Methods blog · Bayesian & Network Inference

BNMCMC

BNMCMC: inferring and visualising Bayesian networks with modern MCMC samplers

Software that runs the Neighbourhood, Hit-and-Run and Metropolis–Hastings samplers for discrete Bayesian network structure learning, with adaptive exploration and interactive network visualisation.

#software#Bayesian networks#MCMC#Hit-and-Run#Neighbourhood sampler#visualisation

Bayesian networks (BNs) are widely used to model biological networks from experimental data, and many packages can infer BN structure. In practice, though, the inference often gets trapped in local optima. Our methods work showed that the Neighbourhood sampler (NS) and the Hit-and-Run (HAR) sampler avoid this far better than the standard Metropolis–Hastings (MH) sampler. BNMCMC puts those samplers in practitioners’ hands.

What BNMCMC does

  • Three samplers in one place: NS, HAR and MH, all targeting the posterior over BN structures under a discrete Bayesian model for categorical data.
  • Adaptive exploration: the samplers exploit adaptive techniques to explore BN space efficiently and to evaluate the posterior probability of candidate networks, which matters for large-scale network inference.
  • Visualisation: a web front-end draws the inferred networks. The planned server version uses a physics-based (force-directed) BN layout in D3.js.

A typical MCMC structure-learning workflow

  1. Prepare a discretised data matrix (samples × variables).
  2. Choose a sampler (NS or HAR for rugged posteriors; MH as a baseline) and the chain length.
  3. Run the chain, discarding burn-in.
  4. Summarise the sampled networks, for example as posterior edge probabilities or the highest-posterior network.
  5. Visualise the result.

Under the hood

BNMCMC is implemented in C#.NET and ASP.NET with jQuery, JavaScript and D3.js for the interface. The standalone version, without the visualisation, comes with a user guide and an example simulation file.

Get it

The same samplers were used to learn causal models of perturbed TGF-β signalling in PathTurbEr.

Read the paper

  1. BNMCMC: a software for inferring and visualizing Bayesian networks using MCMC methods

    AKM Azad, Salem A. Alyami, Jonathan Macgregor Keith

    bioRxiv 414953 · 2018

    PreprintFirst author OA

BibTeX