Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/archive/pre_rewrite_20261010/references.bib

@@ -0,0 +1,229 @@
+% Bibliography normalized from saved DOI-provider and conference BibTeX.
+% See editorial/citation_audit.json for retrieval provenance and transformations.
+
+@article{avilacobos2020benchmarking,
+  author = {Avila Cobos, Francisco and Alquicira-Hernandez, Jos{\'e} and Powell, Joseph E. and Mestdagh, Pieter and De Preter, Katleen},
+  title = {{Benchmarking of cell type deconvolution pipelines for transcriptomics data}},
+  journal = {Nature Communications},
+  volume = {11},
+  number = {1},
+  pages = {5650},
+  year = {2020},
+  doi = {10.1038/s41467-020-19015-1},
+  url = {http://dx.doi.org/10.1038/s41467-020-19015-1}
+}
+
+@article{newman2019cibersortx,
+  author = {Newman, Aaron M. and Steen, Chlo{\'e} B. and Liu, Chih Long and Gentles, Andrew J. and Chaudhuri, Aadel A. and Scherer, Florian and Khodadoust, Michael S. and Esfahani, Mohammad S. and Luca, Bogdan A. and Steiner, David and Diehn, Maximilian and Alizadeh, Ash A.},
+  title = {{Determining cell type abundance and expression from bulk tissues with digital cytometry}},
+  journal = {Nature Biotechnology},
+  volume = {37},
+  number = {7},
+  pages = {773--782},
+  year = {2019},
+  doi = {10.1038/s41587-019-0114-2},
+  url = {http://dx.doi.org/10.1038/s41587-019-0114-2}
+}
+
+@article{chu2022bayesprism,
+  author = {Chu, Tinyi and Wang, Zhong and Pe'er, Dana and Danko, Charles G.},
+  title = {{Cell type and gene expression deconvolution with BayesPrism enables Bayesian integrative analysis across bulk and single-cell RNA sequencing in oncology}},
+  journal = {Nature Cancer},
+  volume = {3},
+  number = {4},
+  pages = {505--517},
+  year = {2022},
+  doi = {10.1038/s43018-022-00356-3},
+  url = {http://dx.doi.org/10.1038/s43018-022-00356-3}
+}
+
+@article{wang2024scsemiprofiler,
+  author = {Wang, Jingtao and Fonseca, Gregory J. and Ding, Jun},
+  title = {{scSemiProfiler: Advancing large-scale single-cell studies through semi-profiling with deep generative models and active learning}},
+  journal = {Nature Communications},
+  volume = {15},
+  number = {1},
+  pages = {5989},
+  year = {2024},
+  doi = {10.1038/s41467-024-50150-1},
+  url = {http://dx.doi.org/10.1038/s41467-024-50150-1}
+}
+
+@article{nehar2020mapping,
+  author = {Nehar-Belaid, Djamel and Hong, Seunghee and Marches, Radu and Chen, Guo and Bolisetty, Mohan and Baisch, Jeanine and Walters, Lynnette and Punaro, Marilynn and Rossi, Robert J. and Chung, Cheng-Han and Huynh, Richie P. and Singh, Prashant and Flynn, William F. and Tabanor-Gayle, Joy-Ann and Kuchipudi, Navya and Mejias, Asuncion and Collet, Magalie A. and Lucido, Anna Lisa and Palucka, Karolina and Robson, Paul and Lakshminarayanan, Santhanam and Ramilo, Octavio and Wright, Tracey and Pascual, Virginia and Banchereau, Jacques F.},
+  title = {{Mapping systemic lupus erythematosus heterogeneity at the single-cell level}},
+  journal = {Nature Immunology},
+  volume = {21},
+  number = {9},
+  pages = {1094--1106},
+  year = {2020},
+  doi = {10.1038/s41590-020-0743-0},
+  url = {http://dx.doi.org/10.1038/s41590-020-0743-0}
+}
+
+@article{ota2021immunexut,
+  author = {Ota, Mineto and Nagafuchi, Yasuo and Hatano, Hiroaki and Ishigaki, Kazuyoshi and Terao, Chikashi and Takeshima, Yusuke and Yanaoka, Haruyuki and Kobayashi, Satomi and Okubo, Mai and Shirai, Harumi and Sugimori, Yusuke and Maeda, Junko and Nakano, Masahiro and Yamada, Saeko and Yoshida, Ryochi and Tsuchiya, Haruka and Tsuchida, Yumi and Akizuki, Shuji and Yoshifuji, Hajime and Ohmura, Koichiro and Mimori, Tsuneyo and Yoshida, Ken and Kurosaka, Daitaro and Okada, Masato and Setoguchi, Keigo and Kaneko, Hiroshi and Ban, Nobuhiro and Yabuki, Nami and Matsuki, Kosuke and Mutoh, Hironori and Oyama, Sohei and Okazaki, Makoto and Tsunoda, Hiroyuki and Iwasaki, Yukiko and Sumitomo, Shuji and Shoda, Hirofumi and Kochi, Yuta and Okada, Yukinori and Yamamoto, Kazuhiko and Okamura, Tomohisa and Fujio, Keishi},
+  title = {{Dynamic landscape of immune cell-specific gene regulation in immune-mediated diseases}},
+  journal = {Cell},
+  volume = {184},
+  number = {11},
+  pages = {3006--3021.e17},
+  year = {2021},
+  doi = {10.1016/j.cell.2021.03.056},
+  url = {http://dx.doi.org/10.1016/j.cell.2021.03.056}
+}
+
+@article{robinson2010edger,
+  author = {Robinson, Mark D. and McCarthy, Davis J. and Smyth, Gordon K.},
+  title = {{edgeR: a Bioconductor package for differential expression analysis of digital gene expression data}},
+  journal = {Bioinformatics},
+  volume = {26},
+  number = {1},
+  pages = {139--140},
+  year = {2010},
+  doi = {10.1093/bioinformatics/btp616},
+  url = {http://dx.doi.org/10.1093/bioinformatics/btp616}
+}
+
+@article{lopez2018scvi,
+  author = {Lopez, Romain and Regier, Jeffrey and Cole, Michael B. and Jordan, Michael I. and Yosef, Nir},
+  title = {{Deep generative modeling for single-cell transcriptomics}},
+  journal = {Nature Methods},
+  volume = {15},
+  number = {12},
+  pages = {1053--1058},
+  year = {2018},
+  doi = {10.1038/s41592-018-0229-2},
+  url = {http://dx.doi.org/10.1038/s41592-018-0229-2}
+}
+
+@article{cui2024scgpt,
+  author = {Cui, Haotian and Wang, Chloe and Maan, Hassaan and Pang, Kuan and Luo, Fengning and Duan, Nan and Wang, Bo},
+  title = {{scGPT: toward building a foundation model for single-cell multi-omics using generative AI}},
+  journal = {Nature Methods},
+  volume = {21},
+  number = {8},
+  pages = {1470--1480},
+  year = {2024},
+  doi = {10.1038/s41592-024-02201-0},
+  url = {http://dx.doi.org/10.1038/s41592-024-02201-0}
+}
+
+@article{theodoris2023geneformer,
+  author = {Theodoris, Christina V. and Xiao, Ling and Chopra, Anant and Chaffin, Mark D. and Al Sayed, Zeina R. and Hill, Matthew C. and Mantineo, Helene and Brydon, Elizabeth M. and Zeng, Zexian and Liu, X. Shirley and Ellinor, Patrick T.},
+  title = {{Transfer learning enables predictions in network biology}},
+  journal = {Nature},
+  volume = {618},
+  number = {7965},
+  pages = {616--624},
+  year = {2023},
+  doi = {10.1038/s41586-023-06139-9},
+  url = {http://dx.doi.org/10.1038/s41586-023-06139-9}
+}
+
+@article{luo2024scdiffusion,
+  author = {Luo, Erpai and Hao, Minsheng and Wei, Lei and Zhang, Xuegong},
+  title = {{scDiffusion: conditional generation of high-quality single-cell data using diffusion model}},
+  journal = {Bioinformatics},
+  volume = {40},
+  number = {9},
+  pages = {btae518},
+  year = {2024},
+  doi = {10.1093/bioinformatics/btae518},
+  url = {http://dx.doi.org/10.1093/bioinformatics/btae518}
+}
+
+@article{benjamini1995fdr,
+  author = {Benjamini, Yoav and Hochberg, Yosef},
+  title = {{Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing}},
+  journal = {Journal of the Royal Statistical Society Series B: Statistical Methodology},
+  volume = {57},
+  number = {1},
+  pages = {289--300},
+  year = {1995},
+  doi = {10.1111/j.2517-6161.1995.tb02031.x},
+  url = {http://dx.doi.org/10.1111/j.2517-6161.1995.tb02031.x}
+}
+
+@misc{mcinnes2018umap,
+  author = {McInnes, Leland and Healy, John and Melville, James},
+  title = {{UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction}},
+  year = {2018},
+  doi = {10.48550/ARXIV.1802.03426},
+  url = {https://arxiv.org/abs/1802.03426},
+  note = {Preprint at arXiv:1802.03426}
+}
+
+@misc{ho2020ddpm,
+  howpublished = {In \emph{Advances in Neural Information Processing Systems} \textbf{33}, 6840--6851 (Curran Associates, Inc.)},
+  author = {Ho, Jonathan and Jain, Ajay and Abbeel, Pieter},
+  title = {{Denoising Diffusion Probabilistic Models}},
+  booktitle = {Advances in Neural Information Processing Systems},
+  volume = {33},
+  pages = {6840--6851},
+  year = {2020},
+  url = {https://proceedings.neurips.cc/paper_files/paper/2020/file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf}
+}
+
+@misc{song2021ddim,
+  howpublished = {In \emph{International Conference on Learning Representations}},
+  author = {Song, Jiaming and Meng, Chenlin and Ermon, Stefano},
+  title = {{Denoising Diffusion Implicit Models}},
+  booktitle = {International Conference on Learning Representations},
+  year = {2021},
+  url = {https://openreview.net/forum?id=St1giarCHLP}
+}
+
+@misc{ostheimer2025sparse,
+  author = {Ostheimer, Phil and Nagda, Mayank and Balinskyy, Andriy and Radig, Jean and Herrmann, Carl and Mandt, Stephan and Kloft, Marius and Fellenz, Sophie},
+  title = {{Sparse Data Diffusion for Scientific Simulations in Biology and Physics}},
+  year = {2025},
+  doi = {10.48550/ARXIV.2502.02448},
+  url = {https://arxiv.org/abs/2502.02448v3},
+  note = {Preprint at arXiv:2502.02448, version 3 (22 January 2026)}
+}
+
+@article{babdor2026immicro,
+  author = {Babdor, Joel and Patel, Ravi K. and Davidson, Brittany and Koser, Kelvin and Noecker, Cecilia and Rahim, Maha K. and Bisanz, Jordan E. and Tenvooren, Iliana and Marquez, Diana and Calvo, Maria and Johri, Vrinda and McCarthy, Elizabeth E. and Shaheed, Avneet and Ekstrand, Christina and Weakley, Allison M. and Yu, Feiqiao B. and Krip, Kristen and Shaikh, Kashif A. and Amatullah, Hajera and Fiehn, Oliver and Turnbaugh, Peter J. and Combes, Alexis J. and Fragiadakis, Gabriela K. and Spitzer, Matthew H.},
+  title = {{Immune-microbiome coordination defines interferon setpoints in healthy humans}},
+  journal = {Cell},
+  volume = {189},
+  number = {10},
+  pages = {3144--3163.e20},
+  year = {2026},
+  doi = {10.1016/j.cell.2026.02.003},
+  url = {https://doi.org/10.1016/j.cell.2026.02.003}
+}
+
+@article{randolph2021genetic,
+  author = {Randolph, Haley E. and Fiege, Jessica K. and Thielen, Beth K. and Mickelson, Clayton K. and Shiratori, Mari and Barroso-Batista, Jo{\~a}o and Langlois, Ryan A. and Barreiro, Luis B.},
+  title = {{Genetic ancestry effects on the response to viral infection are pervasive but cell type specific}},
+  journal = {Science},
+  volume = {374},
+  number = {6571},
+  pages = {1127--1133},
+  year = {2021},
+  doi = {10.1126/science.abg0928},
+  url = {https://doi.org/10.1126/science.abg0928}
+}
+
+@misc{randolph2021scripts,
+  author = {Randolph, Haley E.},
+  title = {{Influenza A response variation scripts}},
+  year = {2021},
+  publisher = {Zenodo},
+  doi = {10.5281/zenodo.4273999},
+  url = {https://doi.org/10.5281/zenodo.4273999},
+  note = {Published 26 November 2021}
+}
+
+@article{zheng2022c1qbehcet,
+  title={Single-cell analyses highlight the proinflammatory contribution of {C1q}-high monocytes to {Behcet's} disease},
+  author={Zheng, Wenjie and Wang, Xiaoman and Liu, Jinjing and Yu, Xin and Li, Lu and Wang, Heping and Yu, Jijun and Pei, Xiaoya and Li, Chaoran and Wang, Zhimian and others},
+  journal={Proceedings of the National Academy of Sciences},
+  volume={119},
+  number={26},
+  pages={e2204289119},
+  year={2022},
+  doi={10.1073/pnas.2204289119}
+}

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/archive/pre_rewrite_20261010/sections/abstract.tex

@@ -0,0 +1 @@
+Bulk RNA sequencing exists for many immune cohorts that can no longer be sequenced cell by cell. A model that proposes single cells from a bulk profile would extend them, but only if three things hold: the generated cells must reflect the bulk they were given, the model must accept bulk measured in a clinic rather than the pseudobulk it was trained on, and the scores must be limited by something that can be fixed. We present \method{}, a diffusion model that writes peripheral blood cells directly in gene space, generating each gene's expression together with whether it is detected. On a held-out lupus cohort, swapping each donor's bulk for one from the opposite disease group inverts the recovered disease effect, from $r=+0.24$ to $-0.27$, and 40 of 42 generated populations follow the bulk they received, so the signal comes from the condition. From raw clinical bulk of 89 donors measured before and after influenza exposure, generated cells recover the exposure effect at $r=0.64$, at least as well as from the model's own training-style input; donor identity is not recovered, in any arm. Capacity is not the limit: a fourfold wider model lowers the training loss by 31\% and improves no biological score. The read-out is. The model's support head is trained as a detection probability, and drawing each gene from that probability instead of cutting it at one half raises differential-gene overlap from 57 to 73 and improves composition, with no tuned parameter. [TODO: ImMicro composition result and the donor-depth follow-up.] We show two uses, a 107-donor healthy reference cohort and a monocyte-subset question in Beh\c{c}et disease, and state where each stops.

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/archive/pre_rewrite_20261010/sections/declarations.tex

@@ -0,0 +1,17 @@
+\section*{Data availability}
+The childhood SLE data are in the Gene Expression Omnibus under GSE135779. We used processed ImMicro bulk RNA-seq data from GSE314922 and CITE-seq data from GSE314416. We used processed Randolph bulk RNA-seq data from GSE162632. The Randolph authors' pseudobulk matrix and sample information came from Zenodo record 4273999 (doi:10.5281/zenodo.4273999). We did not process the raw sequencing reads again. The Source Data folder contains the numbers used in the figures. It also lists the samples, correction and test splits, results, and source files for the paired-study test. The notes list the training data and the saved files used for each comparison. Before submission, the authors must add public archive links for the training recipes, generated expression arrays and saved models.
+
+\section*{Code availability}
+We used the model code, settings and test code from the Model 8 PBISC research repository. This paper package includes the plotting scripts and the numbers they use. Before submission, the authors must add a stable public code link and an archived software version.
+
+\section*{Acknowledgements}
+\textit{Authors must confirm the funding sources, grant numbers and people to thank.}
+
+\section*{Author contributions}
+\textit{Authors must state who did each part of the work and confirm approval of the final paper.}
+
+\section*{Competing interests}
+\textit{All authors must confirm their statements about competing interests.}
+
+\section*{Ethics declarations}
+This study used existing expression data. We did not collect new samples from people. Before submission, the authors must check the reuse rules for each dataset and their institution's ethics requirements. This draft does not claim any ethics approval or consent that has not been confirmed.

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/archive/pre_rewrite_20261010/sections/discussion.tex

@@ -0,0 +1,15 @@
+\section*{Discussion}
+
+A bulk-to-cell model is useful if the cells it writes follow the bulk it is given, if it accepts bulk that was measured rather than summed, and if its limits are known. \method{} meets the first two conditions at the level of condition effects and misses both at the level of the individual donor, and its main limit is a read-out that can be fixed without training.
+
+\textbf{The signal is the condition's.} Swapping, permuting and averaging the bulk change the recovered disease effect in the direction the bulk dictates, and generated populations follow the donor whose bulk they received. This rules out a model that reproduces a cohort-average cell, and it explains an earlier negative result: adding predicted cell-type proportions to the condition changed nothing because the bulk already carries them.
+
+\textbf{Clinical bulk works for what cohorts are used for.} The influenza effect in Randolph is recovered from raw bulk as well as from the model's training-style input, without the correction that condition-level comparisons had suggested was necessary. Group contrasts are what bulk cohorts are collected to test, and that is what transfers. What does not transfer is the donor: summed generated cells are not closer to their own donor than to others, even from pseudobulk, whereas the same measure reaches 31\% on the lupus benchmark. Two conditioning settings inherited from the condition-level experiments---a fixed depth and a fixed cell count---remove information that varies twofold across donors, and the pending native-depth run will decide whether the loss is theirs. [TODO: state the outcome.] Until then, claims built on generated cells should name the donor as the biological unit and the group as the object of inference.
+
+\textbf{Read the support head as what it was trained to be.} The one free setting in the sampler was a threshold on a quantity that is already a probability. Drawing each gene from that probability removes the setting, brings the detection rate within 10\% of the observed one, and improves two of three scores; a swept threshold can exceed it on the third only by producing cells denser than real ones and paying in composition. Two facts should stay attached to this result. The observed detection rate is not a property of the biology alone---it moves by a factor of two across donors of one cohort and by a quarter between cohorts---so any read-out that matches it is matching a protocol. And the drawn read-out still lands 10\% low and compresses the spread between donors, so the support head is calibrated but not perfectly so; a temperature on its logits, fitted once on training data, is the obvious next step and would keep the parameter out of the evaluation.
+
+\textbf{Capacity is not the limit, and it costs control.} A fourfold wider trunk lowers the loss by 31\% and returns nothing in disease agreement at any read-out. It also saturates the support head, so the read-out that moves the small models by 77 genes per cell moves the wide ones by 10; the loss gain is bought with a hard decision that the sampler can no longer soften. Validation loss cannot select a checkpoint for donor-level analysis, and neither, in this range, can parameter count.
+
+\textbf{What the generated object is good for.} It can be labelled, grouped and summed like observed cells, which is what makes donor-level tests and subset-level questions possible at all. In Beh\c{c}et disease it finds the C1Q-high monocyte population and reproduces two of three programme differences inside it; the third, which the observed cells carry in every sample, it does not. In the healthy reference cohort [TODO: composition result]. The honest summary is that the object is enough to ask a within-cell-type question of a bulk cohort and to answer it in part, and that where it fails it fails on magnitude---effect sizes compressed to roughly 40\%, a programme difference flattened---rather than on direction.
+
+\textbf{Limits.} Each width was trained once, and two runs that differ only by an unused input resolve 0.018 in $r$ and 7 in overlap, so small orderings between arms are not interpretable. The lupus cohort was used during development and is not an untouched test set; the two paired cohorts were not. Sampling at 50 steps is stable for gene sets and composition and not converged for the correlation. Cell-type scores depend on one frozen annotator, and unassigned cells stay in every denominator. The loss spends most of its weight on measured zeros, and restricting it to expressed genes collapses the cell-type structure, so the obvious change to the objective is the wrong one; within-cell-type dispersion, which the model under-produces, is not a term at all.