GAME: detecting cisregulatory elements using a genetic algorithm

1
GAME detecting cis-regulatory elements using a
genetic algorithm

• Present Cyrus
• March 22 2007

2
Outline

• Introduction
• Methods
• Results
• Discussion

3
Outline

• Introduction
• Methods
• Results
• Discussion

4
Binding of transcription factor

• TFBSs
• In close proximity (100-1000 bps) to a target
  gene
• The binding sites are short (lt30 bps)
• Share a common sequence of nucleotides
  (consensus)
• Specific enough so that the TF protein does not
  bind to many random locations
• The specificity cannot be absolute in that
  varying binding affinities between the
  transcription and its target sites are required
  for different genes

5
Goal

• Based on the scoring function used by
  BioOptimizer
• BioOptimizer very simple (hill climbing), can
  only make local improvements upon another motif
  discovery method
• Goal
• Introduces an independent method using the
  scoring function
• More exhaustive search
• Eliminates the dependence on other programs
• GAME Genetic algorithm using such function

6
Outline

• Introduction
• Methods
• Results
• Discussion

7
Restricted solution space

• Dataset
• m upstream sequences, each of length li
• Sij is the nucleotide in position j of sequence i
• A motif is presented by
• a matrix of binding site locations A where each
  Aij1 if position j of sequence i is the motif
  starting site and 0 otherwise
• They do not expect much more than one binding
  site per sequence multiple occurrence will be
  handled later
• Then A can be restricted to a vector A(a1,,am),
  where ai0 means no motif

8
An artificial example

• The upper case letters indicate TFBSs
• A is (4, 3, 5, 0, 11) in this example

9
Bayesian scoring function for motifs

• Presented by Jensen et al. (Bioinformatics 2004)
• The original scoring function is closely related
  to the simpler function
• A is the number of predicted sites,
• is the estimated motif abundance out of
• possible site locations

Information content we discussed before!
10
Standard genetic operators

• is used by GAME as the fitness function
• Representation
• Vector A(a1,,am) where 0ltailtli-w1
• Mutation operator
• One point mutation
• A(a1,,ai,,am) ? A(a1 ,,ai,,am) with
  probability r (0.001, according to De Jongs
  suggestion for a small mutation rate)
• Crossover
• Single point crossover

11
Illustration of genetic operators
12
Genetic algorithm configuration

• Population N500
• N individual configurations is randomly grouped
  into N/2 pairs each pair produces 2 childrenN/2
  pairs of children in total
• Tournament selection (replacement) size 2 the
  2N individuals are randomly paired together and
  the one with a better fitness is
  selected into the next generation

13
Additional genetic operators

• Maximum generation 3000
• Converge no changes in 50 generations
• Finally Aopt is obtained from the GA
• Additional operators
• Actually not genetic ones but post-processing
  operators which are only applied on Aopt
• ADJUST
• SHIFT

14
ADJUST

• The first type of local optima
• when a majority of the motif sites have been
  aligned, with a few sites remaining to be aligned
  correctly
• Even if only one motif site has not been aligned
  correctly, there are so many such local optima
  surrounding the true optimum that our standard
  genetic algorithm is easily trapped in one of
  them
• Adjust get rid of this type of local optima

15
ADJUST

• Exhaustively check very possible better site
  position in each sequence
• Until no further improvement

16
SHIFT

• All motif sites are slightly mis-aligned

17
PWM-Scan

• Extract additional motif sites within a set of
  sequences for multiple occurrence
• Start from Aopt, cycles through all remaining
  potential motif sites
• selects any additional sites that give a
  superior fitness score
• Accept a new configuration with additional sites
  Aopt only if

18
The whole framework
GA Part
Post-processing
19
Multiple motifs

• Iterative-masking approach
• the binding sites of a discovered motif are
  masked out of the sequence dataset
• then GAME is re-applied to this masked dataset to
  find additional motifs
• Another issue of repeated sequences trap
• the presence of repeated segments of DNA which
  are not transcription factor binding sites
• be discovered and then masked out by GAME using
  iterative-masking approach

20
Extension to unknown motif width

• considering the motif width w to also be a random
  variable with a prior distribution p(w), such as
  a Poisson(w0) with prior motif width w0
• This variable-width model has a more complicated
  scoring function

21
Outline

• Introduction
• Methods
• Results
• Discussion

22
Simulation evaluation of GAME

• simulations approximating different biological
  scenarios
• 200 sequence datasets
• Number of sequences small (20) or large (100).
• Width of motif short (8 bp) or long (16 bp).
• Degree of conservation high (91) or low (70).
• Data scenario noise-free or noisy (10 contain
  no motifs).

23
Metrics

• F-scores

24
Noise Free
Noisy
In the noise-free scenario, GAME performs best,
in terms of F score, in each condition
In the noisy scenario, GAME generally shows
superior performance (in terms of F score) over
MEME, BioProspector and BioOptimizer
GAME becomes more predominant as conditions
become more difficult (e.g. lower motif
conservation)
25
Real-data applications

• CRP 18 sequences, 105 bps, 23 sites
• ERE 25 sequences, 200 bps, 25 sites (one per
  sequence)
• E2F family 25 sequences, 200 bps, 27 sites
• Five additional datasets for the TFs CREB, MEF2,
  MYOD, SRF and TBP from the recently-published ABS
  database of annotated regulatory binding sites

26
Motif width

• variable-width version of GAME
• find the optimal width within a range of 3 bps on
  either side of the expected motif width w0
• Expected motif width w0
• CRP 22 ERE 13 E2F 11
• 8, 7, 6, 10 and 6 for CREB, MEF2, MYOD, SRF and
  TBP
• Also are the widths determined by experiments

27
GAME gives superior recall and comparable
precision when compared to BioOptimizer, MEME and
BioProspector, which results in a better F score
for GAME in each application
28
(No Transcript)
29
Outline

• Introduction
• Methods
• Results
• Discussion

30
Discussion

• A Bayesian scoring function for motifs are
  adopted
• Deals with sequences without motifs, multiple
  occurrences of motifs, varying motif lengths
• GA is not fully utilized
• Additional operators are not genetic
• The refinement methods are post processing and ad
  hoc (only applied on Aopt)
• An independent and improved method over
  BioOptimizer with global search

31
The End

• Thank you!