Granulator

granulator

Workflows

Standard Workflow

Deconvolute bulk RNA-seq data using multiple reference profiles and benchmark the results against ground truth cell type proportions.

library(granulator)

# Load datasets for deconvolution of PBMC RNA-seq data
load_ABIS()

# Create a list of multiple signature matrices to test simultaneously
sigList <- list(
  ABIS_S0 = sigMatrix_ABIS_S0,
  ABIS_S1 = sigMatrix_ABIS_S1,
  ABIS_S2 = sigMatrix_ABIS_S2,
  ABIS_S3 = sigMatrix_ABIS_S3
)

# Plot signature matrix similarity matrices
plot_similarity(sigMatrix = sigList)

# Deconvolute bulk RNA-seq data using all available methods
decon <- deconvolute(m = bulkRNAseq_ABIS, sigMatrix = sigList)

# Plot cell type proportions for a specific model (e.g., SVR on ABIS_S0)
plot_proportions(deconvoluted = decon, method = 'svr', signature = 'ABIS_S0')

# Plot all estimated cell type proportions across methods and cell types
plot_deconvolute(deconvoluted = decon, scale = TRUE, labels = FALSE)

# Benchmark methods by correlating estimated to measured cell type proportions
bench <- benchmark(deconvoluted = decon, ground_truth = groundTruth_ABIS)

# Plot regression for a specific model
plot_regress(benchmarked = bench, method = 'svr', signature = 'ABIS_S0')

# Plot Pearson correlation between predictions and true proportions
plot_benchmark(benchmarked = bench, metric = 'pcc')

# Perform correlation analysis across selected methods
methods <- c('ols', 'nnls', 'qprog', 'rls', 'svr')
decon_sel <- deconvolute(bulkRNAseq_ABIS, list(ABIS_S2 = sigMatrix_ABIS_S2), methods)
correl <- correlate(deconvoluted = decon_sel)

# Plot correlation heatmap
plot_correlate(correlated = correl, method = "heatmap", legend = TRUE)

Note on inputs/outputs:

• Input: A bulk gene expression matrix (TPM values) and one or more cell-type reference signature matrices.
• Output: Estimated cell-type proportions per sample, benchmarking metrics (PCC, CCC, RMSE), and correlation plots comparing deconvolution methods.

When to Use

• Bulk Deconvolution: Estimating cell type proportions in heterogeneous bulk RNA-seq samples (e.g., PBMCs) using deconvolute().
• Method Benchmarking: Benchmarking multiple deconvolution algorithms (such as SVR, NNLS, OLS, RLS) against known ground truth proportions using benchmark().
• Signature Quality Assessment: Assessing the similarity and quality of reference signature matrices using plot_similarity().
• Consensus Analysis: Analyzing the consistency of predictions across different deconvolution methods when ground truth is unavailable using correlate().

When NOT to Use

• Differential Expression: For differential expression analysis directly, use DESeq2 or edgeR because granulator only infers cell type proportions (which can then be used as covariates in those packages).
• Single-Cell Clustering: For single-cell clustering and cell-type annotation from scratch, use Seurat because granulator is designed for bulk deconvolution using pre-defined signatures.

Data Requirements

• Bulk Expression: A gene (rows) by sample (columns) matrix containing normalized expression values (e.g., TPM), which can be computed from raw counts using get_TPM().
• Reference Profiles: A gene (rows) by cell type (columns) matrix containing normalized cell-type specific expression values (e.g., sigMatrix_ABIS_S0).
• Ground Truth (optional for benchmarking): A sample (rows) by cell type (columns) matrix of measured cell-type percentages (e.g., from FACS).

Key Parameters

• m: Input bulk expression matrix in deconvolute().
• sigMatrix: Reference signature matrix or list of matrices in deconvolute().
• scale (TRUE): Logical indicating whether to scale proportions to standard scores in plot_deconvolute().
• metric ('pcc'): Evaluation metric to plot in plot_benchmark(); options include 'pcc', 'ccc', 'adj.r2', and 'rmse'.
• method ('heatmap'): Visualization style in plot_correlate().

Best Practices

• Normalize raw counts to TPM using get_TPM() before running deconvolution.
• Test multiple reference profiles at different cell-type resolutions (e.g., collapsing highly similar subtypes) to find the most stable deconvolution performance.
• Use the Condition Number of the signature matrix to evaluate its sensitivity to input data variability before running deconvolution.
• Scale estimated proportions to standard scores using scale = TRUE in plot_deconvolute() to compare relative changes across methods with different absolute scales.

Common Pitfalls

• High Collinearity: High collinearity between closely related cell types in the reference matrix leads to unstable predictions. Fix: Collapse highly similar cell types into broader categories (e.g., using sigMatrix_ABIS_S2 instead of sigMatrix_ABIS_S0).
• Negative Proportions: Negative estimated proportions returned by unconstrained methods. Fix: Use constrained methods like nnls or qprogwc, or refine the reference signature matrix.
• Mismatched Gene Identifiers: Mismatched gene identifiers between bulk RNA-seq and reference signature matrices. Fix: Ensure both matrices use the same gene symbol or Entrez ID format before calling deconvolute().

Alternatives

• DESeq2 for differential expression analysis.
• edgeR for differential expression analysis of digital gene expression data.
• limma for linear modeling of gene expression.
• Seurat for single-cell RNA-seq analysis and reference generation.

Citations

• Monaco, G. et al. (2019). RNA-Seq Signatures Normalized by mRNA Abundance Allow Absolute Deconvolution of Human Immune Cell Types. Cell Reports, 26(6), 1627-1640.
• Newman, A. M. et al. (2015). Robust enumeration of cell subsets from tissue expression profiles. Nature Methods, 12(5), 453-457.

References

• Homepage: https://bioconductor.org/packages/granulator
• Vignette: https://bioconductor.org/packages/release/bioc/vignettes/granulator/inst/doc/granulator.html