/******************************************************************************
Train PMC models from positive example files
*******************************************************************************/
void PMCSQS_Scorer::train_pmc_rank_models(Config *config, const FileManager& fm,
int sel_charge, bool overwrite)
{
const bool sample_diagnostic = false;
const vector<int>& spectra_counts = fm.get_spectra_counts();
max_model_charge=0;
int charge;
for (charge=1; charge<spectra_counts.size(); charge++)
{
if (spectra_counts[charge]>=MIN_SPECTRA_FOR_PMCSQS_MODEL)
max_model_charge=charge;
}
const int max_to_read_per_file = 40000;
vector<string> real_names;
init_PMC_feature_names(real_names);
// try and read existing pmc model, otherwise init a new one
string pmc_path = config->get_resource_dir() + "/" + config->get_model_name() + "_PMCR.txt";
ifstream model_stream(pmc_path.c_str());
if (model_stream.is_open() && model_stream.good())
{
model_stream.close();
string pmcr_name = config->get_model_name() + "_PMCR.txt";
const char *path = pmc_path.c_str();
this->read_pmc_rank_models(config,(char *)pmcr_name.c_str());
}
else
{
set_pmc_mass_thresholds();
this->set_frag_pair_sum_offset(MASS_PROTON); // b+y - PM+19
this->set_bin_increment(0.1);
pmc_rank_models.resize(pmc_rank_mass_thresholds.size());
pmc_charge_mz_biases.resize(pmc_rank_mass_thresholds.size());
}
const double prop_train = 0.5;
// It is assumed that the mass thresholds were set according to the training data
// (this is done manually with values encoded in the set_mass_threhsolds function)
for (charge=1; charge<=max_model_charge; charge++)
{
if (sel_charge>0 && charge != sel_charge)
continue;
const int num_sizes = pmc_rank_mass_thresholds[charge].size();
pmc_rank_models[charge].resize(num_sizes+1,NULL);
pmc_charge_mz_biases[charge].resize(num_sizes+1,0);
int size_idx;
for (size_idx=0; size_idx<=num_sizes; size_idx++)
{
if (pmc_rank_models[charge][size_idx] && ! overwrite)
continue;
vector<SingleSpectrumFile *> test_ssfs;
BasicSpecReader bsr;
static QCPeak peaks[5000];
RankBoostDataset train_ds, test_ds, pos_ds, neg_ds;
mass_t min_mass =0;
mass_t max_mass = POS_INF;
if (size_idx>0)
min_mass = pmc_rank_mass_thresholds[charge][size_idx-1];
if (size_idx<num_sizes)
max_mass = pmc_rank_mass_thresholds[charge][size_idx];
// these ranges are given according to pm_with_19
// so files should be selected through select_files and not
// select_file_in_mz_range
FileSet fs;
fs.select_files(fm,min_mass,max_mass,-1,-1,charge);
if (fs.get_total_spectra()<500)
continue;
int num_groups_in_train=0;
int num_groups_in_test=0;
cout << "TRAINING charge " << charge << " size " << size_idx << " (" <<
min_mass << "-" << max_mass << ")" << endl;
fs.randomly_reduce_ssfs(max_to_read_per_file);
const vector<SingleSpectrumFile *>& all_ssf = fs.get_ssf_pointers();
const int num_samples = all_ssf.size();
// first find the bias in number of bins between the true m/z bin and
// the optimal m/z bin
vector<bool> skipped_idxs;
skipped_idxs.resize(num_samples,false);
int skipped_bad_mz=0;
mass_t total_bias=0;
int i;
for (i=0; i<num_samples; i++)
{
SingleSpectrumFile* ssf = all_ssf[i];
BasicSpectrum bs;
bs.num_peaks = bsr.read_basic_spec(config,fm,ssf,peaks);
bs.peaks = peaks;
bs.ssf = ssf;
ssf->peptide.calc_mass(config);
const mass_t true_mz = (ssf->peptide.get_mass()+MASS_H2O+(mass_t)charge)/(mass_t)charge;
if (fabs(true_mz - bs.ssf->m_over_z)>2.5)
{
//cout << setprecision(2) << true_mz << " <---> " << bs.ssf->m_over_z << " skipping" << endl;
skipped_bad_mz++;
skipped_idxs[i]=true;
continue;
}
init_for_current_spec(config,bs);
calculate_curr_spec_pmc_values(bs, bin_increment);
// find the true_mz_bin_idx
const vector<PMCRankStats>& pmc_stats = curr_spec_rank_pmc_tables[charge];
int true_mz_bin_idx=0;
while (true_mz_bin_idx<pmc_stats.size() && pmc_stats[true_mz_bin_idx].m_over_z<true_mz)
true_mz_bin_idx++;
if (true_mz_bin_idx == pmc_stats.size())
true_mz_bin_idx--;
if (true_mz_bin_idx>0 && pmc_stats[true_mz_bin_idx].m_over_z-true_mz>true_mz-pmc_stats[true_mz_bin_idx-1].m_over_z)
true_mz_bin_idx--;
int opt_bin_idx = get_optimal_bin(true_mz_bin_idx, charge);
if (opt_bin_idx <=0 || opt_bin_idx == pmc_stats.size()-1)
{
skipped_bad_mz++;
skipped_idxs[i]=true;
continue;
}
total_bias += (pmc_stats[opt_bin_idx].m_over_z - pmc_stats[true_mz_bin_idx].m_over_z);
if (fabs(pmc_stats[opt_bin_idx].m_over_z - pmc_stats[true_mz_bin_idx].m_over_z)>4.0)
{
cout << "opt bin: " << opt_bin_idx << " (" << pmc_stats[opt_bin_idx].m_over_z << ") ";
cout << "tru bin: " << true_mz_bin_idx << " ("<< pmc_stats[true_mz_bin_idx].m_over_z << ")" << endl;
}
}
mass_t mz_bias = total_bias / (mass_t)(num_samples-skipped_bad_mz);
pmc_charge_mz_biases[charge][size_idx]=mz_bias;
cout << "m/z bias: " << setprecision(4) << mz_bias << endl;
cout << "skipped " << skipped_bad_mz << "/" << num_samples <<
" because of m/z more than 2.5 away from observed..." << endl;
// pmc_charge_mz_biases[charge][size_idx] = 0;
for (i=0; i<num_samples; i++)
{
if (skipped_idxs[i])
continue;
SingleSpectrumFile* ssf = all_ssf[i];
BasicSpectrum bs;
bs.num_peaks = bsr.read_basic_spec(config,fm,ssf,peaks);
bs.peaks = peaks;
bs.ssf = ssf;
const mass_t true_mz = (ssf->peptide.get_mass()+MASS_H2O+(mass_t)charge)/(mass_t)charge;
init_for_current_spec(config,bs);
calculate_curr_spec_pmc_values(bs, bin_increment);
// find the true_mz_bin_idx
const vector<PMCRankStats>& pmc_stats = curr_spec_rank_pmc_tables[charge];
int true_mz_bin_idx=0;
while (true_mz_bin_idx<pmc_stats.size() && pmc_stats[true_mz_bin_idx].m_over_z<true_mz)
true_mz_bin_idx++;
if (true_mz_bin_idx == pmc_stats.size())
true_mz_bin_idx--;
if (true_mz_bin_idx>0 && pmc_stats[true_mz_bin_idx].m_over_z-true_mz>true_mz-pmc_stats[true_mz_bin_idx-1].m_over_z)
true_mz_bin_idx--;
int opt_bin_idx = get_optimal_bin(true_mz_bin_idx, charge);
static vector<RankBoostSample> spec_samples;
fill_RankBoost_smaples_with_PMC(bs, charge, spec_samples);
// select samples and add them to pmc_ds
int good_idx;
vector<int> bad_idxs;
select_training_sample_idxs(charge,spec_samples,bs,good_idx,bad_idxs);
const bool ind_add_to_train = (my_random()<prop_train);
int group_idx;
if (ind_add_to_train)
{
group_idx= num_groups_in_train++;
}
else
{
group_idx= num_groups_in_test++;
test_ssfs.push_back(ssf);
}
RankBoostDataset& ds = (ind_add_to_train ? train_ds : test_ds);
const int pos_index = ds.get_num_samples();
spec_samples[good_idx].group_idx = group_idx;
spec_samples[good_idx].rank_in_group=0;
ds.add_sample(spec_samples[good_idx]);
if (sample_diagnostic)
pos_ds.add_sample(spec_samples[good_idx]);
int j;
for (j=0; j<bad_idxs.size(); j++)
{
const int bad_idx = bad_idxs[j];
if (bad_idx < 0 || bad_idx>= spec_samples.size())
continue;
spec_samples[bad_idx].group_idx=group_idx;
spec_samples[bad_idx].rank_in_group=1;
ds.add_to_phi_vector(ds.get_num_samples(),pos_index);
ds.add_sample(spec_samples[bad_idx]);
if (sample_diagnostic)
neg_ds.add_sample(spec_samples[bad_idx]);
}
}
train_ds.set_num_groups(num_groups_in_train);
test_ds.set_num_groups(num_groups_in_test);
train_ds.compute_total_phi_weight();
train_ds.initialize_potenital_lists();
train_ds.initialzie_real_feature_table(real_names.size());
test_ds.compute_total_phi_weight();
if (pmc_rank_models[charge][size_idx])
delete pmc_rank_models[charge][size_idx];
pmc_rank_models[charge][size_idx] = new RankBoostModel;
RankBoostModel* boost = pmc_rank_models[charge][size_idx];
vector<string> empty;
empty.clear();
boost->init_rankboost_model_feature_names(empty,real_names);
boost->init_rankboost_model_for_training(train_ds,100,25);
train_ds.initialize_real_vote_lists(*boost);
if (sample_diagnostic)
{
boost->summarize_features_pos_neg(pos_ds.get_samples(),neg_ds.get_samples());
}
else
boost->summarize_features(train_ds.get_samples());
boost->train_rankboost_model(train_ds,4000,NULL,&test_ds);
boost->ouput_ranked_feature_list();
// output_pmc_rank_results(fm,charge,test_ssfs);
// exit(0);
ind_initialized_pmcr = true;
// string path;
// path = config->get_resource_dir() + "/" + config->get_model_name() + "_PMCRtt.txt";
// this->write_pmc_rank_models(path.c_str());
}
}
string path;
path = config->get_resource_dir() + "/" + config->get_model_name() + "_PMCR.txt";
this->write_pmc_rank_models(path.c_str());
ind_initialized_pmcr = true;
}
void PMCSQS_Scorer::train_sqs_models(Config *config,
const FileManager& fm_pos,
const char *neg_list,
int specificCharge,
vector<vector<float> > *inputWeights)
{
vector< vector< vector<ME_Regression_Sample> > > samples; // neg, p1, p2, p3 / sizeIndex
FileManager fm_neg;
const vector<int>& spectra_counts = fm_pos.get_spectra_counts();
maximalChargeWithModels_ = (inputWeights ? inputWeights->size()-1 : 3);
int charge;
set_frag_pair_sum_offset(MASS_PROTON); // b+y - PM+19
set_bin_increment(0.1);
this->set_sqs_mass_thresholds();
if (this->pmcMassThresholds_.size() == 0)
{
pmcMassThresholds_=config->get_size_thresholds();
}
vector<vector<float> > classWeights;
if (inputWeights)
{
classWeights = *inputWeights;
}
else
{
classWeights.resize(maximalChargeWithModels_+1);
int i;
for (i=0; i<classWeights.size(); i++)
classWeights[i].resize(maximalChargeWithModels_+1,1.0);
}
const int numSizes = this->sqsMassThresholds_.size();
cout << "NUM SIZE MODELS: " << numSizes+1 << endl;
samples.resize(maximalChargeWithModels_+1);
fm_neg.init_from_list_file(config, neg_list);
const int max_to_read_per_file = 8000;
for (charge=0; charge<=maximalChargeWithModels_; charge++)
{
if (charge>0 && specificCharge>0 && charge != specificCharge)
continue;
int sizeIndex;
for (sizeIndex=0; sizeIndex<=numSizes; sizeIndex++)
{
const mass_t minMass = (sizeIndex == 0 ? 0 : sqsMassThresholds_[sizeIndex-1]);
const mass_t maxMass = (sizeIndex == numSizes ? POS_INF : sqsMassThresholds_[sizeIndex]);
samples[charge].resize(numSizes+1);
BasicSpecReader bsr;
QCPeak peaks[5000];
FileSet fs;
if (charge == 0)
{
fs.select_files_in_mz_range(fm_neg,minMass, maxMass,0);
}
else
{
fs.select_files_in_mz_range(fm_pos, minMass, maxMass, charge);
}
cout << "Found " << fs.get_total_spectra() << " for charge " << charge << " ranges:" <<
minMass << " - " << maxMass << endl;
fs.randomly_reduce_ssfs(max_to_read_per_file);
const vector<SingleSpectrumFile *>& all_ssf = fs.get_ssf_pointers();
const int label = (charge == 0 ? 1 : 0);
const int num_samples = all_ssf.size();
samples[charge][sizeIndex].resize(num_samples);
int i;
for (i=0; i<num_samples; i++)
{
SingleSpectrumFile* ssf = all_ssf[i];
BasicSpectrum bs;
bs.peaks = peaks;
bs.ssf = ssf;
if (charge==0)
{
bs.num_peaks = bsr.read_basic_spec(config,fm_neg,ssf,peaks);
bs.ssf->charge=0;
}
else
bs.num_peaks = bsr.read_basic_spec(config,fm_pos,ssf,peaks);
init_for_current_spec(config,bs);
calculate_curr_spec_pmc_values(bs, bin_increment);
fill_fval_vector_with_SQS(bs, samples[charge][sizeIndex][i]);
samples[charge][sizeIndex][i].label = label;
}
}
}
// cout sample composition
cout << "Sample composition:" << endl;
for (charge=0; charge<=maximalChargeWithModels_; charge++)
{
cout << charge;
int i;
for (i=0; i<samples[charge].size(); i++)
cout << "\t" << samples[charge][i].size();
cout << endl;
}
// create SQS models
this->sqs_models.resize(maximalChargeWithModels_+1);
for (charge =0; charge<=maximalChargeWithModels_; charge++)
{
sqs_models[charge].resize(maximalChargeWithModels_+1);
int j;
for (j=0; j<sqs_models[charge].size(); j++)
sqs_models[charge][j].resize(numSizes+1,NULL);
}
for (charge=1; charge<=maximalChargeWithModels_; charge++)
{
int sizeIndex;
for (sizeIndex=0; sizeIndex<=numSizes; sizeIndex++)
{
ME_Regression_DataSet ds;
cout << endl << "CHARGE " << charge << " SIZE " << sizeIndex << endl;
ds.num_classes=2;
ds.num_features=SQS_NUM_FIELDS;
ds.add_samples(samples[0][sizeIndex]);
ds.add_samples(samples[charge][sizeIndex]);
ds.tally_samples();
if (ds.class_weights[0]<0.0001 || ds.class_weights[1]<0.0001)
{
cout << "Warning: insufficient number of samples, not trianing model for this charge " << charge <<
" size " << sizeIndex << endl;
continue;
}
const double pos_weight = 0.2 + classWeights[charge][sizeIndex]*0.3;
ds.randomly_remove_samples_with_activated_feature(1,SQS_IND_MAX_TAG_LENGTH_ABOVE_4,0.5);
ds.calibrate_class_weights(pos_weight); // charge vs bad spectra
ds.print_feature_summary(cout,SQS_var_names);
sqs_models[charge][0][sizeIndex]=new ME_Regression_Model;
sqs_models[charge][0][sizeIndex]->train_cg(ds,250);
sqs_models[charge][0][sizeIndex]->print_ds_probs(ds);
}
}
////////////////////////////////////////////
// train model vs. model if charge1>charge2
if (1)
{
int charge1,charge2;
for (charge1=2; charge1<=maximalChargeWithModels_; charge1++)
{
for (charge2=1; charge2<charge1; charge2++)
{
int sizeIndex;
for (sizeIndex=0; sizeIndex<=numSizes; sizeIndex++)
{
ME_Regression_DataSet ds;
ds.num_classes=2;
ds.num_features=SQS_NUM_FIELDS;
ds.add_samples(samples[charge1][sizeIndex]);
int i;
for (i=0; i<samples[charge2][sizeIndex].size(); i++)
{
samples[charge2][sizeIndex][i].label=1;
ds.add_sample(samples[charge2][sizeIndex][i]);
samples[charge2][sizeIndex][i].label=0;
}
float relative_weight = classWeights[charge1][sizeIndex]/
(classWeights[charge1][sizeIndex]+classWeights[charge2][sizeIndex]);
ds.tally_samples();
if (ds.class_weights[0]<0.0001 || ds.class_weights[1]<0.0001)
{
cout << "Warning: insufficient number of samples, not trianing model for charge " << charge1 <<
" vs charge " << charge2<< " (size " << sizeIndex << ")" << endl;
continue;
}
ds.calibrate_class_weights(relative_weight);
sqs_models[charge1][charge2][sizeIndex] = new ME_Regression_Model;
cout << endl << "CHARGE " << charge1 << " vs " << charge2 << " size " << sizeIndex << endl;
cout << "Relative weights: " << charge1 << "/(" << charge1 << "+" <<
charge2 << "): " << relative_weight << endl;
ds.print_feature_summary(cout,SQS_var_names);
sqs_models[charge1][charge2][sizeIndex]->train_cg(ds,300);
sqs_models[charge1][charge2][sizeIndex]->print_ds_probs(ds);
}
}
}
}
init_sqs_correct_factors(maximalChargeWithModels_, sqsMassThresholds_.size());
////////////////////////////////////////////
// final report on datasets
cout << endl;
int sizeIndex;
for (sizeIndex=0; sizeIndex<=numSizes; sizeIndex++)
{
cout << endl << "SIZE: " << sizeIndex << endl;
cout << "--------" << endl;
float p_thresh = 0.05;
int d;
for (d=0; d<=maximalChargeWithModels_; d++)
{
vector<int> counts;
vector<int> max_counts;
counts.resize(maximalChargeWithModels_+1,0);
max_counts.resize(maximalChargeWithModels_+1,0);
int i;
for (i=0; i<samples[d][sizeIndex].size(); i++)
{
bool above_thresh=false;
float max_prob=0;
int max_class=0;
int c;
for (c=1; c<=maximalChargeWithModels_; c++)
{
if (! sqs_models[c][0][sizeIndex])
continue;
float prob = sqs_models[c][0][sizeIndex]->p_y_given_x(0,samples[d][sizeIndex][i]);
if (prob>p_thresh)
{
counts[c]++;
above_thresh=true;
if (prob>max_prob)
{
max_prob=prob;
max_class=c;
}
}
}
max_counts[max_class]++;
if (! above_thresh)
counts[0]++;
}
cout << d << "\t";
for (i=0; i<=maximalChargeWithModels_; i++)
cout << fixed << setprecision(4) << max_counts[i]/(float)samples[d][sizeIndex].size() << "\t";
cout << endl;
}
}
ind_initialized_sqs = true;
string path;
path = config->get_resource_dir() + "/" + config->get_model_name() + "_SQS.txt";
write_sqs_models(path.c_str());
}