Deconvolution of Bulk RNA-seq DEGs Using scRNA-seq Reference

This repository contains the analysis pipeline used to deconvolve bulk-tissue RNA-seq differential expression results using a single-cell RNA-seq (scRNA-seq) reference dataset, enabling identification of cell-type–specific gene expression signatures underlying bulk transcriptomic changes.

The workflow integrates:

• Bulk RNA-seq DESeq2 results
• A curated scRNA-seq reference from mouse cortex
• Dimensionality reduction and correlation-based analyses

---

Methodological Background and References

The analysis pipeline implemented in this repository is based on the reference profile–based deconvolution framework described by Marquez-Galera et al. (2022), which leverages publicly available single-cell RNA-seq datasets to interpret cell-type–specific contributions to gene expression signatures derived from bulk-tissue RNA-seq. In this approach, differentially expressed gene (DEG) lists obtained from bulk RNA-seq are projected onto a curated scRNA-seq reference, and linear dimensionality reduction together with correlation-based clustering is used to identify cell-type–specific gene modules embedded within bulk transcriptional signatures.

This strategy was originally applied to dissect sublayer- and cell-type–specific transcriptional changes in the hippocampal CA1 region under physiological and epileptic conditions (Cid et al., 2021). That work demonstrated that bulk RNA-seq signatures reflect a convolution of cell-type–specific expression programs and changes in cellular composition, and showed that reactive microglial and glial gene signatures can be unmasked by integrating bulk RNA-seq DEGs with single-cell reference data, underscoring the importance of cellular heterogeneity in the interpretation of bulk transcriptomic analyses.

The single-cell RNA-seq reference used in this pipeline is derived from the Mouse Whole Cortex and Hippocampus SMART-seq dataset generated by the Allen Institute for Brain Science and described by Yao et al. (2021), which provides a comprehensive transcriptomic taxonomy of neuronal and non-neuronal cell types across the mouse isocortex and hippocampal formation.

The present implementation adapts these published frameworks to cortical bulk RNA-seq data, following the same conceptual and analytical principles while allowing flexible subsetting of the scRNA-seq reference and dynamic visualization of cell-type–specific DEG signatures. Consistent with the original studies, this pipeline is intended for qualitative deconvolution and hypothesis generation rather than quantitative estimation of cell-type proportions.

---

Overview of the Analysis Pipeline

The analysis is organized into four main steps:

1. Identification and filtering of bulk RNA-seq DEGs
2. Construction of the scRNA-seq reference object
3. Subsetting and preprocessing of the scRNA-seq reference to match bulk tissue composition
4. Deconvolution of bulk DEGs using the scRNA-seq reference

Each step is implemented as a standalone, reproducible script.

---

Step 1: Load and Filter Bulk RNA-seq Differentially Expressed Genes (DEGs)

Purpose:
Identify statistically significant DEGs from bulk RNA-seq and prepare ranked gene lists for downstream integration with scRNA-seq data.

Key operations:

• Load DESeq2 results table
• Remove genes without valid gene symbols
• Filter by adjusted p-value (padj < 0.1)
• Rank genes by log fold change
• Split genes by direction of regulation (control vs experimental)
• Select the top 250 DEGs per condition

Inputs:

• S1_Table_DESeq.csv (DESeq2 results)

Outputs:

• ctrl_DEGs: top downregulated genes
• c4_DEGs: top upregulated genes

---

Step 2: Build the scRNA-seq Reference Dataset

Purpose:
Load raw single-cell gene expression data and metadata, and construct a Seurat object for downstream analysis.

Key operations:

• Load cell metadata and gene expression matrix
• Convert data into Seurat-compatible format
• Create a Seurat object without initial filtering
• Retain all genes and cells to preserve reference completeness

Inputs:

• matrix.csv (gene expression counts)
• metadata.csv (cell annotations)

Outputs:

• sc_data: Seurat object containing the scRNA-seq reference dataset

---

Step 3: Subset scRNA-seq Reference to Match Bulk Tissue Composition

Purpose:
Restrict the scRNA-seq reference dataset to cortical cell populations that are biologically relevant to the bulk-tissue RNA-seq experiment.

Key operations:

• Inspect region and subclass metadata distributions
• Subset cells based on neocortical region labels
• Remove hippocampal and non-isocortical subclasses
• Set cell-type subclass labels as active identities
• Normalize, scale, and identify variable genes
• Perform PCA, t-SNE, and UMAP
• Dynamically generate cell-type–adaptive color palettes
• Save dimensionality reduction plots

Inputs:

• sc_data from Step 2

Outputs:

• Filtered and normalized scRNA-seq reference
• PCA, t-SNE, and UMAP plots (.png)
• Extracted plot legends for figure assembly

---

Step 4: Deconvolution of Bulk DEGs Using scRNA-seq Reference

Purpose:
Interpret bulk RNA-seq DEGs by identifying cell-type–specific gene expression patterns using the scRNA-seq reference.

Key operations:

• Intersect bulk DEGs with scRNA-seq gene universe
• Use equally sized DEG sets for comparability
• Scale scRNA-seq expression for DEG genes only
• Perform PCA using DEG-driven expression
• Identify genes contributing to DEG-associated variance
• Compute gene–gene Pearson correlation matrices
• Perform hierarchical clustering
• Visualize correlation structure using heatmaps

Inputs:

• ctrl_DEGs / c4_DEGs from Step 1
• sc_data from Step 3

Outputs:

• DEG-driven PCA plots
• Correlation heatmaps (.pdf)
• Cell-type–specific gene signature visualizations

---

Software and R Packages

This pipeline was developed and tested using:

• R (≥ 4.2)
• Seurat
• data.table
• ggplot2
• cowplot
• patchwork
• gplots

Install required packages using:

install.packages(c("data.table", "ggplot2", "cowplot", "patchwork", "gplots"))
install.packages("Seurat")