Inferring transcriptional networks II

1
Inferring transcriptional networks II
CSCI5461 Functional Genomics, Systems Biology
and Bioinformatics

• Rui Kuang and Chad Myers
• Department of Computer Science and Engineering
• University of Minnesota

2
Announcements

• Project proposals due Wed. 4/8!
• Get started on your projects! (see us if you
  need help with data, etc.)
• Remember to read paper for next time!
• M. Middendorf, E. Ziv, C. H. Wiggins. Inferring
  network mechanisms the Drosophila melanogaster
  protein interaction network. Proc Natl Acad Sci U
  S A., 102(9)31927.

3
Outline for today

• Finish inferring transcriptional networks from
  gene-expression data using Bayesian networks
• Paper discussion (Dynamic Bayes nets)
• Wrap-up of regulatory network inference, new
  directions, etc.

4
A reminder about gene transcription
Transcription Factors (proteins)
RNA polymerase (protein)
C T A A T G T . . .
5
3
3
5
G A T T A C A . . .
Binding sites
Transcription factors recognize transcription
factor binding sites and bind to them, forming a
complex. RNA polymerase binds the complex.
Protein-DNA interaction!
(eukaryotes)
5
Gene Transcription
Transcription Factors (proteins)
RNA polymerase (protein)
C T A A T G T . . .
5
3
3
5
G A T T A C A . . .
Transcription factors recognize transcription
factor binding sites and bind to them, forming a
complex. RNA polymerase binds the complex.
(eukaryotes)
6
Gene Transcription
G A T T A C A . . .
5
3
3
5
C T A A T G T . . .
The two strands are separated
(eukaryotes)
7
Gene Transcription
G A T T A C A . . .
5
3
G A U U A C A
3
5
C T A A T G T . . .
An RNA copy of the 5?3 sequence is created from
the 3?5 template until a termination sequence
is reached
(eukaryotes)
8
Inferring regulatory networks from expression data
Transcriptional regulatory network
Microarray data
9
The sprinkler Bayes net
Prior probability that it is cloudy
Conditional probability that it rains when its
cloudy

• What are the conditional independence relations
  implied?
• What is the joint probability?

Probability that grass is wet when the sprinkler
is off and it rains
10
Applying Bayes nets to transcriptional network
inference
Gene 2
Gene 4
Gene 3
Gene 1
Gene 5
Gene 6
Transcriptional regulatory network
Microarray data

• Challenges
• We dont know either the structure or the
  conditional probabilities!
• Expression data are noisy
• Even if expression data were perfect, they dont
  capture the complete picture (e.g.
  post-transcriptional/translational regulation)

11
Bayesian structure/parameter learning
Likelihood of data given model
Posterior probability of model given data
12
Structure scoring criteria
where
Assuming multinomial distribution with Dirichlet
prior (analytical solution)
-Heckerman,A Tutorial on Learning With Bayesian
Networks and Neapolitan,Learning Bayesian Networks
13
Dynamic Bayesian networks paper discussion
14
Question can we use prior knowledge of
transcription factors to help in network
inference?
Transcription Factors (proteins)
RNA polymerase (protein)
C T A A T G T . . .
5
3
3
5
G A T T A C A . . .
Binding sites
(eukaryotes)
15
Answer Yes!
Module networks identifying regulatory modules
and their condition-specific regulators from gene
expression data. Nature Genetics  34, 166 - 176
(2003)

• Basic idea
• start with known transcription factors
• simultaneously learn regulatory program and
  regulated module groups

16
Preprocessing

• Candidate regulators are chosen from among known
  and suspected transcription factors and signal
  transduction molecules. Informed choice of
  candidate regulators makes algorithm workable
  without selectivity, bad results are likely.

17
Module network procedure

• Genes are partitioned into modules and regulation
  program is sought for each module to explain gene
  expression in module.

18
Post-processing

• Enrichment of annotations for predicted modules
  are sought in literature enrichment of
  regulatory motifs sought within 500 base pairs
  upstream from genes

19
Regulator program
20
Learning module networks/regulatory programs

• Iteration
• Search for regulation program for each module
• Re-assign genes to the module whose program best
  predicts its behavior

21
An example result respiration carbon
regulation module
22
Validating module results

• Compare module gene set with GO terms
• 31 gt 50 functional coherence
• 4 lt 30 functional coherence
• Look for enriched upstream sequence elements
  (regulator binding sites)
• No sequence information was used for defining
  groups!

23
Module summary
24
(No Transcript)
25
Are these learned models predictive?

• Experimental validation
• Knock-out 3 predicted regulators, check
• Does it cause change under the predicted
  condition?
• Does it affect the predicted set of genes?
• Does the function of the predicted regulator
  match the prediction?

26
Results summary

• The method is able to accurately predict (1)
  functions for regulators
• (2) known transcription factor targets
• (3) the conditions under which regulation occurs
• What general principles about successful methods
  for inferring networks can we learn from this
  example?

27
New trends in transcriptional network inference

• Sub-structure learning
• Can we extend these models to infer complete
  networks over all genes?
• Basic idea learn small sub-networks and
  stitch together
• Incorporating perturbations into network
  inference process (more on this later)
• Models that leverage more prior knowledge (e.g.
  TF binding site info)

28
Experimental determination of protein-DNA
interactions
ChIP-chip Chromatin immunoprecipitation chip
(microarray)
(antibodies bind transcription factor of
interest)
(TF-bound sequences hybridized to microarray)
Simon et al., Cell 2001
29
Mapped transcription factor binding sites in
yeast (based on ChIP-chip)
Harbison C., Gordon B., et al. Nature 2004
30
Example MEDUSA- learning regulatory programs
from known TFs and binding sites
Kundaje A, Lianoglou S, Li X, Quigley D, Arias M,
Wiggins CH, Zhang L, Leslie C.Learning
regulatory programs that accurately predict
differential expression with MEDUSA. Ann N Y Acad
Sci. 2007 Dec1115178-202.
31
MEDUSA performance (yeast gene expression)
Evaluation can regulatory programs predict
up/down expression of held-out differentially
expressed genes
Kundaje A, Lianoglou S, Li X, Quigley D, Arias M,
Wiggins CH, Zhang L, Leslie C.Learning
regulatory programs that accurately predict
differential expression with MEDUSA. Ann N Y Acad
Sci. 2007 Dec1115178-202.
32
Summary of transcriptional network inference

• Learning network structure is hard! (model space
  gtgt data)
• Bayesian models provide a reasonable framework
  for learning these models
• Incorporating prior knowledge is key!
• Validation is and will continue to be an issue
  (very few gold standards)