UC6 / Cell-type-resolved pseudobulk DE (COVID-19) — interview input

Pulled-down GitHub issue used as the effective result of the INTERVIEW → GALAXY interview (the live interview mechanics are harness-owned and precede pipeline phase 1). Source: https://github.com/jmchilton/galaxy-brain/issues/29 Paired aspirational target: none yet — workflow not yet extracted from a Galaxy history.

Plain-language science (why this is interesting)

A single-cell RNA-seq experiment measures gene expression in thousands of individual cells from several people (here: some healthy “normal”, some “COVID-19”). Each cell carries two labels: which person it came from (individual) and what cell type it is (B cell, CD4 T cell, monocyte, …).

Pseudobulk = sum the raw counts of all cells of one type within one person into a single profile (“all B cells from patient 7” → one B-cell sample). This bridges single-cell data to trusted bulk statistics (edgeR), with people as the real replicates — the correct unit for a disease comparison (thousands of correlated cells from one person are not independent replicates).

Why per-cell-type DE matters: the interesting question is rarely “is gene X up in disease overall” — it’s “which cell types are actually mounting the response, and do they respond the same way?” In COVID you’d expect monocytes and some T-cell subsets to be transcriptionally inflamed while B cells barely move. Pooling all cell types (what the shipped workflow does) averages these together and can hide a strong cell-type-specific signal. Running DE separately per cell type asks the sharp question, then you can rank cell types by responsiveness and split genes into a shared core (e.g. interferon-stimulated genes lighting up everywhere) vs cell-type-specific sets. That decomposition is the scientific payoff and is invisible from a single pooled contrast.

Purpose

Demonstrate a Galaxy Notebook that goes beyond a single bulk contrast: take an annotated single-cell dataset, aggregate it into pseudobulk profiles, and ask which cell types actually respond to disease and which genes are a shared response vs cell-type-specific. Every embedded figure/table is a genuine on-graph Galaxy tool output; the reusable workflow is extracted from the history provenance graph.

Objective (MVP + stretch)

• MVP (no recomposition needed): Reproduce the shipped IWC pseudobulk→edgeR analysis as an on-graph narrative — pseudobulk aggregation (decoupler), the pooled normal vs COVID_19 edgeR contrast, ranked DEG table, and volcano — but present it single-cell-natively: show the pseudobulk QC plot, explain people-as-replicates, display DEG table + volcano as on-graph artifacts. Novel vs the bulk-DESeq2 vignette (adds the pseudobulk aggregation axis) and vs clustering (ends in statistics).
• Stretch (the substantial deviation — requires recomposition): Fan the analysis out per cell type — subset the pseudobulk count matrix by cell_type, map edgeR over the per-cell-type subsets as a collection, and synthesize a cross-cell-type comparison: a responder-ranking table (significant genes per cell type) and a shared-vs-cell-type-specific gene summary — all from on-graph collection operations.

Why this is a useful demo deviation

• vs the existing bulk DESeq2 vignette (count matrix → ranked table → one volcano, supervised two-condition contrast): adds the pseudobulk aggregation of single cells into per-(cell type × person) samples (a single-cell-specific transform that is itself an on-graph teaching artifact) and, in the stretch, a per-cell-type fan-out producing a collection of contrasts plus a cross-cell-type synthesis. Exercises Galaxy Notebooks’ collection-map-over story, not a single plot.
• vs a plain scanpy clustering tutorial (cluster → UMAP → marker dotplot): clustering only names populations; here we use those names to quantify disease response per cell type. The narrative ends in cross-cell-type statistics, not a UMAP.

Existing analysis anchors (real tool ids / paths)

Primary: iwc/workflows/scRNAseq/pseudobulk-worflow-decoupler-edger/pseudo-bulk_edgeR.ga

• ebi-gxa/decoupler_pseudobulk/decoupler_pseudobulk/1.4.0+galaxy8 (pseudobulk; outputs count_matrix, samples_metadata, genes_metadata, plot_output PNG, filter_by_expr_plot PNG)
• iuc/edger/edger/3.36.0+galaxy5 (DE; outputs outTables, outReport HTML)
• iuc/volcanoplot/volcanoplot/0.0.6
• supporting: iuc/column_remove_by_header/1.0, bgruening/text_processing/{tp_replace_in_line,tp_replace_in_column,tp_awk_tool}/9.3+galaxy1, bgruening/split_file_to_collection/0.5.2, iuc/collection_element_identifiers/0.0.2, param_value_from_file.

Upstream annotation reference (optional chapter): iwc/workflows/scRNAseq/scanpy-clustering/...Scanpy.ga — iuc/scanpy_* 1.10.2+galaxy*, iuc/anndata_* 0.10.9+galaxy0; produces louvain clusters + manual cell-type annotation (the kind of cell_type label the pseudobulk step consumes).

velocyto was evaluated and rejected as anchor: its IWC workflow outputs only a .loom of spliced/unspliced counts (no plotting tool, no scVelo) — no on-graph figures to embed.

Public data candidates (real shipped data + actual labels)

• Primary: Source AnnData file.h5ad — https://zenodo.org/records/13929549/files/Source%20AnnData%20file.h5ad (from pseudo-bulk_edgeR-tests.yml). obs columns: cell_type (groupby), individual (sample key / replicate), disease (factor, values normal and COVID_19), gene_symbol (var), raw counts layer. Test confirms a real DEG table (edgeR_normal-COVID_19, ~1430 lines). Verify exact set of cell_type values and per-cell-type sample counts before finalizing the stretch.
• Optional upstream illustration: scanpy clustering PBMC3k 10X (barcodes.tsv/genes.tsv/matrix.mtx, https://zenodo.org/record/3581213/files/...), annotated to CD4+ T, CD14+, B, CD8+ T, FCGR3A+, NK, Dendritic, Megakaryocytes — use only to show where cell-type labels come from.

Notebook workflow plan (numbered, on-graph)

1. Load the annotated h5ad; show obs structure (cell_type / individual / disease) — narrative + small on-graph inspection table.
2. Decoupler pseudobulk (groupby cell_type, sample_key individual, layer counts, factor disease): embed the on-graph pseudobulk QC plot and the filter-by-expression plot; show count_matrix head.
3. MVP contrast: sanitation chain → edgeR ~ 0 + disease, contrast normal-COVID_19; embed the on-graph DEG table and volcano.
4. Stretch fan-out: split the pseudobulk count_matrix columns by cell_type into a collection (on-graph column/split ops) → map edgeR over the per-cell-type subsets → per-cell-type DEG tables + volcanoes as a collection.
5. Cross-cell-type synthesis: from the per-cell-type DEG tables, build (a) a responder-ranking table (count of FDR-significant genes per cell type) and (b) a shared-vs-specific summary via on-graph set/column ops. Embed both.
6. Extract the reusable workflow by walking the provenance graph backward from the synthesis outputs.

Expected paper/demo artifacts

• Pseudobulk QC + filter plots (genuine decoupler PNGs).
• MVP: one DEG table + one volcano (pooled contrast).
• Stretch: a collection of per-cell-type volcanoes + DEG tables.
• A responder-ranking table and a shared-vs-cell-type-specific gene summary (on-graph).
• The extracted reusable workflow .ga.

Scope and risks (honest)

• Recomposition required for the headline (stretch). The shipped workflow runs ONE pooled ~ 0 + disease contrast across all cell types (single count_matrix → single edgeR run; the split step splits by disease-contrast, of which there’s only one). Per-cell-type DE needs new on-graph steps to subset the pseudobulk matrix by cell type and map edgeR over the subsets. The MVP avoids this and is still novel. (This is structurally the same “irreducible → split it” arc as the manuscript’s Vignette 3.)
• Per-cell-type statistical power. Some cell types may have too few individuals for stable edgeR fits; the responder ranking must carry sample-count caveats. Verify per-cell-type sample counts in the test h5ad.
• Matrix-subsetting mechanics. Splitting a wide pseudobulk matrix by cell-type column-group on-graph needs validation (confirm decoupler’s column naming encodes cell_type parseably).
• Version/data. Pin tool versions; confirm the Zenodo h5ad downloads and that disease has only the two expected levels.

Tasks

• Download Source AnnData file.h5ad; enumerate exact cell_type values and per-(cell_type × disease) individual counts.
• Run the shipped pseudobulk→edgeR workflow as-is; confirm on-graph outputs (pseudobulk plot, DEG table, volcano) render embeddably.
• Build the MVP notebook around the pooled contrast (single-cell framing, people-as-replicates).
• Prototype the stretch: on-graph subset of count_matrix by cell_type → collection → map edgeR over subsets.
• Build cross-cell-type synthesis (responder-ranking; shared-vs-specific via on-graph set/column ops).
• Verify per-cell-type sample sizes are adequate; document low-power caveats.
• Extract reusable workflow from provenance graph; validate round-trip.
• (Optional) Add an upstream “where annotations come from” chapter referencing the scanpy clustering workflow.
• Confirm all displayed artifacts are genuine on-graph tool outputs (no pasted figures).