Title: Application of Hidden Markov Model for Sequence Analysis and Use for Predicting Protein Localization
1Application of Hidden Markov Model for Sequence
Analysis and Use for Predicting Protein
Localization
- By Manchikalapati
- Myerow
- Shivananda
- Monday, April 14, 2003
2Mathematical Modeling
- Mathematical Modeling in biology and chemistry
- Using probabilistic models
- Bayes Theorem and Maximum Likelihood Theorem
- Ex HMM
3What is Markov Chain ?
- A directed graph with a collection of states with
transition probabilities. - Models a random process with finite states.
- Markov Assumption The chain is memory less and
current state probability depends on previous
state. This allows us to predict behavior.
4Hidden Markov Model
- Hidden Markhov Model
- A probabilistic model that is composed of states
which are not observable events. - A statistical model that describes a probability
distribution over a number of possible sequences. - HMM has the following components
- States
- Symbol emission probabilities
- State transition probabilities
- Why Hidden? Only the symbol sequence that a
hidden state emits is observable. - Protein Modeling using HMM.
5What is Hidden? in the Markov Model
- Observed sequence is a probabilistic function of
underlying Markov chain - In HMMs the state sequence is not uniquely
determined by the observed symbol sequence, but
must be inferred probabilistically from it.
6Definition of Profile
- A profile is a description of the consensus of a
multiple sequence alignment.
Alignment Methods
Position Specific Scoring System
Position Independent (Pairwise alignment) Scoring
System Ex BLAST, FASTA
7Profile HMM
- Is a linear state machine consisting of a series
of nodes, each of which corresponds roughly to a
position (column) in the alignment from which it
was built. - The HMM will have a set of positions which would
correspond to the columns in a multiple alignment
and each column can have one of the three states
Insert, Delete and Match. - Profile HMMs can be used to do sensitive
database searching using statistical descriptions
of a sequence family's consensus.
8Profile HMM vs Std Profiles
- Profile HMMs have a formal probabilistic basis
and have a consistent theory behind gap and
insertion scores. - Profile HMMs apply a statistical method to
estimate the true frequency of a residue at a
given position in the alignment from its observed
frequency. - In general, producing good profile HMMs requires
less skill and manual intervention than producing
good standard profiles.
- Standard profile methods use heuristic methods.
- Standard profiles use the observed frequency
itself to assign the score for that residue.
9Three Algorithms of HMM
- The Viterbi algorithm get the most probable
state sequence. - The Forward/Backward algorithm score an
observation sequence against a model. - Expectation/Maximization get the parameters of
the model from the data. - For all HMM applications, the algorithms are
fairly standard. Only the design of the model are
different.
10Application of HMM
- Gene finding
- Chromosome identification
- Protein applications include
- Database searching
- Homology detection
- ExOne could take a single sequence of interest,
and query it against the model to determine if it
contained certain domains of interest.
11HMM and its basic elements
- 1)Match States(M1,M2..)
- 2)Delete State(D1,D2)
- 3)Insert States(I0,I1)
- 4) Begin State
- 5)End State
- 6)Emmision Probabilities
- 7) Transition Probabilites
- 8) Parameters
12Problems DEFINE HMM Architecture
- Problem at hand (given below)defines
architecture(to the left) - Finding Ungapped Motifs -? BLOCKS
- Finding Multiple Motifs?META-MEME
- Finding Protein Familes ? ProfileHMMs(Krogh)
- HMMER2 architecture is used in SAM,HMMER.
13 HMM Profile alignment flow chart in Pfam
14Three Important Questions that HMM should answer
- Scoring
- 1Q) How likely is a given sequence coming from
the model? - Alignment
- 2Q)What is the optimal path for generating a
given sequence - Training
- 3Q) Given a set of sequences how can you learn
about the HMM parameters
15Q1)How likely is the given Seq (ACCY) coming from
the model
- Answer
- Forward Algorithm
- Prob(A in state I0) 0.40.30.12
- Prob(C in state I1) 0.050.060.5 0.015
- Prob(C in state M1) 0.460.01 0.005
- Prob(C in state M2) (0.0050.97) (0.0150.46)
.012 - Prob(Y in state I3) .0120.0150.730.01
1.31x10-7 - Prob(Y in state M3) .0120.970.2 0.002
16Q2)What is the optimal path for generating a
given seq(ACCY)
- Answer Viterbi Algorithim
- 1. The probability that the amino acid A was
generated by state I0 is computed and entered as
the first element of the matrix. - 2. The probabilities that C is emitted in state
M1 (multiplied by the probability of the most
likely transition to state M1 from state I0) and
in state I1 (multiplied by the most likely
transition to state I1 from state I0) are entered
into the matrix element indexed by C and I1/M1. - 3. The maximum probability, max(I1, M1), is
calculated. - 4. A pointer is set from the winner back to state
I0. - 5. Steps 2-4 are repeated until the matrix is
filled. - Prob(A in state I0) 0.40.30.12
- Prob(C in state I1) 0.050.060.5 .015
- Prob(C in state M1) 0.460.01 0.005
- Prob(C in state M2) 0.460.5 0.23
- Prob(Y in state I3) 0.0150.730.01 .0001
- Prob(Y in state M3) 0.970.23 0.22
- The most likely path through the model can now be
found by following the back-pointers.
173Q)Given a set of sequences how do you learn
about HMM params
- The Learning Task
- given
- a model
- a set of sequences (the training set)
- do
- find the most likely parameters to explain the
training sequences - the goal is find a model that generalizes well to
sequences we havent seen before
- Answer Baum-Welch(Forward Backward) Algorithm
- initialize parameters of model
- iterate until convergence
- calculate the expected number of times each
transition or emission is used - adjust the parameters to maximize the
likelihood of these expected values
18HMMER in the Workflow
19Tripartite structure of signal peptide
20Translocation of Signal Peptide and Signal Anchor
signal peptide
After translocation the signal peptide is cleaved
off and the mature protein released,
signal anchor
The signal anchor is not cleaved off and the
protein is anchored to the membrane
21Two HMM Models for Signal Peptides First Model
- (Nielsen, H and Krogh A. Prediction of signal
peptides and signal anchors by a hidden Markov
model. Proc. Sixth Int. Conf on Intelligent
Systems for Molecular Biology, 122-130. AAAI
Press, 1998.) - Model not based on Multiple sequence alignment
(profile) - Compare model to neural network in eukaryotes and
prokaryotes
22The model used for signal peptides. The states in
a shaded box are tied to each other.
23Combined Model
- The model of signal anchors has only two types of
states - (grouped by the shaded boxes) apart from the
Met state. - The final states shown in the shaded box are tied
to each other, and model all residues not in a
signal peptide or an anchor.
24Hidden Markov model (HMM) vs. neural network
(NN)
- Cleavage site location percentage of signal
peptide sequences where the cleavage site was
placed correctly - Discrimination values correlation coefficients
(Mathews 1975). - Protein types signal peptides (sig) cytoplasmic
or nuclearproteins (non-sec), and signal anchors
(anc). - NN simple S-score NN combined Y-score
25Second model for Signal Peptide
- Barash S, Wang W, and Shi Y. Human secretory
signal peptide description by hidden Markov model
and generation of a strong artificial signal
peptide for secreted protein expression. Biochem
and Biophys Res Com 294, 835-842, 2002. - Profile HMM method using HMMER software
26Steps for Model Building with HMMER
- N-terminal region of 416 non-redundant human
secreted proteins - Training in hmmalign all start Met aligned in
first column, 406/416 cleavage sites aligned - Build model with MLL estimation (random model
Swiss Prot 34) - Evaluate alignment model 416/416 start Met,
406/416 cleavage site, 416/416 h-region - Re-estimate HMM with maximum discrimination method
27Model Validation
- Used hmmemit program to generate artificial
sequences of variable bit scores - In vitro validation using secretion test plasmid
constructs using secretory alkP with native
signalP replaced by HMM signal peptides, the
signal strengths correlate with the bit scores
(transcription or translation effect?) - Ranked signal strengths of known natural human
secretory proteins above average serum proteins
such as albumin were found to have high bit
scores
28Conclusion
- HMM and its applicability to sequence analysis
has been discussed - Two different HMM architectures for modeling the
signal peptide have been shown - Both are able to perform the task of separating
secreted proteins from cytoplasmic and nuclear
proteins with excellent discrimination - Discrimination of signal peptides from signal
anchors is a little less clean - Multiple modeling strategies may be beneficial
depending on the nature of the query and
available data for training