v7.1.3 correction patch: corrected labels, regenerated v4 headline table, scope and averaging statements, epoch disclosure, RR-7 row, MAQC 2026 abstract erratum

Correction patch over canonical v7.1 results; no new benchmark result generation. Mission labels: A2 third mission is SpaceX-10/RR-4 quadriceps (OSD-326); A6 has two mission-held-out folds (OSD-397 belongs to RR-1). The v4 headline table is regenerated from v4/evaluation/M1_summary.json; sample scope and averaging rules are stated; fine-tuned FM values are disclosed as best-of-10-epoch selections on the held-out test mission; the skin RR-7 open-validation row is a copy of the LOMO fold and its recorded 0.885 does not reproduce (converged refit 0.805). The card and RESULTS_SUMMARY.md carry the erratum for the MAQC 2026 abstract numbers. Source commit on GitHub (private): integrate/2026-09 6100b97.

README.md

@@ -20,18 +20,19 @@ language:
   - en
 size_categories:
   - 1K<n<10K
-pretty_name: "SpaceBio-Bench / GeneLab Benchmark v7.1.2 Public Fold Package"
+pretty_name: "SpaceBio-Bench / GeneLab Benchmark v7.1.3 Public Fold Package"
 viewer: false
 ---
 
-# SpaceBio-Bench / GeneLab Benchmark v7.1.2 Public Fold Package
+# SpaceBio-Bench / GeneLab Benchmark v7.1.3 Public Fold Package
 
-> **Correction in progress (September 2026).** Two labelling errors in the released benchmark are being corrected.
+> **Corrections (September 2026; applied in v7.1.3, 2026-09-27).** Two labelling errors in the released benchmark are corrected.
 > (1) The gastrocnemius task (A2) lists a third mission as "RR-9"; those eight samples (OSD-326) come from the
 > SpaceX-10 mission (Rodent Research-4) and are quadriceps femoris, not gastrocnemius. RR-9 contributes liver and
 > thymus only. (2) The eye task (A6) evaluated the OSD-397 samples as a separate mission; they belong to RR-1 and
-> share animals with OSD-100, so the task has two mission-held-out folds, not three. Tables that report per-mission
-> results for these two tasks will be reissued with the corrected labels. The source code repository is private
+> share animals with OSD-100, so the task has two mission-held-out folds, not three. This card and `RESULTS_SUMMARY.md`
+> use the corrected labels; fold directory names keep the historical labels (`fold_RR-9_test`, `fold_OSD-397_test`).
+> See the changelog below for the other v7.1.3 corrections and the MAQC 2026 abstract erratum. The source code repository is private
 > while the corrected release is prepared; code and fold definitions are available on request (contact:
 > JangKeun Kim). Every other task and the mission-held-out design are unaffected.
 
@@ -40,12 +41,11 @@ viewer: false
 machine-learning and foundation-model methods generalize spaceflight biological
 signatures across missions.**
 
-Public status: **v7.1.2 public-card/metadata patch over canonical v7.1 results**
+Public status: **v7.1.3 correction patch over canonical v7.1 results** (2026-09-27; earlier patch: v7.1.2 public-card/metadata patch, 2026-06-16)
 
 Dataset freeze: **2026-03-01**
 
-Patch scope: documentation, public metadata, and access guidance. It does not
-introduce new benchmark result generation.
+Patch scope: documentation corrections over canonical v7.1 results (mission labels for OSD-326 and OSD-397, the v4 headline table regenerated from its result JSON, sample scope and averaging rules, fine-tuning epoch selection, the skin RR-7 open-validation row, fGSEA direction statements) and the MAQC 2026 abstract erratum. It does not introduce new benchmark result generation.
 
 Code and full documentation: <https://github.com/jang1563/GeneLab_benchmark>
 
@@ -94,7 +94,7 @@ genelab-benchmark/
 │   ├── fold_MHU-2_test/
 │   ├── fold_RR-6_test/
 │   ├── fold_RR-7_test/
-│   └── fold_RR-7_holdout/       # retrospective open validation; labels public
+│   └── fold_RR-7_holdout/       # identical copy of fold_RR-7_test; labels public
 ├── A6_eye_lomo/
 │   ├── task_info.json
 │   ├── fold_RR-1_test/
@@ -105,10 +105,12 @@ genelab-benchmark/
 └── v6/evaluation/
 ```
 
-`fold_OSD-397_test` is the stable public label for the third A6 eye fold.
+`fold_OSD-397_test` is the third A6 fold directory. Its samples belong to RR-1 (corrected 2026-09-25; they share animals with OSD-100), so A6 has two mission-held-out folds.
 
 The two historical `_holdout` directories also contain public `test_y.csv`
-files. They support retrospective reproducibility, not blind evaluation. See
+files. They support retrospective reproducibility, not blind evaluation.
+`fold_RR-7_holdout` repeats `fold_RR-7_test` exactly; only `fold_RR-23_holdout`
+holds a mission outside the LOMO folds. See
 the [benchmark integrity note](https://github.com/jang1563/GeneLab_benchmark/blob/main/docs/BENCHMARK_INTEGRITY.md).
 
 ## Scope
@@ -117,7 +119,7 @@ the [benchmark integrity note](https://github.com/jang1563/GeneLab_benchmark/blo
 |---|---|
 | Full v1-v7 benchmark surface | 8 tissues |
 | Public source catalog | 24+ NASA OSDR accessions |
-| Processed sample scope | 600+ binary/control samples across release layers |
+| Processed sample scope | 549 profiles in the v1 LOMO tasks A1–A6 (22 folds; 565 with the RR-23 open-validation fold) and 792 in the v4 8-tissue evaluation; counts are RNA-seq profiles, not animals |
 | v4 multi-method evaluation | 8 tissues x 8 classifiers x 4 feature types = 256 evaluations |
 | Public HF fold package | 4 reviewer-facing LOMO tasks plus selected result artifacts |
 
@@ -188,12 +190,12 @@ repository.
 
 | Result surface | Takeaway |
 |---|---|
-| Multi-method benchmark | PCA-LR is the strongest 8-tissue gene-level baseline in v4, with mean AUROC 0.776. |
-| Best tissue rows | Thymus 0.948, colon 0.921, lung 0.901, kidney 0.829 across best method-feature combinations. |
+| Multi-method benchmark | PCA-LR is the strongest gene-level baseline in v4: mean AUROC 0.776 over 8 tissues (0.753 over the 6 LOMO tissues). |
+| Best tissue rows | Highest of 32 method-feature rows per tissue, selected after the fact: LOMO thymus 0.948, gastrocnemius 0.898, kidney 0.829, eye 0.823, skin 0.819, liver 0.766; single-mission colon 0.921 and lung 0.901 use 5-fold CV, not LOMO. |
 | Cross-mission transfer | Thymus and gastrocnemius show the strongest mission-transfer signal; liver and kidney are harder. |
 | Pathway features | Pathway representations rescue some weaker gene-level tissues, especially kidney and eye. |
-| Foundation models | Tested gene-expression foundation models underperform tuned classical baselines on small-n bulk RNA-seq mission shift. |
-| Retrospective open validation | Thymus RR-23 AUROC 0.905; skin RR-7 AUROC 0.885; labels are public. |
+| Foundation models | Tested gene-expression foundation models underperform tuned classical baselines on small-n bulk RNA-seq mission shift. Fine-tuned scGPT and Mouse-Geneformer fold AUROCs keep the best of 10 epochs scored on the held-out test mission itself (no inner validation split; `scripts/scgpt_finetune.py`, `scripts/geneformer_finetune.py`), so they are optimistic. |
+| Retrospective open validation | Thymus RR-23 AUROC 0.905 (n=16); labels are public. The skin RR-7 split is a copy of the LOMO RR-7 fold; its recorded 0.885 does not reproduce (a converged refit gives the LOMO value 0.805), so it is not additional evidence. |
 
 ## Intended Use
 
@@ -212,10 +214,26 @@ manifest.
 | Surface | Public label |
 |---|---|
 | v7.1 GeneLab Benchmark | Canonical historical result surface and citation target |
-| v7.1.2 public-card patch | Documentation and metadata patch over v7.1 results |
+| v7.1.2 public-card patch | Documentation and metadata patch over v7.1 results (2026-06-16) |
+| v7.1.3 correction patch | Label, headline-table and documentation corrections over v7.1 results; MAQC 2026 abstract erratum (2026-09-27) |
 
-This HF dataset card describes the v7.1 public fold package with the v7.1.2
-public-card patch.
+This HF dataset card describes the v7.1 public fold package with the v7.1.3
+correction patch.
+
+## Changelog
+
+**v7.1.3 (2026-09-27), correction patch.**
+
+- Mission labels: the A2 fold archived as RR-9 is the SpaceX-10 / RR-4 quadriceps study (OSD-326); the A6 OSD-397 samples belong to RR-1, so A6 has two mission-held-out folds.
+- v4 headline table regenerated from `v4/evaluation/M1_summary.json`, with a fixed PCA-LR column, a post hoc best-row column and the evaluation scheme (colon and lung use 5-fold stratified CV); the previous table held two values that are not v4 rows and three wrong method labels.
+- Sample scope stated as 549 profiles in 22 LOMO folds for the v1 tasks A1–A6 (565 with the RR-23 fold; 792 in the v4 evaluation); the six-tissue classical reference mean (0.758) is a fixed per-tissue model, not PCA-LR.
+- Fine-tuned scGPT and Mouse-Geneformer values disclosed as best-of-10-epoch selections on the held-out test mission.
+- Skin `fold_RR-7_holdout` documented as a copy of the LOMO RR-7 fold; its recorded logistic-regression AUROC (0.885) does not reproduce, and a converged refit gives 0.805, the LOMO value.
+- Pathway-direction statements corrected to the flight-positive fGSEA sign convention (files regenerated on 2026-08-14).
+
+**Erratum (MAQC 2026 abstract).** The abstract "From Reproducible Pipelines to Reproducible Claims: An Audit of Model Selection and Aggregation in Spaceflight Omics" (Kim and Mason, submitted to MAQC 2026 on 19 August 2026) reported results on the six-tissue task surface of this dataset as released at the time (549 profiles, 22 leave-one-mission-out task folds): fixed PCA-logistic regression six-tissue macro AUROC 0.730; scGPT 0.666 with the best held-out test epoch per fold and 0.599 at a fixed epoch 10; Mouse-Geneformer 0.476 and 0.458; thymus PCA-logistic regression 0.923 as the mean of mission-level AUROCs and 0.631 pooled out of fold. As the abstract stated, the fixed-epoch values are a post hoc sensitivity analysis, not nested epoch selection. The values were computed correctly from the files released at the time, but three problems found afterwards affect them: (1) the scGPT and Mouse-Geneformer inputs were z-scored expression values, while both tokenizers expect raw counts; (2) the gastrocnemius task's third mission is the SpaceX-10 / Rodent Research-4 quadriceps study (OSD-326), not RR-9 gastrocnemius, and the eye task has two mission-held-out folds, not three, because the OSD-397 samples belong to RR-1 and share animals with OSD-100; (3) three MHU-2 thymus flight profiles carry swapped condition labels: the original BioSample records show that the profiles archived as `MHU2_FLT_1G_Rep1-3` are microgravity and the `uG`-named profiles are artificial gravity. Read the abstract's numbers as results on the pre-correction surface; they are not comparable with later results. The reanalysis (raw-count inputs, nested epoch selection, corrected labels, a recorded evaluation contract) is reported separately.
+
+**v7.1.2 (2026-06-16).** Public-card, citation, and metadata patch over canonical v7.1 results.
 
 ## Citation
 
@@ -227,7 +245,7 @@ Please cite the software and benchmark using the GitHub `CITATION.cff` metadata.
   author = {Kim, JangKeun},
   year = {2026},
   url = {https://huggingface.co/datasets/jang1563/genelab-benchmark},
-  note = {v7.1.2 documentation, public-card, and metadata patch over canonical v7.1 results; data freeze 2026-03-01}
+  note = {v7.1.3 correction patch over canonical v7.1 results; data freeze 2026-03-01}
 }
 ```
 

RESULTS_SUMMARY.md

@@ -2,9 +2,11 @@
 
 > **Correction notice (September 2026).** The gastrocnemius (A2) mission listed as RR-9 is the SpaceX-10 / Rodent Research-4
 > quadriceps study (OSD-326), and the eye task (A6) has two mission-held-out folds because the OSD-397 samples belong to RR-1;
-> per-mission tables for A2 and A6 will be reissued with corrected labels. See the correction notice on the dataset card.
+> this summary uses the corrected labels (fold directory names keep the historical ones). See the correction notice on the dataset card.
+>
+> **Erratum (MAQC 2026 abstract).** The abstract "From Reproducible Pipelines to Reproducible Claims: An Audit of Model Selection and Aggregation in Spaceflight Omics" (Kim and Mason, submitted to MAQC 2026 on 19 August 2026) reported results on the six-tissue task surface of this dataset as released at the time (549 profiles, 22 leave-one-mission-out task folds): fixed PCA-logistic regression six-tissue macro AUROC 0.730; scGPT 0.666 with the best held-out test epoch per fold and 0.599 at a fixed epoch 10; Mouse-Geneformer 0.476 and 0.458; thymus PCA-logistic regression 0.923 as the mean of mission-level AUROCs and 0.631 pooled out of fold. As the abstract stated, the fixed-epoch values are a post hoc sensitivity analysis, not nested epoch selection. The values were computed correctly from the files released at the time, but three problems found afterwards affect them: (1) the scGPT and Mouse-Geneformer inputs were z-scored expression values, while both tokenizers expect raw counts; (2) the gastrocnemius task's third mission is the SpaceX-10 / Rodent Research-4 quadriceps study (OSD-326), not RR-9 gastrocnemius, and the eye task has two mission-held-out folds, not three, because the OSD-397 samples belong to RR-1 and share animals with OSD-100; (3) three MHU-2 thymus flight profiles carry swapped condition labels: the original BioSample records show that the profiles archived as `MHU2_FLT_1G_Rep1-3` are microgravity and the `uG`-named profiles are artificial gravity. Read the abstract's numbers as results on the pre-correction surface; they are not comparable with later results. The reanalysis (raw-count inputs, nested epoch selection, corrected labels, a recorded evaluation contract) is reported separately.
 
-Generated: 2026-03-01 (Updated: 2026-03-29 — v5 biological interpretation added; numeric corrections 2026-09-10). ⚠️ This summary covers results through v5; the repository card is versioned separately as a v7.1.2 card/metadata patch over canonical v7.1 results.
+Generated: 2026-03-01 (Updated: 2026-03-29 — v5 biological interpretation added; numeric corrections 2026-09-10; v4 table regenerated from `v4/evaluation/M1_summary.json` and fGSEA directions corrected 2026-09-27; released as the v7.1.3 correction patch). ⚠️ This summary covers results through v5; the repository card is versioned separately as a v7.1.2 card/metadata patch over canonical v7.1 results.
 
 ---
 
@@ -13,23 +15,30 @@ Generated: 2026-03-01 (Updated: 2026-03-29 — v5 biological interpretation adde
 **Scope**: 8 tissues x 8 classifiers x 4 feature types = 256 evaluations
 
 ### Classifiers
-PCA-LR, ElasticNet-LR, Random Forest, XGBoost, SVM-Linear, SVM-RBF, TabNet, LightGBM
+PCA-LR, ElasticNet-LR, Random Forest, XGBoost, SVM-RBF, kNN, MLP, TabNet (`v4/scripts/classifier_registry.py`)
 
 ### Feature Types
-Gene (log2-normalized), Hallmark (ssGSEA), KEGG (ssGSEA), Pathway-combined
+Gene (log2-normalized), Hallmark pathway scores, KEGG pathway scores, Combined (gene + Hallmark)
 
-### Best Results per Tissue
+### Results per Tissue (from `v4/evaluation/M1_summary.json`, regenerated 2026-09-27)
 
-| Tissue | Best AUROC | Method | Feature | perm_p | Significant |
-|--------|-----------|--------|---------|--------|-------------|
-| **Thymus** | **0.948** | PCA-LR | KEGG | <0.05 | Yes* |
-| **Colon** | **0.921** | PCA-LR | KEGG | <0.05 | Yes* |
-| **Lung** | **0.901** | PCA-LR | Gene | <0.05 | Yes* |
-| **Kidney** | **0.829** | ElasticNet-LR | Hallmark | <0.01 | Yes** |
-| **Eye** | 0.823 | PCA-LR | Hallmark | — | — |
-| **Skin** | 0.819 | PCA-LR | Gene | — | — |
-| **Gastrocnemius** | 0.776 | PCA-LR | Gene | — | — |
-| **Liver** | 0.670 | PCA-LR | Gene | — | — |
+Selection rule: the fixed column is PCA-LR on gene features for every tissue, the pre-specified classical reference. The best-row column is the highest AUROC among the tissue's 32 method × feature rows (ties resolved gene → Hallmark → KEGG → combined, as in `v4/scripts/aggregate_results.py`); it is chosen after seeing the results and is therefore optimistic. AUROC is the mean of per-fold AUROC; perm_p is the mean of per-fold permutation p-values (1,000 permutations per fold, uncorrected).
+
+| Tissue | Evaluation | Fixed: PCA-LR / gene, AUROC (perm_p) | Best row: method / features | Best AUROC (perm_p) |
+|---|---|---:|---|---:|
+| Thymus | LOMO, 4 folds | 0.908 (0.038) | PCA-LR / KEGG | 0.948 (0.031) |
+| Gastrocnemius† | LOMO, 3 folds | 0.852 (0.145) | ElasticNet-LR / gene | 0.898 (0.058) |
+| Kidney | LOMO, 3 folds | 0.584 (0.351) | PCA-LR / Hallmark | 0.829 (0.010) |
+| Eye† | LOMO, 3 folds | 0.697 (0.149) | PCA-LR / Hallmark | 0.823 (0.042) |
+| Skin | LOMO, 3 folds | 0.796 (0.005) | ElasticNet-LR / gene | 0.819 (0.004) |
+| Liver | LOMO, 6 folds | 0.683 (0.203) | ElasticNet-LR / KEGG | 0.766 (0.093) |
+| Colon | 5-fold stratified CV (single mission, RR-6) | 0.857 (0.051) | PCA-LR / KEGG | 0.921 (0.033) |
+| Lung | 5-fold stratified CV (single mission, RR-6) | 0.834 (0.127) | ElasticNet-LR / gene | 0.901 (0.028) |
+
+- "Combined" means gene + Hallmark features. For gastrocnemius, skin and lung, ElasticNet-LR on combined features gives the identical AUROC and perm_p.
+- Colon and lung have one mission each (RR-6), so v4 evaluates them with 5-fold stratified CV rather than leave-one-mission-out; rank them separately from the LOMO tissues.
+- † Reported as run. The gastrocnemius fold archived as RR-9 is the SpaceX-10 / RR-4 quadriceps study (OSD-326), and the eye fold OSD-397 belongs to RR-1 (September 2026 correction notice).
+- Regenerated 2026-09-27. The earlier table listed gastrocnemius 0.776 and liver 0.670, which are not v4 rows (0.776 matches the 8-tissue PCA-LR gene mean; 0.670 is the v1 liver gene LOMO value in `evaluation/J5_gene_vs_pathway.json`), named the wrong method for lung, kidney and skin, marked the eye and skin best rows as not significant, and did not state the evaluation scheme for colon and lung.
 
 ### Classifier Rankings (Gene-level mean across 8 tissues)
 
@@ -37,20 +46,20 @@ Gene (log2-normalized), Hallmark (ssGSEA), KEGG (ssGSEA), Pathway-combined
 |------|-----------|----------------|
 | 1 | **PCA-LR** | **0.776** |
 | 2 | ElasticNet-LR | 0.762 |
-| 3 | LightGBM | ~0.72 |
-| 4 | XGBoost | ~0.71 |
-| 5 | Random Forest | ~0.70 |
-| 6 | SVM-Linear | ~0.69 |
+| 3 | XGBoost | 0.709 |
+| 4 | Random Forest | 0.684 |
+| 5 | kNN | 0.666 |
+| 6 | MLP | 0.643 |
 | 7 | TabNet | 0.527 |
 | 8 | SVM-RBF | 0.510 |
 
 ### Key v4 Findings
 
-- **40/256** evaluations significant at p<0.05; **6/8** tissues have >=1 significant result
+- **40/256** evaluations with perm_p < 0.05 (mean of per-fold permutation p-values, uncorrected); **6/8** tissues (all but liver and gastrocnemius) have >=1
 - PCA-LR best overall; deep learning (TabNet) and kernel methods (SVM-RBF) worst
 - Pathway features improve: kidney (0.584->0.829), thymus (0.908->0.948), eye (0.697->0.823)
-- Gene features better for: skin (0.819), lung (0.901)
-- v4 expanded controls: BC/VC included (liver 261 vs v1's 193 samples)
+- Gene features better for: skin (best gene row 0.819 vs best pathway row 0.784) and lung (0.901 vs 0.761; 5-fold CV)
+- v4 expanded controls: basal controls added as ground (v1 already counted vivarium controls): liver 261 vs 193 profiles; all tissues 792 vs 549
 - v1 PCA-LR liver AUROC reproduced exactly (0.5870 vs 0.5871) using task folds
 
 ### v4 Label Encoding
@@ -61,7 +70,7 @@ Gene (log2-normalized), Hallmark (ssGSEA), KEGG (ssGSEA), Pathway-combined
 
 ---
 
-## v1 Results (6 tissues, original analysis)
+## v1 Results (6 tissues, 549 profiles, 22 LOMO folds; original analysis)
 
 ## Hypothesis Results
 
@@ -80,7 +89,7 @@ Gene (log2-normalized), Hallmark (ssGSEA), KEGG (ssGSEA), Pathway-combined
 | **Thymus** | 0.860 | [0.763, 0.953] | 4 | 12 | 1 |
 | **Gastrocnemius** | 0.801 | [0.653, 0.944] | 3 | 6 | 1 |
 | **Skin** | 0.772 | [0.691, 0.834] | 3 | 6 | 2 |
-| **Eye** | 0.754 | [0.688, 0.838] | 3 | 6 | 2 |
+| **Eye** | 0.754 | [0.688, 0.838] | 2 (3 datasets) | 6 (2 within RR-1) | 2 |
 | **Liver** | 0.577 | [0.492, 0.666] | 6 | 30 | 3 |
 | **Kidney** | 0.555 | [0.397, 0.681] | 3 | 6 | 3 |
 
@@ -90,12 +99,14 @@ Thymus vs Liver Δ = 0.283. Permutation tests: Thymus vs Liver p=0.001, Gastro v
 
 | Tissue | AUROC | Raw p | FDR q | Significant? |
 |---|---|---|---|---|
-| Skin | 0.821 | 0.002 | 0.012 | **Yes** |
-| Gastrocnemius | 0.824 | 0.026 | 0.074 | No |
-| Thymus | 0.923 | 0.037 | 0.074 | No |
-| Eye | 0.789 | 0.063 | 0.095 | No |
-| Liver | 0.670 | 0.091 | 0.109 | No |
-| Kidney | 0.432 | 0.281 | 0.281 | No |
+| Skin (LR) | 0.821 | 0.002 | 0.012 | **Yes** |
+| Gastrocnemius (LR) | 0.907 | 0.026 | 0.074 | No |
+| Thymus (PCA-LR) | 0.923 | 0.037‡ | 0.074 | No |
+| Eye (LR) | 0.811 | 0.063 | 0.095 | No |
+| Liver (LR) | 0.653‡ | 0.091‡ | 0.109 | No |
+| Kidney (LR) | 0.593‡ | 0.281‡ | 0.281 | No |
+
+‡ Values of the v1.0 results table (commit eaec2a5), on which the BH-FDR q values were computed; the current result JSONs no longer reproduce them (thymus PCA-LR mean permutation p 0.136 after a rerun; liver LR 0.588, p 0.319; kidney LR 0.521, p 0.391). Corrected 2026-09-27: the table previously paired these p-values with AUROCs from other models and runs (gastrocnemius and eye PCA-LR; liver 0.670 and kidney 0.432 from `evaluation/J5_gene_vs_pathway.json`).
 
 *Note: Only skin survives BH-FDR correction at α=0.05. However, all top-4 tissues have AUROC > 0.7 (GO threshold). High AUROC with modest significance reflects small fold counts (3-4 folds per tissue), not weak signal.*
 
@@ -170,7 +181,7 @@ All pathway F1 ≈ 0.05-0.41 → pathways resist confounder detection (batch-inv
 | Tissue | NES Mean r | Transfer AUROC | N fGSEA missions |
 |---|---|---|---|
 | **Thymus** | **0.619** | **0.860** | 3 |
-| Eye | 0.335 | 0.754 | 3 |
+| Eye | 0.335 | 0.754 | 3 datasets (2 missions) |
 | **Skin** | **0.147** | **0.772** | **3** |
 | Liver | 0.059 | 0.577 | 6 |
 | Gastrocnemius | 0.057 | 0.801* | 2 |
@@ -182,18 +193,23 @@ Excluding gastrocnemius: rank-order correlation for 5 tissues (thymus/eye/skin/l
 
 ---
 
-## Biological Validation (fGSEA Hallmark, all tissues PASS)
+## Biological Validation (fGSEA Hallmark; directions corrected 2026-09-27)
+
+Directions use the flight-positive fGSEA files regenerated on 2026-08-14 (`processed/fgsea/README.md`). The earlier
+readings in this table (thymocyte proliferation, skin cell proliferation and ECM remodelling, retinal metabolic demand,
+literature-concordant liver metabolism) came from the inverted sign. Cell 2020 direction concordance is 28.3% across five
+tissues; per-tissue tables are in `docs/BIOLOGICAL_GROUND_TRUTH.md`.
 
-| Tissue | Top Differentially Enriched Pathways (FLT vs GC) | Consistency |
+| Tissue | Top Differentially Enriched Pathways (FLT vs GC) | Direction in flight (corrected sign) |
 |---|---|---|
-| Liver | OXIDATIVE_PHOSPHORYLATION, FATTY_ACID_METABOLISM | Literature-concordant (direction varies by mission) |
-| Thymus | E2F_TARGETS, G2M_CHECKPOINT, IFN-gamma | Thymocyte proliferation |
-| Gastrocnemius | OXIDATIVE_PHOSPHORYLATION, MYOGENESIS | Muscle metabolism (direction varies by mission) |
-| Kidney | MTORC1_SIGNALING, CHOLESTEROL_HOMEOSTASIS | Renal metabolism |
-| Eye | OXIDATIVE_PHOSPHORYLATION (dominant 3/3 missions) | Retina metabolic demand |
-| Skin | E2F_TARGETS, G2M_CHECKPOINT, EPITHELIAL_MESENCHYMAL_TRANSITION | Cell proliferation + ECM remodeling (2/3 missions consistent) |
+| Liver | OXIDATIVE_PHOSPHORYLATION, FATTY_ACID_METABOLISM | Higher in RR-3 and RR-9, lower in the other four missions |
+| Thymus | E2F_TARGETS, G2M_CHECKPOINT, IFN-gamma | Proliferation lower and IFN-γ higher in every contrast |
+| Gastrocnemius | OXIDATIVE_PHOSPHORYLATION, MYOGENESIS | Opposite signs in the two entries; the second entry is GLDS-326, the RR-4 quadriceps study |
+| Kidney | MTORC1_SIGNALING, CHOLESTEROL_HOMEOSTASIS | Higher in RR-1 and RR-3, lower in RR-7 |
+| Eye | OXIDATIVE_PHOSPHORYLATION (dominant in all 3 datasets, 2 missions) | Lower in all three entries |
+| Skin | E2F_TARGETS, G2M_CHECKPOINT, EPITHELIAL_MESENCHYMAL_TRANSITION | Lower in all three entries |
 
-*Note: "Top Differentially Enriched" = highest |NES| across missions. Enrichment direction (up/down in spaceflight) may vary by mission for liver and gastrocnemius due to mission-specific biological variability. See individual fGSEA result files in `processed/fgsea/` for per-mission NES values and directions.*
+*Note: "Top Differentially Enriched" = highest |NES| across missions. See individual fGSEA result files in `processed/fgsea/` for per-mission NES values and directions.*
 
 ---
 
@@ -213,7 +229,7 @@ Mouse-Geneformer (6L BERT, 56K mouse gene vocab, pretrained on 30M scRNA-seq cel
 
 **Interpretation**: Classical ML wins 6/6 tissues (sign test p=0.016). Geneformer performs near chance level (0.5) on small-n bulk RNA-seq (train n=30-100). This is consistent with literature — foundation models pretrained on single-cell data do not automatically transfer to small-sample bulk transcriptomics tasks.
 
-*Note: Table shows best baseline per tissue for fair comparison. Publication figures use unified PCA-LR baseline (mean 0.743) for cross-figure consistency with Category A/B results.*
+*Note: Baseline = a fixed classical model per tissue (LR-ElasticNet for A1, A2, A3, A5; PCA-LR for A4, A6; `BEST_BASELINE` in `scripts/aggregate_geneformer_results.py`), not the per-tissue maximum; 0.758 is the unweighted mean of the six tissue fold-means. Fine-tuned scGPT and Mouse-Geneformer fold AUROCs keep the best of 10 epochs scored on the held-out test mission itself (no inner validation split; `scripts/scgpt_finetune.py`, `scripts/geneformer_finetune.py`), so they are optimistic. Publication figures use unified PCA-LR baseline (mean 0.743) for cross-figure consistency with Category A/B results.*
 
 ---
 
@@ -221,7 +237,7 @@ Mouse-Geneformer (6L BERT, 56K mouse gene vocab, pretrained on 30M scRNA-seq cel
 
 scGPT-whole_human (12L Transformer, 512d hidden, 8 heads, pretrained on 33M human CellXGene cells) fine-tuned on mouse bulk RNA-seq LOMO folds via ENSMUSG→human gene symbol ortholog mapping. Training: 10 epochs, batch=8, lr=1e-4, freeze=10/12 layers (flash_attn disabled for PyTorch 2.1 compatibility).
 
-**Note on reliability**: Folds with n_test ≤ 8 (MHU-1 thymus, RR-9 gastro) produce highly variable AUROC estimates and should be interpreted with caution. Large-n folds (RR-8 liver n=103, RR-7 kidney n=94, RR-7 skin n=30) are most reliable.
+**Note on reliability**: Folds with n_test ≤ 8 (MHU-1 thymus; A2 fold `RR-9`, which is RR-4 quadriceps, OSD-326) produce highly variable AUROC estimates and should be interpreted with caution. Large-n folds (RR-8 liver n=103, RR-7 kidney n=94, RR-7 skin n=30) are most reliable.
 
 | Task | Tissue | scGPT AUROC | Geneformer AUROC | Baseline AUROC (unified PCA-LR) | Δ vs GF | Δ vs Baseline | Winner |
 |------|--------|------------|-----------------|---------------|---------|--------------|--------|
@@ -237,7 +253,7 @@ scGPT-whole_human (12L Transformer, 512d hidden, 8 heads, pretrained on 33M huma
 
 > ⚠️ **Two tables, two Geneformer columns.** The Geneformer AUROCs above differ from the Tier 2 table for
 > five of six tissues (e.g. gastrocnemius 0.432 here vs 0.382 there, kidney 0.432 vs 0.452). The baseline
-> columns also differ, because Tier 2 reports the **best baseline per tissue** (mean 0.758) while this table
+> columns also differ, because Tier 2 reports a **fixed classical model per tissue** (mean 0.758) while this table
 > reports the **unified PCA-LR baseline** (mean 0.743). The baseline difference is intentional and now
 > labelled; **the Geneformer difference is not yet reconciled against the source runs.** Treat the Tier 2
 > column as the primary Geneformer result until that is resolved, and do not quote a Geneformer number from
@@ -245,13 +261,13 @@ scGPT-whole_human (12L Transformer, 512d hidden, 8 heads, pretrained on 33M huma
 
 **Key observation**: scGPT shows higher variance (std=0.05–0.31) than Geneformer (std=0.05–0.23), partly reflecting ortholog mapping noise from human pretraining. Large-n reliable folds (liver RR-8 n=103: 0.468; kidney RR-7 n=94: 0.557; skin n=30–39: 0.636–0.737) suggest scGPT hovers near chance (0.5) on the most statistically robust estimates.
 
-*Results file: `evaluation/scgpt_whole_human_all_tissues_summary.json`*
+*Results file: `evaluation/scgpt/scgpt_whole_human_all_tissues_summary.json`* (written by `scripts/aggregate_scgpt_results.py`). Fine-tuned scGPT and Mouse-Geneformer fold AUROCs keep the best of 10 epochs scored on the held-out test mission itself (no inner validation split; `scripts/scgpt_finetune.py`, `scripts/geneformer_finetune.py`), so they are optimistic.
 
 ---
 
-## Held-Out Evaluation: A4 Thymus (OSD-515 / RR-23)
+## Retrospective Open Validation: A4 Thymus (OSD-515 / RR-23)
 
-Reserved held-out test set for external benchmark evaluation. Train on 4 missions (MHU-1, MHU-2, RR-6, RR-9; n=67), test on RR-23 (n=16: 7 Flight, 9 GC). 27,541 common genes.
+Retrospective open validation on a fifth mission (labels public; not a blind holdout). Train on the 4 LOMO missions (MHU-1, MHU-2, RR-6, RR-9; n=67), test on RR-23 (n=16: 7 Flight, 9 GC). 27,541 common genes.
 
 | Model | AUROC | 95% CI | p-value |
 |-------|-------|--------|---------|
@@ -260,7 +276,7 @@ Reserved held-out test set for external benchmark evaluation. Train on 4 mission
 | PCA-50 + LogReg | 0.873 | [0.609, 1.000] | 0.011 |
 | Geneformer (Mouse-GF) | 0.556 | [0.265, 0.850] | — |
 
-**Interpretation**: Classical baselines achieve strong held-out performance (AUROC ~0.90, p<0.01), confirming thymus cross-mission generalization beyond LOMO. Geneformer remains near chance on held-out data, consistent with LOMO results (0.495). The held-out confirms thymus as the most robust tissue for spaceflight detection.
+**Interpretation**: Classical baselines reach AUROC 0.87-0.91 (p <= 0.011) on RR-23, a mission outside the four LOMO folds. With n=16, a 95% CI reaching 1.000 and public labels, this is consistent with the thymus LOMO result, not a confirmation of it. Geneformer stays near chance (0.556), as in LOMO (0.495).
 
 ---
 
@@ -362,9 +378,9 @@ RR-8 shows strong recovery with overshoot past baseline (MYC targets V1 +2.49, P
 
 ---
 
-## Held-Out Evaluation: A5 Skin (OSD-254 / RR-7)
+## RR-7 Skin (OSD-254): copy of the LOMO RR-7 fold
 
-Second held-out test set. Train on 2 missions (RR-6, MHU-2; n=72), test on RR-7 (n=30: 10 Flight, 20 GC). 20,110 common genes. RR-7 is a 75-day mission (longest in skin dataset).
+Not a second held-out set: `fold_RR-7_holdout` is byte-identical to the LOMO fold `fold_RR-7_test` (train RR-6 + MHU-2, n=72; test RR-7, n=30: 10 Flight, 10 GC, 10 VC; 20,110 genes after the variance filter). RR-7 includes 25-day and 75-day animals. The LR value below (0.885) differs from the LR value for the same split in `A5_baseline_results.json` (0.805); RF and PCA-LR are identical in both files. A converged refit with the current code (5,853 SAGA iterations) reproduces the LOMO value 0.805 on this split, and the RR-23 value 0.905 reproduces exactly; the RR-7 value 0.885 does not reproduce, and its 8.5 s training time (210 s for the LOMO run, 41 s for the refit) suggests the solver stopped early.
 
 | Model | AUROC | 95% CI | p-value |
 |-------|-------|--------|---------|
@@ -372,14 +388,14 @@ Second held-out test set. Train on 2 missions (RR-6, MHU-2; n=72), test on RR-7
 | PCA-50 + LogReg | 0.840 | [0.679, 0.963] | 0.001 |
 | Random Forest | 0.777 | [0.583, 0.929] | 0.007 |
 
-**Cross-Tissue Held-Out Comparison**:
+**Open-validation rows** (only RR-23 is outside the LOMO folds):
 
 | Tissue | Mission | Duration | Best AUROC | n_test |
 |--------|---------|----------|------------|--------|
 | Thymus | RR-23 | 30 days | 0.905 (LR) | 16 |
-| Skin | RR-7 | 75 days | 0.885 (LR) | 30 |
+| Skin | RR-7 (LOMO fold copy) | 25 and 75 days | 0.885 recorded, not reproducible (LR; converged refit and LOMO run: 0.805) | 30 |
 
-**Interpretation**: Skin held-out confirms strong generalization (AUROC 0.885, p<0.001), exceeding the LOMO mean (0.821). Both held-out tissues achieve AUROC > 0.85, which is consistent with cross-mission spaceflight detection beyond leave-one-out evaluation. These are small held-out sets and the thymus 95% CI reaches 1.000 ([0.672, 1.000], n=16), so this is supporting evidence, not validation.
+**Interpretation**: The skin row adds nothing beyond LOMO: it is the LOMO RR-7 fold scored again, and that fold is already inside the LOMO mean (0.821). Only thymus RR-23 is a mission outside the LOMO folds; it is small (n=16; 95% CI [0.672, 1.000]) and its labels were public, so it is supporting evidence, not validation.
 
 ---
 
@@ -401,7 +417,7 @@ Second held-out test set. Train on 2 missions (RR-6, MHU-2; n=72), test on RR-7
 | Geneformer | 6 tissues, 22 LOMO folds (Mouse-GF) | Complete |
 | scGPT | 6 tissues, 21 LOMO folds (whole_human), mean AUROC=0.666 | Complete |
 | LLM Zero-Shot | 3 providers × 6 tasks (18 evals) | Complete |
-| Held-Out | A4 Thymus (RR-23) + A5 Skin (RR-7) | Complete |
+| Retrospective open validation | A4 Thymus (RR-23); A5 Skin RR-7 is a copy of the LOMO fold | Complete |
 | T1-T3 Temporal | ISS-T/LAR, Recovery, Age×Spaceflight | Complete |
 | J2 DGE Pipeline | 9 missions × 3 pipelines (DESeq2/edgeR/limma-voom) | Complete |
 | **v1 Figures** | **4 main + 4 supplementary (HTML/SVG)** | **Complete** |
@@ -418,13 +434,13 @@ Second held-out test set. Train on 2 missions (RR-6, MHU-2; n=72), test on RR-7
 
 | Model | Liver | Gastro | Kidney | Thymus | Eye | Lung | Colon | Mean |
 |-------|-------|--------|--------|--------|-----|------|-------|------|
-| **PCA-LR** | **0.670** | **0.824** | 0.432 | **0.923** | **0.789** | — | — | **0.758** |
-| scGPT | 0.628 | 0.685 | **0.556** | 0.782 | 0.650 | — | — | 0.667 |
+| **Classical reference**§ | 0.588 | **0.907** | 0.521 | **0.923** | **0.789** | — | — | **0.758** |
+| scGPT | 0.628 | 0.685 | **0.556** | 0.782 | 0.650 | — | — | 0.666 |
 | scFoundation | 0.635** | 0.691* | 0.541 | 0.487 | 0.563 | 0.389 | 0.755 | ~0.58 |
 | UCE (seeded) | 0.459 | 0.578 | 0.489 | 0.632* | 0.550 | 0.555 | 0.449 | ~0.53 |
 | Geneformer | 0.486 | 0.382 | 0.452 | 0.495 | 0.484 | — | — | 0.476 |
 
-*p<0.05, **p<0.01. All FMs underperform PCA-LR baseline.
+*p<0.05, **p<0.01. § Classical reference = the fixed per-tissue model of the Tier 2 comparison (LR-ElasticNet for liver, gastrocnemius, kidney and skin; PCA-LR for thymus and eye). Corrected 2026-09-27: the earlier PCA-LR row mixed values from different runs (liver 0.670 and kidney 0.432 from `evaluation/J5_gene_vs_pathway.json`, gastrocnemius PCA-LR 0.824) with the 0.758 mean of this reference. The classical, scGPT and Geneformer means are six-tissue means including skin (0.821 / 0.691 / 0.557), which has no column; the scGPT six-tissue mean is 0.666 (its 21-fold mean is 0.667). All FMs underperform the classical reference.
 
 ### RRRM-2 scRNA-seq (F5)
 

manifest.json

@@ -14,9 +14,9 @@
       "lane_id": "v7.1",
       "public_label": "canonical historical result surface",
       "status": "canonical_historical_result_surface",
-      "version": "7.1.2",
-      "date_released": "2026-06-05",
-      "summary": "The v1-v7 canonical benchmark result surface plus the v7.1.2 documentation, public-card, and metadata patch. The patch does not introduce new benchmark result generation.",
+      "version": "7.1.3",
+      "date_released": "2026-09-27",
+      "summary": "The v1-v7 canonical benchmark result surface plus the v7.1.2 documentation, public-card, and metadata patch and the v7.1.3 correction patch (mission labels, regenerated v4 headline table, scope and averaging statements, MAQC 2026 abstract erratum). Neither patch introduces new benchmark result generation.",
       "canonical_surfaces": [
         "README.md",
         "docs/CANONICAL_RESULTS_V7_1.md",
@@ -28,11 +28,11 @@
       "huggingface": {
         "repo_id": "jang1563/genelab-benchmark",
         "dataset_card_path": "docs/hf_dataset_card.md",
-        "revision_status": "remote_current_v7_1_2_public_card_patch"
+        "revision_status": "remote_current_v7_1_3_correction_patch"
       },
       "citation": {
         "cff_path": "CITATION.cff",
-        "citation_version": "7.1.2",
+        "citation_version": "7.1.3",
         "status": "software_citation_available"
       }
     },