diff --git a/paper-releases/pbisc-diffusion/updates/20261006T150739Z-deck-figure-plan-decisions/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/figure_plan.json b/paper-releases/pbisc-diffusion/updates/20261006T150739Z-deck-figure-plan-decisions/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/figure_plan.json new file mode 100644 index 0000000000000000000000000000000000000000..b8dd32d538333a66598b0de44efb4a4ec49137da --- /dev/null +++ b/paper-releases/pbisc-diffusion/updates/20261006T150739Z-deck-figure-plan-decisions/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/figure_plan.json @@ -0,0 +1,310 @@ +{ + "source": "Nature Style Manuscript Writing.pdf (paper-release-20260928/reference-papers, 2026-10-06): six main figures, up to 17 supplementary; short abstract and introduction, descriptive results, short discussion with limitations, detailed methods. User decisions of 2026-10-06: no PBISC-VAE (scSemiProfiler, a VAE-based method, is the comparison); LoRA performance first on the healthy cohort with the realism metrics, the disease-cohort comparison later; CYBERSORTx and other bulk deconvolution methods only in the Supplementary (PBISC generates single cells); Figs 4 and 6 left empty until their analyses are done.", + "status_legend": { + "Ready": "slide exists and fits the panel", + "Adapt": "slide exists; change its scope or style", + "Draft": "new draft exists, not yet approved", + "Build": "results exist; no slide yet", + "Run": "needs a new analysis or experiment", + "Plan": "not planned yet" + }, + "figures": [ + { + "n": 1, + "title": "Design and performance of PBISC", + "claim": "Diffusion on each gene's value and detection, conditioned on bulk, writes realistic blood cells.", + "panels": [ + { + "p": "a", + "content": "Backbones compared: scDiffusion, SDD, PBISC-Diffusion", + "evidence": [ + "Backbone: SDD → PBISC-Diffusion" + ], + "status": "Ready" + }, + { + "p": "b", + "content": "Overview: pretraining, diffusion, generation", + "evidence": [ + "Overview draft (Nature style)" + ], + "status": "Draft" + }, + { + "p": "c", + "content": "What the bulk condition contains", + "evidence": [ + "Bulk conditioning" + ], + "status": "Ready" + }, + { + "p": "d", + "content": "Pretraining corpus: assays, conditions, cell types", + "evidence": [ + "Pretraining corpus" + ], + "status": "Ready" + }, + { + "p": "e", + "content": "Width and context: lower loss, same biology", + "evidence": [ + "Width & context ablation: loss saturation", + "Width & context ablation: benchmark" + ], + "status": "Ready" + }, + { + "p": "f", + "content": "Gene panel: 2k, 3k, 6k, 15k", + "evidence": [ + "Gene context: 2k / 3k / 6k / 15k" + ], + "status": "Ready" + } + ], + "candidates": [], + "details": [ + "PBISC-Diffusion: bulk in, cells out", + "Denoiser: joint value + support diffusion" + ] + }, + { + "n": 2, + "title": "Fine-tuning PBISC with LoRA", + "claim": "On a healthy cohort, a few paired samples and a LoRA of minutes give cells close to the real ones.", + "panels": [ + { + "p": "a", + "content": "Real bulk differs from the training input", + "evidence": [ + "Real bulk ≠ pseudobulk (model input)", + "Same shift in three paired cohorts" + ], + "status": "Ready" + }, + { + "p": "b", + "content": "Without fine-tuning, cells fall off the real manifold", + "evidence": [ + "Consequence: cells off-manifold" + ], + "status": "Ready" + }, + { + "p": "c", + "content": "LoRA on paired bulk and cells; how it is tested", + "evidence": [ + "Fine-tuning with paired bulk + sc", + "Evaluation design" + ], + "status": "Ready" + }, + { + "p": "d", + "content": "How generated cells are scored (the metric suite)", + "evidence": [ + "How we score generated cells" + ], + "status": "Ready" + }, + { + "p": "e", + "content": "Healthy cohort: precision, coverage, composition", + "evidence": [ + "5-fold CV: train on the other 4 folds" + ], + "status": "Adapt", + "note": "ImMicro only; LoRA vs zero-shot and calibration" + }, + { + "p": "f", + "content": "Healthy cohort: classifier AUROC, MMD, E_N", + "evidence": [ + "5-fold: realism beyond precision/coverage" + ], + "status": "Adapt", + "note": "ImMicro only" + }, + { + "p": "g", + "content": "Full, LoRA and condition-encoder fine-tuning", + "evidence": [ + "Full vs LoRA vs cond.-encoder-only" + ], + "status": "Adapt", + "note": "ImMicro only" + }, + { + "p": "h", + "content": "How many pairs are needed", + "evidence": [ + "Pairs needed: composition, coverage, training steps" + ], + "status": "Adapt", + "note": "ImMicro only" + }, + { + "p": "i", + "content": "LoRA rank and K choices", + "evidence": [ + "LoRA rank: r = 4, 16, 64", + "K diversification: no gain" + ], + "status": "Ready", + "note": "supplementary candidate" + } + ], + "candidates": [], + "details": [] + }, + { + "n": 3, + "title": "Comparison with scSemiProfiler", + "claim": "PBISC gives better cells than scSemiProfiler, a VAE-based method, for far less compute.", + "panels": [ + { + "p": "a", + "content": "Lupus benchmark against scSemiProfiler", + "evidence": [ + "SLE benchmark: PBISC vs scSemiProfiler", + "SLE benchmark: per cell type", + "Bulk broadcast baseline" + ], + "status": "Ready" + }, + { + "p": "b", + "content": "Five paired cohorts", + "evidence": [ + "Five paired cohorts" + ], + "status": "Ready" + }, + { + "p": "c", + "content": "Five donors: generated cells", + "evidence": [ + "5 donors: LoRA vs scSemiProfiler — cells", + "5 donors: realism beyond precision/coverage", + "5 donors: precision/coverage at k = 3, 5, 10" + ], + "status": "Ready" + }, + { + "p": "d", + "content": "scSemiProfiler copies its representative", + "evidence": [ + "scSemiProfiler copies its representative" + ], + "status": "Ready" + }, + { + "p": "e", + "content": "Same training pairs, 5-fold", + "evidence": [ + "Same training pairs: LoRA vs scSemiProfiler" + ], + "status": "Ready", + "note": "scSemiProfiler CV still running for some cohorts" + }, + { + "p": "f", + "content": "Time to apply to a new cohort", + "evidence": [ + "5 donors: time to apply" + ], + "status": "Ready" + } + ], + "candidates": [], + "details": [] + }, + { + "n": 4, + "title": "Disease signal and rare cell states", + "claim": "Open until the disease-cohort analyses are done, including LoRA on the disease cohort.", + "panels": [], + "candidates": [ + "Pairs needed: donor and disease signal", + "5 donors: LoRA vs scSemiProfiler — disease signal", + "Disease genes at every list length (CAT)" + ], + "details": [] + }, + { + "n": 5, + "title": "Datasets outside training", + "claim": "PBISC applies to cohorts it never saw, and needs their own pairs to do it well.", + "panels": [ + { + "p": "a", + "content": "Other cohorts' pairs do not replace a cohort's own", + "evidence": [ + "Pairs from other cohorts do not transfer", + "Longer training on other cohorts: gap stays" + ], + "status": "Ready" + }, + { + "p": "b", + "content": "Behçet: disease direction from real bulk", + "evidence": [ + "Consequence: disease direction flips" + ], + "status": "Adapt", + "note": "add the LoRA-3 result (C1Q-high BD > HC)" + }, + { + "p": "c", + "content": "Influenza exposure from raw clinical bulk (Randolph)", + "evidence": [], + "status": "Build", + "note": "manuscript result" + }, + { + "p": "d", + "content": "Diseases absent from the corpus vs present", + "evidence": [], + "status": "Build", + "note": "Kawasaki, SLE vs DRESS (2026-10-04)" + } + ], + "candidates": [], + "details": [] + }, + { + "n": 6, + "title": "Discovery from public bulk-only data", + "claim": "Analysis not started; to be planned.", + "panels": [], + "candidates": [ + "Next: cohorts with a paired subset (planned)" + ], + "details": [] + } + ], + "supplementary": { + "slides": [ + "No pretraining: same cohort, from scratch", + "No pretraining: other cohorts, from scratch", + "Rare states: from scratch vs pretrained" + ], + "planned": [ + { + "p": "S", + "content": "Bulk deconvolution methods (CYBERSORTx and others), for reference", + "evidence": [], + "status": "Run", + "note": "different task: PBISC generates single cells" + } + ] + }, + "appendix": [ + "Appendix: scSemiProfiler without output threshold" + ], + "discussion": [ + "Next: other tissues from a few dozen pairs" + ] +} diff --git a/paper-releases/pbisc-diffusion/updates/20261006T150739Z-deck-figure-plan-decisions/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/figure_plan_slide_text.json b/paper-releases/pbisc-diffusion/updates/20261006T150739Z-deck-figure-plan-decisions/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/figure_plan_slide_text.json new file mode 100644 index 0000000000000000000000000000000000000000..48fbcc27ad62be7205009227646bbf9a76a5dbb2 --- /dev/null +++ b/paper-releases/pbisc-diffusion/updates/20261006T150739Z-deck-figure-plan-decisions/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/figure_plan_slide_text.json @@ -0,0 +1,132 @@ +{ + "section": "figure-plan", + "scripts": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/tools/figure_plan.py" + ], + "generated_from": "tools/figure_plan.py (rerun to refresh; do not edit by hand)", + "slides": [ + { + "title": "Manuscript plan: six main figures", + "layout": "full", + "figures": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/figplan_overview.png" + ], + "keywords": [], + "footnote_short": "Nature style: short abstract and introduction, descriptive results, short discussion, detailed methods; up to 17 supplementary figures.", + "footnote": "Nature Style Manuscript Writing.pdf (paper-release-20260928/reference-papers, 2026-10-06): six main figures, up to 17 supplementary; short abstract and introduction, descriptive results, short discussion with limitations, detailed methods. User decisions of 2026-10-06: no PBISC-VAE (scSemiProfiler, a VAE-based method, is the comparison); LoRA performance first on the healthy cohort with the realism metrics, the disease-cohort comparison later; CYBERSORTx and other bulk deconvolution methods only in the Supplementary (PBISC generates single cells); Figs 4 and 6 left empty until their analyses are done. Status: Ready = slide exists and fits the panel; Adapt = slide exists; change its scope or style; Draft = new draft exists, not yet approved; Build = results exist; no slide yet; Run = needs a new analysis or experiment; Plan = not planned yet.", + "numbers": [] + }, + { + "title": "Figure 1 plan: Design and performance of PBISC", + "layout": "full", + "figures": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/figplan_fig1.png" + ], + "keywords": [], + "footnote_short": "Diffusion on each gene's value and detection, conditioned on bulk, writes realistic blood cells. 5 of 6 panels ready.", + "footnote": "Panels: a) Backbones compared: scDiffusion, SDD, PBISC-Diffusion [Ready]. b) Overview: pretraining, diffusion, generation [Draft]. c) What the bulk condition contains [Ready]. d) Pretraining corpus: assays, conditions, cell types [Ready]. e) Width and context: lower loss, same biology [Ready]. f) Gene panel: 2k, 3k, 6k, 15k [Ready]. Details in this section: PBISC-Diffusion: bulk in, cells out; Denoiser: joint value + support diffusion.", + "numbers": [] + }, + { + "title": "Figure 2 plan: Fine-tuning PBISC with LoRA", + "layout": "full", + "figures": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/figplan_fig2.png" + ], + "keywords": [], + "footnote_short": "On a healthy cohort, a few paired samples and a LoRA of minutes give cells close to the real ones. 5 of 9 panels ready.", + "footnote": "Panels: a) Real bulk differs from the training input [Ready]. b) Without fine-tuning, cells fall off the real manifold [Ready]. c) LoRA on paired bulk and cells; how it is tested [Ready]. d) How generated cells are scored (the metric suite) [Ready]. e) Healthy cohort: precision, coverage, composition [Adapt; ImMicro only; LoRA vs zero-shot and calibration]. f) Healthy cohort: classifier AUROC, MMD, E_N [Adapt; ImMicro only]. g) Full, LoRA and condition-encoder fine-tuning [Adapt; ImMicro only]. h) How many pairs are needed [Adapt; ImMicro only]. i) LoRA rank and K choices [Ready; supplementary candidate].", + "numbers": [] + }, + { + "title": "Figure 3 plan: Comparison with scSemiProfiler", + "layout": "full", + "figures": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/figplan_fig3.png" + ], + "keywords": [], + "footnote_short": "PBISC gives better cells than scSemiProfiler, a VAE-based method, for far less compute. 6 of 6 panels ready.", + "footnote": "Panels: a) Lupus benchmark against scSemiProfiler [Ready]. b) Five paired cohorts [Ready]. c) Five donors: generated cells [Ready]. d) scSemiProfiler copies its representative [Ready]. e) Same training pairs, 5-fold [Ready; scSemiProfiler CV still running for some cohorts]. f) Time to apply to a new cohort [Ready].", + "numbers": [] + }, + { + "title": "Figure 4 plan: Disease signal and rare cell states", + "layout": "full", + "figures": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/figplan_fig4.png" + ], + "keywords": [], + "footnote_short": "Open until the disease-cohort analyses are done, including LoRA on the disease cohort.", + "footnote": "Panels: Slides kept in this section until it is planned: Pairs needed: donor and disease signal; 5 donors: LoRA vs scSemiProfiler — disease signal; Disease genes at every list length (CAT).", + "numbers": [] + }, + { + "title": "Figure 5 plan: Datasets outside training", + "layout": "full", + "figures": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/figplan_fig5.png" + ], + "keywords": [], + "footnote_short": "PBISC applies to cohorts it never saw, and needs their own pairs to do it well. 1 of 4 panels ready.", + "footnote": "Panels: a) Other cohorts' pairs do not replace a cohort's own [Ready]. b) Behçet: disease direction from real bulk [Adapt; add the LoRA-3 result (C1Q-high BD > HC)]. c) Influenza exposure from raw clinical bulk (Randolph) [Build; manuscript result]. d) Diseases absent from the corpus vs present [Build; Kawasaki, SLE vs DRESS (2026-10-04)].", + "numbers": [] + }, + { + "title": "Figure 6 plan: Discovery from public bulk-only data", + "layout": "full", + "figures": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/figplan_fig6.png" + ], + "keywords": [], + "footnote_short": "Analysis not started; to be planned.", + "footnote": "Panels: Slides kept in this section until it is planned: Next: cohorts with a paired subset (planned).", + "numbers": [] + }, + { + "title": "Supplementary plan", + "layout": "full", + "figures": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/figplan_supp.png" + ], + "keywords": [], + "footnote_short": "Bulk deconvolution methods appear here only: PBISC generates single cells, a different task.", + "footnote": "Supplementary: No pretraining: same cohort, from scratch; No pretraining: other cohorts, from scratch; Rare states: from scratch vs pretrained. Planned: Bulk deconvolution methods (CYBERSORTx and others), for reference [Run; different task: PBISC generates single cells]. Appendix: Appendix: scSemiProfiler without output threshold.", + "numbers": [] + }, + { + "title": "Overview draft (Nature style)", + "layout": "full", + "figures": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/fig1b_overview_nature.png" + ], + "keywords": [ + "Each cell is learned with its own bulk", + "Diffusion on each gene's value and detection", + "One bulk in, K cells out; optional LoRA" + ], + "footnote_short": "Draft for Fig. 1b; schematic, no data. Plan: manuscript editorial/fig1_nature_redesign_plan.", + "footnote": "Candidate replacement for the engineering diagram (kept in this section as details and for Methods). Source: paper-release-20260928/manuscript/figures/fig01_architecture_nature.svg (scripts/build_figure1_architecture_nature.py).", + "numbers": [] + }, + { + "title": "Pretraining corpus", + "layout": "full", + "figures": [ + "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/fig1d_corpus.png" + ], + "keywords": [ + "41.4 M cells, 22 datasets, 3,199 donors", + "88.5% healthy donors; 23.6% cultured in vitro", + "Labels describe the corpus; never model input" + ], + "footnote_short": "Detailed version: Supplementary Fig. S1 of the manuscript.", + "footnote": "Source: paper-release-20260928/manuscript/source_data/corpus/ (build_corpus_tables.py) and figures/fig01_corpus_nature.png (build_figure1_corpus.py).", + "numbers": [ + { + "value": "41,395,184 cells; 22 datasets; 3,199 donors; healthy 88.5%; in vitro 23.6%", + "source": "manuscript/source_data/corpus/corpus_summary.json" + } + ] + } + ] +} \ No newline at end of file diff --git a/paper-releases/pbisc-diffusion/updates/20261006T150739Z-deck-figure-plan-decisions/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/slides_spec.json b/paper-releases/pbisc-diffusion/updates/20261006T150739Z-deck-figure-plan-decisions/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/slides_spec.json new file mode 100644 index 0000000000000000000000000000000000000000..1795cc87cb2a62604a59df53ffe9336b46827f22 --- /dev/null +++ b/paper-releases/pbisc-diffusion/updates/20261006T150739Z-deck-figure-plan-decisions/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/slides_spec.json @@ -0,0 +1,1857 @@ +{ + "deck": { + "output_pptx": "PBISC-Diffusion_paper_summary_20260930.pptx", + "paths": "figure paths are relative to the deck folder", + "note": "footnote = on-slide text (max 2 lines, may be a shortened copy); footnote_full = section author's full text, put in speaker notes when it differs. Generated slide text: section 9 main slides from data/sec09_slide_text.json (sec09_index), other generated slides from the file named in text_source (data/sec09_extra_slide_text.json, data/sec07_rank_sweep_slide_text.json); refresh with tools/sync_sec09_text.py after rerunning the scripts that write them." + }, + "slides": [ + { + "kind": "title", + "title": "PBISC-Diffusion: bulk-to-single-cell generation", + "subtitle": "Paper figure plan (Nature style) · updated 2026-10-06" + }, + { + "kind": "agenda", + "title": "Outline", + "items": [ + "Fig. 1 Design and performance of PBISC", + "Fig. 2 Fine-tuning PBISC with LoRA", + "Fig. 3 Comparison with scSemiProfiler", + "Fig. 4 Disease signal and rare cell states", + "Fig. 5 Datasets outside training", + "Fig. 6 Discovery from public bulk-only data", + "Supplementary", + "Appendix", + "Discussion" + ] + }, + { + "kind": "content", + "section": "plan", + "section_label": "Plan", + "text_source": { + "file": "data/figure_plan_slide_text.json", + "index": 0 + }, + "title": "Manuscript plan: six main figures", + "layout": "full", + "figures": [ + "figures/figplan_overview.png" + ], + "keywords": [], + "footnote": "Nature style: short abstract and introduction, descriptive results, short discussion, detailed methods; up to 17 supplementary figures.", + "sources": "tools/figure_plan.py; data/figure_plan.json", + "footnote_full": "Nature Style Manuscript Writing.pdf (paper-release-20260928/reference-papers, 2026-10-06): six main figures, up to 17 supplementary; short abstract and introduction, descriptive results, short discussion with limitations, detailed methods. User decisions of 2026-10-06: no PBISC-VAE (scSemiProfiler, a VAE-based method, is the comparison); LoRA performance first on the healthy cohort with the realism metrics, the disease-cohort comparison later; CYBERSORTx and other bulk deconvolution methods only in the Supplementary (PBISC generates single cells); Figs 4 and 6 left empty until their analyses are done. Status: Ready = slide exists and fits the panel; Adapt = slide exists; change its scope or style; Draft = new draft exists, not yet approved; Build = results exist; no slide yet; Run = needs a new analysis or experiment; Plan = not planned yet.", + "numbers": [] + }, + { + "kind": "divider", + "title": "Figure 1. Design and performance of PBISC" + }, + { + "kind": "content", + "section": "plan", + "section_label": "Fig. 1 plan", + "text_source": { + "file": "data/figure_plan_slide_text.json", + "index": 1 + }, + "title": "Figure 1 plan: Design and performance of PBISC", + "layout": "full", + "figures": [ + "figures/figplan_fig1.png" + ], + "keywords": [], + "footnote": "Diffusion on each gene's value and detection, conditioned on bulk, writes realistic blood cells. 5 of 6 panels ready.", + "sources": "tools/figure_plan.py; data/figure_plan.json", + "footnote_full": "Panels: a) Backbones compared: scDiffusion, SDD, PBISC-Diffusion [Ready]. b) Overview: pretraining, diffusion, generation [Draft]. c) What the bulk condition contains [Ready]. d) Pretraining corpus: assays, conditions, cell types [Ready]. e) Width and context: lower loss, same biology [Ready]. f) Gene panel: 2k, 3k, 6k, 15k [Ready]. Details in this section: PBISC-Diffusion: bulk in, cells out; Denoiser: joint value + support diffusion.", + "numbers": [] + }, + { + "kind": "content", + "section": "1", + "section_label": "Fig. 1a", + "title": "Backbone: SDD → PBISC-Diffusion", + "layout": "diagram", + "figures": [ + "figures/sec02_backbone_compare.png" + ], + "keywords": [ + "Kept from SDD: value + bit, skip-MLP", + "Changed: bulk condition, masked loss, Bernoulli read-out" + ], + "footnote": "SDD: Ostheimer et al., arXiv 2502.02448. scDiffusion: Luo et al., Bioinformatics 2024.", + "sources": "tools/sec02_backbone.py; data/sec02_backbone_compare.csv, data/sec02_slide_text.csv, data/refs/", + "footnote_full": "Luo et al., Bioinformatics 2024; Ostheimer et al., arXiv 2502.02448 (code 1d5423c). PBISC: final model f01 6k 1x, part 16. σ = sigmoid; Bernoulli is the final read-out; the logit > 0 cut is also reported.", + "refs": [ + "Ostheimer 2025, arXiv", + "Luo 2024, Bioinformatics" + ], + "refs_full": [ + "Ostheimer et al. Sparse Data Diffusion for scientific simulations in biology and physics. arXiv:2502.02448 (2025)", + "Luo et al. scDiffusion: conditional generation of high-quality single-cell data using diffusion model. Bioinformatics 40(9):btae518 (2024)" + ] + }, + { + "kind": "content", + "section": "plan", + "section_label": "Fig. 1b", + "text_source": { + "file": "data/figure_plan_slide_text.json", + "index": 8 + }, + "title": "Overview draft (Nature style)", + "layout": "full", + "figures": [ + "figures/fig1b_overview_nature.png" + ], + "keywords": [ + "Each cell is learned with its own bulk", + "Diffusion on each gene's value and detection", + "One bulk in, K cells out; optional LoRA" + ], + "footnote": "Draft for Fig. 1b; schematic, no data. Plan: manuscript editorial/fig1_nature_redesign_plan.", + "sources": "tools/figure_plan.py; manuscript figures/fig01_architecture_nature.svg", + "footnote_full": "Candidate replacement for the engineering diagram (kept in this section as details and for Methods). Source: paper-release-20260928/manuscript/figures/fig01_architecture_nature.svg (scripts/build_figure1_architecture_nature.py).", + "numbers": [] + }, + { + "kind": "content", + "section": "1", + "section_label": "Fig. 1c", + "title": "Bulk conditioning", + "layout": "diagram", + "figures": [ + "figures/sec02_bulk_conditioning.png" + ], + "keywords": [ + "Same bulk condition for every generated cell", + "Final: no cell types, proportions, sum constraint", + "18,006 → 512 context into every block" + ], + "footnote": "Final model f01 6k 1x. Bar heights schematic.", + "sources": "tools/sec02_backbone.py; data/sec02_bulk_conditioning.csv, data/sec02_slide_text.csv", + "footnote_full": "Final model f01 6k 1x, part 16. Bar heights schematic." + }, + { + "kind": "content", + "section": "plan", + "section_label": "Fig. 1d", + "text_source": { + "file": "data/figure_plan_slide_text.json", + "index": 9 + }, + "title": "Pretraining corpus", + "layout": "full", + "figures": [ + "figures/fig1d_corpus.png" + ], + "keywords": [ + "41.4 M cells, 22 datasets, 3,199 donors", + "88.5% healthy donors; 23.6% cultured in vitro", + "Labels describe the corpus; never model input" + ], + "footnote": "Detailed version: Supplementary Fig. S1 of the manuscript.", + "sources": "tools/figure_plan.py; manuscript figures/fig01_corpus_nature.png, source_data/corpus/", + "footnote_full": "Source: paper-release-20260928/manuscript/source_data/corpus/ (build_corpus_tables.py) and figures/fig01_corpus_nature.png (build_figure1_corpus.py).", + "numbers": [ + { + "value": "41,395,184 cells; 22 datasets; 3,199 donors; healthy 88.5%; in vitro 23.6%", + "source": "manuscript/source_data/corpus/corpus_summary.json" + } + ] + }, + { + "kind": "content", + "section": "1", + "section_label": "Fig. 1e", + "title": "Width & context ablation: loss saturation", + "layout": "full", + "figures": [ + "figures/sec03_width_loss.png" + ], + "keywords": [ + "Trunk x4: val loss −31% (0.197 → 0.136)", + "ctx x4, prop hint: loss = 1x", + "1x flat E8; trunk x4 still falling" + ], + "footnote": "Old 6k panel (1% floor); one validation set. 16x: memory-floor stop after part 2.", + "say": "All arms share the 6k panel (1% floor), corpus, seed and schedule; validation = 192 recipes x 64 cells. ctx x4, both x4 and 16x were read from server_2 snapshots.", + "sources": "tools/sec03_width_loss.py; data/sec03_width_loss.csv, data/sec03_width_saturation.csv, data/sec03_width_loss_headline.csv" + }, + { + "kind": "content", + "section": "1", + "section_label": "Fig. 1e", + "title": "Width & context ablation: benchmark", + "layout": "full", + "figures": [ + "figures/sec03_width_bench.png" + ], + "keywords": [ + "trunk x4: loss −31%, no DEG gain", + "Prop hint: oracle = predicted, L1 worse", + "1x kept: ties ctx x4, 54M vs 102M" + ], + "footnote": "SLE, 42 donors, 2,980 genes; old 6k panel. ctx x4: better L1 (0.161 vs 0.197).", + "say": "Bars = mean of parts 14-16 with the cut p > 0.5; diamonds = Bernoulli read-out (4 width arms only). Prop hint = 9 L1 cell-type proportions; older panel, not the final model.", + "sources": "tools/sec03_width_bench.py; data/sec03_width_bench.csv, data/sec03_width_bench_summary.csv, data/sec03_width_bench_headline.csv" + }, + { + "kind": "content", + "section": "1", + "section_label": "Fig. 1f", + "title": "Gene context: 2k / 3k / 6k / 15k", + "layout": "full", + "figures": [ + "figures/sec03_panel.png" + ], + "keywords": [ + "6k ≈ 3k: 3k r, DEG inside 6k seed range", + "2k best L1; 15k worst r, L1", + "6k kept: first panel with CLEC9A (cDC1)" + ], + "footnote": "SLE, 42 donors, 1,006 genes in every panel. 6k: 5 seeds; base seed best of 5 on r and DEG.", + "say": "Runs f01_{2k,3k,6k,p15k}_1x, part 16, Bernoulli, EMA, DDIM 50. Different gene axis from the previous slide, so the two slides cannot be compared. Loss curves are divided by each panel's own part-16 value.", + "sources": "tools/sec03_panel.py; data/sec03_panel_bench.csv, data/sec03_panel_seeds.csv, data/sec03_panel_loss.csv, data/sec03_panel_headline.csv" + }, + { + "kind": "content", + "section": "1", + "section_label": "Fig. 1 details", + "title": "PBISC-Diffusion: bulk in, cells out", + "layout": "full", + "figures": [ + "figures/sec01_slide_a.png" + ], + "keywords": [ + "Train pair: K-cell pseudobulk → each member cell" + ], + "footnote": "Fig. 1b redraw. Final model corpus: 33.1M train cells; K = 1–4,000 cells per pseudobulk.", + "footnote_full": "Top: deck redraw of manuscript Fig. 1b (larger labels; cells differ by x_T and the Bernoulli support draw, Methods 'Support read-out'). Bottom: corpus of final run f01_6k_1x_seed20260627 (index fingerprint matches). Cells per dataset = base + overlay rows (Model 7 e2_full_source/e2_full_index manifests); 33.1M cells in the donor-level 80/10/10 train split. Registered K grid = 8,192 recipes/epoch. Plus 3,658 top-up recipes over 16 epochs (196-283 per epoch, K 1-4,000), so every train cell is a target once (ALLCELL_DATA_CONTRACT.json coverage_after_topup).", + "sources": "tools/sec01_overview.py; data/sec01_model_numbers.csv, data/sec01_corpus_datasets.csv, data/sec01_recipe_k.csv", + "refs": [ + "Ostheimer 2025, arXiv", + "Song 2021, ICLR" + ], + "refs_full": [ + "Ostheimer et al. Sparse Data Diffusion for scientific simulations in biology and physics. arXiv:2502.02448 (2025)", + "Song, Meng & Ermon. Denoising Diffusion Implicit Models. ICLR 2021" + ] + }, + { + "kind": "content", + "section": "1", + "section_label": "Fig. 1 details", + "title": "Denoiser: joint value + support diffusion", + "layout": "diagram", + "figures": [ + "figures/sec01_fig1a_denoiser.png" + ], + "keywords": [ + "6,000 genes · 54.2M params · Bernoulli support draw" + ], + "footnote": "Fig. 1a redraw. Final model f01 6k 1x, epoch 16, EMA; 50 DDIM steps.", + "footnote_full": "Deck redraw of manuscript Fig. 1a (larger labels, legend dropped). Purple dashed = soft self-conditioning feedback; gold dashed = context to all blocks. Final model f01_6k_1x_seed20260627, checkpoint_epoch_016, EMA; 7 blocks 1024-1024-512-256-512-1024-1024 (effective_config.json, RUN_STATUS.json). Sampling: 50 DDIM steps, η = 0, drawn support read-out, keep gene if u < σ(logit) (methods.tex l.58, l.99). 16 epochs = 16 disjoint 8,192-recipe slots = one pass over the 131,072-recipe pool (manuscript: part 16).", + "sources": "tools/sec01_overview.py; data/sec01_model_numbers.csv, data/sec01_fig1_label_pt.csv", + "refs": [ + "Ostheimer 2025, arXiv" + ], + "refs_full": [ + "Ostheimer et al. Sparse Data Diffusion for scientific simulations in biology and physics. arXiv:2502.02448 (2025)" + ] + }, + { + "kind": "divider", + "title": "Figure 2. Fine-tuning PBISC with LoRA" + }, + { + "kind": "content", + "section": "plan", + "section_label": "Fig. 2 plan", + "text_source": { + "file": "data/figure_plan_slide_text.json", + "index": 2 + }, + "title": "Figure 2 plan: Fine-tuning PBISC with LoRA", + "layout": "full", + "figures": [ + "figures/figplan_fig2.png" + ], + "keywords": [], + "footnote": "On a healthy cohort, a few paired samples and a LoRA of minutes give cells close to the real ones. 5 of 9 panels ready.", + "sources": "tools/figure_plan.py; data/figure_plan.json", + "footnote_full": "Panels: a) Real bulk differs from the training input [Ready]. b) Without fine-tuning, cells fall off the real manifold [Ready]. c) LoRA on paired bulk and cells; how it is tested [Ready]. d) How generated cells are scored (the metric suite) [Ready]. e) Healthy cohort: precision, coverage, composition [Adapt; ImMicro only; LoRA vs zero-shot and calibration]. f) Healthy cohort: classifier AUROC, MMD, E_N [Adapt; ImMicro only]. g) Full, LoRA and condition-encoder fine-tuning [Adapt; ImMicro only]. h) How many pairs are needed [Adapt; ImMicro only]. i) LoRA rank and K choices [Ready; supplementary candidate].", + "numbers": [] + }, + { + "kind": "content", + "section": "2", + "section_label": "Fig. 2a", + "title": "Real bulk ≠ pseudobulk (model input)", + "layout": "figure_keywords", + "figures": [ + "figures/sec05_condition_space.png" + ], + "keywords": [ + "Real bulk: 2.46 D from own pseudobulk", + "95% of squared offset is one shift", + "Shift removed (diagnostic): own donor 7.5% → 75.7%" + ], + "footnote": "ImMicro, 107 healthy donors. Model input space (log1p CPM). D = median donor-to-donor distance.", + "say": "The two pseudobulk clusters on PC1 are sc pools DB1-3 vs DB4-7, so the pairing lines cross mainly because of sc batch. Shift removal subtracts the mean offset of the other 106 donors; a bulk-only deployment would not have them.", + "sources": "tools/sec05_condition_space.py; data/sec05_condition_space_numbers.csv, data/sec05_condition_space_coords.csv", + "footnote_full": "ImMicro, 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 R² 0.96 on PC1)." + }, + { + "kind": "content", + "section": "2", + "section_label": "Fig. 2a", + "title": "Same shift in three paired cohorts", + "layout": "full", + "figures": [ + "figures/sec05_shift_cohorts.png" + ], + "keywords": [ + "Shared shift in all 3 cohorts" + ], + "footnote": "Model inputs only, no generation. PCA fitted per cohort on sc pseudobulks.", + "sources": "tools/sec05_shift_cohorts.py; data/sec05_shift_cohorts.csv", + "footnote_full": "Model inputs only, no generation. PCA is fitted on each cohort's sc pseudobulks (Arunachalam: 11 PCs, n = 12). D uses same-condition pairs only for Randolph and Arunachalam. Shift removal is a leave-one-out diagnostic." + }, + { + "kind": "content", + "section": "2", + "section_label": "Fig. 2b", + "title": "Consequence: cells off-manifold", + "layout": "two_figures", + "figures": [ + "figures/sec05_cells_umap.png", + "figures/sec05_cells_metrics.png" + ], + "keywords": [ + "Monocytes: raw 30.6% vs real 16.7%", + "Same noise: 19% of T cells become monocytes", + "Precision: raw 0.36 vs pseudobulk 0.92" + ], + "footnote": "ImMicro, 107 donors; same starting noise for every input. Coverage: raw 0.08, sc pseudobulk 0.51.", + "footnote_full": "ImMicro, 107 donors. Final model f01 6k 1× 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.", + "sources": "tools/sec05_cells_offmanifold.py; data/sec05_cells_metrics.csv, data/sec05_cells_umap_coords.csv", + "refs": [ + "Kynkäänniemi 2019, NeurIPS", + "Naeem 2020, ICML" + ], + "refs_full": [ + "Kynkäänniemi et al. Improved precision and recall metric for assessing generative models. NeurIPS 2019", + "Naeem et al. Reliable fidelity and diversity metrics for generative models. ICML 2020" + ] + }, + { + "kind": "content", + "section": "2", + "section_label": "Fig. 2c", + "title": "Fine-tuning with paired bulk + sc", + "layout": "diagram", + "figures": [ + "figures/sec06_ft_method.png" + ], + "keywords": [ + "LoRA 2–14 min vs pretraining 15.9 h" + ], + "footnote": "Base: final model f01 6k 1x E16. Times from 40 LoRA runs on this host.", + "sources": "tools/sec06_finetune_method.py; data/sec06_ft_method.csv", + "footnote_full": "Receipts: immicro/fold0/real_lora and ye2026_sepsis/fewshot_n5/draw0/lora RUN.json. Base: f01_6k_1x E16. Times: 40 local LoRA runs.", + "refs": [ + "Hu 2022, ICLR" + ], + "refs_full": [ + "Hu et al. LoRA: Low-Rank Adaptation of Large Language Models. ICLR 2022" + ] + }, + { + "kind": "content", + "section": "2", + "section_label": "Fig. 2c", + "title": "Evaluation design", + "layout": "full", + "figures": [ + "figures/sec06_eval_design.png" + ], + "keywords": [ + "Donor-level splits; test donors never trained", + "5-fold: other folds · few: 5 donors", + "Giroux few-pair trains 12–14 of 24 pairs" + ], + "footnote": "Bars: Ye 2026 sepsis. QC-failed pairs excluded.", + "sources": "tools/sec06_finetune_method.py; data/sec06_eval_design.csv, data/sec06_eval_design_sepsis_roles.csv", + "footnote_full": "Counts: data//folds.csv and fewshot_n5/DRAWS.json; QC-failed pairs (fold −1) excluded. Bars: Ye 2026 sepsis." + }, + { + "kind": "content", + "section": "2", + "title": "How we score generated cells", + "layout": "full", + "figures": [ + "figures/sec04_metrics_overview.png" + ], + "keywords": [ + "Cells, composition and disease signal, separately", + "Every metric next to its control or ceiling", + "Same donors and reference cells for every method" + ], + "footnote": "Definitions follow the cited papers; scorers: tools/score_ft.py, extra_metrics.py, contrast_extra.py.", + "sources": "tools/sec04_metrics_overview.py; data/sec04_metrics_overview.csv", + "refs_full": [ + "Kynkäänniemi et al. Improved precision and recall metric for assessing generative models. NeurIPS 2019", + "Naeem et al. Reliable fidelity and diversity metrics for generative models. ICML 2020", + "Lopez-Paz & Oquab. Revisiting classifier two-sample tests. ICLR 2017", + "Gretton et al. A kernel two-sample test. JMLR 13:723-773 (2012)", + "Wang 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", + "Roohani, Huang & Leskovec. Predicting transcriptional outcomes of novel multigene perturbations with GEARS. Nat Biotechnol 42:927-935 (2024). doi:10.1038/s41587-023-01905-6", + "Viñas Torné 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", + "Li et al. A systematic comparison of single-cell perturbation response prediction models. Sci Adv 12:eaed3414 (2026). doi:10.1126/sciadv.aed3414", + "Squair et al. Confronting false discoveries in single-cell differential expression. Nat Commun 12:5692 (2021). doi:10.1038/s41467-021-25960-2", + "Law, Chen, Shi & Smyth. voom: precision weights unlock linear model analysis tools for RNA-seq read counts. Genome Biol 15:R29 (2014)", + "Irizarry et al. Multiple-laboratory comparison of microarray platforms. Nat Methods 2:345-350 (2005). doi:10.1038/nmeth756" + ], + "section_label": "Fig. 2d" + }, + { + "kind": "content", + "section": "2", + "section_label": "Fig. 2e", + "sec09_index": 3, + "title": "5-fold CV: train on the other 4 folds", + "layout": "full", + "figures": [ + "figures/sec09_fivefold.png" + ], + "keywords": [ + "LoRA tops coverage 5/5, logFC r 4/4", + "Ji whole blood: LoRA 0.86, others 0.06–0.17", + "Broadcast ahead: Sepsis, Giroux" + ], + "footnote": "Donor-level 5-fold CV. Giroux: sex-linked genes excluded (severity confounded with sex: 4F/0M vs 0F/5M).", + "footnote_full": "Donor-level 5-fold CV; each pair scored once, by the fold model that never trained on its donor (one pooled call per cohort). LoRA r=16 on the real bulk of the other 4 folds' training donors, early stop on inner-val donors. Zero-shot = base model without fine-tuning, fed the test pair's real bulk as measured; zero-shot + calibration = the same model fed gene-wise calibrated bulk (alpha fitted on the fold's training donors); sc-only fine-tune (+ calibration) = full fine-tune on training donors' sc pseudobulk, fed calibrated bulk. scSemiProfiler, 5-fold CV = scSemiProfiler with every training donor of the fold (inner-val included) as representatives, scored for the same pairs and reference cells (ImMicro, Ji). Base f01 6k·1x·E16. LoRA best of 5 arms: coverage 5/5, logFC r 4/4, top-50 DEG 4/4. Broadcast = raw real bulk. Sepsis DEG from pooled_deg. Giroux: moderate vs mild subjects are 4F/0M vs 0F/5M (scored; all paired subjects 4F/1M vs 0F/5M; sex from bulk chrY expression = sc = GEO), so its disease metrics (logFC r, partial r, top-50 DEG, broadcast) exclude 196 sex-linked genes (43 chrY non-PAR + 153 X-inactivation escape / variable-escape genes of Tukiainen et al. 2017; PAR genes kept), 65 of them scoring genes; same labels and cells as before (scores_nosex/giroux2022/). Scores: scores//pooled/summary.json.", + "sources": "tools/sec09_paired_validation.py; data/sec09_fivefold.csv, data/sec09_slide_text.json", + "numbers": [ + { + "value": "LoRA best of 5 arms: coverage 5/5, logFC r 4/4, top-50 DEG 4/4", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//pooled/summary.json (Sepsis DEG: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/ye2026_sepsis/pooled_deg/summary.json; Giroux: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_nosex/giroux2022/pooled/summary.json)" + }, + { + "value": "Ji coverage: zero-shot 0.055, zero-shot + calibration 0.106, sc-only (+ calibration) 0.113, LoRA 0.857", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/ji2025_chili/pooled/summary.json" + }, + { + "value": "Giroux logFC r: broadcast 0.575 vs LoRA 0.461; DEG broadcast 5 vs LoRA 5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_nosex/giroux2022/pooled/summary.json" + }, + { + "value": "LoRA logFC r vs broadcast: Sepsis 0.527 vs 0.316, Imperial 0.819 vs 0.661, Giroux 0.461 vs 0.575, Ji 0.479 vs 0.240", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//pooled/summary.json (Sepsis DEG: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/ye2026_sepsis/pooled_deg/summary.json; Giroux: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_nosex/giroux2022/pooled/summary.json)" + }, + { + "value": "coverage LoRA / zero-shot + calibration / zero-shot / sc-only (+ calibration): ImMicro 0.784 / 0.527 / 0.192 / 0.623, Sepsis 0.829 / 0.431 / 0.067 / 0.726, Imperial 0.850 / 0.639 / 0.112 / 0.555, Giroux 0.928 / 0.906 / 0.281 / 0.908, Ji 0.857 / 0.106 / 0.055 / 0.113", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//pooled/summary.json (Sepsis DEG: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/ye2026_sepsis/pooled_deg/summary.json; Giroux: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_nosex/giroux2022/pooled/summary.json)" + }, + { + "value": "logFC r LoRA / zero-shot + calibration / zero-shot / sc-only (+ calibration): Sepsis 0.527 / 0.123 / -0.012 / 0.296, Imperial 0.819 / 0.679 / 0.644 / 0.808, Giroux 0.461 / 0.269 / 0.334 / 0.344, Ji 0.479 / 0.142 / 0.084 / 0.245", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//pooled/summary.json (Sepsis DEG: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/ye2026_sepsis/pooled_deg/summary.json; Giroux: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_nosex/giroux2022/pooled/summary.json)" + }, + { + "value": "top-50 DEG LoRA / zero-shot + calibration / zero-shot / sc-only (+ calibration): Sepsis 23 / 11 / 9 / 15, Imperial 143 / 111 / 106 / 129, Giroux 5 / 0 / 1 / 0, Ji 15 / 0 / 0 / 0", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//pooled/summary.json (Sepsis DEG: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/ye2026_sepsis/pooled_deg/summary.json; Giroux: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_nosex/giroux2022/pooled/summary.json)" + }, + { + "value": "scSemiProfiler, 5-fold CV: coverage / logFC r / top-50 DEG (n/a = no contrast or rescore pending): ImMicro 0.038 / n/a / n/a, Ji 0.167 / 0.273 / 0", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/immicro/scsemi_cv_20261002/pooled/summary.json; /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/ji2025_chili/scsemi_cv_20261002/pooled/summary.json" + }, + { + "value": "Pooled n (pairs / donors): ImMicro 107/107, Sepsis 82/82, Imperial 157/57, Giroux 24/9, Ji 33/20", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//pooled/summary.json (n_donors); /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data//fewshot_n5/DRAWS.json" + }, + { + "value": "early stopping: About 10 % of each fold's training donors are inner-validation, used only for early stopping", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/README.md" + } + ], + "refs": [ + "Viñas Torné 2026, Nat Biotechnol", + "Li 2026, Sci Adv", + "Law 2014, Genome Biol" + ], + "refs_full": [ + "Viñas Torné 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", + "Li et al. A systematic comparison of single-cell perturbation response prediction models. Sci Adv 12:eaed3414 (2026). doi:10.1126/sciadv.aed3414", + "Law, Chen, Shi & Smyth. voom: precision weights unlock linear model analysis tools for RNA-seq read counts. Genome Biol 15:R29 (2014)" + ] + }, + { + "kind": "content", + "section": "2", + "section_label": "Fig. 2f", + "text_source": { + "file": "data/sec09_extra_slide_text.json", + "index": 1 + }, + "title": "5-fold: realism beyond precision/coverage", + "layout": "full", + "figures": [ + "figures/sec09_extra_fivefold.png" + ], + "keywords": [ + "Every metric: LoRA first in 5/5 cohorts", + "Ji whole blood: E_N 1.9 vs 7.1–10.1", + "Still separable: LoRA AUC 0.87–0.94" + ], + "footnote": "Donor-level 5-fold CV, one pooled call per cohort. E_N: scSemiProfiler's own error metric.", + "footnote_full": "Donor-level 5-fold CV; each pair scored once by the fold model that never trained on its donor (one pooled call per cohort, all cross-fitted pairs). Arms as on the 5-fold slide. Metric definitions as on the 5-donor realism slide. LoRA first among the 4 arms in all 5 cohorts on 12/12 headline metrics (kNN and RF AUC, 1-NN accuracy, MMD² at σ = m and summed, E_N all cells, precision and coverage at k = 3, 5, 10). LoRA kNN AUC ImMicro 0.941, Sepsis 0.886, Imperial 0.866, Giroux 0.881, Ji 0.908; real vs real ImMicro 0.505, Sepsis 0.506, Imperial 0.496, Giroux 0.501, Ji 0.494. E_N LoRA vs other arms: ImMicro 1.56 vs 1.77–4.83; Sepsis 1.20 vs 1.40–5.29; Imperial 1.13 vs 1.19–4.05; Giroux 1.70 vs 2.24–4.48; Ji 1.95 vs 7.08–10.13. scSemiProfiler 5-fold CV not available yet (no extra-metric unit with set scsemi__raw). Scores: scores_extra//pooled/.", + "sources": "tools/sec09_extra_metrics.py; data/sec09_extra_fivefold.csv, data/sec09_extra_scsemi_status.csv; plans/paired_finetune_intradataset_20260927/scores_extra/", + "numbers": [ + { + "value": "LoRA first among 4 arms in all 5 cohorts on 12/12 metrics", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_extra//pooled/{metrics,en}.csv" + }, + { + "value": "Ji E_N: LoRA 1.948; Zero-shot, real bulk 10.125, Zero-shot + calibration 7.076, sc-only fine-tune (+ calibration) 7.215", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_extra//pooled/{metrics,en}.csv" + }, + { + "value": "LoRA kNN AUC (pooled): ImMicro 0.941, Sepsis 0.886, Imperial 0.866, Giroux 0.881, Ji 0.908", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_extra//pooled/{metrics,en}.csv" + }, + { + "value": "scSemiProfiler, 5-fold CV: not available", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/data/sec09_extra_scsemi_status.csv" + } + ], + "refs": [ + "Lopez-Paz 2017, ICLR", + "Gretton 2012, JMLR" + ], + "refs_full": [ + "Lopez-Paz & Oquab. Revisiting classifier two-sample tests. ICLR 2017", + "Gretton et al. A kernel two-sample test. JMLR 13:723-773 (2012)" + ] + }, + { + "kind": "content", + "section": "2", + "section_label": "Fig. 2g", + "title": "Full vs LoRA vs cond.-encoder-only", + "layout": "full", + "figures": [ + "figures/sec07_ft_modes.png" + ], + "keywords": [ + "LoRA ≈ full FT, 42× fewer params", + "Cond.-encoder-only ≈ ZS + cal. on ImMicro", + "LoRA beats both on coverage" + ], + "footnote": "Fold 0, cond.-enc. vs ZS + cal.: coverage 0.73 vs 0.73 / 0.74 vs 0.63, L1 p 0.06 / 0.93 (ImMicro / sepsis).", + "sources": "tools/sec07_ft_modes.py; data/sec07_ft_modes_accuracy.csv, data/sec07_ft_modes_compute.csv, data/sec07_ft_modes_wilcoxon.csv", + "footnote_full": "Base f01_6k_1x E16. ImMicro / sepsis; paired Wilcoxon on donor values, uncorrected. Full vs LoRA L1 p 0.25 / 0.71 (pooled); 42× fewer trainable params (1.3 M). Zero-shot + calibration: α_g = mean CPM(pseudobulk_g) / mean CPM(bulk_g) over the fold's 85–86 / 65–66 paired training donors (train + inner-val); bulk CPM × α, renormalised. The sc-only FT never saw real bulk, so it is fed calibrated bulk. Raw → calibrated zero-shot (pooled): L1 0.45 → 0.20 / 0.51 → 0.40, coverage 0.19 → 0.53 / 0.07 → 0.43, sepsis logFC r −0.01 → 0.12. LoRA vs ZS + cal. (pooled): coverage 0.78 vs 0.53 / 0.83 vs 0.43 (p 7e-19 / 4e-15); L1 0.16 vs 0.20 (p 2e-7) / 0.36 vs 0.40 (p 0.19); logFC r 0.53 vs 0.12. Fold 0, cond.-enc. vs ZS + cal.: L1 0.20 vs 0.17 (p 0.06) / 0.38 vs 0.39 (p 0.93), coverage 0.73 vs 0.73 (p 0.68) / 0.74 vs 0.63 (p 7e-4). Fold 0, LoRA: coverage 0.91 / 0.86; vs cond.-enc. p 4e-5 / 1e-3, vs ZS + cal. p 6e-5 / 3e-4; L1 vs cond.-enc. p 3e-5 / 0.52, vs ZS + cal. p 0.29 / 0.96. sc-only FT (+ cal.), pooled: coverage 0.62 / 0.73, logFC r 0.30. Open = fold-0 pilot; compare same marker only. Per-step time: runs early-stop differently.", + "refs": [ + "Hu 2022, ICLR" + ], + "refs_full": [ + "Hu et al. LoRA: Low-Rank Adaptation of Large Language Models. ICLR 2022" + ] + }, + { + "kind": "content", + "section": "2", + "section_label": "Fig. 2h", + "title": "Pairs needed: composition, coverage, training steps", + "layout": "full", + "figures": [ + "figures/sec08_pairs_realism.png" + ], + "keywords": [ + "No pairs: coverage 0.19 / 0.07; length fix 0.50 / 0.40", + "Calibration flat 0.53 / 0.43; scSemi 0.04", + "LoRA with 5 pairs: coverage 0.84 / 0.77" + ], + "footnote": "ImMicro / sepsis, 5-fold CV (107 / 82 donors). 0 = no pairs; squares = calibration fitted on the same n pairs.", + "footnote_full": "ImMicro / sepsis, donor-level 5-fold CV, pooled scoring (107 / 82 donors). x = LoRA training pairs per fold; each run also early-stops on 7–8 paired inner-val donors; the last point is every training donor of the fold (78–79 / 58–59; the other four folds minus inner-val, never the test fold). x = 0 uses no paired data: the base model fed the real bulk as measured (pale) and after exon-length correction (sand: counts divided by Ensembl union-exon kb before the CPM step; scores//zeroshot_tpm_20261002/pooled). Squares: zero-shot + gene-wise calibration with alpha fitted on exactly the pairs that LoRA point saw (cal_budget_20261001; the last square uses every training pair of the fold). Coverage ImMicro / sepsis: as measured 0.19 / 0.07, length-corrected 0.50 / 0.40, calibration 0.52–0.53 / 0.43–0.44 at every n, LoRA 5 pairs 0.84 / 0.77, every training donor 0.78 / 0.83. Composition L1: length correction does not fix it (ImMicro 0.45 → 0.42, sepsis 0.51 → 0.56). Diamonds (ImMicro; sepsis still running): scSemiProfiler with exactly LoRA's n training + inner-val pairs as representatives (scores//scsemi_budget_20261002/pooled), and with every training donor at the last point (scsemi_cv_20261002/pooled); same donors and reference as the LoRA points.", + "say": "Each LoRA run also early-stops on 7 inner-validation pairs (8 in 2 of the 5 sepsis folds).", + "sources": "tools/sec08_pairs_needed.py; data/sec08_pairs_realism.csv, data/sec08_selected_steps.csv, data/sec08_slide_numbers.csv", + "refs": [ + "Kynkäänniemi 2019, NeurIPS", + "Naeem 2020, ICML" + ], + "refs_full": [ + "Kynkäänniemi et al. Improved precision and recall metric for assessing generative models. NeurIPS 2019", + "Naeem et al. Reliable fidelity and diversity metrics for generative models. ICML 2020" + ] + }, + { + "kind": "content", + "section": "2", + "section_label": "Fig. 2i", + "text_source": { + "file": "data/sec07_rank_sweep_slide_text.json", + "index": 0 + }, + "title": "LoRA rank: r = 4, 16, 64", + "layout": "full", + "figures": [ + "figures/sec07_rank_sweep.png" + ], + "keywords": [ + "ImMicro r64 vs r16: coverage 0.78 → 0.82", + "ImMicro r4 vs r16: L1 0.162 → 0.186", + "Kept r16; r64 precision 0.92 → 0.90" + ], + "footnote": "Sepsis, 58–59 pairs/fold: r64 logFC r 0.53 → 0.55, DEG 23 → 26; r4 0.48, DEG 18. Pooled 5-fold scores.", + "footnote_full": "LoRA fine-tune of the final model (f01 6k·1x·E16) with rank r and alpha = r, same recipe as the deck's r = 16 runs (early stop on 7–8 inner-val pairs, 20k step cap); n = training pairs, nested subsets; donor-level 5-fold CV, each pair scored once by its fold model (pooled call). Trainable parameters r4 324,056, r16 1,296,224, r64 5,184,896. ImMicro n = 5: r4: L1 0.162, P/C 0.937 / 0.789 (L1 vs r16 -0.009, Wilcoxon p 0.00088, 107 donors); r16: L1 0.171, P/C 0.946 / 0.839; r64: L1 0.177, P/C 0.946 / 0.847 (L1 vs r16 +0.005, Wilcoxon p 0.0076, 107 donors) | ImMicro n = 20: r4: L1 0.171, P/C 0.925 / 0.715 (L1 vs r16 -0.001, Wilcoxon p 0.32, 107 donors); r16: L1 0.172, P/C 0.920 / 0.755; r64: L1 0.172, P/C 0.905 / 0.794 (L1 vs r16 -0.001, Wilcoxon p 0.66, 107 donors) | ImMicro n = 78–79 (every training donor of the fold): r4: L1 0.186, P/C 0.933 / 0.712 (L1 vs r16 +0.024, Wilcoxon p 9.8e-09, 107 donors); r16: L1 0.162, P/C 0.924 / 0.784; r64: L1 0.161, P/C 0.904 / 0.816 (L1 vs r16 -0.001, Wilcoxon p 0.89, 107 donors) | Sepsis n = 5: r4: L1 0.426, P/C 0.907 / 0.742, logFC r 0.27, DEG 2 (L1 vs r16 -0.002, Wilcoxon p 0.97, 82 donors); r16: L1 0.428, P/C 0.911 / 0.767, logFC r 0.29, DEG 5; r64: L1 0.429, P/C 0.906 / 0.766, logFC r 0.31, DEG 1 (L1 vs r16 +0.001, Wilcoxon p 0.66, 82 donors) | Sepsis n = 20: r4: L1 0.378, P/C 0.929 / 0.769, logFC r 0.31, DEG 14 (L1 vs r16 -0.003, Wilcoxon p 0.58, 82 donors); r16: L1 0.380, P/C 0.927 / 0.807, logFC r 0.37, DEG 18; r64: L1 0.376, P/C 0.911 / 0.808, logFC r 0.43, DEG 18 (L1 vs r16 -0.004, Wilcoxon p 0.046, 82 donors) | Sepsis n = 58–59 (every training donor of the fold): r4: L1 0.363, P/C 0.923 / 0.789, logFC r 0.48, DEG 18 (L1 vs r16 +0.002, Wilcoxon p 0.77, 82 donors); r16: L1 0.362, P/C 0.926 / 0.829, logFC r 0.53, DEG 23; r64: L1 0.367, P/C 0.907 / 0.824, logFC r 0.55, DEG 26 (L1 vs r16 +0.005, Wilcoxon p 0.66, 82 donors). r64 vs r16 brings small gains in coverage and disease signal but slightly lower precision and no composition change; r4 is worse with many pairs; at n = 5 the ranks are mixed. r = 16 is kept. Scores: scores//lora_rank_20261001/pooled (r4, r64) and scores//pooled{,_deg} (r16); table: scores//rankcal_20261001.md.", + "sources": "tools/sec07_rank_sweep.py; data/sec07_rank_sweep.csv, data/sec07_rank_sweep_runs.csv; plans/paired_finetune_intradataset_20260927/scores//rankcal_20261001.md", + "numbers": [ + { + "value": "ImMicro 78–79 pairs/fold coverage r16 0.784 -> r64 0.816; precision 0.924 -> 0.904; L1 r16 0.162, r4 0.186 (Wilcoxon p 9.8e-09)", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/immicro/pooled/summary.json; /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/immicro/lora_rank_20261001/pooled/summary.json, per_donor.csv" + }, + { + "value": "Sepsis 58–59 pairs/fold logFC r r16 0.527, r64 0.553, r4 0.476; DEG 23 / 26 / 18", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/ye2026_sepsis/pooled_deg/summary.json, contrast.csv; /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/ye2026_sepsis/lora_rank_20261001/pooled/summary.json, contrast.csv" + }, + { + "value": "trainable parameters: r4 324056, r16 1296224, r64 5184896", + "source": "runs//fold//RUN.json (trainable.n_trainable)" + } + ], + "refs": [ + "Hu 2022, ICLR" + ], + "refs_full": [ + "Hu et al. LoRA: Low-Rank Adaptation of Large Language Models. ICLR 2022" + ] + }, + { + "kind": "content", + "section": "2", + "section_label": "Fig. 2i", + "title": "K diversification: no gain", + "layout": "figure_keywords", + "figures": [ + "figures/sec07_k_aug.png" + ], + "keywords": [ + "No gain; keep single K 1000", + "Multi-K slightly worse on ImMicro (p 0.04)", + "Inner-val loss Δ ≤ 1.2e-4" + ], + "footnote": "LoRA r16, 5-fold pooled. p = paired Wilcoxon on donor L1, uncorrected.", + "sources": "tools/sec07_k_aug.py; data/sec07_k_aug.csv, data/sec07_k_aug_wilcoxon.csv, data/sec07_k_aug_inner_val_delta.csv", + "footnote_full": "LoRA r16 on f01_6k_1x E16, 5-fold pooled. p: paired Wilcoxon, donor L1 vs plain LoRA, uncorrected. Recipes arm: K 1000 real bulk + ½-batch pretraining-style pseudobulk recipes made from the training donors' own cells (recipe K drawn from the pretraining grid 250–4,000; RUN.json augmentation); the real bulk itself is not multi-K." + }, + { + "kind": "divider", + "title": "Figure 3. Comparison with scSemiProfiler" + }, + { + "kind": "content", + "section": "plan", + "section_label": "Fig. 3 plan", + "text_source": { + "file": "data/figure_plan_slide_text.json", + "index": 3 + }, + "title": "Figure 3 plan: Comparison with scSemiProfiler", + "layout": "full", + "figures": [ + "figures/figplan_fig3.png" + ], + "keywords": [], + "footnote": "PBISC gives better cells than scSemiProfiler, a VAE-based method, for far less compute. 6 of 6 panels ready.", + "sources": "tools/figure_plan.py; data/figure_plan.json", + "footnote_full": "Panels: a) Lupus benchmark against scSemiProfiler [Ready]. b) Five paired cohorts [Ready]. c) Five donors: generated cells [Ready]. d) scSemiProfiler copies its representative [Ready]. e) Same training pairs, 5-fold [Ready; scSemiProfiler CV still running for some cohorts]. f) Time to apply to a new cohort [Ready].", + "numbers": [] + }, + { + "kind": "content", + "section": "3", + "section_label": "Fig. 3a", + "title": "SLE benchmark: PBISC vs scSemiProfiler", + "layout": "full", + "figures": [ + "figures/sec04_sle_benchmark.png" + ], + "keywords": [ + "42 held-out donors, sc-pseudobulk input", + "Beats scSemiProfiler on all three metrics", + "Bulk logFC alone scores higher (slide 13)" + ], + "footnote": "GSE135779, 42 held-out donors (30 SLE / 12 HC). PBISC: 5 seeds, mean ± SD; scSemiProfiler: 1 run.", + "sources": "tools/sec04_sle_benchmark.py; data/sec04_sle_benchmark.csv, data/sec04_sle_numbers.csv", + "footnote_full": "GSE135779, 42 held-out donors (30 SLE / 12 HC). PBISC f01 6k E16, 5 seeds (mean ± SD); baselines 1 draw. scSemiProfiler also gets real sc of 14 in-cohort donors (1,500 cells each). Dashed: input bulk logFC copied to every type (broadcast slide).", + "refs": [ + "Wang 2024, Nat Commun", + "Xu 2021, Mol Syst Biol" + ], + "refs_full": [ + "Wang 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", + "Xu et al. Probabilistic harmonization and annotation of single-cell transcriptomics data with deep generative models. Mol Syst Biol 17:e9620 (2021)" + ] + }, + { + "kind": "content", + "section": "3", + "section_label": "Fig. 3a", + "title": "SLE benchmark: per cell type", + "layout": "full", + "figures": [ + "figures/sec04_sle_per_celltype.png" + ], + "keywords": [ + "Beats scSemiProfiler in T, B, Mono, NK", + "Loses in Other; DC r ties, fewer DEGs", + "DC: ~15 generated cells per donor" + ], + "footnote": "Same 42 donors, 1,006 genes. PBISC: 5 seeds, mean ± SD; scSemiProfiler: 1 run. Other: top-50 of 12.", + "say": "PBISC's DC top list in the base seed has only 13 genes, because few DC cells are generated.", + "sources": "tools/sec04_sle_benchmark.py; data/sec04_sle_per_celltype.csv, data/sec04_sle_numbers.csv", + "footnote_full": "Same 42 donors, 1,006 genes. PBISC 5 seeds (mean ± SD), baselines 1 draw. Tick %: share of 251,179 real cells. Other top-50 is out of 12 truth genes. Bulk copy: r 0.68, top-50 113 (broadcast slide).", + "refs": [ + "Wang 2024, Nat Commun" + ], + "refs_full": [ + "Wang 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" + ] + }, + { + "kind": "content", + "section": "3", + "section_label": "Fig. 3a", + "title": "Bulk broadcast baseline", + "layout": "full", + "figures": [ + "figures/sec04_sle_broadcast.png" + ], + "keywords": [ + "Bulk logFC copied: r 0.68 vs 0.37", + "Type-specific part: partial r 0.21", + "scSemiProfiler: r 0.32, partial r 0.17" + ], + "footnote": "SLE, 42 donors, 1,006 genes. Partial r: both logFCs residualised on bulk logFC.", + "sources": "tools/sec04_sle_benchmark.py; data/sec04_sle_broadcast.csv, data/sec04_sle_numbers.csv", + "footnote_full": "SLE held-out, 42 donors, 1,006 genes; PBISC 5 seeds. Partial r: truth and method logFC residualised on bulk logFC. Broadcast has no residual, so it is not scored. Split-half and BayesPrism checks: sle_broadcast/RESULTS.md.", + "refs": [ + "Viñas Torné 2026, Nat Biotechnol", + "Ahlmann-Eltze 2025, Nat Methods", + "Chu 2022, Nat Cancer" + ], + "refs_full": [ + "Viñas Torné 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", + "Ahlmann-Eltze, Huber & Anders. Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. Nat Methods 22:1657-1661 (2025). doi:10.1038/s41592-025-02772-6", + "Chu 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)" + ] + }, + { + "kind": "content", + "section": "3", + "section_label": "Fig. 3b", + "sec09_index": 0, + "title": "Five paired cohorts", + "layout": "full", + "figures": [ + "figures/sec09_cohorts.png" + ], + "keywords": [ + "403 paired samples, 275 donors", + "Only identity-pass pairs; donor-level splits", + "Ji bulk: whole blood, not PBMC" + ], + "footnote": "Usable pair = passed identity check and ≥ 200 cells: 403 of 430 pairs; donor-level splits.", + "footnote_full": "Usable pair = passed the pairing identity check and >= 200 cells (fold >= 0 in data//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//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//QC_REPORT.md; ImMicro: GEO GSE314922/GSE314416.", + "sources": "tools/sec09_paired_validation.py; data/sec09_cohorts.csv, data/sec09_quoted_numbers.csv, data/sec09_slide_text.json", + "numbers": [ + { + "value": "403 paired samples (sum of usable rows)", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data//fewshot_n5/DRAWS.json (n_usable_rows)" + }, + { + "value": "275 donors (sum of usable donors)", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data//fewshot_n5/DRAWS.json (n_usable_donors)" + }, + { + "value": "dropped 15 identity-fail, 11 borderline, 1 < 200 cells, of 430", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data//features.npz (identity_pass, identity_check); samples.csv (n_cells); /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/tools/prepare_dataset.py (MIN_CELLS)" + }, + { + "value": "ImMicro 107 pairs / 107 donors", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/immicro/fewshot_n5/DRAWS.json" + }, + { + "value": "Sepsis 82 pairs / 82 donors; sepsis 61 vs control 21", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/ye2026_sepsis/fewshot_n5/DRAWS.json; /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/ye2026_sepsis/pooled/summary.json" + }, + { + "value": "Imperial 157 pairs / 57 donors; COVID+ 113 vs COVID− 34 (+10 recovery, neither)", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/imperial_eskd/fewshot_n5/DRAWS.json; /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/imperial_eskd/pooled/summary.json" + }, + { + "value": "Giroux 24 pairs / 9 donors; moderate 11 vs mild 13", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/giroux2022/fewshot_n5/DRAWS.json; /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/giroux2022/pooled/summary.json" + }, + { + "value": "Ji 33 pairs / 20 donors; ChILI 18 vs ICI control 15", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/ji2025_chili/fewshot_n5/DRAWS.json; /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/ji2025_chili/pooled/summary.json" + } + ] + }, + { + "kind": "content", + "section": "3", + "section_label": "Fig. 3c", + "sec09_index": 1, + "title": "5 donors: LoRA vs scSemiProfiler — cells", + "layout": "full", + "figures": [ + "figures/sec09_fewpair_cells.png" + ], + "keywords": [ + "LoRA-5 > scSemi-5: coverage, precision 5/5", + "Beats zero-shot + calibration: coverage 5/5", + "Lower L1 than calibration: ImMicro only" + ], + "footnote": "5 draws per cohort; test = all other usable pairs. Calibration α from the same 5 donors. scANVI labels.", + "footnote_full": "Base f01 6k·1x·E16 (EMA, DDIM 50) + LoRA r=16, 3,000 steps. LoRA-5 / scSemi-5 = LoRA / scSemiProfiler trained on the same 5 donors (5–15 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 0.83–0.97 vs scSemi-5 0.15–0.28. 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//fewshot_n5/.", + "sources": "tools/sec09_paired_validation.py; data/sec09_fewpair_cells.csv, data/sec09_scsemi_status.csv, data/sec09_slide_text.json", + "numbers": [ + { + "value": "LoRA-5 vs scSemi-5 coverage (shared-draw means; LoRA-better draws/n; verdict): ImMicro 0.792 vs 0.036 (5/5, lora); Sepsis 0.694 vs 0.080 (5/5, lora); Imperial 0.737 vs 0.059 (5/5, lora); Giroux 0.961 vs 0.607 (5/5, lora); Ji 0.707 vs 0.145 (5/5, lora) -> LoRA verdicts 5/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/summary.json and draw_scsemi5/summary.json" + }, + { + "value": "LoRA-5 vs scSemi-5 precision (shared-draw means; LoRA-better draws/n; verdict): ImMicro 0.952 vs 0.518 (5/5, lora); Sepsis 0.906 vs 0.728 (5/5, lora); Imperial 0.853 vs 0.330 (5/5, lora); Giroux 1.000 vs 0.992 (5/5, lora); Ji 0.978 vs 0.840 (5/5, lora) -> LoRA verdicts 5/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/summary.json and draw_scsemi5/summary.json" + }, + { + "value": "LoRA-5 vs scSemi-5 comp_l1 (shared-draw means; LoRA-better draws/n; verdict): ImMicro 0.176 vs 0.233 (5/5, lora); Sepsis 0.466 vs 0.612 (5/5, lora); Imperial 0.503 vs 0.602 (5/5, lora); Giroux 0.256 vs 0.270 (3/5, split); Ji 0.513 vs 0.592 (4/5, lora) -> LoRA verdicts 4/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/summary.json and draw_scsemi5/summary.json" + }, + { + "value": "LoRA-5 vs scSemi-5 coverage_dm (shared-draw means; LoRA-better draws/n; verdict): ImMicro 0.966 vs 0.284 (5/5, lora); Sepsis 0.834 vs 0.158 (5/5, lora); Imperial 0.863 vs 0.152 (5/5, lora); Giroux 0.974 vs 0.238 (5/5, lora); Ji 0.908 vs 0.150 (5/5, lora) -> LoRA verdicts 5/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/summary.json and draw_scsemi5/summary.json" + }, + { + "value": "LoRA-5 vs scSemi-5 precision_dm (shared-draw means; LoRA-better draws/n; verdict): ImMicro 0.975 vs 0.829 (5/5, lora); Sepsis 0.914 vs 0.504 (5/5, lora); Imperial 0.881 vs 0.429 (5/5, lora); Giroux 0.995 vs 0.426 (5/5, lora); Ji 0.974 vs 0.564 (5/5, lora) -> LoRA verdicts 5/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/summary.json and draw_scsemi5/summary.json" + }, + { + "value": "LoRA-5 vs zero-shot + calibration coverage (same-draw means; LoRA-better draws/n; verdict): ImMicro 0.792 vs 0.535 (5/5, lora); Sepsis 0.694 vs 0.428 (5/5, lora); Imperial 0.737 vs 0.634 (5/5, lora); Giroux 0.961 vs 0.944 (5/5, lora); Ji 0.707 vs 0.114 (5/5, lora) -> LoRA verdicts 5/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/summary.json (arms lora5, zs_cal)" + }, + { + "value": "LoRA-5 vs zero-shot + calibration precision (same-draw means; LoRA-better draws/n; verdict): ImMicro 0.952 vs 0.904 (5/5, lora); Sepsis 0.906 vs 0.899 (3/5, split); Imperial 0.853 vs 0.845 (3/5, split); Giroux 1.000 vs 0.996 (4/5, lora); Ji 0.978 vs 0.782 (5/5, lora) -> LoRA verdicts 3/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/summary.json (arms lora5, zs_cal)" + }, + { + "value": "LoRA-5 vs zero-shot + calibration comp_l1 (same-draw means; LoRA-better draws/n; verdict): ImMicro 0.176 vs 0.203 (5/5, lora); Sepsis 0.466 vs 0.446 (2/5, split); Imperial 0.503 vs 0.506 (2/5, split); Giroux 0.256 vs 0.288 (2/5, split); Ji 0.513 vs 0.514 (3/5, split) -> LoRA verdicts 1/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/summary.json (arms lora5, zs_cal)" + }, + { + "value": "LoRA-5 vs zero-shot + calibration coverage_dm (same-draw means; LoRA-better draws/n; verdict): ImMicro 0.966 vs 0.780 (5/5, lora); Sepsis 0.834 vs 0.658 (5/5, lora); Imperial 0.863 vs 0.681 (5/5, lora); Giroux 0.974 vs 0.937 (5/5, lora); Ji 0.908 vs 0.357 (5/5, lora) -> LoRA verdicts 5/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/summary.json (arms lora5, zs_cal)" + }, + { + "value": "LoRA-5 vs zero-shot + calibration precision_dm (same-draw means; LoRA-better draws/n; verdict): ImMicro 0.975 vs 0.940 (5/5, lora); Sepsis 0.914 vs 0.909 (3/5, split); Imperial 0.881 vs 0.868 (3/5, split); Giroux 0.995 vs 0.989 (5/5, lora); Ji 0.974 vs 0.658 (5/5, lora) -> LoRA verdicts 3/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/summary.json (arms lora5, zs_cal)" + }, + { + "value": "LoRA-5 vs zero-shot (real bulk) coverage (same-draw means; LoRA-better draws/n; verdict): ImMicro 0.792 vs 0.190 (5/5, lora); Sepsis 0.694 vs 0.076 (5/5, lora); Imperial 0.737 vs 0.114 (5/5, lora); Giroux 0.961 vs 0.472 (5/5, lora); Ji 0.707 vs 0.066 (5/5, lora) -> LoRA verdicts 5/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/summary.json (arms lora5, zs_raw)" + }, + { + "value": "LoRA-5 vs zero-shot (real bulk) precision (same-draw means; LoRA-better draws/n; verdict): ImMicro 0.952 vs 0.677 (5/5, lora); Sepsis 0.906 vs 0.625 (5/5, lora); Imperial 0.853 vs 0.500 (5/5, lora); Giroux 1.000 vs 0.980 (5/5, lora); Ji 0.978 vs 0.799 (5/5, lora) -> LoRA verdicts 5/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/summary.json (arms lora5, zs_raw)" + }, + { + "value": "LoRA-5 vs zero-shot (real bulk) comp_l1 (same-draw means; LoRA-better draws/n; verdict): ImMicro 0.176 vs 0.445 (5/5, lora); Sepsis 0.466 vs 0.510 (5/5, lora); Imperial 0.503 vs 0.806 (5/5, lora); Giroux 0.256 vs 0.351 (5/5, lora); Ji 0.513 vs 0.721 (5/5, lora) -> LoRA verdicts 5/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/summary.json (arms lora5, zs_raw)" + }, + { + "value": "LoRA-5 vs zero-shot (real bulk) coverage_dm (same-draw means; LoRA-better draws/n; verdict): ImMicro 0.966 vs 0.213 (5/5, lora); Sepsis 0.834 vs 0.073 (5/5, lora); Imperial 0.863 vs 0.151 (5/5, lora); Giroux 0.974 vs 0.549 (5/5, lora); Ji 0.908 vs 0.199 (5/5, lora) -> LoRA verdicts 5/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/summary.json (arms lora5, zs_raw)" + }, + { + "value": "LoRA-5 vs zero-shot (real bulk) precision_dm (same-draw means; LoRA-better draws/n; verdict): ImMicro 0.975 vs 0.671 (5/5, lora); Sepsis 0.914 vs 0.595 (5/5, lora); Imperial 0.881 vs 0.561 (5/5, lora); Giroux 0.995 vs 0.996 (2/5, split); Ji 0.974 vs 0.480 (5/5, lora) -> LoRA verdicts 4/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/summary.json (arms lora5, zs_raw)" + }, + { + "value": "coverage cohort means over all draws: LoRA-5 ImMicro 0.792, Sepsis 0.694, Imperial 0.737, Giroux 0.961, Ji 0.707; scSemi-5 ImMicro 0.036, Sepsis 0.080, Imperial 0.059, Giroux 0.607, Ji 0.145; zero-shot + calibration ImMicro 0.535, Sepsis 0.428, Imperial 0.634, Giroux 0.944, Ji 0.114; zero-shot ImMicro 0.190, Sepsis 0.076, Imperial 0.114, Giroux 0.472, Ji 0.066", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/summary.json and draw_scsemi5/summary.json" + }, + { + "value": "LoRA-5 minus zero-shot + calibration composition L1, per-donor Wilcoxon (draw-averaged): ImMicro -0.027, n=107, p=2.2e-10; Sepsis +0.018, n=82, p=0.79; Imperial +0.016, n=57, p=0.38; Giroux -0.065, n=8, p=0.38; Ji -0.012, n=20, p=0.7 -> gain in 1/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/per_donor.csv; /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data//folds.csv" + }, + { + "value": "LoRA-5 minus zero-shot (real bulk) composition L1, per-donor Wilcoxon (draw-averaged): ImMicro -0.270, n=107, p=2.2e-18; Sepsis -0.044, n=82, p=0.15; Imperial -0.348, n=57, p=1.6e-08; Giroux -0.107, n=8, p=0.039; Ji -0.232, n=20, p=0.017 -> gain in 4/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/per_donor.csv; /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data//folds.csv" + }, + { + "value": "scSemiProfiler draws used / LoRA draws: ImMicro 5/5, Sepsis 5/5, Imperial 5/5, Giroux 5/5, Ji 5/5", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/data/sec09_scsemi_status.csv" + }, + { + "value": "training pairs per draw: ImMicro 5, Sepsis 5, Imperial 12–15, Giroux 12–14, Ji 7–9; test pairs: ImMicro 102, Sepsis 77, Imperial 142–145, Giroux 10–12, Ji 24–26", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data//fewshot_n5/DRAWS.json (n_train_rows, n_test_rows)" + }, + { + "value": "LoRA r = 16; 3000 fixed steps; base runs/f01_6k_1x_seed20260627/checkpoint_epoch_016.pt; sampling: test donors only, 1,500 cells each, seed 20260927, DDIM 50, EMA, Bernoulli read-out", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/runs/immicro/fewshot_n5/draw0/lora/RUN.json; /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/README.md" + }, + { + "value": "P/C definition: precision/coverage/density (PCA-50 + k = 10 balls fitted on the scored donors' observed R1 cells); PCA components in every summary.json = 50", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/tools/score_ft.py" + } + ], + "refs": [ + "Kynkäänniemi 2019, NeurIPS", + "Naeem 2020, ICML", + "Wang 2024, Nat Commun" + ], + "refs_full": [ + "Kynkäänniemi et al. Improved precision and recall metric for assessing generative models. NeurIPS 2019", + "Naeem et al. Reliable fidelity and diversity metrics for generative models. ICML 2020", + "Wang 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" + ] + }, + { + "kind": "content", + "section": "3", + "section_label": "Fig. 3c", + "text_source": { + "file": "data/sec09_extra_slide_text.json", + "index": 0 + }, + "title": "5 donors: realism beyond precision/coverage", + "layout": "full", + "figures": [ + "figures/sec09_extra_fewpair.png" + ], + "keywords": [ + "Classifier AUC: LoRA-5 best in 25/25 draws", + "E_N: LoRA-5 below scSemi-5 in 25/25 draws", + "Still separable: LoRA-5 AUC 0.89–0.95" + ], + "footnote": "E_N: scSemiProfiler's own error metric. Misses only vs calibration, e.g. Imperial E_N 1.28 vs 1.22.", + "footnote_full": "Same 5-donor draws as the previous slides; every metric on exactly the scored cells (scores_extra/, tools/extra_metrics.py). Real side for AUC and MMD² = held-out real cells of the same samples (R2), projected like the generated cells; real vs real = R2 vs R3 (AUC 0.49–0.50). Classifier AUC: 5-fold CV ROC AUC of a kNN classifier (k = 5) separating real from generated cells, 0.5 = indistinguishable; bars start at 0.5. MMD²: unbiased, Gaussian kernel, bandwidth = median distance among real cells. E_N: scSemiProfiler's normalized error (Wang et al. 2024, eq. 17–18), their code's squared L2 on all cells, per-arm PCA-100; real vs real = within-sample half split (0.07–0.38); every generated arm is above 1. LoRA-5 best of the 4 arms in 25/25 draws on kNN AUC, 25/25 on RF AUC, 25/25 on 1-NN accuracy, 23/25 on MMD², 22/25 on E_N. LoRA-5 better than scSemi-5, zero-shot and zero-shot + calibration on both AUCs, 1-NN accuracy and coverage at k = 3, 5, 10 in every draw. E_N scSemi-5 vs LoRA-5 (cohort means): ImMicro 3.65 vs 1.55; Sepsis 3.83 vs 1.46; Imperial 2.04 vs 1.28; Giroux 3.79 vs 1.99; Ji 4.27 vs 2.60; LoRA-5 lower in 25/25 draws. Draws where LoRA-5 is worse (cohort means in brackets): vs zero-shot + calibration: Sepsis precision k5 1/5 (0.778 vs 0.752), precision k10 2/5 (0.906 vs 0.899); Imperial MMD² (σ = m) 1/5 (0.0156 vs 0.0218), MMD² (σ sum) 1/5 (0.0443 vs 0.0673), E_N (all cells) 3/5 (1.28 vs 1.22), precision k3 2/5 (0.610 vs 0.598), precision k5 2/5 (0.726 vs 0.719), precision k10 2/5 (0.853 vs 0.845); Giroux MMD² (σ = m) 1/5 (0.0178 vs 0.0257), MMD² (σ sum) 1/5 (0.0584 vs 0.0838), precision k5 1/5 (0.995 vs 0.981), precision k10 1/5 (1.000 vs 0.996). LoRA-5 AUC per draw 0.88–0.97 (kNN), RF cohort means 0.90–0.96: a classifier still tells generated from real cells. Scores: scores_extra//fewshot_n5/draw/.", + "sources": "tools/sec09_extra_metrics.py; data/sec09_extra_fewpair.csv, data/sec09_extra_quoted_numbers.csv; plans/paired_finetune_intradataset_20260927/scores_extra/ (README.md, TABLES.md)", + "numbers": [ + { + "value": "LoRA-5 best of 4 arms (draws): kNN AUC 25/25, RF AUC 25/25, 1-NN 25/25, MMD² 23/25, E_N 22/25", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_extra//fewshot_n5/draw/{metrics,en}.csv" + }, + { + "value": "E_N LoRA-5 < scSemi-5 in 25/25 draws; cohort means scSemi-5 vs LoRA-5: ImMicro 3.65 vs 1.55; Sepsis 3.83 vs 1.46; Imperial 2.04 vs 1.28; Giroux 3.79 vs 1.99; Ji 4.27 vs 2.60", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_extra//fewshot_n5/draw/en.csv (cellset all, distance sq)" + }, + { + "value": "LoRA-5 kNN AUC cohort means: ImMicro 0.943, Sepsis 0.936, Imperial 0.918, Giroux 0.893, Ji 0.955; per draw 0.878–0.975", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_extra//fewshot_n5/draw/{metrics,en}.csv" + }, + { + "value": "real vs real: kNN AUC ImMicro 0.502, Sepsis 0.501, Imperial 0.498, Giroux 0.495, Ji 0.491; E_N ImMicro 0.382, Sepsis 0.100, Imperial 0.260, Giroux 0.145, Ji 0.073", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_extra//fewshot_n5/draw/{metrics,en}.csv" + }, + { + "value": "LoRA-5 worse than a comparator (draws): vs zero-shot + calibration: Sepsis precision k5 1/5 (0.778 vs 0.752), precision k10 2/5 (0.906 vs 0.899); Imperial MMD² (σ = m) 1/5 (0.0156 vs 0.0218), MMD² (σ sum) 1/5 (0.0443 vs 0.0673), E_N (all cells) 3/5 (1.28 vs 1.22), precision k3 2/5 (0.610 vs 0.598), precision k5 2/5 (0.726 vs 0.719), precision k10 2/5 (0.853 vs 0.845); Giroux MMD² (σ = m) 1/5 (0.0178 vs 0.0257), MMD² (σ sum) 1/5 (0.0584 vs 0.0838), precision k5 1/5 (0.995 vs 0.981), precision k10 1/5 (1.000 vs 0.996)", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_extra//fewshot_n5/draw/{metrics,en}.csv" + }, + { + "value": "cross-check: 720 cohort × arm × metric values equal scores_extra/{fewshot,pooled}_headline.csv", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_extra/fewshot_headline.csv; /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_extra/pooled_headline.csv" + } + ], + "refs": [ + "Lopez-Paz 2017, ICLR", + "Gretton 2012, JMLR", + "Wang 2024, Nat Commun" + ], + "refs_full": [ + "Lopez-Paz & Oquab. Revisiting classifier two-sample tests. ICLR 2017", + "Gretton et al. A kernel two-sample test. JMLR 13:723-773 (2012)", + "Wang 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" + ] + }, + { + "kind": "content", + "section": "3", + "section_label": "Fig. 3c", + "text_source": { + "file": "data/sec09_extra_slide_text.json", + "index": 2 + }, + "title": "5 donors: precision/coverage at k = 3, 5, 10", + "layout": "full", + "figures": [ + "figures/sec09_extra_ksens.png" + ], + "keywords": [ + "Smaller k lowers every set, real included", + "Coverage: LoRA-5 first at all k, 25/25", + "Precision: LoRA-5 first in 20–23/25 draws" + ], + "footnote": "5 donors, 5 draws per cohort; error bars = SD over draws. k = 10 is the value on the main slides.", + "footnote_full": "Precision 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: 0.92–0.99), 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 4 arms across k unchanged; LoRA first at every k in 5/5 cohorts. Scores: scores_extra//fewshot_n5/draw/metrics.csv.", + "sources": "tools/sec09_extra_metrics.py; data/sec09_extra_ksens.csv; plans/paired_finetune_intradataset_20260927/scores_extra/", + "numbers": [ + { + "value": "LoRA-5 first on coverage at k = 3, 5 and 10: 25/25 draws", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_extra//fewshot_n5/draw/{metrics,en}.csv" + }, + { + "value": "LoRA-5 first on precision (draws) at k = 3/5/10: 23/21/20 of 25", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_extra//fewshot_n5/draw/{metrics,en}.csv" + }, + { + "value": "real vs real coverage at k = 3: ImMicro 0.932, Sepsis 0.925, Imperial 0.919, Giroux 0.993, Ji 0.980", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_extra//fewshot_n5/draw/{metrics,en}.csv" + }, + { + "value": "every set falls with smaller k (cohort means): True", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_extra//fewshot_n5/draw/{metrics,en}.csv" + } + ], + "refs": [ + "Naeem 2020, ICML", + "Khayatkhoei 2023, ICML" + ], + "refs_full": [ + "Naeem et al. Reliable fidelity and diversity metrics for generative models. ICML 2020", + "Khayatkhoei & AbdAlmageed. Emergent asymmetry of precision and recall for measuring fidelity and diversity of generative models in high dimensions. ICML 2023" + ] + }, + { + "kind": "content", + "section": "3", + "section_label": "Fig. 3d", + "text_source": { + "file": "data/sec09_scsemi_representative_slide_text.json", + "index": 0 + }, + "title": "scSemiProfiler copies its representative", + "layout": "full", + "figures": [ + "figures/sec09_scsemi_representative.png" + ], + "keywords": [ + "scSemi-5 never beats copying its representative: 5/5", + "Its error tracks the representative's: rho 0.57–0.93", + "LoRA-5 below the copy error in 4/5 cohorts" + ], + "footnote": "Copy = composition of the one training sample scSemiProfiler assigns to each test sample.", + "sources": "plans/paired_finetune_intradataset_20260927/scsemi/tools/representative_dependence.py; tools/sec09_scsemi_representative.py; data/sec09_scsemi_representative.csv", + "footnote_full": "Same 25 five-donor draws. scSemiProfiler gives every test sample exactly one representative, the nearest training sample in its bulk PCA (assignment.json), and runs its encoder on that representative's cells. Copy = composition L1 between the test sample's and its representative's observed cells (same frozen scANVI labels as the scorer). Bars: mean over draws of the per-draw mean; dots = draws. scSemi vs copy: paired Wilcoxon over test samples (ImMicro p 1e-33, Sepsis p 0.014, Imperial p 0.91, Giroux p 0.0073, Ji p 0.46). Upstream's own representative choice (KMeans on the same bulk PCA, the sample nearest each centre; proxy, not a rerun) gives a lower copy error than our 5-donor draws in Imperial, Giroux and Ji, where our draws hold all visits of 5 donors and KMeans spreads the same number of samples over as many donors; equal or higher in ImMicro and sepsis. Copy error: ImMicro (5 samples) ours 0.19, KMeans 0.18, random samples 0.19; Sepsis (5 samples) ours 0.58, KMeans 0.61, random samples 0.58; Imperial (14 samples) ours 0.61, KMeans 0.53, random samples 0.56; Giroux (13 samples) ours 0.25, KMeans 0.16, random samples 0.23; Ji (8 samples) ours 0.57, KMeans 0.46, random samples 0.51. Source: plans/paired_finetune_intradataset_20260927/scsemi/analysis/representative_dependence_20261002/.", + "numbers": [ + { + "value": "ImMicro: copy 0.187, scSemi-5 0.233 (Wilcoxon vs copy p 1e-33), LoRA-5 0.176; rho scSemi 0.57, LoRA 0.35", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scsemi/analysis/representative_dependence_20261002/summary.csv" + }, + { + "value": "Sepsis: copy 0.580, scSemi-5 0.612 (Wilcoxon vs copy p 0.014), LoRA-5 0.466; rho scSemi 0.88, LoRA 0.56", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scsemi/analysis/representative_dependence_20261002/summary.csv" + }, + { + "value": "Imperial: copy 0.608, scSemi-5 0.602 (Wilcoxon vs copy p 0.91), LoRA-5 0.503; rho scSemi 0.93, LoRA 0.55", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scsemi/analysis/representative_dependence_20261002/summary.csv" + }, + { + "value": "Giroux: copy 0.253, scSemi-5 0.267 (Wilcoxon vs copy p 0.0073), LoRA-5 0.258; rho scSemi 0.81, LoRA 0.08", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scsemi/analysis/representative_dependence_20261002/summary.csv" + }, + { + "value": "Ji: copy 0.570, scSemi-5 0.591 (Wilcoxon vs copy p 0.46), LoRA-5 0.511; rho scSemi 0.72, LoRA 0.39", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scsemi/analysis/representative_dependence_20261002/summary.csv" + }, + { + "value": "KMeans proxy, copy error: ImMicro (5 samples) ours 0.19, KMeans 0.18, random samples 0.19; Sepsis (5 samples) ours 0.58, KMeans 0.61, random samples 0.58; Imperial (14 samples) ours 0.61, KMeans 0.53, random samples 0.56; Giroux (13 samples) ours 0.25, KMeans 0.16, random samples 0.23; Ji (8 samples) ours 0.57, KMeans 0.46, random samples 0.51", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scsemi/analysis/representative_dependence_20261002/kmeans_proxy.csv" + } + ], + "refs": [ + "Wang 2024, Nat Commun" + ], + "refs_full": [ + "Wang 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" + ] + }, + { + "kind": "content", + "section": "3", + "section_label": "Fig. 3e", + "text_source": { + "file": "data/sec09_same_pairs_slide_text.json", + "index": 0 + }, + "title": "Same training pairs: LoRA vs scSemiProfiler", + "layout": "full", + "figures": [ + "figures/sec09_same_pairs.png" + ], + "keywords": [ + "Same pairs: LoRA ahead on every metric", + "Coverage 0.78 vs 0.04 (ImMicro)", + "19–20× less GPU time per fold" + ], + "footnote": "5-fold CV; scSemiProfiler representatives = LoRA's training pairs. Sepsis, Imperial, Giroux: running.", + "sources": "tools/sec09_same_pairs.py; data/sec09_same_pairs.csv; plans/paired_finetune_intradataset_20260927/scsemi/analysis/cv_budget_20261005/", + "say": "Here scSemiProfiler gets exactly the pairs LoRA trains on, every training donor of the fold. LoRA is still ahead on every metric and about 20 times cheaper. Sepsis, Imperial and Giroux are still running.", + "footnote_full": "5-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. Still running (server_1 / server_2): Sepsis, Imperial, Giroux. Equal pair budget, ImMicro (n = 5, 10, 20, 40 pairs): scSemiProfiler L1 0.218–0.227 and coverage 0.04–0.04 at every n; LoRA L1 0.171–0.188, coverage 0.76–0.84 (drawn on the 'Pairs needed: composition, coverage, training steps' slide). Source: plans/paired_finetune_intradataset_20260927/scores//{pooled,scsemi_cv_20261002/pooled}/ and scsemi/analysis/cv_budget_20261005/.", + "numbers": [ + { + "value": "ImMicro: L1 LoRA 0.162 / scSemi 0.217; coverage 0.784 / 0.038; precision 0.924 / 0.505; logFC r nan / nan; top-50 DEG None / None; GPU per fold LoRA 14.7 min / scSemi 4.75 h", + "source": "data/sec09_same_pairs.csv" + }, + { + "value": "Ji: L1 LoRA 0.504 / scSemi 0.575; coverage 0.857 / 0.167; precision 0.971 / 0.858; logFC r 0.479 / 0.273; top-50 DEG 15 / 0; GPU per fold LoRA 4.4 min / scSemi 1.48 h", + "source": "data/sec09_same_pairs.csv" + }, + { + "value": "Equal pair budget, ImMicro (n = 5, 10, 20, 40 pairs): scSemiProfiler L1 0.218–0.227 and coverage 0.04–0.04 at every n; LoRA L1 0.171–0.188, coverage 0.76–0.84 (drawn on the 'Pairs needed: composition, coverage, training steps' slide).", + "source": "data/sec09_same_pairs_budget.csv" + } + ] + }, + { + "kind": "content", + "section": "3", + "section_label": "Fig. 3f", + "text_source": { + "file": "data/sec09_time_to_apply_slide_text.json", + "index": 0 + }, + "title": "5 donors: time to apply", + "layout": "full", + "figures": [ + "figures/sec09_time_to_apply.png" + ], + "keywords": [ + "LoRA-5: 3–11 min per draw, every test sample", + "scSemiProfiler-5: 57 min–4.1 h per draw", + "17–59× less GPU time, same 5 donors" + ], + "footnote": "One GB10 GPU; mean of 5 draws. Data prep and scoring not counted. PBISC pretraining (15.9 h) is done once.", + "sources": "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)", + "footnote_full": "Same 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 15–19×; Sepsis 26–47×; Imperial 19–26×; Giroux 54–63×; Ji 11–20× (all 25 draws 11–63×; keyword and figure = mean of each cohort's 5 draws).", + "numbers": [ + { + "value": "ImMicro: LoRA-5 8.4 min (fine-tune 1.9 + generate 6.5), scSemiProfiler-5 141 min (representatives 15 + per-sample models 126); 15–19× per draw; per test sample 4.9 s vs 83 s", + "source": "data/sec09_time_to_apply.csv" + }, + { + "value": "Sepsis: LoRA-5 6.7 min (fine-tune 1.8 + generate 4.9), scSemiProfiler-5 244 min (representatives 30 + per-sample models 214); 26–47× per draw; per test sample 5.2 s vs 190 s", + "source": "data/sec09_time_to_apply.csv" + }, + { + "value": "Imperial: LoRA-5 10.8 min (fine-tune 1.8 + generate 9.0), scSemiProfiler-5 229 min (representatives 32 + per-sample models 197); 19–26× per draw; per test sample 4.5 s vs 96 s", + "source": "data/sec09_time_to_apply.csv" + }, + { + "value": "Giroux: LoRA-5 2.6 min (fine-tune 1.9 + generate 0.7), scSemiProfiler-5 156 min (representatives 117 + per-sample models 38); 54–63× per draw; per test sample 14.8 s vs 872 s", + "source": "data/sec09_time_to_apply.csv" + }, + { + "value": "Ji: LoRA-5 3.4 min (fine-tune 1.8 + generate 1.6), scSemiProfiler-5 57 min (representatives 24 + per-sample models 33); 11–20× per draw; per test sample 8.1 s vs 137 s", + "source": "data/sec09_time_to_apply.csv" + }, + { + "value": "PBISC one-time pretraining 15.9 h (not counted per draw)", + "source": "data/sec06_ft_method.csv (pretrain_wall_hours)" + } + ], + "refs": [ + "Wang 2024, Nat Commun" + ], + "refs_full": [ + "Wang 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" + ] + }, + { + "kind": "divider", + "title": "Figure 4. Disease signal and rare cell states" + }, + { + "kind": "content", + "section": "plan", + "section_label": "Fig. 4 plan", + "text_source": { + "file": "data/figure_plan_slide_text.json", + "index": 4 + }, + "title": "Figure 4 plan: Disease signal and rare cell states", + "layout": "full", + "figures": [ + "figures/figplan_fig4.png" + ], + "keywords": [], + "footnote": "Open until the disease-cohort analyses are done, including LoRA on the disease cohort.", + "sources": "tools/figure_plan.py; data/figure_plan.json", + "footnote_full": "Panels: Slides kept in this section until it is planned: Pairs needed: donor and disease signal; 5 donors: LoRA vs scSemiProfiler — disease signal; Disease genes at every list length (CAT).", + "numbers": [] + }, + { + "kind": "content", + "section": "4", + "section_label": "Fig. 4 candidate", + "title": "Pairs needed: donor and disease signal", + "layout": "figure_keywords", + "figures": [ + "figures/sec08_pairs_signal.png" + ], + "keywords": [ + "No pairs or calibration only: logFC r ≤ 0.12", + "LoRA logFC r above broadcast from 10 pairs", + "LoRA top-50 DEG ≥ broadcast from 40 pairs" + ], + "footnote": "Sepsis: 61 case vs 21 control. logFC r = mean over T / B / Mono / NK. 0 = no pairs.", + "footnote_full": "Sepsis 61 case vs 21 control, pooled 5-fold scoring. Disease logFC r: as measured −0.01, length-corrected 0.11, calibration from the same n pairs 0.08–0.12, LoRA 0.29 / 0.42 / 0.37 / 0.48 / 0.53 at 5 / 10 / 20 / 40 / 58–59 pairs; broadcast 0.32; zero-shot from sc pseudobulk 0.47. Top-50 DEG (Welch, sum over 5 types): as measured 9, length-corrected 11, calibration 7–12, LoRA 5 / 8 / 18 / 22 / 23, broadcast 21. ImMicro donor NK fraction r: as measured 0.49, length-corrected 0.49, calibration 0.50–0.53, LoRA 0.20 → 0.70.", + "sources": "tools/sec08_pairs_needed.py; data/sec08_pairs_signal.csv, data/sec08_slide_numbers.csv", + "refs": [ + "Roohani 2024, Nat Biotechnol", + "Li 2026, Sci Adv" + ], + "refs_full": [ + "Roohani, Huang & Leskovec. Predicting transcriptional outcomes of novel multigene perturbations with GEARS. Nat Biotechnol 42:927-935 (2024). doi:10.1038/s41587-023-01905-6", + "Li et al. A systematic comparison of single-cell perturbation response prediction models. Sci Adv 12:eaed3414 (2026). doi:10.1126/sciadv.aed3414" + ] + }, + { + "kind": "content", + "section": "4", + "section_label": "Fig. 4 candidate", + "sec09_index": 2, + "title": "5 donors: LoRA vs scSemiProfiler — disease signal", + "layout": "full", + "figures": [ + "figures/sec09_fewpair_disease.png" + ], + "keywords": [ + "LoRA-5 > scSemi-5: DEG 2/4, logFC r 2/4", + "Spearman or limma: broadcast ahead 3/4", + "Split-half ceiling 0.96–0.99: not limiting" + ], + "footnote": "Giroux: sex-linked genes excluded (severity confounded with sex: 4F/0M vs 0F/5M). Top-50 DEG over 5 types.", + "footnote_full": "Same draws and verdict rule as the '5 donors: LoRA vs scSemiProfiler — cells' slide; DEG tie = half the draws equal or mean gap < 1 gene. Rank and threshold-free checks (scores_contrast/, tools/contrast_extra.py): Spearman r of the same logFC vectors; split-half ceiling = observed cells of each scored sample split in two halves (5 seeds), r of half-A vs half-B logFC, Spearman-Brown corrected to full data (noise-free ceiling); limma-voom top-50 = edgeR TMM + voom + limma eBayes on the same donor x type pseudobulks, BH FDR < 0.05, top 50 by |logFC|; grey tick = chance overlap (hypergeometric mean) of that arm's lists. Repeated samples of one patient (Imperial, Giroux, Ji) enter both tests as separate samples, as in the scorer. Sepsis: Spearman LoRA-5 0.22 / scSemi-5 0.16 / broadcast 0.33, limma top-50 LoRA-5 16.2 / scSemi-5 11.4 / broadcast 35.4 (chance 2.2), ceiling 0.99; Imperial: Spearman LoRA-5 0.55 / scSemi-5 0.37 / broadcast 0.36, limma top-50 LoRA-5 102.0 / scSemi-5 35.2 / broadcast 59.6 (chance 3.1), ceiling 0.97; Giroux: Spearman LoRA-5 0.19 / scSemi-5 0.27 / broadcast 0.38, limma top-50 LoRA-5 0.8 / scSemi-5 1.6 / broadcast 14.8 (chance 0.5), ceiling 0.97; Ji: Spearman LoRA-5 0.29 / scSemi-5 0.09 / broadcast 0.12, limma top-50 LoRA-5 3.0 / scSemi-5 0.6 / broadcast 6.2 (chance 1.4), ceiling 0.96. Earlier slide lines: logFC r: LoRA-5 > calibrated zero-shot 4/4; Giroux: broadcast r beats all arms. logFC r = mean over T/B/Mono/NK of case-vs-control log2FC r (generated vs observed); top-50 DEG = Welch+BH overlap summed over 5 types (max 250; Giroux 1–63, Ji 75–250 per draw). Giroux test = 2 vs 2 subjects per draw. LoRA-5-better draws (verdict): vs scSemi-5 logFC r Sepsis split 3/5, Imperial lora 5/5, Giroux scSemi-5 0/5, Ji lora 5/5; DEG Sepsis lora 5/5, Imperial lora 5/5, Giroux tie 1/5, Ji tie 1/5. Vs zero-shot + calibration: logFC r Sepsis lora 4/5, Imperial lora 5/5, Giroux lora 4/5, Ji lora 5/5; DEG Sepsis lora 4/5, Imperial lora 5/5, Giroux tie 1/5, Ji tie 1/5. Vs zero-shot (real bulk): logFC r Sepsis lora 5/5, Imperial lora 5/5, Giroux split 3/5, Ji lora 5/5; DEG Sepsis lora 4/5, Imperial lora 5/5, Giroux tie 1/5, Ji tie 1/5. Broadcast = the input real bulk's own logFC. Giroux: moderate vs mild subjects are 4F/0M vs 0F/5M (scored; all paired subjects 4F/1M vs 0F/5M; sex from bulk chrY expression = sc = GEO), so its disease metrics (logFC r, partial r, top-50 DEG, broadcast) exclude 196 sex-linked genes (43 chrY non-PAR + 153 X-inactivation escape / variable-escape genes of Tukiainen et al. 2017; PAR genes kept), 65 of them scoring genes; same labels and cells as before (scores_nosex/giroux2022/). Scores: scores//fewshot_n5/draw{,_scsemi5}/.", + "sources": "tools/sec09_paired_validation.py; data/sec09_fewpair_disease.csv, data/sec09_slide_text.json", + "numbers": [ + { + "value": "Giroux sex by contrast arm (subjects, F/M): scored moderate 4F/0M vs mild 0F/5M; all paired 4F/1M vs 0F/5M", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/giroux2022/features.npz (sex_bulk_expr = sex_sc_expr = sex_geo); /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/giroux2022/folds.csv" + }, + { + "value": "Giroux sex-linked genes excluded from the disease metrics: 196 listed (43 chrY non-PAR + 153 X escape), 65 of 5563 scoring genes", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/sex_linked_genes/sex_linked_genes_exclude.csv; /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/sex_linked_genes/BUILD.json; /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_nosex/giroux2022/pooled/summary.json (gene_exclusion)" + }, + { + "value": "Top-50 DEG LoRA-5 vs scSemi-5 (shared-draw means; LoRA-better / tied draws; verdict): Sepsis 16.4 vs 5.2 (5/5, 0 tied, lora); Imperial 113.6 vs 46.8 (5/5, 0 tied, lora); Giroux 0.2 vs 0.0 (1/5, 4 tied, tie); Ji 0.2 vs 0.0 (1/5, 4 tied, tie)", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw{,_scsemi5}/summary.json and contrast.csv (Giroux: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_nosex/giroux2022/)" + }, + { + "value": "Top-50 DEG LoRA-5 vs zero-shot + calibration (shared-draw means; LoRA-better / tied draws; verdict): Sepsis 16.4 vs 9.4 (4/5, 0 tied, lora); Imperial 113.6 vs 103.8 (5/5, 0 tied, lora); Giroux 0.2 vs 0.0 (1/5, 4 tied, tie); Ji 0.2 vs 0.0 (1/5, 4 tied, tie)", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw{,_scsemi5}/summary.json and contrast.csv (Giroux: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_nosex/giroux2022/)" + }, + { + "value": "Top-50 DEG LoRA-5 vs zero-shot (real bulk) (shared-draw means; LoRA-better / tied draws; verdict): Sepsis 16.4 vs 10.0 (4/5, 0 tied, lora); Imperial 113.6 vs 105.6 (5/5, 0 tied, lora); Giroux 0.2 vs 0.0 (1/5, 4 tied, tie); Ji 0.2 vs 0.0 (1/5, 4 tied, tie)", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw{,_scsemi5}/summary.json and contrast.csv (Giroux: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_nosex/giroux2022/)" + }, + { + "value": "logFC r LoRA-5 vs scSemi-5 (shared-draw means; LoRA-better draws; verdict): Sepsis 0.350 vs 0.402 (3/5, split); Imperial 0.749 vs 0.562 (5/5, lora); Giroux 0.297 vs 0.397 (0/5, other); Ji 0.333 vs 0.153 (5/5, lora)", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw{,_scsemi5}/summary.json and contrast.csv (Giroux: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_nosex/giroux2022/)" + }, + { + "value": "logFC r LoRA-5 vs zero-shot + calibration (shared-draw means; LoRA-better draws; verdict): Sepsis 0.350 vs 0.096 (4/5, lora); Imperial 0.749 vs 0.676 (5/5, lora); Giroux 0.297 vs 0.228 (4/5, lora); Ji 0.333 vs 0.140 (5/5, lora)", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw{,_scsemi5}/summary.json and contrast.csv (Giroux: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_nosex/giroux2022/)" + }, + { + "value": "logFC r LoRA-5 vs zero-shot (real bulk) (shared-draw means; LoRA-better draws; verdict): Sepsis 0.350 vs -0.024 (5/5, lora); Imperial 0.749 vs 0.638 (5/5, lora); Giroux 0.297 vs 0.255 (3/5, split); Ji 0.333 vs 0.087 (5/5, lora)", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw{,_scsemi5}/summary.json and contrast.csv (Giroux: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_nosex/giroux2022/)" + }, + { + "value": "Sepsis broadcast logFC r 0.320 vs best arm 0.402; broadcast DEG 20.0 vs best arm 16.4 (means over all draws; arms: zero-shot, zero-shot + calibration, scSemi-5, LoRA-5)", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/ye2026_sepsis/fewshot_n5/draw{,_scsemi5}/summary.json" + }, + { + "value": "Imperial broadcast logFC r 0.656 vs best arm 0.749; broadcast DEG 74.8 vs best arm 113.6 (means over all draws; arms: zero-shot, zero-shot + calibration, scSemi-5, LoRA-5)", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/imperial_eskd/fewshot_n5/draw{,_scsemi5}/summary.json" + }, + { + "value": "Giroux broadcast logFC r 0.492 vs best arm 0.397; broadcast DEG 0.0 vs best arm 0.2 (means over all draws; arms: zero-shot, zero-shot + calibration, scSemi-5, LoRA-5)", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_nosex/giroux2022/fewshot_n5/draw{,_scsemi5}/summary.json" + }, + { + "value": "Ji broadcast logFC r 0.234 vs best arm 0.333; broadcast DEG 1.0 vs best arm 0.2 (means over all draws; arms: zero-shot, zero-shot + calibration, scSemi-5, LoRA-5)", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores/ji2025_chili/fewshot_n5/draw{,_scsemi5}/summary.json" + }, + { + "value": "DEG maximum per draw: Sepsis 250, Imperial 250, Giroux 1–63, Ji 75–250", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores//fewshot_n5/draw/summary.json (deg_top50_max; Giroux: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_nosex)" + }, + { + "value": "Giroux test subjects per draw: 2 vs 2 (moderate vs mild); test pairs 10–12", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/giroux2022/fewshot_n5/DRAWS.json (train_donors); /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/data/giroux2022/folds.csv (group)" + } + ], + "refs": [ + "Roohani 2024, Nat Biotechnol", + "Viñas Torné 2026, Nat Biotechnol", + "Li 2026, Sci Adv", + "Law 2014, Genome Biol" + ], + "refs_full": [ + "Roohani, Huang & Leskovec. Predicting transcriptional outcomes of novel multigene perturbations with GEARS. Nat Biotechnol 42:927-935 (2024). doi:10.1038/s41587-023-01905-6", + "Viñas Torné 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", + "Li et al. A systematic comparison of single-cell perturbation response prediction models. Sci Adv 12:eaed3414 (2026). doi:10.1126/sciadv.aed3414", + "Law, Chen, Shi & Smyth. voom: precision weights unlock linear model analysis tools for RNA-seq read counts. Genome Biol 15:R29 (2014)" + ] + }, + { + "kind": "content", + "section": "4", + "section_label": "Fig. 4 candidate", + "text_source": { + "file": "data/sec09_cat_curves_slide_text.json", + "index": 0 + }, + "title": "Disease genes at every list length (CAT)", + "layout": "full", + "figures": [ + "figures/sec09_cat_curves.png" + ], + "keywords": [ + "LoRA top model: 3/4 (5 donors), 2/4 (5-fold)", + "Broadcast above LoRA: Giroux", + "Chance at k = 100: 0.012–0.016" + ], + "footnote": "Same-sign overlap of observed and generated top-k genes per cell type; no FDR cut-off.", + "sources": "tools/sec09_cat_curves.py; data/sec09_cat_curves.csv", + "refs": [ + "Irizarry 2005, Nat Methods", + "Webber 2010, ACM TOIS", + "Squair 2021, Nat Commun" + ], + "refs_full": [ + "Irizarry et al. Multiple-laboratory comparison of microarray platforms. Nat Methods 2:345-350 (2005). doi:10.1038/nmeth756", + "Webber, Moffat & Zobel. A similarity measure for indefinite rankings. ACM Trans Inf Syst 28(4):20 (2010). doi:10.1145/1852102.1852106", + "Squair et al. Confronting false discoveries in single-cell differential expression. Nat Commun 12:5692 (2021). doi:10.1038/s41467-021-25960-2" + ], + "footnote_full": "Correspondence 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); 5-fold = pooled call. Broadcast = the input real bulk ranked the same way. Chance = k / (2G), G = genes tested on both sides (0.012–0.016 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 / broadcast 0.069 / chance 0.012; 5-fold LoRA 0.163 / broadcast 0.075; Imperial CAT100 5 donors LoRA 0.243 / scSemi 0.133 / calibration 0.198 / raw 0.179 / broadcast 0.214 / chance 0.012; 5-fold LoRA 0.270 / broadcast 0.225; Giroux CAT100 5 donors LoRA 0.048 / scSemi 0.075 / calibration 0.050 / raw 0.048 / broadcast 0.176 / chance 0.016; 5-fold LoRA 0.068 / broadcast 0.145; Ji CAT100 5 donors LoRA 0.091 / scSemi 0.030 / calibration 0.021 / raw 0.021 / broadcast 0.043 / chance 0.014; 5-fold LoRA 0.152 / broadcast 0.068.", + "numbers": [ + { + "value": "Sepsis CAT100 5 donors LoRA 0.092 / scSemi 0.037 / calibration 0.069 / raw 0.028 / broadcast 0.069 / chance 0.012; 5-fold LoRA 0.163 / broadcast 0.075; Imperial CAT100 5 donors LoRA 0.243 / scSemi 0.133 / calibration 0.198 / raw 0.179 / broadcast 0.214 / chance 0.012; 5-fold LoRA 0.270 / broadcast 0.225; Giroux CAT100 5 donors LoRA 0.048 / scSemi 0.075 / calibration 0.050 / raw 0.048 / broadcast 0.176 / chance 0.016; 5-fold LoRA 0.068 / broadcast 0.145; Ji CAT100 5 donors LoRA 0.091 / scSemi 0.030 / calibration 0.021 / raw 0.021 / broadcast 0.043 / chance 0.014; 5-fold LoRA 0.152 / broadcast 0.068", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_contrast/fewshot_cat.csv; /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scores_contrast/pooled_cat.csv" + } + ] + }, + { + "kind": "divider", + "title": "Figure 5. Datasets outside training" + }, + { + "kind": "content", + "section": "plan", + "section_label": "Fig. 5 plan", + "text_source": { + "file": "data/figure_plan_slide_text.json", + "index": 5 + }, + "title": "Figure 5 plan: Datasets outside training", + "layout": "full", + "figures": [ + "figures/figplan_fig5.png" + ], + "keywords": [], + "footnote": "PBISC applies to cohorts it never saw, and needs their own pairs to do it well. 1 of 4 panels ready.", + "sources": "tools/figure_plan.py; data/figure_plan.json", + "footnote_full": "Panels: a) Other cohorts' pairs do not replace a cohort's own [Ready]. b) Behçet: disease direction from real bulk [Adapt; add the LoRA-3 result (C1Q-high BD > HC)]. c) Influenza exposure from raw clinical bulk (Randolph) [Build; manuscript result]. d) Diseases absent from the corpus vs present [Build; Kawasaki, SLE vs DRESS (2026-10-04)].", + "numbers": [] + }, + { + "kind": "content", + "section": "5", + "section_label": "Fig. 5a", + "layout": "full", + "sources": "tools/sec08_cross_cohort.py; data/sec08_cross_cohort.csv, data/sec08_cross_cohort_steps.csv; plans/paired_crosscohort_20261002/ (README, scores/aggregate.csv)", + "text_source": { + "file": "data/sec08_cross_cohort_slide_text.json", + "index": 0 + }, + "title": "Pairs from other cohorts do not transfer", + "figures": [ + "figures/sec08_cross_cohort.png" + ], + "keywords": [ + "Train on the other cohorts, test the held-out one", + "Calibration ties or beats them on composition in 4/5", + "Own-cohort pairs best on coverage 5/5, disease r 4/4" + ], + "footnote": "Leave-one-cohort-out, K = 1000; train on the other cohorts (+ Liu 2021), test the held-out one.", + "say": "Why one LoRA per cohort: a model trained on the other PBMC cohorts' pairs (three, or four for Ji, plus Liu 2021) rarely beats a zero-pair gene-wise calibration on composition, and no arm without the cohort's own pairs matches the own-cohort LoRA on coverage or disease signal.", + "footnote_full": "Each 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.", + "numbers": [ + { + "value": "ImMicro: L1 zs 0.446 / xcal 0.172 / lora16sat 0.175 / lora64sat 0.199 / fullsat 0.213 / own 0.162; coverage zs 0.19 / xcal 0.57 / lora16sat 0.45 / lora64sat 0.38 / fullsat 0.35 / own 0.78", + "source": "data/sec08_cross_cohort.csv" + }, + { + "value": "Sepsis: L1 zs 0.509 / xcal 0.444 / lora16sat 0.491 / lora64sat 0.570 / fullsat 0.585 / own 0.362; coverage zs 0.07 / xcal 0.46 / lora16sat 0.34 / lora64sat 0.29 / fullsat 0.35 / own 0.83; logFC r zs -0.02 / xcal 0.10 / lora16sat 0.10 / lora64sat 0.23 / fullsat 0.07 / own 0.53", + "source": "data/sec08_cross_cohort.csv" + }, + { + "value": "Imperial: L1 zs 0.803 / xcal 0.651 / lora16sat 0.635 / lora64sat 0.594 / fullsat 0.580 / own 0.376; coverage zs 0.11 / xcal 0.64 / lora16sat 0.68 / lora64sat 0.72 / fullsat 0.61 / own 0.85; logFC r zs 0.64 / xcal 0.69 / lora16sat 0.49 / lora64sat 0.58 / fullsat 0.46 / own 0.82", + "source": "data/sec08_cross_cohort.csv" + }, + { + "value": "Giroux: L1 zs 0.337 / xcal 0.202 / lora16sat 0.431 / lora64sat 0.410 / fullsat 0.465 / own 0.258; coverage zs 0.29 / xcal 0.89 / lora16sat 0.82 / lora64sat 0.81 / fullsat 0.78 / own 0.93; logFC r zs 0.34 / xcal 0.24 / lora16sat 0.30 / lora64sat 0.28 / fullsat 0.20 / own 0.46", + "source": "data/sec08_cross_cohort.csv" + }, + { + "value": "Ji (whole blood): L1 zs 0.725 / xcal 0.565 / lora16sat 0.740 / lora64sat 0.620 / fullsat 1.030 / own 0.504; coverage zs 0.06 / xcal 0.07 / lora16sat 0.34 / lora64sat 0.40 / fullsat 0.08 / own 0.86; logFC r zs 0.07 / xcal 0.13 / lora16sat 0.05 / lora64sat 0.16 / fullsat 0.14 / own 0.48", + "source": "data/sec08_cross_cohort.csv" + } + ] + }, + { + "kind": "content", + "section": "5", + "section_label": "Fig. 5a", + "layout": "full", + "sources": "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_/K1000//RUN.json)", + "text_source": { + "file": "data/sec08_cross_cohort_slide_text.json", + "index": 1 + }, + "title": "Longer training on other cohorts: gap stays", + "figures": [ + "figures/sec08_cross_cohort_steps.png" + ], + "keywords": [ + "To saturation: r16 20–31 min, full 59–107 min", + "Full fine-tune drifts in Sepsis and Ji (Ji L1 0.84 → 1.03)", + "So PBISC fits LoRA on a few pairs per cohort" + ], + "footnote": "Test metrics at every 20k-step milestone (best inner-validation state so far). K = 1000.", + "say": "Each run trained until its inner-validation loss saturated. More steps on other cohorts never closed the coverage gap to the cohort's own pairs: Imperial and Giroux improved a little, and full fine-tuning drifted further in Sepsis and Ji (whole-blood bulk).", + "footnote_full": "Same 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).", + "numbers": [ + { + "value": "ImMicro LoRA r16: 20k L1 0.163 cov 0.45, 28k L1 0.175 cov 0.45", + "source": "data/sec08_cross_cohort_steps.csv" + }, + { + "value": "ImMicro LoRA r64: 20k L1 0.196 cov 0.44, 39k L1 0.199 cov 0.38", + "source": "data/sec08_cross_cohort_steps.csv" + }, + { + "value": "ImMicro Full fine-tune: 20k L1 0.208 cov 0.39, 40k L1 0.211 cov 0.37, 60k L1 0.223 cov 0.35, 65k L1 0.213 cov 0.35", + "source": "data/sec08_cross_cohort_steps.csv" + }, + { + "value": "Sepsis LoRA r16: 20k L1 0.500 cov 0.35, 26k L1 0.491 cov 0.34", + "source": "data/sec08_cross_cohort_steps.csv" + }, + { + "value": "Sepsis LoRA r64: 20k L1 0.550 cov 0.29, 40k L1 0.564 cov 0.29, 59k L1 0.570 cov 0.29", + "source": "data/sec08_cross_cohort_steps.csv" + }, + { + "value": "Sepsis Full fine-tune: 20k L1 0.498 cov 0.33, 40k L1 0.530 cov 0.33, 60k L1 0.546 cov 0.34, 80k L1 0.576 cov 0.34, 100k L1 0.577 cov 0.35, 108k L1 0.585 cov 0.35", + "source": "data/sec08_cross_cohort_steps.csv" + }, + { + "value": "Imperial LoRA r16: 20k L1 0.657 cov 0.68, 40k L1 0.635 cov 0.68", + "source": "data/sec08_cross_cohort_steps.csv" + }, + { + "value": "Imperial LoRA r64: 20k L1 0.622 cov 0.68, 40k L1 0.603 cov 0.71, 60k L1 0.577 cov 0.72, 69k L1 0.594 cov 0.72", + "source": "data/sec08_cross_cohort_steps.csv" + }, + { + "value": "Imperial Full fine-tune: 20k L1 0.632 cov 0.60, 40k L1 0.609 cov 0.60, 60k L1 0.565 cov 0.61, 80k L1 0.576 cov 0.60, 93k L1 0.580 cov 0.61", + "source": "data/sec08_cross_cohort_steps.csv" + }, + { + "value": "Giroux LoRA r16: 20k L1 0.416 cov 0.79, 35k L1 0.431 cov 0.82", + "source": "data/sec08_cross_cohort_steps.csv" + }, + { + "value": "Giroux LoRA r64: 20k L1 0.491 cov 0.80, 40k L1 0.469 cov 0.82, 60k L1 0.430 cov 0.82, 69k L1 0.410 cov 0.81", + "source": "data/sec08_cross_cohort_steps.csv" + }, + { + "value": "Giroux Full fine-tune: 20k L1 0.543 cov 0.75, 40k L1 0.532 cov 0.75, 60k L1 0.499 cov 0.76, 80k L1 0.466 cov 0.76, 100k L1 0.490 cov 0.76, 120k L1 0.465 cov 0.78", + "source": "data/sec08_cross_cohort_steps.csv" + }, + { + "value": "Ji (whole blood) LoRA r16: 20k L1 0.757 cov 0.27, 40k L1 0.740 cov 0.34", + "source": "data/sec08_cross_cohort_steps.csv" + }, + { + "value": "Ji (whole blood) LoRA r64: 20k L1 0.639 cov 0.38, 40k L1 0.668 cov 0.39, 51k L1 0.620 cov 0.40", + "source": "data/sec08_cross_cohort_steps.csv" + }, + { + "value": "Ji (whole blood) Full fine-tune: 20k L1 0.838 cov 0.17, 40k L1 1.006 cov 0.11, 60k L1 1.032 cov 0.08, 80k L1 1.052 cov 0.07, 86k L1 1.030 cov 0.08", + "source": "data/sec08_cross_cohort_steps.csv" + }, + { + "value": "Training time to saturation, LoRA r16: ImMicro 21 min, Sepsis 20 min, Imperial 30 min, Giroux 28 min, Ji 31 min", + "source": "data/sec08_cross_cohort.csv (minutes = RUN.json wall_seconds / 60)" + }, + { + "value": "Training time to saturation, LoRA r64: ImMicro 27 min, Sepsis 44 min, Imperial 54 min, Giroux 53 min, Ji 41 min", + "source": "data/sec08_cross_cohort.csv (minutes = RUN.json wall_seconds / 60)" + }, + { + "value": "Training time to saturation, Full fine-tune: ImMicro 59 min, Sepsis 93 min, Imperial 82 min, Giroux 107 min, Ji 77 min", + "source": "data/sec08_cross_cohort.csv (minutes = RUN.json wall_seconds / 60)" + } + ] + }, + { + "kind": "content", + "section": "5", + "section_label": "Fig. 5b", + "title": "Consequence: disease direction flips", + "layout": "figure_keywords", + "figures": [ + "figures/sec05_behcet_direction.png" + ], + "keywords": [ + "CD16 gap: raw −3.97 vs observed +4.94", + "C1Q flip vanishes at common depth", + "sc pseudobulk: right sign in 5/5 seeds" + ], + "footnote": "Behçet. Bulk GSE198533 (9 BD / 10 HC); sc GSE198616 (4 / 4, different people) gives observed and pseudobulk.", + "say": "Raw bulk gives the wrong sign for CD16 and C1Q in 5/5 seeds. At one common monocyte depth the C1Q gap is −0.08 ± 0.09 pp (1/5 seeds positive), so the C1Q flip is a depth effect; the CD16 flip stays at −3.59. sc pseudobulk of the 8 sc donors gives the observed sign in 5/5 seeds (CD16 +5.48, C1Q +1.82). Observed p is descriptive (4 vs 4). No paired data in this cohort, so no fine-tune here.", + "sources": "tools/sec05_behcet_direction.py; data/sec05_behcet_direction.csv", + "footnote_full": "Behçet, final model. Bulk GSE198533 (BD 9, HC 10), 5,000 cells per sample; observed = sc GSE198616 (BD 4, HC 4, different people; descriptive p). At common depth, raw CD16 is −3.59 (seed 0) and raw C1Q is −0.08 ± 0.09 pp." + }, + { + "kind": "divider", + "title": "Figure 6. Discovery from public bulk-only data" + }, + { + "kind": "content", + "section": "plan", + "section_label": "Fig. 6 plan", + "text_source": { + "file": "data/figure_plan_slide_text.json", + "index": 6 + }, + "title": "Figure 6 plan: Discovery from public bulk-only data", + "layout": "full", + "figures": [ + "figures/figplan_fig6.png" + ], + "keywords": [], + "footnote": "Analysis not started; to be planned.", + "sources": "tools/figure_plan.py; data/figure_plan.json", + "footnote_full": "Panels: Slides kept in this section until it is planned: Next: cohorts with a paired subset (planned).", + "numbers": [] + }, + { + "kind": "content", + "section": "6", + "section_label": "Fig. 6 candidate", + "title": "Next: cohorts with a paired subset (planned)", + "layout": "figure_keywords", + "figures": [ + "figures/sec10_paired_subset_cohorts.png" + ], + "keywords": [ + "Train LoRA on pairs, generate bulk-only", + "Sepsis controls: 21 paired + 54 bulk-only", + "ChILI pre-ICI: 11 bulk, 4 unpaired sc" + ], + "footnote": "Counts from pairing manifests. Model: f01 6k 1x E16 + LoRA r16.", + "say": "Planned: no LoRA cells exist yet for any bulk-only sample. Ji bulk is PAXgene whole blood. Only 7 of the Arunachalam pairs are strict same-visit pairs. Asan is omitted because it has no donor map.", + "sources": "tools/sec10_paired_subset_cohorts.py; data/sec10_cohort_counts.csv, data/sec10_bulk_only_groups.csv, data/sec10_quoted_numbers.csv", + "footnote_full": "Counts from pairing manifests. QC = identity check (Imperial also >=200 cells); sex-flag = bulk sex differs from metadata or is ambiguous. Not shown: Asan (no donor map), Giroux (24 pairs, 17 bulk-only; broadcast beats LoRA). Model: f01 6k 1x E16 + LoRA r16." + }, + { + "kind": "divider", + "title": "Supplementary" + }, + { + "kind": "content", + "section": "plan", + "section_label": "Supplementary", + "text_source": { + "file": "data/figure_plan_slide_text.json", + "index": 7 + }, + "title": "Supplementary plan", + "layout": "full", + "figures": [ + "figures/figplan_supp.png" + ], + "keywords": [], + "footnote": "Bulk deconvolution methods appear here only: PBISC generates single cells, a different task.", + "sources": "tools/figure_plan.py; data/figure_plan.json", + "footnote_full": "Supplementary: No pretraining: same cohort, from scratch; No pretraining: other cohorts, from scratch; Rare states: from scratch vs pretrained. Planned: Bulk deconvolution methods (CYBERSORTx and others), for reference [Run; different task: PBISC generates single cells]. Appendix: Appendix: scSemiProfiler without output threshold.", + "numbers": [] + }, + { + "kind": "content", + "section": "S", + "section_label": "Supplementary", + "text_source": { + "file": "data/sec07_no_pretraining_slide_text.json", + "index": 0 + }, + "title": "No pretraining: same cohort, from scratch", + "layout": "full", + "figures": [ + "figures/sec07_no_pretraining.png" + ], + "keywords": [ + "Same realism; 5 pairs: donor r 0.70 vs 0.17", + "58 pairs: scratch logFC r 0.58 vs 0.49", + "Scratch 61–86 min, LoRA 3–5 min" + ], + "footnote": "Pilot: sepsis fold 0, 17 test donors (12 sepsis, 5 controls). Folds 1–4 not run yet.", + "sources": "tools/sec07_no_pretraining.py; data/sec07_no_pretraining.csv; plans/paired_finetune_intradataset_20260927/scores/ye2026_sepsis/fold0_scratch_20261005/", + "say": "Pilot on one sepsis fold. Training the same model from scratch on the cohort's pairs gives realistic cells, and with 58 pairs a higher disease logFC r, but with few pairs it cannot place donors: what pretraining buys is the donor-specific composition and minutes instead of an hour per cohort. Folds 1-4 are needed before any claim.", + "footnote_full": "Pilot 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'.", + "numbers": [ + { + "value": "5 pairs: pretrained+LoRA L1 0.363 / scratch 0.448; coverage 0.81 / 0.80; logFC r 0.23 / 0.18; donor composition r 0.70 / 0.17; time 3.2 / 61 min; steps (selected/run) 2750/4750 vs 66000/76000", + "source": "data/sec07_no_pretraining.csv" + }, + { + "value": "58 pairs: pretrained+LoRA L1 0.373 / scratch 0.398; coverage 0.86 / 0.86; logFC r 0.49 / 0.58; donor composition r 0.77 / 0.68; time 5.4 / 79 min; steps (selected/run) 6500/8500 vs 100000/100000", + "source": "data/sec07_no_pretraining.csv" + }, + { + "value": "58 + recipes pairs: pretrained+LoRA L1 0.385 / scratch 0.385; coverage 0.87 / 0.88; logFC r 0.45 / 0.54; donor composition r 0.80 / 0.77; time 4.5 / 86 min; steps (selected/run) 3750/5750 vs 100000/100000", + "source": "data/sec07_no_pretraining.csv" + } + ] + }, + { + "kind": "content", + "section": "S", + "section_label": "Supplementary", + "text_source": { + "file": "data/sec08_cross_scratch_slide_text.json", + "index": 0 + }, + "title": "No pretraining: other cohorts, from scratch", + "layout": "full", + "figures": [ + "figures/sec08_cross_scratch.png" + ], + "keywords": [ + "Scratch ≈ LoRA r16: coverage 0.33 vs 0.34", + "Own-cohort pairs far ahead: coverage 0.83", + "Scratch 127 min vs LoRA r16 20 min" + ], + "footnote": "Pilot: held-out sepsis, 82 samples. Genes no training cohort measured are zeroed in every arm.", + "sources": "tools/sec08_cross_scratch.py; data/sec08_cross_scratch.csv; plans/paired_crosscohort_20261002/scores_masked/ye2026_sepsis/, generated_masked/ye2026_sepsis/hoL_ye2026_sepsis/*.log, runs/hoL_ye2026_sepsis/K1000/*/RUN.json; plans/paired_finetune_intradataset_20260927/data/ye2026_sepsis/cells_panel6k.npz (observed share)", + "say": "Held-out sepsis pilot. Training from scratch on the other cohorts' pairs ends where the pretrained arms end: no arm without sepsis's own pairs gets near the own-cohort LoRA. Genes that no training cohort measured are zeroed in every arm, because the scratch model never learned them.", + "footnote_full": "Pilot, one held-out cohort: sepsis (82 samples), training pool ImMicro + Imperial + Giroux + Liu 2021, K = 1,000, same inner-val selection and sampling as the 'Pairs from other cohorts do not transfer' and 'Longer training on other cohorts' slides. From scratch = the same denoiser at random initialisation (--init random), full training at lr 2e-4, inner validation every 1,000 steps (patience 10), milestones every 50k steps without the 1% saturation rule; stopped at 162k (selected 152k; 127 min on one GB10). The pretrained arms: inner validation every 250 steps (patience 8), 20k-step blocks with the 1% rule; all four runs ended by patience early stopping. Every arm is scored with the 46 panel genes that no training cohort measured set to 0 in the generated cells (tools/mask_rescore.py): the scratch model never gets a loss on them and puts ~94% of its UMI there (91–94% across milestones; observed sepsis cells 0.014%); the pretrained arms barely change (e.g. LoRA r16 coverage 0.34 masked vs 0.34 unmasked). Calibration needs no training. Own-cohort LoRA = 5-fold LoRA on sepsis pairs (mean 6.3 min per fold). Scratch milestones: L1 0.492 / 0.497 / 0.510 / 0.499; coverage 0.339 / 0.343 / 0.330 / 0.328; logFC r 0.247 / 0.245 / 0.247 / 0.229. Source: plans/paired_crosscohort_20261002/scores_masked/ye2026_sepsis/ (table_masked.py), runs/hoL_ye2026_sepsis/K1000/.", + "numbers": [ + { + "value": "Calibration: L1 0.444, coverage 0.461, logFC r 0.096, top-50 DEG 10, time none", + "source": "data/sec08_cross_scratch.csv" + }, + { + "value": "LoRA r16: L1 0.491, coverage 0.345, logFC r 0.106, top-50 DEG 4, time 20 min", + "source": "data/sec08_cross_scratch.csv" + }, + { + "value": "LoRA r64: L1 0.568, coverage 0.284, logFC r 0.226, top-50 DEG 5, time 44 min", + "source": "data/sec08_cross_scratch.csv" + }, + { + "value": "Full FT: L1 0.585, coverage 0.340, logFC r 0.084, top-50 DEG 3, time 93 min", + "source": "data/sec08_cross_scratch.csv" + }, + { + "value": "From scratch: L1 0.499, coverage 0.328, logFC r 0.229, top-50 DEG 1, time 127 min", + "source": "data/sec08_cross_scratch.csv" + }, + { + "value": "Own-cohort LoRA: L1 0.362, coverage 0.828, logFC r 0.526, top-50 DEG 28, time 6 min", + "source": "data/sec08_cross_scratch.csv" + }, + { + "value": "Scratch milestones (masked) 50k / 100k / 150k / 162k steps: L1 0.492 / 0.497 / 0.510 / 0.499; coverage 0.339 / 0.343 / 0.330 / 0.328; logFC r 0.247 / 0.245 / 0.247 / 0.229", + "source": "data/sec08_cross_scratch.csv" + } + ] + }, + { + "kind": "content", + "section": "S", + "section_label": "Supplementary", + "text_source": { + "file": "data/secS_scratch_rare_states_slide_text.json", + "index": 0 + }, + "title": "Rare states: from scratch vs pretrained", + "layout": "full", + "figures": [ + "figures/secS_scratch_rare_states.png" + ], + "keywords": [ + "Same pairs: no significant difference on average", + "Cycling T/NK: pretrained closer, 8 of 8", + "C1Q-high needs the cohort's own pairs" + ], + "footnote": "Pilot, one seed: sepsis fold 0 (17 donors), held-out sepsis (82, masked). 1 = observed rate; grey = within 2-fold.", + "say": "Supplementary pilot: one sepsis fold and one held-out cohort, one seed. With the same pairs, the model trained from scratch shows no significant difference from the pretrained model on average over 13 rare states, within about plus or minus 0.3 log2. Pretraining stays closer on cycling T and NK cells, in both panels; scratch is closer on DN2 B cells and where the pretrained model over-calls labels. In this fold, C1Q-high monocytes reach about half the observed rate or more only with the cohort's own pairs, with either initialisation.", + "sources": "plans/rare_states_screen_20261004/out/{sepsis_fold0,sepsis_loco_masked}_*.csv; plans/paired_finetune_intradataset_20260927/data/ye2026_sepsis/folds.csv; plans/paired_finetune_intradataset_20260927/runs/ye2026_sepsis/fold0/scratch_*/RUN.json; plans/paired_crosscohort_20261002/generated_masked/ye2026_sepsis/hoL_ye2026_sepsis/*.log; tools/secS_scratch_rare_states.py; data/secS_scratch_rare_states{,_bootstrap}.csv", + "footnote_full": "Supplementary. Rare-state screen (plans/rare_states_screen_20261004): 13 rare states plus two controls, marker rules from the literature (tools/states.json) or scANVI labels; per arm, generated state % of its parent / observed % (1 = same rate; grey band = within 2-fold). For marker states the observed cells are thinned to the arm's generated depth; the two label states (plasmablast, proliferating T) are not thinned. The figure shows 8 of the 13 states; every average uses all 13 (data/secS_scratch_rare_states_bootstrap.csv). Same cohort: sepsis fold 0, 17 held-out donors, the same training pairs, conditions and sampling seed for both initialisations; from scratch = random initialisation, full training (lr 2e-4, cap 100k steps). Other cohorts: all 82 sepsis donors held out, trained on ImMicro + Imperial + Giroux + Liu pairs; the 46 genes that no training cohort measured are zeroed in every arm (the scratch model is never trained on them). Donor bootstrap of D = |log2| scratch - |log2| pretrained, mean of the 13 rare states (D > 0 = pretrained closer): same cohort 5 pairs +0.22 [-0.08, +0.41]; 58 pairs +0.02 [-0.35, +0.25]; 58 pairs, full FT +0.14 [-0.14, +0.41]; 58 pairs + recipes -0.09 [-0.38, +0.18]; other cohorts, scratch final vs LoRA r16 -0.04 [-0.20, +0.18]; LoRA r64 +0.06 [-0.08, +0.22]; Full FT +0.29 [+0.07, +0.57]; Calibration +0.72 [+0.57, +0.95]; scratch 50k vs LoRA r16 -0.31 [-0.45, -0.11]; LoRA r64 -0.21 [-0.37, -0.02]; Full FT +0.02 [-0.19, +0.30]; Calibration +0.44 [+0.30, +0.66]. A non-significant average is not a shown equivalence (intervals about +/-0.3 log2). Cycling T/NK: pretrained closer in 4 of 4 same-cohort settings (1 significant) and 4 of 4 other-cohort comparisons (4 significant; scratch 0.08x). ISG-high T: pretrained closer in 4 of 4 same-cohort settings (1 significant) but scratch closer in 4 of 4 other-cohort comparisons (0 significant; scratch 1.10x vs pretrained 1.54–1.98x). Same-cohort state-level calls (13 states x 4 settings): pretrained closer 3, scratch closer 7 of 52; of the scratch calls 4 are label over-calls by the pretrained arms (Plasmablast, Proliferating T), 2 are states where both arms are more than 4-fold off (Granulocyte-like), and 1 is DN2 B (58 pairs + recipes -0.70 [-1.21, -0.08]); scratch is also closer on MS1 at 58 pairs (-0.97 [-1.44, +0.03], n.s.). C1Q-high: 0.47x or more with the cohort's own pairs (both initialisations), 0.10x or less without them. Fold 0 is the own-pair LoRA's best fold for C1Q-high and cycling T/NK (58 pairs: 1.03x and 1.00x vs 0.27–0.82x and 0.32–0.64x in folds 1–4); scratch has not been run on folds 1–4. One fold, one held-out cohort, one seed; no multiplicity correction. Full analysis: plans/rare_states_screen_20261004/README.md and out/scratch_vs_lora_20261005/REPORT.md (two verification passes on the same-cohort and unmasked analyses); the masked cross-cohort numbers here come from the REPORT addendum and masked_cross_bootstrap.csv, which supersede the unmasked cross-cohort readings in the REPORT body.", + "numbers": [ + { + "value": "C1Q-high mono: same cohort 5 pairs LoRA 1.07 / scratch 1.01, 58 pairs LoRA 1.05 / scratch 0.47, 58 pairs + recipes LoRA 0.63 / scratch 0.53; other cohorts Own-cohort LoRA (reference) 0.51, Calibration 0.08, LoRA r16 0.04, Full FT 0.08, From scratch 0.07", + "source": "data/secS_scratch_rare_states.csv" + }, + { + "value": "MS1 mono: same cohort 5 pairs LoRA 0.15 / scratch 0.10, 58 pairs LoRA 0.21 / scratch 0.40, 58 pairs + recipes LoRA 0.30 / scratch 0.46; other cohorts Own-cohort LoRA (reference) 0.72, Calibration 0.73, LoRA r16 0.12, Full FT 0.27, From scratch 0.16", + "source": "data/secS_scratch_rare_states.csv" + }, + { + "value": "Cycling mono: same cohort 5 pairs LoRA 0.11 / scratch 0.06, 58 pairs LoRA 0.09 / scratch 0.09, 58 pairs + recipes LoRA 0.10 / scratch 0.09; other cohorts Own-cohort LoRA (reference) 0.24, Calibration 0.26, LoRA r16 0.07, Full FT 0.12, From scratch 0.04", + "source": "data/secS_scratch_rare_states.csv" + }, + { + "value": "Cycling T/NK: same cohort 5 pairs LoRA 1.15 / scratch 0.51, 58 pairs LoRA 0.93 / scratch 0.42, 58 pairs + recipes LoRA 0.99 / scratch 0.48; other cohorts Own-cohort LoRA (reference) 0.52, Calibration 0.48, LoRA r16 0.13, Full FT 0.36, From scratch 0.08", + "source": "data/secS_scratch_rare_states.csv" + }, + { + "value": "ISG-high T: same cohort 5 pairs LoRA 0.46 / scratch 0.15, 58 pairs LoRA 0.37 / scratch 0.17, 58 pairs + recipes LoRA 0.53 / scratch 0.18; other cohorts Own-cohort LoRA (reference) 0.58, Calibration 1.54, LoRA r16 1.71, Full FT 1.98, From scratch 1.10", + "source": "data/secS_scratch_rare_states.csv" + }, + { + "value": "Plasmablast (label): same cohort 5 pairs LoRA 1.87 / scratch 1.12, 58 pairs LoRA 3.47 / scratch 1.24, 58 pairs + recipes LoRA 2.30 / scratch 1.18; other cohorts Own-cohort LoRA (reference) 0.76, Calibration 0.78, LoRA r16 1.38, Full FT 2.21, From scratch 0.87", + "source": "data/secS_scratch_rare_states.csv" + }, + { + "value": "DN2 B: same cohort 5 pairs LoRA 1.11 / scratch 1.63, 58 pairs LoRA 0.35 / scratch 0.63, 58 pairs + recipes LoRA 0.28 / scratch 0.46; other cohorts Own-cohort LoRA (reference) 0.29, Calibration 0.28, LoRA r16 0.14, Full FT 0.29, From scratch 0.27", + "source": "data/secS_scratch_rare_states.csv" + }, + { + "value": "Granulocyte-like: same cohort 5 pairs LoRA 0.05 / scratch 0.13, 58 pairs LoRA 0.04 / scratch 0.04, 58 pairs + recipes LoRA 0.03 / scratch 0.06; other cohorts Own-cohort LoRA (reference) 0.06, Calibration 0.07, LoRA r16 0.02, Full FT 0.01, From scratch 0.02", + "source": "data/secS_scratch_rare_states.csv" + } + ] + }, + { + "kind": "divider", + "title": "Appendix" + }, + { + "kind": "content", + "section": "A", + "section_label": "Appendix", + "text_source": { + "file": "data/secA_scsemi_readout_slide_text.json", + "index": 0 + }, + "title": "Appendix: scSemiProfiler without output threshold", + "layout": "full", + "figures": [ + "figures/secA_scsemi_readout.png" + ], + "keywords": [ + "Official output zeroes values ≤ 10 (fixed)", + "Removed: scSemi coverage up in 5/5 cohorts", + "No LoRA-5 comparison flips (23/23 same)" + ], + "footnote": "Modified read-out, not scSemiProfiler's official output. Main slides use the official one.", + "say": "Appendix only (sensitivity check). The threshold is part of scSemiProfiler's official output, so removing it is a modified run; the main comparison stays the official output. This slide shows that our conclusions do not depend on that read-out.", + "sources": "plans/paired_finetune_intradataset_20260927/scsemi/tools/{resample_output_variants,compare_output_variants}.py; tools/secA_scsemi_readout.py; data/secA_scsemi_readout.csv", + "footnote_full": "Appendix, 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').", + "numbers": [ + { + "value": "ImMicro: coverage LoRA-5 0.79, official 0.04, no threshold 0.23; precision 0.95 / 0.52 / 0.74; L1 0.176 / 0.233 / 0.199", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scsemi/analysis/output_variants_20261003/per_draw.csv" + }, + { + "value": "Sepsis: coverage LoRA-5 0.69, official 0.08, no threshold 0.18; precision 0.91 / 0.73 / 0.77; L1 0.466 / 0.612 / 0.591; logFC r 0.35 / 0.40 / 0.46; top-50 DEG 16.4 / 5.2 / 14.0", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scsemi/analysis/output_variants_20261003/per_draw.csv" + }, + { + "value": "Imperial: coverage LoRA-5 0.74, official 0.06, no threshold 0.42; precision 0.85 / 0.33 / 0.72; L1 0.503 / 0.602 / 0.603; logFC r 0.75 / 0.56 / 0.66; top-50 DEG 113.6 / 46.8 / 94.6", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scsemi/analysis/output_variants_20261003/per_draw.csv" + }, + { + "value": "Giroux: coverage LoRA-5 0.96, official 0.61, no threshold 0.83; precision 1.00 / 0.99 / 1.00; L1 0.256 / 0.270 / 0.273; logFC r 0.30 / 0.40 / 0.45; top-50 DEG 0.2 / 0.0 / 0.0", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scsemi/analysis/output_variants_20261003/per_draw.csv" + }, + { + "value": "Ji: coverage LoRA-5 0.71, official 0.15, no threshold 0.53; precision 0.98 / 0.84 / 0.92; L1 0.513 / 0.592 / 0.575; logFC r 0.33 / 0.15 / 0.19; top-50 DEG 0.2 / 0.0 / 0.0", + "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927/scsemi/analysis/output_variants_20261003/per_draw.csv" + } + ] + }, + { + "kind": "divider", + "title": "Discussion" + }, + { + "kind": "content", + "section": "D", + "section_label": "Discussion", + "text_source": { + "file": "data/sec10_other_tissues_slide_text.json", + "index": 0 + }, + "title": "Next: other tissues from a few dozen pairs", + "layout": "full", + "figures": [ + "figures/sec10_other_tissues.png" + ], + "keywords": [ + "Pretrained: a 3-min LoRA works from 5 pairs", + "Scratch: 58 pairs + recipes come close (86 min)", + "Next: test this route on other tissues" + ], + "footnote": "Pilots: sepsis fold 0 (17 donors); last row held-out sepsis (82). Folds 1–4 and a second cohort pending.", + "say": "Pretraining is what makes PBISC practical: a LoRA of a few minutes on as few as five pairs. In a one-fold sepsis pilot, though, with enough compute, a few dozen paired bulk and single-cell samples plus pseudobulk recipes came close to the same results without corpus pretraining, and the outcome depends on the training protocol. This does not weaken the pretraining result. It suggests that a PBISC-like bulk-to-single-cell model could be built for other organs and tissues that have no large single-cell corpus yet. Folds 1 to 4 and a second cohort will test this first.", + "sources": "tools/sec10_other_tissues.py; data/sec10_other_tissues.csv (from sec07_no_pretraining.csv, sec08_cross_scratch.csv, secS_scratch_rare_states_bootstrap.csv, sec06_ft_method.csv, sec01_model_numbers.csv train_split_cells / corpus_cells); plans/paired_finetune_intradataset_20260927/runs/ye2026_sepsis/fold0/*/RUN.json; plans/paired_crosscohort_20261002/generated_masked/ye2026_sepsis/hoL_ye2026_sepsis/locoL_scratchsat_*.log", + "footnote_full": "Discussion point; the from-scratch evidence is in the Supplementary. Same cohort, same pairs: sepsis fold 0, 17 held-out donors; every run also early-stops on the same 7 inner-validation pairs. From scratch = the same denoiser at random initialisation, full training (lr 2e-4, inner-val selection, cap 100k steps; the 58-pair and 58-pair + recipes runs were still improving at the cap). It still took the architecture, the diffusion settings, the 6k gene panel, the value scaler and the bulk z-score constants from the PBMC checkpoint, so a corpus-free setup (panel and scalers built from the pairs alone) is untested. Recipes = 50% of each batch is pseudobulk built from the training donors' own cells, K from the pretraining grid (250–4,000). Bold = higher by 0.05 or more (one fold, no interval), or less training time per cohort; the pretraining row is not scored. Rare states: mean |log2 gen/obs| over 13 states, scratch minus pretrained -0.09 [-0.38, +0.18] (donor bootstrap, n.s.); per-state calls = significant, marker states only, excluding states that both arms miss by more than 4-fold. Protocol dependence: recipes, the step budget, and the gene space - a scratch model learns only genes measured in its training pairs (trained on other cohorts, it put 91–94% of its UMI on the 46 genes that none of those cohorts measured). Pairs from other cohorts do not replace the target cohort's own: own-cohort LoRA coverage 0.83 on held-out sepsis. Pilot only: folds 1–4 and a second cohort are needed before the claim. Pretraining (15.9 h on 33.1 M train cells of the 41.4 M-cell corpus) is done once and reused by every cohort.", + "numbers": [ + { + "value": "Corpus pretraining: pretrained 33.1 M train cells, 15.9 h, once / scratch none", + "source": "data/sec01_model_numbers.csv train_split_cells; sec06_ft_method.csv pretrain_wall_hours" + }, + { + "value": "Training, sepsis fold 0 (one GPU): pretrained 3–5 min / scratch 61–86 min", + "source": "data/sec07_no_pretraining.csv minutes" + }, + { + "value": "5 pairs: donor composition r: pretrained 0.70 / scratch 0.17", + "source": "data/sec07_no_pretraining.csv comp_r" + }, + { + "value": "58 pairs + recipes: donor composition r: pretrained 0.80 / scratch 0.77", + "source": "data/sec07_no_pretraining.csv comp_r" + }, + { + "value": "58 pairs + recipes: coverage: pretrained 0.87 / scratch 0.88", + "source": "data/sec07_no_pretraining.csv coverage" + }, + { + "value": "58 pairs + recipes: disease logFC r: pretrained 0.45 / scratch 0.54", + "source": "data/sec07_no_pretraining.csv logfc_r" + }, + { + "value": "58 pairs + recipes: 13 rare states, n.s. on average: pretrained closer on ISG-high T / scratch closer on DN2 B", + "source": "data/secS_scratch_rare_states_bootstrap.csv" + }, + { + "value": "Other cohorts' pairs only (sepsis held out): coverage: pretrained 0.34 (LoRA r16) / scratch 0.33", + "source": "data/sec08_cross_scratch.csv coverage (masked)" + } + ] + } + ] +} diff --git a/paper-releases/pbisc-diffusion/updates/20261006T150739Z-deck-figure-plan-decisions/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/figures/figplan_supp.png b/paper-releases/pbisc-diffusion/updates/20261006T150739Z-deck-figure-plan-decisions/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/figures/figplan_supp.png new file mode 100644 index 0000000000000000000000000000000000000000..80eb581de8c4c5126651ddd5a3c255b757072037 --- /dev/null +++ b/paper-releases/pbisc-diffusion/updates/20261006T150739Z-deck-figure-plan-decisions/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/figures/figplan_supp.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88203a76398a932cafaeca304ae9a71eb5834a6a6930332d787a60d8d4c4b282 +size 170373