// update the training data on the current segment
void train_segment(HMMRealignmentInput& window, uint32_t segment_id)
{
// Get the segments
assert(segment_id + 2 < window.anchored_columns.size());
HMMAnchoredColumn& start_column = window.anchored_columns[segment_id];
HMMAnchoredColumn& middle_column = window.anchored_columns[segment_id + 1];
HMMAnchoredColumn& end_column = window.anchored_columns[segment_id + 2];
std::string s_m_base = start_column.base_sequence;
std::string m_e_base = middle_column.base_sequence;
// Set up the the input data for the HMM
std::vector<HMMInputData> input = get_input_for_columns(window, start_column, end_column);
// no training can be performed if there are no reads for this segment
if(input.empty()) {
return;
}
// assume models for all the reads have the same k
const uint32_t k = input[0].read->pore_model[input[0].strand].k;
std::string segment_sequence = join_sequences_at_kmer(s_m_base, m_e_base, k);
for(uint32_t ri = 0; ri < input.size(); ++ri) {
std::vector<HMMAlignmentState> decodes = profile_hmm_align(segment_sequence, input[ri]);
update_training_with_segment(segment_sequence, input[ri]);
}
}
void run_splice_segment(HMMRealignmentInput& window, uint32_t segment_id, const uint32_t k)
{
// The structure of the data looks like this:
// --------------------------------------------------------
// S M E
// where is the start column, M is the middle column and E
// is the end column. We want to call a new consensus from S
// to E. We do this by generating the base sequence from S to E
// and then applying all of the alternatives indicated by the
// start and middle column. We score these alternatives using
// the read strands spanning from S to E. After a new consensus
// has been selected, we re-calculate the alignments of events to
// the middle anchor.
// Get the segments
assert(segment_id + 2 < window.anchored_columns.size());
HMMAnchoredColumn& start_column = window.anchored_columns[segment_id];
HMMAnchoredColumn& middle_column = window.anchored_columns[segment_id + 1];
HMMAnchoredColumn& end_column = window.anchored_columns[segment_id + 2];
std::string s_m_base = start_column.base_sequence;
std::string m_e_base = middle_column.base_sequence;
// The collection of alternative sequences
std::vector<std::string> alts;
for(uint32_t ai = 0; ai < start_column.alt_sequences.size(); ++ai) {
alts.push_back(start_column.alt_sequences[ai]);
}
// set up the input data for the HMM
std::vector<HMMInputData> data = get_input_for_columns(window, start_column, end_column);
if(opt::verbose > 0) {
fprintf(stderr, "correcting segment %u with %zu reads\n", segment_id, data.size());
}
// The current consensus sequence
std::string original = join_sequences_at_kmer(s_m_base, m_e_base, k);
std::string base = original;
// filter out poor quality reads
filter_outlier_data(data, base);
// Only attempt correction if there are any reads here
if(!data.empty()) {
std::string bs_result = run_block_substitution(base, data, alts);
std::string mut_result = run_mutation(bs_result, data);
base = mut_result;
}
if(opt::verbose > 0) {
fprintf(stderr, "ORIGINAL[%d] %s\n", segment_id, original.c_str());
fprintf(stderr, "RESULT[%d] %s\n", segment_id, base.c_str());
}
// Update the sequences for the start and middle segments
// by cutting the new consensus in the middle
// We maintain the k-mer match invariant by requiring the
// sequences to overlap by k-bp
assert(base.length() >= k);
uint32_t midpoint_kmer = (base.length() - k + 1) / 2;
std::string s_m_fixed = base.substr(0, midpoint_kmer + k);
std::string m_e_fixed = base.substr(midpoint_kmer);
assert(s_m_fixed.substr(s_m_fixed.size() - k) == m_e_fixed.substr(0, k));
start_column.base_sequence = s_m_fixed;
middle_column.base_sequence = m_e_fixed;
// Update the event indices in the first column to match
for(uint32_t ri = 0; ri < data.size(); ++ri) {
// Realign to the consensus sequence
std::vector<HMMAlignmentState> decodes = profile_hmm_align(base, data[ri]);
// Get the closest event aligned to the target kmer
int32_t min_k_dist = base.length();
uint32_t event_idx = 0;
for(uint32_t di = 0; di < decodes.size(); ++di) {
int32_t dist = abs(decodes[di].kmer_idx - midpoint_kmer);
if(dist <= min_k_dist) {
min_k_dist = dist;
event_idx = decodes[di].event_idx;
}
}
middle_column.anchors[data[ri].anchor_index].event_idx = event_idx;
}
}
void debug_sequence(const std::string& name, uint32_t seq_id, uint32_t read_id, const HMMInputSequence& sequence, const HMMInputData& data)
{
std::vector<HMMAlignmentState> alignment = profile_hmm_align(sequence, data);
print_alignment(name, seq_id, read_id, sequence, data, alignment);
}
void update_training_with_segment(const HMMInputSequence& sequence, const HMMInputData& data)
{
std::vector<HMMAlignmentState> alignment = profile_hmm_align(sequence, data);
data.read->parameters[data.strand].add_training_from_alignment(sequence, data, alignment);
}
std::vector<AlignmentState> hmm_align(const std::string& sequence, const HMMInputData& data)
{
return profile_hmm_align(sequence, data);
// return khmm_posterior_decode(sequence, state);
}