void DPTreephaser::QueryAllStates(BasecallerRead& data, vector< vector<float> >& query_states, vector<int>& hp_lengths, int max_flows)
{
max_flows = min(max_flows,flow_order_.num_flows());
InitializeState(&path_[0]);
max_flows = min(max_flows, flow_order_.num_flows());
query_states.reserve(data.sequence.size());
query_states.resize(0);
hp_lengths.assign(data.sequence.size(), 0);
char last_nuc = 'N';
int hp_count = 0;
for (vector<char>::iterator nuc = data.sequence.begin(); nuc != data.sequence.end() and path_[0].flow < max_flows; ++nuc) {
if (last_nuc != *nuc and last_nuc != 'N') {
hp_lengths[hp_count] = path_[0].last_hp;
query_states.push_back(path_[0].state);
hp_count++;
}
AdvanceStateInPlace(&path_[0], *nuc, max_flows);
last_nuc = *nuc;
}
hp_lengths[hp_count] = path_[0].last_hp;
query_states.push_back(path_[0].state);
hp_lengths.resize(query_states.size());
data.prediction.swap(path_[0].prediction);
}
void DPTreephaser::QueryState(BasecallerRead& data, vector<float>& query_state, int& current_hp, int max_flows, int query_flow)
{
max_flows = min(max_flows,flow_order_.num_flows());
assert(query_flow < max_flows);
InitializeState(&path_[0]);
query_state.assign(max_flows,0);
char myNuc = 'N';
for (vector<char>::iterator nuc = data.sequence.begin(); nuc != data.sequence.end() and path_[0].flow <= query_flow; ++nuc) {
if (path_[0].flow == query_flow and myNuc != 'N' and myNuc != *nuc)
break;
AdvanceStateInPlace(&path_[0], *nuc, flow_order_.num_flows());
if (path_[0].flow == query_flow and myNuc == 'N')
myNuc = *nuc;
}
// Catching cases where a query_flow without incorporation or query_flow after end of sequence was given
int until_flow = min(path_[0].window_end, max_flows);
if (path_[0].flow == query_flow) {
current_hp = path_[0].last_hp;
for (int flow = path_[0].window_start; flow < until_flow; ++flow)
query_state[flow] = path_[0].state[flow];
}
else
current_hp = 0;
}
void DPTreephaser::Simulate(BasecallerRead& data, int max_flows)
{
InitializeState(&path_[0]);
for (vector<char>::iterator nuc = data.sequence.begin(); nuc != data.sequence.end() and path_[0].flow < max_flows; ++nuc)
AdvanceStateInPlace(&path_[0], *nuc, flow_order_.num_flows());
data.prediction.swap(path_[0].prediction);
}
void DPTreephaser::Simulate(BasecallerRead& data, int max_flows)
{
max_flows = min(max_flows,flow_order_.num_flows());
InitializeState(&path_[0]);
for (int solution_flow = 0; solution_flow < max_flows; ++solution_flow)
for (int hp = 0; hp < data.solution[solution_flow]; ++hp)
AdvanceStateInPlace(&path_[0], flow_order_.int_at(solution_flow), flow_order_.num_flows());
data.prediction.swap(path_[0].prediction);
}
void DPTreephaser::Simulate(BasecallerRead& data, int max_flows,bool state_inphase)
{
InitializeState(&path_[0]);
for (vector<char>::iterator nuc = data.sequence.begin(); nuc != data.sequence.end()
and path_[0].flow < max_flows; ++nuc) {
AdvanceStateInPlace(&path_[0], *nuc, flow_order_.num_flows());
path_[0].sequence.push_back(*nuc); // Needed to simulate diagonal states correctly
if (state_inphase and path_[0].flow < max_flows)
data.state_inphase.at(path_[0].flow) = path_[0].state.at(path_[0].flow);
}
data.prediction.swap(path_[0].prediction);
}
void DPTreephaser::SimulateRecalibrated(BasecallerRead& data, int max_flows)
{
InitializeState(&path_[0]);
// Generate predicted signal
for (vector<char>::iterator nuc = data.sequence.begin(); nuc != data.sequence.end() and path_[0].flow < max_flows; ++nuc) {
int flow_s = path_[0].flow;
AdvanceStateInPlace(&path_[0], *nuc, flow_order_.num_flows());
if (path_[0].flow < flow_order_.num_flows()) {
//update flow2base_pos
path_[0].base_pos++;
path_[0].flow2base_pos[path_[0].flow] = path_[0].base_pos;
for(int flow_inter = flow_s+1; flow_inter < path_[0].flow; flow_inter++)
path_[0].flow2base_pos[flow_inter] = path_[0].flow2base_pos[flow_s];
}
}
// Apply signal distortion according to HP recalibration model
if (pm_model_available_) {
for (int flow = 0; flow < flow_order_.num_flows(); ++flow) {
int hp_length = 0;
if(flow == 0)
hp_length = path_[0].flow2base_pos[0];
else
hp_length =path_[0].flow2base_pos[flow] - path_[0].flow2base_pos[flow-1];
if(hp_length < 0)
hp_length = 0;
if(hp_length > MAX_HPXLEN)
hp_length = MAX_HPXLEN;
path_[0].prediction[flow] = path_[0].prediction[flow] * (*As_)[flow][flow_order_.int_at(flow)][hp_length]
+ (*Bs_)[flow][flow_order_.int_at(flow)][hp_length];
}
}
data.prediction.swap(path_[0].prediction);
}
void DPTreephaser::Solve(BasecallerRead& read, int max_flows, int restart_flows)
{
static const char nuc_int_to_char[5] = "ACGT";
assert(max_flows <= flow_order_.num_flows());
// Initialize stack: just one root path
for (int p = 1; p < kNumPaths; ++p)
path_[p].in_use = false;
InitializeState(&path_[0]);
path_[0].path_metric = 0;
path_[0].per_flow_metric = 0;
path_[0].residual_left_of_window = 0;
path_[0].dot_counter = 0;
path_[0].in_use = true;
//path_[0].sequence.reserve(2*flow_order_.num_flows()); //Done in InitializeState
int space_on_stack = kNumPaths - 1;
float sum_of_squares_upper_bound = 1e20; //max_flows; // Squared distance of solution to measurements
if (restart_flows > 0) {
// The solver will not attempt to solve initial restart_flows
// - Simulate restart_flows instead of solving
// - If it turns out that solving was finished before restart_flows, simply exit without any changes to the read.
restart_flows = min(restart_flows, flow_order_.num_flows());
for (vector<char>::iterator nuc = read.sequence.begin(); nuc != read.sequence.end() and path_[0].flow < restart_flows; ++nuc) {
int flow_s = path_[0].flow;
AdvanceStateInPlace(&path_[0], *nuc, flow_order_.num_flows());
if (path_[0].flow < flow_order_.num_flows()) {
path_[0].sequence.push_back(*nuc);
//update flow2base_pos
path_[0].base_pos++;
path_[0].flow2base_pos[path_[0].flow] = path_[0].base_pos;
for(int flow_inter = flow_s+1; flow_inter < path_[0].flow; flow_inter++)
path_[0].flow2base_pos[flow_inter] = path_[0].flow2base_pos[flow_s];
}
}
if (path_[0].flow < restart_flows-10) { // This read ended before restart_flows. No point resolving it.
read.prediction.swap(path_[0].prediction);
return;
}
for (int flow = 0; flow < path_[0].window_start; ++flow) {
float residual = 0;
if(pm_model_enabled_==false){
residual = read.normalized_measurements[flow] - path_[0].prediction[flow];
}
else{
int hp_length = 0;
if(flow==0) hp_length = path_[0].flow2base_pos[0];
else hp_length =path_[0].flow2base_pos[flow] - path_[0].flow2base_pos[flow-1];
if(hp_length<0) hp_length = 0;
residual = read.normalized_measurements[flow]
- (path_[0].prediction[flow] * (*As_)[flow][flow_order_.int_at(flow)][hp_length]
+ (*Bs_)[flow][flow_order_.int_at(flow)][hp_length]);
}
path_[0].residual_left_of_window += residual * residual;
}
}
// Initializing variables
//read.solution.assign(flow_order_.num_flows(), 0);
read.sequence.clear();
read.sequence.reserve(2*flow_order_.num_flows());
read.prediction.assign(flow_order_.num_flows(), 0);
// Main loop to select / expand / delete paths
while (1) {
// ------------------------------------------
// Step 1: Prune the content of the stack and make sure there are at least 4 empty slots
// Remove paths that are more than 'maxPathDelay' behind the longest one
if (space_on_stack < kNumPaths-3) {
int longest_path = 0;
for (int p = 0; p < kNumPaths; ++p)
if (path_[p].in_use)
longest_path = max(longest_path, path_[p].flow);
if (longest_path > kMaxPathDelay) {
for (int p = 0; p < kNumPaths; ++p) {
if (path_[p].in_use and path_[p].flow < longest_path-kMaxPathDelay) {
path_[p].in_use = false;
space_on_stack++;
}
}
}
}
// If necessary, remove paths with worst perFlowMetric
while (space_on_stack < 4) {
// find maximum per flow metric
float max_per_flow_metric = -0.1;
int max_metric_path = kNumPaths;
for (int p = 0; p < kNumPaths; ++p) {
if (path_[p].in_use and path_[p].per_flow_metric > max_per_flow_metric) {
max_per_flow_metric = path_[p].per_flow_metric;
max_metric_path = p;
}
}
// killing path with largest per flow metric
if (!(max_metric_path < kNumPaths)) {
printf("Failed assertion in Treephaser\n");
for (int p = 0; p < kNumPaths; ++p) {
if (path_[p].in_use)
printf("Path %d, in_use = true, per_flow_metric = %f\n", p, path_[p].per_flow_metric);
else
printf("Path %d, in_use = false, per_flow_metric = %f\n", p, path_[p].per_flow_metric);
}
fflush(NULL);
}
assert (max_metric_path < kNumPaths);
path_[max_metric_path].in_use = false;
space_on_stack++;
}
// ------------------------------------------
// Step 2: Select a path to expand or break if there is none
TreephaserPath *parent = NULL;
float min_path_metric = 1000;
for (int p = 0; p < kNumPaths; ++p) {
if (path_[p].in_use and path_[p].path_metric < min_path_metric) {
min_path_metric = path_[p].path_metric;
parent = &path_[p];
}
}
if (!parent)
break;
// ------------------------------------------
// Step 3: Construct four expanded paths and calculate feasibility metrics
assert (space_on_stack >= 4);
TreephaserPath *children[4];
for (int nuc = 0, p = 0; nuc < 4; ++p)
if (not path_[p].in_use)
children[nuc++] = &path_[p];
float penalty[4] = { 0, 0, 0, 0 };
for (int nuc = 0; nuc < 4; ++nuc) {
TreephaserPath *child = children[nuc];
AdvanceState(child, parent, nuc_int_to_char[nuc], max_flows);
// Apply easy termination rules
if (child->flow >= max_flows) {
penalty[nuc] = 25; // Mark for deletion
continue;
}
if (child->last_hp > kMaxHP) {
penalty[nuc] = 25; // Mark for deletion
continue;
}
if ((int)parent->sequence.size() >= (2 * flow_order_.num_flows() - 10)) {
penalty[nuc] = 25; // Mark for deletion
continue;
}
child->path_metric = parent->residual_left_of_window;
child->residual_left_of_window = parent->residual_left_of_window;
float penaltyN = 0;
float penalty1 = 0;
for (int flow = parent->window_start; flow < child->window_end; ++flow) {
float residual = 0;
if(pm_model_enabled_==false){
residual = read.normalized_measurements[flow] - child->prediction[flow];
}
else{
int hp_length = 0;
if(flow==0)
hp_length = parent->flow2base_pos[0];
else
hp_length = parent->flow2base_pos[flow] - parent->flow2base_pos[flow-1];
if(hp_length<0) hp_length = 0;
residual = read.normalized_measurements[flow]
- (child->prediction[flow] * (*As_)[flow][flow_order_.int_at(flow)][hp_length]
+ (*Bs_)[flow][flow_order_.int_at(flow)][hp_length]);
}
float residual_squared = residual * residual;
// Metric calculation
if (flow < child->window_start) {
child->residual_left_of_window += residual_squared;
child->path_metric += residual_squared;
} else if (residual <= 0)
child->path_metric += residual_squared;
if (residual <= 0)
penaltyN += residual_squared;
else if (flow < child->flow)
penalty1 += residual_squared;
}
penalty[nuc] = penalty1 + kNegativeMultiplier * penaltyN;
penalty1 += penaltyN;
if (child->flow>0)
child->per_flow_metric = (child->path_metric + 0.5 * penalty1) / child->flow;
} //looping over nucs
// Find out which nuc has the least penalty (the greedy choice nuc)
int best_nuc = 0;
if (penalty[best_nuc] > penalty[1])
best_nuc = 1;
if (penalty[best_nuc] > penalty[2])
best_nuc = 2;
if (penalty[best_nuc] > penalty[3])
best_nuc = 3;
// ------------------------------------------
// Step 4: Use calculated metrics to decide which paths are worth keeping
for (int nuc = 0; nuc < 4; ++nuc) {
TreephaserPath *child = children[nuc];
// Path termination rules
if (penalty[nuc] >= 20)
continue;
if (child->path_metric > sum_of_squares_upper_bound)
continue;
// This is the only rule that depends on finding the "best nuc"
if (penalty[nuc] - penalty[best_nuc] >= kExtendThreshold)
continue;
float dot_signal = 0;
if(pm_model_enabled_==false){
dot_signal = (read.normalized_measurements[child->flow] - parent->prediction[child->flow]) / child->state[child->flow];
}
else{
int hp_length = 0;
if(child->flow==0) hp_length = parent->flow2base_pos[0];
else hp_length = parent->flow2base_pos[child->flow] - parent->flow2base_pos[child->flow-1];
if(hp_length<0) hp_length = 0;
dot_signal = (read.normalized_measurements[child->flow] - (parent->prediction[child->flow]
* (*As_)[child->flow][flow_order_.int_at(child->flow)][hp_length]
+ (*Bs_)[child->flow][flow_order_.int_at(child->flow)][hp_length]))
/ child->state[child->flow];
}
child->dot_counter = (dot_signal < kDotThreshold) ? (parent->dot_counter + 1) : 0;
if (child->dot_counter > 1)
continue;
// Path survived termination rules and will be kept on stack
child->in_use = true;
space_on_stack--;
// Fill out the remaining portion of the prediction
memcpy(&child->prediction[0], &parent->prediction[0], parent->window_start*sizeof(float));
for (int flow = child->window_end; flow < max_flows; ++flow)
child->prediction[flow] = 0;
// Fill out the solution
child->sequence = parent->sequence;
child->sequence.push_back(nuc_int_to_char[nuc]);
//calculate starting base position for each flow
child->base_pos = parent->base_pos;
child->base_pos++;
child->flow2base_pos = parent->flow2base_pos;
for(int flow_inter = parent->flow+1; flow_inter < child->flow; flow_inter++){
child->flow2base_pos[flow_inter] = child->flow2base_pos[parent->flow];
}
child->flow2base_pos[child->flow] = child->base_pos;
}
// ------------------------------------------
// Step 5. Check if the selected path is in fact the best path so far
// Computing sequence squared distance
float sum_of_squares = parent->residual_left_of_window;
for (int flow = parent->window_start; flow < max_flows; flow++) {
float residual = 0;
if(pm_model_enabled_==false){
residual = read.normalized_measurements[flow] - parent->prediction[flow];
}
else{
int hp_length = 0;
if(flow==0) hp_length = parent->flow2base_pos[0];
else hp_length = parent->flow2base_pos[flow] - parent->flow2base_pos[flow-1];
if(hp_length<0) hp_length = 0;
residual = read.normalized_measurements[flow] - (parent->prediction[flow]
* (*As_)[flow][flow_order_.int_at(flow)][hp_length]
+ (*Bs_)[flow][flow_order_.int_at(flow)][hp_length]);
}
sum_of_squares += residual * residual;
}
// Updating best path
if (sum_of_squares < sum_of_squares_upper_bound) {
read.prediction.swap(parent->prediction);
read.sequence.swap(parent->sequence);
sum_of_squares_upper_bound = sum_of_squares;
}
parent->in_use = false;
space_on_stack++;
} // main decision loop
}
