Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/sections/results_fig1.tex

@@ -0,0 +1,13 @@
+\subsection*{\method{} generates single blood cells in gene space from one bulk profile}
+
+\textbf{\method{} works directly on genes.} Like Sparse Data Diffusion (SDD)\cite{ostheimer2025sparse} and unlike scDiffusion\cite{luo2024scdiffusion}, it diffuses each gene's expression value together with its detection (Fig.~\ref{fig:fig1}a). We added a bulk condition and a loss over measured genes only. At the last step, detection is drawn from each gene's predicted probability (the drawn, or Bernoulli, read-out), not cut at 0.5.
+
+\textbf{Each cell is learned with its own bulk.} We pair each training cell with a pseudobulk of $K$ cells from its own sample, itself included, so pretraining needs no measured bulk (Fig.~\ref{fig:fig1}b and Supplementary Fig.~2). An optional low-rank adaptation (LoRA)\cite{hu2022lora} on a few paired samples adapts the model to a cohort (Fig.~\ref{fig:fig2}).
+
+\textbf{The condition carries the bulk and nothing about cell types.} All cells generated from one bulk share one condition of 18,006 numbers (Fig.~\ref{fig:fig1}c). It holds the bulk expression of the 6,000 panel genes, six scalars for depth and profile shape and two masks of measured genes, but no cell-type labels or proportions. The condition is mapped to a 512-dimensional context that enters every denoiser block. Nothing forces the generated cells to sum to the bulk.
+
+\textbf{The pretraining corpus is large and mostly healthy.} It holds 41.4 million blood cells from 22 scRNA-seq datasets and 3,199 donors (Fig.~\ref{fig:fig1}d, Supplementary Fig.~1 and Supplementary Note~1). Of these cells, 88.5\% came from healthy donors and 23.6\% had been cultured in vitro.
+
+\textbf{A wider model fits better but carries no more disease signal.} We chose the design on a systemic lupus erythematosus (SLE) cohort of children and adults outside the corpus (GSE135779; 42 donors, 30 with SLE)\cite{nehar2020mapping}, generating cells from each donor's pseudobulk. We scored logFC r averaged over six cell types, the top-50 differentially expressed gene (DEG) overlap summed over them and composition L1, the error in cell-type fractions (Methods). On an earlier 6,000-gene panel, a four times wider denoiser cut the validation loss by 31\% (0.197 to 0.136). Yet with the drawn read-out its logFC r was 0.268 against 0.286 for the base model, and its top-50 DEG 62 against 68 (Fig.~\ref{fig:fig1}e). With the 0.5 cut its logFC r was 0.020 higher (0.265 against 0.245), about one run-to-run standard deviation (0.018). A four times wider condition context matched the base model on loss and disease signal and lowered composition L1 (0.161 against 0.197, drawn read-out), with nearly twice the parameters. Adding cell-type proportions to the condition did not help: true and predicted proportions scored the same, and composition L1 rose (0.245 against 0.199, 0.5 cut). We kept the base width.
+
+\textbf{Six thousand genes keep a rare-cell marker without losing signal.} We then lowered the gene panels' detection floor from 1\% to 0.1\% of cells to admit markers of rare cell types. We scored panels of 2,000, 3,000, 6,000 and 15,114 genes on the 1,006 SLE benchmark genes in every panel. The 3,000- and 6,000-gene models tied on disease signal: the 3,000-gene model's logFC r (0.358) and top-50 DEG (72) lay within the 6,000-gene model's range over five sampling seeds (0.358--0.399 and 72--81; Fig.~\ref{fig:fig1}f). The 2,000- and 3,000-gene panels had lower composition L1 (0.131 and 0.164, against 0.181 for 6,000 genes). The 15,114-gene panel had the lowest logFC r (0.344) and the highest composition L1 (0.297). We kept 6,000 genes, the smallest panel that contains \textit{CLEC9A}, a marker of type 1 conventional dendritic cells. The final model has 54.2 million parameters and was trained on the 33.1 million cells of the training split.

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/sections/results_fig2.tex

@@ -0,0 +1,31 @@
+% Results, Fig. 2 (healthy ImMicro only; no disease metrics). Written 2026-10-10 from deck slides 17-27 and DECK/data:
+% a sec05_shift_cohorts.csv, sec05_condition_space_numbers.csv; b sec05_cells_metrics.csv; c sec06_ft_method.csv,
+% sec06_eval_design.csv; d sec04_metrics_overview.csv; e fig23_healthy.csv, sec09_fewpair_cells.csv, sec09_quoted_numbers.csv;
+% f sec09_extra_fewpair.csv, sec09_extra_fivefold.csv, sec09_extra_quoted_numbers.csv; g sec07_ft_modes_{accuracy,compute,wilcoxon}.csv;
+% h sec08_pairs_realism.csv, sec08_pairs_signal.csv, sec08_selected_steps*.csv; i sec07_rank_sweep.csv, sec07_k_aug.csv.
+% Tightened 2026-10-10 (copy edit, ~820 words); repaired after check (restored: fold scope, all-pair K-aug comparator 0.78,
+% E_N definition, 5-draw means, P = 0.25, signed-rank, same metrics, same-pair calibration, early stopping, bulk depth K).
+% Numbers cut here are still given in the Fig. 2 legend (5,508 genes, 22 layers, 102 test donors, 22 fold-0 donors),
+% Supplementary Fig. 9 (9.48 M), Fig. 10 (78-79 pairs, step 3,250) and Fig. 12 (all-pair rank values, K = 250-4,000).
+% Review fixes 2026-10-11: LoRA-5 vs calibration composition per draw (sec09_fewpair_cells.csv: ImMicro 5/5, sepsis 2/5,
+% Imperial 2/5, Giroux 2/5, Ji 3/5); cohort-mean null 0.1606 (sec08_pairs_realism.csv ref_null); condition-encoder-only
+% on fold 0 only; 40 fine-tune runs = ImMicro, sepsis, Giroux, Ji x (5 draws + 5 folds) (sec06_ft_method.csv).
+\subsection*{Five paired samples and a few minutes of LoRA fine-tuning adapt \method{} to real bulk}
+
+\textbf{Real bulk is not the pseudobulk the model learned from.} ImMicro pairs bulk RNA-seq and scRNA-seq of peripheral blood mononuclear cells (PBMCs) from 107 healthy donors\cite{babdor2026immicro}. In the first 30 principal components of the model's input, real bulk lay a median 2.46 donor distances from the donor's own pseudobulk (3.19 over all genes; Fig.~\ref{fig:fig2}a). One shift shared by all donors made up 95\% of the squared offset (91\% over all genes). Only 7.5\% of real bulks lay nearest their own donor's pseudobulk (chance 0.9\%), but 75.7\% did once the mean shift of the other donors was removed (a diagnostic, not a correction the method applies). Two more paired PBMC cohorts also showed one shared shift (71\% and 92\% of the squared offset)\cite{randolph2021genetic,arunachalam2020systems}. The donor signal survives beneath this shift.
+
+\textbf{Without fine-tuning, cells from real bulk fall off the real-cell manifold.} Fed the real bulk as measured (zero-shot), the model made 30.6\% monocytes, against 16.7\% in real cells and 14.3\% from the donors' pseudobulks (Fig.~\ref{fig:fig2}b). From the same starting noise, 19\% of cells that pseudobulk made into T cells became monocytes. Precision, the share of generated cells inside a real cell's neighbourhood\cite{kynkaanniemi2019improved}, fell from 0.92 with pseudobulk to 0.36, and coverage\cite{naeem2020reliable} from 0.51 to 0.08. Both inputs carried each donor's own single-cell depth and cell count, unlike the fixed bulk depth of $K=1{,}000$ cells used for fine-tuning and testing below (Fig.~\ref{fig:fig2}e).
+
+\textbf{LoRA adapts the model to a cohort's bulk in minutes.} We fine-tuned the model with LoRA\cite{hu2022lora} to generate each paired sample's cells from its real bulk. LoRA freezes the pretrained weights and trains a rank-16 update of 1.30 million weights, 2.4\% of the model's 54.2 million (Fig.~\ref{fig:fig2}c). A fine-tuning run took 2--14 min on one GPU (40 runs of the five-donor and 5-fold designs in four cohorts). In the main setting, LoRA-5 trains on the pairs of 5 donors for a fixed 3,000 steps and is tested on all other donors from their real bulk, over 5 draws. Donor-level 5-fold CV is reported separately. The comparators are zero-shot and zero-shot + calibration, which rescales each gene by a factor fitted on the same training pairs (Methods). All methods are scored with the same metrics against held-out real cells of the test donors (Fig.~\ref{fig:fig2}d).
+
+\textbf{Five pairs are enough to cover the real cells.} In ImMicro, LoRA-5 raised coverage to 0.79, from 0.19 for zero-shot and 0.54 for zero-shot + calibration (mean of 5 draws; Fig.~\ref{fig:fig2}e). Precision rose to 0.95 (0.68 and 0.90), and composition L1 fell to 0.18 (0.45 and 0.20). LoRA-5 beat calibration on all three in every draw (per-donor L1 lower by 0.027; Wilcoxon signed-rank test, $P=2.2\times10^{-10}$, 107 donors). In 5-fold CV, reported separately, LoRA reached coverage 0.78 and L1 0.16. A model fine-tuned on the same donors' cells with pseudobulk instead of real bulk reached only 0.62 and 0.24 when fed calibrated bulk. Most of the gain thus needs the paired real bulk.
+
+\textbf{LoRA cells are the most realistic variant, but still separable from real cells.} A $k$-nearest-neighbour classifier trained to tell generated from real cells\cite{lopezpaz2017revisiting} reached an area under the receiver operating characteristic curve (AUC) of 0.94 on LoRA-5 cells, against 0.97 with calibration and 1.00 for zero-shot. Real cells score 0.50 against each other (Fig.~\ref{fig:fig2}f). The squared maximum mean discrepancy (MMD$^2$) between generated and real cells\cite{gretton2012kernel} fell to 0.023 (0.043 and 0.103), and scSemiProfiler's own normalised error $E_N$\cite{wang2024scsemiprofiler} to 1.5 (2.9 and 7.2).
+
+\textbf{LoRA matches full fine-tuning with 42 times fewer trainable weights.} In 5-fold CV, training all weights gave a similar composition L1 (0.157 against 0.162 for LoRA; Wilcoxon $P=0.25$, 107 donors) but lower coverage (0.70 against 0.78) and precision (0.89 against 0.92) (Fig.~\ref{fig:fig2}g). Training only the condition encoder, which reads the bulk, was run on fold 0 alone. There it fell short of LoRA (L1 0.20 against 0.15; coverage 0.73 against 0.91) and did no better than calibration (coverage 0.73 for both; Supplementary Fig.~9).
+
+\textbf{More pairs mainly refine donor-level composition.} Pairs-needed curves vary the training pairs per fold in 5-fold CV, each run early-stopped on 7 inner-validation pairs; their 5-pair point is not the LoRA-5 draws (Fig.~\ref{fig:fig2}h and Supplementary Fig.~10). Five pairs already gave coverage 0.84 (calibration from the same pairs 0.53), and more pairs did not raise it. Composition L1 changed little (0.17 with 5 pairs, 0.16 with all), and in these healthy donors predicting each donor's composition as the other donors' mean also gives 0.16 (Supplementary Fig.~10). The across-donor correlation of generated and observed NK-cell fractions, which such a mean cannot capture, rose from 0.20 with 5 pairs to 0.70 with 40, passing zero-shot (0.49) from 20 pairs on.
+
+\textbf{Rank 16 and one bulk depth suffice.} With all training pairs, rank 64 raised coverage only slightly and lowered precision, and rank 4 worsened L1. With 5 pairs per fold, the ranks were mixed: rank 4 lowered L1 (0.162 against 0.171) but also coverage (0.79 against 0.84). With all training pairs, neither training at several bulk depths nor adding pseudobulk from the training donors' cells helped (coverage 0.77 for both, against 0.78). We kept rank 16 and $K=1{,}000$ (Fig.~\ref{fig:fig2}i and Supplementary Fig.~12).
+
+\textbf{The realism gain held in all five paired cohorts.} LoRA-5 had the highest coverage and the lowest classifier AUC in all 25 draws, and LoRA ranked first on every realism metric in 5-fold CV (AUC 0.87--0.94; Supplementary Figs~3--6). With 5 donors, however, LoRA-5 beat calibration on composition only in ImMicro (5 of 5 draws; 2 or 3 of 5 elsewhere), and in sepsis calibration had the lower mean L1 (0.45 against 0.47; Supplementary Table~2).

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/sections/results_fig3.tex

@@ -0,0 +1,9 @@
+\subsection*{LoRA cells are closer to real cells than scSemiProfiler's, in minutes instead of days}
+
+\textbf{Same five donors, better cells.} With the same five paired donors, \method{} with LoRA was ahead of scSemiProfiler on all six cell measures (Fig.~\ref{fig:fig3}a). scSemiProfiler\cite{wang2024scsemiprofiler} combines bulk for every sample with scRNA-seq of a few representative samples; we ran its official code with its default training settings and scored its official output. In each of five draws on the healthy ImMicro cohort\cite{babdor2026immicro}, scSemiProfiler-5 used LoRA-5's five training donors as representatives, and both methods then generated cells for the other 102 donors, given only their real bulk. LoRA-5 covered more of the real cells (coverage\cite{naeem2020reliable} 0.79 against 0.04) and had higher precision\cite{kynkaanniemi2019improved} (0.95 against 0.52). Its composition L1 (0.176 against 0.233), MMD$^2$\cite{gretton2012kernel} (0.023 against 0.186) and $E_N$ (1.5 against 3.6) were lower. A classifier\cite{lopezpaz2017revisiting} still told LoRA-5 cells from real ones (AUC 0.94) but separated scSemiProfiler-5 cells almost perfectly (0.999). By cohort means, LoRA-5 was ahead on all six measures in all five paired cohorts (Supplementary Figs~3 and~4). By draws, it was ahead on composition L1 in 22 of 25 (Giroux 3 of 5, Ji 4 of 5) and on the other five measures in all 25. A modified read-out without scSemiProfiler's fixed output threshold, not its official output, changed the winner in none of 23 cohort-by-metric comparisons with LoRA-5 (Supplementary Note~4 and Supplementary Fig.~17).
+
+\textbf{More pairs do not close the gap.} scSemiProfiler stayed behind when it received more representatives (Fig.~\ref{fig:fig3}b). Inside the 5-fold CV design, it received the same $n$ pairs per fold as LoRA ($n = 5$, 10, 20 or 40; not the five-donor draws) plus the seven pairs that LoRA used for early stopping. Its coverage stayed at 0.04 at every $n$, against 0.76--0.84 for LoRA. LoRA's composition L1 was lower at every $n$ and also with every training donor, the 5-fold CV result reported separately (0.16 against 0.22). scSemiProfiler tracked the donors' NK-cell fractions better only with five pairs (Pearson r 0.27 against 0.20; 0.49 against 0.70 with 40). In the 5-fold CV of all five cohorts, LoRA had the higher coverage and precision everywhere and the lower L1 in all but Giroux (Supplementary Fig.~7 and Supplementary Table~3).
+
+\textbf{scSemiProfiler copies its representative.} scSemiProfiler-5 did worse than simply copying the composition of the representative it assigns to each test sample (L1 0.233 against 0.187; Fig.~\ref{fig:fig3}c). For each test sample, scSemiProfiler picks the training sample nearest in bulk as its representative and derives the test cells from that sample's cells; its error tracked the copy's over test donors (Spearman 0.57). LoRA-5 fell below the copy's error (L1 0.176) and tracked it less (Spearman 0.35). Across cohorts, scSemiProfiler-5 was never significantly better than the copy and was significantly worse in three of five ($P \le 0.014$, paired Wilcoxon test), whereas LoRA-5 was below the copy in four of five (Supplementary Fig.~8). A proxy for scSemiProfiler's own choice of representatives gave a lower copy error in Imperial, Giroux and Ji, where it spreads the same number of samples over more donors, and a similar or higher one in ImMicro and sepsis (Supplementary Note~5).
+
+\textbf{Minutes instead of days.} For the same five donors, LoRA-5 needed 17--59 times less GPU time per cohort than scSemiProfiler-5 (Fig.~\ref{fig:fig3}d; Methods). On average, a five-donor draw took scSemiProfiler-5 57 min to 4.1 h, depending on the cohort, and LoRA-5 3--11 min. A cohort's five draws took 5--20 h for scSemiProfiler-5 against 13--54 min for LoRA-5. A 5-fold CV took scSemiProfiler up to 1.6 GPU days (sepsis), against 37 min for LoRA. The 25 five-donor draws and 25 fold models together took 7.7 GPU days for scSemiProfiler and 5.7 GPU hours for LoRA (Supplementary Table~4). The gap is built into the designs. scSemiProfiler trains a model for every representative and then one for every test sample, so its time grows with the number of samples. In an ImMicro draw, 126 of its 141 min on average went to the 102 per-sample models. LoRA-5 trains once, in 1.7--2.0 min, and then only samples cells.

paper-releases/pbisc-diffusion/updates/20261010T153319Z-manuscript-rewrite-from-deck/diffusion/dataset/server_1/manuscript/sections/results_fig4.tex

@@ -0,0 +1,17 @@
+\subsection*{LoRA cells carry disease signal beyond the bulk, ahead of scSemiProfiler and cell-type deconvolution}
+
+\textbf{Four public paired cohorts carry a disease contrast.} Five cohorts gave 403 usable pairs from 275 donors (Fig.~\ref{fig:fig4}a and Supplementary Table~1); all but the healthy ImMicro cohort\cite{babdor2026immicro} carry a contrast. The sepsis cohort\cite{ye2026sepsis} has 61 sepsis vs 21 control samples, Imperial\cite{gisby2022multiomics,stephenson2025temporal} 113 COVID-19-positive vs 34 negative samples from haemodialysis patients, and Giroux\cite{giroux2022differential} 11 moderate vs 13 mild samples from nine COVID-19 outpatients. Ji\cite{ji2025chili} compares 18 samples from patients with checkpoint-inhibitor liver injury with 15 from treated controls and has whole-blood bulk. In Giroux, severity is confounded with sex, so its disease metrics exclude 65 sex-linked genes\cite{tukiainen2017landscape}.
+
+\textbf{Partial r measures what single cells add to the bulk.} logFC r is the Pearson r, across genes, between generated and observed case-vs-control log fold changes (logFC) per cell type, averaged over T, B, monocyte and NK cells (Methods). Partial r removes the part inherited from the bulk by regressing the input bulk's logFC out of both\cite{vinastorne2026systema}. It is the cell-type signal that the bulk does not carry. Top-50 DEG counts genes shared by the observed and generated top-50 lists (false discovery rate below 0.05), summed over these types and dendritic cells.
+
+\textbf{With five paired donors, LoRA-5 leads scSemiProfiler-5 in most draws.} In the main comparison (five draws of five donors per cohort; Fig.~\ref{fig:fig4}b), LoRA-5\cite{hu2022lora} was ahead of scSemiProfiler-5\cite{wang2024scsemiprofiler} on logFC r in 13 of 20 draws and on partial r, the signal beyond the bulk, in 14 of 20. Mean partial r (LoRA-5 vs scSemiProfiler-5) was 0.60 vs 0.43 in Imperial, 0.30 vs 0.09 in Ji, 0.30 vs 0.33 in sepsis and 0.19 vs 0.24 in Giroux. Mean logFC r was, in the same order, 0.75 vs 0.56, 0.33 vs 0.15, 0.35 vs 0.40 and 0.30 vs 0.40 (Supplementary Table~2). LoRA-5 led in every Imperial and Ji draw and in three of five sepsis draws, though its sepsis means were lower. In Giroux, the exception, scSemiProfiler-5 was ahead; each test set there holds only four subjects, two per group. On top-50 DEG, LoRA-5 was ahead in 12 of 20 draws; the other 8, in Giroux and Ji, were ties at zero. It also beat both zero-shot versions on mean logFC r and partial r in all four cohorts.
+
+\textbf{With all training pairs, LoRA reaches a disease logFC r of 0.82 in Imperial.} In 5-fold CV, reported separately (Fig.~\ref{fig:fig4}c), LoRA cells in Imperial matched the observed COVID-19 logFC at $r=0.82$ (0.90 in T cells), with partial r 0.69 and 143 shared top-50 DEG, against 0.71, 0.54 and 87 for scSemiProfiler given the same pairs. LoRA had the highest logFC r of the \method{} versions in all four cohorts, above both zero-shot versions and an sc-only fine-tune trained without the paired bulk. It led scSemiProfiler on logFC r and partial r in all cohorts but Giroux and on top-50 DEG in all four (Supplementary Table~3).
+
+\textbf{The lead holds without a significance cut-off.} Correspondence at the top (CAT)\cite{irizarry2005multiple} needs no such cut-off. It ranks each cell type's genes by the absolute case-vs-control moderated $t$ of limma-voom\cite{law2014voom} and reports the share of the observed top $k$ genes also in the generated top $k$ with the same sign, averaged over the four types. At $k=100$, LoRA-5 had the highest CAT in three of four cohorts and beat scSemiProfiler-5 in the same three (Imperial 0.24 vs 0.13, sepsis 0.09 vs 0.04, Ji 0.09 vs 0.03), against chance of 0.012--0.016 (Fig.~\ref{fig:fig4}d). Rank-biased overlap\cite{webber2010similarity}, which weights the top of both rankings most, picked the same leader in every cohort. In 5-fold CV, where Fig.~\ref{fig:fig4}d compares only the \method{} versions, LoRA was highest at $k=100$ in sepsis and Ji, and the sc-only fine-tune in Imperial and Giroux.
+
+\textbf{Cell-type deconvolution with the same pairs adds little beyond the bulk.} BayesPrism\cite{chu2022bayesprism} and CIBERSORTx\cite{newman2019cibersortx}, two cell-type deconvolution methods, estimate cell-type fractions and per-type expression, not single cells. Both received the same five donors' cells as reference and the real bulk as mixture (Fig.~\ref{fig:fig4}e). On logFC r, LoRA-5 led BayesPrism in 15 of 20 draws and CIBERSORTx in all 17 scored draws; BayesPrism led in Giroux. On partial r, LoRA-5 was ahead in 18 of 20 and 17 of 17 draws; deconvolution's cohort means reached at most 0.19 and were below LoRA-5's in every cohort (0.19--0.60). Top-50 DEG was mixed: LoRA-5 was ahead in 6 of 20 and 6 of 17 draws and behind in 8 of 20 and 4 of 17; deconvolution led in sepsis, LoRA-5 in Imperial. LoRA-5 was also closer to the observed composition in 22 of 25 and 24 of 25 draws. BayesPrism recovers the observed mix from pseudobulk, so its composition error comes from the real bulk (Supplementary Fig.~14).
+
+\textbf{Even without fine-tuning, the pretrained model beats scSemiProfiler on a lupus cohort.} A separate benchmark used GSE135779\cite{nehar2020mapping}, an SLE cohort of children and adults absent from the pretraining corpus. Each of 42 donors (30 SLE, 12 controls) was given as its own pseudobulk; scSemiProfiler also received 14 other donors' cells. Over five sampling seeds (mean $\pm$ s.d.), \method{} was ahead on all four metrics (Fig.~\ref{fig:fig4}f). On 1,006 genes and six cell types, logFC r was 0.371 $\pm$ 0.018 vs 0.318, top-50 DEG 76.2 $\pm$ 3.7 vs 56 and partial r 0.208 $\pm$ 0.012 vs 0.167. Composition L1 was 0.181 $\pm$ 0.003 vs 0.264. Per cell type, \method{} led in T, B, monocyte and NK cells and trailed in the small `Other' group (logFC r 0.10 vs 0.27). On dendritic cells its logFC r was similar (0.29 vs 0.28), but it shared fewer top-50 DEG (2.6 vs 9), as it generated only about 15 of these cells per donor. These donors also guided the choice of model width and gene panel (Fig.~\ref{fig:fig1}e,f), making this a development cohort, not an untouched test set.
+
+\textbf{Rank 16 is enough, and the sepsis disease signal needs pairs.} In sepsis (5-fold CV), rank-16 LoRA matched full fine-tuning (Supplementary Fig.~13), and logFC r stayed at or below 0.12 without pairs or with calibration alone. On the sepsis pairs-needed curves, at equal budgets of 5 to 40 pairs, scSemiProfiler's logFC r was higher than LoRA's (0.52--0.58 vs 0.29--0.48), but its top-50 DEG stayed at 2--6 (LoRA 5--22; Supplementary Fig.~11).