23
Computational discovery of gene modules and regulatory networks Ziv Bar-Joseph et al (2003) Presented By: Dan Baluta

Computational discovery of gene modules and regulatory networks

  • Upload
    joey

  • View
    41

  • Download
    0

Embed Size (px)

DESCRIPTION

Computational discovery of gene modules and regulatory networks. Ziv Bar-Joseph et al (2003). Presented By: Dan Baluta. Agenda. Introduction Goal of Paper Methods & Results Conclusions Critique Discussion. Introduction. Interest in figuring out regulatory networks - PowerPoint PPT Presentation

Citation preview

Page 1: Computational discovery of gene modules and regulatory networks

Computational discovery of gene modules and regulatory networksZiv Bar-Joseph et al (2003)

Presented By: Dan Baluta

Page 2: Computational discovery of gene modules and regulatory networks

Agenda

Introduction Goal of Paper Methods & Results Conclusions Critique Discussion

Page 3: Computational discovery of gene modules and regulatory networks

Introduction

Interest in figuring out regulatory networks Genome-wide EXPRESSION data sets DNA-binding data (LOCATION analysis) PROBLEMS!!!

Page 4: Computational discovery of gene modules and regulatory networks

Theory…

Integration of expression and location data ought to result in a more accurate assignment of genes to regulators, when compared to either types of data sets on their own.

Basic result of combining data is more information.

Page 5: Computational discovery of gene modules and regulatory networks

Primary Goal

Develop an algorithm that combines expression and location data to discover gene modules and regulatory networks.

Page 6: Computational discovery of gene modules and regulatory networks

Methods GRAM Algorithm

Genetic RegulAtory Modules As opposed to ‘GRM’ algorithm???

GRAM Validation Comparing results Chromatin-IP (CHIP) experiments MIPS category enrichment analysis DNA binding motif analysis

Targeted data analysis using GRAM Transcriptional regulation of rapamycin response

Page 7: Computational discovery of gene modules and regulatory networks

GRAM Algorithm

Binding Data Expression Data

Step 1: search all possible (pairwise) combinations of transcript regulators.

Pull out sets of genes that share binding transcriptional regulators.

STRINGENT BINDING CRITERIA USED (p < .001).

Step 2: Reduce the gene sets from Step 1, by filtering out all genes from

each set that do not have highly (positively) correlated expression levels. Reduced sets act as ‘seeds’

for gene ‘modules’.

Step 3: Revisit binding data and adds genes sharing binding transcriptional regulators to gene modules from Step

2, using RELAXED BINDING CRITERIA (p < .01).

Page 8: Computational discovery of gene modules and regulatory networks

Methods

GRAM applied to binding data for 106 transcription factors and 500 expression experiments from Saccharomyces cerevisiae.

List of 500 Expression Experiments

Page 9: Computational discovery of gene modules and regulatory networks

Results: GRAM Algorithm

106 gene modules found. Containing 655 distinct genes. Regulated by 68 transcription factors

(TFs).

Page 10: Computational discovery of gene modules and regulatory networks

Results: GRAM Algorithm

(~ 35%)

Page 11: Computational discovery of gene modules and regulatory networks

Validation: Comparing Results Picked up many more (2.5x) regulator-

gene interactions than binding data alone would have predicted.

How do we know these are not all false-positives?

Page 12: Computational discovery of gene modules and regulatory networks

Validation: CHIP Experiments Allows you to determine if a given gene

actually binds to a specific TF. Used IP experiments for Stb1 and 36

randomly chosen genes to characterize sensitivity and specificity.

Page 13: Computational discovery of gene modules and regulatory networks

Validation: CHIP Experiments

GRAM pulled out 3 TF-gene relationships that were… A) Validated by the IP results. B) NOT pulled out using binding data alone.

GRAM did NOT pull out TF-gene relationships that were not also validated by the IP results.

IP experiments indeed showed reduction in false negatives, and a lack of increase in false positives.

Page 14: Computational discovery of gene modules and regulatory networks

Validation: MIPS Categories

Gene modules ought to belong to same MIPS categories.

Gene modules derived using GRAM were 3X more likely to be enriched for genes in the same MIPS category than groups of genes derived from binding/location data alone.

Page 15: Computational discovery of gene modules and regulatory networks

Validation: DNA Binding Motifs

Genes linked to specific TFs ought to have the same binding motifs upstream of them as those associated with their TFs.

TRANSFAC database was used to determine whether genes in GRAM modules were more likely to be independently determined to be co-regulated vs. groups of genes from binding data alone.

Page 16: Computational discovery of gene modules and regulatory networks

Validation: DNA Binding Motifs

GRAM modules did indeed display higher percentage of genes containing the appropriate motif in the upstream region of DNA.

Further validation of GRAM algorithm.

Page 17: Computational discovery of gene modules and regulatory networks

Validation: Rapamycin Response

Rapamycin inhibits Tor kinase signaling Mimics nutrient starvation

Selected 14 TFs and performed genome-wide location analysis on them

Ran GRAM algorithm using location data plus expression data from literature

Page 18: Computational discovery of gene modules and regulatory networks

Validation: Rapamycin Response

Found 39 Gene Modules. 23 had significant MIPS category

enrichment. Added 192 pairs of gene-TF interactions

that location data alone missed. Generated 4 novel hypotheses.

Page 19: Computational discovery of gene modules and regulatory networks

Software Availability

Provide link to Java Application

QuickTime™ and aTIFF (LZW) decompressor

are needed to see this picture.

QuickTime™ and aTIFF (LZW) decompressor

are needed to see this picture.

QuickTime™ and aTIFF (LZW) decompressor

are needed to see this picture.

Page 20: Computational discovery of gene modules and regulatory networks

Conclusions

GRAM provides a means of discovering putative regulatory networks that other data sets cannot detect independently.

Integrating data sets provides us with more information than is available with either set independently.

Page 21: Computational discovery of gene modules and regulatory networks

Critiques

No solid measure of sensitivity and specificity. Argue that GRAM is more sensitive, but without specificity measure, how do we know that these are not all false-positives?

Looked for positive correlations as indicative of activation. Did not look at negatively correlated expression -- potentially an important loss of information.

Software does not appear to work OOB with sample data provided.

Page 22: Computational discovery of gene modules and regulatory networks

Discussion Topics Can this method be applied to other higher-level organisms?

Should it be? How can this model be improved to include more

information? e.g. can we look at negatively correlated expression data?

Should society consider other projects, on the scale of HGP, to extract more data on organisms in a standardized and systematic way?

Pairwise data is used in many cases in biology to infer system-level interactions, which in reality are multivariate. Is using this pair-wise data wise? Is there an alternative?

Could adding multiple species sets improve our results? i.e. Use metagenes instead of genes?

Page 23: Computational discovery of gene modules and regulatory networks

The End