HughesBoone Journal Club

1
Hughes/Boone Journal Club

• Physical Network Models
• Yeang, Ideker, and Jaakkola
• JCB 2004

2
Background regulatory networks
Friedman et al, Hartemink et al (2001, 2002)
Idea Build Bayesian Network (BN) of expression
data, links in the network represent direct
dependencies. Problem What genes are the
regulators? Can only determine direction of
links in certain cases.
BN
Segal et al (2003) module networks
Idea Build BN/DT of expression data given a list
of regulators. Problem Only models feed-forward
regulation. Difficult to recover direction of
regulation. Difficult to connect modules
together.
Beer and Tavazoie (2004) regulatory programs
Idea Build BN of cluster membership given
promoter sequence. Problem Putative regulators
are motifs, not TFs. Difficult to connect the
clusters together.
3
Basic idea
Protein-DNA link (ChIP-chip)
Protein
Most parsimonious explanation
4
Basic idea
Given
Consistent explanations
Dg1, g4 down-regulated
Dg3, g4 up-regulated
Infer
g1 up-regulates g3, g3 up-regulates g4.
5
Method

• Search over all possible configurations of
  variables representing the
• presence (exists or not),
• direction (A a B or B a A),
• type (up- or down-regulation).
• of each edge in the skeleton graph to find the
  configuration that best explains the available
  data, i.e., has the highest score.

6
Data Sources

• Protein-DNA
• ChIP chip (Lee et al 2002)
• Protein-protein
• DIP (small and large-scale experiments)
• Knock-out effects
• Hughes compendium (Hughes et al 2000)

7
Scoring

• presence
• Protein-DNA Likelihood ratio derived from (Lee
  et al 2002) p-value. Thresholded.
• Protein-protein Criteria from (Deane et al 2001)
• presence in small-scale database,
• report by multiple sources,
• co-expression,
• binding of paralogs.
• Knock-out effects Likelihood ratio based on
  (Hughes et al 2000) p-value. Thresholded.
• direction and type
• Fit of values with knock-out effects.

• Notes on scheme
• biased against protein-DNA and protein-protein
  interactions,
• multiple explanations of same knock-out effect
  increase the score,
• uses fudge factors to avoid having to explain
  every effect.
• uses path length restriction (lt6 or lt4)

8
Search strategy

• Precomputes consistent explanations for each
  knock-out effect to make search more efficient
• Uses max-product approximation algorithm to make
  search tractable.

9
Basic idea
Given
Consistent explanations
Dg1, g4 down-regulated
10
Results mating response pathway
Using

• 60 genes
• 37 protein-DNA interactions
• 30 YPD protein-protein interactions
• 13 deletion mutants DSTE2, DSTE4, DSTE18, DFUS3,
  DSTE7, DSTE11, DSTE5, DSTE12, DKSS1, DSTE10,
  DSST2, DSIN3, DTUP1
• 149 pairwise knock-out interactions

11
Results mating response pathway
Variant part of network
Predicting knock-out effects
Invariant part of network
All 106 effects explained
12
Results large scale analysis
Invariant part of network
13
Results large scale analysis
Variant part of network
14
Evaluation
Assumptions of model are clear but limiting
Needs a lot of high-quality data, i.e., not very
modular.