Dynamic Bayesian networks

1
Dynamic Bayesian networks

• Elinor Velasquez, Fang Qi,
• Jianfeng Zhu, Scott Christley,
• Yue Fan, Oliver Wienand

2
Outline

• Definition of Bayesian network (BN).
• Definition of dynamical Bayesian network (DBN).
• Software used.
• Results.

3
Definition of Bayesian network (BN)

• A directed acyclic graph whose nodes correspond
  to random variables, , and edges are
  conditional dependence relations between a node
  and its parents.
• The joint distribution for all variables is
  represented as the product of the conditional
  dependence relations.

4
Definition of dynamical Bayesian network

• Dynamical Bayesian network (DBN) models the
  stochastic evolution of a set of random variables
  over time.
• The problem of learning a DBN is understood as
  follows find a network graph that best matches a
  given dataset of time series.

5
Example of DBN

• The nodes represent the genes and gene products.
• The edges represent that the parent nodes
  regulate (promotion, inhibition) the child node.

6
Software used

• Banjo Bayesian Network Inference with Java
  Objects

http//www.cs.duke.edu/amink/software/banjo/
7
Software used

• Graphviz - Graph Visualization Software

http//www.graphviz.org
8
Results

• Used wild type and knock out data sets.
• Compared the simulated annealing versus the
  greedy methods.
• Simulated annealing method gave the better
  result.
• Wild type plus knock out data sets gave the
  better result.
• Better results meant optimized on the
  reproduction of the Segment Polarity Network.

9
Solid green edges match the Brandys Segment
Polarity Network. Dashed blue edges new
predictions not on Brandys network but match
H.G.Othmers Segment Polarity Network. Dotted red
edges new predictions not on either networks.
However, most edges indicate a protein regulating
its own gene prior biological knowledge could
exclude them. Arrows are promotion. Boxes are
inhibition. Weighted edges close to zero
indicate Banjo could not differentiate between
inhibition or promotion.