void StablePairFinder::run(const std::vector<ConsensusMap>& input_maps,
ConsensusMap& result_map)
{
// empty output destination:
result_map.clear(false);
// sanity checks:
if (input_maps.size() != 2)
{
throw Exception::IllegalArgument(__FILE__, __LINE__, __PRETTY_FUNCTION__,
"exactly two input maps required");
}
checkIds_(input_maps);
// set up the distance functor:
double max_intensity = max(input_maps[0].getMaxInt(),
input_maps[1].getMaxInt());
Param distance_params = param_.copy("");
distance_params.remove("use_identifications");
distance_params.remove("second_nearest_gap");
FeatureDistance feature_distance(max_intensity, false);
feature_distance.setParameters(distance_params);
// keep track of pairing:
std::vector<bool> is_singleton[2];
is_singleton[0].resize(input_maps[0].size(), true);
is_singleton[1].resize(input_maps[1].size(), true);
typedef pair<double, double> DoublePair;
DoublePair init = make_pair(FeatureDistance::infinity,
FeatureDistance::infinity);
// for every element in map 0:
// - index of nearest neighbor in map 1:
vector<UInt> nn_index_0(input_maps[0].size(), UInt(-1));
// - distances to nearest and second-nearest neighbors in map 1:
vector<DoublePair> nn_distance_0(input_maps[0].size(), init);
// for every element in map 1:
// - index of nearest neighbor in map 0:
vector<UInt> nn_index_1(input_maps[1].size(), UInt(-1));
// - distances to nearest and second-nearest neighbors in map 0:
vector<DoublePair> nn_distance_1(input_maps[1].size(), init);
// iterate over all feature pairs, find nearest neighbors:
// TODO: iterate over SENSIBLE RT (and m/z) window -- sort the maps beforehand
// to save a lot of processing time...
// Once done, remove the warning in the description of the 'use_identifications' parameter
for (UInt fi0 = 0; fi0 < input_maps[0].size(); ++fi0)
{
const ConsensusFeature& feat0 = input_maps[0][fi0];
for (UInt fi1 = 0; fi1 < input_maps[1].size(); ++fi1)
{
const ConsensusFeature& feat1 = input_maps[1][fi1];
if (use_IDs_ && !compatibleIDs_(feat0, feat1)) // check peptide IDs
{
continue; // mismatch
}
pair<bool, double> result = feature_distance(feat0, feat1);
double distance = result.second;
// we only care if distance constraints are satisfied for "best
// matches", not for second-best; this means that second-best distances
// can become smaller than best distances
// (e.g. the RT is larger than allowed (->invalid pair), but m/z is perfect and has the most weight --> better score!)
bool valid = result.first;
// update entries for map 0:
if (distance < nn_distance_0[fi0].second)
{
if (valid && (distance < nn_distance_0[fi0].first))
{
nn_distance_0[fi0].second = nn_distance_0[fi0].first;
nn_distance_0[fi0].first = distance;
nn_index_0[fi0] = fi1;
}
else
{
nn_distance_0[fi0].second = distance;
}
}
// update entries for map 1:
if (distance < nn_distance_1[fi1].second)
{
if (valid && (distance < nn_distance_1[fi1].first))
{
nn_distance_1[fi1].second = nn_distance_1[fi1].first;
nn_distance_1[fi1].first = distance;
nn_index_1[fi1] = fi0;
}
else
{
nn_distance_1[fi1].second = distance;
}
}
}
}
// if features from the two maps are nearest neighbors of each other, they
// can become a pair:
for (UInt fi0 = 0; fi0 < input_maps[0].size(); ++fi0)
{
UInt fi1 = nn_index_0[fi0]; // nearest neighbor of "fi0" in map 1
// cout << "index: " << fi0 << ", RT: " << input_maps[0][fi0].getRT()
// << ", MZ: " << input_maps[0][fi0].getMZ() << endl
// << "neighbor: " << fi1 << ", RT: " << input_maps[1][fi1].getRT()
// << ", MZ: " << input_maps[1][fi1].getMZ() << endl
// << "d(i,j): " << nn_distance_0[fi0].first << endl
// << "d2(i): " << nn_distance_0[fi0].second << endl
// << "d2(j): " << nn_distance_1[fi1].second << endl;
// criteria set by the parameters must be fulfilled:
if ((nn_distance_0[fi0].first < FeatureDistance::infinity) &&
(nn_distance_0[fi0].first * second_nearest_gap_ <= nn_distance_0[fi0].second))
{
// "fi0" satisfies constraints...
if ((nn_index_1[fi1] == fi0) &&
(nn_distance_1[fi1].first * second_nearest_gap_ <= nn_distance_1[fi1].second))
{
// ...nearest neighbor of "fi0" also satisfies constraints (yay!)
// cout << "match!" << endl;
result_map.push_back(ConsensusFeature());
ConsensusFeature& f = result_map.back();
f.insert(input_maps[0][fi0]);
f.getPeptideIdentifications().insert(f.getPeptideIdentifications().end(),
input_maps[0][fi0].getPeptideIdentifications().begin(),
input_maps[0][fi0].getPeptideIdentifications().end());
f.insert(input_maps[1][fi1]);
f.getPeptideIdentifications().insert(f.getPeptideIdentifications().end(),
input_maps[1][fi1].getPeptideIdentifications().begin(),
input_maps[1][fi1].getPeptideIdentifications().end());
f.computeConsensus();
double quality = 1.0 - nn_distance_0[fi0].first;
double quality0 = 1.0 - nn_distance_0[fi0].first * second_nearest_gap_ / nn_distance_0[fi0].second;
double quality1 = 1.0 - nn_distance_1[fi1].first * second_nearest_gap_ / nn_distance_1[fi1].second;
quality = quality * quality0 * quality1; // TODO other formula?
// incorporate existing quality values:
Size size0 = max(input_maps[0][fi0].size(), size_t(1));
Size size1 = max(input_maps[1][fi1].size(), size_t(1));
// quality contribution from first map:
quality0 = input_maps[0][fi0].getQuality() * (size0 - 1);
// quality contribution from second map:
quality1 = input_maps[1][fi1].getQuality() * (size1 - 1);
f.setQuality((quality + quality0 + quality1) / (size0 + size1 - 1));
is_singleton[0][fi0] = false;
is_singleton[1][fi1] = false;
}
}
}
// write out unmatched consensus features
for (UInt input = 0; input <= 1; ++input)
{
for (UInt index = 0; index < input_maps[input].size(); ++index)
{
if (is_singleton[input][index])
{
result_map.push_back(input_maps[input][index]);
if (result_map.back().size() < 2) // singleton consensus feature
{
result_map.back().setQuality(0.0);
}
}
}
}
// canonical ordering for checking the results, and the ids have no real meaning anyway
result_map.sortByMZ();
// protein IDs and unassigned peptide IDs are added to the result by the
// FeatureGroupingAlgorithm!
}
void LabeledPairFinder::run(const vector<ConsensusMap>& input_maps, ConsensusMap& result_map)
{
if (input_maps.size() != 1)
throw Exception::IllegalArgument(__FILE__, __LINE__, __PRETTY_FUNCTION__, "exactly one input map required");
if (result_map.getFileDescriptions().size() != 2)
throw Exception::IllegalArgument(__FILE__, __LINE__, __PRETTY_FUNCTION__, "two file descriptions required");
if (result_map.getFileDescriptions().begin()->second.filename != result_map.getFileDescriptions().rbegin()->second.filename)
throw Exception::IllegalArgument(__FILE__, __LINE__, __PRETTY_FUNCTION__, "the two file descriptions have to contain the same file name");
checkIds_(input_maps);
//look up the light and heavy index
Size light_index = numeric_limits<Size>::max();
Size heavy_index = numeric_limits<Size>::max();
for (ConsensusMap::FileDescriptions::const_iterator it = result_map.getFileDescriptions().begin();
it != result_map.getFileDescriptions().end();
++it)
{
if (it->second.label == "heavy")
{
heavy_index = it->first;
}
else if (it->second.label == "light")
{
light_index = it->first;
}
}
if (light_index == numeric_limits<Size>::max() || heavy_index == numeric_limits<Size>::max())
{
throw Exception::IllegalArgument(__FILE__, __LINE__, __PRETTY_FUNCTION__, "the input maps have to be labeled 'light' and 'heavy'");
}
result_map.clear(false);
// sort consensus features by RT (and MZ) to speed up searching afterwards
typedef ConstRefVector<ConsensusMap> RefMap;
RefMap model_ref(input_maps[0].begin(), input_maps[0].end());
model_ref.sortByPosition();
//calculate matches
ConsensusMap matches;
//settings
double rt_pair_dist = param_.getValue("rt_pair_dist");
double rt_dev_low = param_.getValue("rt_dev_low");
double rt_dev_high = param_.getValue("rt_dev_high");
double mz_dev = param_.getValue("mz_dev");
DoubleList mz_pair_dists = param_.getValue("mz_pair_dists");
bool mrm = param_.getValue("mrm").toBool();
//estimate RT parameters
if (param_.getValue("rt_estimate") == "true")
{
//find all possible RT distances of features with the same charge and a good m/z distance
vector<double> dists;
dists.reserve(model_ref.size());
for (RefMap::const_iterator it = model_ref.begin(); it != model_ref.end(); ++it)
{
for (RefMap::const_iterator it2 = model_ref.begin(); it2 != model_ref.end(); ++it2)
{
for (DoubleList::const_iterator dist_it = mz_pair_dists.begin(); dist_it != mz_pair_dists.end(); ++dist_it)
{
double mz_pair_dist = *dist_it;
if (it2->getCharge() == it->getCharge()
&& it2->getMZ() >= it->getMZ() + mz_pair_dist / it->getCharge() - mz_dev
&& it2->getMZ() <= it->getMZ() + mz_pair_dist / it->getCharge() + mz_dev)
{
dists.push_back(it2->getRT() - it->getRT());
}
}
}
}
if (dists.empty())
{
cout << "Warning: Could not find pairs for RT distance estimation. The manual settings are used!" << endl;
}
else
{
if (dists.size() < 50)
{
cout << "Warning: Found only " << dists.size() << " pairs. The estimated shift and std deviation are probably not reliable!" << endl;
}
//--------------------------- estimate initial parameters of fit ---------------------------
GaussFitter::GaussFitResult result(-1, -1, -1);
//first estimate of the optimal shift: median of the distances
sort(dists.begin(), dists.end());
Size median_index = dists.size() / 2;
result.x0 = dists[median_index];
//create histogram of distances
//consider only the maximum of pairs, centered around the optimal shift
Size max_pairs = model_ref.size() / 2;
Size start_index = (Size) max((SignedSize)0, (SignedSize)(median_index - max_pairs / 2));
Size end_index = (Size) min((SignedSize)(dists.size() - 1), (SignedSize)(median_index + max_pairs / 2));
double start_value = dists[start_index];
double end_value = dists[end_index];
double bin_step = fabs(end_value - start_value) / 99.999; //ensure that we have 100 bins
Math::Histogram<> hist(start_value, end_value, bin_step);
//std::cout << "HIST from " << start_value << " to " << end_value << " (bin size " << bin_step << ")" << endl;
for (Size i = start_index; i <= end_index; ++i)
{
hist.inc(dists[i]);
}
//cout << hist << endl;
dists.clear();
//determine median of bins (uniform background distribution)
vector<Size> bins(hist.begin(), hist.end());
sort(bins.begin(), bins.end());
Size bin_median = bins[bins.size() / 2];
bins.clear();
//estimate scale A: maximum of the histogram
Size max_value = hist.maxValue();
result.A = max_value - bin_median;
//overwrite estimate of x0 with the position of the highest bin
for (Size i = 0; i < hist.size(); ++i)
{
if (hist[i] == max_value)
{
result.x0 = hist.centerOfBin(i);
break;
}
}
//estimate sigma: first time the count is less or equal the median count in the histogram
double pos = result.x0;
while (pos > start_value && hist.binValue(pos) > bin_median)
{
pos -= bin_step;
}
double sigma_low = result.x0 - pos;
pos = result.x0;
while (pos<end_value&& hist.binValue(pos)> bin_median)
{
pos += bin_step;
}
double sigma_high = pos - result.x0;
result.sigma = (sigma_high + sigma_low) / 6.0;
//cout << "estimated optimal RT distance (before fit): " << result.x0 << endl;
//cout << "estimated allowed deviation (before fit): " << result.sigma*3.0 << endl;
//--------------------------- do gauss fit ---------------------------
vector<DPosition<2> > points(hist.size());
for (Size i = 0; i < hist.size(); ++i)
{
points[i][0] = hist.centerOfBin(i);
points[i][1] = max(0u, hist[i]);
}
GaussFitter fitter;
fitter.setInitialParameters(result);
result = fitter.fit(points);
cout << "estimated optimal RT distance: " << result.x0 << endl;
cout << "estimated allowed deviation: " << fabs(result.sigma) * 3.0 << endl;
rt_pair_dist = result.x0;
rt_dev_low = fabs(result.sigma) * 3.0;
rt_dev_high = fabs(result.sigma) * 3.0;
}
}
// check each feature
for (RefMap::const_iterator it = model_ref.begin(); it != model_ref.end(); ++it)
{
for (DoubleList::const_iterator dist_it = mz_pair_dists.begin(); dist_it != mz_pair_dists.end(); ++dist_it)
{
double mz_pair_dist = *dist_it;
RefMap::const_iterator it2 = lower_bound(model_ref.begin(), model_ref.end(), it->getRT() + rt_pair_dist - rt_dev_low, ConsensusFeature::RTLess());
while (it2 != model_ref.end() && it2->getRT() <= it->getRT() + rt_pair_dist + rt_dev_high)
{
// if in mrm mode, we need to compare precursor mass difference and fragment mass difference, charge remains the same
double prec_mz_diff(0);
if (mrm)
{
prec_mz_diff = fabs((double)it2->getMetaValue("MZ") - (double)it->getMetaValue("MZ"));
if (it->getCharge() != 0)
{
prec_mz_diff = fabs(prec_mz_diff - mz_pair_dist / it->getCharge());
}
else
{
prec_mz_diff = fabs(prec_mz_diff - mz_pair_dist);
}
}
bool mrm_correct_dist(false);
double frag_mz_diff = fabs(it->getMZ() - it2->getMZ());
//cerr << it->getRT() << " charge1=" << it->getCharge() << ", charge2=" << it2->getCharge() << ", prec_diff=" << prec_mz_diff << ", frag_diff=" << frag_mz_diff << endl;
if (mrm &&
it2->getCharge() == it->getCharge() &&
prec_mz_diff < mz_dev &&
(frag_mz_diff < mz_dev || fabs(frag_mz_diff - mz_pair_dist) < mz_dev))
{
mrm_correct_dist = true;
//cerr << "mrm_correct_dist" << endl;
}
if ((mrm && mrm_correct_dist) || (!mrm &&
it2->getCharge() == it->getCharge() &&
it2->getMZ() >= it->getMZ() + mz_pair_dist / it->getCharge() - mz_dev &&
it2->getMZ() <= it->getMZ() + mz_pair_dist / it->getCharge() + mz_dev
))
{
//cerr << "dist correct" << endl;
double score = sqrt(
PValue_(it2->getMZ() - it->getMZ(), mz_pair_dist / it->getCharge(), mz_dev, mz_dev) *
PValue_(it2->getRT() - it->getRT(), rt_pair_dist, rt_dev_low, rt_dev_high)
);
// Note: we used to copy the id from the light feature here, but that strategy does not generalize to more than two labels.
// We might want to report consensus features where the light one is missing but more than one heavier variant was found.
// Also, the old strategy is inconsistent with what was done in the unlabeled case. Thus now we assign a new unique id here.
matches.push_back(ConsensusFeature());
matches.back().setUniqueId();
matches.back().insert(light_index, *it);
matches.back().clearMetaInfo();
matches.back().insert(heavy_index, *it2);
matches.back().setQuality(score);
matches.back().setCharge(it->getCharge());
matches.back().computeMonoisotopicConsensus();
}
++it2;
}
}
}
//compute best pairs
// - sort matches by quality
// - take highest-quality matches first (greedy) and mark them as used
set<Size> used_features;
matches.sortByQuality(true);
for (ConsensusMap::const_iterator match = matches.begin(); match != matches.end(); ++match)
{
//check if features are not used yet
if (used_features.find(match->begin()->getUniqueId()) == used_features.end() &&
used_features.find(match->rbegin()->getUniqueId()) == used_features.end()
)
{
//if unused, add it to the final set of elements
result_map.push_back(*match);
used_features.insert(match->begin()->getUniqueId());
used_features.insert(match->rbegin()->getUniqueId());
}
}
//Add protein identifications to result map
for (Size i = 0; i < input_maps.size(); ++i)
{
result_map.getProteinIdentifications().insert(result_map.getProteinIdentifications().end(), input_maps[i].getProteinIdentifications().begin(), input_maps[i].getProteinIdentifications().end());
}
//Add unassigned peptide identifications to result map
for (Size i = 0; i < input_maps.size(); ++i)
{
result_map.getUnassignedPeptideIdentifications().insert(result_map.getUnassignedPeptideIdentifications().end(), input_maps[i].getUnassignedPeptideIdentifications().begin(), input_maps[i].getUnassignedPeptideIdentifications().end());
}
// Very useful for checking the results, and the ids have no real meaning anyway
result_map.sortByMZ();
}