From 6323fb8673d067cd3465cce8786b076ceec05c6b Mon Sep 17 00:00:00 2001
From: rpoplin <rpoplin@348d0f76-0448-11de-a6fe-93d51630548a>
Date: Tue, 3 May 2011 18:00:22 +0000
Subject: [PATCH] misc cleanup in VQSR

git-svn-id: file:///humgen/gsa-scr1/gsa-engineering/svn_contents/trunk@5732 348d0f76-0448-11de-a6fe-93d51630548a
---
 .../ContrastiveRecalibrator.java              |  7 +++----
 .../GaussianMixtureModel.java                 |  8 ++++----
 .../MultivariateGaussian.java                 |  1 +
 .../VariantDataManager.java                   |  2 +-
 .../VariantRecalibratorEngine.java            | 19 +++++--------------
 5 files changed, 14 insertions(+), 23 deletions(-)

diff --git a/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/ContrastiveRecalibrator.java b/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/ContrastiveRecalibrator.java
index 692aae3a3..8311faba6 100755
--- a/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/ContrastiveRecalibrator.java
+++ b/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/ContrastiveRecalibrator.java
@@ -210,7 +210,7 @@ public class ContrastiveRecalibrator extends RodWalker<ExpandingArrayList<Varian
 
     public void onTraversalDone( final ExpandingArrayList<VariantDatum> reduceSum ) {
         dataManager.setData( reduceSum );
-        dataManager.normalizeData();
+        dataManager.normalizeData(); // Each data point is now (x - mean) / standard deviation
         final GaussianMixtureModel goodModel = engine.generateModel( dataManager.getTrainingData() );
         engine.evaluateData( dataManager.getData(), goodModel, false );
         final GaussianMixtureModel badModel = engine.generateModel( dataManager.selectWorstVariants( VRAC.PERCENT_BAD_VARIANTS ) );
@@ -250,12 +250,10 @@ public class ContrastiveRecalibrator extends RodWalker<ExpandingArrayList<Varian
 
         createArrangeFunction( stream );
 
-        stream.println("pdf(\"" + RSCRIPT_FILE + ".pdf\")");
+        stream.println("pdf(\"" + RSCRIPT_FILE + ".pdf\")"); // Unfortunately this is a huge pdf file, BUGBUG: need to work on reducing the file size
 
         for(int iii = 0; iii < USE_ANNOTATIONS.length; iii++) {
             for( int jjj = iii + 1; jjj < USE_ANNOTATIONS.length; jjj++) {
-                //stream.println("png(\"" + RSCRIPT_FILE + "." + USE_ANNOTATIONS[iii] + "." + USE_ANNOTATIONS[jjj] + ".png\", type=\"cairo\", width = 960, height = 960)");
-                //stream.println("pdf(\"" + RSCRIPT_FILE + "." + USE_ANNOTATIONS[iii] + "." + USE_ANNOTATIONS[jjj] + ".pdf\")");
                 logger.info( "Building " + USE_ANNOTATIONS[iii] + " x " + USE_ANNOTATIONS[jjj] + " plot...");
 
                 final ExpandingArrayList<VariantDatum> fakeData = new ExpandingArrayList<VariantDatum>();
@@ -266,6 +264,7 @@ public class ContrastiveRecalibrator extends RodWalker<ExpandingArrayList<Varian
                     minAnn2 = Math.min(minAnn2, datum.annotations[jjj]);
                     maxAnn2 = Math.max(maxAnn2, datum.annotations[jjj]);
                 }
+                // Create a fake set of data which spans the full extent of these two annotation dimensions in order to calculate the model PDF projected to 2D
                 for(double ann1 = minAnn1; ann1 <= maxAnn1; ann1+=0.1) {
                     for(double ann2 = minAnn2; ann2 <= maxAnn2; ann2+=0.1) {
                         final VariantDatum datum = new VariantDatum();
diff --git a/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/GaussianMixtureModel.java b/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/GaussianMixtureModel.java
index 64bb3755a..58f05633d 100755
--- a/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/GaussianMixtureModel.java
+++ b/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/GaussianMixtureModel.java
@@ -42,9 +42,6 @@ public class GaussianMixtureModel {
         empiricalMu = new double[numAnnotations];
         empiricalSigma = new Matrix(numAnnotations, numAnnotations);
         isModelReadyForEvaluation = false;
-    }
-
-    public void cacheEmpiricalStats() {
         Arrays.fill(empiricalMu, 0.0);
         empiricalSigma.setMatrix(0, empiricalMu.length - 1, 0, empiricalMu.length - 1, Matrix.identity(empiricalMu.length, empiricalMu.length).times(200.0).inverse());
     }
@@ -75,6 +72,7 @@ public class GaussianMixtureModel {
 
         int ttt = 0;
         while( ttt++ < numIterations ) {
+            // Estep: assign each variant to the nearest cluster
             for( final VariantDatum datum : data ) {
                 double minDistance = Double.MAX_VALUE;
                 MultivariateGaussian minGaussian = null;
@@ -89,6 +87,7 @@ public class GaussianMixtureModel {
                 datum.assignment = minGaussian;
             }
 
+            // Mstep: update gaussian means based on assigned variants
             for( final MultivariateGaussian gaussian : gaussians ) {
                 gaussian.zeroOutMu();
                 int numAssigned = 0;
@@ -188,7 +187,7 @@ public class GaussianMixtureModel {
         for( final MultivariateGaussian gaussian : gaussians ) {
             pVarInGaussianLog10[gaussianIndex++] = gaussian.pMixtureLog10 + gaussian.evaluateDatumLog10( datum );
         }
-        return MathUtils.log10sumLog10(pVarInGaussianLog10);
+        return MathUtils.log10sumLog10(pVarInGaussianLog10); // Sum(pi_k * p(v|n,k))
     }
 
     public double evaluateDatumMarginalized( final VariantDatum datum ) {
@@ -197,6 +196,7 @@ public class GaussianMixtureModel {
         int numIter = 10;
         final double[] pVarInGaussianLog10 = new double[gaussians.size()];
         for( int iii = 0; iii < datum.annotations.length; iii++ ) {
+            // marginalize over the missing dimension by drawing X random values for the missing annotation and averaging the lod
             if( datum.isNull[iii] ) {
                 for( int ttt = 0; ttt < numIter; ttt++ ) {
                     datum.annotations[iii] = Normal.staticNextDouble(0.0, 1.0);
diff --git a/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/MultivariateGaussian.java b/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/MultivariateGaussian.java
index b8e8fd28a..78935fb2d 100755
--- a/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/MultivariateGaussian.java
+++ b/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/MultivariateGaussian.java
@@ -94,6 +94,7 @@ public class MultivariateGaussian {
     }
 
     public void precomputeDenominatorForVariationalBayes( final double sumHyperParameterLambda ) {
+        // Variational Bayes calculations from Bishop
         cachedSigmaInverse = sigma.inverse();
         cachedSigmaInverse.timesEquals( hyperParameter_a );
         double sum = 0.0;
diff --git a/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/VariantDataManager.java b/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/VariantDataManager.java
index b46fe536f..fd7a548d1 100755
--- a/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/VariantDataManager.java
+++ b/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/VariantDataManager.java
@@ -73,7 +73,7 @@ public class VariantDataManager {
             throw new UserException.BadInput( "Found annotations with zero variance. They must be excluded before proceeding." );
         }
 
-        // trim data by standard deviation threshold and place into two sets: data and failingData
+        // trim data by standard deviation threshold and mark failing data for exclusion later
         for( final VariantDatum datum : data ) {
             boolean remove = false;
             for( final double val : datum.annotations ) {
diff --git a/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/VariantRecalibratorEngine.java b/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/VariantRecalibratorEngine.java
index 2a2565df8..b3284efc4 100755
--- a/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/VariantRecalibratorEngine.java
+++ b/java/src/org/broadinstitute/sting/playground/gatk/walkers/variantrecalibration/VariantRecalibratorEngine.java
@@ -29,7 +29,6 @@ public class VariantRecalibratorEngine {
 
     public VariantRecalibratorEngine( final VariantRecalibratorArgumentCollection VRAC ) {
         this.VRAC = VRAC;
-        initialize( this.VRAC );
     }
 
     public GaussianMixtureModel generateModel( final List<VariantDatum> data ) {
@@ -50,36 +49,28 @@ public class VariantRecalibratorEngine {
         }
     }
 
-    /////////////////////////////
-    // Private Methods used for initialization
-    /////////////////////////////
-
-    private void initialize( final VariantRecalibratorArgumentCollection VRAC ) {
-    }
-
     /////////////////////////////
     // Private Methods used for generating a GaussianMixtureModel
     /////////////////////////////
 
     private void variationalBayesExpectationMaximization( final GaussianMixtureModel model, final List<VariantDatum> data ) {
 
-        model.cacheEmpiricalStats();
         model.initializeRandomModel( data, VRAC.NUM_KMEANS_ITERATIONS );
 
         // The VBEM loop
         model.normalizePMixtureLog10();
         model.expectationStep( data );
-        double currentLikelihood;
+        double currentChangeInMixtureCoefficients;
         int iteration = 0;
         logger.info("Finished iteration " + iteration );
         while( iteration < VRAC.MAX_ITERATIONS ) {
             iteration++;
             model.maximizationStep( data );
-            currentLikelihood = model.normalizePMixtureLog10();
-            model.expectationStep( data );
-            logger.info("Current change in mixture coefficients = " + String.format("%.5f", currentLikelihood));
+            currentChangeInMixtureCoefficients = model.normalizePMixtureLog10();
+            model.expectationStep(data);
+            logger.info("Current change in mixture coefficients = " + String.format("%.5f", currentChangeInMixtureCoefficients));
             logger.info("Finished iteration " + iteration );
-            if( iteration > 2 && currentLikelihood < MIN_PROB_CONVERGENCE ) {
+            if( iteration > 2 && currentChangeInMixtureCoefficients < MIN_PROB_CONVERGENCE ) {
                 logger.info("Convergence!");
                 break;
             }