// approximately unimodal under most pictures of the world
// scan at maximum and build up a picture of the likelihood using log(n)*constant measurements, interpolate to linearize
void PosteriorInference::InterpolateFrequencyScan(ShortStack &total_theory, bool update_frequency, int strand_key) {
unsigned int num_reads = ResizeToMatch(total_theory);
float fnum_reads = (float) num_reads;
UpdateMaxFreqFromResponsibility(total_theory, strand_key);
int eval_start = (int)(max_freq * num_reads);
vector<unsigned int> samples;
FibInterval(samples, eval_start, num_reads);
unsigned int i_last = 0;
eval_at_frequency[i_last] = (float)i_last / fnum_reads;
log_posterior_by_frequency[i_last] = total_theory.PosteriorFrequencyLogLikelihood(eval_at_frequency[i_last], data_reliability, strand_key);
int bottom = log_posterior_by_frequency[i_last];
int top = bottom;
for (unsigned int i_dx = 1; i_dx < samples.size(); i_dx++) {
unsigned int i_eval = samples[i_dx];
eval_at_frequency[i_eval] = (float)i_eval / fnum_reads;
log_posterior_by_frequency[i_eval] = total_theory.PosteriorFrequencyLogLikelihood(eval_at_frequency[i_eval],data_reliability, strand_key);
top = log_posterior_by_frequency[i_eval];
for (unsigned int i_mid = i_last + 1; i_mid < i_eval; i_mid++) {
int delta_low = i_mid - i_last;
int delta_hi = i_eval - i_last;
eval_at_frequency[i_mid] = (float)i_mid / fnum_reads;
log_posterior_by_frequency[i_mid] = (top * delta_low + bottom * delta_hi) / (delta_low + delta_hi);
}
bottom = top;
i_last = i_eval;
}
FindMaxFrequency(update_frequency);
scan_done = true;
};
// do a hard classification as though the reads were independent
// i.e. look more like the data in the BAM file
void PosteriorInference::StartAtHardClassify(ShortStack &total_theory, bool update_frequency, float start_frequency) {
// just to allocate
ResizeToMatch(total_theory);
if (update_frequency) {
max_freq = start_frequency;
max_ll = total_theory.PosteriorFrequencyLogLikelihood(max_freq, data_reliability, ALL_STRAND_KEY);
}
total_theory.UpdateResponsibility(max_freq, data_reliability);
}
void PosteriorInference::DoPosteriorFrequencyScan(ShortStack &total_theory, bool update_frequency, int strand_key) {
//cout << "ScanningFrequency" << endl;
// posterior frequency inference given current data/likelihood pairing
unsigned int num_reads = ResizeToMatch(total_theory);
float fnum_reads = (float) num_reads;
for (unsigned int i_eval = 0; i_eval < eval_at_frequency.size(); i_eval++) {
eval_at_frequency[i_eval] = (float)i_eval / fnum_reads;
log_posterior_by_frequency[i_eval] = total_theory.PosteriorFrequencyLogLikelihood(eval_at_frequency[i_eval], data_reliability, strand_key);
}
// if doing monomorphic eval, set frequency to begin with and don't update
FindMaxFrequency(update_frequency);
scan_done = true;
// log_posterior now contains all frequency information inferred from the data
}
void PosteriorInference::UpdateMaxFreqFromResponsibility(ShortStack &total_theory, int strand_key) {
// skip time consuming scan and use responsibilities as cluster entry
float max_freq = total_theory.FrequencyFromResponsibility(strand_key);
max_ll = total_theory.PosteriorFrequencyLogLikelihood(max_freq, data_reliability, strand_key);
scan_done = false; // didn't come from scan
}