Archive 4 checkpoints (1-4)

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

@@ -0,0 +1,26 @@
+% Declarations: data and code availability, acknowledgements, author contributions, competing interests, ethics.
+% Written 2026-10-10. Accessions checked against plans/paired_finetune_intradataset_20260927/data/<cohort>/QC_REPORT.md
+% and MANIFEST.json, the datalake GEO series titles (GSE314922 = bulk, GSE314416 = CITE-seq), the CELLxGENE source
+% citation of the AIFI atlas, plans/c1q_crossdisease_20261004/README.md (Kawasaki), the Randolph and Arunachalam source
+% records of plans/utility_studies_final6k_20260925, and the software versions in run receipts and
+% scores_contrast/*/limma/versions.txt; BayesPrism 2.2.3 from the DESCRIPTION file in the R library that
+% plans/bayesprism_paired_20261006/tools/run_bp.R loads. Author-specific statements are left as \todo.
+% Fragment only: \input by main.tex after the Methods.
+
+\section*{Data availability}
+All data analysed in this study are public. We used the processed expression data that the authors deposited and did not reprocess sequencing reads. The five paired cohorts are available as follows. ImMicro has bulk RNA-seq in the Gene Expression Omnibus (GEO) under GSE314922 and CITE-seq under GSE314416\cite{babdor2026immicro}. Ye 2026 sepsis has bulk RNA-seq under GSE279448 and scRNA-seq under GSE279452\cite{ye2026sepsis}; its CITE-seq series GSE279451 was not used. Imperial has bulk RNA-seq in Zenodo record 6497251 (doi:10.5281/zenodo.6497251)\cite{gisby2022multiomics} and CITE-seq gene expression in the COVID-19 Cell Atlas file \texttt{eskd\_covid19.h5ad} (\url{https://covid19.cog.sanger.ac.uk/eskd_covid19.h5ad})\cite{stephenson2025temporal}. Giroux is GEO SuperSeries GSE206284, with bulk RNA-seq under GSE206263 and scRNA-seq under GSE206283\cite{giroux2022differential}. Ji has whole-blood bulk RNA-seq under GSE287540 and scRNA-seq under GSE287541\cite{ji2025chili}. The cohort of Liu et al.\ (GSE161918)\cite{liu2021timeresolved} supplied extra bulk RNA-seq and CITE-seq training pairs, used only in Supplementary Notes~2 and 3. Its single cells are also part of the pretraining corpus, and it was never a test cohort. The lupus scRNA-seq cohort is GSE135779\cite{nehar2020mapping}. Beh\c{c}et disease bulk RNA-seq is GSE198533 and its scRNA-seq GSE198616\cite{zheng2022c1qbehcet}. Kawasaki disease scRNA-seq, for the planned Fig.~\ref{fig:fig5}e, is GSE168732\cite{wang2021kawasaki}. The two further paired cohorts of Fig.~\ref{fig:fig2}a are those of Randolph et al., with bulk RNA-seq under GSE162632 and the authors' pseudobulk inputs in Zenodo record 4273999 (doi:10.5281/zenodo.4273999)\cite{randolph2021genetic}, and of Arunachalam et al., with bulk RNA-seq under GSE152418 and scRNA-seq under GSE155673\cite{arunachalam2020systems}. The Human Immune Health Atlas of the Allen Institute for Immunology (AIFI), on which the cell-type annotator was trained, is in CZ CELLxGENE Discover (collection 77f9d7e9-5675-49c3-abed-ce02f39eef1b)\cite{gong2025multiomic}. The 22 public datasets of the pretraining corpus are summarised in Supplementary Note~1 and Supplementary Fig.~1. \todo{list the 22 corpus datasets with their accessions (manuscript source\_data/corpus/corpus\_datasets.csv, where several CELLxGENE entries lack a collection identifier) in Supplementary Note~1 or a Supplementary Table} \todo{deposit and link the pretrained model and LoRA weights, the generated cells (including the bulk-only sets of Fig.~\ref{fig:fig6}), the sample splits, conditions and scores} Source data are provided with this paper. \todo{assemble the Source Data file from the deck's data tables}
+
+\section*{Code availability}
+Code for pretraining, LoRA fine-tuning, sampling, scoring and every figure is available at \todo{public repository URL} and archived at \todo{Zenodo DOI of the release}. The model was trained and sampled with Python 3.12.14 and PyTorch 2.14.0. The cell-type annotator used scvi-tools 1.4.1, and the classifier tests used scikit-learn 1.9.0. The correspondence-at-the-top (CAT) analysis used R 4.5.2 with limma 3.66.0 and edgeR 4.8.1. scSemiProfiler was run from the authors' repository at commit c2aea02 with the compatibility patch described in Methods. Cell-type deconvolution used BayesPrism 2.2.3 and the official CIBERSORTx Docker images (Fractions and HiRes). \todo{full software list for the Reporting Summary}
+
+\section*{Acknowledgements}
+\todo{authors to add funding sources, grant numbers and the people to thank}
+
+\section*{Author contributions}
+\todo{authors to state who did each part of the work and to confirm that all authors approved the final manuscript}
+
+\section*{Competing interests}
+\todo{all authors to confirm their competing-interest statements}
+
+\section*{Ethics declarations}
+This study reanalysed published, de-identified expression data, and no new samples were collected from people. Approvals and consent for the original samples are those reported by the source studies. \todo{authors to confirm the reuse terms of every dataset and their institution's requirements; this draft claims no approval that has not been confirmed}

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

@@ -0,0 +1,23 @@
+% Discussion. Written 2026-10-10 to frame the Results drafts (results_fig1.tex ... results_fig6.tex) and
+% Supplementary Notes 2 and 3 (supp_notes.tex). Numbers come from those drafts and were checked against the deck
+% (slides 17, 19, 23, 26, 30, 33-34, 38-41, 45, 49, 52, 59, 61, 68, 70-72, 76) and DECK/data: fig23_healthy.csv,
+% sec05_condition_space_numbers.csv, sec05_shift_cohorts.csv, sec05_cells_metrics.csv, sec08_pairs_realism.csv,
+% sec08_pairs_signal.csv, sec09_extra_fewpair.csv, sec09_extra_fivefold.csv, sec09_time_to_apply.csv, fig4_disease.csv,
+% secS_deconvolution.csv, fig5_rare_states.csv (50/56 below observed, 19/56 within 2-fold and 28/56 zero-shot closer
+% recounted), sec05_behcet_direction.csv, sec08_cross_cohort.csv, sec10_other_tissues.csv, sec07_no_pretraining.csv.
+% Tightened 2026-10-10 (copy edit only: repeated definitions removed, no number changed or added).
+% Review fixes 2026-10-11: classifier AUC labelled by design (LoRA-5 cohort means of five draws, sec09_extra_fewpair.csv;
+% 5-fold CV 0.87-0.94, sec09_extra_fivefold.csv, slide 61); Supplementary Fig. 15 cited; pilot donor composition r
+% defined (mean over T, B, monocyte, NK; sepsis fold 0; sec07_no_pretraining.csv, slide 70).
+% Fragment only: \input by main.tex after the Results.
+\section*{Discussion}
+
+\textbf{A laboratory with bulk RNA-seq and a handful of paired samples can now generate single cells for every sample.} With five donors' pairs, LoRA fitted \method{} to a cohort's bulk in about 2 min, and a five-donor draw took 3 to 11 GPU minutes on average per cohort. Of the methods tested, LoRA-5 cells covered the most real cells and were the hardest to tell from real ones, in all 25 draws over five cohorts. In 107 healthy donors, their coverage was 0.79, against 0.04 for scSemiProfiler-5 given the same five donors. The cells also carried disease signal beyond the bulk: partial r favoured LoRA-5 over scSemiProfiler-5 in 14 of 20 draws. It also favoured LoRA-5 over cell-type deconvolution by BayesPrism (18 of 20 draws) and CIBERSORTx (17 of 17 scored draws), which return fractions and per-type expression, not cells. On top-50 DEG overlap, that comparison was mixed, with deconvolution ahead in sepsis. LoRA-5 kept rare states closer to their observed fractions than scSemiProfiler-5 in 51 of 56 state-by-cohort pairs. With all training pairs, in 5-fold CV reported separately, LoRA's logFC r reached 0.82 for COVID-19 in haemodialysis patients (scSemiProfiler 0.71).
+
+\textbf{Pretraining supplies the cells, and a few pairs teach the cohort's bulk.} From pseudobulk, the pretrained model already made cells that lay among real ones (precision 0.92 in these healthy donors). For real bulk it lacked the cohort's shift from bulk to pseudobulk, which is mostly shared by all donors (95\% of the squared offset; 71\% and 92\% in two further cohorts), so a few pairs can teach it. In the pairs-needed curves of these donors, five pairs per fold already gave as much coverage as all training pairs (0.84 against 0.78). Without pairs, the pretrained model can even reverse a disease direction (Fig.~\ref{fig:fig5}d). The shift is also cohort-specific. LoRA trained to saturation on other cohorts' pairs fell short of the cohort's own LoRA (5-fold CV, all training pairs) on coverage in 5 of 5 cohorts and on disease logFC r in 4 of 4 (Supplementary Note~2 and Supplementary Fig.~15). \method{} therefore fits one LoRA per cohort.
+
+\textbf{The generated cells are not yet real cells, and our tests have limits.} A classifier still told LoRA cells from real ones (AUC 0.89 to 0.95 for LoRA-5, means of five draws in each of the five cohorts; 0.87 to 0.94 in 5-fold CV). Most rare states were generated below their observed fraction (50 of 56 pairs for LoRA-5; 19 of 56 within 2-fold), and zero-shot came closer than LoRA-5 on 28 of the 56 fractions. In Giroux, whose five-donor test sets hold two subjects per group, scSemiProfiler-5 led on disease signal (logFC r 0.40 against 0.30). In sepsis, its means over the five draws were also higher (logFC r 0.40 against 0.35, partial r 0.33 against 0.30), although LoRA-5 led in three of five draws. In the sepsis pairs-needed curves, at equal budgets of 5 to 40 pairs per fold, scSemiProfiler's logFC r was higher (0.52 to 0.58 against 0.29 to 0.48), but its top-50 DEG overlap stayed at 2 to 6 (LoRA 5 to 22). The SLE cohort of children and adults (GSE135779; Fig.~\ref{fig:fig4}f) helped choose the model's width and gene panel, so it is a development cohort; the final model was fixed before the paired-cohort experiments reported here. All paired-cohort scores rest on one frozen scANVI annotator and the 6,000-gene panel, and all data come from blood.
+
+\textbf{Three directions follow.} First, a pilot on one sepsis fold suggests that a few dozen pairs could train a similar model for tissues without a large single-cell corpus (Supplementary Note~3 and Supplementary Fig.~16). Trained from random weights on 58 pairs plus pseudobulk from the training donors' cells, the same denoiser matched the pretrained LoRA trained the same way on coverage (0.88 against 0.87) and exceeded it on disease logFC r (0.54 against 0.45). Pretraining kept clear advantages. The randomly initialised model was further from the observed fractions of cycling T and NK cells in 8 of 8 comparisons, with 5 or 58 pairs and with other cohorts' pairs. From five pairs in that fold, it could not place donors (mean across-donor correlation of T, B, monocyte and NK fractions, 0.17 against 0.70). It also trained for 61 to 86 min instead of 3 to 5 min. It reused the pretrained gene panel and scaling. Folds 1 to 4 and a second cohort are pending. Second, Fig.~\ref{fig:fig6} applies the LoRA-5 models to 122 sepsis and 164 haemodialysis samples that have bulk only. \todo{bulk-only findings of Fig.~6} Third, with the released model\todo{release links}, a laboratory can fine-tune a LoRA on its own pairs on one GPU and generate cells for all its bulk samples.
+
+In short, five donors' pairs and minutes of fine-tuning turn a bulk blood cohort into the most realistic single cells of the methods tested, with disease signal beyond the bulk.

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/sections/fig1.tex

@@ -0,0 +1,9 @@
+\mainfigure{fig:fig1}{A compact gene-space diffusion model turns one bulk profile into single blood cells}
+\figpanel{a}{figures/deck_20261010/sec02_backbone_compare.pdf}
+\figpanelh{b}{figures/deck_20261010/fig01_architecture_nature.pdf}{0.38}
+\figpanel{c}{figures/deck_20261010/sec02_bulk_conditioning.pdf}
+\figpanel{d}{figures/deck_20261010/fig01_corpus_nature.pdf}
+\figpanelpair{e}{figures/deck_20261010/sec03_width_loss.pdf}{figures/deck_20261010/sec03_width_bench.pdf}{0.49}{0.49}
+\figpanel{f}{figures/deck_20261010/sec03_panel.pdf}
+\noindent\todo{Panels b and d reuse manuscript images that print their own panel letters (a--c inside b, d inside d); remove those letters when the 180 mm figure is drawn.}\par
+\figlegend{\textbf{Design and performance of \method{}.} \textbf{a}, Three diffusion backbones: scDiffusion\cite{luo2024scdiffusion}, Sparse Data Diffusion (SDD)\cite{ostheimer2025sparse} and \method{}; orange, changed from SDD. \textbf{b}, Overview. Top left, $K$ cells of one sample form a pseudobulk, and each cell is learned with it, without cell-type or disease labels. Top right, noise is added to each gene's value and detection; a denoiser conditioned on the bulk in every block removes it step by step, and detection is drawn from its probability, leaving exact zeros. Bottom, one measured bulk gives one shared condition, and independent noise draws give cells, labelled after generation; an optional low-rank adaptation (LoRA) uses a few paired samples (Fig.~\ref{fig:fig2}). Mini-plots are schematic; details in Supplementary Fig.~2. \textbf{c}, The final model's condition (18,006 numbers) and its 512-dimensional context, which enters all seven denoiser blocks. A measured bulk is given $K=1{,}000$; block heights are schematic. \textbf{d}, Pretraining corpus: cell shares by assay, condition family within assay, and cell type (Supplementary Fig.~1). Labels describe the corpus and are never model input. \textbf{e}, Width and context on an earlier 6,000-gene panel (1\% detection floor). Left, training and validation loss per schedule part for 1x (base), ctx x4 and trunk x4 (condition context or denoiser layers four times wider), both x4, 16x (stopped in part 3 by a memory limit) and 1x with a cell-type proportion hint; parameters in brackets; circles, first part within 1\% of the final validation loss. Right, systemic lupus erythematosus cohort of children and adults (GSE135779\cite{nehar2020mapping}; 42 donors, 2,980 genes): bars, mean of parts 14--16 with detection cut at 0.5; dots, single parts; diamonds, drawn (Bernoulli) read-out; hint pred and hint oracle, predicted and true proportions. \textbf{f}, Gene panels selected at a 0.1\% detection floor, scored with the drawn read-out on the 1,006 benchmark genes in every panel, same donors. 6,000 genes, mean $\pm$ s.d. of five sampling seeds (dots); other panels, one seed. Right, validation loss divided by each panel's part-16 value; circles as in \textbf{e}. Orange dashed lines in \textbf{e} and \textbf{f}, scSemiProfiler\cite{wang2024scsemiprofiler} on the same donors (Fig.~\ref{fig:fig4}f); \textbf{e} and \textbf{f} use different genes and are not comparable.}

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/sections/fig2.tex

@@ -0,0 +1,23 @@
+% Figure 2 block. Panels are the deck's vector PDFs placed as-is.
+\mainfigure{fig:fig2}{A few paired samples and minutes of LoRA fine-tuning adapt \method{} to real bulk}
+\figpanelpair{a}{figures/deck_20261010/sec05_condition_space.pdf}{figures/deck_20261010/sec05_shift_cohorts.pdf}{0.37}{0.61}
+\figpanelpair{b}{figures/deck_20261010/sec05_cells_umap.pdf}{figures/deck_20261010/sec05_cells_metrics.pdf}{0.50}{0.48}
+\figpanelh{c}{figures/deck_20261010/sec06_ft_method.pdf}{0.25}
+\figpanelh{}{figures/deck_20261010/sec06_eval_design.pdf}{0.25}
+\figpanelh{d}{figures/deck_20261010/sec04_metrics_overview.pdf}{0.24}
+\figpanelh{e}{figures/deck_20261010/fig2_healthy_cells.pdf}{0.27}
+\figpanelh{f}{figures/deck_20261010/fig2_healthy_realism.pdf}{0.27}
+\figpanelh{g}{figures/deck_20261010/fig2_healthy_modes.pdf}{0.27}
+\figpanelh{h}{figures/deck_20261010/fig2_healthy_pairs.pdf}{0.29}
+\figpanelh{i}{figures/deck_20261010/fig2_healthy_rank.pdf}{0.27}
+\figlegend{\textbf{Fine-tuning \method{} with LoRA on healthy donors.} ImMicro, 107 healthy donors, unless stated.
+\textbf{a}, Each donor's real bulk (blue) and pseudobulk (black or open, by single-cell pool), joined by lines, in the model's input space (log1p CPM, 5,508 genes; PCA fitted on pseudobulks). Bars, own donor's pseudobulk nearest (PCA-30; all genes), as measured and with the other donors' mean shift removed (diagnostic). Right, distance to own pseudobulk, share of the squared offset in one shared shift, and own-donor matching in three paired cohorts (Randolph\cite{randolph2021genetic}, 178 samples; Arunachalam\cite{arunachalam2020systems}, 12). D, median distance between two donors' pseudobulks.
+\textbf{b}, Cells of the pretrained model without fine-tuning, from each donor's real bulk or pseudobulk, from the same starting noise, beside held-out real cells (UMAP\cite{mcinnes2018umap} fitted on real cells; it overstates overlap). Right, monocyte share, precision and coverage; real cells, scored against other real cells of the same donors. Inputs use each donor's single-cell depth and cell count; \textbf{e}--\textbf{i} use $K=1{,}000$.
+\textbf{c}, LoRA fine-tuning (rank 16 in 22 linear layers; 1.30 M trainable weights; merged after training) and the donor-level evaluation designs for sepsis (bars) and all five paired cohorts (table); `few pairs' is LoRA-5, the 5-donor design of \textbf{e}.
+\textbf{d}, Metrics used in Figs~2--4.
+\textbf{e}, Coverage, precision and composition L1 (six scANVI cell types) for zero-shot, zero-shot + calibration, sc-only fine-tuning (5-fold only; trained on pseudobulk, fed calibrated bulk) and LoRA. Left, 5 donors (LoRA-5; bars, means; dots, 5 draws of 102 test donors). Right, 5-fold CV (one pooled call), reported separately.
+\textbf{f}, Classifier AUC ($k$-nearest-neighbour; 0.5, indistinguishable), MMD$^2$ and $E_N$ (log scale); dashed, real against real cells.
+\textbf{g}, Full and LoRA fine-tuning, 5-fold pooled and on fold 0 alone (22 test donors; plus condition-encoder-only), and trainable weights.
+\textbf{h}, Pairs-needed curves (5-fold design, pooled): x, training pairs per fold (0, no pairs; plus 7 inner-validation pairs); calibration fitted on the same pairs; NK fraction r, across-donor Pearson r, generated against observed; selected step, best inner-validation step (dots, folds; lines, medians; cap 20,000).
+\textbf{i}, LoRA rank 4, 16 and 64 at 5, 20 and all training pairs; open symbols, several bulk depths ($K=250$--$4{,}000$) and added pseudobulk recipes, with all pairs (5-fold, pooled).
+Precision and coverage use $k=10$ nearest-neighbour balls in a 50-component PCA of held-out real cells.}