Important
This repo is no longer maintained and its contents have been superseded by functions in https://github.com/mkoslovsky/MicroBVS

Bayesian variable selection for Dirichlet-Multinomial regression. For $n$ samples, let $Y$ be an $n \times q$ matrix with $q$ taxa. Let $X$ be an $n \times p$ matrix with $p$ covariates. By applying spike-and-slab priors to the Dirichlet-Multinomial regression coefficients, we obtain concise taxa/covariate associations. The methodology is further described in the manuscript:

Wadsworth, W. D., Argiento, R., Guindani, M., Galloway-Pena, J., Samuel, S. A., & Vannucci, M. (2016). An Integrative Bayesian Dirichlet-Multinomial Regression Model for the Analysis of Taxonomic Abundances in Microbiome data.

The code has been updated on 10/02/17 with an option to perform a stochastic search algorithm for Bayesian variable selection instead of the Gibbs sampler. The stochastic search variable selection approach provides gains in computational speed.

Updated on 07/28/22: The code is no longer actively maintained and has been superseded by functions in https://github.com/mkoslovsky/MicroBVS (see vignettes).

The repository contains

1. C code implementing an MCMC sampler for Dirichlet-Multinomial Bayesian Variable Selection (dmbvs) using spike-and-slab priors,

2. R code to wrap and run the sampler from within R,

3. a function for simulating data,

4. and a "start-to-finish" script (example_analysis_script.R) demonstrating usage of the code. This script gives reasonable default settings for the hyperparameters and MCMC parameters for the example simulated data. Settings may change for other data.

• The R code requires the dirmult and MASS packages. Please ensure those are installed first.

• The C code relies on the GNU Scientific Library (GSL). GSL must be installed and modifications may need to be made to the 'library' and 'include' paths in the Makefile.

• The main C file (dmbvs.c), as well as the Makefile to be used for compilation of the code, can be found under the directory code/.

• On Linux and Mac the C code may be compiled with the Makefile from R using:

# must be in package's root directory
setwd("dmbvs")
system("cd code; make")

• If compilation has been successful there will be an executable called dmbvs.x the code/ directory. Data may be simulated and the MCMC code run using:

source(file.path("code", "wrapper.R"))
source(file.path("code", "helper_functions.R"))
simdata = simulate_dirichlet_multinomial_regression(n_obs = 100, n_vars = 50,
                                                    n_taxa = 50, n_relevant_vars = 5,
                                                    n_relevant_taxa = 5)
results = dmbvs(XX = simdata$XX[,-1], YY = simdata$YY, 
                intercept_variance = 10, slab_variance = 10, 
                bb_alpha = 0.02, bb_beta = 1.98, GG = 1100L, thin = 10L, burn = 100L,
                exec = file.path(".", "code", "dmbvs.x"), output_location = ".")

• Tested on:
  • Mac with R version 3.2.2 and GSL version 2.1 with the Apple LLVM version 7.0.2 compiler
  • Red Hat Linux 6.6 with R version 3.1.2 and GSL version 2.1 and gcc version 5.2.0