scCompoundDE

Overview

scCompoundDE decomposes pseudo-bulk differential expression (DE) signals into two orthogonal components:

• Transcriptional — cell-intrinsic expression changes: the gene would be DE even if subtype proportions were held fixed.
• Compositional — signal arising from a shift in the relative abundance of cell subtypes: the gene appears DE only because high-expressing subtypes became more common.

Standard pseudo-bulk tools (DESeq2, edgeR, limma-voom) cannot distinguish between these two sources. Every published pseudo-bulk DE result potentially conflates them.

The problem

Donor A (disease):  80% exhausted T cells (high PDCD1, TOX)
Donor B (healthy):  80% naive T cells     (high IL7R, CCR7)

Standard pseudo-bulk DE reports:
  → PDCD1  UP    p = 1e-8   ← but is this transcriptional or compositional?
  → IL7R   DOWN  p = 1e-6   ← same question

scCompoundDE answers:
  → PDCD1  TC_ratio = 0.12  → compositional artefact ⚠
  → IL7R   TC_ratio = 0.09  → compositional artefact ⚠
  → IFNG   TC_ratio = 0.91  → transcriptional signal ✓

The TC_ratio score

For each gene, compoundDE() returns a TC_ratio in [0, 1]:

TC_ratio	Classification	Meaning
≥ 0.8	Transcriptional	Real cell-intrinsic biology
≤ 0.2	Compositional	Subtype proportion artefact
0.2–0.8	Mixed	Both mechanisms contribute

Installation

# From Bioconductor
BiocManager::install("scCompoundDE")

# Development version from GitHub
BiocManager::install("SubhadipJana1409/scCompoundDE")

Quick start

library(scCompoundDE)

result <- compoundDE(
    sce,
    broad_type   = "T_cell",      # broad population to decompose
    subtype_col  = "cell_subtype",# fine-grained subtype labels
    broad_col    = "cell_type",   # broad label column
    donor        = "donor",
    condition    = "condition",
    min_cells    = 10L,
    min_subtypes = 2L
)

result
# CDEResult
#   Broad type     : T_cell
#   Subtypes       : Exhausted, Naive, Memory
#   Genes tested   : 8,342
#   Significant    : 1,204 (adj.P.Val < 0.05)
#   ── Decomposition ──────────────────
#   Transcriptional: 831 genes (TC_ratio >= 0.8)
#   Compositional  : 203 genes (TC_ratio <= 0.2)
#   Mixed          : 170 genes

# Get only transcriptionally-driven genes
real_genes <- filterGenesBySource(result, source = "transcriptional")

# Visualise
plotDecomposition(result)  # T_score vs C_score scatter
plotTCRatio(result)        # TC_ratio histogram
plotProportion(result)     # subtype proportions per condition

Performance & Validation

We validated scCompoundDE on the widely used Kang et al. (2018) IFN-β PBMC dataset:

• Scale: 35,635 genes, 29,056 cells, 8 donors, 8 cell subtypes.

Biological Accuracy (The Ground Truth)

The algorithm correctly identified 9 out of 10 canonical interferon-stimulated genes (e.g., ISG15, IFIT1, MX1) and successfully classified them as transcriptionally driven. This confirms the method reliably isolates true, cell-intrinsic biological responses from confounding factors.

Artefact Detection

Out of 5,975 significant genes (FDR < 0.05), 547 genes were flagged as purely compositional artefacts (TC_ratio ≤ 0.2). These are genes that standard pseudo-bulk tools would mistakenly report as cell-intrinsic DEGs, but are actually just the result of subtype proportion shifts (e.g., an influx of a specific cell type) across conditions. Examples of these artefacts include ACTB, FTH1, GAPDH, and S100A8.

Decomposition scatter plot

Each gene's transcriptional (T) vs compositional (C) z-scores. Blue = transcriptional, red = compositional, orange = mixed.

TC_ratio distribution

Histogram of TC_ratio values across all significant genes. Vertical lines mark the classification thresholds (0.2 and 0.8).

Subtype proportion shifts

Mean subtype proportions per condition. Shifts here drive the compositional component of the DE signal.

The algorithm

1. Filter sparse donor-subtype combinations (min_cells).
2. Compute subtype proportions π(donor, subtype, condition).
3. Run a separate limma-voom DE model per subtype.
4. Run a broad pseudo-bulk DE (all subtypes collapsed).
5. Decompose each gene's logFC into:
   • T_g = Σ_k π̄_k · logFC_gk (proportion-weighted transcript shift)
   • C_g = Σ_k Δπ_k · μ_gk (expression-weighted proportion shift)
6. Z-score both components and compute TC_ratio_g = |T_z| / (|T_z| + |C_z|).
7. Classify genes as transcriptional / compositional / mixed.

Key functions

Function	Description
compoundDE()	Full decomposition pipeline → CDEResult
filterGenesBySource()	Extract gene lists by classification
plotDecomposition()	Scatter: T_score vs C_score per gene
plotProportion()	Stacked bar: subtype proportions per condition
plotTCRatio()	Histogram of TC_ratio with threshold lines

CDEResult S4 class

deTable(result)              # DataFrame with all stats + TC_ratio + source
subtypeProportions(result)   # matrix: samples × subtypes
subtypeDE(result)            # list of per-subtype DE DataFrames
tcRatio(result)              # named numeric: TC_ratio per gene

Citation

Jana S (2026). scCompoundDE: Compositional and Transcriptional Decomposition of Pseudo-Bulk Differential Expression for Single-Cell RNA-seq. R package version 0.99.0. https://github.com/SubhadipJana1409/scCompoundDE

License

MIT © Subhadip Jana