Archive 4 checkpoints (1-4)

.gitattributes

@@ -71,3 +71,5 @@ paper-releases/pbisc-diffusion/updates/20261001T145641Z-lora-rank-sweep-calibrat
 paper-releases/pbisc-diffusion/updates/20261001T145641Z-lora-rank-sweep-calibration-budget/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_paper_summary_20260930.pptx filter=lfs diff=lfs merge=lfs -text
 paper-releases/pbisc-diffusion/updates/20261002T002920Z-deck-extra-metrics-scsemi-prep/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_paper_summary_20260930.pdf filter=lfs diff=lfs merge=lfs -text
 paper-releases/pbisc-diffusion/updates/20261002T002920Z-deck-extra-metrics-scsemi-prep/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_paper_summary_20260930.pptx filter=lfs diff=lfs merge=lfs -text
+paper-releases/pbisc-diffusion/updates/20261002T022002Z-deck-time-to-apply-scsemi-reps/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_paper_summary_20260930.pdf filter=lfs diff=lfs merge=lfs -text
+paper-releases/pbisc-diffusion/updates/20261002T022002Z-deck-time-to-apply-scsemi-reps/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_paper_summary_20260930.pptx filter=lfs diff=lfs merge=lfs -text

paper-releases/pbisc-diffusion/updates/20261002T022002Z-deck-time-to-apply-scsemi-reps/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_paper_summary_20260930.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0016014c1001731abe40a49a025a6e12709ad457e885e3ac94851e55189a7929
+size 5647990

paper-releases/pbisc-diffusion/updates/20261002T022002Z-deck-time-to-apply-scsemi-reps/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/PBISC-Diffusion_paper_summary_20260930.pptx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe22865ddb3ac01d0c052f3dc121d2873d14f485542b37949d29fb4cd236a775
+size 6353209

paper-releases/pbisc-diffusion/updates/20261002T022002Z-deck-time-to-apply-scsemi-reps/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/README.md

@@ -0,0 +1,104 @@
+# PBISC-Diffusion paper summary deck (2026-09-30, r00)
+
+34-slide English summary of the PBISC-Diffusion (Model 8) paper experiments: title,
+outline, 10 sections (25 content slides), then a Backup divider and 7 backup slides.
+Short keyword lines only; the figures carry the results. Section 10 is planned work.
+
+- `PBISC-Diffusion_paper_summary_20260930.pptx`: the deck
+- `PBISC-Diffusion_paper_summary_20260930.pdf`: LibreOffice PDF of the deck
+- `data/slides_spec.json`: slide order, titles, keywords, footnotes, speaker-note text
+- `generate_deck.py`: builds the pptx from `data/slides_spec.json` only (python-pptx)
+- `figures/`: all slide figures (PNG + PDF), made by `tools/sec*.py`
+- `data/sec*.csv`, `data/sec09_slide_text.json`, `data/sec09_extra_slide_text.json`,
+  `data/sec07_rank_sweep_slide_text.json`, `data/sec09_time_to_apply_slide_text.json`,
+  `data/sec09_scsemi_representative_slide_text.json`: every plotted or quoted number with its source path
+- `render/`: PDF and 70-dpi slide JPGs for checking
+
+## Rebuild
+
+Everything is CPU only; no GPU or checkpoint is touched.
+
+```bash
+# prerequisite of tools/sec09_scsemi_representative.py (CPU; copies server_2 inputs once):
+#   python3 ../../plans/paired_finetune_intradataset_20260927/scsemi/tools/representative_dependence.py
+cd /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00
+# figures + data tables (sec03_common.py, style.py and scsemi_results.py are shared helpers, not run directly)
+for f in sec01_overview sec02_backbone sec03_width_loss sec03_width_bench sec03_panel \
+         sec04_sle_benchmark sec05_condition_space sec05_cells_offmanifold \
+         sec05_behcet_direction sec05_shift_cohorts sec06_finetune_method sec07_ft_modes \
+         sec07_k_aug sec07_rank_sweep sec08_pairs_needed sec09_paired_validation sec09_extra_metrics \
+         sec09_time_to_apply sec09_scsemi_representative sec10_paired_subset_cohorts; do
+  python3 "tools/$f.py" || break
+done
+python3 tools/sync_sec09_text.py      # copy generated slide text (sec09, sec09_extra, sec07_rank_sweep,
+                                      # sec09_time_to_apply, sec09_scsemi_representative) into the spec
+python3 generate_deck.py              # -> PBISC-Diffusion_paper_summary_20260930.pptx
+soffice --headless --convert-to pdf --outdir render PBISC-Diffusion_paper_summary_20260930.pptx
+pdftoppm -jpeg -r 70 render/PBISC-Diffusion_paper_summary_20260930.pdf render/slide
+cp render/PBISC-Diffusion_paper_summary_20260930.pdf .
+```
+
+`generate_deck.py` stops with an error if a figure is missing, a title does not fit on one
+line, a footnote needs more than 2 lines, or any text contains Korean or placeholder words.
+
+## Notes
+
+- Section 9 (few-pair scSemiProfiler comparison) uses all 25 scSemiProfiler draws
+  (5 cohorts x 5 draws). If scores change, rerun `tools/sec09_paired_validation.py`, then
+  `tools/sync_sec09_text.py`, then check the short footnotes it lists and rebuild.
+- Section 9 update (user decisions 2026-10-01): (1) the gene-wise calibration comparator is back,
+  named "Zero-shot + calibration" (few-pair `zs_cal`, alpha from the same 5 donors; 5-fold `zs__cal`)
+  and the sc-only control is "sc-only fine-tune (+ calibration)" (`pseudo_full__cal`); (2) Giroux's
+  disease metrics (logFC r, partial r, top-50 DEG, broadcast) exclude 196 sex-linked genes because its
+  severity is confounded with sex, read from `plans/paired_finetune_intradataset_20260927/scores_nosex/`
+  (see its README). Both are stated in the slide footnotes.
+- On-slide footnotes are one short line; the full source and method text is in the speaker
+  notes (`footnote_full` in the spec).
+- `DRAFT_1150_*`: frozen copy shown at the 2026-09-30 12:00 meeting (23/25 scSemiProfiler
+  draws, long footnotes). Superseded by the files above.
+- Story rule (user decision 2026-09-30): the deck and paper do not mention PBISC-VAE, reference
+  calibration (refcal), the TPM input variant or any calibration arm (calibration arm reinstated
+  2026-10-01 as "Zero-shot + calibration"; never "refcal"). The no-fine-tune baseline is
+  zero-shot on real bulk as measured; the sc-only control is the sc-only fine-tune fed raw bulk.
+  `tools/sec04_immicro.py` (refcal vs scSemiProfiler on ImMicro) is kept but no longer in the deck.
+  The pre-revision deck, scripts and figures are in `_archive/2026-09-30_before_no_vae_refcal/`.
+
+- Added 2026-10-02 (from `plans/paired_finetune_intradataset_20260927/scores_extra/` and `scores/<cohort>/rankcal_20261001.*`):
+  slide 23 "5 donors: realism beyond precision/coverage" (classifier AUC, MMD², scSemiProfiler's E_N per cohort;
+  few-pair arms, mean ± SD over draws) and three backups: slide 30 LoRA rank r = 4 / 16 / 64 (ImMicro, sepsis),
+  slide 31 the same realism metrics for the 5-fold pooled design, slide 32 precision/coverage at k = 3 / 5 / 10
+  (few-pair; the 5-fold k ranking is in the speaker notes). Scripts: `tools/sec09_extra_metrics.py` (checks every
+  value against `scores_extra/{fewshot,pooled}_headline.csv`) and `tools/sec07_rank_sweep.py` (checks
+  `rankcal_20261001.json` against the scoring calls). Their slide text is generated (`text_source` in the spec) and
+  synced by `tools/sync_sec09_text.py`. Pre-update deck: `_archive/2026-10-01_before_extra_metrics/`.
+- scSemiProfiler 5-fold and equal-budget results are picked up automatically when they appear
+  (`tools/scsemi_results.py`: arm `scsemi__raw` in `scores/<cohort>/scsemi_cv_20261002/pooled*/`, arms
+  `scsemi_n{5,10,20,40}__raw` in `scores/<cohort>/scsemi_budget_20261002/pooled*/`, extra metrics below
+  `scores_extra/<cohort>/scsemi_cv_20261002/`; a call is used only for the same donors and reference cells as the
+  main pooled call). When present: a "scSemiProfiler, all training pairs" bar on slide 24 (Giroux disease metrics only
+  after a `scores_nosex/giroux2022/scsemi_cv_20261002/` rescore, else "rescore pending"), a scSemiProfiler line across
+  n on slides 18–19 ("all" = the 5-fold run), and a scSemiProfiler row on slide 31. When absent, each script prints a
+  NOTE and the figures are unchanged. After they arrive: rerun the four scripts, sync, check the hand-written keywords
+  of slides 18–19 (not generated) and rebuild. The presence path was tested with synthetic arms in a scratch copy
+  (`DECK_SCSEMI_PLAN_ROOT`, `DECK_FIGDIR`, `DECK_OUTDATA` redirect the lookup and outputs).
+
+- Added 2026-10-02, second update (pre-update deck: `_archive/2026-10-02_before_time_to_apply/`). Slide numbers
+  from 24 on moved by one: "All pairs, 5-fold" is now slide 25, the backups are 28-34 (LoRA rank 31, 5-fold realism
+  32, k = 3 / 5 / 10 33). In the note above, read slide 24 -> 25 and slide 31 -> 32.
+  - Slide 24 "5 donors: time to apply" (`tools/sec09_time_to_apply.py`): GPU time per draw of LoRA-5 vs
+    scSemiProfiler-5 on the same 25 draws, read from receipts and queue logs on both hosts (LoRA RUN.json + sampling
+    job wall clock; scSemiProfiler model receipts + export `total_target_seconds`). The script stops if a draw lacks
+    a complete record. Data prep, scoring and PBISC's one-time pretraining are not counted (footnote).
+  - Backup slide 34 "scSemiProfiler copies its representative" (`tools/sec09_scsemi_representative.py`, from
+    `plans/paired_finetune_intradataset_20260927/scsemi/analysis/representative_dependence_20261002/`): composition
+    L1 of copying the assigned representative's observed composition vs scSemiProfiler-5 and LoRA-5, and their
+    Spearman with that copy error. The KMeans proxy for upstream's own representative choice (not a rerun) is in the
+    speaker notes only.
+
+## Sources
+
+Final model: `runs/f01_6k_1x_seed20260627` (epoch 16, EMA). Scores and receipts come from
+`plans/utility_studies_final6k_20260925/`, `plans/paired_finetune_intradataset_20260927/`,
+`plans/immicro_scsemi_comparison_20260927/`, `plans/floor01_campaign_20260923/`,
+`analysis/scsemi_benchmark_diffusion_20260913/` and the run folders named in each
+`data/sec*.csv`. Manuscript: `/home/sj_server_1/PBISC/paper-release-20260928/manuscript/`.

paper-releases/pbisc-diffusion/updates/20261002T022002Z-deck-time-to-apply-scsemi-reps/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/tools/sec09_scsemi_representative.py

@@ -0,0 +1,144 @@
+#!/usr/bin/env python3
+"""Backup slide for section 9 of the 2026-09-30 paper-summary deck: scSemiProfiler's composition follows its representative.
+
+CPU only. Reads plans/paired_finetune_intradataset_20260927/scsemi/analysis/representative_dependence_20261002/
+{per_target.csv, summary.csv, kmeans_proxy.csv} (made by scsemi/tools/representative_dependence.py; rerun that first)
+and draws:
+  a  composition L1 per cohort: copying the representative's observed composition, scSemiProfiler-5, LoRA-5
+     (bar = mean over 5 draws of the per-draw mean, whiskers = SD, dots = draws)
+  b  Spearman, over test samples, of each method's L1 with the copy error
+The KMeans proxy (upstream's own representative choice, not a rerun) goes to the speaker notes only.
+
+Outputs (deck folder; DECK_FIGDIR / DECK_OUTDATA override the folders):
+  figures/sec09_scsemi_representative.{png,pdf}
+  data/sec09_scsemi_representative.csv              every plotted value with its source
+  data/sec09_scsemi_representative_slide_text.json  title, keywords, footnotes, quoted numbers (synced by
+                                                    tools/sync_sec09_text.py through text_source)
+"""
+from __future__ import annotations
+
+import os
+
+os.environ["CUDA_VISIBLE_DEVICES"] = ""  # plotting only
+
+import json
+import sys
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+
+DECK = Path("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/"
+            "2026-09-30--paper-summary--r00")
+sys.path.insert(0, str(DECK / "tools"))
+import sec09_paired_validation as S9  # noqa: E402  (shared section 9 names and colours; sets the deck rcParams)
+import style  # noqa: E402
+import matplotlib.pyplot as plt  # noqa: E402
+from matplotlib.patches import Patch  # noqa: E402
+
+A = S9.T / "scsemi" / "analysis" / "representative_dependence_20261002"
+FIGDIR = Path(os.environ.get("DECK_FIGDIR", str(DECK / "figures")))
+OUTDATA = Path(os.environ.get("DECK_OUTDATA", str(DECK / "data")))
+ORDER, NAME, ROWNAME = S9.ORDER, S9.NAME, S9.ROWNAME
+FS, SUB, EDGE = S9.FS, S9.SUB, S9.EDGE
+ARMS = [("copy", "copy_l1", style.ARM["broadcast"], "Copy the representative's composition"),
+        ("scsemi", "scsemi_l1", style.ARM["scsemi"], "scSemiProfiler, 5 donors"),
+        ("lora", "lora_l1", style.ARM["lora"], "LoRA, 5 donors")]
+
+
+def main() -> int:
+    per = pd.read_csv(A / "per_target.csv")
+    summ = pd.read_csv(A / "summary.csv").set_index("cohort")
+    prox = pd.read_csv(A / "kmeans_proxy.csv").set_index("cohort")
+    by_draw = per.groupby(["cohort", "draw"])[[c for _, c, _, _ in ARMS]].mean()
+    # the plotted pooled means must equal the analysis summary
+    for ds in ORDER:
+        for _, c, _, _ in ARMS:
+            if not np.isclose(per[per.cohort == ds][c].mean(), summ.loc[ds, c]):
+                raise SystemExit(f"{ds} {c}: per_target and summary disagree")
+    fig, (ax, bx) = plt.subplots(1, 2, figsize=(12.3, 4.6), gridspec_kw={"width_ratios": [2.0, 1.0], "wspace": 0.12})
+    h, gap = 0.27, 0.4
+    yt, rows = [], []
+    for i, ds in enumerate(ORDER):
+        y0 = i * (3 * h + gap)
+        for k, (key, col, colour, lab) in enumerate(ARMS):
+            v = by_draw.loc[ds][col].to_numpy()
+            y = y0 + k * h
+            ax.barh(y, v.mean(), h * 0.88, color=colour, edgecolor="none")
+            ax.errorbar(v.mean(), y, xerr=v.std(ddof=1), fmt="none", ecolor=EDGE, elinewidth=1, capsize=2.5)
+            ax.scatter(v, np.full(v.size, y), s=12, facecolor="white", edgecolor=EDGE, linewidth=0.8, zorder=3)
+            rows += [dict(panel="a", cohort=ds, series=key, draw=j, value=float(x),
+                          source=str(A / "per_target.csv")) for j, x in enumerate(v)]
+        yt.append(y0 + h)
+        p = summ.loc[ds, "scsemi_vs_copy_wilcoxon_p"]
+        diff = summ.loc[ds, "scsemi_minus_copy"]
+        verdict = "worse" if (p < 0.05 and diff > 0) else ("better" if (p < 0.05 and diff < 0) else "no difference")
+        ax.text(0.79, y0 + h, f"scSemi vs copy:\n{verdict}", va="center", ha="left", fontsize=FS - 3, color=SUB,
+                linespacing=1.1)
+    nt = per.groupby(["cohort", "draw"]).size().groupby(level=0).agg(["min", "max"])
+    ax.set_yticks(yt, [f"{ROWNAME[d]}\ntest {nt.loc[d, 'min']}" + (f"–{nt.loc[d, 'max']}" if nt.loc[d, 'max'] != nt.loc[d, 'min'] else "")
+                       + " per draw" for d in ORDER], fontsize=FS - 2)
+    ax.invert_yaxis()
+    ax.set_xlim(0, 1.02)
+    ax.set_xticks([0, 0.2, 0.4, 0.6])
+    ax.set_xlabel("Composition L1 ↓")
+    ax.tick_params(axis="y", length=0)
+    for i, ds in enumerate(ORDER):
+        y0 = i * (3 * h + gap)
+        for k, (key, col) in enumerate((("scsemi", "rho_scsemi_copy"), ("lora", "rho_lora_copy"))):
+            v = summ.loc[ds, col]
+            colour = style.ARM[key]
+            bx.barh(y0 + (k + 1) * h, v, h * 0.88, color=colour, edgecolor="none")
+            bx.text(v + 0.02, y0 + (k + 1) * h, f"{v:.2f}", va="center", fontsize=FS - 3, color=SUB)
+            rows.append(dict(panel="b", cohort=ds, series=key, draw=None, value=float(v), source=str(A / "summary.csv")))
+    bx.set_yticks(yt, [""] * len(yt))
+    bx.tick_params(axis="y", length=0)
+    bx.invert_yaxis()
+    bx.set_xlim(0, 1.08)
+    bx.set_xlabel("Spearman with the copy error")
+    fig.legend(handles=[Patch(color=c, label=l) for _, _, c, l in ARMS], loc="lower center", ncol=3,
+               bbox_to_anchor=(0.45, -0.08), fontsize=FS - 2, handlelength=1.2, columnspacing=1.6)
+    style.save(fig, "sec09_scsemi_representative", FIGDIR)
+    OUTDATA.mkdir(exist_ok=True)
+    pd.DataFrame(rows).to_csv(OUTDATA / "sec09_scsemi_representative.csv", index=False)
+
+    s = summ.loc[list(ORDER)]
+    worse = [d for d in ORDER if s.loc[d, "scsemi_vs_copy_wilcoxon_p"] < 0.05 and s.loc[d, "scsemi_minus_copy"] > 0]
+    better = [d for d in ORDER if s.loc[d, "scsemi_vs_copy_wilcoxon_p"] < 0.05 and s.loc[d, "scsemi_minus_copy"] < 0]
+    lora_below = [d for d in ORDER if s.loc[d, "lora_l1"] < s.loc[d, "copy_l1"]]
+    kw = [f"scSemi-5 never beats copying its representative: {len(ORDER) - len(better)}/{len(ORDER)}",
+          f"Its error tracks the representative's: rho {s.rho_scsemi_copy.min():.2f}–{s.rho_scsemi_copy.max():.2f}",
+          f"LoRA-5 below the copy error in {len(lora_below)}/{len(ORDER)} cohorts"]
+    numbers = [{"value": f"{NAME[d]}: copy {s.loc[d, 'copy_l1']:.3f}, scSemi-5 {s.loc[d, 'scsemi_l1']:.3f} "
+                         f"(Wilcoxon vs copy p {s.loc[d, 'scsemi_vs_copy_wilcoxon_p']:.2g}), LoRA-5 {s.loc[d, 'lora_l1']:.3f}; "
+                         f"rho scSemi {s.loc[d, 'rho_scsemi_copy']:.2f}, LoRA {s.loc[d, 'rho_lora_copy']:.2f}",
+                "source": str(A / "summary.csv")} for d in ORDER]
+    proxy = "; ".join(f"{NAME[d]} ({int(prox.loc[d, 'n_representative_samples'])} samples) "
+                      f"ours {prox.loc[d, 'copy_l1_our_draws']:.2f}, KMeans {prox.loc[d, 'copy_l1_kmeans']:.2f}, "
+                      f"random samples {prox.loc[d, 'copy_l1_random_samples']:.2f}" for d in ORDER)
+    numbers.append({"value": "KMeans proxy, copy error: " + proxy, "source": str(A / "kmeans_proxy.csv")})
+    full = ("Same 25 five-donor draws. scSemiProfiler gives every test sample exactly one representative, the nearest "
+            "training sample in its bulk PCA (assignment.json), and runs its encoder on that representative's cells. "
+            "Copy = composition L1 between the test sample's and its representative's observed cells (same frozen "
+            "scANVI labels as the scorer). Bars: mean over draws of the per-draw mean; dots = draws. scSemi vs copy: "
+            "paired Wilcoxon over test samples (" + ", ".join(f"{NAME[d]} p {s.loc[d, 'scsemi_vs_copy_wilcoxon_p']:.2g}"
+                                                             for d in ORDER) + "). "
+            "Upstream's own representative choice (KMeans on the same bulk PCA, the sample nearest each centre; proxy, "
+            "not a rerun) gives a lower copy error than our 5-donor draws in Imperial, Giroux and Ji, where our draws hold "
+            "all visits of 5 donors and KMeans spreads the same number of samples over as many donors; equal or higher "
+            "in ImMicro and sepsis. Copy error: " + proxy + ". "
+            "Source: plans/paired_finetune_intradataset_20260927/scsemi/analysis/representative_dependence_20261002/.")
+    txt = {"title": "scSemiProfiler copies its representative", "layout": "full",
+           "figures": [str(FIGDIR / "sec09_scsemi_representative.png")], "keywords": kw,
+           "footnote_short": "Copy = composition of the one training sample scSemiProfiler assigns to each test sample.",
+           "footnote": full, "numbers": numbers}
+    out = {"section": "09-scsemi-representative", "scripts": [str(Path(__file__).resolve())],
+           "generated_from": "tools/sec09_scsemi_representative.py (rerun to refresh; do not edit by hand)",
+           "slides": [txt]}
+    (OUTDATA / "sec09_scsemi_representative_slide_text.json").write_text(json.dumps(out, indent=1, ensure_ascii=False))
+    print("\n".join(kw))
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())