Datasets:

junidude14
/

PBISC-data

paper-releases/pbisc-diffusion/updates/20261012T130738Z-strictpairs-implementation/diffusion/dataset/server_1/model8/plans/paper_figures_strictpairs_20261012/archive/20261012T130738Z-strictpairs-implementation/snapshot/manuscript/sections/fig6.tex ADDED Viewed

	@@ -0,0 +1,3 @@

+\mainfigure{fig:fig6}{Bulk-only reproduction (to be continued)}
+\tbcpanel{Waiting for the user-prepared bulk-only reproduction data.}
+\figlegend{\textbf{Pending panel.} No new bulk-only analysis or discovery claim is included in this draft.}

	@@ -0,0 +1,19 @@

+% Introduction updated from JH Word edits on 2026-10-12; tracked insertions accepted and deletions omitted.
+% Minimal grammar and mechanism wording corrections plus citations are recorded in editorial/jh_edit_integration_20261012.json.
+\section*{Introduction}
+Bulk RNA sequencing of blood and tissue remains the most widely used transcriptomic assay in clinical research, including large clinical trials, biobanks and disease cohorts. Each bulk profile, however, is a mixture: changes in cell-type abundance are confounded with changes in cell-state-specific expression, and disease signals carried by minority populations are diluted. Single-cell RNA sequencing resolves these components, but cost and sample logistics still limit it to tens or hundreds of donors per study.
+Computational deconvolution has been the principal bridge between the two. Digital cytometry methods such as CIBERSORTx\cite{newman2019cibersortx}, MuSiC\cite{wang2019music}, Bisque\cite{jew2020bisque}, DWLS\cite{tsoucas2019dwls}, Scaden\cite{menden2020scaden} and TAPE\cite{chen2022tape} estimate cell-type proportions from bulk using single-cell-derived signatures, and BayesPrism\cite{chu2022bayesprism}, BLADE\cite{barbosa2021blade} and bMIND\cite{wang2021bmind} additionally infer cell-type-specific expression. Despite of the highly advanced calculation algorithms, the accuracy of each model's output highly depends on data transformation, single cell profile reference choice and annotation granularity, degrades for closely related and rare cell types, and changes rank when the reference differs from the bulk cohort\cite{avilacobos2020benchmarking,sturm2019evaluation,dietrich2026omnideconv,community2024assessment}. Also these methods' outputs are limited to cell-type proportions and expression averages, thus cannot represent heterogeneity within a cell type.
+Advances in machine learning have enabled generative approaches to infer single-cell expression profiles from bulk transcriptomes. Bulk2Space\cite{liao2022bulk2space} and BulkTrajBlend\cite{zeng2024omicverse} use estimated cell-type composition to guide generation from single-cell reference data, whereas scSemiProfiler\cite{wang2024scsemiprofiler} adapts profiles from selected cohort representatives using bulk-derived constraints on aggregate expression. However, these approaches remain dependent on selected references, whose coverage and representativeness can influence the inferred cellular landscape\cite{liao2022bulk2space,zeng2024omicverse,wang2024scsemiprofiler}. Moreover, methods that organize single cell generation around pre-defined cell-types or clusters also depend on how these populations are defined\cite{liao2022bulk2space,zeng2024omicverse}, which motivate approaches that do not require cell-type identities or proportions to be specified in advance.
+Diffusion-based methods, including scDiffusion\cite{luo2024scdiffusion}, have expanded the possibilities for generating cellular profiles. However, realistic single-cell generation still requires faithful reproduction of the sparse patterns of zero and nonzero expression observed in experimental data. Meanwhile, Sparse Data Diffusion introduced ``sparsity bits'' to represent these patterns explicitly alongside expression magnitudes, improving the fidelity of sparse data generation in biological and physical datasets\cite{ostheimer2025sparse}. This approach provides a basis for jointly modeling which genes are detected and their expression levels when reconstructing cellular heterogeneity.
+A further obstacle for such bulk transcriptome deconvolution is the scarcity of experimentally paired bulk and single-cell measurements. Large single-cell datasets can provide abundant bulk-to-single-cell training sets by aggregating cells into pseudobulk profiles, but these simulated profiles do not fully reproduce the expression distribution of measured bulk RNA-seq due to technical and sample-preparation differences, such as differences in sequencing depth\cite{wang2024scsemiprofiler} and cell loss during dissociation of solid tissues\cite{denisenko2020systematic}. Systematic differences between pseudobulk and measured bulk can therefore limit the transfer of models to experimental data\cite{delafuente2025sweetwater}. Learning broadly from available single-cell resources and subsequently adapting to real bulk measurements with a small amount of paired data offers a practical strategy for addressing this limitation.
+Here we introduce PBISC, a bulk-conditioned diffusion model that combines pretraining on over 33 million single cells across 22 scRNA-seq datasets with low-rank adaptation (LoRA)\cite{hu2022lora} using paired bulk and single-cell data from just five donors. The pretraining step establishes a shared model of cellular expression across cohorts, reducing reliance on selecting a separate single-cell reference for each target. We incorporated the sparsity-bit concept from Sparse Data Diffusion\cite{ostheimer2025sparse} to jointly model gene-detection patterns and expression levels. With PBISC, we were able to generate cells from independent random initializations under a shared bulk condition, without requiring cell-type annotations or predicted cell-type proportions during training or generation.
+To adapt PBISC to measured bulk RNA-seq, we fine-tuned the pretrained model using experimentally paired bulk and single-cell profiles, with as few as five paired samples in the evaluated settings. During adaptation, measured bulk profiles served as conditioning inputs, while observed single-cell profiles from the corresponding biological samples provided the training targets. We then applied the adapted model to generate single-cell profiles for the remaining cohort using bulk measurements alone.
+In various evaluation standards, we were able to reproduce the distributions of paired experimental single-cell profiles, as assessed by precision\cite{kynkaanniemi2019improved} and coverage\cite{naeem2020reliable}. PBISC-generated deconvoluted dataset preserved disease-associated expression changes more accurately than comparator methods in most evaluated settings. Furthermore, PBISC reproduced literature-defined rare cell populations\cite{reyes2020immune,nehar2020mapping} observed in the single-cell data of the evaluated cohorts, without specifying these populations as separate generation categories. Together, these results support extending analyses of cellular heterogeneity and disease-associated transcriptional programs to large cohorts primarily profiled by bulk RNA-seq.

	@@ -0,0 +1,8 @@

+\subsection*{Scoring and statistical units}
+Observed and generated cells are read out under the same fixed-seed count and frozen cell-annotation pipeline. The healthy comparison reports precision, coverage, composition L1, $k$-nearest-neighbour classifier AUC, MMD$^2$ and scSemiProfiler's normalised error $E_N$ on the same held-out samples and measured-gene axis for every arm\cite{kynkaanniemi2019improved,naeem2020reliable,lopezpaz2017revisiting,gretton2012kernel,wang2024scsemiprofiler}. Real-reference UMAP and marker panels are displays; separation metrics use the frozen high-dimensional embedding. Sampling uncertainty is shown across the five outer folds or disease draws, while donor-level intervals and tests cluster all visits from one donor. Repeated cells or visits are never independent biological replicates.
+Disease expression scores compare generated and observed case--control contrasts within cell type: gene-wise logFC correlation, bulk-adjusted partial correlation, significant top-50 DEG overlap and correspondence at the top. Case and control summaries are first formed per donor; repeated visits stay in that donor's cluster. Where deconvolution provides fractions and cell-type expression, those outputs are compared only on supported composition and expression measures. BayesPrism and CIBERSORTx do not generate individual cells, so precision, coverage, classifier AUC, MMD$^2$ and $E_N$ do not apply to them\cite{chu2022bayesprism,newman2019cibersortx}. Donor-aware continuous-expression limma is the primary disease sensitivity where estimable; the historical sample-level Welch results remain separately labelled.
+Rare states use fixed required-marker rules, actual donor identifiers and the same cell-count budget and declared depth target for observed and generated samples. A missing required marker makes a state ineligible for that cohort. The state rate is undefined when its parent cell count is zero, while an observed zero among a nonzero parent remains zero. Every generated zero-state draw remains in the aggregation. Donor ranking is an abundance-ranking diagnostic and is not a binary state-detection claim. Confidence intervals resample donor IDs, not visits or cells. Beh\c{c}et bulk and single-cell donors are mostly unmatched, and its any-C1Q-gene rule is weaker than the multi-marker paired-state definition; results are descriptive only\cite{zheng2022c1qbehcet}.
+Operational compute includes cohort-specific model adaptation and generation receipts and excludes population pretraining, data preparation and scoring. Historical timing from a different support design is not substituted for the pending strict timing. Figure~\ref{fig:fig6} awaits the user-prepared bulk-only reproduction data and makes no discovery claim.

	@@ -0,0 +1,64 @@

+% Methods, part 1: model, data and pretraining. Written 2026-10-10 from the paper-summary deck (slides 6-14, 42-43)
+% and checked against the code and run records:
+%   M8/src/pbisc_sdd/{model.py, diffusion.py, data.py, gene_selection.py, panel_ladder.py, ema.py, ladder_experiment.py}
+%   M8/scripts/{train_allcell_sdd.py, sample_sdd_standalone.py, prepare_gse135779_scsemi_target42_conditions.py}
+%   M8/runs/f01_6k_1x_seed20260627/{effective_config.json, RUN_STATUS.json, ALLCELL_DATA_CONTRACT.json, history.json}
+%   M8/artifacts/panel_ladder_floor01_1m_20260923/manifest.json; Model 7 E2 index and conditioner (data_e2.py, data.py)
+%   manuscript/source_data/corpus/{README.md, corpus_summary.json}
+%   M8/plans/{aifi_scanvi_rebuild_20260916, utility_studies_final6k_20260925/sle, floor01_campaign_20260923,
+%     arch_ablation_6k_20260913, prop_condition_6k_20260916}; DECK/data/sec01_*, sec02_*, sec03_*, sec04_sle_*.
+% Exact settings and formulas that do not fit here are in supp_methods_model.tex (Supplementary Methods).
+% Fragment only: \input by main.tex; this file opens the Methods section.
+\section*{Methods}
+\subsection*{Overview of \method{}}
+\method{} generates single-cell expression profiles from one bulk RNA-seq profile. It takes the bulk counts, the number of cells $K$ that the bulk stands for and the genes that the assays measured, and draws each requested cell independently. All cells from one bulk share one condition; they differ only in their starting noise and in a random draw that decides which genes they detect (Fig.~\ref{fig:fig1}). The model never receives cell-type labels, cell-type proportions, disease labels or donor identities, and nothing forces the generated cells to add up to the bulk.
+The model adapts sparse data diffusion (SDD)\cite{ostheimer2025sparse}, which diffuses each gene's expression value together with its detection support (whether a measured gene has a positive count) in gene space, with a skip-connected multilayer perceptron as the denoiser. We kept these parts, added the bulk condition, restricted the loss to genes that the assay measured, and replaced SDD's final hard cut on the support with a draw from the predicted detection probability (Fig.~\ref{fig:fig1}).
+\subsection*{Pretraining corpus and data splits}
+The corpus held 41,395,184 cells from 22 public scRNA-seq datasets of peripheral blood mononuclear cells (PBMCs), with 3,199 donors, 6,507 samples and 432 count files (Fig.~\ref{fig:fig1}; Supplementary Note~1 and Supplementary Fig.~1\todo{cite the 22 corpus datasets in Supplementary Note 1}). Two of the datasets are enriched, one for circulating haematopoietic stem and progenitor cells and one for B cells. We used the raw unique molecular identifier (UMI) counts of every cell, on a fixed axis of 20,097 protein-coding genes (GENCODE release 49). Each count file carries a binary mask of the genes its assay measured, so a measured zero is never confused with a missing gene.
+We split each dataset by donor (by sample where the donor was unknown) into training, validation and test sets of 80\%, 10\% and 10\% of donors. The training split held 33,141,380 cells from 5,232 samples. No sample is in two splits, but in two datasets the corpus donor key is finer than the person: 26 subjects of GSE274680 are keyed by visit, and the one patient of GSE132802 by sample (time point or culture). These 27 people have samples in more than one split, and 38 validation or test samples (659,112 cells) come from a person who also has training samples.
+\subsection*{Pseudobulk recipes and training pairs}
+A recipe is a set of $K$ cells drawn at random, without replacement, from one training sample; its pseudobulk is the sum of their raw counts. Recipe sizes were $K\in\{250, 500, 1{,}000, 2{,}000, 4{,}000\}$ with relative weights 1:2:3:4:1. A fixed pool of 131,072 training recipes was registered before training and divided into 16 disjoint slots of 8,192. Each epoch, also called a schedule part, used one slot, so the 16 epochs passed once through the pool. Square-root weights kept large platforms, datasets and samples from dominating (Supplementary Methods).
+All $K$ cells of a recipe became training targets, each paired with the pseudobulk of its full recipe, itself included (Supplementary Fig.~2). The registered recipes reached 94.0\% of training cells. The other 1,995,930 cells were placed, each exactly once, in 3,658 top-up recipes built from single samples ($K$ from 1 to 4,000) and spread over the 16 epochs. Pretraining thus used every training cell at least once: 134,730 recipes and 195,639,930 target-cell occurrences in all.
+\subsection*{Gene panel}
+We drew 128 cells at random from each of 8,192 training recipes (1,048,576 cells), scaled each cell to 10,000 counts and took $\log(1+x)$. A gene was eligible if it was measured in at least 10\% of these cells and detected in at least 0.1\%; 15,114 of 20,097 genes passed. Eligible genes were ranked with immune and interferon marker genes first (55 genes), then by dispersion, $\log(1+\mathrm{variance}/\mathrm{mean})$, standardised within 20 bins of mean expression. Panels of 2,000, 3,000, 6,000 and 15,114 genes are prefixes of this ranking, so each contains the smaller ones (Fig.~\ref{fig:fig2}). The 0.1\% floor replaced a 1\% floor, which excluded genes specific to cell types rarer than 1\% of PBMCs. The final model uses the first 6,000 genes, the smallest panel that holds \textit{CLEC9A} (rank 3,793), a marker of type 1 conventional dendritic cells.
+\subsection*{Bulk condition}
+For a bulk or pseudobulk with counts $b_g$ over the 20,097 genes, total $L=\sum_g b_g$ and nominal cell count $K$, the gene feature is the bulk scaled to the depth of one average cell, $u_g=\log\!\bigl(1+b_g/\max(L/K,10^{-6})\bigr)$. Six scalars describe depth and profile shape: $\log(1+L)$, $\log(1+L/K)$, the fraction of genes with $b_g>0$, the mean and standard deviation of $\log(1+b_g)$, and the share of $L$ held by the 1\% highest-count genes. Features were standardised with constants from 2,048 training recipes and clipped to $[-8,8]$, and the 6,000 panel genes were kept afterwards. Two masks complete the condition: $f_g$, the fraction of the $K$ cells whose assay measured gene $g$, and $m_g$, the measured-gene mask of the assay that the generated cells should match:
+\begin{equation}
+c=\bigl[\,z(u_{\mathcal G}),\; z(a),\; 2f_{\mathcal G}-1,\; 2m_{\mathcal G}-1\,\bigr]\in\mathbb R^{3G+6},
+\end{equation}
+where $\mathcal G$ is the panel of $G=6{,}000$ genes, $a$ holds the six scalars and $z(\cdot)$ is the clipped standardisation; $c$ has 18,006 entries (Fig.~\ref{fig:fig1}).
+\subsection*{Joint diffusion of values and detection support}
+Each measured gene of a cell with raw counts $x_g$ contributes a value and a detection sign,
+\begin{equation}
+v_g=2\,\mathrm{clip}\!\left(\tfrac{1}{8}\log(1+x_g),\,0,\,1\right)-1,\qquad s_g=+1\ \text{if}\ x_g>0,\ \text{else}\ -1.
+\end{equation}
+Unmeasured genes were set to 0 in both halves and never entered the loss. The clean state $z_0=[v,s]\in\mathbb R^{2G}$ was noised in closed form\cite{ho2020ddpm} as $z_t=\alpha(t)z_0+\sigma(t)\varepsilon$, with $t\sim\mathcal U(0,1)$ per cell, $\varepsilon\sim\mathcal N(0,I)$, $\alpha^2+\sigma^2=1$ and a cosine schedule\cite{nichol2021improved} (offset 0.008; Supplementary Methods). The denoiser predicts clean values $\hat v_g$ and detection logits $\ell_g$. Over the set $M$ of measured entries of a batch, the loss is
+\begin{equation}
+\mathcal L=\frac{1}{|M|}\sum_{(i,g)\in M}\Bigl[(\hat v_{ig}-v_{ig})^2+\mathrm{BCE}\bigl(\ell_{ig},\mathbb I[x_{ig}>0]\bigr)\Bigr],
+\end{equation}
+with BCE the binary cross-entropy on logits. The value term includes measured zeros, and the support head learns the log-odds that a gene is detected. In a random half of the batches, a first prediction made without gradients, $[\hat v,\tanh(\ell/2)]$, was fed back as an extra input (self-conditioning\cite{chen2023analog}); otherwise that input was zero.
+\subsection*{Denoiser architecture}
+The denoiser is a skip-connected multilayer perceptron (Supplementary Fig.~2). It takes the noisy state and the self-conditioning input (24,000 values) through seven blocks of widths 1,024, 1,024, 512, 256, 512, 1,024 and 1,024. Each block adds a linear map of its input to a linear map of the shared context passed through a sigmoid linear unit (SiLU), then applies layer normalisation and SiLU. The outputs of blocks 5, 6 and 7 are summed with those of blocks 3, 2 and 1, their mirror blocks of equal width. The context, of 512 dimensions, combines the encoded time (128 sinusoidal features) and the encoded condition. A shared 1,024-wide layer feeds two heads of 6,000 outputs, for values and detection logits. The model has 54,185,184 parameters and no dropout.
+\subsection*{Pretraining settings}
+We used AdamW\cite{loshchilov2019decoupled} with a learning rate of $2\times10^{-4}$ and no weight decay, batches of 256 target cells, gradient clipping at norm 5, bfloat16 mixed precision (PyTorch autocast, with 32-bit weights) and seed 20260627. An exponential moving average (EMA) of the weights was updated after every step with decay 0.999. Training ran for 16 epochs, or 764,249 steps. After each epoch the loss was computed with the EMA weights on a fixed validation set of 192 recipes from validation donors, 64 cells each, with fixed noise. It fell at every epoch, from 0.153 to 0.137. The EMA weights after epoch 16 are the pretrained model used throughout. Pretraining took 15.9 h on one NVIDIA GB10 GPU and was run once.
+\subsection*{Sampling and the drawn support read-out}
+Generation used the EMA weights and 50 deterministic steps of the denoising diffusion implicit model (DDIM) sampler\cite{song2021ddim} ($\eta=0$). Each step's clean prediction was also the next step's self-conditioning input. Each cell's starting noise came from its own seeded generator on the CPU, so it is the same whatever the batch size or hardware (Supplementary Methods).
+At the last step, $p_g=\mathrm{sigmoid}(\ell_g)$ is the probability that gene $g$ is detected. We kept gene $g$ when a uniform draw $u_g$ fell below $p_g$ and the output mask allowed it; otherwise its value was exactly zero. A kept gene takes the count-like value $\exp\{8\,\mathrm{clip}[(\hat v_g+1)/2,0,1]\}-1$. This drawn (Bernoulli) read-out has no free parameter; a cell's expected number of detected genes is the sum of $p_g$ over the genes that the output mask allows. The pretraining design rescoring in Fig.~\ref{fig:fig2} uses saved epoch-16 Bernoulli outputs. Unless stated otherwise we generated 1,500 cells per bulk. A measured bulk has no cell count, so it was given $K=1{,}000$ at a fixed depth (see ``Paired bulk and single-cell cohorts'').
+\subsection*{Held-out pretraining design rescore}
+Saved architecture and gene-panel outputs for the 42 held-out SLE donors were rescored separately. On the common 1,006-gene axis, each donor contributed 100 real reference cells, 100 disjoint real query cells and 100 generated cells per arm. One fixed Poisson read-out, a real-reference 50-component PCA and $k=10$ local neighbourhoods defined precision and coverage. The architecture and panel families had different training detection floors and are not a factorial comparison. One saved generation seed per arm makes these descriptive scores.