Using Bayesian Network to Analyze Expression Data With Bayesware Discover

1
Using Bayesian Network to Analyze Expression Data
With Bayesware Discover

• Course Bioinformatics
• By Fang Yu

2
1. Introduction of Bayesware

• Bayesware Discoverer is an automated modeling
  tool based on Bayesian networks. It transforms a
  database into a Bayesian network, by searching
  for the most probable model responsible for the
  observed data.
• The aim of Bayesware Discoverer is to provide a
  Knowledge Discovery tool able to extract reusable
  knowledge from databases, using sounds and
  accountable statistical methods.
• Website to download www.bayesware.com

3
2. Function of Bayesware

• Naive Bayes Classifier Create and learn a Naive
  Bayes Classifier, one of the most common form of
  supervised learning.
• Bayesian Network ClassifierCreate and learn a
  Bayesian Network Classifier, a form of supervised
  classification that does not assume attribute
  independence.
• Bayesian Network GeneratorCreate and learn a
  Bayesian Network, a form of unsupervised learning
  discovering dependency models in your data.
• Batch PredictionPredict the value of one or more
  variables for a database of cases.
• Cross ValidationEvaluate the predictive accuracy
  of of model.

4
3. Spellman Data Set

• 42 genes (each column represent a gene)
• 77 rows (observation in one column is observation
  on relevant genes)
• Data is composed of -1, 0, 1 according to the
  original observation value

5
4. Result With Bayesian Network (I)
6
5. Result With Bayesian Network (II)
7
6. Summary of Results from Bayesian Network

• Nodes 42
• Log-likelihood for whole model -2588
• Nodes independent of other nodes 3
• Independent Nodes YLR013W, YLR389C, YNC145W
• The whole model is separated into four groups
• 8 root nodes involved, the nodes except the
  independent nodes are YBR088C, YHR143W, YFR030W,
  YAR071W, YGR086C, YBL002W

8
7. Comparing with result of WINMINE

• With result form winmine with default setting
• More nodes are included (39 vs. 37)
• Both are separated in four groups, final
  structure is close but with minor difference
• Independent Nodes YLR459W, UJR137C, YLR013W,
  YLR389C, YNL145W.
• Result from WinMine has 9 root nodes, the root
  nodes for WinMine except independent nodes are
  YDR225W, YKL163W, YHR143W, YBR088C.

• With result from WinMine using higher complexity
• less nodes are involved (39 vs. 40)
• Simpler than the result of WinMine
• WinMine result have 4 root nodes YBR088C,
  YKL163W, YLR389C, YLR459W
• Result from WinMine has 9 root nodes, the root
  nodes in WinMine result except independent nodes
  are YDR225W, YKL163W, YHR143W, YBR088C.

9
8. Result from Cross validation

• Good Performance from viewpoint of prediction
• Overall Accuracy 41.8
• Individual Accuracy varies largely from 20 to
  93
• Eg Variable Yjr137c
• Correct 60.0
• Incorrect 17.0
• Accuracy 77.922
• Std. Dev 4.727

10
9. Advantage of Bayesware

• Easy to use
• Bayesware does not expecting the use to have any
  particular methodological background.
• Bayesware offers Graphic User Interface on two
  main elements databases and networks. Various
  statistic (including conditional probability,
  marginal probability, variance, factor, etc) for
  every node is given out after justifying the
  network using the marginal likelihood.
• High Flexibility
• Adding prior information Using the locally
  search algorithm to train subsets of all the
  possible Bayesian networks, Bayesware discover
  let the user to add their prior knowledge through
  identifying an order of evaluating variables in
  the database. The higher the order of a variable
  is located, the larger the number of variables
  will be tested as its possible parents.
• Incomplete Database With BC (Bound and Collapse)
  methodology, Bayesware Discoverer can learn
  conditional probabilities from possibly
  incomplete databases.
• Missing Data Each missing datum is replaced by a
  set of possible databases consistent with the
  available data.

11
10. Limitation of Bayesware

• The free version is a restricted version, which
  only can handle 700 cases.
• It can only deal with the discrete data, when the
  data is continuous, it need good decretization
  rule. The discretization rule offered in
  Bayesware sometimes may bring trouble in
  data-training.
• The prior information is partly added since the
  relation between nodes can not be specified
  directly but through the searching order.
• Observational data is collected with no design
  and answers are given out without knowing the
  question.
• No option for different complexity to train
  network
• No source code available.

12
11. Conclusion

• Bayesware is suitable for our analysis in
  expression data
• The result from Bayesware is close to what we get
  from other software.
• The different among results by using different
  software comes mainly from the searching
  algorithm.
• Thank You !