Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261006T133650Z-fig1-nature-style-candidate/diffusion/dataset/server_1/manuscript/editorial/fig1_nature_redesign_plan_20261006.md

@@ -0,0 +1,92 @@
+# Figure 1 redesign plan: Nature-style overview (2026-10-06)
+
+## What the references do
+
+Five Nature-family Fig. 1 overviews were studied:
+- Nicheformer, *Nat Methods* 2025
+- Squidiff, *Nat Methods* 2025 (a diffusion model)
+- scGPT, *Nat Methods* 2024
+- GEARS, *Nat Biotechnol* 2024
+- scSemiProfiler, *Nat Commun* 2024
+- Bulk2Space, *Nat Commun* 2022
+
+They share these traits:
+
+1. **Story before machinery.** Panel a is the workflow: data, then training, then use, read left to right. The model's
+   inside comes second and is simplified. Layer widths, tensor sizes and equations go to Methods or Extended Data.
+2. **Data drawn as data.** Cells × genes appear as small coloured tiles, a bulk profile as one column, cells as coloured
+   dots or a UMAP cloud, and noise as a grey cloud.
+3. **Few symbols.** At most x₀, x_t, x_T and one condition symbol. No "3G+6" or "→ 512".
+4. **A row of uses.** Small plots show what the output is for, such as composition, effects and subsets.
+5. **Quiet visual language.**
+   - Type: Arial/Helvetica, 5–7 pt labels, 8 pt bold lower-case panel letters, sentence case.
+   - Colour: two or three accents, with the same colour for the same thing in every panel.
+   - Boxes: rounded, light tints, thin lines, and headers as small banners or plain bold text.
+
+## What the current Fig. 1 does differently
+
+The current Fig. 1 is an engineering diagram:
+- it shows "3G + 6 → MLP → 512", "256: internal width", "[v̂, tanh(ℓ/2)]", "4G", "x_t bypass" and "repeat · 2G";
+- it gives no view of the training data or of what the generated cells are for;
+- in the manuscript text, the read-out that the paper recommends (drawing detection from the support head's probability)
+  appears only as "hard zero/nonzero choice".
+
+## Plan
+
+Canvas 180 mm wide, Arial. Panels a–c are new; d is the existing corpus strip (`fig01_corpus`, unchanged).
+
+**a, Pretraining on pseudobulk (top left).**
+- Corpus badge: 41.4 M cells, 22 datasets, 3,199 donors (numbers from `source_data/corpus/corpus_summary.json`).
+- One sample's cells shown as a cells × genes tile block. K cells are picked (K = 250–4,000) and summed into a bulk column,
+  which carries a depth value and the measured-gene mask.
+- One highlighted cell row is the training target, paired with the bulk of its own recipe.
+- Message: every cell is learned together with the bulk it belongs to; no cell-type labels.
+
+**b, Diffusion on each gene's value and detection (top right).**
+- A cell is drawn as two aligned strips over the genes: expression value (blue bars) and detection (on/off dots).
+  Unmeasured genes are hatched grey.
+- Forward: noise added in gene space, x₀ → x_T. Reverse: a denoiser steps back from x_T.
+- Denoiser: a stack of blocks with skip arcs. It receives the noisy cell, its own previous estimate (dashed loop) and the
+  time step, and the bulk condition enters every block (gold lines, same gold as a).
+- Two outputs per gene: a value and a detection probability. Read-out: draw detection from the probability, keep the
+  value where detected, zero elsewhere. This gives counts with exact zeros. The paper's recommended read-out is drawn as
+  a probability bar followed by a draw.
+
+**c, From a measured bulk to cells (bottom row, full width).**
+- A clinical blood sample's bulk RNA-seq becomes a bulk column, with depth and that assay's measured-gene mask.
+- Optional adaptation, in a dashed box: a few paired bulk + single-cell samples give a LoRA update in minutes
+  (the paired fine-tuning result in the paper summary deck).
+- The pretrained model (mini b icon) receives one shared condition and K independent noise draws, and outputs K cells:
+  a UMAP-like cloud coloured by cell type, labelled after generation.
+- Uses, as three mini-plots: cell-type composition (stacked bars), cell-type-specific disease effect (generated vs
+  observed scatter) and a rare subset (a highlighted cluster).
+- Note: summed cells are compared with the bulk, not forced to match it, and there are no cell-type quotas.
+
+**d, Pretraining corpus.** The existing strip (assay shares, condition families, cell types).
+
+**Colour semantics** (kept across a–d):
+- gold: bulk / condition
+- blue: expression value
+- teal: detection
+- grey: noise and unmeasured genes
+- slate: model blocks
+- purple: LoRA adaptation
+- cell-type colours as in `fig01_corpus` (T blue, monocyte green, B ochre, NK light blue)
+
+**Text budget:** at most about 12 words per box. Numbers appear only where they are facts of the method (K range,
+corpus size). Everything else is explained in Methods.
+
+## Open choices for the authors
+
+- LoRA in c: the manuscript text does not yet describe paired fine-tuning. Keep the box if the paper adopts it; else drop it.
+- Mini-plots in c are schematic (no data). If they should show real results, they can be drawn from the source tables.
+
+## Files
+
+- `scripts/build_figure1_architecture_nature.py` writes `figures/fig01_architecture_nature.svg` (the editable master,
+  180 x 95 mm, Arial) and renders the PDF and PNG.
+- `scripts/corpus/build_figure1_corpus.py` writes `figures/fig01_corpus_nature.*`, the panel-d strip: Arial, bold title,
+  no corpus totals because panel a carries them.
+- `figures/drafts/fig01_nature_preview_20261006.png` stacks a-c over d as a preview.
+- `fig01_architecture.svg`, `fig01_corpus.*` (the current panel c) and `sections/overview_figure.tex` stay unchanged until
+  the new version is approved.