Hidden Markov Models. Ivan Gesteira Costa Filho IZKF Research Group Bioinformatics RWTH Aachen Adapted from:

Transcription

1 Hidden Markov Models Ivan Gesteira Costa Filho IZKF Research Group Bioinformatics RWTH Aachen Adapted from:

2 Outline CG-islands The Fair Bet Casino Hidden Markov Model Decoding Algorithm Forward-Backward Algorithm HMM Parameter Estimation Viterbi training Baum-Welch algorithm

3 CG-Islands Given 4 nucleotides: probability of occurrence is ~ 1/4. Thus, probability of occurrence of a dinucleotide is ~ 1/16. However, the frequencies of dinucleotides in DNA sequences vary widely. In particular, CG is typically underepresented (frequency of CG is typically < 1/16)

4 Why CG-Islands? CG is the least frequent dinucleotide because C in CG is easily methylated and has the tendency to mutate into T afterwards However, the methylation is suppressed around genes in a genome. So, CG appears at relatively high frequency within these CG islands So, finding the CG islands in a genome is an important problem

5 Markov Chains Given a sequence x = AGCTCTC...T and transition probabilities a AG =p(x i =G x i-1 =A) What is p(x)?

6 Markov Chains p(x)=p(x L,x L-1,x L-2,...,x 1 ) =p(x L x L-1,x L-2,,x 1 ).p(x L-1 x L-2,...,x 1 )...p(x 1 ) Markov Assumption p(x i x i-1,x i-2,...,x 1 ) = p(x i x i-1 ) then p(x)=p(x L x L-1 ).p(x L-1 x L-2 )...p(x 1 ) =p(x 1 )Π i=1 p(x i-1 x i-2 )

7 Markov Chains Start State Given a sequence x = bagctctc...te Transition probabilities a AG =p(x i =G x i-1 =A) p(x) =p(x 1 )Π i=1 a p(x 1 )=1 p(x 2 x 1 )=1/4 x i x i-1

8 Using Markov Model for CG discrimination Use curated sequences of CG islands (non- CG islands) to build two markov models (model + and model - ) Estimate a st =c st / Σ t' c st'

9 Markov Model for CG discrimination For a given sequence x calculate log (p(x model + )/ p(x model - )) non-cpg islands cpg islands

10 What if we don't known any CG islands?

11 CG Islands and the Fair Bet Casino The CG islands problem can be modeled after a problem named The Fair Bet Casino The game is to flip coins, which results in only two possible outcomes: Head or Tail. The Fair coin will give Heads and Tails with same probability ½. The Biased coin will give Heads with prob. ¾.

12 The Fair Bet Casino (cont d) Thus, we define the probabilities: P(H F) = P(T F) = ½ P(H B) = ¾, P(T B) = ¼ The crooked dealer changes between Fair and Biased coins with probability 10%

13 The Fair Bet Casino Problem Input: A sequence x = x 1 x 2 x 3 x n of coin tosses made by two possible coins (F or B). Output: A sequence π = π 1 π 2 π 3 π n, with each π i being either F or B indicating that x i is the result of tossing the Fair or Biased coin respectively.

14 Problem Fair Bet Casino Problem Any observed outcome of coin tosses could have been generated by any sequence of dices π! Need to incorporate a way to grade different sequences differently. Decoding Problem Find sequence π with maximum probability

15 Hidden Markov Model (HMM) Can be viewed as an abstract machine with k hidden states that emits symbols from an alphabet Σ. Each state has its own probability distribution, and the machine switches between states according to this probability distribution. While in a certain state, the machine makes 2 decisions: What state should I move to next? What symbol - from the alphabet Σ - should I emit?

16 HMM for Fair Bet Casino (cont d) HMM model for the Fair Bet Casino Problem

17 Why Hidden? Observers can see the emitted symbols of an HMM but have no ability to know which state the HMM is currently in. Thus, the goal is to infer the most likely hidden states of an HMM based on the given sequence of emitted symbols.

18 HMM Parameters Σ: set of emission characters. Ex.: Σ = {H, T} for coin tossing Σ = {1, 2, 3, 4, 5, 6} for dice tossing Q: set of hidden states, each emitting symbols from Σ. Q={F,B} for coin tossing

19 HMM Parameters (cont d) A = (a kl ): a Q x Q matrix of probability of changing from state k to state l. a FF = 0.9 a FB = 0.1 a BF = 0.1 a BB = 0.9 E = (e k (b)): a Q x Σ matrix of probability of emitting symbol b while being in state k. e F (0) = ½ e F (1) = ½ e B (0) = ¼ e B (1) = ¾

20 HMM for Fair Bet Casino The Fair Bet Casino in HMM terms: Σ = {0, 1} (0 for Tails and 1 Heads) Q = {F,B} F for Fair & B for Biased coin. Transition Probabilities A *** Emission Probabilities E Fair Biased Tails(0) Heads(1) Fair a FF = 0.9 a FB = 0.1 Biased a BF = 0.1 a BB = 0.9 Fair Biased e F (0) = ½ e F (1) = ½ e B (0) = ¼ e B (1) = ¾

21 Hidden Paths A path π = π 1 π n in the HMM is defined as a sequence of states. Consider path π = FFFBBBBBFF and sequence x = Probability that x i was emitted from state π i X π = F F F B B B B B F F F P(x i π i ) ½ ½ ½ ¾ ¾ ¾ ¼ ¾ ½ ½ ½ P(π i π i-1 ) 9 / 10 9 / 10 1 / 10 9 / 10 9 / 10 9 / 10 9 / 10 1 / 10 9 / 10 9 / 10 9 / 10 Transition probability from state π i-1 to state π i

22 P(x,π) Calculation P(x,π): Probability that sequence x was generated by the path π: n P(x,π) = Π i=1 P(x i π i ) P(π i+1 π i ) n = Π i=1 e πi (x i ) a π i, π i+1

23 HMM - Algorithms We will describe algorithms that allow us to compute: argmax π P(π,x) - Most probable path for a given string x (Viterbi path) P(π i = k x) - Posterior probability that the i th state is k, given x A more refined measure of which states x may be in

24 Decoding Problem Goal: Find an optimal hidden path of states given observations. Input: Sequence of observations x = x 1 x n generated by an HMM M(Σ, Q, A, E) Output: A path that maximizes P(x,π) over all possible paths π.

25 Building Manhattan for Decoding Problem Andrew Viterbi used the Manhattan grid model to solve the Decoding Problem. Explore the fact that the solution for (π 1 π n ) is based on the solution from (π 1 π n-1 ) A node S i,k in the graph encodes the probability that P(x 1,..x i,π i1,,π i =k). The Viterbi algorithm finds the path that maximizes P(x,π) among all possible paths.

26 Edit Graph for Decoding Problem S1,1 1 S2,1 S3,1 Sn,1 states 2 S1,2 S2,2 S3,2 2 Sn,2 S1,k S2,k K S3,k Sn,k x 1 x 2 x 3 x n observations

27 Edit Graph for Decoding Problem S1,1 states S1,2 S1,k x 1 Initialization: s 1,l = e l (x i ) Recursions: s i+1,l = e l (x i+1 ) max j Є Q {s ji a jl }

28 Edit Graph for Decoding Problem S1,1 S2,1 states S1,2 S2,2 S1,k S2,k x 1 x 2 Initialization: s 1,l = e l (x i ) Recursions: s i+1,l = e l (x i+1 ) max j Є Q {s ji a jl }

29 Edit Graph for Decoding Problem S1,1 S2,1 states S1,2 S2,2 S1,k S2,k x 1 x 2 Initialization: s 1,l = e l (x i ) Recursions: s i+1,l = e l (x i+1 ) max j Є Q {s ji a jl }* (keep pointer to max)

30 Edit Graph for Decoding Problem S1,1 S2,1 S3,1 Sn,1 states S1,2 S2,2 S3,2 S,n,k S1,k S2,k S3,k Sn,k x 1 x 2 x 3 x n Initialization: s 1,l = e l (x i ) Recursions: s i+1,l = e l (x i+1 ) max j Є Q {s ji a jl }* (keep pointer to max)

31 Edit Graph for Decoding Problem S1,1 S2,1 S3,1 Sn,1 states S1,2 S2,2 S3,2 S,n,k S1,k S2,k S3,k Sn,k x 1 x 2 x 3 x n Let π * be the optimal path. Then, 1. find P(x,π * ) = max l Є Q {s n,l } 2. backtrack on pointers

32 Edit Graph for Decoding Problem S1,1 1 S2,1 S3,1 Sn,1 states 2 S1,2 S2,2 S3,2 2 Sn,2 S1,k S2,k K S3,k Sn,k x 1 x 2 x 3 x n Let π * be the optimal path. Then, 1. find P(x,π * ) = max l Є Q {s n,l } 2. backtrack on pointers

33 Viterbi Algorithm Computational complexity n*k 2 The value of the product can become extremely small, which leads to overflowing. To avoid overflowing, use log value instead. s i+1,k = loge l (x i+1 ) + max k Є Q {s i,k + log(a kl )}

34 Posterior Problem Given: a sequence of coin tosses generated by an HMM. Goal: find the probability that the dealer was using a biased coin at a particular time (posterior problem).

35 Posterior Algorithm We want to compute P(π i = k x), the probability distribution on the i th position, given x We start by computing P(π i = k, x) = P(x 1 x i, π i = k, x i+1 x N ) = P(x 1 x i, π i = k) P(x i+1 x N x 1 x i, π i = k) = P(x 1 x i, π i = k) P(x i+1 x N π i = k) Forward, f i k Backward, b k i Then, P(π i = k x) = P(π i = k, x) / P(x)

36 Backward Algorithm (derivation) Define backward probability b k,i as the probability of being in state π i = k and emitting the suffix x i+1 x n b i k = P(x i+1 x N π i = k) starting from i th state = k, generate rest of x = Σ πi+1 π N P(x i+1,x i+2,, x N, π i+1,, π N π i = k) = Σ l Σ πi+1 π P(x N i+1,x i+2,, x N, π i+1 = l, π i+2,, π N π i = k) = Σ l e l (x i+1 ) a kl Σ πi+1 πn P(x i+2,, x N, π i+2,, π N π i+1 = l) = Σ l e l (x i+1 ) a kl b l (i+1)

37 Backward Algorithm We can compute b i k for all k, i, using dynamic programming/edit graph Initialization: b n,k = 1, for all k Iteration: b k (i) = Σ l e l (x i+1 ) a kl b l (i+1) Termination: P(x) = Σ l a 0l e l (x 1 ) b l (1)

38 Forward Algorithm Define forward probability f k,i as the probability of being in state π i = k and emitting the prefix x 1 x i. Initialization: f 0 (0) = 1 Iteration: f k (i) = e k (x i ) Σ l f l (i 1) a lk Termination: P(x) = Σ k f k (N)

39 Posterior Decoding We can now calculate P(π i = k x) and perform the poster decoding f k (i) b k (i) P(π i = k x) = π^i P(x) = argmax k P(π i = k x) State 1 x 1 x 2 x 3 x N l P(π i =l x) k

40 Example: Fair dice P(π i = fair x) for a given sequence of dices x

41 HMM Parameter Estimation So far, we have assumed that the transition and emission probabilities are known. However, in most HMM applications, the probabilities are not known. It s very hard to estimate the probabilities.

42 HMM Parameter Estimation Problem Given HMM with states and alphabet (emission characters) Independent training sequences x 1, x m Find HMM parameters Θ (that is, a kl, e k (b)) that maximize P(x 1,, x m Θ) the joint probability of the training sequences.

43 Maximize the likelihood P(x 1,, x m Θ) as a function of Θ is called the likelihood of the model. The training sequences are assumed independent, therefore P(x 1,, x m Θ) = Π i P(x i Θ) The parameter estimation problem seeks Θ that realizes max Θ i P ( x i Θ) In practice the log likelihood is computed to avoid underflow errors

44 Two situations Known paths for training sequences CpG islands marked on training sequences One evening the casino dealer allows us to see when he changes dice Unknown paths CpG islands are not marked Do not see when the casino dealer changes dice

45 Known paths A kl = # of times each k l is taken in the training sequences E k (b) = # of times b is emitted from state k in the training sequences Compute a kl and e k (b) as maximum likelihood estimators: a = A / A e kl k ( b) kl = E k l' kl' ( b) / b' E k ( b' )

46 Pseudocounts Some state k may not appear in any of the training sequences. This means A kl = 0 for every state l and a kl cannot be computed with the given equation. To avoid this overfitting use predetermined pseudocounts r kl and r k (b). A kl = # of transitions k l + r kl E k (b) = # of emissions of b from k + r k (b) The pseudocounts reflect our prior biases about the probability values.

47 Unknown paths: Viterbi training Idea: use Viterbi decoding to compute the most probable path for training sequence x Start with some guess for initial parameters and compute π* the most probable path for x using initial parameters. Iterate until no change in π* : 1. Determine A kl and E k (b) as before 2. Compute new parameters a kl and e k (b) using the same formulas as before 3. Compute new π* for x and the current parameters

48 Viterbi training analysis The algorithm converges precisely There are finitely many possible paths. New parameters are uniquely determined by the current π*. There may be several paths for x with the same probability, hence must compare the new π* with all previous paths having highest probability. Does not maximize the likelihood Π x P(x Θ) but the contribution to the likelihood of the most probable path Π x P( Θ, π*) In general performs less well than Baum-Welch

49 Overview introduction to Hidden Markov models basic evaluation algorithms Viterbi Path, Posterior Decoding basic training methods supervised case unsupervised case: Viterbi & Baum Welch (not seem today)

50 References Original slides from bio algorithms Some slides were adapted from CS 262 course at Stanford given by Serafim Batzoglou