Archive 4 checkpoints (1-4)

.gitattributes

@@ -113,3 +113,4 @@ paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-
 paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/supplementary.pdf filter=lfs diff=lfs merge=lfs -text
 paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/figures/deck_20261010/sec01_slide_a.pdf filter=lfs diff=lfs merge=lfs -text
 paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/archive/pre_rewrite_20261010/main_before_rewrite.pdf filter=lfs diff=lfs merge=lfs -text
+paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/archive/pre_rewrite_20261010/supplementary_before_rewrite.pdf filter=lfs diff=lfs merge=lfs -text

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

@@ -0,0 +1,51 @@
+\section{Supplementary Methods}
+\label{sec:supp-methods}
+
+\subsection{Pretraining corpus}
+The registered corpus holds 41,395,184 cells from 22 public datasets, 3,199 donors and 6,507 samples (Supplementary Fig.~\ref{fig:supp_corpus}). By platform, 74.8\% of cells come from 10x Genomics assays, 23.4\% from Parse, 1.7\% from ScaleBio and 0.1\% from Seq-Well. Most cells (88.5\%) come from donors without a recorded disease. This group includes 9.2 million cells from cytomegalovirus-seropositive donors, which the source metadata record as cytomegalovirus infection, and blood drawn after influenza vaccination. The rest come from autoimmune disease (7.1\%: systemic lupus erythematosus, rheumatoid arthritis, ACPA-positive individuals at risk of rheumatoid arthritis and juvenile dermatomyositis), infection (3.4\%: COVID-19, post-COVID-19 condition and influenza), hematologic disorders (1.0\%: myelodysplastic syndrome, myelodysplastic/myeloproliferative neoplasms and cytopenia) and drug reaction with eosinophilia and systemic symptoms ($<$0.1\%). One dataset is enriched for circulating hematopoietic progenitors and one for B cells. About a quarter of the cells (23.6\%) were cultured before sequencing, almost all of them in a 90-cytokine panel on peripheral blood mononuclear cells from 12 healthy donors. Where the source metadata lacked a condition or an assay, we took it from the GEO series records (GSE274680, GSE132802, GSE163121 and GSE314010) or from the data provider's description. Cell types come from each dataset's own labels, mapped to a shared two-level hierarchy: T cells 59.8\%, monocytes 17.8\%, B cells 9.0\%, NK cells 8.1\%, progenitors 2.9\% and dendritic cells 1.1\%, with 0.7\% unlabelled. These condition and cell-type labels describe the corpus only; the model never receives them.
+
+\begin{figure}[p]
+\centering
+% Authored at 180 mm like Fig. 1; 1.17\textwidth keeps the smallest labels near 4.5 pt and the legend on the page.
+\makebox[\textwidth][c]{\includegraphics[width=1.17\textwidth]{figures/figS_corpus.pdf}}
+\caption{\textbf{The pretraining corpus.}
+\textbf{a}, Shares of cells by split, by sample handling (in vitro: cultured or stimulated before sequencing) and by assay, and condition shares within each assay. Healthy includes cytomegalovirus-seropositive donors and post-vaccination draws.
+\textbf{b}, The 22 datasets: cells (log scale), donors, samples, and condition and cell-type (L1) shares.
+\textbf{c}, Cell types from each dataset's own labels, mapped to a shared two-level hierarchy: L1 shares and L2 cell counts (log scale).
+\textbf{d}, What each cell record carries. Condition and cell-type labels describe the corpus and are never model inputs.}
+\label{fig:supp_corpus}
+\end{figure}
+
+\subsection{Continuous noise schedule and reverse steps}
+Before clipping, the cumulative cosine coefficient was
+\begin{equation}
+ A(t)=\frac{\cos^2\!\left[\frac{t+0.008}{1.008}\frac{\pi}{2}\right]}{\cos^2\!\left[\frac{0.008}{1.008}\frac{\pi}{2}\right]}.
+\end{equation}
+The code clips $A(t)$ to $[10^{-12},1-10^{-12}]$. It then computes $\lambda(t)=\operatorname{clip}(\log A(t)-\log[1-A(t)],-20,20)$. The coefficients used by the model are $\alpha(t)=\sqrt{\operatorname{sigmoid}[\lambda(t)]}$ and $\sigma(t)=\sqrt{\operatorname{sigmoid}[-\lambda(t)]}$. The schedule uses float32 when model states use lower precision. Sinusoidal time features use 64 frequencies $\omega_k=10000^{-k/63}$. They join $\cos(2\pi t\omega_k)$ and $\sin(2\pi t\omega_k)$ for $k=0,\ldots,63$.
+
+For the denoiser prediction $\hat z_0=[\operatorname{clip}(\hat v,-1,1),\tanh(\ell/2)]$, reverse steps add no new noise. Between adjacent times $t>t'$, they use
+\begin{equation}
+ \hat\varepsilon=\frac{z_t-\alpha(t)\hat z_0}{\max[\sigma(t),10^{-8}]},\qquad
+ z_{t'}=\alpha(t')\hat z_0+\sigma(t')\hat\varepsilon.
+\end{equation}
+The grid has 51 endpoints $t_k=1-k/50$. It requires 50 denoiser calls. The final call, at $t=1/50$, returns its clean prediction directly. The previous clean prediction supplies self-conditioning during sampling. The final count-like output is
+\begin{equation}
+ \hat x_g=\left\{\exp\left[8\,\operatorname{clip}\left(\frac{\hat v_g+1}{2},0,1\right)\right]-1\right\}
+ \mathbb I(\ell_g>0)\,m_g.
+\end{equation}
+The code uses a condition check to set excluded values to exactly zero. It does not add Poisson or negative-binomial count draws. Nonzero generated values are therefore continuous, with a maximum of $\exp(8)-1$. A positive support prediction can still give a zero output when the clipped value reaches its lower endpoint. We therefore store separate masks for positive support and nonzero output.
+
+\subsection{Gene-module definitions}
+The IFN-core module contains \textit{ISG15, IFI6, IFI27, IFI35, IFI44, IFI44L, IFIT1, IFIT2, IFIT3, IFITM1, IFITM2, IFITM3, IRF7, IRF9, ISG20, MX1, MX2, OAS1, OAS2, OAS3, OASL, RSAD2, STAT1} and \textit{XAF1}. All 24 genes appear in the fixed gene list. The shared 1,945-gene SLE evaluation set lacks \textit{IRF9}, so its score uses 23 genes. The 11,589-gene set adds measured \textit{IRF9}.
+
+The B/plasmablast module in the fixed gene list contains \textit{MS4A1, CD19, CD27, CD38, CD79A, CD79B, BANK1, CD74, HLA-DRA, HLA-DRB1, TNFRSF13B, TNFRSF17, MZB1, XBP1, JCHAIN, PRDM1} and \textit{IRF4}. The original requested set also contained \textit{IGKC, IGHG1} and \textit{IGHG3}. These were absent from the fixed gene list, so the reported coverage uses 17 as its denominator. The shared gene set lacks \textit{CD27, TNFRSF17} and \textit{XBP1}. The 6,000- and 11,589-gene sets include measured \textit{CD27} and \textit{XBP1}, giving coverage of 16/17. Coverage means that genes can be tested. It does not mean that the model recovered their expression. These lists came from a saved pSS analysis for the related manuscript. The saved SLE records confirmed module names and used-gene counts but did not separately retain all original gene symbols.
+
+\subsection{How to read the single-recipe check}
+The $K=1000$ case was validation recipe \texttt{4ffc27115a79d477baa87e74}, sample \texttt{RV8959683}, donor \texttt{CV19-I-O-51}. Its original labels gave 614 monocytes, 227 T cells, 68 NK cells, 46 erythrocytes and 20 NKT cells. They also gave seven dendritic cells, seven platelets, six B/plasma cells, four hematopoietic stem/progenitor cells and one neutrophil. All 1,000 cells defined the target cell proportions. An earlier plot showed only 64 cells. We selected this case after an earlier inspection of nearest cells. It is not a random cohort sample. The reference and target sets shared no cells or donors. The independent classifier used 1,974 genes. This differed from both the 1,945-gene SLE set and the 1,750-gene set used in the earlier plot.
+
+We normalized reference profiles to 10,000 and applied $\log(1+x)$. We then averaged each class and applied $L_2$ normalization to its centroid. Query profiles used the same transform. Each label came from the class with the highest cosine similarity. An added check assigned Unknown when the top-two similarity difference was below 0.05. We separately reported cell proportions from original labels and from classifier-assigned real-cell labels. This showed errors in the reference labels assigned by the classifier. Neither comparison corrects unknown labeling errors on generated cells.
+
+\subsection{Saved records and limits of the tested models}
+The fixed manifest for the five checkpoints lists hashes for checkpoints and training settings. It also records the fixed gene identifiers, assay masks and per-cell seeds. Training checkpoints save both raw and EMA parameters. All reported reverse-sampling results use EMA parameters. The main analysis keeps Unknown cells and every other generated row. Each donor's generated file keeps outputs on the model's own gene set and exact mappings into the shared evaluation set. Source data include donor-selection rules, pseudobulks, per-gene tests, cell-proportion tables and embedding coordinates.
+
+The code uses separate transforms for gene selection, bulk conditioning, diffusion values and evaluation. Gene selection and evaluation use normalized-log expression. Conditioning uses bulk and depth features from the full gene list. Diffusion predicts clipped log raw counts. Replacing any of these transforms changes the experiment. The 20,097-gene result comes only from a 128-update pilot. The main results for the three gene sets cover only the stated training checkpoints. Neither the pilot nor the planned 16-slot schedule shows final performance after full-gene training has settled.

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

@@ -0,0 +1,237 @@
+% Compile independently with pdfLaTeX; all dependencies are local.
+\documentclass[pdflatex,sn-nature]{sn-jnl}
+\usepackage{graphicx}
+\usepackage{amsmath,amssymb}
+\usepackage{booktabs,array,longtable}
+\usepackage{microtype}
+\newcommand{\method}{PBISC-Diffusion}
+\renewcommand{\thesection}{S\arabic{section}}
+\renewcommand{\thefigure}{S\arabic{figure}}
+\renewcommand{\thetable}{S\arabic{table}}
+\renewcommand{\theequation}{S\arabic{equation}}
+\begin{document}
+\raggedbottom
+\title{Supplementary Information: Diffusion for single-cell gene expression from bulk RNA data}
+\input{authors}
+\maketitle
+\raggedbottom
+
+% S1
+\input{sections/supplementary_methods}
+
+% S2
+\section{Gene-panel size at matched early exposure}
+\label{sec:supp_panel}
+Every experiment in the main text uses the 6,000-gene panel. This note records the comparison that chose it, made on early checkpoints of the base architecture; the 2,000- and 11,589-gene panels have not yet been trained to the full schedule. [TODO: 2,000- and 11,589-gene panels retrained to part 16 with the base architecture and scored on the genes common to all three.]
+
+We defined nested panels of 2,000, 6,000, 11,589 and 20,097 genes from a training-only ranking (Supplementary Fig.~\ref{fig:panel_scope}). The 11,589-gene panel holds every gene that passes fixed measurement and detection thresholds; the largest panel is the full gene list. A larger panel raises input size, output size and parameter count together, so panel comparisons test these changes jointly. In the childhood SLE test data the three trained panels contain 1,945, 5,851 and 11,352 measured genes, and their shared 1,945 genes define the common task. A 128-update pilot completed on all four panels, from 23.46 to 162.48 million parameters and 0.599 to 3.814~GiB of peak allocated CUDA memory (Supplementary Fig.~\ref{fig:panel_scope}c,d).
+
+We scored five checkpoints on 44 donors (33 cases, 11 controls), generating 204,925 cells per checkpoint to match the selected real-cell count (Supplementary Table~\ref{tab:training_updates}). With matched training exposure, coverage of the real expression space fell from 0.918 at 2K to 0.798 at 6K and 0.636 at 11,589 genes, while precision rose from 0.765 to 0.926 and 0.944; the harmonic mean was highest at 6K (Supplementary Table~\ref{tab:interim_results}, Supplementary Fig.~\ref{fig:distribution}a--g). Longer training of the 2K model raised the mean disease-effect correlation from 0.436 to 0.689 and lowered RMSE from 0.1216 to 0.1017, while coverage rose and precision fell. Broader panels widen the testable gene set; on the common genes they did not improve disease effects at matched exposure. Cell-type overlap in generated data moved in opposite directions across cell types: T-cell differential-gene Jaccard fell from 0.396 to 0.304 over three parts while monocyte overlap rose from 0.211 to 0.318 (Supplementary Fig.~\ref{fig:deg}, Supplementary Table~\ref{tab:deg_by_type}). Summed expression agreed on average and not per donor: the generated-to-real count ratio averaged 1.001 across donors with a range of 0.504 to 1.630 (Supplementary Fig.~\ref{fig:distribution}h), and composition error stayed between 0.100 and 0.139 across checkpoints. For a fixed 1,000-cell bulk, more generated cells reduced seed variation but did not correct proportions: total variation distance was $0.111\pm0.011$, $0.133\pm0.012$ and $0.148\pm0.014$ for the three panels, against 0.023 for ideal independent sampling of the reference proportions (Supplementary Fig.~\ref{fig:composition}).
+
+The clinical analysis of this note uses 1,945 shared measured genes and 44 donors. A five-type classifier replaces the unavailable original labeling model. Its cell-space plot is fitted on real data. It uses 1,464 variable genes and 33,792 cells across six sources. The separate fixed-bulk check uses 1,974 shared measured genes, ten reference types and one 1,000-cell condition; its 40,000 generated cells come from four checkpoints and ten seed streams. Both experiments also allow Unknown labels. These cell sets and labeling rules are kept separate from the held-out benchmark of the main text. Supplementary Figs.~\ref{fig:supp_welch} and \ref{fig:supp_edger} show all four cell types that met the testing rules under the Welch analysis and an edgeR check; the two use different expression scales and significance rules, so their RMSE values cannot be compared directly.
+
+\begin{table}[htbp]
+\centering
+\small
+\caption{Training updates and measured genes for each early checkpoint. E1 means one schedule slot. It does not mean a pass through all training cells. The first three rows compare gene sets after the same number of target occurrences.}
+\label{tab:training_updates}
+\begin{tabular}{lrrrr}
+\toprule
+Checkpoint & Output genes & Measured genes & Updates & Target occurrences\\
+\midrule
+2K E1 & 2,000 & 1,945 & 47,765 & 12,227,262\\
+6K E1 & 6,000 & 5,851 & 47,765 & 12,227,262\\
+11.6K E1 & 11,589 & 11,352 & 47,765 & 12,227,262\\
+2K E2 & 2,000 & 1,945 & 95,530 & 24,454,489\\
+2K E3 & 2,000 & 1,945 & 143,293 & 36,681,498\\
+\bottomrule
+\end{tabular}
+\end{table}
+
+\input{tables/results_summary}
+
+\begin{figure}[tbp]
+\centering
+\includegraphics[width=\textwidth]{figures/fig02_panel_scope.pdf}
+\caption{\textbf{Gene panels and their cost.}
+\textbf{a}, Training-only measurement and detection counts for the 20,097 fixed genes; dashed lines mark the thresholds (3,276 measured and 327 detected of 32,768 sampled cell occurrences) that define the 11,589-gene panel. One is added to both counts for the log scale.
+\textbf{b}, Output-panel sizes and the measured subset in the childhood SLE cohort; hatched parts are genes absent from that assay.
+\textbf{c}, Trainable parameters. \textbf{d}, Peak CUDA memory allocated and reserved during 128 optimizer updates at batch size 256, measured by the PyTorch allocator on one GB10 host. Source data contain the gene list and the resource table.}
+\label{fig:panel_scope}
+\end{figure}
+
+\begin{figure}[tbp]
+\centering
+\includegraphics[width=\textwidth]{figures/fig03_distribution.pdf}
+\caption{\textbf{Expression overlap and count agreement measure different properties.}
+\textbf{a--f}, UMAP of real cells and five checkpoints in one reference fitted to real cells; 5,632 cells per source (128 per donor). Generated cells are drawn over the grey real-cell reference; a cross marks the sampled Unknown cell in the 6K panel.
+\textbf{g}, Coverage and precision in the 50-dimensional reference space, using each real cell's distance to its fifteenth real neighbour as the radius.
+\textbf{h}, Generated-to-real total-count ratios on the common 1,945 genes; each point is a donor ($n=44$), black lines are donor means and the dashed line marks exact agreement. Source data provide all coordinates and donor-level measures.}
+\label{fig:distribution}
+\end{figure}
+
+\begin{figure}[tbp]
+\centering
+\includegraphics[width=\textwidth]{figures/fig04_deg.pdf}
+\caption{\textbf{Disease effects and significant-gene sets respond differently to panel size and training.}
+\textbf{a,b}, Generated versus real disease effects for T cells and monocytes across five checkpoints; each point is a common gene passing the analysis rules and dashed lines mark equal effects. Effects are differences in mean $\log(1+10^4\,\mathrm{counts}/\mathrm{library})$. T-cell and monocyte tests use 33/11 and 32/10 case/control donors.
+\textbf{c}, Jaccard overlap of full significant-gene sets for four cell types (Welch, BH FDR $<0.10$).
+\textbf{d,e}, Monocyte interferon-module effect and case--control AUC over the same 23 covered genes; dashed lines show the real reference.}
+\label{fig:deg}
+\end{figure}
+
+\begin{figure}[tbp]
+\centering
+\includegraphics[width=\textwidth]{figures/fig05_composition.pdf}
+\caption{\textbf{A fixed bulk separates sampling variation from cell-proportion error.}
+\textbf{a}, Total variation distance from the original labels for one fixed 1,000-cell bulk; points are means over ten generation seeds and error bars their standard deviations. $N=64,256,1{,}000$ are prefixes of the same draws. The dotted curve and band show the mean and 2.5th--97.5th percentiles of 10,000 ideal independent label-sampling simulations.
+\textbf{b}, At $N=1{,}000$, error against original labels and against the same real cells relabelled by the generated-cell classifier.
+\textbf{c}, Original, relabelled and mean generated proportions (percentages); every generated row stays in the denominator, including Unknown. This ten-type classifier and 1,974-gene set are separate from the main evaluation. Source data include every seed and the simulation summaries.}
+\label{fig:composition}
+\end{figure}
+
+\begin{figure}[p]
+\centering
+\includegraphics[width=\textwidth,height=.73\textheight,keepaspectratio]{figures/fig04_full_deg_welch_paper.pdf}
+\caption{\textbf{Welch effect comparisons on shared genes.} Rows show T cells, B cells, monocytes and NK cells. Columns show 2K E1, 6K E1, 11.6K E1, 2K E2 and 2K E3. Each point is a shared gene that met the testing rules. The horizontal axis shows the real-data effect. The vertical axis shows the generated-data effect. Each effect is the case mean minus the control mean for normalized-log donor expression. Dashed diagonal lines mark equal effects. Pearson correlations are calculated across genes. T, B and NK comparisons use 33 case and 11 control donors. Monocytes use 32 and 10. Each row uses the same genes and donors across checkpoints. These results describe effects from one training seed. They do not test whether one model is better than another.}
+\label{fig:supp_welch}
+\end{figure}
+
+\begin{figure}[p]
+\centering
+\includegraphics[width=\textwidth,height=.73\textheight,keepaspectratio]{figures/fig04_full_deg_edger_robustness.pdf}
+\caption{\textbf{An edgeR check on shared genes.} Layout and donor counts follow Supplementary Fig.~\ref{fig:supp_welch}. Effects are edgeR log$_2$ fold changes. We select genes using real-data expression and test the same genes in generated pseudobulks. For this check, continuous generated outputs are rounded only after summing within each donor and cell type. Dashed lines mark equal effects. Gene-wise significance uses the edgeR quasi-likelihood test with BH FDR $<0.05$ and absolute log$_2$ fold change $>1$. The scatter shows all shared genes that met the expression rule. This count-based check does not show that diffusion outputs follow an integer count model.}
+\label{fig:supp_edger}
+\end{figure}
+\clearpage
+\begingroup
+\footnotesize
+\setlength{\tabcolsep}{3pt}
+\input{tables/results_deg_by_type}
+\endgroup
+
+% S3
+\section{Sampler steps, seeds and the resolution of a score}
+\label{sec:supp_sampler}
+Five sampling seeds on one checkpoint give standard deviations of 0.005 in effect correlation, 4.7 in differential-gene overlap and 0.002 in composition, which is the resolution of any single score (Supplementary Table~\ref{tab:sampling}). Differential genes and composition are flat from 25 steps upward. Effect correlation drifts downward with more steps, from 0.258 at 50 steps to 0.240 at 200, more than two seed standard deviations, so the 50-step setting used throughout is stable for gene sets and composition but not converged for the correlation. With the same starting noise, 50-step and 200-step cells correlate at 0.996, and within-type expression variance rises from 0.63 to 0.67 of the observed variance as steps increase. Two independently trained copies of the base model---the base arm and a run that differs only by nine extra conditioning inputs the model does not use---differ by $0.009\pm0.018$ in effect correlation, $-1.6\pm7.2$ in top-50 overlap and $0.038\pm0.023$ in composition $L_1$ paired over all 16 parts; 0.018, 7 and 0.023 are treated as the smallest interpretable differences between runs.
+
+\input{tables/results_sampling}
+
+% S4
+\section{Width factorial, paired within schedule part}
+\label{sec:supp_width}
+The main text compares the four width arms as means over parts 14--16 under each read-out. Here they are compared within schedule part at the inherited read-out and averaged over parts 6--16, after every arm has left its warm-up (Supplementary Table~\ref{tab:factorial}, Supplementary Fig.~\ref{fig:supp_architecture}). Averaged over the eleven parts, effect correlation is 0.247, 0.235, 0.234 and 0.238 across the four models, a spread smaller than the run-to-run resolution of 0.018. Paired within part, the $4\times$ trunk costs $11.5\pm12.9$ top-50 genes at the narrow condition and gains nothing at the wide one. Composition separates the arms instead, and the two widths interact: widening the condition improves composition at the $1\times$ trunk ($-0.027\pm0.024$) and degrades it at the $4\times$ trunk ($+0.087\pm0.039$), so the worst composition belongs to the largest model ($0.277$) and the best to the narrow trunk with the wide condition ($0.163$, and $0.145$ at its best part). Adding a nine-dimensional cell-type proportion hint to the condition changed no score beyond the resolution, and supplying the true proportions instead of predicted ones changed the outputs by less than 0.001. Detection rates at part 16 and the inherited read-out, on the 2,980-gene panel, are 0.9246, 0.9259, 0.9344 and 0.9360 zero fraction for the four arms against 0.9155 observed; across the arms the ordering is the reverse of the disease scores, so detection rate does not order the arms, whereas within one arm it orders the read-outs (Supplementary Note~\ref{sec:supp_threshold}).
+
+\input{tables/results_factorial}
+
+\begin{figure}[p]
+\centering
+\includegraphics[width=\textwidth]{figures/figS_architecture.pdf}
+\caption{\textbf{Held-out scores of the four width arms against schedule part, inherited read-out.}
+The same 6,000-gene model trained with the denoiser trunk at $1\times$ or $4\times$ and the conditioning embedding at 512 or 2,048, with every other setting fixed. Each point is one checkpoint scored on the held-out lupus cohort after restriction to the same 2,980 genes, under one frozen scANVI annotator and its six-type gate.
+\textbf{a}, Mean log fold-change Pearson correlation across cell types.
+\textbf{b}, Top-50 differential-gene overlap, summed over cell types.
+\textbf{c}, Per-donor composition $L_1$, plotted with the axis inverted so that better is up.
+Horizontal lines mark baselines scored on the same donors, genes and annotator. Points are single checkpoints from one training seed and one sampling seed; two independently trained base models resolve about 0.018 in \textbf{a} and 7 in \textbf{b}.}
+\label{fig:supp_architecture}
+\end{figure}
+
+% S5
+\section{The support threshold sweep}
+\label{sec:supp_threshold}
+Sampling ends with one decision per gene: in the inherited read-out a predicted value is kept when its support logit exceeds zero. Sweeping that threshold on one checkpoint, with everything else held constant, moves the median detection rate from 170 to 708 genes per cell and the effect correlation from 0.231 to 0.327 (Supplementary Table~\ref{tab:threshold}). Two properties of the sweep decide how it is used in the main text. First, the threshold that matches the observed detection rate on the scoring panel (mean 251.9 genes per cell) is $-1.55$, where $r=0.291$; the setting that maximizes $r$, $-2.0$ to $-3.0$, produces 13\% to 71\% more detected genes than observed, and composition worsens monotonically as the threshold falls. The correlation is therefore rewarded past the point of fidelity, because denser pseudobulks give the donor-level test less noise. Second, coverage of the observed cell space peaks at $-1.5$ (0.821) and falls to 0.787 by $-3.0$, so cells denser than observed agree less with observed cells. These are the reasons the main text reports the drawn read-out, which has no threshold, and treats $-2.0$ as a swept setting rather than a calibrated one. Earlier drafts of this work described $-2.0$ as detection-matched; the zero fractions that supported that description were measured on the model's 6,000-gene panel, not the scoring panel, and the corrected values are the ones in the table.
+
+\input{tables/results_threshold}
+
+% S6
+\section{Restricting the value loss to expressed genes}
+\label{sec:supp_loss}
+The value term of the loss covers every measured gene, so most of its weight falls on measured zeros. We retrained the base arm with that term restricted to genes that are expressed in the target cell, changing nothing else, and ran the full 16-part schedule. The restricted model is worse at every part. At part 16 it reaches effect correlation 0.187, top-50 overlap 18 and composition $L_1$ 1.325, against 0.241, 57 and 0.227 for the model trained on all measured genes at the same read-out; composition $L_1$ is at or above 0.75 from part 5 onward and exceeds 1.0 from part 12. Training the value head on measured zeros as well as on expressed genes is what holds the cell-type structure together, so the weight the loss spends on zeros is not free capacity to reclaim.
+
+% S7
+\section{Scoring on the model's own panel}
+\label{sec:supp_ownpanel}
+The 2,980 genes scored in the main text are the genes shared by the model's panel and the panel of the semi-profiling comparison. To measure what that choice costs, we rebuilt the observed cells on the model's own panel from the raw cohort data and rescored seven generated sets on 5,847 genes. Effect correlation falls by 0.033 to 0.061 for every set, differential-gene overlap moves by $-5$ to $+20$, and composition $L_1$ changes by at most 0.003, since it is computed from labels rather than expression. A control that rescores the shared 2,980 genes with the wide-panel annotator attributes the drop to the gene set ($-0.026$ to $-0.049$) rather than to the annotator ($-0.012$ to $+0.004$). The ordering of the sets by effect correlation survives the change (Spearman $\rho=0.964$); the ordering by differential-gene overlap does not ($\rho=0.559$). The shared panel is not a neutral sample of the model's output: it was selected as the most variable genes of this cohort, so the shared genes carry 3.1 times the donor-to-donor variance of the model-only genes, and 21 model-only genes are zero in all 251,179 observed cells. The comparison method cannot be scored on the wider panel, which is why the main text uses the shared one.
+
+% S8
+\section{Condition alignment with ten paired donors}
+\label{sec:supp_calibration}
+Before any cell was generated from separately measured bulk, we asked whether ten paired bulk and single-cell donors could bring a study's bulk conditions close to its pseudobulk conditions, keeping the model, scaler, depth rule and mask rule frozen; prediction for each remaining donor uses that donor's bulk alone. ImMicro supplied 107 healthy baseline donors and Randolph 178 ex vivo measurements from 89 donors, with 100 different ten-donor selections. Affine calibration reduced the mean error of the learned 512-dimensional condition from 5.5989 to 0.1741 in ImMicro and from 2.1579 to 0.4380 in Randolph, against 4.7087 and 1.3609 for changing only the available gene set (Supplementary Fig.~\ref{fig:cohort_calibration}a,b). Exact-sample retrieval rose from 2.79\% to 38.33\% and from 11.22\% to 20.70\% (Supplementary Fig.~\ref{fig:cohort_calibration}c,d). Lower average error did not guarantee the right match: mean-ratio scaling gave higher embedding error than the affine fit in ImMicro (0.1993 versus 0.1741) but higher retrieval (50.73\% versus 38.33\%), and ratio-plus-offset was better on both counts in Randolph (0.3013 and 40.40\%). Retrieval also fell after the learned encoder, from 84.19\% and 41.43\% in input space to 38.33\% and 20.70\%. Ten pairs therefore give useful partial correction of assay differences at the condition level. The main text tests what this means for generated cells; the gene-wise mean-ratio multiplier, the estimator with the highest retrieval here, is the one used there. These condition-level results were obtained with the 2K E3 checkpoint, and the 6K and 11,589-gene E1 checkpoints were tested on the first 20 selections (Supplementary Fig.~\ref{fig:supp_calibration_robustness}).
+
+The Randolph reference used the authors' published pseudobulks after they removed their labeled ``highly infected'' cluster. We did not build new pseudobulks from all cells. The reference therefore reflects the authors' retained cells and processing steps.
+
+Let $B_{ig}$ and $P_{ig}$ be paired bulk and pseudobulk expression on a study's fixed set of shared measured genes. Each is normalized to a total of $10^6$. Calibration gave each donor equal total weight. The main ten-pair experiment used one measurement per donor. Let $\overline B_g$ and $\overline P_g$ be the weighted bulk and pseudobulk means. Let $V_g$ be the weighted bulk population variance and $C_g$ be the weighted paired covariance. The prior slope was
+\begin{equation}
+ a_{0g}=\begin{cases}\overline P_g/\overline B_g,&\overline B_g\geq1\text{ and }\overline P_g\geq1,\\1,&\text{otherwise.}\end{cases}
+\end{equation}
+For $V_g>[10^{-6}\max(|\overline B_g|,1)]^2$, the fitted coefficients were
+\begin{equation}
+ a_g=\max\!\left(0,\frac{C_g/V_g+\lambda a_{0g}}{1+\lambda}\right),\qquad
+ b_g=\overline P_g-a_g\overline B_g.
+\end{equation}
+For smaller variance, $a_g=a_{0g}$ and the same offset formula applied. The fit minimizes donor-weighted squared error plus the penalty $\lambda V_g(a_g-a_{0g})^2$. The offset has no penalty, and the slope must be nonnegative. Multiplying the penalty by training variance makes $\lambda$ unit-free. We chose one $\lambda$ for all genes from $\{0,0.1,1,10,100\}$. We left out each calibration donor in turn to compare these choices. Each inner fit recomputed its own expression statistics, prior slope and variance check. We chose the lowest mean $\log(1+x)$ expression MSE after clipping and normalization. Ties within relative and absolute tolerance $10^{-12}$ favored the larger $\lambda$. Outer test donors and their condition vectors did not help choose this value. In the main ten-pair analysis, $\lambda=0$ was selected in 97 of 100 ImMicro selections and 90 of 100 Randolph selections. The ridge candidate therefore often became least squares with nonnegative slopes and no positive penalty.
+
+For $\mathcal C(x)=10^6x/\sum_gx_g$, the final affine prediction was
+\begin{equation}
+ \widehat P_i=\mathcal C\!\left([a\odot B_i+b]_+\right).
+\end{equation}
+If no entry was positive before normalization, we used the calibration-mean pseudobulk and recorded this replacement. We recorded the fraction and total mass of negative values before clipping. We also saved positive totals before normalization, zero inputs changed to positive values, coefficient limits and replacement counts. Genes remained in the evaluation even when their mean expression failed the prior-ratio rule.
+
+Ordinary least squares with nonnegative slopes used $\lambda=0$ and the same near-zero-variance rule. Mean-ratio multiplication used $\mathcal C(a_0\odot B_i)$. Offset-only correction added the mean difference $\overline P-\overline B$ before final clipping and normalization. Ratio plus offset used the calibration mean of $P_i-\mathcal C(a_0\odot B_i)$ as its offset. It therefore fitted the remaining difference after multiplication and renormalization. It did not jointly fit a paired regression slope and offset. Support-only correction kept the shared measured bulk genes without fitting any expression coefficients. We tested separate mean-pseudobulk and mean-condition controls. The first averaged expression with all measured pseudobulk genes before applying the fixed nonlinear encoder. The second averaged encoded profiles. These operations need not give the same vector.
+
+We placed predictions in the fixed 20,097-gene order. Entries outside the shared measured set received zero fillers and false measurement flags. The original pseudobulk target kept its full measured gene set. All methods shared a fixed output mask from the single-cell assay. This mask came from the calibration data schema, not from test expression. Uncorrected bulk kept its original measured gene set. The fixed model normalized over the full gene list and computed six scalar features before selecting model genes. We did not refit the saved scaler. The main condition measures used the full $3G+6$ input and the learned 512-dimensional encoding. We also measured errors separately for each input block. This kept fixed mask entries from hiding gene-expression errors.
+
+The main comparison and the added gene-set comparisons included all nine methods. Pair-budget and held-out-group checks used six methods chosen before evaluation. They were uncorrected bulk, support-only correction, mean-ratio scaling, ridge affine correction, mean pseudobulk and mean condition. The matching reference set contained all 107 ImMicro samples or all 178 Randolph samples. It stayed fixed across methods and calibration selections. We split credit equally when several references tied for the closest match. Randolph donor matching accepted either condition of the correct donor. Exact-sample matching required the correct donor and condition. We also saved the true pair's rank and its distance margin against the nearest other sample or donor. These checks separate a match to a person's condition from closeness to the cohort mean. Percentiles across calibration selections show variation due to the selected donors within these studies. The selections share donors and test samples. We did not treat them as independent biological repeats.
+
+After the main analysis, we added a check of the pooled Randolph mean-condition control. For each test sample, we used the five calibration donors with the same known NI or IAV condition. We averaged their fixed reference embeddings. This used condition labels and calibration embeddings, not test pseudobulk expression. Across the same 100 main selections, mean MSE was 1.0472, compared with 3.0969 for the pooled mean. We then divided each method's selection-level MSE by this condition-specific control's MSE. Mean ratios were 0.4212 for ridge affine, 0.3643 for mean-ratio scaling and 0.2897 for ratio-plus-offset. This check shows how mixing conditions affects the main denominator. It did not change the original nine methods, fitted parameters, main normalization or selected models.
+
+\begin{figure}[tbp]
+\centering
+\includegraphics[width=\textwidth]{figures/fig06_cohort_calibration.pdf}
+\caption{\textbf{Ten-pair calibration reduces condition error without giving exact agreement.}
+Expression is corrected before it enters the frozen conditioning pipeline; the scores measure its learned encoding.
+\textbf{a,b}, Mean squared error of the 512-dimensional condition embedding, divided within each selection by the error of predicting the calibration-mean condition; the dashed line marks a ratio of one. Both studies use 100 selections of ten donors and ten pairs, testing 97 ImMicro or 79 Randolph donors.
+\textbf{c,d}, Exact-sample top-one retrieval against fixed reference sets of 107 or 178 pseudobulk conditions; points are selection means and lines the 2.5th--97.5th percentiles across selections. All methods share the splits, the frozen model and a fixed depth.}
+\label{fig:cohort_calibration}
+\end{figure}
+
+\clearpage
+\begin{figure}[p]
+\centering
+\includegraphics[width=\textwidth,height=.75\textheight,keepaspectratio]{figures/figS_calibration_methods.pdf}
+\caption{\textbf{All planned correction methods, evaluated before and after condition encoding.}
+\textbf{a,b}, MSE across the full $3G+6$ condition input for ImMicro and Randolph.
+\textbf{c,d}, MSE in the learned 512-dimensional encoding. For each selection, error is divided by the calibration-mean-condition error at the same stage. Points show means across the same 100 ten-pair calibration selections per study. Horizontal ranges show their 2.5th--97.5th percentiles. The dashed reference is one. All nine methods were chosen before evaluation. The mean-pseudobulk control averages expression before the nonlinear transform. The mean-condition control averages after that transform, so these controls can differ. The single-cell reference keeps its original measured genes. Source Data provide separate input-gene, scalar and mask errors. These ranges show how results change with the chosen calibration donors. They do not measure variation across independent studies or trained models.}
+\label{fig:supp_calibration_methods}
+\end{figure}
+
+\begin{figure}[p]
+\centering
+\includegraphics[width=\textwidth,height=.73\textheight,keepaspectratio]{figures/figS_calibration_size.pdf}
+\caption{\textbf{Number of calibration pairs and condition agreement on test donors.}
+\textbf{a,b}, Encoded-condition MSE divided by calibration-mean-condition MSE for ImMicro and Randolph.
+\textbf{c,d}, Exact-sample matching by Euclidean distance against the fixed full-cohort reference set. Every run uses the fixed 2K E3 checkpoint. Each calibration donor contributes exactly one pair. The two Randolph conditions are balanced as closely as the pair budget allows. No measurement from a calibration donor enters testing. Lines show selection means, and shading shows 2.5th--97.5th selection percentiles. Budgets of 5, 20 and 40 pairs use 30 selections each. The main ten-pair analysis uses 100. We evaluate all remaining donors, so the test-set size changes with the pair budget. All runs use references with their original measured genes, fixed depth and the same saved scaler. This plot does not establish a minimum calibration size for all studies or a rule for declaring equal conditions.}
+\label{fig:supp_calibration_size}
+\end{figure}
+
+\begin{figure}[p]
+\centering
+\includegraphics[width=\textwidth,height=.78\textheight,keepaspectratio]{figures/figS_calibration_robustness.pdf}
+\caption{\textbf{Checks using separate experimental groups and different fixed checkpoints.}
+\textbf{a--d}, Calibration using random donors or holding out an entire group. ImMicro is on the left and Randolph on the right. Each held-out-group fit uses ten calibration pairs from outside the test pool or experimental day. Each of ten repeats combines all seven ImMicro groups or 15 Randolph groups, weighted by their test-donor counts. Normalized error is then recomputed from the two combined MSEs. Random-donor results use 100 calibration selections.
+\textbf{e--h}, The same first 20 ten-pair selections tested with fixed 2K E3, 6K E1 and 11,589-gene E1 checkpoints. These checkpoints differ in both training amount and gene set. This is therefore not a comparison of gene-set size alone. The first and third rows show encoded MSE divided by calibration-mean-condition MSE. The second and fourth rows show exact-sample matching by Euclidean distance. Lines show selection means, and shading shows the 2.5th--97.5th selection percentiles. All references keep their original measured single-cell genes. These checks do not measure generated-cell performance.}
+\label{fig:supp_calibration_robustness}
+\end{figure}
+
+% S9
+\section{Subset-level audit: selection and scoring rules}
+\label{sec:supp_behcet}
+Two rules had to change before the C1Q-high monocyte analysis could be run on generated cells. The rule that defines C1Q-high-like cells, the 95th percentile of control monocytes, falls exactly on zero in both generated sets, with about 10,400 tied cells, so it reduces to ``any C1Q count at all'' and no longer controls the selected fraction; in the observed cells the same rule gives a strictly positive threshold with no ties. The module scores also require a depth-free statistic: C1Q-high-like monocytes are 1.49 times deeper than the other monocytes, and a per-cell mean-expression score pushes every module upward, which reverses the sign of the pro-inflammatory result in the observed data. The pseudobulk fold change with expression-matched background subtraction described in Methods is the statistic used.
+
+The read-out changes this biological result as well, and not uniformly. At the inherited read-out the selected subset is 0.85\% of control monocytes and all 19 samples hold fewer than 20 such cells; at the swept threshold of $-2.0$ it is 4.68\% against 5.00\% observed, with no sample below 32 cells. The threshold tightens the two reproduced programmes and brings a narrower MHC-II and Fc$\gamma$-receptor module from twice the observed difference to the observed difference; it also removes the phagocytosis signal, which is $+0.093$ and $+0.121$ at the inherited read-out and $+0.011$ and $-0.036$ at $-2.0$ against $+0.157$ and $+0.097$ observed. Three of four read-outs favour the lower threshold and the pre-registered one favours the default. [TODO: the same audit under the drawn read-out.]
+
+% S10
+\section{Source Data and saved records}
+The Source Data package contains unchanged numerical tables from the fixed September 2026 analyses. It also contains gene lists, the short pilot's resource measurements and independent check reports. A SHA-256 manifest links each copied file to its original record. Figure scripts read only these tables. They do not rerun generation or fit a new UMAP. We mark manuscript summary tables separately from copied source tables. The manuscript upload package does not contain raw single-cell expression matrices or checkpoints. The authors must add public data and code archive identifiers before submission.
+
+\end{document}

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

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:edf373e44050f63623b48a73f07cca90dbe15ab5808e1b03eabebbd001543d00
+size 2516919

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

@@ -0,0 +1,19 @@
+\begin{table}[t]
+\centering
+\footnotesize
+\setlength{\tabcolsep}{2.2pt}
+\caption{Capacity lowers the loss and does not move the biology; it also removes the read-out's control. Four models differ only in denoiser trunk width and conditioning width, trained for 16 schedule parts on the 6,000-gene panel with the same data, seed, batch size, learning rate and schedule. Scores are means over parts 14--16 on the held-out lupus split under three read-outs of the same checkpoints; $r$ and DEG higher-better, composition $L_1$ lower-better. Validation loss is the denoising objective at part 16. Response is the change in detected genes per cell when the support threshold is lowered by one unit, measured at part 16: the two wide-trunk models barely respond, so their sparsity cannot be set by a threshold at all.}
+\label{tab:architecture}
+\begin{tabular}{lrrrrrrrrrrrrrr}
+\toprule
+ & \multicolumn{2}{c}{Width} & & Val. & \multicolumn{3}{c}{Cut at $p>0.5$} & \multicolumn{3}{c}{Cut at $\tau=-2.0$} & \multicolumn{3}{c}{Drawn from $p$} & Response \\
+\cmidrule(lr){2-3}\cmidrule(lr){6-8}\cmidrule(lr){9-11}\cmidrule(lr){12-14}
+Model & trunk & cond. & Params & loss & $r$ & DEG & $L_1$ & $r$ & DEG & $L_1$ & $r$ & DEG & $L_1$ & genes/unit \\
+\midrule
+Base           & $1\times$ & 512 & 54.2M & 0.1968 & 0.245 & 59.7 & 0.199 & \textbf{0.326} & 65.7 & 0.208 & 0.286 & 67.7 & 0.197 & 77 \\
+Wide condition & $1\times$ & 2,048 & 102.1M & 0.1969 & 0.245 & 61.0 & 0.161 & 0.304 & \textbf{70.0} & \textbf{0.158} & 0.278 & 68.0 & 0.161 & 79 \\
+Wide trunk     & $4\times$ & 512 & 240.0M & \textbf{0.1355} & 0.265 & 60.7 & 0.181 & 0.271 & 64.3 & 0.182 & 0.268 & 61.7 & 0.181 & 10 \\
+Both wide      & $4\times$ & 2,048 & 312.6M & 0.1356 & 0.253 & 58.7 & 0.264 & 0.262 & 62.0 & 0.264 & 0.259 & 59.3 & 0.264 & 6 \\
+\bottomrule
+\end{tabular}
+\end{table}