CCREPE

Compositionality Corrected by REnormalization and PErmutation

What is CCREPE?

CCREPE (Compositionality Corrected by REnormalization and PErmutation) is an R package for measuring and assessing the statistical significance of correlations in compositional microbiome data. Standard correlation methods (Pearson, Spearman) produce biased results with compositional data because the values sum to a constant. CCREPE corrects for this bias.

CCREPE answers the question: β€œWhich pairs of microbial features are truly correlated with each other, accounting for the compositional nature of microbiome data?”

When to Use CCREPE

Use CCREPE when you want to:

  • Identify positively or negatively co-occurring microbial taxa
  • Build microbial co-occurrence networks
  • Correlate microbial features with each other or with metabolites
  • Correct for compositional bias in standard correlation analyses
Note

Other tools for microbiome correlation analysis include SpiecEasi and flashWeave. CCREPE’s approach is based on NC (North Carolina) renormalization, which is especially suited to relative abundance data.

Installation

From GitHub

devtools::install_github("biobakery/ccrepe")

Basic Usage

library(ccrepe)

# Load species abundance table
# Rows = samples, columns = species
abundance <- read.table("species_abundance.tsv",
                         sep = "\t",
                         header = TRUE,
                         row.names = 1)

# Run CCREPE with default settings (Spearman correlation)
ccrepe_results <- ccrepe(
  x     = abundance,
  sim.score = cor,          # similarity function
  iterations = 1000,        # number of permutations
  min.subj = 10             # minimum samples for inclusion
)

Cross-dataset correlation

# Correlate species abundance with metabolite abundance
metabolites <- read.table("metabolites.tsv",
                           sep = "\t",
                           header = TRUE,
                           row.names = 1)

ccrepe_cross <- ccrepe(
  x = abundance,
  y = metabolites,          # second dataset
  sim.score = cor,
  iterations = 1000
)

Output

CCREPE returns a list of matrices:

Output Description
sim.score Correlation coefficient matrix
p.values Raw p-values for each correlation
q.values FDR-adjusted p-values (Benjamini-Hochberg)
z.stat Z-statistics

Accessing results

# Get significant correlations (q < 0.05)
sig_pairs <- which(ccrepe_results$q.values < 0.05, arr.ind = TRUE)

# Build a table of significant correlations
library(tibble)
sig_table <- tibble(
  feature1 = rownames(ccrepe_results$q.values)[sig_pairs[, 1]],
  feature2 = colnames(ccrepe_results$q.values)[sig_pairs[, 2]],
  correlation = ccrepe_results$sim.score[sig_pairs],
  q_value = ccrepe_results$q.values[sig_pairs]
)

Custom Similarity Functions

CCREPE supports any custom similarity or distance function:

# Use a custom similarity metric (e.g., Mutual Information)
library(infotheo)

mi_score <- function(x, y) {
  mutinformation(discretize(x), discretize(y))
}

ccrepe_mi <- ccrepe(
  x         = abundance,
  sim.score = mi_score,
  iterations = 1000
)

Tips & Gotchas

Warning

Compositionality β€” Never use standard Pearson or Spearman correlations directly on relative abundance data. The compositional constraint (all features sum to 1) creates spurious negative correlations. CCREPE corrects for this.

Tip

More iterations = more accurate p-values β€” The default 1000 iterations is a minimum. Use 10,000+ for publication-quality results, at the cost of longer computation time.

Tip

Filter rare features β€” Remove taxa present in fewer than 10% of samples before running CCREPE to reduce multiple testing burden and improve statistical power.

Tip

Visualize results as a network β€” Use the igraph or ggraph R packages to visualize significant CCREPE correlations as a co-occurrence network.

Further Reading