Skip to contents

Infer a direct association network using Part Mutual Information

Source: R/network.R
pmi_network.Rd

Constructs a network where edges represent direct associations between variables. For each pair (i, j), PMI is computed conditioning on all other variables (full-order conditioning) or on a specified subset. Statistical significance is assessed via permutation testing.

Usage

pmi_network(
  data,
  method = c("binning", "knn"),
  bins = 5,
  k = 3,
  threshold = 0.05,
  n_permutations = 100,
  adjust = c("fdr", "bonferroni", "none"),
  mc_cores = 1,
  verbose = TRUE,
  ...
)

Arguments

data

Numeric matrix or data.frame (rows = samples, columns = variables). Variable/column names are used as node labels.

method

PMI estimation method: "binning" or "knn".

bins

Number of bins for binning estimator (default 5).

k

Number of neighbors for kNN estimator (default 3).

threshold

Significance threshold for p-values (default 0.05).

n_permutations

Number of permutations for significance testing (default 100). Use at least 1000 for published results.

adjust

Method for multiple testing correction: "none", "bonferroni", or "fdr" (default "fdr").

mc_cores

Number of cores for parallel computation (default 1). Windows users should use 1.

verbose

Print progress messages (default TRUE).

...

Additional arguments passed to pmi.

Value

A list with components:

adjacency

Adjacency matrix (binary, 0/1) of significant edges.

pmi_matrix

Matrix of PMI values for all pairs.

pvalue_matrix

Matrix of permutation p-values.

pvalue_adjusted

Matrix of adjusted p-values.

References

Zhao et al. (2016) PNAS 113(18):5130-5135.

Examples

set.seed(42)
n <- 50
g <- rnorm(n)
x <- g + rnorm(n, 0, 0.5)
y <- g + rnorm(n, 0, 0.5)
z <- x + rnorm(n, 0, 0.3)
w <- rnorm(n)
dat <- data.frame(X = x, Y = y, Z = z, W = w)
net <- pmi_network(dat, n_permutations = 20, threshold = 0.1)
#> Computing PMI for 4 variables...
#>   Pair 1/6 (17%)
#>   Pair 2/6 (33%)
#>   Pair 3/6 (50%)
#>   Pair 4/6 (67%)
#>   Pair 5/6 (83%)
#>   Pair 6/6 (100%)
#> Found 1 edges at threshold 0.100 (fdr corrected).
print(net$adjacency)
#>   X Y Z W
#> X 0 0 1 0
#> Y 0 0 0 0
#> Z 1 0 0 0
#> W 0 0 0 0