Using Bayesian Networks to Analyze Gene Expression Data

Project Goal

Massive amounts of gene expression profiles are quickly accumulating. This is due to the development of biotechnologies such as array based hybridization. Our goal is to develop algorithms and computational tools that can extract meaningful information about gene regulation and function from this data. We use methods for learning Bayesian networks to recover the structure of regulatory interactions between the different genes.
We present here some preliminary results.

Presentations

Using Bayesian Networks to Analyze Gene Expression Data - an overview of our project.

A tutorial on Bayesian Networks

An interactive tour of our results

We present here the results of our learning methods on data from the Yeast cell cycle analysis project published by Spellman et al. (1998) in Molecular Biology of the Cell. We thank the lab at Stanford for making this data available, and the lab members for the courteous help they gave us.
We applied our methods to a data set of 800 genes that were found to be regulated by the cell cycle. These genes were clustered into 8 clusters in Spellman et al. (see their figure ). We learned our genetic network using no prior knowledge or assumptions ("Tabula Rasa"). We present an interactive tour of the learned networks .

All 800 genes

Instructions:

The learned genetic network is displayed in the upper left window. It is possible to scroll to different parts of the network.
Each node represents the expressed RNA level of some gene .
The edges represent relations between these genes. The edges are colored by level of confidence in the Markov relation (see below).

Place the mouse over a node to see the ORF name and protein name (if known).

Place the mouse over an edge to see the level of confidence for this relation.

Clicking on a node displays additional information about the ORF in the bottom frame.

This includes links to some tables showing the strength of relations between the ORF and other genes in the network.
They include:

Markov Relations - shows all genes that are direct relatives of the queried gene.

Order (before) - shows all genes that appear before the queried gene in the learned network. This implies they might be causally (partly) responsible for the expression level of this gene, perhaps through a chain of other genes.

Order (After) - shows all genes that appear after the queried gene in the learned network. This implies that their expression level might be causally affected by the level of this gene, again, perhaps through a chain of interactions involving other genes.

Edges - shows all genes that have edges common with the queried gene. The edge direction is marked as:

->OtherGene - if the edge points from the queried gene to the OtherGene

OtherGene-> - if the edge points from the OtherGene to the queried gene.

--OtherGene - if the direction of the edge can not be determined.

Those relations will appear in the right-hand frame. The numbers (between 0.0 - 1.0), as well as the size of the bars on the right, indicate the level of confidence in this relation.

There is also a link to the SGD entry of the queried ORF.

Papers

Using Bayesian Networks to Analyze Expression Data N. Friedman, I. Nachman and D. Pe'er (A technical report describing this work. Submitted to RECOMB 2000).

Learning Bayesian Network Structure from Massive Datasets: The "Sparse Candidate algorithm". N. Friedman, I. Nachman and D. Pe'er (UAI 99)

Data Analysis with Bayesian Networks: A Bootstrap Approach. N. Friedman, M. Goldszmidt, and A. Wyner. (UAI 99)

Who are we

Nir Friedman

Iftach Nachman

Dana Pe`er

In collaboration with

Michal Linial Life Sciences Hebrew University

Moises Goldszmidt SRI international