Paper Presentation

1
April 10, 2006
Paper Presentation
- Training HMM structure with genetic algorithm
for biological sequence analysis
Rui Min Topic in Bioinformatics, Dr. Charles Yan
2
Overview

• An automatic means of optimizing the structure of
  HMMs
• Genetic algorithm (GA) for optimizing the HMM
  structure
• Experiments on two models
• Promoter model of C.jejuni
• Coding region model of C.jejuni
• Conclusion

3
Train HMM structure by GA

• Problems
• Biologically interpretable structure
• Controllable complexity
• Method
• Combine Baum-Wetch training with GA, called GA
  for hidden Markov models (GA-HMM).

4
Flowchart
5
Genetic Operations (I)

• Selection
• Roulette wheel selection
• Stochastic universal sampling

6
Genetic Operations (II)

• Mutation

7
Genetic Operations (III)

• Crossover

8
Selective Baum-Welch

• The Log-likelihood of model k

9
Fitness value

• Fitness

10
Experiment I promoter model of C.jejuni

• Parameters

11
Structure
12
Comparison
13
Experiment II coding region model of C.jejuni
14
Experiment II coding region model of C.jejuni
15
Conclusion

• Drawbacks
• Biologically interpretable structure
• No novel types of architecture
• No large HMM structures
• Those may be the future works
• Merit
• Capability of dealing with substructures
• GA has an application on bioinformatics

16
Unstated Aspects

• Too many constant parameters
• Probability of population for Baum-Welch training
• Percentage of training/validation
• Iteration times
• Are they best?
• Unclear parameters
• Terminal condition
• The distribution of results, t-test?
• The specific way to crossover, single?

17
Questions Discussion