1. Multiple alignment using hidden Markove models November 21, 2001 Kim Hye Jin Intelligent Multimedia Lab marisan@postech.ac.kr

2. Outline Introduction Methods and algorithm Result Discussion IM lab

3. Introduction Why HMM? –Mathematically consistent description of insertions and deletions –Theoretical insight into the difficulties of combining disparate forms of information Ex) sequences / 3D structures –Possible to train models from initially unaligned sequences Introduction| why HMM IM lab

4. Methods and algorithms State transition –State sequence is a 1 st order Markov chain –Each state is hidden –match/Insert/delete state Symbol emission Methods and algorithms|HMMs States transition Symbol emission IM lab

5. Deletion state Match state Insertion state IM lab Methods and algorithms|HMMs

6. Methods and algorithms Methods and algorithms|HMMs IM lab Replacing arbitrary scores with probabilities relative to consensus Model M consists of N states S 1 …S N. Observe sequence O consists of T symbols O 1 … O N from an alphabet x a ij : a transition from S i to S j b j (x) : emission probabilities for emission of a symbol x from each state S j

7. Methods and algorithms Methods and algorithms|HMMs IM lab Model of HMM : example of ACCY

8. Methods and algorithms Methods and algorithms|HMMs IM lab Forward algorithm - a sum rather than a maximum

9. Methods and algorithms Methods and algorithms|HMMs IM lab Viterbi algorithm -the most likely path through the model -following the back pointers

10. Methods and algorithms Methods and algorithms|HMMs IM lab Baum-Welch algorithm –A variation of the forward algorithm –Reasonable guess for initial model and then calculates a score for each sequence in the training set using EM algorithms Local optima problem: –forward algorithm /Viterbi algorithm –Baum-welch algorithm

11. Methods and algorithms Methods and algorithms|HMMs IM lab Simulated annealing –support global suboptimal –kT = 0 : standard Viterbi training procesure –kT goes down while in training

12. Methods and algorithms Methods and algorithms|HMMs IM lab ClustalW

13. Methods and algorithms Methods and algorithms|HMMs IM lab ClustalX

14. Results IM lab len : consensus length of the alignment ali : the # structurally aligned sequences %id: the percentage sequence identity Homo: the # homologues identified in and extraced from SwissProt 30 %id : the average percentage sequence identity in the set of homologues

15. Results IM lab

16. Discussion IM lab HMM -a consistent theory for insertion and deletion penality -EGF : fairly difficult alignments are well done ClusterW -progressive alignment -Disparaties between the sequence identity of the structures and the sequence identity of the homologoues -Large non-correlation between score and quality

17. Discussion IM lab The ability of HMM to sensitive fold recognition is apparent