Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261008T050202Z-scsemi-imperial-deck/diffusion/dataset/server_1/model8/plans/bayesprism_paired_20261006/README.md

@@ -0,0 +1,127 @@
+# Bulk deconvolution with the same paired donors (BayesPrism, CIBERSORTx) vs PBISC LoRA
+
+Started 2026-10-06; results final 2026-10-08. CPU only, no GPU.
+
+User requests:
+- 2026-10-06: compare BayesPrism; then add CIBERSORTx and the paired COVID cohort (Imperial). The paired COVID cohort
+  already in the benchmark was judged enough.
+- 2026-10-06 user decision: deconvolution methods go to the paper's Supplementary only, because PBISC generates single
+  cells, a different task.
+
+Deck: two Supplementary slides in `presentations/2026-09-30--paper-summary--r00`, made by `tools/secS_deconvolution.py`:
+"Deconvolution with the same pairs: composition" and "...: disease signal".
+
+## Question and design
+
+Can standard bulk deconvolution, given the same paired donors that PBISC LoRA fine-tunes on, recover what PBISC
+recovers on the paired benchmark (`plans/paired_finetune_intradataset_20260927`)?
+
+Everything except the method is the benchmark's own:
+- **Splits** (`tools/common.py split_rows`)
+  - 5-donor draws: reference = the draw's 5 training donors.
+  - 5-fold CV: reference = each fold's LoRA training plus inner-validation rows (from its RUN.json).
+  - Mixture = the test samples' real bulk raw counts.
+  - Only paired samples are used.
+- **Truth and scorer**: score_ft.py functions imported unchanged, on the same donors, observed cells, case/control
+  groups and `--exclude-genes` (Giroux: sex-linked genes excluded, from `scores_nosex/`).
+- **Gates** (fail closed):
+  - The observed composition equals the stored per_donor values.
+  - The recomputed bulk-broadcast logFC r and top-50 DEG overlap equal the stored values for every split.
+  - Older pooled scorings have no DEG rows; for those, only the r gate applies.
+- **Cohorts**: ImMicro (healthy; composition only), sepsis (Ye 2026), Ji ChILI, Imperial ESKD COVID, Giroux.
+
+### Methods
+
+**BayesPrism** (`tools/run_bp.R`; R library `utility_studies_final6k_20260925/deconv/Rlib`):
+- Settings:
+  - input.type GEP.
+  - States = "six-type gate | AIFI L2", with states under 20 cells merged into "type|minor".
+  - outlier.cut 0.01, outlier.fraction 0.1.
+- Gene modes:
+  - `full` (main): BayesPrism's recommended set, canonical protein-coding minus chrX/Y/M and RPS/RPL/MRPS/MRPL,
+    measured by the sc assay; 15-19k genes.
+  - `score` (sensitivity): the scorer's ~5.5-6k scoring genes. This handicaps composition.
+- Outputs: θ initial ("first", Gibbs) and θ updated ("final", after the reference update), plus Z (posterior
+  cell-type expression).
+- θ is in mRNA units. It is converted to cell units by dividing by each type's mean UMI per reference cell, then
+  renormalising.
+
+**CIBERSORTx** (`tools/build_cibersortx_inputs.py`, `tools/run_cibersortx.sh`):
+- Runs from the official Docker images `cibersortx/fractions` and `cibersortx/hires`.
+- The images are amd64-only. On server_1 (aarch64) they run under qemu binfmt emulation, registered with
+  `tonistiigi/binfmt` (user-approved 2026-10-06; undo with `--uninstall amd64`).
+- **Fractions** (single-cell reference mode with S-mode batch correction):
+  - Reference: up to 400 cells per type, seeded.
+  - Flags: `--fraction 0 --replicates 5 --sampling 0.5 --G.min 300 --G.max 500 --q.value 0.01 --QN FALSE`.
+- **HiRes**:
+  - Runs on the scoring genes, with the Fractions signature and S-mode source GEPs.
+  - HiRes regresses each gene on the 6 type fractions, so splits with 8 or fewer test samples are skipped. That
+    affects the Ji and Giroux CV folds.
+  - HiRes imputes a different gene subset per type and writes exactly 1 in every sample for the other genes. The
+    scorer treats those rows as not expressed (zero) and not covered (`score_bp.py load_cx`).
+  - In 3 of 20 disease draws (Ji draw 3, Giroux draws 1 and 4), HiRes imputed no gene of the four contrast types.
+    Those draws have no CIBERSORTx disease score.
+- **Credentials**:
+  - Read at run time from `/home/sj_server_1/cibersortx.key` into a mode-600 env file that only the container reads.
+  - The binary echoes its options, so its output is redacted before logging.
+  - Checked 2026-10-08: no token or username in any plan file (counts only).
+
+## Results (main: BayesPrism gene mode full)
+
+Full tables:
+- `results/RESULTS_full.md` (BayesPrism, main)
+- `results/RESULTS_score.md` (BayesPrism, score-gene sensitivity)
+- `results/RESULTS_cibersortx.md`
+- Machine-readable versions: `results/TABLE_{fewshot,cv}_{full,score,cibersortx}.csv`
+- Per split: `results/<cohort>/<comparison>/{summary.json, per_donor.csv, contrast.csv}`
+
+Five-donor draws (25 draws; 20 with a case/control contrast). Number of draws where LoRA-5 beats the method:
+
+| | composition L1 | logFC r | partial r beyond bulk | top-50 DEG |
+|---|---|---|---|---|
+| BayesPrism (better θ) | 22/25 | 15/20 | 18/20 | 6/20 (BayesPrism ahead 8, ties 6) |
+| CIBERSORTx | 24/25 | 17/17 | 17/17 | 6/17 (many ties at 0) |
+
+- **Composition.** Deconvolution is far off in most cohorts. For example, in Giroux BayesPrism's L1 is 1.46 vs
+  LoRA-5's 0.26, and CIBERSORTx's is 0.67.
+- **Pipeline check.** On sc pseudobulk mixtures, BayesPrism recovers the observed mRNA shares (Giroux T 0.386 vs
+  0.375; ImMicro T 0.54 vs 0.60). The error therefore comes from real bulk, not from the pipeline.
+- **Disease signal.** BayesPrism Z is ahead of LoRA-5 on logFC r only in Giroux, where the plain bulk broadcast is
+  best of all. CIBERSORTx HiRes is below the broadcast everywhere.
+- **DEG.** Both deconvolutions are ahead on top-50 DEG in sepsis (27 and 25 vs 16); LoRA-5 is ahead in Imperial (114
+  vs 72 and 31).
+- **5-fold CV, pooled.**
+  - LoRA has the lowest composition L1 in ImMicro, sepsis, Ji and Imperial.
+  - In Giroux, scSemiProfiler is slightly lower (0.243 vs 0.258).
+  - scSemiProfiler's 5-fold column comes from `scores[_nosex]/<ds>/scsemi_cv_20261002/pooled` for all five cohorts.
+    Imperial finished on server_2 on 2026-10-08 13:45 and was finalized on server_1 from sha-checked fold copies.
+    In Imperial, LoRA leads on every metric (L1 0.376 vs 0.520, logFC r 0.819 vs 0.705, DEG 143 vs 87).
+
+## Rebuild
+
+```bash
+P=/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/bayesprism_paired_20261006
+PY=/home/sj_server_1/miniforge3/envs/cxg_corpus/bin/python
+# per cohort: aggregates and split inputs (CPU)
+CUDA_VISIBLE_DEVICES= $PY $P/tools/build_aggregates.py DATASET full
+CUDA_VISIBLE_DEVICES= $PY $P/tools/build_split.py DATASET {fewshot|cv} IDX full
+GENES=full bash $P/tools/run_all.sh LANE CORES DATASET ...        # BayesPrism, bounded units
+CUDA_VISIBLE_DEVICES= $PY $P/tools/build_cibersortx_inputs.py DATASET
+bash $P/tools/run_cibersortx.sh LANE CORES DATASET ...            # CIBERSORTx (qemu on aarch64; hours per split)
+bash $P/tools/finalize_full.sh                                     # score + tables, BayesPrism (full and score modes)
+RESCORE_FULL_CV="cohorts with new scSemiProfiler CV" bash $P/tools/finalize_cx.sh   # CIBERSORTx (+ refresh CV)
+```
+
+`remote_cache/` holds sha-checked copies of inputs that live only on spark-0110: condition NPZs and the Imperial
+few-shot RUN.json (`common.local_copy`).
+
+## Changes
+
+- 2026-10-08:
+  - `score_bp.py`:
+    - HiRes placeholder rows (all exactly 1) are no longer read as expression.
+    - The covered variant is per type for CIBERSORTx.
+    - Giroux's scSemiProfiler CV is read from `scores_nosex/`, matching LoRA's gene exclusion.
+    - The BayesPrism metrics were rechecked identical after the change.
+  - `aggregate_bp.py`: labels its table by method.
+  - New: `finalize_cx.sh`.

paper-releases/pbisc-diffusion/updates/20261008T050202Z-scsemi-imperial-deck/diffusion/dataset/server_1/model8/plans/bayesprism_paired_20261006/results/TABLE_cv_cibersortx.csv

@@ -0,0 +1,6 @@
+cohort,bp_l1_final,bp_l1_first,lora_l1,bp_r,bp_r_covered,lora_r,broadcast_r,broadcast_r_covered,bp_partial,lora_partial,bp_deg,lora_deg,broadcast_deg,scsemi_l1,scsemi_r,scsemi_deg
+immicro,0.5065033643252109,0.5065033643252109,0.16172890350786212,,,,,,,,,,,0.2170087237986556,,
+ye2026_sepsis,0.607149641152929,0.607149641152929,0.3618246783833512,-0.035832909681658,-0.12522813744518627,0.5267612711229187,0.31594581626202867,0.601490367831679,-0.0739317360248297,0.46874787884206276,0.0,23.0,21.0,0.5022299158951089,0.43526410070392063,6.0
+ji2025_chili,0.914073829946989,0.914073829946989,0.5044992874562503,,,,,,,,,,,0.575388570715078,0.2730887293849816,0.0
+imperial_eskd,0.5918271288164071,0.5918271288164071,0.37638560875745664,0.21113497684457377,0.3693097049105145,0.8187448084273437,0.6606053132085035,0.5831498842343769,0.015667623661280307,0.6899314750539278,1.0,143.0,84.0,0.5203155086290455,0.7053598964433999,87.0
+giroux2022,0.5914100321768722,0.5914100321768722,0.2580442754044374,,,,,,,,,,,0.24307690277174857,0.5372927269740906,0.0

paper-releases/pbisc-diffusion/updates/20261008T050202Z-scsemi-imperial-deck/diffusion/dataset/server_1/model8/plans/bayesprism_paired_20261006/results/TABLE_cv_full.csv

@@ -0,0 +1,6 @@
+cohort,bp_l1_final,bp_l1_first,lora_l1,bp_r,bp_r_covered,lora_r,broadcast_r,broadcast_r_covered,bp_partial,lora_partial,bp_deg,lora_deg,broadcast_deg,scsemi_l1,scsemi_r,scsemi_deg
+immicro,0.5745628784374615,0.9085350102150449,0.16172890350786212,,,,,,,,,,,0.2170087237986556,,
+ye2026_sepsis,0.6029270442284831,0.4750523423944086,0.3618246783833512,0.2554670801971232,0.2533386702661269,0.5267612711229187,0.31594581626202867,0.2531577574908379,0.09471562043764327,0.46874787884206276,37.0,23.0,21.0,0.5022299158951089,0.43526410070392063,6.0
+ji2025_chili,0.8832571878464304,1.054524674985768,0.5044992874562503,0.1857896942054881,0.20352470183066124,0.47928798340225065,0.2398746224459366,0.21227266152493962,0.02042361532730819,0.45204372964017614,4.0,15.0,4.0,0.575388570715078,0.2730887293849816,0.0
+imperial_eskd,0.8870870830413495,1.0215635566013643,0.37638560875745664,0.5684615839866303,0.5843506673991777,0.8187448084273437,0.6606053132085035,0.6388431994811914,0.11228628082302743,0.6899314750539278,91.0,143.0,84.0,0.5203155086290455,0.7053598964433999,87.0
+giroux2022,1.8027760246078957,0.8778647964495544,0.2580442754044374,0.40594570910829286,0.40810658821060636,0.46094848304079133,0.5752659929360795,0.5750505878963892,0.02157078325029122,0.3230347269696575,3.0,5.0,5.0,0.24307690277174857,0.5372927269740906,0.0

paper-releases/pbisc-diffusion/updates/20261008T050202Z-scsemi-imperial-deck/diffusion/dataset/server_1/model8/plans/bayesprism_paired_20261006/tools/build_archive_request.py

@@ -0,0 +1,157 @@
+#!/usr/bin/env python3
+"""Archive-maintenance request (paper-release-20260928/maintenance/POLICY.md, project pbisc-diffusion) for the
+2026-10-08 deconvolution + scSemiProfiler update. CPU only; reads files, writes the request and one manifest.
+
+Payload (server_1):
+  this plan     tools, README, results/ (tables, summaries, RESULTS*.md), logs (credential-free; checked by count),
+                small inputs metadata (SPLIT.json, state_types.tsv, AGGREGATES*.json, INPUT.json, hires_genes.txt),
+                BayesPrism theta + genes_used (both gene modes), CIBERSORTx fractions / signature / phenoclasses /
+                run logs
+  paired FT     new scSemiProfiler score units (Giroux and sepsis 5-fold, sepsis equal budget, Giroux without
+                sex-linked genes, contrast and realism extras) and today's finalize / scoring logs
+  deck          every deck file; unchanged ones are skipped by update.py (same sha already archived)
+Large outputs kept in place, listed with bytes + sha256 in results/LARGE_OUTPUTS_SHA256.tsv (archived): BayesPrism Z
+(full mode), CIBERSORTx HiRes, source GEPs, S-mode adjusted mixtures, inferred reference copies, per-state aggregates.
+Inference locations: the pooled scSemiProfiler NPZs (receipt = POOL_MANIFEST.json); fold NPZs are registered already.
+usage: build_archive_request.py OUT.json
+"""
+from __future__ import annotations
+
+import glob
+import hashlib
+import json
+import sys
+from pathlib import Path
+
+P = Path(__file__).resolve().parents[1]
+FT = Path("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/paired_finetune_intradataset_20260927")
+DECK = Path("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/"
+            "2026-09-30--paper-summary--r00")
+LARGE_CX = ("CIBERSORTx_refsample_inferred_refsample.txt", "CIBERSORTx_Mixtures_Adjusted.txt",
+            "CIBERSORTx_cell_type_sourceGEP.txt")
+
+
+def sha(p: Path) -> str:
+    h = hashlib.sha256()
+    with open(p, "rb") as f:
+        for b in iter(lambda: f.read(1 << 22), b""):
+            h.update(b)
+    return h.hexdigest()
+
+
+def files(root: Path, patterns, exclude=lambda p: False):
+    out = set()
+    for pat in patterns:
+        for f in glob.glob(str(root / pat), recursive=True):
+            q = Path(f)
+            if q.is_file() and not q.is_symlink() and "__pycache__" not in q.parts and not exclude(q):
+                out.add(q)
+    return sorted(out)
+
+
+def main(out: str) -> int:
+    assets = []
+
+    def add(paths, kind):
+        for p in paths:
+            assets.append({"host": "server_1", "source": str(p), "kind": kind, "state": "completed"})
+
+    # ---- large outputs: manifest only
+    large = (files(P, ["inputs/*/*_full/bp/Z_type_*.tsv.gz", "cibersortx/*/*/out_hires/CIBERSORTxHiRes_*.txt",
+                       "data/*/aggregates*.npz"])
+             + [q for n in LARGE_CX for q in files(P, [f"cibersortx/*/*/out_fractions/{n}"])])
+    man = P / "results" / "LARGE_OUTPUTS_SHA256.tsv"
+    man.write_text("path\tbytes\tsha256\n" + "".join(f"{q.relative_to(P)}\t{q.stat().st_size}\t{sha(q)}\n" for q in large))
+    # ---- this plan
+    add(files(P, ["tools/*.py", "tools/*.sh", "tools/*.R"]), "code")
+    add(files(P, ["README.md"]), "document")
+    add(files(P, ["results/**/*.csv", "results/**/*.json", "results/*.tsv"]), "table")
+    add(files(P, ["results/*.md"]), "report")
+    add(files(P, ["logs/*.log"]), "log")
+    add(files(P, ["inputs/*/*/SPLIT.json", "inputs/*/*/state_types.tsv", "data/*/AGGREGATES*.json",
+                  "cibersortx/*/*/INPUT.json", "cibersortx/*/*/hires_genes.txt"]), "metadata")
+    add(files(P, ["inputs/*/*/bp/theta_*.csv", "inputs/*/*/bp/genes_used.txt"]), "data")
+    add(files(P, ["cibersortx/*/*/out_fractions/CIBERSORTx_Adjusted.txt",
+                  "cibersortx/*/*/out_fractions/CIBERSORTx_sigmatrix_Adjusted.txt",
+                  "cibersortx/*/*/out_fractions/CIBERSORTx_refsample_inferred_phenoclasses*.txt",
+                  "cibersortx/*/*/out_hires/fractions_results.txt"]), "data")
+    add(files(P, ["cibersortx/*/*/out_fractions/.run.log", "cibersortx/*/*/hires_run.log"]), "log")
+    # ---- paired fine-tuning plan: new scSemiProfiler units
+    units = ["scores/giroux2022/scsemi_cv_20261002", "scores_nosex/giroux2022/scsemi_cv_20261002",
+             "scores/ye2026_sepsis/scsemi_cv_20261002", "scores/ye2026_sepsis/scsemi_budget_20261002",
+             "scores/imperial_eskd/scsemi_cv_20261002"] + \
+            [f"scores_contrast/{d}/scsemi_cv_20261002" for d in ("ye2026_sepsis", "giroux2022", "imperial_eskd")] + \
+            [f"scores_extra/{d}/scsemi_cv_20261002" for d in ("immicro", "ye2026_sepsis", "ji2025_chili", "giroux2022",
+                                                              "imperial_eskd")]
+    for u in units:
+        fs = files(FT, [f"{u}/**/*"])
+        add([q for q in fs if q.suffix in (".csv", ".json", ".tsv", ".parquet")], "table")
+        add([q for q in fs if q.suffix == ".md"], "report")
+    add(files(FT, ["logs/finalize_scsemi_*server1_20261008*.log", "logs/score-giroux2022-scsemi_cv_20261002-*_20261008*.log",
+                   "logs/score-ye2026_sepsis-scsemi_budget_20261002-*_20261008*.log", "logs/pool-*_20261008*.log",
+                   "logs/scores_nosex-giroux2022-scsemi_cv-pooled_20261008*.log", "logs/extra-scsemi-cv_20261008*.log",
+                   "logs/finalize_scsemi_cv_imperial_server1_20261008*.log",
+                   "logs/score-imperial_eskd-scsemi_cv_20261002-*_20261008*.log",
+                   "logs/extra-imperial-scsemi-cv_20261008*.log",
+                   "scores_contrast/_logs/cx-sepsis-giroux-scsemi-cv_20261008*.log",
+                   "scores_contrast/_logs/cx-imperial-scsemi-cv_20261008*.log"]), "log")
+    # ---- deck
+    add(files(DECK, ["*.pptx", "*.pdf"]), "report")
+    add(files(DECK, ["README.md"]), "document")
+    add(files(DECK, ["generate_deck.py", "tools/*.py"]), "code")
+    add(files(DECK, ["figures/*.png", "figures/*.pdf", "figures/*.svg"]), "figure")
+    add(files(DECK, ["data/*.csv"]), "table")
+    add(files(DECK, ["data/*.json"]), "config")
+    # ---- inference locations: pooled scSemiProfiler NPZs
+    locs = []
+    for ds, sub in (("giroux2022", "pooled_scsemi_cv_20261002"), ("ye2026_sepsis", "pooled_scsemi_cv_20261002"),
+                    ("ye2026_sepsis", "pooled_scsemi_budget_20261002"), ("imperial_eskd", "pooled_scsemi_cv_20261002")):
+        d = FT / "generated" / ds / sub
+        m = json.loads((d / "POOL_MANIFEST.json").read_text())
+        for arm, a in m["arms"].items():
+            got = sha(Path(a["path"]))
+            if got != a["sha256"]:
+                raise SystemExit(f"{a['path']}: sha differs from POOL_MANIFEST.json")
+            locs.append({"artifact_role": "inference_output", "cohort": ds, "experiment": f"scsemi_{sub[7:]}",
+                         "host": "server_1", "path": a["path"], "receipt": str(d / "POOL_MANIFEST.json"),
+                         "sha256": got, "output_sha256": got, "state": "completed",
+                         "note": f"scSemiProfiler {arm}, five test folds pooled (pool_folds.py); fold NPZs registered earlier"})
+    # fold NPZs: the server_1 copies used by the finalize (sha-checked against the server_2 receipts), and Imperial
+    # fold 4 on server_2, finished 2026-10-08 13:45 KST
+    for ds in ("giroux2022", "imperial_eskd"):
+        for k in range(5):
+            npz = FT / "generated" / ds / f"fold{k}" / "scsemi__raw.npz"
+            rec = npz.with_suffix(".log")
+            want = json.loads(rec.read_text())["output_sha256"]
+            if sha(npz) != want:
+                raise SystemExit(f"{npz}: sha differs from its receipt")
+            locs.append({"artifact_role": "inference_output", "cohort": ds, "experiment": "scsemi_cv_20261002",
+                         "host": "server_1", "path": str(npz), "receipt": str(rec), "sha256": want,
+                         "output_sha256": want, "state": "completed",
+                         "note": "sha-checked copy of the spark-0110 original, used for the server_1 finalize"})
+            if ds == "imperial_eskd" and k == 4:
+                locs.append({"artifact_role": "inference_output", "cohort": ds, "experiment": "scsemi_cv_20261002",
+                             "host": "server_2", "path": str(npz), "receipt": str(rec), "sha256": want,
+                             "output_sha256": want, "state": "completed",
+                             "note": "original; run finished 2026-10-08 13:45 KST (queue scsemi-queue-cv-gi-s2, success)"})
+    req = {"project_id": "pbisc-diffusion",
+           "summary": ("2026-10-08: bulk deconvolution with the same paired donors (BayesPrism recommended genes + score-gene "
+                       "sensitivity; official CIBERSORTx Fractions S-mode + HiRes under qemu on server_1) scored with the "
+                       "paired benchmark's own truth and scorer (plans/bayesprism_paired_20261006, README + RESULTS*.md); "
+                       "new scSemiProfiler results (5-fold CV of Giroux, sepsis and Imperial, finalized on server_1 from "
+                       "sha-checked fold copies; Giroux without sex-linked genes; sepsis equal budget; contrast and realism "
+                       "extras), so scSemiProfiler 5-fold now covers all five cohorts; paper-summary deck updated (two "
+                       "Supplementary slides, Fig. 3 scSemiProfiler slides refreshed; 64 slides). Large deconvolution "
+                       "outputs stay on server_1, listed with sha256 in results/LARGE_OUTPUTS_SHA256.tsv."),
+           "assets": assets, "inference_locations": locs}
+    Path(out).write_text(json.dumps(req, indent=1))
+    kinds = {}
+    for a in assets:
+        kinds[a["kind"]] = kinds.get(a["kind"], 0) + 1
+    print(json.dumps({"assets": len(assets), "by_kind": kinds, "locations": len(locs), "large_outputs": len(large),
+                      "large_bytes": sum(q.stat().st_size for q in large)}, indent=1))
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main(sys.argv[1]))