An Iterative Bayesian Updating Method for Biological Pathway Prediction

1
An Iterative Bayesian Updating Method for
Biological Pathway Prediction

• Corey Powell

2
Biological Pathway Prediction

• Input A list of genes that work together in a
  biological pathway.
• Goal Determine all genes in the biological
  pathway.

3
Desirable Properties for a Biological Pathway
Prediction Method

• The method should be able to integrate data from
  multiple data sources
• The method should give probabilistic assessments
  of pathway membership.
• The method should identify genes that might be
  misclassified as being in the pathway.

4
The Method

• Determine prior estimates for the probability
  that each gene is in the pathway.
• Repeat
• Determine score functions for each gene and
  data source.
• Compute positive and negative conditional
  distributions for the score function from each
  data source.
• Update the probability estimates for each gene
  using the positive and negative conditional
  distributions together with Bayes Theorem and
  the Naïve Bayes assumption.

5
Testing the Method

• The method was tested on the Ribosome, Oxidative
  Phosphorylation, Proteasome, and Cell Cycle
  pathways from KEGG, together with random
  pathways consisting of 25, 50, 75, and 100 genes.
• The pathway probabilities were estimated using
  1) gene coexpression data, 2) protein interaction
  data, and 3) both 1) and 2).
• The testing was based on a 5-fold cross
  validation, with a gene classified as being in
  the pathway if the probability was greater than
  50.

6
General Behavior of the Method

• Pathway probabilities tended to converge either
  to 0 or to 1. The convergence was very rapid for
  pathways that were well represented in the data.
• The method exploded if run for too many
  iterations.

7
Results

• For the random pathways, the method would rarely
  classify a gene as belonging to the pathway, and
  this classification was almost always incorrect.
• The performance of the method was best on the
  Ribosome, followed by the Proteasome, Oxidative
  Phosphorylation, and Cell Cycle. Combining gene
  coexpression data and protein interaction data
  resulted in sometimes dramatic improvements in
  precision.