Archive 4 checkpoints (1-4)

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

@@ -0,0 +1,273 @@
+ 7  none  fig 12.29x5.70 in  Experimental overview
+ 8  none  fig 12.45x5.67 in  Pretraining data: 33.14 million training cells
+ 9  none  fig 12.45x5.67 in  Full source corpus: training, validation and test
+10  none  fig 12.45x5.67 in  From 41.40M source cells to 33.14M training cells
+11  none  fig 12.45x5.67 in  Recipe K and training exposure are different counts
+12  below fig 12.38x5.33 in  Corpus scale and generation workflow
+13  none  fig 12.45x4.07 in  From Gaussian noise to generated cell states
+16  none  fig 12.45x5.67 in  Where context width and trunk width change
+17  none  fig 12.45x5.67 in  Predicted and oracle hints: one trained denoiser
+18  none  fig 12.45x5.67 in  Architecture: full and late validation progression
+19  below fig  9.57x4.63 in  Width & context ablation: benchmark
+20  none  fig 12.45x5.67 in  Architecture arms: dimensions and endpoints
+21  below fig  9.57x4.63 in  Gene context: 2k / 3k / 6k / 15k
+22  none  fig 12.45x5.67 in  Gene panels: progression relative to each endpoint
+23  none  fig 12.45x5.67 in  Gene-panel arms: exact sizes and training provenance
+24  none  fig 12.45x5.67 in  All ablation arms: parameters and training time
+25  none  fig 12.15x5.70 in  Real-cell repeat: a sampling reference
+26  none  fig 12.45x5.48 in  Native-pseudobulk model comparisons
+29  none  fig 11.76x5.70 in  Bulk and pooled cells: different expression values
+30  none  fig 12.45x5.19 in  Paired calibration reduces same-donor expression error
+31  below fig 10.47x4.63 in  Consequence: cells off-manifold
+32  none  fig 12.33x5.70 in  Adaptation recovers the observed cell mixture
+33  none  fig 12.45x5.35 in  Adaptation: mixture error and expression agreement
+34  none  fig 12.15x5.70 in  Healthy paired data: full FT and LoRA
+35  none  fig 12.45x5.40 in  Full FT and LoRA ranks: quality and parameters
+36  none  fig 11.45x5.70 in  What does “all pairs” actually mean?
+37  none  fig 12.19x5.70 in  Healthy pair budget: correlation in all cell types
+38  none  fig 12.19x5.70 in  Healthy pair budget: error in all cell types
+39  right fig  9.20x5.17 in  Five pairs, no extra validation: LoRA vs scSemi
+40  none  fig 12.17x5.70 in  Composition: error and donor-level wins
+43  right fig  9.20x5.18 in  Disease signals with five training donors
+44  right fig  9.20x4.80 in  5-fold CV: disease signal
+45  below fig 11.93x4.63 in  Disease genes at every list length (CAT)
+46  below fig 10.54x4.63 in  Cell-type deconvolution, same pairs: disease signal
+47  none  fig 12.45x5.70 in  Disease gene rankings: complete CV comparison
+48  none  fig 11.99x5.70 in  Disease integrity depends on cohort and endpoint
+49  none  fig 12.19x5.70 in  Five-pair Sepsis: metrics across all support draws
+52  none  fig 12.33x5.70 in  Figure 5 targets begin with the source paper
+53  none  fig 12.18x5.70 in  Ye sepsis: published markers in paired scRNA
+54  none  fig 12.33x5.70 in  Transcriptional programs and rare-population identity
+60  none  fig 12.45x2.62 in  Corpus reference: assay and source composition
+61  below fig 12.45x4.61 in  Backbone: SDD → PBISC-Diffusion
+62  below fig 12.45x4.61 in  Bulk conditioning
+63  below fig 11.68x5.33 in  Denoiser: joint value + support diffusion
+64  below fig 12.45x4.71 in  Fine-tuning with paired bulk + sc
+65  below fig 12.35x4.63 in  Evaluation design
+66  below fig 11.19x4.63 in  How we score generated cells
+67  none  fig 12.45x5.31 in  Does the model follow its bulk conditioning?
+68  below fig 10.30x4.63 in  SLE benchmark: PBISC vs scSemiProfiler
+69  below fig 10.79x4.63 in  SLE benchmark: per cell type
+70  none  fig 12.45x5.38 in  Disjoint-cell controls test the bulk-broadcast baseline
+71  none  fig 12.45x5.60 in  SLE: what is added beyond the bulk broadcast?
+72  below fig 12.45x5.24 in  Same shift in three paired cohorts
+73  none  fig 12.45x5.64 in  Input corrections on one real-reference UMAP
+74  none  fig 12.45x4.86 in  Where generated-cell density differs from real cells
+75  none  fig 12.45x4.89 in  Input calibration changes cell-type composition
+76  none  fig 12.45x5.63 in  Same noise, different input: matched-cell shifts
+77  none  fig 12.45x4.77 in  Matched noise separates cell state from cell identity
+78  none  fig 12.45x5.63 in  Input changes can switch the generated cell label
+79  none  fig 12.45x5.58 in  Nearest real cells: distance and label agreement
+80  none  fig 12.34x5.70 in  Composition agreement does not ensure state agreement
+81  none  fig 12.45x5.47 in  Gene-length effects in the bulk input
+82  none  fig 12.45x5.47 in  Gene-length effects within label-stable cells
+83  none  fig 12.45x5.41 in  Raw and TPM shifts persist within cell labels
+84  none  fig 12.40x5.70 in  Reference calibration leaves specific gene-state shifts
+85  none  fig 12.45x5.45 in  Donor trait axes and donor identification differ
+86  none  fig 12.45x4.61 in  Reference choice changes different quality endpoints
+88  none  fig 12.45x4.72 in  Technical repeats and follow-up are different controls
+89  none  fig 12.45x5.32 in  Annotation agreement and difficult categories
+90  none  fig 12.45x5.40 in  Depth matching changes precision and coverage
+91  none  fig 12.45x4.86 in  Detection and library size are separate diagnostics
+92  right fig  9.20x5.18 in  Healthy cohort: LoRA cells cover the real ones
+93  right fig  9.20x4.79 in  Healthy cohort: realism beyond precision/coverage
+94  right fig  9.20x5.14 in  Healthy cohort: full vs LoRA vs condition encoder
+95  below fig  9.37x4.63 in  Representative-dependent composition
+96  below fig 11.39x4.63 in  Full vs LoRA vs cond.-encoder-only
+97  none  fig  9.32x5.70 in  Pair-budget summary: composition, coverage and steps
+98  right fig  9.20x5.09 in  Healthy cohort: LoRA rank and K
+99  right fig  9.20x5.09 in  Healthy paired-data learning curves
+100  right fig  9.20x4.66 in  Time: days for scSemiProfiler, minutes for LoRA
+101  below fig 10.77x4.63 in  5 donors: time to apply
+102  below fig  9.99x4.63 in  Pairs needed: composition, coverage, training steps
+103  below fig 10.08x4.63 in  LoRA rank: r = 4, 16, 64
+104  right fig  8.31x5.70 in  K diversification: no gain
+105  below fig 12.45x3.82 in  Five paired cohorts
+106  none  fig 12.45x4.70 in  Sepsis: exact-five reconstruction across five draws
+107  none  fig 12.45x5.18 in  Sepsis: realism beyond precision and coverage
+108  none  fig 12.45x4.70 in  Imperial: first completed exact-five comparison
+109  below fig 10.01x4.63 in  5 donors: LoRA vs scSemiProfiler — cells
+110  below fig 10.04x4.63 in  5 donors: realism beyond precision/coverage
+111  below fig 10.02x4.63 in  5 donors: precision/coverage at k = 3, 5, 10
+112  below fig 10.09x4.63 in  5-fold: realism beyond precision/coverage
+113  below fig 11.74x4.63 in  Same training pairs: LoRA vs scSemiProfiler
+114  below fig 11.85x4.63 in  scSemiProfiler and the representative-copy baseline
+115  none  fig 10.46x5.70 in  Sepsis: patient-aware disease-gene agreement
+116  none  fig 10.46x5.70 in  Imperial: patient-aware disease-gene agreement
+117  below fig 10.60x4.63 in  Sepsis: fine-tuning variants
+118  right fig  8.44x5.70 in  Pairs needed: donor and disease signal
+119  below fig 10.53x4.63 in  Cell-type deconvolution, same pairs: composition
+120  none  fig 12.45x4.95 in  Deconvolution: expression coverage limits the comparison
+121  none  fig 12.45x5.08 in  K=4,000 recipe: paired reconstruction quality
+122  none  fig 12.45x5.08 in  K=4,000 recipe: disease signal does not improve uniformly
+123  none  fig 12.45x5.22 in  K=4,000 recipe: calibration gains are selective
+124  none  fig 12.45x4.66 in  Influenza: response order versus response genes
+125  none  fig 10.30x5.70 in  COVID: calibration on a shared real-cell map
+126  none  fig 10.82x5.70 in  COVID monocytes: direction after depth matching
+127  none  fig 11.44x5.70 in  COVID monocytes: direction, magnitude and null
+128  none  fig 11.44x5.70 in  COVID CD14 monocytes: direction, magnitude and null
+129  none  fig 12.45x5.02 in  COVID: biological endpoints give a mixed verdict
+130  none  fig 11.83x5.70 in  Rare-state frequencies in observed cells
+131  none  fig 11.35x5.70 in  Rare-state misses: rules 1–8
+132  none  fig 11.72x5.70 in  Rare-state misses: rules 9–15
+133  none  fig 12.06x5.70 in  Rare-state fidelity: zero-aware reanalysis
+134  none  fig 10.76x5.70 in  Sepsis rare states across paired-data budgets
+135  none  fig 12.45x5.62 in  Sepsis rare-state direction across support budgets
+136  none  fig 11.35x5.70 in  Rare-state screen in DRESS, Kawasaki and SLE
+137  right fig  7.24x5.70 in  Consequence: disease direction flips
+138  none  fig 12.45x5.18 in  C1Q: single-gene detection versus co-detection
+139  none  fig 12.45x4.54 in  Behçet: K=4,000 changes direction more than identity
+140  none  fig 12.45x5.59 in  Behcet: one-gene and multi-gene C1Q rules
+141  below fig  9.46x4.63 in  Pairs from other cohorts do not transfer
+142  below fig  9.44x4.63 in  Longer training on other cohorts: gap stays
+143  below fig 10.61x4.63 in  No pretraining: same cohort, from scratch
+144  below fig 10.53x4.63 in  No pretraining: other cohorts, from scratch
+145  below fig 10.20x4.63 in  Rare states: from scratch vs pretrained
+146  below fig 10.12x4.63 in  Appendix: scSemiProfiler without output threshold
+147  none  fig 12.45x5.28 in  Bulk-only exploration: calibration populations differ
+149  right fig  9.20x5.19 in  Rare-state fractions: original analysis
+150  right fig  9.20x5.01 in  Rare-state ranking: original analysis
+151  below fig  9.23x4.98 in  Rare states: original five-fold analysis
+152  right fig  8.01x5.70 in  Bulk-only cohort plan: original snapshot
+153  none  fig 12.45x4.73 in  Historical bulk-only search: no confirmed new finding
+155  below fig 12.45x4.62 in  Next: other tissues from a few dozen pairs
+{"slides": 155, "overview": 7, "training_corpus": 8, "full_source_support": 9, "source_bindings": 996}
+convert /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_expanded_master_20261014_figure1.pptx as a Impress document -> /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_expanded_master_20261014_figure1.pdf using filter : impress_pdf_Export:{"UseTaggedPDF":{"type":"boolean","value":"false"},"PDFUACompliance":{"type":"boolean","value":"false"}}
+{"pass": true, "counts": {"spec": 155, "pptx": 155, "pdf": 155}, "failures": []}
+ 7  none  fig 12.29x5.70 in  Experimental overview
+ 8  none  fig 12.45x5.67 in  Pretraining data: 33.14 million training cells
+ 9  none  fig 12.45x5.67 in  Full source corpus: training, validation and test
+10  none  fig 12.45x5.67 in  From 41.40M source cells to 33.14M training cells
+11  none  fig 12.45x5.67 in  Recipe K and training exposure are different counts
+12  below fig 12.38x5.33 in  Corpus scale and generation workflow
+13  none  fig 12.45x4.07 in  From Gaussian noise to generated cell states
+16  none  fig 12.45x5.67 in  Where context width and trunk width change
+17  none  fig 12.45x5.67 in  Predicted and oracle hints: one trained denoiser
+18  none  fig 12.45x5.67 in  Architecture: full and late validation progression
+19  below fig  9.57x4.63 in  Width & context ablation: benchmark
+20  none  fig 12.45x5.67 in  Architecture arms: dimensions and endpoints
+21  below fig  9.57x4.63 in  Gene context: 2k / 3k / 6k / 15k
+22  none  fig 12.45x5.67 in  Gene panels: progression relative to each endpoint
+23  none  fig 12.45x5.67 in  Gene-panel arms: exact sizes and training provenance
+24  none  fig 12.45x5.67 in  All ablation arms: parameters and training time
+25  none  fig 12.15x5.70 in  Real-cell repeat: a sampling reference
+26  none  fig 12.45x5.48 in  Native-pseudobulk model comparisons
+29  none  fig 11.76x5.70 in  Bulk and pooled cells: different expression values
+30  none  fig 12.45x5.19 in  Paired calibration reduces same-donor expression error
+31  below fig 10.47x4.63 in  Consequence: cells off-manifold
+32  none  fig 12.33x5.70 in  Adaptation recovers the observed cell mixture
+33  none  fig 12.45x5.35 in  Adaptation: mixture error and expression agreement
+34  none  fig 12.15x5.70 in  Healthy paired data: full FT and LoRA
+35  none  fig 12.45x5.40 in  Full FT and LoRA ranks: quality and parameters
+36  none  fig 11.45x5.70 in  What does “all pairs” actually mean?
+37  none  fig 12.19x5.70 in  Healthy pair budget: correlation in all cell types
+38  none  fig 12.19x5.70 in  Healthy pair budget: error in all cell types
+39  right fig  9.20x5.17 in  Five pairs, no extra validation: LoRA vs scSemi
+40  none  fig 12.17x5.70 in  Composition: error and donor-level wins
+43  right fig  9.20x5.18 in  Disease signals with five training donors
+44  right fig  9.20x4.80 in  5-fold CV: disease signal
+45  below fig 11.93x4.63 in  Disease genes at every list length (CAT)
+46  below fig 10.54x4.63 in  Cell-type deconvolution, same pairs: disease signal
+47  none  fig 12.45x5.70 in  Disease gene rankings: complete CV comparison
+48  none  fig 11.99x5.70 in  Disease integrity depends on cohort and endpoint
+49  none  fig 12.19x5.70 in  Five-pair Sepsis: metrics across all support draws
+52  none  fig 12.33x5.70 in  Figure 5 targets begin with the source paper
+53  none  fig 12.18x5.70 in  Ye sepsis: published markers in paired scRNA
+54  none  fig 12.33x5.70 in  Transcriptional programs and rare-population identity
+60  none  fig 12.45x2.62 in  Corpus reference: assay and source composition
+61  below fig 12.45x4.61 in  Backbone: SDD → PBISC-Diffusion
+62  below fig 12.45x4.61 in  Bulk conditioning
+63  below fig 11.68x5.33 in  Denoiser: joint value + support diffusion
+64  below fig 12.45x4.71 in  Fine-tuning with paired bulk + sc
+65  below fig 12.35x4.63 in  Evaluation design
+66  below fig 11.19x4.63 in  How we score generated cells
+67  none  fig 12.45x5.31 in  Does the model follow its bulk conditioning?
+68  below fig 10.30x4.63 in  SLE benchmark: PBISC vs scSemiProfiler
+69  below fig 10.79x4.63 in  SLE benchmark: per cell type
+70  none  fig 12.45x5.38 in  Disjoint-cell controls test the bulk-broadcast baseline
+71  none  fig 12.45x5.60 in  SLE: what is added beyond the bulk broadcast?
+72  below fig 12.45x5.24 in  Same shift in three paired cohorts
+73  none  fig 12.45x5.64 in  Input corrections on one real-reference UMAP
+74  none  fig 12.45x4.86 in  Where generated-cell density differs from real cells
+75  none  fig 12.45x4.89 in  Input calibration changes cell-type composition
+76  none  fig 12.45x5.63 in  Same noise, different input: matched-cell shifts
+77  none  fig 12.45x4.77 in  Matched noise separates cell state from cell identity
+78  none  fig 12.45x5.63 in  Input changes can switch the generated cell label
+79  none  fig 12.45x5.58 in  Nearest real cells: distance and label agreement
+80  none  fig 12.34x5.70 in  Composition agreement does not ensure state agreement
+81  none  fig 12.45x5.47 in  Gene-length effects in the bulk input
+82  none  fig 12.45x5.47 in  Gene-length effects within label-stable cells
+83  none  fig 12.45x5.41 in  Raw and TPM shifts persist within cell labels
+84  none  fig 12.40x5.70 in  Reference calibration leaves specific gene-state shifts
+85  none  fig 12.45x5.45 in  Donor trait axes and donor identification differ
+86  none  fig 12.45x4.61 in  Reference choice changes different quality endpoints
+88  none  fig 12.45x4.72 in  Technical repeats and follow-up are different controls
+89  none  fig 12.45x5.32 in  Annotation agreement and difficult categories
+90  none  fig 12.45x5.40 in  Depth matching changes precision and coverage
+91  none  fig 12.45x4.86 in  Detection and library size are separate diagnostics
+92  right fig  9.20x5.18 in  Healthy cohort: LoRA cells cover the real ones
+93  right fig  9.20x4.79 in  Healthy cohort: realism beyond precision/coverage
+94  right fig  9.20x5.14 in  Healthy cohort: full vs LoRA vs condition encoder
+95  below fig  9.37x4.63 in  Representative-dependent composition
+96  below fig 11.39x4.63 in  Full vs LoRA vs cond.-encoder-only
+97  none  fig  9.32x5.70 in  Pair-budget summary: composition, coverage and steps
+98  right fig  9.20x5.09 in  Healthy cohort: LoRA rank and K
+99  right fig  9.20x5.09 in  Healthy paired-data learning curves
+100  right fig  9.20x4.66 in  Time: days for scSemiProfiler, minutes for LoRA
+101  below fig 10.77x4.63 in  5 donors: time to apply
+102  below fig  9.99x4.63 in  Pairs needed: composition, coverage, training steps
+103  below fig 10.08x4.63 in  LoRA rank: r = 4, 16, 64
+104  right fig  8.31x5.70 in  K diversification: no gain
+105  below fig 12.45x3.82 in  Five paired cohorts
+106  none  fig 12.45x4.70 in  Sepsis: exact-five reconstruction across five draws
+107  none  fig 12.45x5.18 in  Sepsis: realism beyond precision and coverage
+108  none  fig 12.45x4.70 in  Imperial: first completed exact-five comparison
+109  below fig 10.01x4.63 in  5 donors: LoRA vs scSemiProfiler — cells
+110  below fig 10.04x4.63 in  5 donors: realism beyond precision/coverage
+111  below fig 10.02x4.63 in  5 donors: precision/coverage at k = 3, 5, 10
+112  below fig 10.09x4.63 in  5-fold: realism beyond precision/coverage
+113  below fig 11.74x4.63 in  Same training pairs: LoRA vs scSemiProfiler
+114  below fig 11.85x4.63 in  scSemiProfiler and the representative-copy baseline
+115  none  fig 10.46x5.70 in  Sepsis: patient-aware disease-gene agreement
+116  none  fig 10.46x5.70 in  Imperial: patient-aware disease-gene agreement
+117  below fig 10.60x4.63 in  Sepsis: fine-tuning variants
+118  right fig  8.44x5.70 in  Pairs needed: donor and disease signal
+119  below fig 10.53x4.63 in  Cell-type deconvolution, same pairs: composition
+120  none  fig 12.45x4.95 in  Deconvolution: expression coverage limits the comparison
+121  none  fig 12.45x5.08 in  K=4,000 recipe: paired reconstruction quality
+122  none  fig 12.45x5.08 in  K=4,000 recipe: disease signal does not improve uniformly
+123  none  fig 12.45x5.22 in  K=4,000 recipe: calibration gains are selective
+124  none  fig 12.45x4.66 in  Influenza: response order versus response genes
+125  none  fig 10.30x5.70 in  COVID: calibration on a shared real-cell map
+126  none  fig 10.82x5.70 in  COVID monocytes: direction after depth matching
+127  none  fig 11.44x5.70 in  COVID monocytes: direction, magnitude and null
+128  none  fig 11.44x5.70 in  COVID CD14 monocytes: direction, magnitude and null
+129  none  fig 12.45x5.02 in  COVID: biological endpoints give a mixed verdict
+130  none  fig 11.83x5.70 in  Rare-state frequencies in observed cells
+131  none  fig 11.35x5.70 in  Rare-state misses: rules 1–8
+132  none  fig 11.72x5.70 in  Rare-state misses: rules 9–15
+133  none  fig 12.06x5.70 in  Rare-state fidelity: zero-aware reanalysis
+134  none  fig 10.76x5.70 in  Sepsis rare states across paired-data budgets
+135  none  fig 12.45x5.62 in  Sepsis rare-state direction across support budgets
+136  none  fig 11.35x5.70 in  Rare-state screen in DRESS, Kawasaki and SLE
+137  right fig  7.24x5.70 in  Consequence: disease direction flips
+138  none  fig 12.45x5.18 in  C1Q: single-gene detection versus co-detection
+139  none  fig 12.45x4.54 in  Behçet: K=4,000 changes direction more than identity
+140  none  fig 12.45x5.59 in  Behcet: one-gene and multi-gene C1Q rules
+141  below fig  9.46x4.63 in  Pairs from other cohorts do not transfer
+142  below fig  9.44x4.63 in  Longer training on other cohorts: gap stays
+143  below fig 10.61x4.63 in  No pretraining: same cohort, from scratch
+144  below fig 10.53x4.63 in  No pretraining: other cohorts, from scratch
+145  below fig 10.20x4.63 in  Rare states: from scratch vs pretrained
+146  below fig 10.12x4.63 in  Appendix: scSemiProfiler without output threshold
+147  none  fig 12.45x5.28 in  Bulk-only exploration: calibration populations differ
+149  right fig  9.20x5.19 in  Rare-state fractions: original analysis
+150  right fig  9.20x5.01 in  Rare-state ranking: original analysis
+151  below fig  9.23x4.98 in  Rare states: original five-fold analysis
+152  right fig  8.01x5.70 in  Bulk-only cohort plan: original snapshot
+153  none  fig 12.45x4.73 in  Historical bulk-only search: no confirmed new finding
+155  below fig 12.45x4.62 in  Next: other tissues from a few dozen pairs
+{"slides": 155, "overview": 7, "training_corpus": 8, "full_source_support": 9, "source_bindings": 996}
+convert /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_expanded_master_20261014_figure1.pptx as a Impress document -> /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_expanded_master_20261014_figure1.pdf using filter : impress_pdf_Export:{"UseTaggedPDF":{"type":"boolean","value":"false"},"PDFUACompliance":{"type":"boolean","value":"false"}}
+Overwriting: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_expanded_master_20261014_figure1.pdf
+{"pass": true, "counts": {"spec": 155, "pptx": 155, "pdf": 155}, "failures": []}