Archive 2 checkpoints (1-2)

paper-releases/20260928/diffusion/data/cohorts/ye2026_sepsis/QC_REPORT.md

@@ -0,0 +1,114 @@
+# Ye 2026 sepsis: paired PBMC bulk and scRNA (QC report)
+
+Study: Ye Q et al., Nat Immunol 2026. Bulk is GEO GSE279448 and scRNA is GSE279452. CITE-seq (GSE279451) was not used.
+Built 2026-09-27 on server_1, CPU only. Processing directory: `/home/sj_server_1/gblee/datalake/paired-pbmc-bulk-sc/processing/ye2026_sepsis_v1_20260927/`.
+
+## Result
+
+**85 pairs** (one bulk, one sc library, one donor each). **82 pass** the identity checks and **3 fail**. Failed rows are kept and flagged (`identity_pass` in features.npz). Exclude them from training and testing unless the authors confirm the labels.
+
+| group | pairs | retained cells per sample, median (min to max) |
+|---|---|---|
+| adult_sepsis | 48 | 6,935 (2,657 to 14,291) |
+| ped_sepsis | 16 | 6,699 (3,472 to 9,427) |
+| adult_HC | 7 | 9,812 (3,720 to 11,075) |
+| adult_ICU_ctrl | 6 | 7,320 (2,107 to 15,576) |
+| ped_HC | 4 | 9,068 (6,519 to 10,918) |
+| ped_ICU_ctrl | 4 | 7,883 (6,570 to 11,535) |
+| all | 85 | 7,172 (2,107 to 15,576); 610,713 cells in total |
+
+The expected count was 80. Exact title matches give exactly 80. Five more pairs differ only in how the sc title is written:
+
+- `RES-PS26`, `RES-S164`, `RES-S174` and `RES-S177` differ from the bulk IDs only in letter case.
+- sc `Skin-152` is bulk `Skin-S152` (the "S" is missing).
+
+All five agree with their bulk on GEO sex, age, group, infection source and collection site. All five pass both identity checks. The `id_match` field marks them.
+
+## Pairing rule
+
+- The sample ID is the GEO title without its `bulkRNA ` or `scRNA ` prefix. It is matched exactly first, then case-insensitively, then with the missing "S" restored.
+- An accepted pair must also agree on GEO sex, age, group, infection source and collection site. All 85 agree.
+- There are no time points. Each ID appears once in each series, and GEO has no visit field. So donor = sample.
+- Only the 133 GEX GSMs are used. The 133 VDJ GSMs (BCR/TCR contigs) are not paired.
+
+**Not paired:**
+
+- 48 sc libraries have no bulk: 45 adult sepsis, 2 adult ICU and 1 pediatric sepsis (`Abd-PS18`).
+- 122 bulk samples have no sc: 68 adult sepsis, 36 adult HC, 17 adult ICU and 1 pediatric ICU.
+
+Full lists are in MANIFEST.json (`excluded`) and in `processing/.../pairing/`. No paired sc library failed processing.
+
+## Identity checks
+
+**Sex.** The sex call uses XIST against RPS4Y1 + DDX3Y + UTY + KDM5D, in CPM on the source gene axis.
+
+- 84 of 85 pairs agree across bulk, sc and GEO.
+- **Uri-S112** disagrees: bulk is male (XIST 2 CPM, Y 491 CPM), sc is female (XIST 96 CPM, Y 0.06 CPM), and GEO says male.
+- The bulk of Uri-S20 is female but shows a small Y signal (55 CPM). This may be minor admixture. The pair still passes.
+- Outside the pairs, bulk Res-S34 is labelled male but its expression is female. Bulk HC23 and HC30 show both XIST and Y signal. All three are unpaired.
+
+**Expression, as specified.** For each bulk, Spearman on log CPM over the top 2,000 genes by mean, against all 133 sc pseudobulks.
+
+- Only 6 of 85 bulks pick their own pair (median own rank 15).
+- After centering each modality, 32 of 85 do.
+- This gene set mostly tracks disease state and cell composition, not the donor, so this test cannot tell donors apart here.
+
+**Expression, donor-sensitive.** This is the test that decides pass or fail. It uses no pairing information.
+
+- Method: within each modality, remove the top 10 sample PCs. Then use the 2,000 genes with the highest remaining variance in both modalities, and compute Pearson correlation.
+- Result: 81 of 85 bulks pick their own sc as the top match, and 82 of 85 sc pick their own bulk. The result is the same with 20 or 30 PCs removed.
+- Passing pairs have r = 0.58 (median; minimum 0.25). The three failures have r near 0.
+
+**Verdict rule.** Pass means sex agrees and the own pair ranks 1st in one direction and within the top 5 in the other. Result: 82 pass, 3 fail, none borderline.
+
+| failed pair | why | best alternative match |
+|---|---|---|
+| Uri-S112 | sex differs; own pair ranks 42 and 78 | sc Uri-S112 matches bulk **Uri-S104** best (r 0.56). Uri-S104 is GEO female, 63 y, has no sc library. |
+| Abd-S8 | own pair ranks 112 and 186 | bulk Abd-S8 and sc **Abd-S80** are each other's best match (r 0.49). Abd-S80 is male, 67 y, abdominal like Abd-S8, but GEO lists a different collection site. It has no bulk. |
+| Abd-PS13 | own pair ranks 77 and 102 | no strong alternative (best r 0.22) |
+
+These alternative matches are leads, not proof. The pairs were not re-labelled.
+
+## Processing
+
+The run used the frozen ImMicro strict-QC code (bundle v5, copied unchanged and sha-checked), the same 20,097-gene GENCODE v49 axis and the same mapping rules.
+
+**Bulk**
+
+- Per-GSM htseq-count tables with 60,664 unversioned Ensembl gene IDs.
+- 19,954 of the 20,097 canonical genes are present. 143 are missing and are marked as unmeasured. 112 source IDs are unknown.
+- All values are integers. There are no duplicate genes.
+- The five htseq `__` counters are kept as metadata. `__no_feature` is 11 to 22% of assigned + no_feature + ambiguous reads. This argues against a strand mismatch, but GEO does not state the htseq strand setting.
+- GEO protocol: PBMC total RNA (Trizol); "standard Illumina protocols" (no kit named; polyA vs rRNA depletion not stated); NovaSeq 6000; HISAT2 2.0.4 to GRCh38; htseq-count.
+
+**scRNA**
+
+- One 10x library per sample, so no demultiplexing. Barcodes are unique; all carry the suffix `-1`.
+- The feature table is the same for all libraries: 37,487 Ensembl genes, in non-genomic order.
+- 19,864 canonical genes are present and 233 are missing. On the 6k panel, 5,975 of 6,000 genes are measured.
+- 3 genes have a symbol that disagrees with the ID. They are kept under their ID, as in ImMicro.
+- Counts are integer UMIs, and every library conserves its total when mapped to the canonical axis.
+- Chemistry: GEO says the GEX is 10x 3' v3.1, but the data show 5'. Every library uses the 737K-august-2016 barcode list (the 3' v3.1 list is different). A pooled 98.2% (median 99.5%) of TCR contig cells are also GEX barcodes. So the GEX came from the same 5' reaction as the V(D)J libraries. The matrices were written by Cell Ranger 6.1.1.
+- The GEO matrices are Cell Ranger cell calls, not the authors' Seurat-filtered cells. Across all 133 libraries (980,095 cells), no cell has under 500 UMIs. But 23,106 cells have 200 genes or fewer and 20,487 have 20% mito or more, which the authors' stated filters would have removed.
+- Retained cells follow the ImMicro `flag_conservative_review_pass` rule: at least 200 genes, at least 500 UMIs and at most 20% mito. This keeps 610,713 of 632,805 cells (96.5%) in the paired libraries. The 5-MAD outlier flags are stored as review-only columns and do not remove cells.
+- GEO provides no author cell-type labels, so `author_cell_type` is empty.
+- One library, Abd-PS05, has 32% of its UMIs in immunoglobulin V genes. These are outside the protein-coding axis, so they are excluded from its counts. This is likely a plasmablast expansion.
+
+## Files
+
+| file | contents |
+|---|---|
+| `features.npz` | One row per pair, in bulk GSM order. Keys follow ImMicro. `bulk_raw` is 0 where not measured; ImMicro uses NaN there. |
+| `cells_panel6k.npz` | CSR int32 counts for the retained cells: 610,713 × 6,000, 255.6 M nonzero, compressed. |
+| `cells_obs.parquet` | One row per retained cell, in the same order. |
+| `samples.csv` | Row index to sample and group. |
+| `pairs.csv` | GEO fields, sex and expression checks, and the verdict for each pair. |
+| `identity_*.csv` | Full sex and expression check tables. |
+| `MANIFEST.json` | Sources with sha256, code with sha256, and all checks. |
+
+Checks passed:
+
+- For every row, the summed panel counts equal `pseudo_raw_sum[:, positions]` exactly.
+- Cell counts equal `qc_cell_counts`.
+- There are no counts outside the measured masks.
+- The sha256 of `gene_positions` matches.

paper-releases/20260928/diffusion/data/cohorts/ye2026_sepsis/cells_full.BUILD.json

@@ -0,0 +1,10 @@
+{
+ "dataset": "ye2026_sepsis",
+ "n_cells": 610713,
+ "nnz": 934565850,
+ "sha256": "0b3606b6ab70342eca37c2ac8068d86593774bb921f756db4295b5d1d0a33f3f",
+ "checks": {
+  "panel_columns_equal_cells_panel6k": true,
+  "row_sums_equal_pseudo_raw_sum": true
+ }
+}