Archive 2 checkpoints (1-2)

paper-releases/20260928/diffusion/data/cohorts/giroux2022/QC_REPORT.md

@@ -0,0 +1,100 @@
+# Giroux 2022 COVID-19: paired PBMC bulk and scRNA (QC report)
+
+Study: Giroux NS et al., Sci Rep 2022 (Duke outpatients, early COVID-19). SuperSeries GSE206284. Bulk is GEO GSE206263 and scRNA is GSE206283.
+Built 2026-09-28 on server_1, CPU only. Processing directory: `/home/sj_server_1/gblee/datalake/paired-pbmc-bulk-sc/processing/giroux2022_v1_20260928/`.
+
+## Result
+
+**26 pairs from 10 subjects.** One row is one subject on one sampling day. **24 pass** the identity checks and **2 fail** (subject 40067F, see below). Failed rows are kept and flagged (`identity_pass`); `prepare_dataset.py` gives them fold -1.
+
+| group | subjects | pairs | days | retained cells per sample, median (min to max) |
+|---|---|---|---|---|
+| mild | 5 | 13 | 0, 7, 14 | 11,344 (5,914 to 15,650) |
+| severe | 5 | 13 | 0, 7, 14 | 10,860 (6,510 to 14,268) |
+| all | 10 | 26 | d0: 10, d7: 7, d14: 9 | 10,958 (5,914 to 15,650); 276,615 cells in total |
+
+The expected count was 39 pairs from 18 subjects. **GEO metadata alone does give exactly 39** (mild 15, severe 13, exposed 6, healthy 5). But the only published bulk counts are the authors' table, and it holds just 49 of the 71 bulk GSMs:
+
+- It has **no exposed subjects** at all: 20 bulk GSMs are missing.
+- It has no column for 450905 day 7 or 0B943B day 3.
+- 5 of its 7 healthy columns (77DA77, 22CBD6, 4F2EAD, F8FF1A, 56E219) use IDs that appear nowhere in GEO. The 5 healthy subjects with scRNA (E0FA1E, 3F3322, 94D44E, 52BA23, 655A91) have no column.
+
+So 13 of the 39 GEO-level pairs have no bulk counts. The raw FASTQs for all 71 bulk GSMs are in SRA (PRJNA849921). NCBI offers no generated count table for this series.
+
+**Healthy leads (not pairs).** Each of the 5 unknown healthy bulk columns matches exactly one healthy sc subject, and they are each other's best match. The residualised r is 0.42 to 0.71, against 0.33 or less for the next best. Sex agrees for all 5. The same mapping also follows from row order: the healthy rows of the author sample sheet line up position by position with the healthy GSMs in GEO (77DA77 = E0FA1E, 22CBD6 = 3F3322, 4F2EAD = 94D44E, F8FF1A = 52BA23, 56E219 = 655A91). No published ID links them, so they are not paired. See `MANIFEST.json` (`leads_not_pairs`).
+
+## Pairing rule
+
+- The GEO title is `<subject>, <day>, <group>, <serostatus>` in both series. It is checked against the characteristics. The sc `library` field equals the subject, and the day suffix of each author library name (e.g. `26-7`) equals the GEO day.
+- A pair is an sc GSM and a bulk GSM with the same subject **and** the same day. Both must agree on GEO age, sex, group and serostatus; all do. The bulk GSM must also have an author-table column. Columns are named `<subject>-<day>`, or `<subject>` for healthy.
+- `sample_id` = `<subject>_d<day>`. `donor_ids` = subject. `timepoints` = day. Groups are the GEO labels, lower-cased. The paper and the author sheet call the "severe" group "moderate"; this is kept in `author_cohort`.
+
+**Not paired** (full lists: `MANIFEST.json` `excluded`, and `processing/.../pairing/`):
+
+- 15 sc libraries. 13 have a bulk GSM but no author column: 450905 d7, 0B943B d3, 7768E4 d14, 61BBAD d0/d7/d14, 80A16A d0/d14 and the 5 healthy. Two (7768E4 d2 and d28) have no bulk GSM on that day.
+- 23 author columns. 17 have no sc library: 1A9B20, 75A2B6, DF309F, 0E1F8E (3 days each), 3F05F3 d0/d7, 40067F d21, and healthy 062CD8 and E9CFE0. 6 have no GEO GSM: the 5 healthy above, plus 3F05F3-14 (GEO lists 3F05F3 on days 0, 7 and 28).
+
+## Identity checks
+
+**Sex.** XIST against Y genes, in CPM on the source axis (the Ye 2026 rule). The bulk uses DDX3Y + UTY + KDM5D, because RPS4Y1 was removed from the author table (see below). The sc uses all 4 genes; a 3-gene sc call gives the same result. **26 of 26 pairs agree** across bulk, sc and GEO (15 male, 11 female rows).
+
+**Donor-sensitive expression.** Within each modality, remove the top **k = 5** sample PCs. Then use the 2,000 genes with the highest remaining variance in both, and compute Pearson over all 49 bulk columns × 41 sc libraries.
+
+- We used k = 5, not 10, because there are only 41 sc and 49 bulk samples (Ye 2026 had 133 and 207). Results for k = 3 and k = 10 are also reported.
+- chrY genes are left out, so this test does not depend on the sex check.
+- A subject's other days are the same person, so the ranks are over **subjects**, each scored by its best sample.
+- Result: at k = 5 the own subject is the top match for **26 of 26** rows in both directions. At k = 3 it is 26 and 25; at k = 10, 24 and 20. The own library is the top match for 23 of 26.
+- Own-pair r: median 0.53, minimum 0.26.
+- The simple Spearman test on the 2,000 highest-mean genes does not separate donors (4 of 26 top-1), as in Ye 2026.
+
+**Same day.** The subject test cannot see a day swap within a subject. So the own day is also ranked among the subject's paired days, on both the residualised and the centred correlation, in both directions.
+
+- A row fails when another day wins in all four. We added this rule after looking at the data, because the task requires the same sampling day.
+- **40067F days 0 and 14 fail.** Bulk 40067F-0 matches sc day 14 (r 0.63) far better than sc day 0 (0.28), and bulk 40067F-14 prefers sc day 0.
+- The interferon score (12 ISGs) points the same way. In the other 9 subjects, both assays peak on day 0. For 40067F, the bulk peaks on "day 14" and the sc on "day 0".
+- Most likely one assay's day 0 and day 14 labels are swapped. We cannot tell which, so the rows were not re-labelled.
+
+**Verdict.** Pass means sex agrees, the own subject ranks 1st in one direction and within the top 3 in the other, and there is no day failure. Result: 24 pass, 2 fail, none borderline. `identity_notes` records weaker day disagreements (82CCF5 d0 and 5 others) and the two low-depth bulks.
+
+## Processing
+
+This build uses the frozen ImMicro strict-QC code (bundle v5, copied unchanged and sha-checked) and the same 20,097-gene GENCODE v49 axis, following the Ye 2026 build.
+
+**Bulk**
+
+- The author table `bulk_RNA-seq_all_counts.csv` has 26,045 **hg19 gene symbols** (no Ensembl IDs) × 49 columns. All values are **integers**, with no duplicates.
+- Symbols were mapped with the bundle's GeneMapper: 18,741 canonical genes are measured. Three many-to-one symbol groups were summed, following the bundle's rule for verified counts. 37 ambiguous and 1,625 unknown symbols were dropped.
+- **The authors removed genes by name.** No symbol contains RPL or RPS, and none starts with MT. That removes ribosomal proteins, MRPL/MRPS, mitochondrial genes, MTOR and similar (ICT1 = MRPL58 survived under its old name).
+  - These 275 canonical genes are `bulk_valid = False`, not zero. 121 of them are on the 6k panel.
+  - In the sc they carry a median 36% (22 to 42%) of UMIs. Compare bulk and sc depth only on `pair_valid` genes.
+- Assigned reads per column: median 13.5 M. Two paired bulks are shallow: **180E1A-0 at 0.65 M** and **7085CA-0 at 1.06 M**. Both still pass the identity checks.
+- GEO protocol: Zymo Direct-zol total RNA, NuGEN Ovation SoLo, NovaSeq 6000, SortMeRNA, STAR to hg19, featureCounts. The strandedness option is not stated.
+
+**scRNA**
+
+- There is one 10x library per subject-day (5' v1, Cell Ranger 3.1.0, hg19), so no demultiplexing is needed.
+- The 41 feature tables are identical (32,738 genes). All barcodes come from the 737K-2016 list (`-1` suffix), which fits 5' v1.
+- 18,976 canonical genes are measured. 90 genes whose symbol disagrees with the ID are kept under the ID, as in ImMicro. Counts are integer UMIs, and every library conserves its total.
+- The GEO matrices are Cell Ranger cells with a 500-UMI floor, not the authors' Seurat subset (≤5% mito, ≤2,500 genes). Retained cells follow the ImMicro rule `flag_conservative_review_pass`. This keeps 276,615 of 284,224 cells in the paired libraries (97.3%); the authors' rule would keep only 183,798. There are no author cell-type labels.
+
+**6k panel.** 5,806 of 6,000 genes are measured in the sc, 5,640 in the bulk, and 5,563 in both. Ye 2026 had 5,975 and 5,980, so the older hg19 annotations lose more genes here.
+
+## Files
+
+| file | contents |
+|---|---|
+| `features.npz` | One row per pair, in bulk GSM order. Keys follow Ye 2026, plus `timepoints`, `serostatus_geo`, `bulk_count_column`, `sc_author_library` and `author_cohort`. It has no `infection_source` or `collection_site`. `bulk_raw` is 0 where not measured. |
+| `cells_panel6k.npz` | CSR int32: 276,615 × 6,000, 98.6 M nonzero. |
+| `cells_full.npz` | The same cells on the full axis: CSR int32, 356.1 M nonzero, 2.85 GB (`cells_full.BUILD.json`). |
+| `cells_obs.parquet` | One row per retained cell. The columns are identical to Ye 2026. |
+| `samples.csv`, `pairs.csv`, `identity_*.csv` | Row index, GEO fields, sex and expression checks, and the verdict for each pair. |
+| `MANIFEST.json` | Sources and code with sha256, all checks, the excluded lists and the healthy leads. |
+
+**Checks passed:**
+
+- The summed panel counts equal `pseudo_raw_sum[:, positions]` exactly, per row. The full-axis cells sum to `pseudo_raw_sum`, and their panel columns equal `cells_panel6k`; this was re-checked on the written file.
+- Cell counts equal `qc_cell_counts`, and there are no counts outside the masks.
+- The `gene_positions` sha256 matches.
+- `prepare_dataset.py --dataset giroux2022` runs: 5 subject-level folds over the 9 usable subjects.
+
+**Caveat for the scorer.** `tools/ft_common.py load_features` refuses datasets with more than one sample per donor. This dataset has 2 to 3 per subject, so the scorer needs an adapter.

paper-releases/20260928/diffusion/data/cohorts/giroux2022/cells_full.BUILD.json

@@ -0,0 +1,11 @@
+{
+ "dataset": "giroux2022",
+ "n_cells": 276615,
+ "nnz": 356147396,
+ "sha256": "44613c8b250fc8a8fe39490ecb0cd7fba308dd67b98ce611c1b5912c9274fc18",
+ "checks": {
+  "panel_columns_equal_cells_panel6k": true,
+  "row_sums_equal_pseudo_raw_sum": true
+ },
+ "builder": "processing/giroux2022_v1_20260928/code/giroux2022_full_cells.py (build_full_cells.read_library, same checks, streamed; re-verified on the written file)"
+}