vector<int> findMinHeightTrees(int n, vector<pair<int, int>>& edges)
{
vector<vector<int>> adjacency_list(n);
vector<summary> summaries(n);
for (size_t e = 0; e < edges.size(); e++)
{
adjacency_list[edges[e].first].push_back(edges[e].second);
adjacency_list[edges[e].second].push_back(edges[e].first);
}
first_pass(-1, 0, adjacency_list, summaries);
second_pass(-1, 0, 0, adjacency_list, summaries);
vector<int> result;
int min_max = 100000;
for (int i = 0; i < n; i++)
{
min_max = min(min_max, summaries[i].get_max());
}
for (int i = 0; i < n; i++)
{
if (summaries[i].get_max() == min_max)
{
result.push_back(i);
}
}
return result;
}
void train_one_round(const Fast5Map& name_map, size_t round)
{
const PoreModelMap& current_models = PoreModelSet::get_models(opt::trained_model_type);
// Initialize the training summary stats for each kmer for each model
ModelTrainingMap model_training_data;
for(auto current_model_iter = current_models.begin(); current_model_iter != current_models.end(); current_model_iter++) {
// one summary entry per kmer in the model
std::vector<StateSummary> summaries(current_model_iter->second.get_num_states());
model_training_data[current_model_iter->first] = summaries;
}
// Open the BAM and iterate over reads
// load bam file
htsFile* bam_fh = sam_open(opt::bam_file.c_str(), "r");
assert(bam_fh != NULL);
// load bam index file
std::string index_filename = opt::bam_file + ".bai";
hts_idx_t* bam_idx = bam_index_load(index_filename.c_str());
assert(bam_idx != NULL);
// read the bam header
bam_hdr_t* hdr = sam_hdr_read(bam_fh);
// load reference fai file
faidx_t *fai = fai_load(opt::genome_file.c_str());
hts_itr_t* itr;
// If processing a region of the genome, only emit events aligned to this window
int clip_start = -1;
int clip_end = -1;
if(opt::region.empty()) {
// TODO: is this valid?
itr = sam_itr_queryi(bam_idx, HTS_IDX_START, 0, 0);
} else {
fprintf(stderr, "Region: %s\n", opt::region.c_str());
itr = sam_itr_querys(bam_idx, hdr, opt::region.c_str());
hts_parse_reg(opt::region.c_str(), &clip_start, &clip_end);
}
#ifndef H5_HAVE_THREADSAFE
if(opt::num_threads > 1) {
fprintf(stderr, "You enabled multi-threading but you do not have a threadsafe HDF5\n");
fprintf(stderr, "Please recompile nanopolish's built-in libhdf5 or run with -t 1\n");
exit(1);
}
#endif
// Initialize iteration
std::vector<bam1_t*> records(opt::batch_size, NULL);
for(size_t i = 0; i < records.size(); ++i) {
records[i] = bam_init1();
}
int result;
size_t num_reads_realigned = 0;
size_t num_records_buffered = 0;
Progress progress("[methyltrain]");
do {
assert(num_records_buffered < records.size());
// read a record into the next slot in the buffer
result = sam_itr_next(bam_fh, itr, records[num_records_buffered]);
num_records_buffered += result >= 0;
// realign if we've hit the max buffer size or reached the end of file
if(num_records_buffered == records.size() || result < 0) {
#pragma omp parallel for
for(size_t i = 0; i < num_records_buffered; ++i) {
bam1_t* record = records[i];
size_t read_idx = num_reads_realigned + i;
if( (record->core.flag & BAM_FUNMAP) == 0) {
add_aligned_events(name_map, fai, hdr, record, read_idx, clip_start, clip_end, round, model_training_data);
}
}
num_reads_realigned += num_records_buffered;
num_records_buffered = 0;
}
if(opt::progress) {
fprintf(stderr, "Realigned %zu reads in %.1lfs\r", num_reads_realigned, progress.get_elapsed_seconds());
}
} while(result >= 0);
assert(num_records_buffered == 0);
progress.end();
// open the summary file
std::stringstream summary_fn;
summary_fn << "methyltrain" << opt::out_suffix << ".summary";
FILE* summary_fp = fopen(summary_fn.str().c_str(), "w");
fprintf(summary_fp, "model_short_name\tkmer\tnum_matches\tnum_skips\t"
"num_stays\tnum_events_for_training\twas_trained\t"
"trained_level_mean\ttrained_level_stdv\n");
// open the tsv file with the raw training data
std::stringstream training_fn;
training_fn << "methyltrain" << opt::out_suffix << ".round" << round << ".events.tsv";
std::ofstream training_ofs(training_fn.str());
// write out a header for the training data
StateTrainingData::write_header(training_ofs);
// iterate over models: template, complement_pop1, complement_pop2
for(auto model_training_iter = model_training_data.begin();
model_training_iter != model_training_data.end(); model_training_iter++) {
// Initialize the trained model from the input model
auto current_model_iter = current_models.find(model_training_iter->first);
assert(current_model_iter != current_models.end());
std::string model_name = model_training_iter->first;
std::string model_short_name = current_model_iter->second.metadata.get_short_name();
// Initialize the new model from the current model
PoreModel updated_model = current_model_iter->second;
uint32_t k = updated_model.k;
const std::vector<StateSummary>& summaries = model_training_iter->second;
// Generate the complete set of kmers
std::string gen_kmer(k, 'A');
std::vector<std::string> all_kmers;
for(size_t ki = 0; ki < summaries.size(); ++ki) {
all_kmers.push_back(gen_kmer);
mtrain_alphabet->lexicographic_next(gen_kmer);
}
assert(gen_kmer == std::string(k, 'A'));
assert(all_kmers.front() == std::string(k, 'A'));
assert(all_kmers.back() == std::string(k, 'T'));
// Update means for each kmer
#pragma omp parallel for
for(size_t ki = 0; ki < summaries.size(); ++ki) {
assert(ki < all_kmers.size());
std::string kmer = all_kmers[ki];
// write the observed values to a tsv file
#pragma omp critical
{
for(size_t ei = 0; ei < summaries[ki].events.size(); ++ei) {
summaries[ki].events[ei].write_tsv(training_ofs, model_short_name, kmer);
}
}
bool is_m_kmer = kmer.find('M') != std::string::npos;
bool update_kmer = opt::training_target == TT_ALL_KMERS ||
(is_m_kmer && opt::training_target == TT_METHYLATED_KMERS) ||
(!is_m_kmer && opt::training_target == TT_UNMETHYLATED_KMERS);
bool trained = false;
// only train if there are a sufficient number of events for this kmer
if(update_kmer && summaries[ki].events.size() >= opt::min_number_of_events_to_train) {
// train a mixture model where a minority of k-mers aren't methylated
ParamMixture mixture;
float incomplete_methylation_rate = 0.05f;
std::string um_kmer = mtrain_alphabet->unmethylate(kmer);
size_t um_ki = mtrain_alphabet->kmer_rank(um_kmer.c_str(), k);
// Initialize the training parameters. If this is a kmer containing
// a methylation site we train a two component mixture, otherwise
// just fit a gaussian
float major_weight = is_m_kmer ? 1 - incomplete_methylation_rate : 1.0f;
mixture.log_weights.push_back(log(major_weight));
mixture.params.push_back(current_model_iter->second.get_parameters(ki));
if(is_m_kmer) {
// add second unmethylated component
mixture.log_weights.push_back(std::log(incomplete_methylation_rate));
mixture.params.push_back(current_model_iter->second.get_parameters(um_ki));
}
if(opt::verbose > 1) {
fprintf(stderr, "INIT__MIX %s\t%s\t[%.2lf %.2lf %.2lf]\t[%.2lf %.2lf %.2lf]\n", model_training_iter->first.c_str(), kmer.c_str(),
std::exp(mixture.log_weights[0]), mixture.params[0].level_mean, mixture.params[0].level_stdv,
std::exp(mixture.log_weights[1]), mixture.params[1].level_mean, mixture.params[1].level_stdv);
}
ParamMixture trained_mixture = train_gaussian_mixture(summaries[ki].events, mixture);
if(opt::verbose > 1) {
fprintf(stderr, "TRAIN_MIX %s\t%s\t[%.2lf %.2lf %.2lf]\t[%.2lf %.2lf %.2lf]\n", model_training_iter->first.c_str(), kmer.c_str(),
std::exp(trained_mixture.log_weights[0]), trained_mixture.params[0].level_mean, trained_mixture.params[0].level_stdv,
std::exp(trained_mixture.log_weights[1]), trained_mixture.params[1].level_mean, trained_mixture.params[1].level_stdv);
}
#pragma omp critical
updated_model.states[ki] = trained_mixture.params[0];
if (model_stdv()) {
ParamMixture ig_mixture;
// weights
ig_mixture.log_weights = trained_mixture.log_weights;
// states
ig_mixture.params.emplace_back(trained_mixture.params[0]);
if(is_m_kmer) {
ig_mixture.params.emplace_back(current_model_iter->second.get_parameters(um_ki));
}
// run training
auto trained_ig_mixture = train_invgaussian_mixture(summaries[ki].events, ig_mixture);
LOG("methyltrain", debug)
<< "IG_INIT__MIX " << model_training_iter->first.c_str() << " " << kmer.c_str() << " ["
<< std::fixed << std::setprecision(5) << ig_mixture.params[0].sd_mean << " "
<< ig_mixture.params[1].sd_mean << "]" << std::endl
<< "IG_TRAIN_MIX " << model_training_iter->first.c_str() << " " << kmer.c_str() << " ["
<< trained_ig_mixture.params[0].sd_mean << " "
<< trained_ig_mixture.params[1].sd_mean << "]" << std::endl;
// update state
#pragma omp critical
{
updated_model.states[ki] = trained_ig_mixture.params[0];
}
}
trained = true;
}
#pragma omp critical
{
fprintf(summary_fp, "%s\t%s\t%d\t%d\t%d\t%zu\t%d\t%.2lf\t%.2lf\n",
model_short_name.c_str(), kmer.c_str(),
summaries[ki].num_matches, summaries[ki].num_skips, summaries[ki].num_stays,
summaries[ki].events.size(), trained, updated_model.states[ki].level_mean, updated_model.states[ki].level_stdv);
}
// add the updated model into the collection (or replace what is already there)
PoreModelSet::insert_model(opt::trained_model_type, updated_model);
}
}
// cleanup records
for(size_t i = 0; i < records.size(); ++i) {
bam_destroy1(records[i]);
}
// cleanup
sam_itr_destroy(itr);
bam_hdr_destroy(hdr);
fai_destroy(fai);
sam_close(bam_fh);
hts_idx_destroy(bam_idx);
fclose(summary_fp);
}
