A New Dynamic Bayesian Network DBN Approach for Identifying Gene Regulatory Networks from Time Cours

1
A New Dynamic Bayesian Network (DBN) Approach for
Identifying Gene Regulatory Networks from Time
Course Microarray Data
By Min Zou and Suzanne Conzen

• Jim Vallandingham

2
Dynamic Bayesian Networks (DBN)

• For modeling time-series data
• Such as microarray data
• capture the fact that time flows forward
• Interested in how genes regulate each other over
  time

3
Dynamic Bayesian Networks (DBN)
4
Dynamic Bayesian Networks (DBN)
Time ?
5
Problems with DBNs

• Lack a way to determine biologically relevant
  transcriptional time lag
• Current methods assume same time lag for all
  potential regulator-target pairs
• Results in low accuracy of predicting gene
  relationships
• Excessive computational cost
• Prevents use of DBNs with large scale datasets

6
New DBN Method Improvements

• Determine biologically relevant transcriptional
  time lag
• Look at initial regulation of regulator and
  potential target to determine time lag
• Analyzed for each relationship
• Will improve relation predictions
• Reduce computational cost
• Only consider genes (up / down) regulated before
  or at the same time as potential target
• Reduces search space
• Reduces cost

7
General Outline of Method
8
Hypothetical Example

• Used to illustrate novel DBN approach
• 4 hypothetical genes A D
• 6 time points T1 T6
• Evenly spaced indicative of actual data sets.
• Process broken into 3 major steps

9
Hypothetical Example

• Step 1 Selection of potential regulators for
  each gene
• First , determine time points of changes in
  expression for each gene
• Used Thresholds to determine when regulation has
  occurred
• 1.2 fold ? up-regulation
• 0.7 fold ? down-regulation
• Find Potential Regulators
• Pick only genes with earlier or simultaneous
  changes as regulator candidates
• Used to reduce number of nodes considered

10
Hypothetical Example
Dynamic Expression profile for Gene D
11
Hypothetical Example
Dynamic Expression profiles for Genes A D
12
Hypothetical Example
13
Hypothetical Example

• Step 2 Estimation of biologically relevant
  transcriptional time lag
• Time between expression changes of potential
  regulator and target genes represents a
  biologically relevant time period
• Can vary from 0 (simultaneous) to many steps
• Using this time period should result in an
  increase of correct relationships

14
Hypothetical Example

• Step 2, cont
• Looking at D as target gene and A-C as potential
  regulators
• A 2 time units
• B 2 time units
• C 1 time unit
• Group potential regulators based on time lag

A B
C
15
Hypothetical Example
16
Hypothetical Example
17
Hypothetical Example

• Step 3 Gene regulatory network modeling
• Use DBN to predict gene regulatory network
• For DBN variables
• Use 2 if expression level is equal to or higher
  than average expression level over all time
  points
• Use 1 if expression level is lower than average
  level
• Focus of DBN is to predict correlation, not
  expression value for any given point

18
Hypothetical Example

• Step 3, cont
• Generate subgroups of groups of potential
  regulators based on user defined minimum and
  maximum regulators
• For Hypothetical Example
• Subsets for group AB ? A, B, A,B
• Assuming maximum 2, minimum 1
• Subset for group C ? C

19
Hypothetical Example

• Step 3, cont
• For each subset, using transcriptional time lag
  from step 2 to organize expression data into NxM
  matrix
• N number of potential regulators target
• T number of time points from original sampling
• t estimated transcriptional time lag
• From step 2
• M number of time points in the data matrix
• T t

20
Hypothetical Example

• Step 3, cont
• Expression value of potential regulators at time
  Tn are aligned with expression value of the
  target gene at time Tn t
• t may vary for the different expression data
  matrices

21
Hypothetical Example
1 below average 2 at or above average
22
Hypothetical Example

• Step 3, cont
• Matrix used to find conditional probabilities of
  the expression the of target gene in relation to
  its potential regulator gene(s)
• Pick subset of potential regulator(s) with
  highest log marginal likelihood score as the
  estimated regulator(s)

23
Hypothetical Example
Discrete Expression Values
Target D
Regulator A
Conditional probabilities
24
Biological Experimentation

• Comparison between their method and standard DBN
  from Murphys BNT Toolkit
• Used Chous yeast cell cycle data as input
• Large number of time steps (16) and small time
  intervals (10 min)
• Previously established relationships used to
  confirm accuracy.

25
Biological Experimentation
Partial view of previously known regulator-target
pairs
Simon,I., Barnett,J., et al(2001) Serial
regulation of transcriptional regulators in the
yeast cell cycle. Cell, 106, 697708.
26
Biological Experimentation

• Two experiments performed
• Prior knowledge used
• Used 9 transcription factors to be considered
  regulators of the 116 genes total
• Looked for the targets of these TFs
• No prior knowledge used
• Allowed any TF to target pairing

27
Biological Experimentation

• Metrics
• Correctly Identified Relationships
• Association matches previously found pathways
• Misdirected
• Association in reverse order of known
  relationship
• Specificity
• Percentage of correctly predicted known gene
  relationships / total number of predicted gene
  relationships
• Computational time
• Time required to generate DBN

28
Experiment 1 Results
9 TFs considered as regulators
29
Experiment 1 Results

• All misdirected relationships of BNT corrected in
  new approach.
• 1 Misdirection in new approach due to earlier
  initial up regulation of target rather than
  regulator
• For 70 of known yeast interactions the regulator
  expression change is earlier or simultaneous with
  target

30
Experiment 1 Results

• Using estimated transcriptional time lags gave
  all relationships found only in new method
  stronger statistical correlations
• Indicates that a biologically relevant and
  variable time lag is better suited to finding
  associations than static time step determined by
  sampling rate
• 8 uniquely identified by BNT
• 3 had better correlation at one time step(10 min)
  rather than the estimated (0 min)
• 5 had earlier expression changes of target than
  regulators

31
Experiment 2 Results
No previous knowledge considered
32
Experiment 2 Results

• 12 correct relations found in new approach and
  not BNT
• 10 of these found because of the use of
  biological relevant transcriptional time lag

33
Summary

• New method does increase relationship prediction
  over standard method in BNT
• Also significantly reduces computational time
• Requires min max number of regulators
• Requires thresholds for up and down regulation
• Pre / simultaneous regulator assumption only
  holds true for 70 of the relationships
• Missing relevant relationships

34
(No Transcript)
35
Relationships Found Exp 1
36
(No Transcript)