CS5238 Combinatorial methods in bioinformatics

1
CS5238 Combinatorial methods in bioinformatics

• Topic Gene Finding
• Promoter Recognition
• Cen Cen, Er Inn Inn, Miao Xiaoping,
• Piyush Kanti Bhunre, Yin Jun

1 November 2002
2
Outline of Presentation

• Biological Background
• Gene Finding
• Promoter Recognition
• Dragon Promoter Finder
• Open Problem and Future Research
• New Algorithm
• Conclusion

3
Biological Background

• What is gene?
• A sequence of DNA that encodes a protein or an
  RNA molecule.
• Gene has 4 regions Coding region, 5 UTR, 3 UTR
  and regulatory region (promoter regulate the
  transcription process)
• Human genome 3G bp, but only 3 is coding
  region.

4
Central Dogma

• Central Dogma- process where DNA sequence
  generates a protein
• Transcription Translation
• Promoter responsible for initiation and
  regulation of transcription
• RNA-polymerase binds to a TATA base sequence in
  promoter region

5
Central Dogma
6
Promoter Region

• Core Promoter
• TATA-box
• Initiator (Inr)
• Downstream promoter element
• 3 types of core promoter
• TATA-box
• TATA-less, Inr-containing
• Inr DPE
• Upstream promoter elements
• TSS -where transcription starts on DNA

The biology of eukaryotic promoter prediction a
review by Pedersen, A.G. et. al.
7
Outline of Presentation

• Biological Background
• Gene Finding
• Promoter Recognition
• Dragon Promoter Finder
• Open Problem and Future Research
• New Algorithm
• Conclusion

8
What is Gene Finding?

• Generate predictions of gene locations from
  primary genomic sequence (DNA sequence) by
  computational methods.
• Task of gene finding separate the coding
  regions, non-coding regions and intergenic
  regions.
• Input A seq of DNA, X x1x2xxn, where xi
  belongs to A, C, G, T
• Output Correct labeling of each element in X as
  a belonging to CR, NCR, Intergenic Region

9
Gene Finding

• 3 major kinds of gene finding strategies
• Content-based overall properties of the
  sequence when making predictions
• Site-based make use of presence or absence of a
  specific sequence, pattern or consensus
• Comparative sequence homology (database
  searching)
• Combinatorial approach - GeneMachine
• GRAIL, FGENEH, MZEF, GenScan, GeneID, GeneParser,
  HMMgene and so on.

10
Gene Finding Open Problems

• Overlapping genes no existing method that can
  deal with this problem
• Alternative splicing, alternative
  transcription/translation problem
• Sequencing errors
• Difficult to identify promoter region (PR)
  polyA (high true pos high false pos)

11
Outline of Presentation

• Biological Background
• Gene Finding
• Promoter Recognition
• Dragon Promoter Finder
• Open Problem and Future Research
• New Algorithm
• Conclusion

12
Promoter Recognition

• Accurate PR can help to
• Detect a respective gene more easily
• Determine the 5 ends of the respective gene more
  precisely
• Localize the regions that contain numerous
  different transcription control components
• Developing a perfect predictive model of PR is
  challenging

13
Main Approach to PR

• Pattern-driven strategy
• Collect a set of real binding sites to build
  characteristics definition, representation,
  pattern or profile from them
• Recognition of individual potential binding sites
  by using their characteristic profiles
• Assembling the candidates binding sites
  following some descriptions and rules about how
  these arrangements should be done.

14
Problem

• Given a collection of known binding sites, how to
  develop a representation of those sites, which is
  useful to search for them in new sequence?
• Consensus sequences
• Positional Weight Matrices (PWM)
• Hidden Markov profiles
• Multilayer neural networks and so on

15
Promoter Recognition Program

• Statistical approach artificial intelligence
  techniques -
• Dragon Promoter Finder (DPF)
• PromoterInspector
• Promoter 2.0

16
Accuracy Metric for PR

• A common measure of prediction accuracy
• Sensitivity Specificity
• TP TN
• SE SP
• TP FN TN FP
• Evaluation largely influenced by training set and
  test sets

17
Prediction of Promoter
2 x 2 contingency table
18
Example of Prediction - DPF

• Promoter positions - exact positions of the TSS
• 2360, 2585, 4125, 5026, 5734, 7090, 8567, 10641,
• -2700, -12561, -12855
• PREDICTED TRANSCRIPTION START SITES
• gi_59865_emb_X02138.1_HEHSV1SU Herpes simplex
  virus type 1 _HSV1_ short unique region DNA
• Sequence length 12979 of bases A2286,
  C4271, G4078, T2344
• Predicted TSS
• Forward strand
• 4125 5733 7093 8567 10641
• of guesses 5
• Reverse complement strand
• -12561 -2698
• of guesses 2

19
MeasurementDragon Promoter Finder, BIC-KRDL
Singapore
SE 7/11 0.64 SP 6479/6479 1
20
Outline of Presentation

• Biological Background
• Gene Finding
• Promoter Recognition
• Dragon Promoter Finder
• Open Problem and Future Research
• New Algorithm
• Conclusion

21
Dragon Promoter Finder -Introduction

• Dragon Promoter Finder( DPF)
• locates RNA polymerase II promoters in DNA
  sequences of vertebrates
• predicts Transcription Start Site (TSS)
  positions.
• strand specific
• Components
• nonlinear promoter recognition models
• signal procession
• artificial neural networks (ANNs )
• sensors.

22
Introduction (cont)

• The latest version
• Dragon Promoter Finder Ver. 1.3
• Main difference in new version
• models are now specialized for CG-rich and for
  CG-poor sequences.

23
Structure

• Overall Model
• comprises a collection of a number of basic
  models
• Basic Model
• made up of two sub-models, A and B
• trained for different ranges of system
  sensitivity
• trained separately for the best performance. 
• Sub-Model

24
Overall Model
25
Basic Model

• A composite collection of basic models
• Possess identical structure
• Trained for narrow specificity range.
• Data procession in each model is analogous.

26
Basic Model
27
Sub-model
28
Sub-model

• Three Sensors
• Specific functional regions of a gene promoter,
  coding-exon, intron
• Represented as positional distributions of
  overlapping pentamers
• ANNs

29
Sensors

• Pentamers
• All sequences of 5 consecutive nucleotides.
• AAAAA,AAAAC,AAAAG 451024 pentamers
• Selected the most significant 256 pentamers from
  1024 pentamers according to statistical relevance
• Positional weight matrices (PWM)
• The positional distribution of selected pentamers
  
• Generate PWMs for each of the 3 functional
  groups, promoter, exon intron, by counting the
  frequencies of all selected pentamers at each
  position.

30

• How to analyze the content of a data window
• Sequence Wn1n2nL-1nL, ni belongs toA, C, G, T
• Sequence P of successive overlapping pentamers
  pj P p1p2 pL5pL4.

S score for each data window The higher the s,
the more likely the data window represents the
respective functional region. These scores are
input to nonlinear signal processing block
(SPB) Output from SPB is then input to ANN
The jth pentamer at position i The frequency
of the jth pentamer at position i
31
ANNs

• Inputs scores (outputs of sensors)
• A multi-sensor integration.
• Trained by the Bayesian regularization method to
  separate promoter regions from the non-promoter
  regions.
• The threshold that best separated promoters from
  non promoter was selected
• ANN output threshold promoter region
  TSS at a position 50bp before the data windows
  end

32
Evaluation

• Successfully recognize both CpG island-related
  and CpG island-nonrelated promoters.
• Its performance on several large sets(A,B,and
  human chromosome 22) is reasonably consistent
• On the average, its expected maximum
  sensitivities is approximately 66 percent.
• In general, the DPF produces many times fewer FP
  predictions than comparative systems at the same
  sensitivity level.

33
Comparison
34
Outline of Presentation

• Biological Background
• Gene Finding
• Promoter Recognition
• Dragon Promoter Finder
• Open Problem and Future Research
• New Algorithm
• Conclusion

35
Open Problem Future Research

• Open problem
• Lack of biological information on transcription
  process
• Characteristics of promoter - low ratio of
  accuracy
• Future research work
• Designing specific algorithm for either classes
  of promoters or species-specific promoters
• Comparative sequence analysis
• Combinatorial approach
• Data mining tools

36
Outline of Presentation

• Biological Background
• Gene Finding
• Promoter Recognition
• Dragon Promoter Finder
• Open Problem and Future Research
• New Algorithm
• Conclusion

37
Gene Recognition Algorithm

• Using Dynamic Programming Approach
• Presented by Yin Jun

38
Dynamic Programming Algorithm

• Existing Dynamic Programming Algorithm for Gene
  Finding
• Snyder and Stormos method
• GeneParser
• Solovyev et als method
• FGENEH
• MORGANs DP algorithm

39
Goal of those Algorithm

• Divide DNA sequence into alternate intron and
  exon regions.
• Define a score for each kind of division. Try to
  find a kind of division which has the maximum
  score. The higher the score, the better the
  division.

40
Advantage and Disadvantage of Snyder and Stormos
algorithm

• Advantage
• the donor and the acceptor site
• HMM hidden status
• Disadvantage
• Cannot recognize promoter
• 3-mer based

41
Our Algorithm

• Combine the ideas of Dragon Promoter Finder and
  Snyder and Stormos algorithm
• Can deal with promoters
• Use pentamer instead of 3-mer, more efficient
• Dynamic Programming

42
Training Phase

• Pentamer 5 consecutive bases
• For example ACGGT
• There are 451024 different kind of pentamers
• Divide a DNA sequence into pentamers
• From training data, we can obtain the probability
  for each kind of pentamer to become a promoter,
  an intron or an exon

43
Probability Table
44
Principle of Division (1)

• Good (red promoter green intron blue exon)
• Bad (low sum of probability)

C
A
A
B
B
C
B
C
D
D
D
C
A
A
B
B
C
B
C
D
D
D
45
Principle of Division (2)

• Good (red promoter green intron blue exon)
• Bad (too frequent mutation)

C
A
A
B
B
C
B
C
D
D
D
C
A
A
B
B
C
B
C
D
D
D
46
Mutation Penalty

• M(x, x) should be 0, x? 1, 2, 3
• 1 promoter
• 2 intron
• 3 exon
• Example

47
Notation

• P(p, r) Probability for pentamer p belongs
  region r
• Obtain from training data
• M(s, t) Mutation penalty
• Parameters to specify
• pi (1in) The i th pentamer in the DNA
  sequence
• Input data (testing data)
• a(pi) Region assignment result a(pi)?1, 2, 3
• Output data

48
Score Function

• For division assignment a, its score is
• We use dynamic programming algorithm to find the
  best division assignment, whose score is the
  highest

49
Bases

• Let F(i, j, s, t) be the optimal score for the
  consecutive segment of pentamers from i th to j
  th, where i th pentamer is assigned region s, j
  th pentamer is assigned region t
• Bases

50
Recursive Definition

• Recursive Definition
• Finally, we get F(1, n, s, t) where s, t ?1, 2,
  3
• Pick up the highest score from the 9 scores

51
Time Complexity

• There are 9n2/2O(n2) entries in the dynamic
  programming table
• Filling each entry needs average n/2O(n) time
• The total time complexity is O(n3)

52
Outline of Presentation

• Biological Background
• Gene Finding
• Promoter Recognition
• Dragon Promoter Finder
• Open Problem and Future Research
• New Algorithm
• Conclusion

53
Conclusion

• Significant achievement in promoter recognition
  technique algorithms contributes to major
  advances in gene finding.
• There is still room for improvement in promoter
  recognition.
• A new algorithm is proposed for gene recognition.