Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/sections/results_fig5.tex

@@ -0,0 +1,17 @@
+% Results, Figure 5: rare cell states in the paired disease cohorts.
+% Sources: deck slides 49-52 (main panels), 48 (panel plan). The from-scratch rare-state result (slide 72) stays in
+% Supplementary Note 3 / Supplementary Fig. 16 and the Discussion only (brief rule 1.12; removed here 2026-10-11).
+% Data: DECK/data/fig5_rare_states.csv, DECK/data/fig5_rare_states_slide_text.json, DECK/data/sec05_behcet_direction.csv.
+\subsection*{LoRA cells keep rare cell states closer to their observed fractions than scSemiProfiler's}
+
+\textbf{Rare states are what single cells add.} Bulk averages rare cell states away, although many of them mark disease. We scored 14 states from the literature, each defined by marker genes or by an immune-atlas label\cite{gong2025multiomic}. They include MS1 monocytes\cite{reyes2020immune}, cycling monocytes\cite{ye2026sepsis}, C1Q-high monocytes\cite{zheng2022c1qbehcet}, T cells high in interferon-stimulated genes (ISG-high T cells)\cite{nehar2020mapping}, DN2 B cells\cite{jenks2018distinct}, plasmablasts\cite{mathew2020deep} and granulocyte-like cells that resemble low-density granulocytes\cite{bennett2003interferon}. A state's fraction is its share of the parent cell type, or of all cells for plasmablasts, haematopoietic stem and progenitor cells and granulocyte-like cells. We used the disease benchmark's cells and frozen scANVI labels\cite{xu2021scanvi} (Fig.~\ref{fig:fig4}) and thinned the observed parent cells of marker states to the generated depth, because marker detection depends on depth. In the four paired disease cohorts the 14 states gave 56 state-by-cohort pairs, each with at least 30 observed state cells at native depth and with state cells in at least 3 samples after thinning to the depth of LoRA's cells (Methods).
+
+\textbf{Closer to the observed fractions.} In the main comparison, LoRA-5 was closer to the observed fraction than scSemiProfiler-5 in 51 of 56 pairs (Fig.~\ref{fig:fig5}a; mean of five draws per cohort). It was within 2-fold (half to twice the observed fraction) in 19 of 56 pairs, against 5 of 56. scSemiProfiler-5 never generated the state in 6 pairs; LoRA-5 generated it in all 56. Most states were still generated below their observed fraction (50 of 56 pairs for LoRA-5), and zero-shot came closer than LoRA-5 on 28 of the 56 fractions.
+
+\textbf{More often in the right samples.} LoRA-5 placed rare cells in the right samples more often than scSemiProfiler-5 (Fig.~\ref{fig:fig5}b). Donor ranking, the Spearman correlation over test samples between generated and observed state fractions, had a median of 0.24 for LoRA-5 against 0.18 for scSemiProfiler-5. LoRA-5 ranked samples better than scSemiProfiler-5 in 43 of 56 pairs and better than zero-shot in 35 of 56. The case-versus-control difference kept its observed sign in 73\% of state-by-draw cases for LoRA-5, against 61\% for scSemiProfiler-5 and 66\% for zero-shot.
+
+\textbf{The same with every training pair.} Reported separately, 5-fold CV, in which scSemiProfiler received LoRA's training and inner-validation pairs, gave the same picture (Fig.~\ref{fig:fig5}c). LoRA was within 2-fold in 25 of 56 pairs against 9 for scSemiProfiler, had a median donor ranking of 0.32 against 0.24, and ranked samples better in 38 of 56 pairs.
+
+\textbf{Without pairs, real bulk can flip a disease direction.} Beh\c{c}et disease has no designed pairs: its bulk (9 patients, 10 controls) and single-cell samples come mostly from different people\cite{zheng2022c1qbehcet}, so we used the pretrained model without fine-tuning (Fig.~\ref{fig:fig5}d). From the real bulk, the model reversed the patient-minus-control gap in the CD16 share of monocytes ($-3.97$ against $+4.94$ percentage points observed) and in C1Q-high monocytes (any C1Q gene detected; $-0.77$ against $+1.32$ points), in 5 of 5 seeds. The observed gaps come from 4 patients and 4 controls (descriptive Mann--Whitney $P=0.029$ for CD16 and $P=0.49$ for C1Q-high). The C1Q-high reversal was a depth effect: at one common monocyte depth its gap shrank to $-0.08$ points (mean of 5 seeds; exact permutation $P=0.60$; observed $+1.21$). At that depth the CD16 reversal remained ($-3.59$ points, one seed). From each single-cell donor's own pseudobulk, the same model gave the observed sign for both in 5 of 5 seeds (CD16 $+5.48$, C1Q-high $+1.82$ points). This points to the shift from pseudobulk to real bulk (Fig.~\ref{fig:fig2}a), not the generator, as the cause of the flip. The bulk and single-cell donors mostly differ, so donor differences may also contribute. Paired samples teach the model this shift (Fig.~\ref{fig:fig2}). \todo{add the LoRA-3 result: Beh\c{c}et C1Q-high direction with three paired donors (three patients whose bulk and single cells were matched afterwards by expression; tested on the 16 bulks without single cells; say that calibration fitted on the same three pairs gives the same direction and that cells with two or more C1Q genes stay under-generated)}
+
+\textbf{Cohorts absent from the corpus come next.} Fig.~\ref{fig:fig5}e will score MS1 monocytes in Kawasaki disease\cite{wang2021kawasaki} and ISG-high T cells in the SLE cohort of children and adults (GSE135779)\cite{nehar2020mapping}. \todo{panel e: states from cohorts absent from the corpus (Kawasaki MS1 monocytes, SLE ISG-high T cells)}

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/sections/results_fig6.tex

@@ -0,0 +1,5 @@
+\subsection*{Cells for public bulk-only samples (to be continued)}
+
+\textbf{The use case is bulk without single cells.} The sepsis and Imperial cohorts also hold many samples that were sequenced as bulk only. The sepsis cohort has 122 such samples (68 adult sepsis, 36 adult healthy controls, 17 adult and 1 paediatric intensive-care controls) and Imperial has 164 (143 COVID-19 positive, 19 negative, 2 recovery); 45 of the Imperial donors have no single cells at all. For each of these samples we generated cells with the five LoRA-5 models of Fig.~\ref{fig:fig4}b, so every sample has five generations, each from a model fine-tuned on the pairs of a different set of five donors. We generated them twice. At bulk depth $K=1{,}000$, each of the five models generated 1,500 cells per sample from the same sampling seed. LoRA-5 models retrained on the same donors at $K=4{,}000$ generated 4,000 cells per sample, each model with its own seed. All cells were labelled with the same frozen annotator as the paired benchmark (Methods).
+
+\todo{Figure 6 and this subsection are to be continued: report the analysis of the bulk-only cells (clinical associations of the rare states scored in Fig.~\ref{fig:fig5}, each finding gated by the paired calibration of its draw, replicated across the five draws, and tested against a bulk marker score).}

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

@@ -0,0 +1,389 @@
+% Supplementary Figures 1-17 (fixed order of the brief). Written 2026-10-10 from the paper-summary deck
+% (slides 9, 13, 14, 44-46, 58-72, 74) and DECK/data/*.csv. Panels are the deck's vector graphics, copied to
+% figures/deck_20261010/ (Supplementary Fig. 1 is the manuscript's figures/figS_corpus.pdf).
+% Main-text figures are cited in plain text (Fig. 3d), because supplementary.tex compiles on its own.
+\setcounter{figure}{0}
+
+% ---------------------------------------------------------------------------------------------------------------
+% Supplementary Fig. 1. Source: manuscript source_data/corpus/corpus_summary.json, corpus_datasets.csv, README.md;
+% deck slide 9.
+\suppfigure{sfig:1-corpus}{The pretraining corpus}
+\figpanelh{}{figures/figS_corpus.pdf}{0.72}
+\figlegend{\textbf{The pretraining corpus holds 41.4 million cells from 22 public datasets, 3,199 donors and 6,507 samples.}
+\textbf{a}, Shares of cells by split (donor-level split within each dataset, which leaves 27 of the 3,199 donors with samples in more than one split; training 80.1\%, validation 10.5\%, test
+9.4\%), 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. Other assays:
+Seq-Well and 10x assays of unreported or mixed chemistry.
+\textbf{b}, The 22 datasets: cells (log scale), assay, 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: raw counts over 20,097 genes with a measured-gene mask, and the sample,
+platform, dataset and donor fields that define the training pseudobulks and the splits. Condition and cell-type labels
+describe the corpus and are never model inputs. Numbers and sources are given in Supplementary Note~1.}
+
+% ---------------------------------------------------------------------------------------------------------------
+% Supplementary Fig. 2. Source: deck slides 13 and 14; DECK/data/sec01_model_numbers.csv, sec01_recipe_k.csv,
+% sec01_corpus_datasets.csv.
+\suppfigure{sfig:2-model}{Model details: training pairs and denoiser}
+\figpanel{a}{figures/deck_20261010/sec01_slide_a.pdf}
+\figpanel{b}{figures/deck_20261010/sec01_fig1a_denoiser.pdf}
+\figlegend{\textbf{\method{} learns each cell together with the pseudobulk it came from, and generates cells from one
+measured bulk.}
+\textbf{a}, Top, generation. One measured bulk sample (its profile over $G$ genes, depth, number of cells $K$ and assay
+mask $m$) becomes one condition of $3G+6$ values (18,006 for $G = 6{,}000$), which a multilayer perceptron maps to a
+512-dimensional context shared by all $K$ cells. The same denoiser draws each cell independently with 50 deterministic
+denoising diffusion implicit model (DDIM) steps\cite{song2021ddim}; cells differ only in their starting noise and in the random draw that decides which genes
+are detected. Cell types are labelled after generation, so no cell-type counts are fixed, and the sum of the cells is
+compared with the bulk as a check, not a constraint. Bottom, training data of the final model: cells per corpus dataset
+(log scale; 41.4 million cells, 33.1 million in the donor-level training split) and training recipes per epoch. A
+recipe is a pseudobulk of $K$ cells from one sample ($K$ = 250, 500, 1,000, 2,000 or 4,000; 745, 1,489, 2,234, 2,979
+and 745 recipes per epoch), and each member cell is a training target with that pseudobulk as its condition. Top-up
+recipes (196 to 283 per epoch, $K$ from 1 to 4,000; 3,658 in all) make every training cell a target at least once over the 16
+epochs.
+\textbf{b}, The denoiser\cite{ostheimer2025sparse}. Each gene has a value channel and a support (detection) channel, so a
+cell is a vector of $2G$ values, and training adds Gaussian noise directly in this space. A conditioned multilayer
+perceptron with seven blocks, whose width narrows from 1,024 to 256 and widens back, and skip sums (54.2 million
+parameters) takes the noisy cell joined with a soft self-conditioning copy of its previous prediction ($4G$ values).
+The context, which mixes the bulk condition with the diffusion time, enters every block. It predicts gene values and
+support logits. A gene is kept when a uniform random number falls below the sigmoid of its logit (Bernoulli read-out);
+kept values are rescaled and masked by the assay, giving non-negative expression with exact zeros. Sampling uses the
+epoch-16 exponential moving average weights.}
+
+% ---------------------------------------------------------------------------------------------------------------
+% Supplementary Fig. 3. Source: deck slide 58; DECK/data/sec09_fewpair_cells.csv (scores[_nosex]/<cohort>/fewshot_n5/).
+\suppfigure{sfig:3-fewpair-cells}{5 donors, all five cohorts: cell metrics}
+\figpanel{}{figures/deck_20261010/sec09_fewpair_cells.pdf}
+\figlegend{\textbf{With the same five paired donors, LoRA-5 covers the real cells better than scSemiProfiler-5 in all
+five cohorts.} Main comparison (5 donors). In each draw, LoRA-5 (the pretrained model fine-tuned with low-rank adaptation (LoRA)\cite{hu2022lora},
+rank 16, for 3,000 steps on all pairs of five donors) and scSemiProfiler-5\cite{wang2024scsemiprofiler} (given the
+same five donors' single cells as representatives) generate every other usable sample of the cohort from its real
+bulk. Five draws per cohort (dots; bars, mean $\pm$ s.d.). Training pairs and test samples per draw are shown on the
+left; Imperial, Giroux and Ji have several samples per donor, Giroux test sets hold two moderate and two mild subjects,
+and Ji bulk is whole blood. Zero-shot: the pretrained model fed the real bulk as measured. Zero-shot + calibration: the
+same model fed the bulk after gene-wise calibration fitted on the draw's training pairs ($\alpha_g$, mean pseudobulk counts per million (CPM)
+over mean bulk CPM). Dashed line: the pretrained model fed each test sample's own single-cell pseudobulk, a reference
+that needs that sample's single cells. Coverage, share of real cells with a generated cell inside their
+$k$-nearest-neighbour ball\cite{naeem2020reliable}; precision, share of generated cells inside the ball of at least one
+real cell\cite{kynkaanniemi2019improved}; $k = 10$, in a 50-component principal component analysis (PCA) fitted on real cells. Composition L1, L1
+distance between generated and observed fractions of six scANVI\cite{xu2021scanvi} cell types (lower is better).
+LoRA-5 is ahead of scSemiProfiler-5 on coverage (0.69 to 0.96 against 0.04 to 0.61) and precision in 5 of 5 cohorts.
+It has lower composition L1 in 4 of 5 cohorts by the draw verdict (lower in at least 4 of 5 draws; Giroux 3 of 5 draws)
+and in 5 of 5 by cohort means. LoRA-5 covers more real cells than zero-shot + calibration in 5 of 5 cohorts, but its
+precision is higher in only 3 of 5 by the same draw verdict (sepsis and Imperial: 3 of 5 draws each). Its per-donor
+composition L1 is significantly lower than that of zero-shot + calibration in ImMicro only (difference $-0.03$,
+$P = 2\times10^{-10}$; paired Wilcoxon test over test donors, uncorrected); in sepsis and Imperial it is slightly higher ($+0.02$; $P = 0.79$ and $0.38$). It is significantly lower than that of zero-shot in four cohorts (all but sepsis, where it is lower by 0.04, $P = 0.15$). After depth matching (real and generated cells of the same donor thinned
+to the same number of detected genes) the ranking is unchanged (coverage 0.83 to 0.97 for LoRA-5, 0.15 to 0.28 for
+scSemiProfiler-5). Per-cohort values: Supplementary Table~2.}
+
+% ---------------------------------------------------------------------------------------------------------------
+% Supplementary Fig. 4. Source: deck slide 59; DECK/data/sec09_extra_fewpair.csv, sec09_extra_quoted_numbers.csv.
+\suppfigure{sfig:4-fewpair-realism}{5 donors: realism beyond precision and coverage}
+\figpanel{}{figures/deck_20261010/sec09_extra_fewpair.pdf}
+\figlegend{\textbf{Classifier, kernel and scSemiProfiler's own error metrics also rank LoRA-5 first in most draws, although a
+classifier still tells its cells from real cells.} Same 25 five-donor draws as Supplementary Fig.~3 (5 donors, main
+comparison), scored on exactly the same cells. Classifier AUC, cross-validated (5-fold) area under the receiver
+operating characteristic curve of a $k$-nearest-neighbour classifier ($k = 5$) trained to separate generated cells from
+held-out real cells of the same samples\cite{lopezpaz2017revisiting}; 0.5 means indistinguishable, and bars start at
+0.5. MMD$^2$, unbiased squared maximum mean discrepancy with a Gaussian kernel whose bandwidth is the median distance
+among real cells\cite{gretton2012kernel}. $E_N$, scSemiProfiler's normalised error\cite{wang2024scsemiprofiler},
+computed with its code (squared L2 on all cells, 100-component PCA per arm; log scale). Lower is better for all three.
+Grey dot-dash line: real against real (two disjoint sets of real cells of the same samples; AUC 0.49 to 0.50; $E_N$
+0.07 to 0.38 for a half split within each sample). Dashed line: the pretrained model fed each test sample's own
+single-cell pseudobulk. Dots, draws; bars, mean $\pm$ s.d. LoRA-5 has the lowest AUC of the four arms in 25 of 25
+draws (a random-forest classifier and 1-nearest-neighbour accuracy agree, 25 of 25), the lowest MMD$^2$ in 23 and the
+lowest $E_N$ in 22. Where it is not lowest, zero-shot + calibration is (for example Imperial $E_N$ 1.28 against 1.22).
+Against scSemiProfiler-5, LoRA-5 has lower $E_N$ in 25 of 25 draws (cohort means 1.28 to 2.60 against 2.04 to 4.27).
+LoRA-5 AUC per draw is 0.88 to 0.97, so a classifier still separates generated from real cells. Per-cohort values:
+Supplementary Table~2.}
+
+% ---------------------------------------------------------------------------------------------------------------
+% Supplementary Fig. 5. Source: deck slide 60; DECK/data/sec09_extra_ksens.csv.
+\suppfigure{sfig:5-ksens}{5 donors: precision and coverage at $k$ = 3, 5 and 10}
+\figpanel{}{figures/deck_20261010/sec09_extra_ksens.pdf}
+\figlegend{\textbf{The ranking of the methods does not depend on the neighbourhood size $k$.} Precision and coverage of
+Supplementary Fig.~3 recomputed with $k$ = 3, 5 and 10 nearest-neighbour balls in the same 50-component PCA ($k$ = 10
+reproduces Supplementary Fig.~3 exactly). 5 donors (main comparison): mean over the five draws per cohort; error bars,
+s.d. Grey squares: real against real (two disjoint sets of real cells of the same samples). Both metrics fall with
+smaller $k$ for every set, the real control included (real coverage at $k$ = 3 is 0.92 to 0.99), so each arm should be
+compared with the grey real-against-real line rather than with 1. The expected coverage of a perfect generator is 0.88, 0.97 and 1.00 at
+$k$ = 3, 5 and 10 (ref.~\cite{naeem2020reliable}, eq.~9). Nearest-neighbour precision and recall behave asymmetrically
+in high dimensions\cite{khayatkhoei2023emergent}, so we checked that the ranking holds across $k$. LoRA-5 is first on
+coverage at every $k$ in 25 of 25 draws and first on precision in 23, 21 and 20 of 25 draws at $k$ = 3, 5 and 10. In
+the other draws (sepsis, Imperial and Giroux, at most two draws per cohort and $k$) zero-shot + calibration is ahead on
+precision. In the donor-level 5-fold cross-validation (CV) design (reported separately; not drawn) the coverage ranking is the same at every $k$, and
+LoRA is first at every $k$ in 5 of 5 cohorts.}
+
+% ---------------------------------------------------------------------------------------------------------------
+% Supplementary Fig. 6. Source: deck slide 61; DECK/data/sec09_extra_fivefold.csv, sec09_extra_scsemi_status.csv.
+\suppfigure{sfig:6-fivefold-realism}{5-fold CV: realism metrics, all cohorts}
+\figpanel{}{figures/deck_20261010/sec09_extra_fivefold.pdf}
+\figlegend{\textbf{With every fold's training pairs, LoRA is first on every realism metric in all five cohorts (5-fold
+CV, reported separately).} Donor-level 5-fold cross-validation (CV): each sample is generated by the fold model that
+never trained on its donor, and each cohort is scored in one pooled call (ImMicro 107 samples from 107 donors, sepsis
+82 from 82, Imperial 157 from 57, Giroux 24 from 9, Ji 33 from 20; Ji bulk is whole blood). This design is reported
+separately from the 5-donor comparison. Arms: zero-shot (real bulk as measured); zero-shot + calibration ($\alpha_g$
+fitted on the fold's training pairs); sc-only fine-tune + calibration (the pretrained model fully fine-tuned on
+pseudobulks of the training donors' single cells, without their real bulk, and fed the calibrated bulk);
+scSemiProfiler with every training donor of the fold, inner-validation donors included, as representatives; and LoRA
+trained on the pairs of the other four folds. Metrics as in Supplementary Fig.~4; grey dot-dash, real against real;
+dashed, the pretrained model fed each sample's own single-cell pseudobulk. LoRA is first of the five arms in all five
+cohorts on 12 of 12 metrics (nearest-neighbour and random-forest classifier AUC, 1-nearest-neighbour accuracy, two
+MMD$^2$ kernel variants, $E_N$, and precision and coverage at $k$ = 3, 5 and 10; only the nearest-neighbour AUC, MMD$^2$ and $E_N$ are drawn). LoRA classifier AUC is 0.87 to 0.94 (real
+against real 0.49 to 0.51), so generated cells remain separable from real ones. $E_N$ of LoRA against the other arms:
+ImMicro 1.56 against 1.77 to 4.83; sepsis 1.20 against 1.40 to 5.29; Imperial 1.13 against 1.19 to 4.05; Giroux 1.70
+against 2.24 to 4.48; Ji 1.95 against 4.13 to 10.13. Per-cohort values: Supplementary Table~3.}
+
+% ---------------------------------------------------------------------------------------------------------------
+% Supplementary Fig. 7. Source: deck slide 62; DECK/data/sec09_same_pairs.csv (scores/<cohort>/{pooled,
+% scsemi_cv_20261002/pooled}); time per fold from the same file (hours_min, hours_max).
+\suppfigure{sfig:7-same-pairs}{Same training pairs (5-fold CV): LoRA and scSemiProfiler}
+\figpanel{}{figures/deck_20261010/sec09_same_pairs.pdf}
+\figlegend{\textbf{Given exactly LoRA's training pairs, scSemiProfiler still trails on coverage in every cohort and on composition in 4 of 5, and
+needs 19 to 62 times more GPU time.} Donor-level 5-fold CV, reported separately from the 5-donor comparison. For every
+fold, scSemiProfiler's representatives are exactly LoRA's training and inner-validation pairs (the other four folds;
+checked per fold), and both methods generate the fold's samples from their real bulk only. One pooled scoring call per
+method over every usable sample, with the same donors and reference cells. scSemiProfiler: upstream code (commit
+c2aea02), default settings, official output. Composition L1 over six scANVI cell types (lower is better); coverage at
+$k$ = 10; disease logFC r, the Pearson correlation between generated and observed case-versus-control log fold
+changes per cell type, averaged over T, B, monocytes and NK cells (ImMicro is healthy and has no contrast; Giroux
+excludes sex-linked genes); GPU hours per fold on one NVIDIA GB10 GPU (range over the five folds; log scale). LoRA time
+is fine-tuning plus sampling 1,500 cells per test sample; scSemiProfiler time is pretraining of every representative
+plus one model per test sample; data preparation and scoring are not counted for either. LoRA is ahead on coverage
+(0.78 to 0.93 against 0.04 to 0.35) and on precision (not drawn) in 5 of 5 cohorts, on composition in 4 of 5 and on
+logFC r in 3 of 4, reaching logFC r 0.82 in Imperial. scSemiProfiler is ahead in Giroux on composition (0.24 against
+0.26) and logFC r (0.54 against 0.46). Per fold, scSemiProfiler took 1.3 to 8.5 GPU hours and LoRA 3 to 15 min
+(Supplementary Table~4).}
+
+% ---------------------------------------------------------------------------------------------------------------
+% Supplementary Fig. 8. Source: deck slide 63; DECK/data/sec09_scsemi_representative.csv and its slide-text JSON
+% (scsemi/analysis/representative_dependence_20261002/summary.csv).
+% Quoted L1 values = summary.csv, pooled over test-sample-by-draw rows (the Wilcoxon pairs; also Supp Note 5's basis).
+% Bars and Supp Table 2 = mean of the five per-draw means (per_target.csv; checked 2026-10-11): equal in ImMicro, sepsis
+% and Imperial at 3 decimals; Giroux copy/scSemi/LoRA 0.256/0.270/0.256, Ji 0.571/0.592/0.513 (max difference 0.0030).
+\suppfigure{sfig:8-representative}{scSemiProfiler copies its representative}
+\figpanel{}{figures/deck_20261010/sec09_scsemi_representative.pdf}
+\figlegend{\textbf{scSemiProfiler-5's composition error follows the representative it is assigned, and it is never
+significantly better than copying that representative.} Same 25 five-donor draws (5 donors, main comparison).
+scSemiProfiler assigns every test sample exactly one representative, the nearest training sample in its bulk
+principal-component space, and generates the test sample from that representative's cells. Copy: the representative's
+observed composition used as the prediction for the test sample (same frozen scANVI labels as the scorer). Left,
+composition L1 over six cell types (bars, mean over draws of the per-draw mean; dots, draws; error bars, s.d.); the
+text beside the bars gives the result of a paired Wilcoxon test of scSemiProfiler-5 against the copy over every test
+sample of every draw. Right, Spearman correlation over test samples between each method's composition error and the
+copy's error. L1 values below are means over all these test samples; in Giroux and Ji, whose test sets differ in size
+between draws, they differ from the bars and from Supplementary Table~2 by up to 0.003. scSemiProfiler-5 is worse than the copy in ImMicro (0.233 against 0.187, $P = 10^{-33}$), sepsis (0.612 against 0.580,
+$P = 0.014$) and Giroux (0.267 against 0.253, $P = 0.0073$), and not significantly different in Imperial (0.602 against 0.608,
+$P = 0.91$) and Ji (0.591 against 0.570, $P = 0.46$). Its error tracks the copy's error ($\rho$ = 0.57 to 0.93) more
+closely than LoRA-5's error does ($\rho$ = 0.08 to 0.56). LoRA-5 is below the copy in four of five cohorts (ImMicro
+0.176, sepsis 0.466, Imperial 0.503, Ji 0.511) and close to it in Giroux (0.258 against 0.253). Supplementary Note~5
+discusses scSemiProfiler's own choice of representatives.}
+
+% ---------------------------------------------------------------------------------------------------------------
+% Supplementary Fig. 9. Source: deck slides 25 and 64; DECK/data/sec07_ft_modes_accuracy.csv,
+% sec07_ft_modes_compute.csv, sec07_ft_modes_wilcoxon.csv.
+\suppfigure{sfig:9-ft-modes}{Full fine-tuning, LoRA and condition-encoder-only fine-tuning}
+\figpanel{}{figures/deck_20261010/sec07_ft_modes.pdf}
+\figlegend{\textbf{LoRA matches full fine-tuning with 42 times fewer trainable weights, and tuning only the condition
+encoder is not enough.} ImMicro (107 healthy donors) and sepsis (61 sepsis and 21 control donors). Donor-level 5-fold
+CV with every training pair of the fold, each run early-stopped on inner-validation pairs (reported separately from
+the 5-donor comparison). Filled circles: pooled over the five folds (107 and 82 donors). Open diamonds: fold-0 pilot
+(22 and 17 test donors); compare like markers only. Arms: full fine-tuning (all 54.2 million weights); LoRA at rank 16
+(1.30 million trainable weights); condition-encoder-only fine-tuning (9.48 million; fold 0 only); sc-only fine-tune +
+calibration; zero-shot + calibration ($\alpha_g$ from the fold's training pairs); zero-shot. logFC r is shown for the
+pooled sepsis contrast only. Bottom: trainable parameters (log scale) and wall time per training step on one NVIDIA
+GB10 GPU; total time differs because runs stop early at different steps. Pooled, LoRA against full fine-tuning:
+coverage 0.78 against 0.70 (ImMicro) and 0.83 against 0.79 (sepsis); composition L1 0.162 against 0.157 ($P = 0.25$)
+and 0.362 against 0.372 ($P = 0.71$); sepsis logFC r 0.53 against 0.53. LoRA against zero-shot + calibration:
+coverage 0.78 against 0.53 ($P = 7\times10^{-19}$) and 0.83 against 0.43 ($P = 4\times10^{-15}$); sepsis logFC r 0.53
+against 0.12. In fold 0, condition-encoder-only fine-tuning matches the coverage of zero-shot + calibration on ImMicro
+(0.73 against 0.73) and exceeds it on sepsis (0.74 against 0.63, $P = 7\times10^{-4}$), but stays below LoRA (0.91 and
+0.86; $P = 4\times10^{-5}$ and $P = 10^{-3}$). $P$, paired Wilcoxon test on donor values, uncorrected.}
+
+% ---------------------------------------------------------------------------------------------------------------
+% Supplementary Fig. 10. Source: deck slide 65; DECK/data/sec08_pairs_realism.csv, sec08_selected_steps.csv,
+% sec08_slide_numbers.csv.
+\suppfigure{sfig:10-pairs-realism}{Pairs needed: composition, coverage and training steps}
+\figpanel{}{figures/deck_20261010/sec08_pairs_realism.pdf}
+\figlegend{\textbf{Five pairs already give most of the coverage; more pairs mainly lower the composition error.}
+ImMicro (107 donors) and sepsis (82 donors), donor-level 5-fold CV with pooled scoring. The x axis is the number of
+LoRA training pairs per fold: 5, 10, 20 or 40 (nested subsets) or every training donor of the fold (78 to 79 and 58 to
+59). These pairs-needed curves sit inside the 5-fold design and are not the 5-donor draws. Every run also early-stops
+on 7 inner-validation pairs (7 to 8 in sepsis), and the test fold is never trained on. At x = 0 no paired data are
+used: the pretrained model is fed the real bulk as measured (pale blue) or after exon-length correction (sand; counts
+divided by Ensembl union-exon length in kb before CPM normalisation). Squares: zero-shot + calibration with $\alpha_g$
+fitted on exactly the pairs that the LoRA point saw. Diamonds: scSemiProfiler with exactly LoRA's training and
+inner-validation pairs as representatives. Dashed: the pretrained model fed each sample's own single-cell pseudobulk.
+Dot-dash: cohort-mean null (every donor predicted with the mean observed composition of the other donors; L1 0.16 in ImMicro and 0.52 in sepsis). Right: training step chosen by
+early stopping (dots, folds; bar, median; cap 20,000). Coverage, ImMicro and sepsis: as measured 0.19 and 0.07;
+exon-length-corrected 0.50 and 0.40; calibration 0.52 to 0.53 and 0.43 to 0.44 at every n; LoRA 0.84 and 0.77 with 5
+pairs and 0.78 and 0.83 with every training donor; scSemiProfiler 0.04 and 0.10 at every n. Composition L1:
+exon-length correction does not fix it (ImMicro 0.45 to 0.42, sepsis 0.51 to 0.56). LoRA reaches 0.17 and 0.43 with 5 pairs and
+0.16 and 0.36 with every training donor; calibration 0.19 to 0.20 and 0.40 to 0.41 (lower than LoRA with 5 pairs in
+sepsis); scSemiProfiler 0.22 to 0.23 and 0.50 to 0.54. With 5 pairs, early stopping
+selects a median of 3,250 (ImMicro) and 2,750 (sepsis) steps, close to the fixed 3,000 steps of LoRA-5.}
+
+% ---------------------------------------------------------------------------------------------------------------
+% Supplementary Fig. 11. Source: deck slide 45; DECK/data/sec08_pairs_signal.csv, sec08_slide_numbers.csv.
+\suppfigure{sfig:11-pairs-signal}{Pairs needed: donor and disease signal}
+\figpanel{}{figures/deck_20261010/sec08_pairs_signal.pdf}
+\figlegend{\textbf{With more pairs, LoRA captures donor-level composition and the sepsis disease signal; without pairs,
+or with calibration alone, the disease logFC r stays near zero.} Same design and arms as Supplementary Fig.~10
+(donor-level 5-fold CV, pooled; x, LoRA training pairs per fold, not the 5-donor draws). Left: ImMicro donor NK
+fraction r, the Pearson correlation over 107 donors between generated and observed NK-cell fractions. Middle: sepsis
+disease logFC r (61 sepsis against 21 control donors), the Pearson correlation between generated and observed
+case-versus-control log fold changes per cell type, averaged over T, B, monocytes and NK cells. Right: sepsis top-50
+differentially expressed gene (DEG) overlap. Per cell type, genes with a false discovery rate below 0.05 (Welch test on donor pseudobulks,
+Benjamini--Hochberg correction\cite{benjamini1995fdr}) are ranked by absolute logFC, and the genes shared by the
+observed and generated top-50 lists are counted and summed over T, B, monocytes, NK and dendritic cells (at most 250).
+Without pairs or with calibration alone, logFC r is at most 0.12 (as measured $-0.01$; exon-length-corrected 0.11;
+calibration 0.08 to 0.12). LoRA logFC r rises from 0.29 with 5 pairs to 0.53 with every training donor (58 to 59
+pairs), and top-50 DEG from 5 to 23. With 5 or 10 pairs, though, LoRA shares fewer top-50 DEG with the observed
+cells (5 and 8) than the arms without pairs (9 as measured, 11 exon-length-corrected) or calibration from the same
+pairs (12 and 10). The donor NK fraction r rises from 0.20 with 5 pairs to 0.70 with 40 or more
+pairs (as measured 0.49; calibration 0.50 to 0.53). scSemiProfiler given the same pairs reaches a higher sepsis logFC r
+at 5 to 40 pairs (0.52 to 0.58, against 0.29 to 0.48 for LoRA) but not with every training donor (0.44 against 0.53).
+Its top-50 DEG stays at 2 to 6, and its ImMicro NK fraction r is 0.27 to 0.49. Dashed: the pretrained model fed each
+sample's own single-cell pseudobulk (logFC r 0.47, top-50 DEG 22, NK fraction r 0.60).}
+
+% ---------------------------------------------------------------------------------------------------------------
+% Supplementary Fig. 12. Source: deck slides 66 and 67; DECK/data/sec07_rank_sweep.csv, sec07_rank_sweep_runs.csv,
+% sec07_k_aug.csv, sec07_k_aug_wilcoxon.csv, sec07_k_aug_inner_val_delta.csv.
+\suppfigure{sfig:12-rank-k}{LoRA rank and $K$ diversification}
+\figpanel{a}{figures/deck_20261010/sec07_rank_sweep.pdf}
+\figpanelh{b}{figures/deck_20261010/sec07_k_aug.pdf}{0.27}
+\figlegend{\textbf{Rank 16 and a single $K$ of 1,000 are good defaults.} Donor-level 5-fold CV with pooled scoring
+(ImMicro 107 donors, sepsis 82; reported separately from the 5-donor comparison).
+\textbf{a}, LoRA rank $r$ = 4, 16 (used throughout) and 64, with the LoRA hyperparameter $\alpha$ equal to the rank, so that the low-rank update is scaled by $\alpha/r = 1$ (324,056,
+1,296,224 and 5,184,896 trainable parameters), at 5, 20 or every training pair per fold (nested subsets). Each run is
+early-stopped on 7 to 8 inner-validation pairs (cap 20,000 steps). With every training donor, rank 64 raises ImMicro
+coverage from 0.78 to 0.82 but lowers precision from 0.92 to 0.90, and rank 4 raises composition L1 from 0.162 to 0.186
+(paired Wilcoxon $P = 9.8\times10^{-9}$). In sepsis, rank 64 gives logFC r 0.55 and top-50 DEG 26 against 0.53 and 23
+for rank 16; rank 4 gives 0.48 and 18. With 5 pairs the ranks are mixed.
+\textbf{b}, $K$ diversification with every training pair (rank 16): plain LoRA on real-bulk conditions at $K$ = 1,000;
+multi-$K$ (conditions at $K$ = 250 to 4,000); and LoRA with recipes (half of each batch is pseudobulk built from the
+training donors' own cells, $K$ from the pretraining grid). Neither variant helps: composition L1 0.162 against 0.168
+(multi-$K$, $P = 0.04$) and 0.161 (recipes, $P = 0.75$) in ImMicro, and 0.362 against 0.366 ($P = 0.70$) and 0.365
+($P = 0.73$) in sepsis; sepsis logFC r 0.53 against 0.51 and 0.50. The inner-validation loss changes by at most
+$1.2\times10^{-4}$. $P$, paired Wilcoxon test on donor composition L1 against rank 16 (\textbf{a}) or plain LoRA
+(\textbf{b}), uncorrected.}
+
+% ---------------------------------------------------------------------------------------------------------------
+% Supplementary Fig. 13. Source: deck slide 44; DECK/data/fig4_disease.csv (design = variants).
+\suppfigure{sfig:13-sepsis-variants}{Sepsis: fine-tuning variants}
+\figpanel{}{figures/deck_20261010/fig4_disease_variants.pdf}
+\figlegend{\textbf{On the sepsis disease signal, LoRA at rank 16 matches full fine-tuning; rank 64 is slightly higher,
+and multi-$K$ training or recipes do not help.} Sepsis (61 sepsis and 21 control donors), donor-level 5-fold CV with
+every training pair of the fold, pooled scoring with the same scorer and donors for every variant (reported separately
+from the 5-donor comparison). LoRA at rank 16 (dark bar; 1.30 million trainable weights) is the model used throughout;
+full fine-tuning trains all 54.2 million weights. Multi-$K$: LoRA trained on conditions at $K$ = 250 to 4,000. Recipes:
+half of each batch is pseudobulk built from the training donors' own cells. logFC r, the Pearson correlation between
+generated and observed case-versus-control log fold changes per cell type, averaged over T, B, monocytes and NK cells.
+Partial r beyond bulk, the same correlation after both log fold changes are residualised on the input bulk's log fold
+change: the cell-type signal that the bulk itself does not carry. Top-50 DEG overlap, as in Supplementary Fig.~11.
+Values (logFC r, partial r, top-50 DEG): rank 4, 0.48, 0.42, 18; rank 16, 0.53, 0.47, 23; rank 64, 0.55, 0.49, 26; full
+fine-tuning, 0.53, 0.47, 24; multi-$K$, 0.51, 0.45, 23; recipes, 0.50, 0.43, 23.}
+
+% ---------------------------------------------------------------------------------------------------------------
+% Supplementary Fig. 14. Source: deck slide 46; DECK/data/secS_deconvolution.csv and its slide-text JSON
+% (plans/bayesprism_paired_20261006/results/RESULTS_full.md, RESULTS_cibersortx.md).
+\suppfigure{sfig:14-deconv-composition}{Cell-type deconvolution with the same pairs: composition}
+\figpanel{}{figures/deck_20261010/secS_deconv_composition.pdf}
+\figlegend{\textbf{Given the same paired donors, \method{} with LoRA estimates the cell-type mix more closely than
+cell-type deconvolution with BayesPrism or CIBERSORTx in almost every draw.} BayesPrism\cite{chu2022bayesprism} and
+CIBERSORTx\cite{newman2019cibersortx} are cell-type deconvolution methods: they return cell-type fractions and
+per-type expression, not single cells. Each was given the same paired donors as \method{} with LoRA and scored on the
+benchmark's own splits, observed truth and scorer. Left, 5 donors (main comparison; mean of five draws; error bars,
+s.d.): the reference is the draw's five training donors' observed cells, and the mixture is each test sample's real
+bulk. Right, 5-fold CV (reported separately; pooled): the reference is each fold's LoRA training and inner-validation
+pairs, and scSemiProfiler with the same pairs is added. BayesPrism used its recommended gene set (protein-coding genes
+without chrX, chrY, mitochondrial and ribosomal genes; 15,000 to 19,000 genes), cell types from the six-type gate with
+cell states from the level-2 labels of the Allen Institute for Immunology (AIFI) immune health atlas\cite{gong2025multiomic}, and the better of its initial and final fractions, in
+cell units. CIBERSORTx used its official Docker images in single-cell reference mode with S-mode batch correction (up
+to 400 reference cells per type); fractions were converted to cell units with each type's mean unique molecular identifier (UMI) count per reference cell.
+Composition L1 over the six types (lower is better). \method{} with LoRA is closer than BayesPrism in 22 of 25 draws
+and closer than CIBERSORTx in 24 of 25. All three BayesPrism wins are sepsis draws, where BayesPrism is also closer
+on average (0.45 against 0.47); the one CIBERSORTx win is a Ji draw. With 5-fold CV it has the lowest L1 in 4 of 5 cohorts (in Giroux
+scSemiProfiler is lower, 0.24 against 0.26). On pseudobulk mixtures BayesPrism recovers the observed mRNA shares (for
+example T cells 0.39 against 0.38), so its error comes from real bulk, whose difference from pseudobulk \method{} learns
+from the pairs.}
+
+% ---------------------------------------------------------------------------------------------------------------
+% Supplementary Fig. 15. Source: deck slides 68 and 69; DECK/data/sec08_cross_cohort.csv, sec08_cross_cohort_steps.csv
+% (plans/paired_crosscohort_20261002/).
+\suppfigure{sfig:15-cross-cohort}{Pairs from other cohorts do not transfer}
+\figpanel{a}{figures/deck_20261010/sec08_cross_cohort.pdf}
+\figpanel{b}{figures/deck_20261010/sec08_cross_cohort_steps.pdf}
+\figlegend{\textbf{A model trained on other cohorts' pairs does not replace the target cohort's own pairs, even when
+trained to saturation.} Leave-one-cohort-out design with conditions at $K$ = 1,000. Each held-out cohort is scored on
+all its usable samples against the same frozen reference as its within-cohort benchmark. Arms without pairs from the
+held-out cohort: zero-shot (real bulk as measured); calibration from other cohorts ($\alpha_g$ fitted on the training
+cohorts' pairs, cohort-balanced; no training); and LoRA at rank 16, LoRA at rank 64 and full fine-tuning, trained on the
+other peripheral blood mononuclear cell (PBMC) cohorts' pairs plus 44 pairs of Liu 2021\cite{liu2021timeresolved} (training only, because its cells are in
+the pretraining corpus). These three arms used cohort-balanced batches and inner validation on 10\% of each training
+cohort's donors (none from Giroux, where 10\% of each group rounds to zero), and trained in 20,000-step blocks until a block improved the best
+inner-validation loss by less than 1\% of the total improvement over the pretrained model (patience early stopping applied
+within blocks). Ji's bulk is whole
+blood and is never trained on. Reference: LoRA at rank 16 on the cohort's own pairs (5-fold CV, pooled). Giroux logFC r
+excludes the sex-linked genes (65 of its scoring genes).
+\textbf{a}, Values at the last milestone; numbers in the bars are training steps (k, thousands).
+\textbf{b}, The same runs at every 20,000-step milestone (best inner-validation state so far); dashed, calibration from
+other cohorts; dotted, LoRA on the cohort's own pairs.
+The own-cohort LoRA is best on coverage in 5 of 5 cohorts, on disease logFC r in 4 of 4 and on composition in 4 of 5
+(in Giroux, calibration from other cohorts is lower, 0.20 against 0.26). No milestone of any
+other-cohort arm reaches its coverage. On composition, calibration from other cohorts ties or beats the trained
+other-cohort arms in 4 of 5 cohorts (full fine-tuning is best in Imperial). Training to saturation took 20 to 31 min
+for rank 16 and 59 to 107 min for full fine-tuning on one GPU, and full fine-tuning drifted in sepsis and Ji (Ji
+composition L1 0.84 to 1.03). See Supplementary Note~2.}
+
+% ---------------------------------------------------------------------------------------------------------------
+% Supplementary Fig. 16. Source: deck slides 70-72 and 76; DECK/data/sec07_no_pretraining.csv, sec08_cross_scratch.csv,
+% secS_scratch_rare_states.csv, secS_scratch_rare_states_bootstrap.csv, sec10_other_tissues.csv.
+\suppfigure{sfig:16-no-pretraining}{Training without pretraining: one-seed sepsis pilots}
+\figpanelh{a}{figures/deck_20261010/sec07_no_pretraining.pdf}{0.2}
+\figpanelh{b}{figures/deck_20261010/sec08_cross_scratch.pdf}{0.2}
+\figpanelh{c}{figures/deck_20261010/secS_scratch_rare_states.pdf}{0.22}
+\figlegend{\textbf{Trained from random initialisation on enough pairs, the same model gives realistic cells; pretraining
+adds donor composition from five pairs, rare cycling cells and minutes instead of an hour of training.} Pilot with one
+seed; folds 1 to 4 and a second cohort have not been run. From scratch: the same denoiser (54.2 million weights) at
+random initialisation, trained in full on the pairs (cap 100,000 steps in the same-cohort runs; the other-cohort run in
+\textbf{b} and \textbf{c} early-stopped at 162,000 steps). It still takes the architecture, the
+6,000-gene panel, the value scaler and the bulk z-score constants from the pretrained checkpoint. LoRA has rank 16.
+\textbf{a}, Same cohort, same pairs: sepsis fold 0 (17 test donors; 12 sepsis, 5 controls), 5 or 58 training pairs plus
+the same 7 inner-validation pairs, and 58 pairs with recipes (half of each batch is pseudobulk built from the
+training donors' own cells). Donor comp.\ r, Pearson correlation over the 17 donors between generated and observed
+fractions of T, B, monocytes and NK cells, averaged over the four. Train time on one NVIDIA GB10 GPU.
+Dashed, zero-shot. Coverage is similar (0.80 to 0.88). With 5 pairs the scratch model cannot place donors (donor
+comp.\ r 0.17 against 0.70); with 58 pairs it reaches a higher logFC r (0.58 against 0.49). Scratch training takes 61
+to 86 min, LoRA 3 to 5 min.
+\textbf{b}, Other cohorts' pairs only: all 82 sepsis donors held out, training on ImMicro, Imperial, Giroux and Liu 2021
+pairs; the 46 panel genes that no training cohort measured are set to zero in every arm. Scratch ends where the
+pretrained arms end (coverage 0.33 against 0.34 for LoRA rank 16), far below the own-cohort LoRA (0.83).
+\textbf{c}, Rare states, generated over observed rate (log scale; grey band, within 2-fold of the observed rate); 8 of 13
+states shown. With the same pairs there is no significant difference on average over the 13 states; the pretrained
+model is closer for cycling T/NK cells in 8 of 8 comparisons; C1Q-high monocytes reach about half the observed rate or
+more only with the cohort's own pairs. With other cohorts' pairs, scratch is closer, though not significantly, for
+interferon-stimulated-gene-high (ISG-high) T cells (1.10 against 1.54 to 1.98 times the observed rate). Fold 0 is the
+pretrained LoRA's best fold for C1Q-high monocytes and cycling T/NK cells (58 pairs: 1.03 and 1.00 times the observed
+rate, against 0.27 to 0.82 and 0.32 to 0.64 in folds 1 to 4). See Supplementary Note~3.}
+
+% ---------------------------------------------------------------------------------------------------------------
+% Supplementary Fig. 17. Source: deck slide 74; DECK/data/secA_scsemi_readout.csv and its slide-text JSON
+% (scsemi/analysis/output_variants_20261003/).
+\suppfigure{sfig:17-scsemi-readout}{scSemiProfiler without its output threshold (modified, not the official output; sensitivity only)}
+\figpanel{}{figures/deck_20261010/secA_scsemi_readout.pdf}
+\figlegend{\textbf{Removing scSemiProfiler's output threshold raises its coverage but changes none of the 23 comparisons
+with LoRA-5.} Modified read-out, not scSemiProfiler's official output; every other figure uses the official output.
+The official output (upstream commit c2aea02) is the decoder's zero-inflated negative binomial mean with every value of
+10 normalised units or less set to zero; this cut is fixed in the code, not an option. Here the same trained per-sample
+models were read out again without the cut (no retraining; each test sample's own seed) and scored exactly as the
+official arm: the same 25 five-donor draws (5 donors, main comparison), donors and reference cells, with Giroux disease
+metrics without sex-linked genes. A control read-out with the cut reproduces the official output (pseudobulk r of at
+least 0.996 over 1,790 test samples; coverage within 0.015). Bars, mean of five draws; dots, draws. Coverage, precision
+and composition L1 as in Supplementary Fig.~3; disease logFC r as in Supplementary Fig.~11 (none for healthy ImMicro).
+Without the cut, scSemiProfiler's coverage rises in 5 of 5 cohorts (for example Imperial 0.06 to 0.42 and Ji 0.15 to
+0.53). On the mean of five draws, LoRA-5 stays ahead on coverage, precision and composition in all five cohorts and on
+top-50 DEG (not drawn) in sepsis and Imperial. In Giroux and Ji every method shares almost no top-50 DEG
+with the observed cells (0.2 against 0), a tie by the rule in Methods (mean difference below one gene). scSemiProfiler stays ahead on logFC r in sepsis and Giroux, as with the
+official output. None of
+the 23 comparisons (three cell metrics in five cohorts; logFC r and top-50 DEG in four) changes sides. See
+Supplementary Note~4.}

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

@@ -0,0 +1,36 @@
+% Supplementary Methods for Methods part 3 (methods_eval.tex): scoring details, rare-state rules, Behcet depth control,
+% LoRA-3 (todo) and bulk-only generation. Written 2026-10-10; every value checked against:
+%   M8/plans/paired_finetune_intradataset_20260927/tools/{score_ft.py, ft_common.py, extra_metrics.py,
+%     collect_extra_metrics.py, contrast_extra.py, contrast_extra_limma.R}; scores_contrast/*/limma/versions.txt
+%   M8/plans/rare_states_screen_20261004/tools/{states.json, screen.py}; M8/plans/rare_states_paired_20261008/tools/rare_paired.py
+%   M8/plans/behcet_c1q_repro_20260921/conditions/MANIFEST.json; M8/plans/utility_studies_final6k_20260925/
+%     (Behcet input comparison, sc-pseudobulk seed runs, probe and common-depth tools)
+%   M8/plans/behcet_c1q_lora3_20261004/README.md (todo only)
+%   M8/plans/bulkonly_lora5_20261006/README.md, M8/plans/bulkonly_lora5_k4000_20261006/README.md
+% No table floats: a float would take a Supplementary Table number and shift the fixed numbering.
+% Fragment only: to be \input by supplementary.tex under \section*{Supplementary Methods}, after supp_methods_model.
+
+\subsection*{Precision and coverage: subsets, other $k$ and depth matching}
+The observed cells of each scored sample were put in a fixed order and permuted once, with a generator seeded by 20260925 and the first eight bytes of the SHA-256 hash of the sample identifier. R1 holds the first $n$ cells and R2 the next $n$, with $n=\min(150,\lfloor N/2\rfloor)$ for a sample with $N$ cells. The generated cells for precision and coverage are $n$ of the sample's generated cells, drawn without replacement by a generator seeded in the same way from 20260926. The 50-component principal component analysis (PCA), fitted on the R1 cells of the scored samples, used the randomised solver (seed 20260925). An R1 cell's radius is the distance to its $k$-th nearest other R1 cell. With $k=3$ and $k=5$ the radii were recomputed on the same cells. For identical distributions and equal set sizes, the expected coverage is 0.88, 0.97 and 1.00 at $k=3$, 5 and 10 (equation 9 in Naeem et al.\cite{naeem2020reliable}), and precision and recall behave asymmetrically in high dimensions\cite{khayatkhoei2023emergent}; arms were therefore compared with the real-versus-real row rather than with 1. For depth matching, the generated and R1 cells of a sample were ranked by their number of detected genes, and cells of equal rank were paired. In each pair, the cell with more detected genes was thinned by sampling its UMIs without replacement down to the other cell's number of detected genes (seed 20260929). The PCA fitted on the unthinned R1 cells was kept, and the radii were recomputed on the thinned R1 cells.
+
+\subsection*{Classifier tests, MMD$^2$ and $E_N$}
+The classifier scores and the squared maximum mean discrepancy (MMD$^2$) use the scorer's 50-component embedding and the generated cells of precision and coverage. Their real side is R2, which the PCA did not see, with as many cells per sample as the generated side. The $k$-nearest-neighbour classifier ($k=5$) and a random forest of 300 fully grown trees, each split drawing from the square root of the number of features, were scored by stratified, shuffled 5-fold cross-validation (seed 20261001). Their score is the mean area under the receiver operating characteristic curve (ROC AUC) over folds. The leave-one-out 1-nearest-neighbour accuracy pools both sets and predicts each cell's origin from its nearest other cell\cite{lopezpaz2017revisiting}. For all three, 0.5 means indistinguishable. The real-versus-real control compares R2 with a third disjoint set R3, the next cells of the sample's permutation, of the same size. MMD$^2$ is the unbiased estimator\cite{gretton2012kernel} with the kernel $\exp(-\lVert x-y\rVert^2/(2\sigma^2))$, where $\sigma$ is the median pairwise Euclidean distance among up to 4,000 R1 cells (seed 20261001); $\sigma/2$, $2\sigma$ and the sum of the three values were also reported. These scores were also computed after depth matching.
+
+$E_N$, scSemiProfiler's normalised error, follows equations 17 and 18 of the scSemiProfiler paper and the distances of its code\cite{wang2024scsemiprofiler}. For each arm, all observed cells of the scored samples and all their generated cells were transformed to $\log(1+\mathrm{CP10K})$, with CP10K the counts per 10,000, on the scoring genes and embedded together in a 100-component PCA. For sample $i$, $D_i$ is the mean, over all its observed and generated cells, of the squared Euclidean distance to the nearest cell of the other set. The lower bound $L$ is the same quantity between two random halves of the whole embedded matrix, estimated from up to 20,000 query cells per direction when the matrix is large. The upper bound $U$ is the mean of $D$ between the observed cells of each sample and those of a randomly chosen other sample. Then $E_N=(\bar D-L)/(U-L)$, where $\bar D$ is the mean of $D_i$ over samples. The real-versus-real control splits each sample's observed cells into two random halves. scSemiProfiler sets a representative's error to $L$; no scored sample is a representative, so no such value enters.
+
+\subsection*{Correspondence at the top and rank-biased overlap}
+For each arm and cell type, observed and generated cells were analysed separately on their per-sample pseudobulks, the same summed counts that enter the Welch test. Genes with a non-zero total on that side were kept. Trimmed mean of M-values (TMM) normalisation in edgeR\cite{robinson2010edger} was followed by voom, lmFit and eBayes in limma\cite{law2014voom,ritchie2015limma}, with case versus control as the only design factor. Genes were ranked by absolute moderated $t$, ties kept in the fixed gene order. CAT($k$) was computed at $k=10$, 20, 50, 100, 200 and 500; a generated side whose test was undefined scores 0. For each draw or call, CAT is the mean over T cells, B cells, monocytes and NK cells. Five-donor curves average the five draws, and 5-fold curves come from the pooled call. The extrapolated rank-biased overlap (equation 32 in Webber et al.\cite{webber2010similarity}) with $p=0.98$ compared the rankings by moderated $t$, from the most up-regulated gene downwards and from the most down-regulated gene upwards. We used R 4.5.2, limma 3.66.0 and edgeR 4.8.1.
+
+\subsection*{Rare-state rules}
+A marker state is a cell of the parent type in which at least the stated number of the listed genes have a count above zero. Parents are first-level labels unless level-2 labels are named, and genes that a cohort's single-cell data did not measure were dropped from its rules. C1Q-high monocytes have 2 of \textit{C1QA}, \textit{C1QB} and \textit{C1QC}\cite{zheng2022c1qbehcet}. MS1 monocytes have \textit{IL1R2} and 2 of \textit{RETN}, \textit{ALOX5AP}, \textit{CYP1B1}, \textit{MCEMP1}, \textit{PLAC8}, \textit{CLU} and \textit{CD163}\cite{reyes2020immune}. Cycling monocytes have 1 of \textit{MKI67}, \textit{TOP2A}, \textit{TYMS}, \textit{CENPF}, \textit{UBE2C} and \textit{BIRC5}\cite{ye2026sepsis}. Cycling T and NK cells have 2 of the same six genes, and cycling activated T cells have 1 of them together with 1 of \textit{CD38}, \textit{HLA-DRA} and \textit{HLA-DRB1}\cite{mathew2020deep}. Proliferating T cells are T cells with the level-2 label proliferating T cell. ISG-high T cells have 4 of \textit{ISG15}, \textit{IFI6}, \textit{IFI27}, \textit{IFI44L}, \textit{MX1}, \textit{OAS1}, \textit{IFIT1}, \textit{IFIT3}, \textit{RSAD2} and \textit{SIGLEC1}\cite{nehar2020mapping}; ISG-high monocytes, the positive control, have 4 of the same ten genes. DN2 B cells are B cells with \textit{ITGAX} or \textit{TBX21}, with 2 of \textit{ITGAX}, \textit{TBX21}, \textit{FCRL5}, \textit{FGR}, \textit{TFEC}, \textit{SOX5} and \textit{ZEB2}, and without \textit{CR2}\cite{jenks2018distinct}. Plasmablasts are cells of any type with 3 of \textit{JCHAIN}, \textit{MZB1}, \textit{TNFRSF17}, \textit{TXNDC5}, \textit{PRDM1}, \textit{IRF4} and \textit{IGLL5}, or, by label, cells with the level-2 label plasma cell\cite{mathew2020deep}. Haematopoietic stem and progenitor cells are cells of any type with 2 of \textit{CD34}, \textit{PRSS57}, \textit{SPINK2} and \textit{SMIM24}\cite{stephenson2021single}. Cytotoxic CD4 T cells are naive CD4 T, memory CD4 T or regulatory T cells (level-2 labels) with 3 of \textit{GZMH}, \textit{GZMB}, \textit{GZMA}, \textit{PRF1}, \textit{NKG7}, \textit{GNLY}, \textit{FGFBP2}, \textit{CX3CR1} and \textit{ZNF683}, and without \textit{CCR7}\cite{hashimoto2019single}. Granulocyte-like cells are cells of any type with 2 of \textit{ELANE}, \textit{MPO}, \textit{CTSG}, \textit{CAMP}, \textit{LCN2}, \textit{MMP9}, \textit{CSF3R}, \textit{CXCR2} and \textit{FCGR3B}\cite{bennett2003interferon}. The negative control, platelet signal in monocytes, needs 2 of \textit{PPBP}, \textit{PF4}, \textit{TUBB1}, \textit{GP9} and \textit{NRGN}.
+
+For marker states, the thinning factor of the observed parent cells was the smaller of 1 and the ratio of the generated to the observed median library of the parent cells, both summed over the measured panel genes. Each observed count of the state's marker genes was replaced by a binomial draw with this probability, from a generator seeded by 20261004, the cohort and the draw. The detection AUROC (area under the receiver operating characteristic curve) is the Mann--Whitney $U$ statistic of the generated fractions of samples above against below the median observed fraction, divided by the product of the two group sizes. Samples at the median were left out, and each side needed at least two samples.
+
+\subsection*{Beh\c{c}et disease: seeds and depth control}
+Real-bulk conditions read each bulk as the sum of 5,000 cells ($K=5{,}000$, equal to the number of cells generated); 255 genes of the 20,097-gene axis were not measured by this bulk assay. The seeds were 20260913 plus $n$ times 1,000,000,007 from the real bulk and 20260926 plus $n$ times 1,000,000,007 from the single-cell pseudobulks, for $n$ from 0 to 4. For the depth control, every count of sample $s$ was replaced by a binomial draw with probability $f_s=\min(1,T/m_s)$, where $m_s$ is the sample's median monocyte library on the 5,671 shared genes and $T=498.5$ UMIs is the smallest $m_s$ among the generated sets (seed 20260925). For generated cells, the C1Q-high share at this depth was computed without a read-out draw, as the mean over the sample's monocytes of $1-\exp(-f_s\lambda)$, where $\lambda$ is the cell's summed generated rate of \textit{C1QA}, \textit{C1QB} and \textit{C1QC}; this is the expected share after the Poisson read-out and the thinning. The CD16 share at common depth was recomputed after relabelling the thinned counts with the frozen annotator, for seed 0. Observed cells were thinned in the same way and kept their labels. The exact permutation test used each bulk sample's mean over the five seeds of its common-depth C1Q-high share and all 92,378 ways to assign the 19 bulk samples to 9 patients and 10 controls.
+
+\subsection*{Beh\c{c}et disease: LoRA on three pairs}
+\todo{Methods of the LoRA-3 arm (Fig.~5d, to be added with the panel; source plans/behcet\_c1q\_lora3\_20261004). Three patient bulk samples (BDV4, BDV5, BDV6) were matched afterwards to single-cell samples by expression (plans/behcet\_pair\_recovery\_20261004); there is no control pair. LoRA of rank 16 with the LoRA-5 protocol (learning rate $5\times10^{-4}$, 3,000 fixed steps, seed 20260927) on the real-bulk conditions at $K=1{,}000$ and, as a follow-up, $K=4{,}000$. Test: the 16 bulks without single cells (6 patients, 10 controls), 5,000 cells per sample, seeds 20260913 and 20261004. Comparators: zero-shot from the real bulk and zero-shot + calibration fitted on the same three pairs. Scoring as above on the 5,658 genes also measured in all eight single-cell libraries; observed cells thinned to each arm's monocyte depth.}
+
+\subsection*{Cells for bulk-only samples}
+The bulk-only samples are every bulk sample of sepsis and Imperial without a single-cell partner; 45 of the Imperial donors have no single cells at all. Conditions used the cohort's harmonised bulk counts, the fixed-depth arithmetic of the paired benchmark and the cohort's single-cell gene mask as output mask; rebuilt in the same way, the paired samples' conditions matched the benchmark byte for byte. Four samples whose bulk sex call disagreed with the metadata (three in sepsis, one in Imperial) were generated and flagged. At $K=1{,}000$ we used the merged LoRA-5 model of each five-donor draw unchanged and generated 1,500 cells per sample with the benchmark sampler and seed 20260927 in every draw. A given cell slot of a sample therefore starts from the same noise in all five draws. The Poisson read-out used seed 20260916. At $K=4{,}000$, LoRA (rank 16, learning rate $5\times10^{-4}$, 3,000 fixed steps, seed 20260927)\cite{hu2022lora} was retrained on each draw's five training donors with their real-bulk conditions at $K=4{,}000$ and generated 4,000 cells per sample. Draw $j$ used base seed $20260927+1{,}000{,}003\,j$, so no two draws share noise, and one Poisson read-out over the whole set with seed $20260916+j$. Large sets were generated in row parts seeded so that every cell keeps the seed it would have in one file. Per draw this gave 183,000 sepsis and 246,000 Imperial cells at $K=1{,}000$, and 488,000 and 656,000 at $K=4{,}000$. All cells were labelled by the frozen annotator.