Archive 4 checkpoints (1-4)

.gitattributes

@@ -108,3 +108,4 @@ paper-releases/pbisc-diffusion/updates/20261008T121550Z-deck-fig4-rules-no-broad
 paper-releases/pbisc-diffusion/updates/20261008T121550Z-deck-fig4-rules-no-broadcast/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/20261008T125804Z-deck-fig2-3-healthy-fig5-rare/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/20261008T125804Z-deck-fig2-3-healthy-fig5-rare/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/20261008T125804Z-deck-fig2-3-healthy-fig5-rare/diffusion/dataset/server_1/model8/plans/rare_states_paired_20261008/out/per_donor.csv filter=lfs diff=lfs merge=lfs -text

paper-releases/pbisc-diffusion/updates/20261008T125804Z-deck-fig2-3-healthy-fig5-rare/diffusion/dataset/server_1/model8/plans/rare_states_paired_20261008/README.md

@@ -0,0 +1,58 @@
+# Rare cell states in the four paired disease cohorts (2026-10-08)
+
+**User request (2026-10-08, Fig. 5):** show rare-cell fractions and how well rare states are found, on the paired
+disease cohorts.
+- Main comparison: 5 donors (LoRA-5 vs scSemiProfiler with the same 5 donors as representatives), with zero-shot
+  shown for reference.
+- 5-fold CV is reported separately.
+- No new generation. CPU only.
+
+## Method (`tools/rare_paired.py`)
+
+- **Cohorts:** Ye 2026 sepsis, Imperial ESKD COVID, Giroux 2022, Ji 2025 ChILI
+  (`plans/paired_finetune_intradataset_20260927`).
+- **5-donor draws** j = 0..4, arms:
+  - `zs_raw` (`fewshot_n5/zs__raw_all.npz`);
+  - `zs_cal`, `lora5` (`draw<j>/lora__raw.npz`);
+  - `scsemi5` (`draw<j>/scsemi5__raw.npz`).
+- **5-fold, pooled:** `zs__raw`, `zs__cal`, `real_lora__raw` (`pooled/`), `scsemi__raw` (`pooled_scsemi_cv_20261002/`).
+- **NPZ handling:**
+  - NPZs are read on server_1, or streamed from spark-0110 (`ssh -n`).
+  - Each sha256 must equal the receipt.
+  - The benchmark's Poisson read-out (whole NPZ, seed 20260916) is reproduced cell by cell against the recorded
+    `umi_scoring_genes`, so the benchmark's own frozen scANVI labels apply.
+- **States:** the 15 literature marker rules of `plans/rare_states_screen_20261004/tools/states.json`.
+  - A marker rule is ≥ k marker genes detected in a parent type; a label rule uses AIFI L2.
+  - The negative control (platelet signal in monocytes) is left out of the figures.
+- **Depth:** for marker states, observed parent cells are binomially thinned (seed 20261004) to the generated parent
+  median library on the measured panel genes. There is one constant per unit, arm and state.
+- **Metrics** per unit, arm and state:
+  - pooled generated % and observed % within the parent;
+  - generated / observed ratio;
+  - Spearman over donors;
+  - detection AUROC (top vs bottom half of observed %);
+  - case-vs-control difference of donor means, observed vs generated, with its sign.
+- **Runner:** `tools/run_all.sh` runs 96 extractions and the analysis in one bounded user scope (MemoryMax 24G, 4
+  threads). Finished extractions are skipped.
+  - Fix of 2026-10-08: a remote `ssh` inside the job loop read the job list from stdin. The loop now reads on fd 3 and
+    `ssh -n` is used, and the run was resumed. All 97 jobs finished, 0 failures.
+
+**Outputs:** `out/summary.csv`, `out/contrasts.csv`, `out/per_donor.csv` (1,440 unit × arm × state rows).
+`work/<ds>/<unit>/<arm>.npz` holds marker counts per labelled cell, with a receipt (NPZ sha, host, reproduction).
+
+## Results
+
+Deck: `tools/fig5_rare_states.py`, Fig. 5a–c. There are 56 state × cohort pairs with ≥ 30 observed state cells in
+≥ 3 donors. A state a method never generates counts as that method's miss.
+
+| | 5 donors (main) | 5-fold (separate) |
+|---|---|---|
+| LoRA closer to the observed fraction than scSemiProfiler | 51/56 | — |
+| within 2-fold of observed: LoRA / scSemiProfiler | 19 / 5 | 25 / 9 |
+| better donor ranking (Spearman) than scSemiProfiler | 43/56 | 38/56 |
+| median Spearman: LoRA / scSemiProfiler | 0.24 / 0.18 | 0.32 / 0.24 |
+| disease direction kept (states × draws): LoRA / scSemiProfiler / zero-shot | 73% / 61% / 66% | — |
+
+**Caveat, stated on the slides:**
+- Zero-shot, real bulk, is closer than LoRA-5 on some fractions; the counts are in the 5a notes.
+- Most states are generated below their observed fraction (log2 ratio < 0).

paper-releases/pbisc-diffusion/updates/20261008T125804Z-deck-fig2-3-healthy-fig5-rare/diffusion/dataset/server_1/model8/plans/rare_states_paired_20261008/out/per_donor.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4db7f868ec59f7f3bfdd7cf84e18248bc26fbee7876063a35be3d952cc33d004
+size 12021378

paper-releases/pbisc-diffusion/updates/20261008T125804Z-deck-fig2-3-healthy-fig5-rare/diffusion/dataset/server_1/model8/plans/rare_states_paired_20261008/tools/run_all.sh

@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+# Every extraction (4 cohorts x (5 draws x 4 arms + 4 five-fold arms) = 96) and the analysis, one after another, in one
+# bounded transient user unit (CPU only: MemoryMax 24G, no swap, 4 threads, Nice 10). Finished extractions are skipped,
+# so a rerun resumes. Remote NPZs are streamed from spark-0110 and sha-checked by rare_paired.py.
+# usage: run_all.sh            (log: logs/run_all_<stamp>.log)
+set -uo pipefail
+P=/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/rare_states_paired_20261008
+PY=/home/sj_server_1/miniforge3/envs/cxg_corpus/bin/python
+LOG=$P/logs/run_all_$(date +%Y%m%dT%H%M%S).log
+avail=$(awk '/MemAvailable/ {print int($2/1048576)}' /proc/meminfo)
+(( avail - 32 > 24 )) || { echo "MemAvailable ${avail} GiB - 32 <= 24 GiB; not starting" >&2; exit 3; }
+JOBS=$P/logs/jobs_$(date +%Y%m%dT%H%M%S).txt
+: > "$JOBS"
+for ds in ye2026_sepsis imperial_eskd giroux2022 ji2025_chili; do
+  for j in 0 1 2 3 4; do for a in zs_raw zs_cal scsemi5 lora5; do echo "extract $ds fewpair $j $a" >> "$JOBS"; done; done
+  for a in zs__raw zs__cal scsemi__raw real_lora__raw; do echo "extract $ds fivefold - $a" >> "$JOBS"; done
+done
+echo "analyze" >> "$JOBS"
+systemd-run --user --scope --quiet -p MemoryMax=24G -p MemorySwapMax=0 -p CPUQuota=400% nice -n 10 \
+  env CUDA_VISIBLE_DEVICES= PYTHONDONTWRITEBYTECODE=1 OMP_NUM_THREADS=4 OPENBLAS_NUM_THREADS=4 \
+  bash -c 'n_fail=0; while IFS= read -r line <&3; do echo "JOB $(date +%FT%T) $line"; '"$PY"' '"$P"'/tools/rare_paired.py $line </dev/null \
+    || { n_fail=$((n_fail+1)); echo "FAIL $line"; }; done 3< '"$JOBS"'; echo "DONE failed=$n_fail jobs=$(wc -l < '"$JOBS"')"' >> "$LOG" 2>&1
+echo "log: $LOG"; tail -2 "$LOG"