Adding ability to specify in VQSR a list of bad sites to use when training the negative model. Just add bad=true to the list of rod tags for your bad sites track.

public/java/src/org/broadinstitute/sting/gatk/walkers/variantrecalibration/TrainingSet.java

@@ -14,6 +14,7 @@ public class TrainingSet {
     public String name;
     public boolean isKnown = false;
     public boolean isTraining = false;
+    public boolean isAntiTraining = false;
     public boolean isTruth = false;
     public boolean isConsensus = false;
     public double prior = 0.0;
@@ -22,17 +23,24 @@ public class TrainingSet {
 
     public TrainingSet( final String name, final Tags tags ) {
         this.name = name;
+
+        // Parse the tags to decide which tracks have which properties
         if( tags != null ) {
             isKnown = tags.containsKey("known") && tags.getValue("known").equals("true");
             isTraining = tags.containsKey("training") && tags.getValue("training").equals("true");
+            isAntiTraining = tags.containsKey("bad") && tags.getValue("bad").equals("true");
             isTruth = tags.containsKey("truth") && tags.getValue("truth").equals("true");
             isConsensus = tags.containsKey("consensus") && tags.getValue("consensus").equals("true");
             prior = ( tags.containsKey("prior") ? Double.parseDouble(tags.getValue("prior")) : prior );
         }
-        if( !isConsensus ) {
+
+        // Report back to the user which tracks were found and the properties that were detected
+        if( !isConsensus && !isAntiTraining ) {
             logger.info( String.format( "Found %s track: \tKnown = %s \tTraining = %s \tTruth = %s \tPrior = Q%.1f", this.name, isKnown, isTraining, isTruth, prior) );
-        } else {
+        } else if( isConsensus ) {
             logger.info( String.format( "Found consensus track: %s", this.name) );
+        } else {
+            logger.info( String.format( "Found bad sites training track: %s", this.name) );
         }
     }
 }

public/java/src/org/broadinstitute/sting/gatk/walkers/variantrecalibration/VariantDataManager.java

@@ -84,7 +84,6 @@ public class VariantDataManager {
                 remove = remove || (Math.abs(val) > VRAC.STD_THRESHOLD);
             }
             datum.failingSTDThreshold = remove;
-            datum.usedForTraining = 0;
         }
     }
 
@@ -118,7 +117,6 @@ public class VariantDataManager {
         for( final VariantDatum datum : data ) {
             if( datum.atTrainingSite && !datum.failingSTDThreshold && datum.originalQual > VRAC.QUAL_THRESHOLD ) {
                 trainingData.add( datum );
-                datum.usedForTraining = 1;
             }
         }
         logger.info( "Training with " + trainingData.size() + " variants after standard deviation thresholding." );
@@ -129,27 +127,37 @@ public class VariantDataManager {
     }
 
     public ExpandingArrayList<VariantDatum> selectWorstVariants( double bottomPercentage, final int minimumNumber ) {
-        Collections.sort( data );
+        // The return value is the list of training variants
         final ExpandingArrayList<VariantDatum> trainingData = new ExpandingArrayList<VariantDatum>();
-        final int numToAdd = Math.max( minimumNumber, Math.round((float)bottomPercentage * data.size()) );
+
+        // First add to the training list all sites overlapping any bad sites training tracks
+        for( final VariantDatum datum : data ) {
+            if( datum.atAntiTrainingSite && !datum.failingSTDThreshold && !Double.isInfinite(datum.lod) ) {
+                trainingData.add( datum );
+            }
+        }
+        final int numBadSitesAdded = trainingData.size();
+        logger.info( "Found " + numBadSitesAdded + " variants overlapping bad sites training tracks." );
+
+        // Next, sort the variants by the LOD coming from the positive model and add to the list the bottom X percent of variants
+        Collections.sort( data );
+        final int numToAdd = Math.max( minimumNumber - trainingData.size(), Math.round((float)bottomPercentage * data.size()) );
         if( numToAdd > data.size() ) {
             throw new UserException.BadInput( "Error during negative model training. Minimum number of variants to use in training is larger than the whole call set. One can attempt to lower the --minNumBadVariants arugment but this is unsafe." );
-        }
+        } else if( numToAdd == minimumNumber - trainingData.size() ) {
-        if( numToAdd == minimumNumber ) {
             logger.warn( "WARNING: Training with very few variant sites! Please check the model reporting PDF to ensure the quality of the model is reliable." );
             bottomPercentage = ((float) numToAdd) / ((float) data.size());
         }
-        int index = 0;
+        int index = 0, numAdded = 0;
-        int numAdded = 0;
         while( numAdded < numToAdd ) {
             final VariantDatum datum = data.get(index++);
             if( !datum.failingSTDThreshold && !Double.isInfinite(datum.lod) ) {
+                datum.atAntiTrainingSite = true;
                 trainingData.add( datum );
-                datum.usedForTraining = -1;
                 numAdded++;
             }
         }
-        logger.info("Training with worst " + (float) bottomPercentage * 100.0f + "% of passing data --> " + trainingData.size() + " variants with LOD <= " + String.format("%.4f", data.get(index).lod) + ".");
+        logger.info( "Additionally training with worst " + (float) bottomPercentage * 100.0f + "% of passing data --> " + (trainingData.size() - numBadSitesAdded) + " variants with LOD <= " + String.format("%.4f", data.get(index).lod) + "." );
         return trainingData;
     }
 
@@ -162,10 +170,11 @@ public class VariantDataManager {
                 returnData.add(datum);
             }
         }
-        // add an extra 5% of points from bad training set, since that set is small but interesting
+
+        // Add an extra 5% of points from bad training set, since that set is small but interesting
         for( int iii = 0; iii < Math.floor(0.05*numToAdd); iii++) {
             final VariantDatum datum = data.get(GenomeAnalysisEngine.getRandomGenerator().nextInt(data.size()));
-            if( datum.usedForTraining == -1 && !datum.failingSTDThreshold ) { returnData.add(datum); }
+            if( datum.atAntiTrainingSite && !datum.failingSTDThreshold ) { returnData.add(datum); }
             else { iii--; }
         }
 
@@ -236,6 +245,7 @@ public class VariantDataManager {
         datum.atTrainingSite = false;
         datum.prior = 2.0;
         datum.consensusCount = 0;
+
         for( final TrainingSet trainingSet : trainingSets ) {
             for( final VariantContext trainVC : tracker.getVariantContexts( ref, trainingSet.name, null, context.getLocation(), false, false ) ) {
                 if( trainVC != null && trainVC.isNotFiltered() && trainVC.isVariant() &&
@@ -248,6 +258,10 @@ public class VariantDataManager {
                     datum.prior = Math.max( datum.prior, trainingSet.prior );
                     datum.consensusCount += ( trainingSet.isConsensus ? 1 : 0 );
                 }
+                if( trainVC != null ) {
+                    datum.atAntiTrainingSite = datum.atAntiTrainingSite || trainingSet.isAntiTraining;
+                }
+
             }
         }
     }

public/java/src/org/broadinstitute/sting/gatk/walkers/variantrecalibration/VariantDatum.java

@@ -14,13 +14,13 @@ public class VariantDatum implements Comparable<VariantDatum> {
     public double lod;
     public boolean atTruthSite;
     public boolean atTrainingSite;
+    public boolean atAntiTrainingSite;
     public boolean isTransition;
     public boolean isSNP;
     public boolean failingSTDThreshold;
     public double originalQual;
     public double prior;
     public int consensusCount;
-    public int usedForTraining;
     public String contig;
     public int start;
     public int stop;

public/java/src/org/broadinstitute/sting/gatk/walkers/variantrecalibration/VariantRecalibrator.java

@@ -175,7 +175,6 @@ public class VariantRecalibrator extends RodWalker<ExpandingArrayList<VariantDat
                     datum.originalQual = vc.getPhredScaledQual();
                     datum.isSNP = vc.isSNP() && vc.isBiallelic();
                     datum.isTransition = datum.isSNP && VariantContextUtils.isTransition(vc);
-                    datum.usedForTraining = 0;
 
                     // Loop through the training data sets and if they overlap this loci then update the prior and training status appropriately
                     dataManager.parseTrainingSets( tracker, ref, context, vc, datum, TRUST_ALL_POLYMORPHIC );
@@ -328,7 +327,8 @@ public class VariantRecalibrator extends RodWalker<ExpandingArrayList<VariantDat
 
                 stream.print("data <- c(");
                 for( final VariantDatum datum : randomData ) {
-                    stream.print(String.format("%.3f, %.3f, %.3f, %d, %d,", datum.annotations[iii], datum.annotations[jjj], (datum.lod < lodCutoff ? -1.0 : 1.0), datum.usedForTraining, (datum.isKnown ? 1 : -1)));
+                    stream.print(String.format("%.3f, %.3f, %.3f, %d, %d,", datum.annotations[iii], datum.annotations[jjj], (datum.lod < lodCutoff ? -1.0 : 1.0),
+                            (datum.atAntiTrainingSite ? -1 : (datum.atTrainingSite ? 1 : 0)), (datum.isKnown ? 1 : -1)));
                 }
                 stream.println("NA,NA,NA,NA,1)");
                 stream.println("d <- matrix(data,ncol=5,byrow=T)");