"""Deal with representations of Markov Models. """ # standard modules import copy import math import random # biopython from Bio.Seq import MutableSeq class MarkovModelBuilder: """Interface to build up a Markov Model. This class is designed to try to separate the task of specifying the Markov Model from the actual model itself. This is in hopes of making the actual Markov Model classes smaller. So, this builder class should be used to create Markov models instead of trying to initiate a Markov Model directly. """ # the default pseudo counts to use DEFAULT_PSEUDO = 1 def __init__(self, state_alphabet, emission_alphabet): """Initialize a builder to create Markov Models. Arguments: o state_alphabet -- An alphabet containing all of the letters that can appear in the states o emission_alphabet -- An alphabet containing all of the letters for states that can be emitted by the HMM. """ self._state_alphabet = state_alphabet self._emission_alphabet = emission_alphabet # the probabilities for transitions and emissions # by default we have no transitions and all possible emissions self.transition_prob = {} self.emission_prob = self._all_blank(state_alphabet, emission_alphabet) # the default pseudocounts for transition and emission counting self.transition_pseudo = {} self.emission_pseudo = self._all_pseudo(state_alphabet, emission_alphabet) def _all_blank(self, first_alphabet, second_alphabet): """Return a dictionary with all counts set to zero. This uses the letters in the first and second alphabet to create a dictionary with keys of two tuples organized as (letter of first alphabet, letter of second alphabet). The values are all set to 0. """ all_blank = {} for first_state in first_alphabet.letters: for second_state in second_alphabet.letters: all_blank[(first_state, second_state)] = 0 return all_blank def _all_pseudo(self, first_alphabet, second_alphabet): """Return a dictionary with all counts set to a default value. This takes the letters in first alphabet and second alphabet and creates a dictionary with keys of two tuples organized as: (letter of first alphabet, letter of second alphabet). The values are all set to the value of the class attribute DEFAULT_PSEUDO. """ all_counts = {} for first_state in first_alphabet.letters: for second_state in second_alphabet.letters: all_counts[(first_state, second_state)] = self.DEFAULT_PSEUDO return all_counts def get_markov_model(self): """Return the markov model corresponding with the current parameters. Each markov model returned by a call to this function is unique (ie. they don't influence each other). """ transition_prob = copy.deepcopy(self.transition_prob) emission_prob = copy.deepcopy(self.emission_prob) transition_pseudo = copy.deepcopy(self.transition_pseudo) emission_pseudo = copy.deepcopy(self.emission_pseudo) return HiddenMarkovModel(transition_prob, emission_prob, transition_pseudo, emission_pseudo) def set_equal_probabilities(self): """Reset all probabilities to be an average value. This resets the values of all allowed transitions and all allowed emissions to be equal to 1 divided by the number of possible elements. This is useful if you just want to initialize a Markov Model to starting values (ie. if you have no prior notions of what the probabilities should be -- or if you are just feeling too lazy to calculate them :-). Warning 1 -- this will reset all currently set probabilities. Warning 2 -- This just sets all probabilities for transitions and emissions to total up to 1, so it doesn't ensure that the sum of each set of transitions adds up to 1. """ # first set the transitions new_trans_prob = float(1) / float(len(self.transition_prob)) for key in self.transition_prob: self.transition_prob[key] = new_trans_prob # now set the emissions new_emission_prob = float(1) / float(len(self.emission_prob)) for key in self.emission_prob: self.emission_prob[key] = new_emission_prob def set_random_probabilities(self): """Set all probabilities to randomly generated numbers. This will reset the value of all allowed transitions and emissions to random values. Warning 1 -- This will reset any currently set probabibilities. Warning 2 -- This does not check to ensure that the sum of all of the probabilities is less then 1. It just randomly assigns a probability to each """ for key in self.transition_prob: self.transition_prob[key] = random.random() for key in self.emission_prob: self.emission_prob[key] = random.random() # --- functions to deal with the transitions in the sequence def allow_all_transitions(self): """A convenience function to create transitions between all states. By default all transitions within the alphabet are disallowed; this is a way to change this to allow all possible transitions. """ # first get all probabilities and pseudo counts set # to the default values all_probs = self._all_blank(self._state_alphabet, self._state_alphabet) all_pseudo = self._all_pseudo(self._state_alphabet, self._state_alphabet) # now set any probabilities and pseudo counts that # were previously set for set_key in self.transition_prob: all_probs[set_key] = self.transition_prob[set_key] for set_key in self.transition_pseudo: all_pseudo[set_key] = self.transition_pseudo[set_key] # finally reinitialize the transition probs and pseudo counts self.transition_prob = all_probs self.transition_pseudo = all_pseudo def allow_transition(self, from_state, to_state, probability = None, pseudocount = None): """Set a transition as being possible between the two states. probability and pseudocount are optional arguments specifying the probabilities and pseudo counts for the transition. If these are not supplied, then the values are set to the default values. Raises: KeyError -- if the two states already have an allowed transition. """ # check the sanity of adding these states for state in [from_state, to_state]: assert state in self._state_alphabet, \ "State %s was not found in the sequence alphabet" % state # ensure that the states are not already set if ((from_state, to_state) not in self.transition_prob and (from_state, to_state) not in self.transition_pseudo): # set the initial probability if probability is None: probability = 0 self.transition_prob[(from_state, to_state)] = probability # set the initial pseudocounts if pseudocount is None: pseudcount = self.DEFAULT_PSEUDO self.transition_pseudo[(from_state, to_state)] = pseudocount else: raise KeyError("Transtion from %s to %s is already allowed." % (from_state, to_state)) def destroy_transition(self, from_state, to_state): """Restrict transitions between the two states. Raises: KeyError if the transition is not currently allowed. """ try: del self.transition_prob[(from_state, to_state)] del self.transition_pseudo[(from_state, to_state)] except KeyError: raise KeyError("Transition from %s to %s is already disallowed." % (from_state, to_state)) def set_transition_score(self, from_state, to_state, probability): """Set the probability of a transition between two states. Raises: KeyError if the transition is not allowed. """ if (from_state, to_state) in self.transition_prob: self.transition_prob[(from_state, to_state)] = probability else: raise KeyError("Transition from %s to %s is not allowed." % (from_state, to_state)) def set_transition_pseudocount(self, from_state, to_state, count): """Set the default pseudocount for a transition. To avoid computational problems, it is helpful to be able to set a 'default' pseudocount to start with for estimating transition and emission probabilities (see p62 in Durbin et al for more discussion on this. By default, all transitions have a pseudocount of 1. Raises: KeyError if the transition is not allowed. """ if (from_state, to_state) in self.transition_pseudo: self.transition_pseudo[(from_state, to_state)] = count else: raise KeyError("Transition from %s to %s is not allowed." % (from_state, to_state)) # --- functions to deal with emissions from the sequence def set_emission_score(self, seq_state, emission_state, probability): """Set the probability of a emission from a particular state. Raises: KeyError if the emission from the given state is not allowed. """ if (seq_state, emission_state) in self.emission_prob: self.emission_prob[(seq_state, emission_state)] = probability else: raise KeyError("Emission of %s from %s is not allowed." % (emission_state, seq_state)) def set_emission_pseudocount(self, seq_state, emission_state, count): """Set the default pseudocount for an emission. To avoid computational problems, it is helpful to be able to set a 'default' pseudocount to start with for estimating transition and emission probabilities (see p62 in Durbin et al for more discussion on this. By default, all emissions have a pseudocount of 1. Raises: KeyError if the emission from the given state is not allowed. """ if (seq_state, emission_state) in self.emission_pseudo: self.emission_pseudo[(seq_state, emission_state)] = count else: raise KeyError("Emission of %s from %s is not allowed." % (emission_state, seq_state)) class HiddenMarkovModel: """Represent a hidden markov model that can be used for state estimation. """ def __init__(self, transition_prob, emission_prob, transition_pseudo, emission_pseudo): """Initialize a Markov Model. Note: You should use the MarkovModelBuilder class instead of initiating this class directly. Arguments: o transition_prob -- A dictionary of transition probabilities for all possible transitions in the sequence. o emission_prob -- A dictionary of emissions probabilities for all possible emissions from the sequence states. o transition_pseudo -- Pseudo-counts to be used for the transitions, when counting for purposes of estimating transition probabilities. o emission_pseduo -- Pseudo-counts fo tbe used for the emissions, when counting for purposes of estimating emission probabilities. """ self._transition_pseudo = transition_pseudo self._emission_pseudo = emission_pseudo self.transition_prob = transition_prob self.emission_prob = emission_prob # a dictionary of the possible transitions from one letter to the next # the keys are the letter, and the values are lists of letters that # can be transitioned to self._transitions_from = \ self._calculate_from_transitions(self.transition_prob) def _calculate_from_transitions(self, trans_probs): """Calculate which 'from transitions' are allowed for each letter. This looks through all of the trans_probs, and uses this dictionary to determine allowed transitions. It converts this information into a dictionary, whose keys are the transition letters and whose values are a list of allowed letters to transition to. """ from_transitions = {} # loop over all of the different transitions for trans_key in trans_probs: # if the letter to 'transition from' already exists, add the # new letter which can be 'transitioned to' to the list try: from_transitions[trans_key[0]].append(trans_key[1]) # otherwise create the list and add the letter except KeyError: from_transitions[trans_key[0]] = [] from_transitions[trans_key[0]].append(trans_key[1]) return from_transitions def get_blank_transitions(self): """Get the default transitions for the model. Returns a dictionary of all of the default transitions between any two letters in the sequence alphabet. The dictionary is structured with keys as (letter1, letter2) and values as the starting number of transitions. """ return self._transition_pseudo def get_blank_emissions(self): """Get the starting default emmissions for each sequence. This returns a dictionary of the default emmissions for each letter. The dictionary is structured with keys as (seq_letter, emmission_letter) and values as the starting number of emmissions. """ return self._emission_pseudo def transitions_from(self, state_letter): """Get all transitions which can happen from the given state. This returns all letters which the given state_letter is allowed to transition to. An empty list is returned if no letters are possible. """ try: return self._transitions_from[state_letter] except KeyError: return [] def viterbi(self, sequence, state_alphabet): """Calculate the most probable state path using the Viterbi algorithm. This implements the Viterbi algorithm (see pgs 55-57 in Durbin et al for a full explanation -- this is where I took my implementation ideas from), to allow decoding of the state path, given a sequence of emissions. Arguments: o sequence -- A Seq object with the emission sequence that we want to decode. o state_alphabet -- The alphabet of the possible state sequences that can be generated. """ # calculate logarithms of the transition and emission probs log_trans = self._log_transform(self.transition_prob) log_emission = self._log_transform(self.emission_prob) viterbi_probs = {} pred_state_seq = {} state_letters = state_alphabet.letters # --- initialization # # NOTE: My index numbers are one less than what is given in Durbin # et al, since we are indexing the sequence going from 0 to # (Length - 1) not 1 to Length, like in Durbin et al. # # v_{0}(0) = 1 viterbi_probs[(state_letters[0], -1)] = 1 # v_{k}(0) = 0 for k > 0 for state_letter in state_letters[1:]: viterbi_probs[(state_letter, -1)] = 0 # --- recursion # loop over the training squence (i = 1 .. L) for i in range(0, len(sequence)): # now loop over all of the letters in the state path for main_state in state_letters: # e_{l}(x_{i}) emission_part = log_emission[(main_state, sequence[i])] # loop over all possible states possible_state_probs = {} for cur_state in self.transitions_from(main_state): # a_{kl} trans_part = log_trans[(cur_state, main_state)] # v_{k}(i - 1) viterbi_part = viterbi_probs[(cur_state, i - 1)] cur_prob = viterbi_part + trans_part possible_state_probs[cur_state] = cur_prob # finally calculate the viterbi probability using the max max_prob = max(possible_state_probs.values()) viterbi_probs[(main_state, i)] = (emission_part + max_prob) # now get the most likely state for state in possible_state_probs: if possible_state_probs[state] == max_prob: pred_state_seq[(i - 1, main_state)] = state break # --- termination # calculate the probability of the state path # loop over all letters all_probs = {} for state in state_letters: # v_{k}(L) viterbi_part = viterbi_probs[(state, len(sequence) - 1)] # a_{k0} transition_part = log_trans[(state, state_letters[0])] all_probs[state] = viterbi_part * transition_part state_path_prob = max(all_probs.values()) # find the last pointer we need to trace back from last_state = '' for state in all_probs: if all_probs[state] == state_path_prob: last_state = state assert last_state != '', "Didn't find the last state to trace from!" # --- traceback traceback_seq = MutableSeq('', state_alphabet) loop_seq = range(0, len(sequence)) loop_seq.reverse() cur_state = last_state for i in loop_seq: traceback_seq.append(cur_state) cur_state = pred_state_seq[(i - 1, cur_state)] # put the traceback sequence in the proper orientation traceback_seq.reverse() return traceback_seq.toseq(), state_path_prob def _log_transform(self, probability): """Return log transform of the given probability dictionary. When calculating the Viterbi equation, we need to deal with things as sums of logs instead of products of probabilities, so that we don't get underflow errors.. This copies the given probability dictionary and returns the same dictionary with everything transformed with a log. """ log_prob = copy.copy(probability) for key in log_prob: log_prob[key] = math.log(log_prob[key]) return log_prob