public override IPeakScoringModel Train(IList <IList <float[]> > targets, IList <IList <float[]> > decoys, LinearModelParams initParameters,
bool includeSecondBest = false, bool preTrain = true, IProgressMonitor progressMonitor = null)
{
return(ChangeProp(ImClone(this), im =>
{
int nWeights = initParameters.Weights.Count;
var weights = new double [nWeights];
for (int i = 0; i < initParameters.Weights.Count; ++i)
{
weights[i] = double.IsNaN(initParameters.Weights[i]) ? double.NaN : DEFAULT_WEIGHTS[i];
}
var parameters = new LinearModelParams(weights);
ScoredGroupPeaksSet decoyTransitionGroups = new ScoredGroupPeaksSet(decoys);
ScoredGroupPeaksSet targetTransitionGroups = new ScoredGroupPeaksSet(targets);
targetTransitionGroups.ScorePeaks(parameters.Weights);
if (includeSecondBest)
{
ScoredGroupPeaksSet secondBestTransitionGroups;
targetTransitionGroups.SelectTargetsAndDecoys(out targetTransitionGroups, out secondBestTransitionGroups);
foreach (var secondBestGroup in secondBestTransitionGroups.ScoredGroupPeaksList)
{
decoyTransitionGroups.Add(secondBestGroup);
}
}
decoyTransitionGroups.ScorePeaks(parameters.Weights);
im.UsesDecoys = decoys.Count > 0;
im.UsesSecondBest = includeSecondBest;
im.Parameters = parameters.RescaleParameters(decoyTransitionGroups.Mean, decoyTransitionGroups.Stdev);
}));
}
private const int MAX_TRAINING_MEMORY = 512 * 1024 * 1024; // 512 MB
/// <summary>
/// Calculate new weight factors for one iteration of the refinement process. This is the heart
/// of the MProphet algorithm.
/// </summary>
/// <param name="iteration">Iteration number (special processing happens for iteration 0).</param>
/// <param name="targetTransitionGroups">Target transition groups.</param>
/// <param name="decoyTransitionGroups">Decoy transition groups.</param>
/// <param name="includeSecondBest">Include the second best peaks in the targets as additional decoys?</param>
/// <param name="weights">Array of weights per calculator.</param>
/// <param name="decoyMean">Output mean of decoy transition groups.</param>
/// <param name="decoyStdev">Output standard deviation of decoy transition groups.</param>
/// <param name="colinearWarning">Set to true if colinearity was detected.</param>
private void CalculateWeights(
int iteration,
ScoredGroupPeaksSet targetTransitionGroups,
ScoredGroupPeaksSet decoyTransitionGroups,
bool includeSecondBest,
double[] weights,
out double decoyMean,
out double decoyStdev,
ref bool colinearWarning)
{
if (includeSecondBest)
{
ScoredGroupPeaksSet secondBestTransitionGroups;
targetTransitionGroups.SelectTargetsAndDecoys(out targetTransitionGroups, out secondBestTransitionGroups);
foreach (var secondBestGroup in secondBestTransitionGroups.ScoredGroupPeaksList)
{
decoyTransitionGroups.Add(secondBestGroup);
}
}
// Select true target peaks using a q-value cutoff filter.
var qValueCutoff = (iteration == 0 ? 0.15 : 0.02);
var truePeaks = targetTransitionGroups.SelectTruePeaks(qValueCutoff, Lambda, decoyTransitionGroups);
var decoyPeaks = decoyTransitionGroups.SelectMaxPeaks();
// Omit first feature during first iteration, since it is used as the initial score value.
weights[0] = (iteration == 0) ? double.NaN : 0;
var featureCount = weights.Count(w => !double.IsNaN(w));
// Copy target and decoy peaks to training data array.
int totalTrainingPeaks = truePeaks.Count + decoyTransitionGroups.Count;
// Calculate the maximum number of training peaks (8 bytes per score - double, featurCount + 1 scores per peak)
int maxTrainingPeaks = MAX_TRAINING_MEMORY / 8 / (featureCount + 1);
var trainData = new double[Math.Min(totalTrainingPeaks, maxTrainingPeaks), featureCount + 1];
if (totalTrainingPeaks < maxTrainingPeaks)
{
for (int i = 0; i < truePeaks.Count; i++)
{
CopyToTrainData(truePeaks[i].Features, trainData, weights, i, 1);
}
for (int i = 0; i < decoyPeaks.Count; i++)
{
CopyToTrainData(decoyPeaks[i].Features, trainData, weights, i + truePeaks.Count, 0);
}
}
else
{
double proportionTrue = truePeaks.Count * 1.0 / totalTrainingPeaks;
int truePeakCount = (int)Math.Round(maxTrainingPeaks * proportionTrue);
int i = 0;
foreach (var peak in truePeaks.RandomOrder())
{
if (i < truePeakCount)
{
CopyToTrainData(peak.Features, trainData, weights, i, 1);
}
else
{
break;
}
i++;
}
int decoyPeakCount = maxTrainingPeaks - truePeakCount;
i = 0;
foreach (var peak in decoyPeaks.RandomOrder())
{
if (i < decoyPeakCount)
{
CopyToTrainData(peak.Features, trainData, weights, i + truePeakCount, 0);
}
else
{
break;
}
i++;
}
}
// Use Linear Discriminant Analysis to find weights that separate true and decoy peak scores.
int info;
double[] weightsFromLda;
alglib.fisherlda(
trainData,
trainData.GetLength(0),
trainData.GetLength(1) - 1,
2,
out info,
out weightsFromLda);
// Check for colinearity.
if (info == 2)
{
colinearWarning = true;
}
// Unpack weights array.
for (int i = 0, j = 0; i < weights.Length; i++)
{
if (!double.IsNaN(weights[i]))
{
weights[i] = weightsFromLda[j++];
}
}
// Recalculate all peak scores.
targetTransitionGroups.ScorePeaks(weights);
decoyTransitionGroups.ScorePeaks(weights);
// If the mean target score is less than the mean decoy score, then the
// weights came out negative, and all the weights and scores must be negated to
// restore the proper ordering.
if (targetTransitionGroups.Mean < decoyTransitionGroups.Mean)
{
for (int i = 0; i < weights.Length; i++)
{
weights[i] *= -1;
}
targetTransitionGroups.ScorePeaks(weights);
decoyTransitionGroups.ScorePeaks(weights);
}
decoyMean = decoyTransitionGroups.Mean;
decoyStdev = decoyTransitionGroups.Stdev;
}
/// <summary>
/// Train the model by iterative calculating weights to separate target and decoy transition groups.
/// </summary>
/// <param name="targets">Target transition groups.</param>
/// <param name="decoys">Decoy transition groups.</param>
/// <param name="initParameters">Initial model parameters (weights and bias)</param>
/// <param name="includeSecondBest"> Include the second best peaks in the targets as decoys?</param>
/// <param name="preTrain">Use a pre-trained model to bootstrap the learning.</param>
/// <param name="progressMonitor"></param>
/// <returns>Immutable model with new weights.</returns>
public override IPeakScoringModel Train(IList <IList <float[]> > targets, IList <IList <float[]> > decoys, LinearModelParams initParameters,
bool includeSecondBest = false, bool preTrain = true, IProgressMonitor progressMonitor = null)
{
if (initParameters == null)
{
initParameters = new LinearModelParams(_peakFeatureCalculators.Count);
}
return(ChangeProp(ImClone(this), im =>
{
targets = targets.Where(list => list.Count > 0).ToList();
decoys = decoys.Where(list => list.Count > 0).ToList();
var targetTransitionGroups = new ScoredGroupPeaksSet(targets);
var decoyTransitionGroups = new ScoredGroupPeaksSet(decoys);
// Bootstrap from the pre-trained legacy model
if (preTrain)
{
var preTrainedWeights = new double[initParameters.Weights.Count];
for (int i = 0; i < preTrainedWeights.Length; ++i)
{
if (double.IsNaN(initParameters.Weights[i]))
{
preTrainedWeights[i] = double.NaN;
}
}
int standardEnabledCount = GetEnabledCount(LegacyScoringModel.StandardFeatureCalculators, initParameters.Weights);
int analyteEnabledCount = GetEnabledCount(LegacyScoringModel.AnalyteFeatureCalculators, initParameters.Weights);
bool hasStandards = standardEnabledCount >= analyteEnabledCount;
var calculators = hasStandards ? LegacyScoringModel.StandardFeatureCalculators : LegacyScoringModel.AnalyteFeatureCalculators;
for (int i = 0; i < calculators.Length; ++i)
{
if (calculators[i].GetType() == typeof(MQuestRetentionTimePredictionCalc))
{
continue;
}
SetCalculatorValue(calculators[i].GetType(), LegacyScoringModel.DEFAULT_WEIGHTS[i], preTrainedWeights);
}
targetTransitionGroups.ScorePeaks(preTrainedWeights);
decoyTransitionGroups.ScorePeaks(preTrainedWeights);
}
// Iteratively refine the weights through multiple iterations.
var calcWeights = new double[initParameters.Weights.Count];
Array.Copy(initParameters.Weights.ToArray(), calcWeights, initParameters.Weights.Count);
double decoyMean = 0;
double decoyStdev = 0;
bool colinearWarning = false;
// This may take a long time between progress updates, but just measure progress by cycles through the training
IProgressStatus status = new ProgressStatus(Resources.MProphetPeakScoringModel_Train_Training_peak_scoring_model);
if (progressMonitor != null)
{
progressMonitor.UpdateProgress(status);
}
for (int iteration = 0; iteration < MAX_ITERATIONS; iteration++)
{
if (progressMonitor != null)
{
if (progressMonitor.IsCanceled)
{
throw new OperationCanceledException();
}
progressMonitor.UpdateProgress(status =
status.ChangeMessage(string.Format(Resources.MProphetPeakScoringModel_Train_Training_peak_scoring_model__iteration__0__of__1__, iteration + 1, MAX_ITERATIONS))
.ChangePercentComplete((iteration + 1) * 100 / (MAX_ITERATIONS + 1)));
}
im.CalculateWeights(iteration, targetTransitionGroups, decoyTransitionGroups,
includeSecondBest, calcWeights, out decoyMean, out decoyStdev, ref colinearWarning);
GC.Collect(); // Each loop generates a number of large objects. GC helps to keep private bytes under control
}
if (progressMonitor != null)
{
progressMonitor.UpdateProgress(status.ChangePercentComplete(100));
}
var parameters = new LinearModelParams(calcWeights);
parameters = parameters.RescaleParameters(decoyMean, decoyStdev);
im.Parameters = parameters;
im.ColinearWarning = colinearWarning;
im.UsesSecondBest = includeSecondBest;
im.UsesDecoys = decoys.Count > 0;
}));
}
private const int MAX_TRAINING_MEMORY = 512 * 1024 * 1024; // 512 MB
/// <summary>
/// Calculate new weight factors for one iteration of the refinement process. This is the heart
/// of the MProphet algorithm.
/// </summary>
/// <param name="documentPath">The path to the current document for writing score distributions</param>
/// <param name="targetTransitionGroups">Target transition groups.</param>
/// <param name="decoyTransitionGroups">Decoy transition groups.</param>
/// <param name="includeSecondBest">Include the second best peaks in the targets as additional decoys?</param>
/// <param name="nonParametricPValues">Non-parametric p values used in selecting true peaks if true</param>
/// <param name="qValueCutoff">The q value cut-off for true peaks in the training</param>
/// <param name="weights">Array of weights per calculator.</param>
/// <param name="decoyMean">Output mean of decoy transition groups.</param>
/// <param name="decoyStdev">Output standard deviation of decoy transition groups.</param>
/// <param name="colinearWarning">Set to true if colinearity was detected.</param>
private int CalculateWeights(string documentPath,
ScoredGroupPeaksSet targetTransitionGroups,
ScoredGroupPeaksSet decoyTransitionGroups,
bool includeSecondBest,
bool nonParametricPValues,
double qValueCutoff,
double[] weights,
out double decoyMean,
out double decoyStdev,
ref bool colinearWarning)
{
if (includeSecondBest)
{
ScoredGroupPeaksSet secondBestTransitionGroups;
targetTransitionGroups.SelectTargetsAndDecoys(out targetTransitionGroups, out secondBestTransitionGroups);
foreach (var secondBestGroup in secondBestTransitionGroups.ScoredGroupPeaksList)
{
decoyTransitionGroups.Add(secondBestGroup);
}
}
// Select true target peaks using a q-value cutoff filter.
var truePeaks = targetTransitionGroups.SelectTruePeaks(decoyTransitionGroups, qValueCutoff, Lambda, nonParametricPValues);
var decoyPeaks = decoyTransitionGroups.SelectMaxPeaks();
WriteDistributionInfo(documentPath, targetTransitionGroups, decoyTransitionGroups); // Only if asked to do so in command-line arguments
// Better to let a really poor model through for the user to see than to give an error message here
if (((double)truePeaks.Count) * 10 * 1000 < decoyPeaks.Count) // Targets must be at least 0.01% of decoys (still rejects zero)
{
throw new InvalidDataException(string.Format(Resources.MProphetPeakScoringModel_CalculateWeights_Insufficient_target_peaks___0__with__1__decoys__detected_at__2___FDR_to_continue_training_, truePeaks.Count, decoyPeaks.Count, qValueCutoff * 100));
}
if (((double)decoyPeaks.Count) * 1000 < truePeaks.Count) // Decoys must be at least 0.1% of targets
{
throw new InvalidDataException(string.Format(Resources.MProphetPeakScoringModel_CalculateWeights_Insufficient_decoy_peaks___0__with__1__targets__to_continue_training_, decoyPeaks.Count, truePeaks.Count));
}
var featureCount = weights.Count(w => !double.IsNaN(w));
// Copy target and decoy peaks to training data array.
int totalTrainingPeaks = truePeaks.Count + decoyTransitionGroups.Count;
// Calculate the maximum number of training peaks (8 bytes per score - double, featurCount + 1 scores per peak)
int maxTrainingPeaks = MAX_TRAINING_MEMORY / 8 / (featureCount + 1);
var trainData = new double[Math.Min(totalTrainingPeaks, maxTrainingPeaks), featureCount + 1];
if (totalTrainingPeaks < maxTrainingPeaks)
{
for (int i = 0; i < truePeaks.Count; i++)
{
CopyToTrainData(truePeaks[i].Features, trainData, weights, i, 1);
}
for (int i = 0; i < decoyPeaks.Count; i++)
{
CopyToTrainData(decoyPeaks[i].Features, trainData, weights, i + truePeaks.Count, 0);
}
}
else
{
double proportionTrue = truePeaks.Count * 1.0 / totalTrainingPeaks;
int truePeakCount = (int)Math.Round(maxTrainingPeaks * proportionTrue);
int i = 0;
foreach (var peak in truePeaks.RandomOrder(ArrayUtil.RANDOM_SEED))
{
if (i < truePeakCount)
{
CopyToTrainData(peak.Features, trainData, weights, i, 1);
}
else
{
break;
}
i++;
}
int decoyPeakCount = maxTrainingPeaks - truePeakCount;
i = 0;
foreach (var peak in decoyPeaks.RandomOrder(ArrayUtil.RANDOM_SEED))
{
if (i < decoyPeakCount)
{
CopyToTrainData(peak.Features, trainData, weights, i + truePeakCount, 0);
}
else
{
break;
}
i++;
}
}
// Use Linear Discriminant Analysis to find weights that separate true and decoy peak scores.
int info;
double[] weightsFromLda;
alglib.fisherlda(
trainData,
trainData.GetLength(0),
trainData.GetLength(1) - 1,
2,
out info,
out weightsFromLda);
// Check for colinearity.
if (info == 2)
{
colinearWarning = true;
}
// Unpack weights array.
for (int i = 0, j = 0; i < weights.Length; i++)
{
if (!double.IsNaN(weights[i]))
{
weights[i] = weightsFromLda[j++];
}
}
// Recalculate all peak scores.
targetTransitionGroups.ScorePeaks(weights);
decoyTransitionGroups.ScorePeaks(weights);
// If the mean target score is less than the mean decoy score, then the
// weights came out negative, and all the weights and scores must be negated to
// restore the proper ordering.
if (targetTransitionGroups.Mean < decoyTransitionGroups.Mean)
{
for (int i = 0; i < weights.Length; i++)
{
weights[i] *= -1;
}
targetTransitionGroups.ScorePeaks(weights);
decoyTransitionGroups.ScorePeaks(weights);
}
decoyMean = decoyTransitionGroups.Mean;
decoyStdev = decoyTransitionGroups.Stdev;
return(truePeaks.Count);
}
/// <summary>
/// Train the model by iterative calculating weights to separate target and decoy transition groups.
/// </summary>
/// <param name="targetsIn">Target transition groups.</param>
/// <param name="decoysIn">Decoy transition groups.</param>
/// <param name="initParameters">Initial model parameters (weights and bias)</param>
/// <param name="iterations">Optional specific number of iterations to use in training</param>
/// <param name="includeSecondBest">Include the second best peaks in the targets as decoys?</param>
/// <param name="preTrain">Use a pre-trained model to bootstrap the learning.</param>
/// <param name="progressMonitor">Used to report progress to the calling context</param>
/// <param name="documentPath">The path to the current document for writing score distributions</param>
/// <returns>Immutable model with new weights.</returns>
public override IPeakScoringModel Train(IList <IList <float[]> > targetsIn,
IList <IList <float[]> > decoysIn,
LinearModelParams initParameters,
int?iterations = null,
bool includeSecondBest = false,
bool preTrain = true,
IProgressMonitor progressMonitor = null,
string documentPath = null)
{
if (initParameters == null)
{
initParameters = new LinearModelParams(_peakFeatureCalculators.Count);
}
return(ChangeProp(ImClone(this), im =>
{
// This may take a long time between progress updates, but just measure progress by cycles through the training
IProgressStatus status = new ProgressStatus(Resources.MProphetPeakScoringModel_Train_Training_peak_scoring_model);
if (progressMonitor != null)
{
progressMonitor.UpdateProgress(status);
}
var targets = targetsIn.Where(list => list.Count > 0);
var decoys = decoysIn.Where(list => list.Count > 0);
var targetTransitionGroups = new ScoredGroupPeaksSet(targets, targetsIn.Count);
var decoyTransitionGroups = new ScoredGroupPeaksSet(decoys, decoysIn.Count);
// Iteratively refine the weights through multiple iterations.
var calcWeights = new double[initParameters.Weights.Count];
Array.Copy(initParameters.Weights.ToArray(), calcWeights, initParameters.Weights.Count);
double qValueCutoff = 0.01; // First iteration cut-off - if not pretraining, just start at 0.01
// Start with scores calculated from the initial weights
if (!preTrain)
{
targetTransitionGroups.ScorePeaks(calcWeights);
decoyTransitionGroups.ScorePeaks(calcWeights);
}
// Bootstrap from the pre-trained legacy model
else
{
qValueCutoff = 0.15;
var preTrainedWeights = new double[initParameters.Weights.Count];
for (int i = 0; i < preTrainedWeights.Length; ++i)
{
if (double.IsNaN(initParameters.Weights[i]))
{
preTrainedWeights[i] = double.NaN;
}
}
int standardEnabledCount = GetEnabledCount(LegacyScoringModel.StandardFeatureCalculators, initParameters.Weights);
int analyteEnabledCount = GetEnabledCount(LegacyScoringModel.AnalyteFeatureCalculators, initParameters.Weights);
bool hasStandards = standardEnabledCount >= analyteEnabledCount;
var calculators = hasStandards ? LegacyScoringModel.StandardFeatureCalculators : LegacyScoringModel.AnalyteFeatureCalculators;
for (int i = 0; i < calculators.Length; ++i)
{
if (calculators[i].GetType() == typeof(MQuestRetentionTimePredictionCalc))
{
continue;
}
SetCalculatorValue(calculators[i].GetType(), LegacyScoringModel.DEFAULT_WEIGHTS[i], preTrainedWeights);
}
targetTransitionGroups.ScorePeaks(preTrainedWeights);
decoyTransitionGroups.ScorePeaks(preTrainedWeights);
}
double decoyMean = 0;
double decoyStdev = 0;
bool colinearWarning = false;
int iterationCount = iterations ?? MAX_ITERATIONS;
int truePeaksCount = 0;
var lastWeights = new double[calcWeights.Length];
for (int i = 0; i < iterationCount; i++)
{
int percentComplete = 0;
double decoyMeanNew, decoyStdevNew;
bool colinearWarningNew = colinearWarning;
int truePeaksCountNew = im.CalculateWeights(documentPath,
targetTransitionGroups,
decoyTransitionGroups,
includeSecondBest,
i == 0, // Use non-parametric q values for first round, when normality assumption may not hold
qValueCutoff,
calcWeights,
out decoyMeanNew,
out decoyStdevNew,
ref colinearWarningNew);
if (progressMonitor != null)
{
if (progressMonitor.IsCanceled)
{
throw new OperationCanceledException();
}
// Calculate progress, but wait to make sure convergence has not occurred before setting it
string formatText = qValueCutoff > 0.02
? Resources.MProphetPeakScoringModel_Train_Training_scoring_model__iteration__0__of__1__
: Resources.MProphetPeakScoringModel_Train_Training_scoring_model__iteration__0__of__1_____2______peaks_at__3_0_____FDR_;
percentComplete = (i + 1) * 100 / (iterationCount + 1);
status = status.ChangeMessage(string.Format(formatText, i + 1, iterationCount, truePeaksCountNew, qValueCutoff))
.ChangePercentComplete(percentComplete);
}
if (qValueCutoff > 0.02)
{
// Tighten the q value cut-off for "truth" to 2% FDR
qValueCutoff = 0.02;
// And allow the true peaks count to go down in the next iteration
// Though it rarely will
truePeaksCountNew = 0;
}
else if (truePeaksCountNew < truePeaksCount)
{
// The model has leveled off enough to begin losing discriminant value
if (qValueCutoff > 0.01)
{
// Tighten the q value cut-off for "truth" to 1% FDR
qValueCutoff = 0.01;
// And allow the true peaks count to go down in the next iteration
truePeaksCountNew = 0;
}
else
{
if (progressMonitor != null)
{
progressMonitor.UpdateProgress(status =
status.ChangeMessage(string.Format(Resources.MProphetPeakScoringModel_Train_Scoring_model_converged__iteration__0_____1______peaks_at__2_0_____FDR_, i + 1, truePeaksCount, qValueCutoff))
.ChangePercentComplete(Math.Max(95, percentComplete)));
}
calcWeights = lastWeights;
break;
}
}
truePeaksCount = truePeaksCountNew;
Array.Copy(calcWeights, lastWeights, calcWeights.Length);
decoyMean = decoyMeanNew;
decoyStdev = decoyStdevNew;
colinearWarning = colinearWarningNew;
if (progressMonitor != null)
{
progressMonitor.UpdateProgress(status);
}
}
if (progressMonitor != null)
{
progressMonitor.UpdateProgress(status.ChangePercentComplete(100));
}
var parameters = new LinearModelParams(calcWeights);
parameters = parameters.RescaleParameters(decoyMean, decoyStdev);
im.Parameters = parameters;
im.ColinearWarning = colinearWarning;
im.UsesSecondBest = includeSecondBest;
im.UsesDecoys = decoysIn.Count > 0;
}));
}
