Motif Discovery in DNA Sequences with EC

1
Motif Discovery in DNA Sequences with EC

• Cyrus Chan

2
Biological Background
Transcription Factor
Expressed
Coexpressed Genes
TFBS motifs
3
Problem Description
Upstream Sequences
4
TFBS Identification

• Experimental methods (in vivo)
• most accurate and reliable
• time-consuming and expensive.
• Computational methods
• based on the upstream sequences
• fast and inexpensive
• de novo TFBS identification

5
Deterministic Methods

• A quick look at deterministic methods
• Enumeration of all possible motifs
• Approximate matching
• (l, d)-motif discovery problem
• Generalized suffix trees O(Nk2ld?d)
• Disadvantages
• Over predict a large amount of output
• TFBS motifs are weakly conserved

d?
6
Conventional Methods

• Multiple Sequence Alignment methods
• Do not lose any information
• Do not generalize the sequence data
• Limited help for biological understanding
• Large
• Machine learning methods
• EM, Gibbs sampling, HMM, Neural networks
• Do not produce biologically meaningful results
• Prior knowledge weight matrix
• Local search local optima

7
Evolutionary Computation

• Why use EC for motif discovery?
• Global search though also not guarantee optimal
  solutions
• Good scaling
• Flexibility of scoring
• Flexibility of representation

Review Lones et al GECCO05
8
EC methods for Motif Discovery

• FMGA Finding Motifs by Genetic Algorithm
• MDGA Motif Discovery Using A Genetic Algorithm
• GACluster Identification of Weak Motifs in
  Multiple Biological Sequences using Genetic
  Algorithm
• St-GA Motif Discovery in Upstream Sequences of
  Coordinately Expressed Genes
• Discovery, validation, and genetic dissection of
  transcription factor binding sites by comparative
  and functional genomics

9
FMGA Liu et al BIBE04

• Framework

IUPAC ambiguity codes
10
FMGA Liu et al BIBE04

• Consensus led consensus generated randomly
• Fitness function

m is the index of sequences, i is the position
within the motif , n is the index of motif
patterns, k is the length of motif pattern, j is
number of matched regions in the sequence
11
FMGA Liu et al BIBE04

• IUPAC ambiguity codes
• Total fitness score function
• where L is the total number of sequences

12
FMGA Liu et al BIBE04

• Operators
• Mutation create a weight matrix from the matched
  motif patterns
• Mutate those not completely conserved randomly

13
FMGA Liu et al BIBE04

• Operators
• Crossover one-point crossover
• Ambiguity codes penalty

14
FMGA Liu et al BIBE04

• Rearrangement
• If the predicted motif pattern is unchanged for
  more than K generations (e.g., K 10)
• For diversity

15
FMGA Liu et al BIBE04

• Experiments
• Compared with MEME and Gibbs sampler. FMGA have
  better prediction results than the others.
• For the computation time, FMGA is faster than
  MEME and slower than Gibbs sampler.
• Comments
• Early method
• Biological meaning AAA., TTT

16
MDGA Che et al GECCO05

• Positions-led like Gary B Fogels method
• Representation concatenated binary encoded
  string
• Fitness information content
• Pseudo count (db) is used

Where fb is the observed frequency of nucleotide
b on the column and pb is the background
frequency of the same nucleotide. The summation
is taken over the four possible types of
nucleotides (b). W is the motif width.
17
MDGA Che et al GECCO05

• Selection
• Roulette wheel mechanism
• The phase problem
• shifting all starting positions to the left or
  right by a small number
• Crossover operators
• single-point and double-point
• Mutation
• bitwise mutation operator

18
MDGA Che et al GECCO05

• Experiment Results
• Crossover operators
• Mutation Rate 0.01
• Compared with the Gibbs Sampler and the
  BioProspector
• on the set of 18 sequences
• Computational time is better than AlignACE

19
MDGA Che et al GECCO05
Measured by deviation from the true starting
positions ER

• Experiment Results

Gibbs Sampler
Bio-Prospector
MDGA
20
MDGA Che et al GECCO05

• Conclusion
• better prediction accuracy
• search spaces with a better strategy
• shorter running time for long sequences
• The assumption
• each sequence contains a motif
• zero to more in real cases

21
GACluster Paul et al GECCO06

• Consensus-led with GA Alignment technique
• Tackles with (l, d) motif discovery problem as
  well as weakly conserved motif discovery
• The framework

Fitness Evaluation Cluster Alignment score of
subsequences
22
GACluster Paul et al GECCO06

• Consensus from subsequences
• Focus on weakly conserved motifs
• Fitness Evaluation
• the alignment score (Information Content)
• Non-linear combination

23
GACluster Paul et al GECCO06

• Fitness Evaluation (cont.)
• Example
• AT, AC, AG, AA, AC, TC, AG, TG
• Clustering and Scoring

24
GACluster Paul et al GECCO06
Min d 1
Min d 1
Min d 1
25
GACluster Paul et al GECCO06

• Fitness
• Offspring generation
• One point crossover
• Single mutation
• Dealing with Poly-A and TATA box
• Reduce the fitness

26
GACluster Paul et al GECCO06

• Experiments
• CRP motifs
• The method with 3 motifs found outperforms the
  binary GA, which finds no real motifs
• MCB
• True motifs ACGCGT, ACGCGA, CCGCGT, TCGCGA,
  ACGCGT, ACGCGT Consensus WCGCGW
• GACluster ACGCGT, ACGCGT, ACGCGT, ACGCGT,
  ACGCGT, ACGCGT Consensus ACGCGT
• Binary GA TTTCGA, TCACCA, TCACGT, TGACGA,
  TCACGA, TAACGG None are true motifs

True Motifs
27
GACluster Paul et al GECCO06

• Discussion
• starting population is very important
• prevent loss of some initial motifs and to keep
  diversity
• Conclusions
• Drawback Like other computational methods of
  motif discovery, our method looks for similar
  subsequences in multiple biological sequences
  many of these similar subsequences have no
  biological significance.

28
St-GA Stine et al CEC03

• Consensus-led
• Representation Structured GA with binary
  encoding( two bits for each base)
• A lt S1, S2 gt S1activation level S2 expression
  level
• A (ai, aij), (ai 0,1 , i 0.. .14)
  (aij 0,1, i 0.. .14 j 0 . . . l ) .
• The interpreted string is constructed from A by
  concatenating each two bit block of S2 ai0,
  ai1 for which ai S1 1

Various Lengths of motifs
29
St-GA Stine et al CEC03

• Use BLAST for alignment
• bl2seq align the motif (subsequence) against the
  sequences
• Measures from bl2seq P is the number of
  sequence
• Also as Fitness
• Results
• Can not work with motif shorter than 7
• (l, d)-motif poor results when dgt1

30
Comparative and functional genomics Gertz et al
Genome Res vol.15 05

• Biological Experiment Oriented
• in vivo Validation of de novo motifs
• GA is used as the computational method
• Consensus-led randomly generated
• Fitness
• Information content I
• Conservation C Scores from other species
• Fitness

31
Summary

• Representation Consensus VS Positions
• Motif generation From subsequences VS Randomly
  generated (enumeration)
• Encoding Binary VS Natural
• Evaluation methods Alignment VS No alignment
• Fitness function Information content,
  Similarity
• Prior Knowledge

32
The End

• Thank you!
• Q A