and N_gen are distinct; corpus labels are descriptive.",
+5,Figure 1,Corpus composition and training exposure,"Complete existing concept: K_cond, depth and N_gen are distinct; corpus labels are descriptive.",
+6,Figure 2,Native-pseudobulk model comparisons,Completed held-out rescoring. Architecture 1% floor and gene-panel 0.1% floor are separate sweeps.,
+7,Figure 2,Gene-panel performance and gene coverage,"Existing panel benchmark; 6k has five sampling seeds, other panels one. Do not claim universal 6k superiority.",12.0
+8,Figure 2,Healthy real bulk: full fine-tuning and LoRA rank,Completed original donor-level five-fold comparison; r16 is a quality/parameter compromise.,
+9,Figure 2,Existing learning curve: matched data access,Historical curve: n training + 7 validation pairs; early stopping. These are not the new total-n budgets.,31.0
+10,Figure 2,Additional pairs improve donor-level information,Historical n + 7 validation design. The fixed-3000-step strict curve is a separate protocol update.,26.0
+11,Figure 3,Measured bulk differs from pseudobulk input,Existing paired-input diagnostic; mean-shift removal uses paired pseudobulks and is not a deployment method.,17.0
+12,Figure 3,Healthy donors: completed exact-five comparison,"ImMicro: exactly 5 pairs, no extra validation, 3000-step LoRA; five draws, remaining 102 donors per draw.",30.0
+13,Figure 3,Healthy cell-type composition from saved scores,"CPU redraw of completed exact-five draws. Each donor is averaged over held-out draws; descriptive, no new p-values.",
+14,Figure 3,Representative-dependent composition errors,"Existing healthy five-draw diagnostic. This evaluates composition dependence, not literal copying of every cell.",32.0
+15,Figure 3,Operational cost after supplied pretraining,"Existing measurements: adaptation + target generation; pretraining, GPU wait and scoring excluded.",34.0
+16,Figure 4,Paired cohorts and their evaluation units,"Existing paired-cohort map. Five donors means 5 pairs in Sepsis, 12-15 Imperial, 12-14 Giroux, 7-9 Ji.",37.0
+17,Figure 4,Disease signal: completed five-donor draws,Existing five-donor protocol; sample counts differ by cohort. Original sample-level estimator; corrected update separate.,38.0
+18,Figure 4,Disease signal: completed five-fold validation,Original donor-level split and original signal estimator. These results are distinct from the exact-five update.,39.0
+19,Figure 4,Disease-gene agreement across list lengths,Existing CAT curves; ranking agreement and chance line. Patient-aware reanalysis reuses the same predictions.,40.0
+20,Figure 4,Cell-type expression against deconvolution,"Completed BayesPrism/CIBERSORTx comparisons; same five-donor support, method-supported expression metrics.",41.0
+21,Figure 4,Composition against deconvolution,Completed deconvolution outputs already exist. Verified identical Sepsis supports use adoption instead of method reruns.,46.0
+22,Figure 5,Rare-state frequencies in the observed cells,"Saved corrected analysis of original five-donor outputs; median across draws, state share within parent population.",
+23,Figure 5,Rare-state fidelity after zero-aware reanalysis,"Original five-donor predictions; zero draws retained and actual donor IDs. 59 state/cohort combinations, including controls.",
+24,Figure 5,Behcet: one-gene and multi-gene C1Q rules,Descriptive unmatched reference includes training donors. Same-three-pair calibration; 2 identity-pass + 1 uncertain pair.,
+25,Work status,Remaining work has three different meanings,Existing evidence stays visible; incomplete replacements are not presented as completed new results.,
+26,Supporting evidence,Existing control-cell embedding diagnostic,Existing same-noise input diagnostic; not the new LoRA/scSemi OOF embedding. UMAP overlap is not a quantitative score.,19.0
+27,Supporting evidence,Generated-cell quality across all cohorts,Completed five-donor comparisons. Multi-visit cohorts used more than five paired samples.,58.0
+28,Supporting evidence,Precision and coverage across neighbourhood sizes,"Completed sensitivity at k=3,5,10; no new sensitivity campaign required.",60.0
+29,Supporting evidence,Same training pairs: five-fold method comparison,Completed original five-fold LoRA/scSemi comparison; not the new total-n curve.,62.0
+30,Supporting evidence,Pairs from other cohorts do not replace local pairs,Completed leave-one-cohort-out test at K=1000; target-cohort sc was excluded from training.,68.0
+31,Supporting evidence,Longer cross-cohort training leaves a gap,Existing saturation experiment; no additional saturation run introduced here.,69.0
+32,Supporting evidence,From-scratch control: current evidence and limits,"Completed Sepsis fold0 pilot, not all folds. Pretraining supports speed/few-pair adaptation, not universal superiority.",70.0
+33,Supporting evidence,scSemiProfiler readout sensitivity,Completed modified-readout check; main results use official output. Models were not retrained for the sensitivity.,74.0
+34,Supporting evidence,SLE pseudobulk development benchmark,"Completed pseudobulk benchmark; separate from external paired real-bulk validation. PBISC five seeds, scSemi one run.",42.0

paper-releases/pbisc-diffusion/updates/20261013T005008Z-reuse-audit-results-deck/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/review_20261013_healthy_composition.csv

@@ -0,0 +1,6 @@
+method,T cell,B cell,Monocyte,NK cell,DC,Other
+Observed,0.6390479323723364,0.07479609654330255,0.166538380766508,0.09119709084308379,0.022818037313248645,0.00560246216152067
+Raw,0.4526230529595016,0.08467912772585667,0.3353769470404984,0.07074143302180683,0.027246105919003092,0.029333333333333295
+Paired cal.,0.7067220145379024,0.08552377985462092,0.1310463136033229,0.04961090342679124,0.011994600207684288,0.015102388369678057
+scSemiProfiler,0.690181412253375,0.07505337487019725,0.11744174454828656,0.05893572170301138,0.05212731048805812,0.006260436137071619
+PBISC LoRA,0.6918943925233645,0.08173219106957422,0.12600903426791274,0.06422201453790234,0.027388785046728936,0.0087535825545171

paper-releases/pbisc-diffusion/updates/20261013T005008Z-reuse-audit-results-deck/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/review_20261013_rare_observed.csv

@@ -0,0 +1,16 @@
+state,ye2026_sepsis,imperial_eskd,giroux2022,ji2025_chili
+C1Q-high monocyte (reference),5.329627661748478,4.265919811320755,1.167883211678832,4.003724394785848
+Cycling T/NK,0.8088838022178573,0.1641599699821769,0.7065166868085897,0.4877638550163989
+Cycling activated T (CD38/HLA-DR),1.4280325483644989,0.8014029649870511,1.0161920714684534,1.419489203113011
+Proliferating T (label),1.8591367139084984,1.0111062879164858,1.1166945840312674,1.5130984643179766
+Plasmablast (label),1.535064935064935,1.2067150881068744,0.296969696969697,0.5246887319549619
+Plasmablast (markers),2.952380952380953,1.9356418619676328,0.98,
+MS1 monocyte (IL1R2+),13.283176060528776,1.526693125080015,0.063552589768033,15.595903165735567
+Cycling monocyte (Ye 2026),2.232600057351194,1.2057783018867925,1.618599175985874,0.819366852886406
+ISG-high monocyte (positive control),7.814485387547649,27.67688679245283,34.44962942629701,2.839851024208566
+ISG-high T,2.021602160216021,11.121443970072283,5.400463245154212,0.2429060395274373
+DN2 / ABC B,1.911932240328892,2.071775312066574,1.740506329113924,1.550023485204321
+HSPC,0.0683982683982684,0.1224107465991101,0.0555555555555555,0.2747626431622187
+Cytotoxic CD4 T,1.1717249327698809,0.9929078014184396,0.6581009088060169,0.2162356969096315
+Granulocyte-like (LDG),0.8658008658008658,0.2982734857491557,0.4242424242424242,0.4869555755053184
+Platelet signal in monocytes (negative control),9.212198221092756,2.8837536807066955,8.284671532846716,14.972067039106143