A Hidden Markov model for
progressive multiple alignment
-Ari Loytynoja and Michel
C.Milinkovitch
Presnted by Santosh Kumar Kodicherla
HMM Applications
 Hidden Markov Model is used to find optimal
value in many applications like:
1. In Membrane Helix
2. In finding a dice whether its Fair dice or
not.
3.Decesion tree applications, Neural
Networks etc.
Working of HMM for Simple Pair
wise Alignment
 We check The two sequences and built the
unknown parent. (Similarity is maximum).
 This forms the basis for Current Algorithm.
Parent
Seq1
Seq2
Steps in HMM Works
Alignments
 Pairwise Alignment
PDGIVTSIGSNLTIACRVS
PPLASSSLGATIRLSCTLS
Multiple Alignment
DREIYGAVGSQVTLHCSFW
TQDERKLLHTTASLRCSLK
PAWLTVSEGANATFTCSLS
LPDWTVQNGKNLTLQCFAD
LDKKEAIQGGIVRVNCSVP
SSFTHLDQGERLNLSCSIP
DAQFEVIKGQTIEVRCESI
LSSKVVESGEDIVLQCAVN
PAVFKDNPTEDVEYCCVAD
Systems and Models
 Building Multiple alignment with Decreasing
Similarity.
 Compute probabilistic alignment
 Keep Track of child pointers.
 For each site Vector probabilities of alternate
characters A/C/G/T/- is calculated.
 New node generated is aligned with another
internal sequence and cont.
 Once root node is defined for multiple alignments
,we use recursive back tracking to generate
multiple alignments.
Substitution Model

Consider Seqx, Seqy- generate Seqz(Parent)
Terms:
Pa(Xi) –Probability Seq Xi has character ‘ a ‘.
If a char is observed it is given a prob=1.
Character ‘a’ has a background probability qa
a Evolves b, this represented as Sab.
Comparing characters, Substitution.
GAP:
Pxi,yi= represents prob. Xi,Yi are aligned and generate Zi.
For all the character states ‘a’ in Zk-
– pxi ,y j = pzk (xi , y j ) =∑pzk=a(xi , y j ).

pzk=a(xi , y j ) = qa ∑b sab pb(xi ) ∑b sab pb(y j )
Steps in Algorithm:
1. Look back HM Model.
2. Pair wise alignment
3. Calculate Posterior Probability.
4. Multiple Alignment
5. Testing Algorithm
 Look back HM Model
– Defines 3 states,
Match M, x-insert ,y-insert.
-Calculate probabilities of
Moving from M to X or Y
represented as δ.
-Probability to stay at insert ‘ε ‘.
-Probability to move back to M.
Pair wise alignment :
 In Dynamic prog, we
define matrix and makes
recursive calls, by
choosing best
path.
 Use Backtracking to find
the best path.
 Veterbi path to get the
best alignment path.
 Used to find the parent
vector which represents
both childs.
Forward and backward recursions.
Multiple Alignment Observations.
 The pair wise algorithm works progressively
from tip of the node to root of tree.
 Once root node is defined multiple
alignments can be generated.
 If a gap is introduced in the process , the
recursive call does not proceed.
 At a given column most of sequences are
well aligned except few which may contain
Gaps.
Testing the new Algorithm