rgm: Random Graphical Models for data from multiple environments

rgm is an R package that performs joint Bayesian inference of multiple Gaussian (or Gaussian-copula) graphical models that share structure through a low-dimensional latent space. It is the reference implementation of the random graphical model of Vinciotti, Wit & Richter (2026, JABES).

When you have multivariate measurements collected at several environments (body sites, tissues, ecological habitats, time periods, treatment groups, …) and you expect the underlying interaction networks to be related but not identical, rgm lets you estimate all of them jointly while quantifying how similar each pair of environments is.

Key features

• Joint inference across environments. Estimates per-environment graphs \(G^{(b)}\) and precisions \(\Omega_b\) simultaneously, with a structural prior that pools strength when environments are similar.
• Latent space of environments. Each environment is assigned a 2-D latent location whose posterior gives an interpretable similarity embedding (close ⇒ shared edges).
• Edge covariates. Optional edge-level covariates X enter through a global probit coefficient \(\beta\) (e.g. taxonomic distance, anatomical proximity).
• Counts via copula. The Gaussian-copula extension (method = "gcgm") handles zero-inflated count data such as microbiome OTU tables.
• Full posterior. Returns posterior samples for graphs, precisions, intercepts, latent locations, and covariate effects, so you can quantify uncertainty on any derived quantity.

Installation

# from CRAN (once 1.1.0 is back online — submitted 2026-05)
install.packages("rgm")

# or development version from GitHub
install.packages("remotes")
remotes::install_github("franciscorichter/rgm", build_vignettes = TRUE)

Quick start

library(rgm)

# Simulate B=8 related environments with p=20 nodes, n=200 obs each.
sim <- sim.rgm(n = 200, p = 20, B = 8)

# Fit RGM. Defaults: empty initial graph, GGM likelihood.
fit <- rgm(data = sim$data, iter = 2000, burnin = 500, method = "ggm")

# Posterior edge probabilities, n.edge x B
edge_prob <- apply(fit$sample.graphs, c(1, 2), mean)

# Posterior-mean latent locations
cloc <- apply(fit$sample.loc, c(1, 2), mean)

# Diagnostic plots (all-in-one)
plots <- post_processing_rgm(simulated_data = sim, results = fit)
plots$rgm_recovery
plots$edge_prob

For count data (microbiome, single-cell) use method = "gcgm" and supply the discrete-Weibull marginal parameters via gcgm.dwpar. See the vignette for a full walkthrough:

vignette("rgm")

What’s new in 1.1.0

• Drops the huge dependency. The default initial graph is now empty; pass initial.graphs = to keep a graphical-lasso warm start of your own.
• New exported function post_processing_rgm() returning a set of ggplot diagnostics.
• Cleaner mvtnorm import; namespace regenerated; build artifacts removed from version control.

See NEWS.md for the full changelog.

Reference

Vinciotti, V., Wit, E. C., & Richter, F. (2026). Random Graphical Model of Microbiome Interactions in Related Environments. Journal of Agricultural, Biological and Environmental Statistics, 31(1), 46–59. https://doi.org/10.1007/s13253-024-00638-6

License

MIT (see LICENSE). Bug reports: https://github.com/franciscorichter/rgm/issues.