Abstract
Background
Hidden Markov models are widely employed by numerous bioinformatics programs used today. Applications range widely from comparative gene prediction to timeseries analyses of microarray data. The parameters of the underlying models need to be adjusted for specific data sets, for example the genome of a particular species, in order to maximize the prediction accuracy. Computationally efficient algorithms for parameter training are thus key to maximizing the usability of a wide range of bioinformatics applications.
Results
We introduce two computationally efficient training algorithms, one for Viterbi training and one for stochastic expectation maximization (EM) training, which render the memory requirements independent of the sequence length. Unlike the existing algorithms for Viterbi and stochastic EM training which require a twostep procedure, our two new algorithms require only one step and scan the input sequence in only one direction. We also implement these two new algorithms and the already published linearmemory algorithm for EM training into the hidden Markov model compiler HMMCONVERTER and examine their respective practical merits for three small example models.
Conclusions
Bioinformatics applications employing hidden Markov models can use the two algorithms in order to make Viterbi training and stochastic EM training more computationally efficient. Using these algorithms, parameter training can thus be attempted for more complex models and longer training sequences. The two new algorithms have the added advantage of being easier to implement than the corresponding default algorithms for Viterbi training and stochastic EM training.
Background
Hidden Markov models (HMMs) and their variants are widely used for analyzing biological sequence data. Bioinformatics applications range from methods for comparative gene prediction (e.g. [1,2]) to methods for modeling promoter grammars (e.g. [3]), identifying protein domains (e.g. [4]), predicting protein interfaces (e.g. [5]), the topology of transmembrane proteins (e.g. [6]) and residueresidue contacts in protein structures (e.g. [7]), querying pathways in protein interaction networks (e.g. [8]), predicting the occupancy of transcription factors (e.g. [9]) as well as inference models for genomewide association studies (e.g. [10]) and disease association tests for inferring ancestral haplotypes (e.g. [11]).
Most of these bioinformatics applications have been set up for a specific type of analysis and a specific biological data set, at least initially. The states of the underlying HMM and the implemented prediction algorithms determine which type of data analysis can be performed, whereas the parameter values of the HMM are chosen for a particular data set in order to optimize the corresponding prediction accuracy. If we want to apply the same method to a new data set, e.g. predict genes in a different genome, we need to adjust the parameter values in order to make sure the performance accuracy is optimal.
Manually adjusting the parameters of an HMM in order to get a high prediction accuracy can be a very time consuming task which is also not guaranteed to improve the performance accuracy. A variety of training algorithms have therefore been devised in order to address this challenge. These training algorithms require as input and starting point a socalled training set of (typically partly annotated) data. Starting with a set of (typically userchosen) initial parameter values, the training algorithm employs an iterative procedure which subsequently derives new, more refined parameter values. The iterations are stopped when a termination criterion is met, e.g. when a maximum number of iterations have been completed or when the change of the loglikelihood from one iteration to the next become sufficiently small. The model with the final set of parameters is then used to test if the performance accuracy has been improved. This is typically done by analyzing a test set of annotated data which has no overlap with the training set by comparing the predicted to the known annotation.
Of the training algorithms used in bioinformatics applications, the Viterbi training
algorithm [12,13] is probably the most commonly used, see e.g. [1416]. This is due to the fact that it is easy to implement if the Viterbi algorithm [17] is used for generating predictions. In each iteration of Viterbi training, a new
set of parameter values ϕ is derived from the counts of emissions and transitions in the Viterbi paths Π^{∗ }for the set of training sequences
Another wellknown training algorithm for HMMs is BaumWelch training [21] which is an expectation maximization (EM) algorithm [22]. In each iteration, a new set of parameter values is derived from the estimated number
of counts of emissions and transitions by considering all possible state paths (rather than only a single Viterbi path) for every training sequence.
The iterations are typically stopped after a fixed number of iterations or as soon
as the change in the loglikelihood is sufficiently small. For BaumWelch training,
the likelihood P(
Stochastic expectation maximization (EM) training or Monte Carlo EM training [32] is another iterative procedure for training the parameters of HMMs. Instead of considering only a single Viterbi state path for a given training sequence as in Viterbi training or all state paths as in BaumWelch training, stochastic EM training considers a fixednumber of K state paths Π^{s }which are sampled from the posterior distribution P(ΠX) for every training sequence X in every iteration. Sampled state paths have already been used in several bioinformatics applications for sequence decoding, see e.g. [2,33] where sampled state paths are used in the context of gene prediction to detect alternative splice variants.
All three above training algorithms, i.e. Viterbi training, BaumWelch training and stochastic EM training, can be combined with the traditional checkpointing algorithm [3436] in order to trade time for memory requirements.
We here introduce two new algorithms that make Viterbi training and stochastic EM training computationally more efficient. Both algorithms have the significant advantage of rendering the memory requirement independent of the sequence length for HMMs while keeping the time requirement the same (for Viterbi training) or modifying it by a factor of M K/(M + K), i.e. decreasing it when only one state path K = 1 is sampled for a model of M states (for stochastic EM training). Both algorithms are inspired by the linearmemory algorithm for BaumWelch training which requires only a unidirectional rather than bidirectional movement along the input sequence and which has the added advantage of being considerably easier to implement. We present a detailed description of the two new algorithms for Viterbi training and stochastic EM training. In addition, we implement all three algorithms, i.e. the new algorithms for Viterbi training and stochastic EM training and the previously published linearmemory algorithm for BaumWelch training, into our HMMcompiler HMMCONVERTER[37] and examine the practical features of these these three algorithms for three small example HMMs.
Methods and Results
Definitions and notation
In order to simplify the notation in the following, we will assume without loss of generality that we are dealing with a 1storder HMM where the Start state and the End state are the only silent states. Our description of the existing and the new algorithms easily generalize to higherorder HMMs, HMMs with more silent states (provided there exists no circular path in the HMM involving only silent states) and nHMMs, i.e. HMMs which read n unaligned input sequences rather than a single input sequence at a time. An HMM is defined by
• a set of states
• a set of transition probabilities
• a set of emission probabilities ℰ = {e_{i}(y)i ∈
We also define:
• T_{max }is the maximum number of states that any state in the model is connected to, also called the model's connectivity.
•
• Π = (π_{0}, π_{1}, ... , π_{L+1}) denotes a state path in the HMM for an input sequence X of length L, i.e. state π_{i }is assigned to sequence position x_{i}. Π^{∗ }denotes a Viterbi path and Π^{s }a state path that has been sampled from the posterior distribution P(ΠX ) of the corresponding sequence X.
A linearmemory algorithm for Viterbi training
Of the HMMbased methods that provide automatic algorithms for parameter training, Viterbi training [13] is the most popular. This is primarily due to the fact that Viterbi training is readily implemented if the Viterbi algorithm is used to generate predictions. Similar to BaumWelch training [21,22], Viterbi training is an iterative training procedure. Unlike BaumWelch training, however, which considers all state paths for a given training sequence in each iteration, Viterbi training only considers a single state path, namely a Viterbi path, when deriving new sets of parameters. In each iteration, a new set of parameter values is derived from the counts of emissions and transitions in the Viterbi paths [17] of the training sequences. The iterations are terminated as soon as the Viterbi paths of the training sequences no longer change.
In the following,
• let
• in particular let
• let
• in particular let
In the following, the superscript q will indicate from which iteration the underlying parameters derive. If we consider
all N sequences of a training set
These equations assume that we know the values of
One straightforward way to determine
A linearmemory version of the Viterbi algorithm, called the Hirschberg algorithm [38], has been known since 1975. It can be used to derive Viterbi paths in memory that is linearized with respect to the length of one of the input sequences while increasing the time requirement by at most a factor of two. The Hirschberg algorithm, however, only applies to nHMMs with n ≥ 2, i.e. HMMs which read two or more unaligned input sequences at a time. One significant disadvantage of the Hirschberg algorithm is that it is considerably more difficult to implement than the Viterbi algorithm. Only few HMMbased applications in bioinformatics actually employ it, see e.g. [1,37,39]. We will see in the following how we can devise a linearmemory algorithm for Viterbi training that does not involve the Hirschberg algorithm and that can be applied to all nHMMs including n = 1.
We now introduce a linearmemory algorithm for Viterbi training. The idea for this algorithm stems from the following observations:
(V1) If we consider the description of the Viterbi algorithm [17], in particular the recursion, we realize that the calculation of the Viterbi values can be continued by retaining only the values for the previous sequence position.
(V2) If we have a close look at the description of the traceback procedure [17], we realize that we only have to remember the Viterbi matrix elements at the previous sequence position in order to deduce the state from which the Viterbi matrix element at the current sequence position and state was derived.
(V3) If we want to derive the Viterbi path Π from the Viterbi matrix, we have to start at the end of the sequence in the End state M.
Observations (V1) and (V2) imply that local information suffices to continue the calculation of the Viterbi matrix elements (V1) and to derive a previous state (V2) if we already are in a particular state and sequence position, whereas observation (V3) reminds us that in order to derive the Viterbi path, we have to start at the end of the training sequence. Given these three observations, it is not obvious how we can come up with a computationally more efficient algorithm for training with Viterbi paths. In order to realize that a more efficient algorithm exists, one also has to also note that:
(V4) While calculating the Viterbi matrix elements in the memoryefficient way outlined in (V1), we can simultaneously keep track of the previous state from which the Viterbi matrix element at every current state and sequence position was derived. This is possible because of observation (V2) above.
(V5) In every iteration q of the training procedure, we only need to know the values of
Given all observations (V1) to (V5), we can now formally write down an algorithm which
calculates
• T_{i,j }(k, m) denotes the number of times the transition from state i to state j is used in a Viterbi state path that finishes at sequence position k in state m,
• E_{i}(y, k, m) denotes the number of times that state i reads symbol y in a Viterbi state path that finishes at sequence position k in state m,
• v_{i}(k) denotes the Viterbi matrix element for state i and sequence position k, i.e. v_{i}(k) is the probability of the Viterbi state path, i.e. the state path with the highest overall probability, that starts at the beginning of the sequence in the Start state and finishes in state i as sequence position k,
• i, j, n ∈
• δ_{i,j }is the deltafunction with δ_{i,j }= 1 for i = j and δ_{i,j }= 0 else.
Initialization: at the start of training sequence X = (x_{1},..., x_{L}) and for all m ∈
Recursion: loop over all positions k from 1 to L in the training sequence X and loop, for each such sequence position k, over all states m ∈
where l denotes the state at the previous sequence position k − 1 from which the Viterbi matrix element v_{m}(k) for state m and sequence position k derives, i.e.
Termination: at the end of the input sequence, i.e. for k = L and for m = M the silent End state, set
where l denotes the state at the sequence position L from which the Viterbi matrix element v_{M }(L) for the End state M and sequence position L derives, i.e.
The above algorithm yields
Theorem 1: The above algorithm yields
Proof: We will prove these statements via induction with respect to the sequence position k.
(1) Induction start at k = 0: This corresponds to the initialization step in the algorithm. T_{i,j }(0, m) = 0 and E_{i}(y, 0, m) = 0 for all m ∈
(2) Induction step k − 1 → k for k ∈ {1,...L − 1} if the state at sequence position k = L is not the End state M : This case corresponds to the recursion in the algorithm. We assume that
We need to distinguish two cases (a) and (b). Let l denote the state at sequence position k − 1 from which the Viterbi matrix element v_{m}(k) for state m and sequence position k derives, i.e.
• Case (a):
Emissions (i): m = i and y = x_{k }: In this case, E_{i}(y, k, m) = E_{i}(y, k − 1, l) + 1. As we know that E_{i}(y, k − 1, l) is the number of times that state i reads symbol y in a Viterbi path ending in state l at sequence position k − 1, we need to add 1 count for reading symbol y = x_{k }by state m = i at the next sequence position k in order to obtain E_{i}(y, k, m).
Transitions (ii): l = i and m = j: In this case, T_{i,j }(k, m) = T_{i,j}(k − 1, l) + 1. As we know that T_{i,j }(k − 1, l) is the number of times that a transition from state i to state j is used in a Viterbi path ending in state l at sequence position k − 1, we need to add 1 count for the transition from state l = i to state m = j which brings us from sequence position k − 1 to k in order to get T_{i,j }(k, m).
• Case (b):
Emissions (i): m ≠ i or y ≠ x_{k }: In this case, E_{i}(y, k, m) = E_{i}(y, k − 1, l). We know that E_{i}(y, k − 1, l) is the number of times that state i reads symbol y in a Viterbi path ending in state l at sequence position k − 1. If we go from state l at position k − 1 to state m at position k and read symbol x_{k }and if m ≠ i or y ≠ x_{k }, we do not need to modify the number of counts as we know that state i at position k does not read symbol y, i.e. E_{i}(y, k, m) = E_{i}(y, k − 1, l).
Transitions (ii): l ≠ i or m ≠ j: In this case, T_{i,j }(k, m) = T_{i,j }(k − 1, l). We know that T_{i,j }(k − 1, l) is the number of times that a transition from state i to state j is used in a Viterbi path ending in state l at sequence position k − 1. If we make a transition from state l at position k − 1 to state m at position k and if l ≠ i or m ≠ j, we do not need to modify the number of counts as we know this is not a transition from state i to state j, i.e. T_{i,j }(k, m) = T_{i,j }(k − 1, l).
(3) If the state at sequence position k = L is the End state M : This case corresponds to the termination step in the algorithm. As in (2), we need
to distinguish two cases (a) and (b), but now only for the transition counts. Let
l denote the state at sequence position L from which the Viterbi matrix element v_{M }(L) for the End state M and sequence position L derives, i.e.
Emissions (i): In this case, E_{i}(y, L, M) = E_{i}(y, L, l). As we know that E_{i}(y, L, l) is the number of times that state i reads symbol y in a Viterbi path ending in state l at sequence position L, we do not need to modify this number of counts when going to the silent End state at the same sequence position L as silent states do not read any symbols from the input sequence. As we are now at
the end of the input sequence X and the Viterbi path Π^{∗}(X), we have
• Case (a):
Transitions (i): l = i and M = j: In this case, T_{i,j }(L, M) = T_{i,j }(L, l) + 1. As we know that T_{i,j }(L, l) is the number of times that a transition from state i to state j is used in a Viterbi path ending in state l at sequence position L, we need to add 1 count for the transition from state l = i to the End state M = j at sequence position L. Note that this transition of state does not incur a change of sequence position
as the End state is a silent state. As we are now at the end of the input sequence X and the Viterbi path Π^{∗}(X), we have
• Case (b):
Transitions (i): l ≠ i or M ≠ j: In this case, T_{i,j }(L, M ) = T_{i,j }(L, l). We know that T_{i,j }(L, l) is the number of times that a transition from state i to state j is used in a Viterbi path ending in state l at sequence position L. If we make a transition from state l at position L to the End state M at sequence position L and if l ≠ i or M ≠ j, we do not make a transition from state i to state j and thus do not need to modify the number of counts, i.e. T_{i,j }(L, M) = T_{i,j }(L, l). Also in case (a), we are now at the end of the input sequence X and the Viterbi path Π^{∗}(X ) and thus have
End of proof.
As is clear from the above description of the algorithm, the calculation of the v_{m}, T_{i,j }and E_{i }values for sequence position k requires only the respective values for the previous sequence position k − 1, i.e. the memory requirement can be linearized with respect to the sequence length.
For an HMM with M states and a training sequence of length L and for every free parameter of the HMM that we want to train, we thus need in every
iteration
For training one free parameter in the HMM with the above algorithm, each iteration
requires
Table 1. Theoretical computational requirements
A linearmemory algorithm for stochastic EM training
One alternative to Viterbi training is BaumWelch training [21], which is an expectation maximization (EM) algorithm [22]. As Viterbi training, BaumWelch training is an iterative procedure. In each iteration of BaumWelch training, the estimated number of counts for each transition and emission is derived by considering all possible state paths for a given training sequence in the model rather than only the single Viterbi path. As discussed in the introduction, there already exists an efficient algorithm for BaumWelch training which linearizes the memory requirement with respect to the sequence length and which is also relatively easy to implement.
One variant of BaumWelch training is called stochastic EM algorithm [32]. Unlike Viterbi training which considers only a single state path and unlike BaumWelch training which considers all possible state paths for every training sequence, the stochastic EM algorithm derives new parameter values from a fixed number of K state paths (each of which is denoted Π^{s}(X)) that are sampled for each training sequence from the posterior distribution P(ΠX). Similar to Viterbi and BaumWelch training, the stochastic EM algorithm employs an iterative procedure. As for BaumWelch training, the iterations are stopped once a maximum number of iterations have been reached or once the change in the loglikelihood is sufficiently small.
In strict analogy to the notation we introduced for Viterbi training,
As usual, the superscript q indicates from which iteration the underlying parameters of the HMM derive. If we
consider all N sequences of the training set
These expressions are strictly analogous to equations 1 and 2 that we introduced for
Viterbi training. As before, these assume that we know the values of
Obtaining the counts from the forward algorithm and stochastic backtracing
It is wellknown that we can obtain the above counts T_{i,j }(X, Π^{s}(X)) and E_{i}(y, X, Π^{s}(X)) for a given training sequence X, iteration q and a sampled state path Π^{s}(X) by using a combination of the forward algorithm and stochastic backtracing [13,32]. For this, we first calculate all values in the twodimensional forward matrix using the forward algorithm and then invoke the stochastic backtracing procedure to sample a statepath Π^{s}(X) from the posterior distribution P(ΠX).
We will now explain these two algorithms in detail in order to facilitate the introduction of our new algorithm. In the following,
• f_{i}(k) denotes the sum of probabilities of all state paths that have read training sequence X up to and including sequence position k and that end in state i, i.e. f_{i}(k) = P(x_{1}, ..., x_{k }, s(x_{k }) = i), where s(x_{k}) denotes the state that reads sequence position x_{k }from input sequence X. We call f_{i}(k) the forward probability for sequence position k and state i.
• p_{i}(k, m) denotes the probability of selecting state m as the previous state while being in state i at sequence position k (i.e. sequence position k has already been read by state i), i.e. p_{i}(k, m) = P(π_{k−1 }= mπ_{k }= i). For a given sequence position k and state i, p_{i}(k, m) defines a probability distribution over previous states as
The forward matrix is calculated using the forward algorithm [13]:
Initialization: at the start of the input sequence, consider all states m ∈
Recursion: loop over all positions k from 1 to L in the input sequence and loop, for each such sequence position k, over all states m ∈
Termination: at the end of the input sequence, i.e. for k = L and m = M the End state, set
Once we have calculated all forward probabilities f_{i}(k) in the twodimensional forward matrix, i.e. for all states i in the model and all positions k in the given training sequence X, we can then use the stochastic backtracing procedure [13] to sample a state path from the posterior distribution P(ΠX).
The stochastic backtracing starts at the end of the input sequence, i.e. at sequence position k = L, in the End state, i.e. i = M , and selects state m as the previous state with probability:
This procedure is continued until we reach the start of the sequence and the Start state. The resulting succession of chosen previous states corresponds to one state path Π^{s}(X) that was sampled from the posterior distribution P(ΠX ).
The denominator in equation (4) corresponds to the sum of probabilities of all state paths that finish in state i at sequence position k, whereas the nominator corresponds to the sum of probabilities of all state paths that finish in state i at sequence position k and that have state m as the previous state.
When being in state i at sequence position k, we can therefore use this ratio to sample which previous state m we should have come from.
As this stochastic backtracing procedure requires the entire matrix of forward values
for all states and all sequence positions, the above algorithm for sampling a state
path requires
If our computer has enough memory to use the forward algorithm and the stochastic
backtracing procedure described above, each iteration in the training algorithm would
require
For many bioinformatics applications, however, where the number of states in the model M is large, the connectivity T_{max }of the model high or the training sequences are long, these memory and time requirements are too large to allow automatic parameter training using stochastic EM training.
Obtaining the counts in a more efficient way
Our previous observations (V1) to (V5) that led to the linearmemory algorithm for Viterbi training can be replaced by similar observations for stochastic EM training:
(S1) If we consider the description of the forward algorithm above, in particular the recursion in Equation (3), we realize that the calculation of the forward values can be continued by retaining only the values for the previous sequence position.
(S2) If we have a close look at the description of the stochastic backtracing algorithm, in particular the sampling step in Equation (4), we observe that the sampling of a previous state only requires the forward values for the current and the previous sequence position. So, provided we are at a particular sequence position and in a particular state, we can sample the state at the previous sequence position, if we know all forward values for the previous sequence position.
(S3) If we want to sample a state path Π^{s}(X) from the posterior distribution P(ΠX), we have to start at the end of the sequence in the End state, see the description above and Equation (4) above. (The only valid alternative for sampling state paths from the posterior distribution would be to use the backward algorithm [13] instead of the forward algorithm and to then start the stochastic backtracing procedure at the start of the sequence in the Start state.)
Observations (S1) and (S2) above imply that local information suffices to continue the calculation of the forward values (S1) and to sample a previous state (S2) if we already are in a particular state and sequence position, whereas observation (S3) reminds us that in order to sample from the correct probability distribution, we have to start the sampling at the end of the training sequence. Given these three observations, it is  as before for Viterbi training  not obvious how we can come up with a computationally more efficient algorithm. In order to realize that a more efficient algorithm does exist, one also has to note that:
(S4) While calculating the forward values in the memoryefficient way outlined in (S1) above, we can simultaneously sample a previous state for every combination of a state and a sequence position that we encounter in the calculating of the forward values. This is possible because of observation (S2) above.
(S5) In every iteration q of the training procedure, we only need to know the values of
Given all observations (S1) to (S5) above, we can now formally write down a new algorithm
which calculates
Initialization: at the start of the training sequence X and for all states m ∈
Recursion: loop over all positions k from 1 to L in the training sequence X and loop, for each such sequence position k, over all states m ∈
where l denotes the state at previous sequence position k − 1 that was sampled from the probability distribution p_{m}(k, n), n ∈ S, while being in state m at sequence position k.
Termination: at the end of the input sequence, i.e. for k = L and m = M the End state, set
where l now denotes the state at sequence position L that was sampled from the probability distribution p_{M}(L, n), n ∈
The above algorithm yields
Theorem 2: The above algorithm yields
Proof: The proof for this theorem is very similar to the proof of theorem 1 for Viterbi training and therefore omitted. The key differences are, first, that l here corresponds to the state at the previous sequence position that is sampled from a probability distribution rather than deterministically determined and, second, that Π^{s }here corresponds to a sampled state path rather than a deterministically derived Viterbi path Π*.
End of proof.
As is clear from the above algorithm, the calculation of the f_{m}, p_{m}, T_{i,j }and E_{i }values for sequence position k requires only the respective values for the previous sequence position k − 1, i.e. the memory requirement can be linearized with respect to the sequence length.
For an HMM with M states, a training sequence of length L and for every free parameter to be trained, we thus need
For training one free parameter in the HMM with the above algorithm, each iterations
requires
For sampling K state paths for the same input sequence and training one free parameter, we thus need
This can be directly compared to the algorithm described in 2.1 with requires
Examples
The algorithms that we introduce here can be used to train any HMM. The previous sections discuss the theoretical properties of the different parameter training methods in detail which are summarized in Table 1.
Even though the theoretical properties of the respective algorithms are independent of any particular HMM, the outcome of the different types of parameter training in terms of prediction accuracy and parameter convergence may very well depend on the features of a particular HMM. This is because the quantities that can be shown to be (locally) optimized by some training algorithms do not necessarily translate into an optimized prediction accuracy as defined by us here.
In order to investigate how well the different methods do in practice in terms of prediction accuracy and parameter convergence, we implemented Viterbi training, BaumWelch training and stochastic EM training for three small example HMMs. For each model, we implemented the linearmemory algorithm for BaumWelch training published earlier as well as the linearmemory algorithms for Viterbi training and stochastic EM training presented here.
In the first step, we use each model with the original parameter values to generate the sequences of the data set. We then randomly choose initial parameter values to initialize the HMM for parameter training. Each type of parameter training is performed three times using 2/3 of the unannotated data set as training set and the remaining 1/3 of the data set for performance evaluation, i.e. we perform three crossevaluation experiments for each model.
Example 1: The dishonest casino
As first case, we consider the wellknown example of the dishonest casino [13], see Figure 1. This casino consists of a fair (state F) and a loaded dice (state L). The fair dice
generates numbers from
Figure 1. HMM of the dishonest casino. Symbolic representation of the HMM of the dishonest casino. States are shown as circles, transitions are shown as directed arrows. Please refer to the text for more details.
The data set for this model consists of 300 sequences of 5000 bp length each. The results of the training experiments are shown in Figures 2 and 3.
Figure 2. Performance for the dishonest casino. The average performance as function of the number of iterations for each training algorithm. The performance is defined as the product of the sensitivity and specificity and the average is the average of three crossevaluation experiments. For stochastic EM training, a fixed number of state paths were sampled for each training sequence in each iteration (stochastic EM 1: one sampled state path, stochastic EM 3: three sampled state paths, stochastic EM 5: five sampled state paths). The error bars correspond to the standard deviation of the performance from the three crossevaluation experiments. Please refer to the text for more information.
Figure 3. Parameter convergence for the dishonest casino. Average differences of the trained and known parameter values as function of the number of iterations for each training algorithm. For a given number of iterations, we first calculate the average value of the absolute differences between the trained and known value of each emission parameter (left figure) or transition parameter (right figure) and then take the average over the three experiments from the threefold crossevaluation. The error bars correspond to the standard deviation from the three crossevaluation experiments. The algorithms have the same meaning as in Figure 2. Please refer to the text for more information.
Example 2: The extended dishonest casino
In order to investigate a HMM with a more complicated regular grammar, we extended the above example of the dishonest casino so it can now use the loaded dice (state L) only in multiples of two and the fair dice (state F) only in multiples of three, see Figure 4.
Figure 4. HMM of the extended dishonest casino. Symbolic representation of the HMM of the extended dishonest casino. States are shown as circles, transitions are shown as directed arrows. Please refer to the text for more details.
This extended HMM has seven states, the silent Start and End states, two F states and three L states, 11 transition probabilities and 30 emission probabilities. Parameterizing the HMM's probabilities yields two independent transition probabilities and 10 independent emission probabilities to be trained, i.e. 12 parameter values. In order to avoid premature termination of parameter training, we use pseudocounts of 1 for every parameter to be trained.
The data set for this model consists of 300 sequences of 5000 bp length each. The results for this extended model are shown in Figures 5 and 6.
Figure 5. Performance for the extended dishonest casino. The average performance as function of the number of iterations for each training algorithm. The performance is defined as the product of the sensitivity and specificity and the average is the average of three crossevaluation experiments. For stochastic EM training, a fixed number of state paths were sampled for each training sequence in each iteration (stochastic EM 1: one sampled state path, stochastic EM 3: three sampled state paths, stochastic EM 5: five sampled state paths). The error bars correspond to the standard deviation of the performance from the three crossevaluation experiments. Please refer to the text for more information.
Figure 6. Parameter convergence for the extended dishonest casino. Average differences of the trained and known parameter values as function of the number of iterations for each training algorithm. For a given number of iterations, we first calculate the average value of the absolute differences between the trained and known value of each emission parameter (left figure) or transition parameter (right figure) and then take the average over the three crossevaluation experiments. The error bars correspond to the standard deviation from the three crossevaluation experiments. The algorithms have the same meaning as in Figure 5. Please refer to the text for more information.
Example 3: The CpG island model
In order to study the features for the different training algorithms for a bioinformatics application, we also investigate an HMM that can be used to detect CpG islands in sequences of genomic DNA [13], see Figure 7. The model consists of 10 states, the silent Start and End states, four nonsilent states to model regions inside CpG islands (states A^{+}, C^{+}, G^{+ }and T^{+}) and four nonsilent states to model regions outside CpG islands (states A^{−}, C^{−}, G^{− }and T^{−}). The emission probabilities for each of the eight nonsilent states is a deltafunction so that any particular state (say A^{+ }or A^{−}) has an emission probability of 1 for reading the corresponding DNA nucleotide (in this case A) and a probability of zero for all other nucleotides, i.e. e_{X }+ (Y) = e_{X− }(Y) = δ_{X,Y }for X, Y ∈ {A, C, G, T}. This implies that none of the emission probabilities of this model thus requires training. With a total of 80 transition probabilities the model is, however, highly connected as any nonsilent state is connected in both directions to any other nonsilent state. Parameterizing these transition probabilities results in 33 parameters, 32 of which were determined in training (the transition probability to go to the End state was fixed). In order to avoid premature termination of parameter training, we use pseudocounts of 1 for every parameter to be trained.
Figure 7. CpG island HMM. Symbolic representation of the CpG island HMM. States are shown as circles, transitions are shown as directed arrows. Every nonsilent state can be reached from the Start state and has a transition to the End state. In addition, every nonsilent state is connected in both directions to all nonsilent states. For clarity, we here only show the transitions from the perspective of the A^{+ }state. Please refer to the text for more details.
The data set for this model consists of 180 sequences of 5000 bp length each. Figures 8 and 9 show the resulting performance.
Figure 8. Performance for the CpG island model. The average performance as function of the number of iterations for each training algorithm. The performance is defined as the product of the sensitivity and specificity and the average is the average of three crossevaluation experiments. For stochastic EM training, a fixed number of state paths were sampled for each training sequence in each iteration (stochastic EM 1: one sampled state path, stochastic EM 3: three sampled state paths, stochastic EM 5: five sampled state paths). The error bars correspond to the standard deviation of the performance from the three crossevaluation experiments. Please refer to the text for more information.
Figure 9. Parameter convergence for the CpG island model. Average differences of the trained and known parameter values as function of the number of iterations for each training algorithm. For a given number of iterations, we first calculate the average value of the absolute differences between the trained and known value of each transition parameter (this model does not have any emission parameters that require training) and then take the average over the three crossevaluation experiments. The error bars correspond to the standard deviation from the three crossevaluation experiments. The algorithms have the same meaning as in Figure 8. Please refer to the text for more information.
Prediction accuracy and parameter convergence
Our primary goal is to investigate how the prediction accuracy of the different training algorithms varies as function of the number of iterations. The prediction accuracy or performance is defined as the product of the sensitivity and specificity. Figures 2, 5 and 8 show the prediction accuracy as function of the number of iterations for all three training methods for the respective model.
Another important goal of parameter training is to recover the original parameter values of the corresponding model. We therefore also investigate how well the trained parameter values converge to the original parameter values, see Figures 3, 6 and 9 show the average differences between the trained and known parameter values as function of the number of iterations for each training algorithm and the respective model. Every data point is calculated by first determining the average value of the absolute differences between the trained and known value of each emission parameter (left figures) or transition parameter (right figures) and then taking the average over the three experiments from the threefold crossevaluation.
For the dishonest casino and the extended dishonest casino, stochastic EM training performs best, both in terms of performance and parameter convergence. It is interesting to note that the results for sampling one, three or five state paths per training sequence and per iteration are essentially the same within error bars. For these two models, Viterbi training converges fastest, i.e. the Viterbi paths remain the same from one iteration to the next, but the point of convergence is suboptimal in terms of performance and in particular in terms of parameter convergence. BaumWelch training does better than Viterbi training for these two models, but not as well as stochastic BM training as it requires more iterations to reach a lower prediction accuracy and worse parameter convergence and as it exhibits the largest variation with respect to the three crossevaluation experiments. The latter is due to many highscoring, suboptimal state paths. For the CpG island model, all training algorithms do almost equally well, with Viterbi training converging fastest. Table 2 summarizes the CPU time per iteration for the different training algorithms and models. For all three models, stochastic EM training is faster than BaumWelch training for one, three or five sampled state paths per training sequence. Viterbi training is even a bit more time efficient than stochastic EM training when sampling one state path per training sequence.
Table 2. CPU time use for different models
Based on the results from these three small example models, we would thus recommend using stochastic EM training for parameter training.
Conclusion and discussion
A wide range of bioinformatics applications are based on hidden Markov models. Having computationally efficient algorithms for training the free parameters of these models is key to optimizing the performance of these models and to adapting the models to new data sets, e.g. biological data sets from a different organism.
We here introduce two new algorithms which render the memory requirements for Viterbi
training and stochastic EM training independent of the sequence length. This is achieved
by replacing the usual bidirectional twostep procedure (which involves first calculating
the Viterbi matrix and then retrieving the Viterbi path (in case of Viterbi training)
or first calculating the forward matrix and the backward matrix before estimating
counts (in case of BaumWelch training)) by a onestep procedure which scans each
training sequence only in a onedirectional way. For an HMM with M states and a connectivity of T_{max}, a training sequence of length L and one iteration, our new algorithm reduces the memory requirement of Viterbi training
from
We have implemented the new algorithms for Viterbi training and stochastic EM training as well as the linearmemory algorithm for BaumWelch training into our HMMcompiler HMMCONVERTER[37] which can be used to set up a variety of HMMbased applications and which is freely available under the GNU General Public License version 3 (GPLv3). Please see http://people.cs.ubc.ca/~irmtraud/training webcite for more information and the source code.
We hope that the new parameter training algorithms introduced here will make parameter training for HMMbased applications easier, in particular those in bioinformatics.
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
TYL and IMM devised the new algorithms, TYL implemented them, TYL and IMM conducted the experiments, evaluated the experiments and wrote the manuscript. All authors read and approved the final manuscript.
Acknowledgements
Both authors would like to thank the anonymous referees for providing useful comments. We would also like to thank Anne Condon for giving us helpful feedback on our manuscript. Both authors gratefully acknowledge support by a Discovery Grant of the Natural Sciences and Engineering Research Council, Canada, and by a Leaders Opportunity Fund of the Canada Foundation for Innovation to I.M.M.
References

Meyer I, Durbin R: Gene structure conservation aids similarity based gene prediction.
Nucleic Acids Research 2004, 32(2):776783. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Stanke M, Keller O, Gunduz I, Hayes A, Waack S, Morgenstern B: AUGUSTUS: ab initio prediction of alternative transcripts.
Nucleic Acids Research 2006, 34:W435W439. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Won K, Sandelin A, Marstrand T, Krogh A: Modeling promoter grammars with evolving hidden Markov models.
Bioinformatics 2008, 24(15):16691675. PubMed Abstract  Publisher Full Text

Finn R, Tate J, Mistry J, Coggill P, Sammut S, Hotz H, Ceric G, Forslund K, Eddy S, Sonnhammer E, Bateman A: The Pfam protein families database.
Nucleic Acids Research 2008, 36:281288. Publisher Full Text

Nguyen C, Gardiner K, Cios K: A hidden Markov model for predicting protein interfaces.
Journal of Bioinformatics and Computational Biology 2007, 5(3):739753. PubMed Abstract  Publisher Full Text

Krogh A, Larsson B, von Heijne G, Sonnhammer E: Predicting transmembrane protein topology with a hidden Markov model: application to complete genomes.
Journal of Molecular Biology 2001, 305(3):567580. PubMed Abstract  Publisher Full Text

Bjöorkholm P, Daniluk P, Kryshtafovych A, Fidelis K, Andersson R, Hvidsten T: Using multidata hidden Markov models trained on local neighborhoods of protein structure to predict residueresidue contacts.
Bioinformatics 2009, 25(10):12641270. PubMed Abstract  PubMed Central Full Text

Qian X, Sze S, Yoon B: Querying pathways in protein interaction networks based on hidden Markov models.
Journal of Computational Biology 2009, 16(2):145157. PubMed Abstract  Publisher Full Text

Drawid A, Gupta N, Nagaraj V, Gélinas C, Sengupta A: OHMM: a Hidden Markov Model accurately predicting the occupancy of a transcription factor with a selfoverlapping binding motif.
BMC Bioinformatics 2009, 10:208. PubMed Abstract  BioMed Central Full Text  PubMed Central Full Text

king F, Sterne J, Smith G, Green P: Inference from genomewide association studies using a novel Markov model.
Genetic Epidemiology 2008, 32(6):497504. PubMed Abstract  Publisher Full Text

Su S, Balding D, Coin L: Disease association tests by inferring ancestral haplotypes using a hidden markov model.
Bioinformatics 2008, 24(7):972978. PubMed Abstract  Publisher Full Text

Juang B, Rabiner L: A segmental kmeans algorithm for estimating parameters of hidden Markov models.
IEEE Transactions on Acoustics, Speech, and Signal Processing 1990, 38(9):16391641. Publisher Full Text

Durbin R, Eddy S, Krogh A, Mitchison G: Biological sequence analysis: Probabilistic models of proteins and nucleic acids. Cambridge: Cambridge University Press; 1998.

Besemer J, Lomsazde A, Borodovsky M: GeneMarkS: a selftraining method for prediction of gene starts in microbial genomes. Implications for finding sequence motifs in regulatory regions.
Nucleic Acids Research 2001, 29(12):26072618. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Lunter G: HMMoC  a compiler for hidden Markov models.
Bioinformatics 2007, 23(18):24852487. PubMed Abstract  Publisher Full Text

TerHovhannisyan V, Lomsadze A, Cherno Y, Borodovsky M: Gene prediction in novel fungal genomes using an ab initio algorithm with unsupervised training.
Genome Research 2008, 18:19791990. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Viterbi A: Error bounds for convolutional codes and an assymptotically optimum decoding algorithm.
IEEE Trans Infor Theor 1967, 260269. Publisher Full Text

Keibler E, Arumugam M, Brent MR: The Treeterbi and Parallel Treeterbi algorithms: efficient, optimal decoding for ordinary, generalized and pair HMMs.
Bioinformatics 2007, 23(5):545554. PubMed Abstract  Publisher Full Text

Sramek R, Brejova B, Vinar T: Online Viterbi algorithm for analysis of long biological sequences.
Algorithms in Bioinformatics, Lecture Notes in Bioinformatics 2007, 4645:240251. Publisher Full Text

Lifshits Y, Mozes S, Weimann O, ZivUkelson M: Speeding Up HMM Decoding and Training by Exploiting Sequence Repetitions.
Algorithmica 2009, 54(3):379399. Publisher Full Text

Baum L: An equality and associated maximization technique in statistical estimation for probabilistic functions of Markov processes.

Dempster A, Laird N, Rubin D: Maximum likelihood from incomplete data via the EM algorithm.

Larsen T, Krogh A: EasyGene  a prokaryotic gene finder that ranks ORFs by statistical significance.
BMC Bioinformatics 2003, 4:21. PubMed Abstract  BioMed Central Full Text  PubMed Central Full Text

Jensen JL: A Note on the Linear Memory BaumWelch Algorithm.
Journal of Computational Biology 2009, 16(9):12091210. PubMed Abstract  Publisher Full Text

Khreich W, Granger E, Miri A, Sabourin R: On the memory complexity of the forwardbackward algorithm.
Pattern Recognition Letters 2010, 31(2):9199. Publisher Full Text

Elliott RJ, Aggoun L, Moon JB: Hidden Markov Models. Estimation and Control. Berlin, Germany: SpringerVerlag; 1995.

Sivaprakasam S, Shanmugan SK: A forwardonly recursion based hmm for modeling burst errors in digital channels.
IEEE Global Telecommunications Conference 1995, 2:10541058.

Turin W: Unidirectional and parallel BaumWelch algorithms.
IEEE Trans Speech Audio Process 1998, (6):516523. Publisher Full Text

Miklós I, Meyer I: A linear memory algorithm for BaumWelch training.
BMC Bioinformatics 2005, 6:231. PubMed Abstract  PubMed Central Full Text

Churbanov A, WintersHilt S: Implementing EM and Viterbi algorithms for Hidden Markov Model in linear memory.
BMC Bioinformatics 2008, 9:224. PubMed Abstract  BioMed Central Full Text  PubMed Central Full Text

Hobolth A, Jensen JL: Applications of hidden Markov models for characterization of homologous DNA sequences with common genes.
Journal of Computational Biology 2005, 12:186203. PubMed Abstract  Publisher Full Text

Bishop CM: Pattern Recognition and Machine Learning. Berlin, Germany: SpringerVerlag; 2006.
chap. 11.1.6

Cawley SL, Pachter L: HMM sampling and applications to gene finding and alternative splicing.
Bioinformatics 2003, 19(2):ii36ii41. PubMed Abstract  Publisher Full Text

Grice JA, Hughey R, Speck D: Reduced space sequence alignment.
Computer Applications in the Biosciences 1997, 13:4553. PubMed Abstract

Tarnas C, Hughey R: Reduced space hidden Markov model training.
Bioinformatics 1998, 14(5):401406. PubMed Abstract  Publisher Full Text

Wheeler R, Hughey R: Optimizing reducedspace sequence analysis.
Bioinformatics 2000, 16(12):10821090. PubMed Abstract  Publisher Full Text

Lam TY, Meyer I: HMMConverter 1.0: a toolbox for hidden Markov models.
Nucleic Acids Research 2009, 37(21):e139. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Hirschberg D: A linear space algorithm for computing maximal common subsequences.
Commun ACM 1975, 18:341343. Publisher Full Text

Meyer IM, Durbin R: Comparative ab initio prediction of gene structures using pair HMMs.
Bioinformatics 2002, 18(10):13091318. PubMed Abstract  Publisher Full Text