UC6 — Cell-type-resolved pseudobulk DE (COVID-19) — drive debrief

Driven via /drive-scenario against local Galaxy (backend 8080, dev client 5173), 2026-06-20.

TL;DR

• MVP is a faithful reproduction. The shipped IWC pseudo-bulk_edgeR.ga runs end-to-end on the COVID-19 PBMC AnnData: pseudobulk QC plot, expression-filter plot, pooled normal vs COVID-19 edgeR contrast → DEG table (1,430 rows, byte-matches the test fixture) + volcano. All genuine on-graph outputs.
• The stretch machinery works; the biology does not survive. Splitting the pseudobulk matrix by cell type → a 6-element collection → mapped edgeR (a clean paired collection map-over) is solid engineering and the spec’s headline deliverable. But this dataset has 7 individuals (3 normal, 4 COVID); split across 6 cell types → 2–4 samples/arm.
• Result: no robust cell-type-specific signal claimable. The four best-powered cell types (B, T, monocyte, neutrophil) return 0 genes at FDR<0.05. The only “significant” cell type is platelet (2v2) with 10 hits, and those hits are hemoglobin genes (HBB/HBA1/HBA2, FDR 1e-22…1e-33) — RBC ambient-contamination markers in an anucleate cell type, amplified by small-n edgeR overconfidence. Textbook false positive: smallest sample, biologically implausible markers, absurd significance.
• The honest biology pointer: monocyte’s strongest gene is TNFAIP3 (NF-κB / inflammation — the expected COVID-monocyte response) at P≈8e-5 but FDR≈0.11: pointing the right way, underpowered at n=3.
• Transferable win = the decomposition workflow + the power/contamination teaching story. Same red-flag lesson as UC5: a lone extreme result from the smallest stratum is an artifact to debunk, not a finding.

Artifacts

• History b2486a20bc56b90f (39+ datasets).
• Page a676e8f07209a3be, slug uc6-pseudobulk-celltype-de-covid19 (/published/page?id=a676e8f07209a3be). Honest banner + 6 sections, all genuine on-graph directives (history_dataset_as_image / _as_table / _collection_display / _embedded).
• MVP workflow UC6_pseudobulk_edgeR_MVP.ga (20 steps; version-migrated shipped IWC workflow; re-import round-trips clean, invocation produced fixture-matching DEG).
• Figures (usecases dir): uc6_report_full.png (whole page), uc6_report_top.png (headline + banner + QC), uc6_pseudobulk_qc.png, uc6_platelet_volcano_artifact.png.

Data structure (verified, gates the stretch)

Source AnnData file.h5ad (548 MB, Zenodo 13929549). obs: cell_type, individual, disease (normal / COVID-19), raw counts layer, gene_symbol var. 7 individuals: Control_#1/2/3 (normal); SARS_CoV2_pos_Mild, Severe_#1/2/3 (COVID). 6 cell types · 34 pseudobulk samples. Per-(cell type × disease) sample counts:

samples_metadata carries only the disease factor — cell_type is encoded only in the sample-name suffix {individual}_{cell_type}. Suffix parsing is unambiguous (all 34 samples resolved; cell-type suffixes are mutually non-overlapping).

Pipeline as driven

1. Upload h5ad via /api/tools/fetch (Zenodo CDN, fast).
2. MVP — version-migrate shipped .ga (see drift below), import, invoke with the test yml params (groupby=cell_type, sample_key=individual, layer=counts, factor=disease, formula=~ 0 + disease, gene_symbol). Volcano step dropped from the workflow (major version break) and run separately.
3. Per-cell-type subset — for each cell type, tp_awk_tool keeps the gene column + matrix columns with header suffix _{ct}$ → 6 count matrices; same suffix-filter on the factor file → 6 factor files. Assemble two paired list collections (id = cell_type).
4. Mapped edgeR — edger over (matrices, factors) collections, broadcasting the shared genes_metadata + contrast file → list:list of 6 per-cell-type DEG tables.
5. Synthesis (on-graph) — map tp_awk over the flat DEG list to (a) count FDR<0.05 rows and (b) emit sig-gene names; collapse_dataset (add_name=cell_type) → responder ranking table and sig-genes-by-cell-type table.
6. Volcanoes — volcanoplot 4.0.3 on the pooled table and mapped over the per-cell-type DEG collection (PDF) → graphicsmagick_image_convert PDF→PNG for inline Page images.

Version drift handled (shipped pins → installed)

The shipped .ga pins exact versions not installed. Patched tool_id+content_id+ tool_version per step and filled new optional params:

• decoupler_pseudobulk galaxy8→galaxy10 — added num_pseudo_replicates, seed (optional → "").
• tp_replace_in_line/column, tp_awk_tool 9.3+galaxy1→9.5+galaxy3 — added per-repeat sed_options ("").
• edger galaxy5→galaxy7; collection_element_identifiers 0.0.2→0.0.3.
• volcanoplot 0.0.6→4.0.3+galaxy0 — full form restructure (with_header|{header,fdr_col, pval_col,lfc_col,label_col}, labels|labels); dropped from workflow, run standalone.

Gotcha: import keys off content_id, not just tool_id — patch both or the version error persists.

What I’d change / caveats

• Power is the story. 7 donors split 6 ways is below the threshold for per-cell-type pseudobulk DE. Even 3v3/3v4 yields 0 FDR hits. The responder ranking must carry per-arm counts (it does, in prose) — the lone non-zero entry being the smallest stratum is the tell.
• Platelet contamination — anucleate platelets carry ambient RBC RNA; HBB/HBA dominate. A real analysis would drop platelet (and likely erythroid) before per-cell-type DE, or add a contamination/ambient-RNA filter.
• Generic on-graph split unavailable. anndata_manipulate / scanpy_filter (the tutorial’s split-by-obs path) and any transpose tool are not installed, so the clean “split AnnData → collection → map the whole workflow” approach wasn’t possible; used awk-per-celltype column subsetting + API-assembled collections instead. Because the collections were API-built (no creating job), auto-extraction can’t reconstruct the stretch as one clean .ga — hence MVP .ga shipped + stretch documented as recipe.
• volcanoplot 4.0.3 emits PDF only; history_dataset_as_pdf renders as a live <embed> (blank in headless capture) → converted PDF→PNG on-graph for figures.
• as_table widgets are scroll-truncated to ~3 rows in static screenshots; live page is interactive, prose carries full numbers.

Key IDs

History b2486a20bc56b90f; input h5ad 709ee9fb7b09a260 (hid1). MVP: pseudobulk QC aaa9bb27aeb7b84b (hid5), filter 300f24fa15eda72e (hid6), pooled DEG 6875c703d2301e12 (hid15), volcano-table 07156755f40944e6 (hid18), report 28e1382d92a13bc5 (hid14), MVP volcano PNG 25352d0f50169864. Stretch: matrices coll 76fc6a61d2847f9c, factors 599c929992aefd6c, flat DEG 75c5bc32ad0ce8b8, mapped DEG list:list 4011851c940e7469, responder ranking 45bc061a4d279b5b, sig-genes 7972115ae4a0dcf6, volcano PNG coll dba3bb6dfae451b8 (platelet fb5da8d37e7135b8, monocyte fda86291ff9ae1d6).