Reconstructing Gene Networks

1
Reconstructing Gene Networks

• Presented by Andrew Darling
• Based on article
• Research Towards Reconstruction of Gene Networks
  from Expression Data by Supervised Learning
• - Soinov, Krestyaninova, Brazma

2
Outline

• Why study another microarray algorithm?
• Background info
• Methods
• Results
• Discussion
• Conclusion

3
Why study another microarray algorithm?

• Study of microarray data continues
• Still unclear on what the data means
• Still unclear on how the genome works
• Confirm existing knowledge about gene networks
  using existing datasets
• Proof of concept in a new algorithm using
  existing knowledge and datasets
• This algorithm actually explains its reasoning

4
Background information

• What is a gene network?
• What is supervised learning?
• What are decision trees / classifiers?
• Why use classifiers?

5
What is a gene network?

• A model of a genes affecting other genes
• What other genes affect a given gene
• How other genes affect a given gene
• Positive, negative, complicated
• Several model types graphs, nodes, edges
• Boolean ( on off )
• Bayesian network ( conditional probability )
• Differential equations ( derivatives, integrals )

6
Gene network - example
7
What is supervised learning?

• The paper was unclear on the subject
• Perhaps a reference to the type of algorithm used
  
• It may have involved human interaction with the
  software
• Possibly, the software produced the classifiers
  in the form of a decision tree, then users
  interpreted the output into classification rules

8
What are decision trees / classifiers?

• Acyclic directed graph - tree
• Each graph explains what other genes affect a
  specific gene
• Inner nodes are gene products of other genes
• Edges are thresholds of concentration of the gene
  products of the other genes rules of the tree
• Leaf nodes are effects on transcription of the
  specific gene
• Each graph is a classifier for a specific gene

9
Classifiers model of gene networks

• Expression of gene is function of transcription
• Transcription of gene is in discrete states
• Expressed more than average
• Expressed less than average
• Transcription state affected by amount of other
  gene products (expression of other genes)
• Use yeast cell cycle data to test algorithm and
  previous knowledge to judge accuracy

10
Why use classifiers?

• The products affecting a specific gene are listed
  in the tree
• Allows for continuous values for concentrations
• Each additional dataset refines the decision
  information
• Decision trees are easy to read and interpret

11
Classifier - example
12
Methods

• Use induction algorithm to generate decision
  trees
• Program called C4.5
• Apply program three ways
• Regulation of target gene as a function of other
  genes at same time (simultaneous)
• Regulation of target gene as a function of other
  genes at previous times (time delay)
• Regulation as a function of change of other genes
  (changes)

13
Results - given

• These genes
• and yeast datasets
• Spellman, Cho,
• Cdc28
• Alpha-factor

14
Results produced this
15
Results with this accuracy
16
Discussion

• Some concern about the accuracy between 70 and
  94 on systems with known interactions
• Does that imply that the microarray data is wrong
  or the algorithm is flawed?

17
Conclusions

• Decision trees and classifiers seem a better way
  to explain gene expression
• This paper did not do a good job of explaining
  how to make / use them
• Reference to the algorithm itself was almost
  specious