Hidden Markov Models for Multiple Sequence Analysis

1
Hidden Markov Models for Multiple Sequence
Analysis

• Jeff Phillips
• 10.14.01

2
Resources

• Baldi, Pierre and Brunak, Soren. Bioinformatics
  The Machine Learning Approach. 2001.
• Clote, Peter and Backofen, Rolf. Computational
  Molecular Biology. 2000.
• Shamir, Ron. Lecture in Algorithms for Molecular
  Biology. http//www.math.tau.ac.il/rshamir/algmb/
  scribe/html/lec06/

3
Outline

• Description of General Technique
• Application to DNA/RNA
• Application to Proteins
• Conclusions

4
Algorithm Overview

• Used on Family of sequences.
• Build profile HMM from family. (Baum-Welch
  Algorithm)
• From profile HMM use Viterbi Algorithm to fit new
  sequence.
• Results aligns all sequences.
• Score determines membership in family

5
Algorithm Overview
Family of sequences
Baum-Welch Algorithm
Profile HMM
Viterbi Algorithm
many-to-one not many-to-many
6
Algorithm Overview

• What is a Profile HMM?

Squares match states (correct
match) Diamonds insert states (need to
insert) Circle delete states (need to
remove item)
7
Algorithm Overview

• Align GGCT, ACCGAT, CT
• GGCT m0, m1, m2, m3, m4, m5
• ACCGAT m0, i0, m1, d2, m3, i3, i3, m4, m5
• CT m0, m1, d2, d3, m4, m5
• . G G C . . T
• a C C g a T
• . C - - . . T

8
RNA/DNA Analysis

• Long sequences (100-200 nucleotides)
• Only four symbols A,G,C,T
• ? Hard to create large profile HMMs
• Free model (align long sequences)
• Tied model (take advantage of periodicity)
• Wheel model (take advantage of periodicity)

9
Symbology

• AT -- A or T
• G -- not G (A, C, or T)
• C -- C
• GTATGACGC

10
Periodicity

• Note periodicity.
• AG in phase
• CT in anti-phase
• AGCT about every 10 base pairs.
• 10 indicates structural patterns.
• 9 would indicate triplet reading frame patterns.

11
Tied Model
Repeated HMM
di
mi
End
Start
setup
tail
ii
12
Tied Model

• GCACATCGCATGCTATCGC
• Periodic in 10 base pairs, so find pattern every
  period.
• Tie sections of HMM together
• Thicker the lines, the higher probability of
  transition

13
Wheel Model

• TATC in 8,9,10
• Outside arrows represent start
• Loop and skip arrows work like inserts and
  deletes.
• Thicker arrows means higher probability

14
Wheel Model

• Emissions of 9 state wheel with no skip or loop
  arrows.
• 3-period reading frame pattern visible.
• AGATCG on 123, 456, 789
• 10 state wheel accounts for these patterns with
  skip and loop arrows.

15
Protein Alignment

• Successfully applied to
• globins, immunoglobulins, kinases, and
  G-protein-coupled receptors
• Multiple alignment.
• Family classification.
• Profile HMM is long sequence, so we need a lot of
  data (large families).

16
Multiple Sequence Alignment

• Use Free model. We usually have enough data.
• Apply tied model in repeated structure.
• Sometimes stronger patterns will emerge.

17
Sugar Transport Proteins
- H. Mamitsuka 96
18
Family Classification

• From profile HMM score new sequence.
• Negative log-likelihoods of the complete Viterbi
  paths.
• Compare score to scores of known sequences in
  family.

19
Family Classification
20
Available Software

• HMMpro www.netid.com
• HMMER www.genetics.wustl.edu/eddy
• SAM www.cse.ucsc.edu/research/compbio/sam.html

21
Conclusion
Advantages

• From many-to-many problem to many-to-one problem.
• One profile HMM can multiply align and classify.
• Seems to notice some structural patterns.
• Available software.

22
Conclusion
Disadvantages

• Requires large amount of base data.
• Limited to first order properties.