Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/sections/fig3.tex

@@ -0,0 +1,7 @@
+\mainfigure{fig:fig3}{LoRA cells are closer to real cells than scSemiProfiler's, in minutes instead of days}
+\figpanel{a}{figures/deck_20261010/fig3_healthy_fewpair.pdf}
+\figpanel{b}{figures/deck_20261010/fig3_healthy_pairs.pdf}
+\figpanelh{c}{figures/deck_20261010/fig3_healthy_copy.pdf}{0.22}
+\figpanelh{d}{figures/deck_20261010/fig3_time_days.pdf}{0.26}
+\figpanelh{}{figures/deck_20261010/sec09_time_to_apply.pdf}{0.26}
+\figlegend{\textbf{LoRA-5 against scSemiProfiler-5 on healthy donors, and the GPU time of both methods.} \textbf{a}, ImMicro (107 healthy donors), five draws: LoRA-5 is fine-tuned on five donors' pairs, scSemiProfiler-5 (official code and output) uses the same donors as representatives, and both generate cells for the other 102 donors from their real bulk. Zero-shot (no pairs, real bulk) for reference. Coverage and precision: $k = 10$ nearest-neighbour balls in 50 principal components of real cells; higher is better. Composition L1 over six cell types, classifier AUC ($k$-nearest-neighbour two-sample test; 0.5, indistinguishable), MMD$^2$ and $E_N$ (scSemiProfiler's normalised error; log scale): lower is better. Dashed lines (AUC, MMD$^2$, $E_N$), real against real cells of the same donors. Bars, mean of five draws; dots, draws. \textbf{b}, ImMicro, 5-fold cross-validation (CV) design: per fold, LoRA trains on $n$ pairs, early-stopped on seven inner-validation pairs, and scSemiProfiler gets the same $n$ pairs plus those seven as representatives; each donor is scored by the fold model that never saw it. Open symbols, every training donor of the fold (78--79): the 5-fold CV result, reported separately. Donor NK fraction r, Pearson r over the 107 donors between generated and observed NK-cell fractions. \textbf{c}, Left, composition L1 of copying the observed composition of the representative that scSemiProfiler assigns to each test donor, of scSemiProfiler-5 and of LoRA-5 (five draws of 102 test donors; bars, mean; dots, draws). Right, Spearman correlation over test donors between each method's L1 and the copy's. \textbf{d}, GPU time on one GB10 GPU, each job alone on it, from run receipts. Top, GPU hours per cohort: all five five-donor draws (left, main comparison) and the five fold models of the 5-fold CV (right, reported separately); log scale; dotted line, one day. Bottom, GPU time per five-donor draw: scSemiProfiler's training on its representatives and its one model per test sample; LoRA's fine-tuning (3,000 steps) and generation of 1,500 cells per test sample. Bars, means; dots, draws; error bars, s.d.; right, GPU seconds per test sample (log scale). Only per-cohort work is counted: not data preparation, scoring or what each method supplies (the pretrained model, scSemiProfiler's code).}

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/sections/fig4.tex

@@ -0,0 +1,8 @@
+\mainfigure{fig:fig4}{LoRA cells carry disease signal beyond the bulk, ahead of scSemiProfiler and cell-type deconvolution}
+\figpanel{a}{figures/deck_20261010/sec09_cohorts.pdf}
+\figpanel{b}{figures/deck_20261010/fig4_disease_fewpair.pdf}
+\figpanel{c}{figures/deck_20261010/fig4_disease_fivefold.pdf}
+\figpanel{d}{figures/deck_20261010/sec09_cat_curves.pdf}
+\figpanel{e}{figures/deck_20261010/secS_deconv_disease.pdf}
+\figpanelpair{f}{figures/deck_20261010/sec04_sle_benchmark.pdf}{figures/deck_20261010/sec04_sle_per_celltype.pdf}{0.48}{0.48}
+\figlegend{\textbf{Disease signal in four paired cohorts and in lupus.} \textbf{a}, Five paired cohorts: usable pairs (identity check passed, at least 200 cells), donors and contrasts (ESKD, end-stage kidney disease; ICI, immune checkpoint inhibitor; ChILI, ICI-induced liver injury). \textbf{b}, Main comparison with five donors: LoRA-5, scSemiProfiler-5 (same five donors as representatives), zero-shot (real bulk as measured) and zero-shot + calibration (gene-wise factors from the same pairs); other usable samples are generated from their real bulk. Bars, means; dots, draws. logFC r, Pearson r across genes between generated and observed case-vs-control logFC, mean over T, B, monocyte and NK cells; partial r, the same after both logFCs are residualised on the input bulk's logFC; top-50 DEG, genes shared by the observed and generated top-50 lists (Welch test, Benjamini--Hochberg FDR $<0.05$), summed over these types and dendritic cells. \textbf{c}, Donor-level 5-fold CV, reported separately, one pooled score per method: each sample comes from the fold model that never trained on it; scSemiProfiler's representatives are LoRA's training and inner-validation pairs; sc-only fine-tune, fine-tuned on the training donors' cells with pseudobulk conditions and fed calibrated bulk. \textbf{d}, Correspondence at the top (CAT): share of the observed top-$k$ genes (ranked by absolute limma-voom moderated t) also in the generated top-$k$ with the same sign, mean over the four types. Top, five donors (mean of draws); bottom, 5-fold CV; dashed, chance ($k/(2G)$ for $G$ tested genes). \textbf{e}, Cell-type deconvolution (fractions and per-type expression, not single cells) with the five donors' cells as reference and real bulk as mixture: BayesPrism posterior and CIBERSORTx HiRes expression, scored as in \textbf{b}; mean over draws; 3 of 20 draws have no CIBERSORTx score. In \textbf{b}--\textbf{e}, Giroux metrics exclude 65 sex-linked genes. \textbf{f}, SLE cohort GSE135779 (42 donors: 30 SLE, 12 controls; children and adults), each given as its own pseudobulk, without fine-tuning; \method{}, mean $\pm$ s.d. of five sampling seeds (dots); scSemiProfiler, one run with 14 other donors' cells as representatives. Left, logFC r (mean) and top-50 DEG (sum) over six cell types on 1,006 and 2,279 genes, partial r and composition L1 (lower is better); right, per cell type, 1,006 genes.}

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/sections/fig5.tex

@@ -0,0 +1,9 @@
+% Figure 5: rare cell states in the paired disease cohorts. Panels a-d are deck figures (slides 49-52);
+% panel e is planned (figure plan status Build). Data: DECK/data/fig5_rare_states.csv, DECK/data/sec05_behcet_direction.csv.
+\mainfigure{fig:fig5}{LoRA cells keep rare cell states closer to their observed fractions than scSemiProfiler's}
+\figpanel{a}{figures/deck_20261010/fig5_rare_fraction.pdf}
+\figpanel{b}{figures/deck_20261010/fig5_rare_donors.pdf}
+\figpanel{c}{figures/deck_20261010/fig5_rare_fivefold.pdf}
+\figpanel{d}{figures/deck_20261010/sec05_behcet_direction.pdf}
+\tbcpanel{\textbf{e}, planned: rare states in cohorts absent from the pretraining corpus, MS1 monocytes in Kawasaki disease and T cells high in interferon-stimulated genes (ISG-high T cells) in the systemic lupus erythematosus (SLE) cohort of children and adults (GSE135779).}
+\figlegend{\textbf{Rare cell states in four paired disease cohorts.} Fourteen literature-defined states (12 marker-gene rules, 2 atlas labels; ISG-high monocytes as a positive control) in four cohorts give 56 state-by-cohort pairs, each with at least 30 observed state cells at native depth and with state cells in at least 3 samples after thinning to the depth of LoRA's cells. Fraction, share of the parent type or of all cells; observed parent cells of marker states were thinned to the generated depth. \textbf{a}, Main comparison, five paired donors (mean of 5 draws; other samples generated from their real bulk). Left, fraction error (absolute $\log_2$ ratio of generated to observed fraction) of LoRA-5 against scSemiProfiler-5 given the same donors' single cells; below the diagonal, LoRA-5 is closer; shaded band, LoRA-5 within 2-fold; `not generated', a state the method never generates, counted as its miss. Right, $\log_2$ ratio for ten disease-linked state-by-cohort pairs, with zero-shot (real bulk, no pairs); grey band, within 2-fold. \textbf{b}, Left, donor ranking: Spearman correlation over test samples between generated and observed state fractions. Right, share of state-by-draw cases whose case-minus-control difference of sample means has the observed sign (contrasts in Supplementary Table~1); dotted line, 50\%. \textbf{c}, 5-fold cross-validation, reported separately (each sample from the fold model that never trained on it; scSemiProfiler given LoRA's training and inner-validation pairs). Left, fraction error; right, donor ranking. \textbf{d}, Beh\c{c}et disease, pretrained model without fine-tuning (no designed pairs). Patient (BD) minus control (HC) difference, in percentage points (pp), in the share of monocytes that are CD16 monocytes (left) or C1Q-high (right; any of \textit{C1QA}, \textit{C1QB} or \textit{C1QC} detected, looser than in \textbf{a}--\textbf{c}). Rows: real bulk (GSE198533; 9 BD, 10 HC; 5,000 cells per sample); pseudobulk of the single-cell donors; observed cells (GSE198616; 4 BD, 4 HC; mostly not the bulk donors). Hollow circles, seeds; filled circles, mean $\pm$ s.d. of 5 seeds; diamonds, real-bulk and observed rows after thinning to one common monocyte depth (C1Q-high, 5 seeds; CD16, seed 0); dashed line, observed; $P$, descriptive Mann--Whitney test (4 vs 4). \todo{add the LoRA-3 arm}. \textbf{e}, \todo{panel e: states from cohorts absent from the corpus (Kawasaki MS1 monocytes, SLE ISG-high T cells)}.}

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/sections/fig6.tex

@@ -0,0 +1,3 @@
+\mainfigure{fig:fig6}{Discovery from public bulk-only data (to be continued)}
+\tbcpanel{Planned panels: \textbf{a}, the bulk-only samples generated (sepsis 122, COVID-19 haemodialysis 164; $K=1{,}000$ and $K=4{,}000$; five LoRA-5 draws per sample). \textbf{b}, findings from the generated cells.}
+\figlegend{\textbf{Discovery from public bulk-only data.} \todo{Figure 6 is to be continued.}}