Signature Matrices and Cross-Signature Deconvolution in deconvMe

Background: Reference-Based vs Reference-Free Deconvolution

Cell-type deconvolution methods can be broadly categorized into two main approaches:

Reference-Based Methods

Reference-based methods rely on signature matrices (also called reference matrices or basis matrices) to estimate the proportions of different cell types in a heterogeneous DNA methylation sample. A signature matrix is a table where each row corresponds to a CpG site (or region), and each column corresponds to a cell type. The entries represent the expected methylation value (often as beta values) for each CpG in each pure cell type. Signature matrices are typically constructed from reference datasets where DNA methylation has been measured in sorted or purified cell populations. The choice of CpGs and the quality of the reference data are critical for accurate deconvolution.

Reference-Free Methods

Reference-free methods attempt to deconvolve cell-type proportions without requiring prior knowledge of cell-type-specific methylation signatures. These methods use computational approaches such as non-negative matrix factorization (NMF) or independent component analysis (ICA) to identify latent cell-type components directly from the mixed methylation data.

Why deconvMe Focuses on Reference-Based Methods

deconvMe exclusively implements reference-based deconvolution methods because they generally provide better performance and more reliable results compared to reference-free approaches. Reference-based methods offer several advantages:

• Higher accuracy: Pre-defined signature matrices based on purified cell populations provide more reliable cell-type identification
• Better interpretability: Results can be directly mapped to known cell types with biological meaning
• Consistent results: The same signature matrix applied to different datasets produces comparable results
• Validation capability: Results can be validated against known cell-type proportions when available

Accessing Signature Matrices in deconvMe

library(deconvMe)
library(dplyr)
library(tidyr)
library(ggplot2)
options(matrixStats.useNames.NA = "deprecated")

# Load example data
library(minfiData)
methyl_set <- minfiData::MsetEx
beta_matrix <- minfi::getBeta(minfi::ratioConvert(methyl_set))

All signature matrices follow a unified design - a data.frame with a "CpGs" column and cell types as additional columns:

# Get signature matrices
epidish_sig <- get_epidish_signature_matrix("blood")
houseman_sig <- get_houseman_signature_matrix()
methatlas_sig <- get_methatlas_signature_matrix()
methylresolver_sig <- get_methylresolver_signature_matrix()

# Check structure
cat("EpiDISH:", dim(epidish_sig), "\n")

## EpiDISH: 333 8

cat("Houseman:", dim(houseman_sig), "\n")

## Houseman: 450 7

cat("MethAtlas:", dim(methatlas_sig), "\n")

## MethAtlas: 6105 26

cat("MethylResolver:", dim(methylresolver_sig), "\n")

## MethylResolver: 419 12

Custom CpG Subsets

You can use custom CpG subsets for deconvolution:

# Get overlapping CpGs between methods
custom_cpgs <- intersect(epidish_sig$CpGs, houseman_sig$CpGs)
cat("Overlapping CpGs:", length(custom_cpgs), "\n")

## Overlapping CpGs: 13

# Run with custom CpGs
result_custom <- run_epidish(beta_matrix, cpg_subset = custom_cpgs)

Cross-Signature Deconvolution

deconvMe supports cross-signature deconvolution, allowing you to use signature matrices from one method with the deconvolution algorithm of another method.

Supported Methods

Method	Cross-Signature Support
EpiDISH	✅ All methods
MethAtlas	✅ All methods
MethylResolver	✅ All methods
Houseman	❌ Not supported
MethylCC	❌ Not supported (no classical signature matrix)

Examples

# EpiDISH with Houseman signature
result_epidish_houseman <- run_epidish(beta_matrix, reference = houseman_sig)

# MethAtlas with EpiDISH signature
result_methatlas_epidish <- run_methatlas(beta_matrix, reference = epidish_sig)

# MethylResolver with MethAtlas signature
result_mr_methatlas <- run_methylresolver(
  beta_matrix, 
  reference = methatlas_sig,
  alpha = 1
)

## Beginning LTS Deconvolution For This Mixture...
##   |                                                                              |                                                                      |   0%  |                                                                              |============                                                          |  17%  |                                                                              |=======================                                               |  33%  |                                                                              |===================================                                   |  50%  |                                                                              |===============================================                       |  67%  |                                                                              |==========================================================            |  83%  |                                                                              |======================================================================| 100%
## Completed LTS Deconvolution For This Mixture...

Comparing Results

# Compare EpiDISH with different signatures
result_epidish_blood <- run_epidish(beta_matrix, reference = "blood")$estF
result_epidish_houseman <- run_epidish(beta_matrix, reference = houseman_sig)$estF

# Rename cell types to unified names
colnames(result_epidish_blood) <- deconvMe::rename_cell_types(colnames(result_epidish_blood))
colnames(result_epidish_houseman) <- deconvMe::rename_cell_types(colnames(result_epidish_houseman))

result_epidish_blood_long <- result_epidish_blood |> 
  as.data.frame() |> 
  tibble::rownames_to_column('sample') |>
  tidyr::pivot_longer(-sample, names_to = 'celltype') |>
  dplyr::mutate('method'='EpiDISH')

result_epidish_houseman_long <- result_epidish_houseman |> 
  as.data.frame() |> 
  tibble::rownames_to_column('sample') |>
  tidyr::pivot_longer(-sample, names_to = 'celltype') |>
  dplyr::mutate('method'='Houseman')

results <- dplyr::bind_rows(result_epidish_blood_long, result_epidish_houseman_long) 

deconvMe::results_aggregated_boxplot(results) + 
  ggplot2::labs(fill = "signature")+
  ggtitle('Deconvolution with EpiDISH, with two different signatures')