/// <summary>
/// http://www.broadinstitute.org/gsa/wiki/index.php/Understanding_the_Unified_Genotyper%27s_VCF_files
/// See section on Strand Bias
/// </summary>
// From GATK source:
//double forwardLod = forwardLog10PofF + reverseLog10PofNull - overallLog10PofF;
//double reverseLod = reverseLog10PofF + forwardLog10PofNull - overallLog10PofF;
//
//// strand score is max bias between forward and reverse strands
//double strandScore = Math.max(forwardLod, reverseLod);
//
//// rescale by a factor of 10
//strandScore *= 10.0;
//
//attributes.put("SB", strandScore);
private static double[] AssignBiasScore(StrandBiasStats overallStats, StrandBiasStats fwdStats, StrandBiasStats rvsStats)
{
var forwardBias = (fwdStats.ChanceVarFreqGreaterThanZero * rvsStats.ChanceFalsePos) /
overallStats.ChanceVarFreqGreaterThanZero;
var reverseBias = (rvsStats.ChanceVarFreqGreaterThanZero * fwdStats.ChanceFalsePos) /
overallStats.ChanceVarFreqGreaterThanZero;
if (overallStats.ChanceVarFreqGreaterThanZero == 0)
{
forwardBias = 1;
reverseBias = 1;
}
var p = Math.Max(forwardBias, reverseBias);
return(new[] { p, MathOperations.PtoGATKBiasScale(p) });
}
public static StrandBiasStats DeepCopy(StrandBiasStats originalStats)
{
if (originalStats == null)
{
return(null);
}
var newStats = new StrandBiasStats(originalStats.Support, originalStats.Coverage)
{
ChanceFalseNeg = originalStats.ChanceFalseNeg,
ChanceFalsePos = originalStats.ChanceFalsePos,
ChanceVarFreqGreaterThanZero = originalStats.ChanceVarFreqGreaterThanZero,
Frequency = originalStats.Frequency,
};
return(newStats);
}
public static void PopulateStats(StrandBiasStats stats, double noiseFreq, double minDetectableSNP,
StrandBiasModel strandBiasModel)
{
if (stats.Support == 0)
{
if (strandBiasModel == StrandBiasModel.Poisson)
{
stats.ChanceFalsePos = 1;
stats.ChanceVarFreqGreaterThanZero = 0;
stats.ChanceFalseNeg = 0;
}
else if (strandBiasModel == StrandBiasModel.Extended)
{
//the chance that we observe the SNP is (minDetectableSNPfreq) for one observation.
//the chance that we do not is (1- minDetectableSNPfreq) for one observation.
//the chance that we do not observe it, N times in a row is:
stats.ChanceVarFreqGreaterThanZero = (Math.Pow(1 - minDetectableSNP, stats.Coverage)); //used in SB metric
//liklihood that variant really does not exist
//= 1 - chance that it does but you did not see it
stats.ChanceFalsePos = 1 - stats.ChanceVarFreqGreaterThanZero; //used in SB metric
//Chance a low freq variant is at work in the model, and we did not observe it:
stats.ChanceFalseNeg = stats.ChanceVarFreqGreaterThanZero;
}
}
else
{
// chance of these observations or less, given min observable variant distribution
stats.ChanceVarFreqGreaterThanZero = Poisson.Cdf(stats.Support - 1, stats.Coverage * noiseFreq); //used in SB metric
stats.ChanceFalsePos = 1 - stats.ChanceVarFreqGreaterThanZero; //used in SB metric
stats.ChanceFalseNeg = Poisson.Cdf(stats.Support, stats.Coverage * minDetectableSNP);
}
//Note:
//
// Type 1 error is when we rejected the null hypothesis when we should not have. (we have noise, but called a SNP)
// Type 2 error is when we accepected the alternate when we should not have. (we have a variant, but we did not call it.)
//
// Type 1 error is our this.ChanceFalsePos aka p-value.
// Type 2 error is out this.ChanceFalseNeg
}
public static StrandBiasResults DeepCopy(StrandBiasResults originalSBresults)
{
if (originalSBresults == null)
{
return(null);
}
var sb = new StrandBiasResults()
{
BiasAcceptable = originalSBresults.BiasAcceptable,
BiasScore = originalSBresults.BiasScore,
GATKBiasScore = originalSBresults.GATKBiasScore,
VarPresentOnBothStrands = originalSBresults.VarPresentOnBothStrands,
CovPresentOnBothStrands = originalSBresults.CovPresentOnBothStrands,
TestAcceptable = originalSBresults.TestAcceptable,
TestScore = originalSBresults.TestScore,
ForwardStats = StrandBiasStats.DeepCopy(originalSBresults.ForwardStats),
OverallStats = StrandBiasStats.DeepCopy(originalSBresults.OverallStats),
ReverseStats = StrandBiasStats.DeepCopy(originalSBresults.ReverseStats),
StitchedStats = StrandBiasStats.DeepCopy(originalSBresults.StitchedStats),
};
return(sb);
}
public static void PopulateDiploidStats(StrandBiasStats stats, double noiseFreq, double minDetectableSNP)
{
//expectation we ought to see the 20% variant on this strand:
//save ourself some time here..
if (stats.Frequency >= minDetectableSNP)
{
stats.ChanceFalseNeg = 1; // TP if we called it
stats.ChanceFalsePos = 0; //FP if we called if
stats.ChanceVarFreqGreaterThanZero = 1;
return;
}
//trickier case, when we barely see it but we dont have enough reads...
var binomialHetAltExpected = new MathNet.Numerics.Distributions.Binomial(minDetectableSNP, (int)stats.Coverage);
//this is a real variant ( a false neg if we filtered it)
stats.ChanceFalseNeg = Math.Max(binomialHetAltExpected.CumulativeDistribution(stats.Support), 0); //if this was a het variant, would we ever see it this low?
//chance this is due to noise ( a false pos if we left it in)
stats.ChanceFalsePos = Math.Max(0.0, 1 - Poisson.Cdf(stats.Support, stats.Coverage * 0.1)); //chance this varaint is due to noise, we could see this much or more
stats.ChanceVarFreqGreaterThanZero = stats.ChanceFalseNeg;
}
public void TestPopulateDiploidStats()
{
double noiseFreq = 0.01;
double diploidThreshold = 0.20;
//Cases where the variant obviously exisits
StrandBiasStats stats = new StrandBiasStats(100, 100); //#observations, coverage
StrandBiasCalculator.PopulateDiploidStats(stats, noiseFreq, diploidThreshold);
Assert.Equal(stats.ChanceFalseNeg, 1, 3);
Assert.Equal(stats.ChanceFalsePos, 0, 3);
Assert.Equal(stats.ChanceVarFreqGreaterThanZero, 1, 3);
stats = new StrandBiasStats(50, 100); //#observations, coverage
StrandBiasCalculator.PopulateDiploidStats(stats, noiseFreq, diploidThreshold);
Assert.Equal(stats.ChanceFalseNeg, 1, 3);
Assert.Equal(stats.ChanceFalsePos, 0, 3);
Assert.Equal(stats.ChanceVarFreqGreaterThanZero, 1, 3);
stats = new StrandBiasStats(20, 100); //#observations, coverage
StrandBiasCalculator.PopulateDiploidStats(stats, noiseFreq, diploidThreshold);
Assert.Equal(stats.ChanceFalseNeg, 1, 3);
Assert.Equal(stats.ChanceFalsePos, 0, 3);
Assert.Equal(stats.ChanceVarFreqGreaterThanZero, 1, 3);
//
//Cases where the variant becomes less obvious
stats = new StrandBiasStats(15, 100); //#observations, coverage
StrandBiasCalculator.PopulateDiploidStats(stats, noiseFreq, diploidThreshold);
Assert.Equal(stats.ChanceFalseNeg, 0.129, 3); //Chance this is a real variant ( a false neg if we filtered it)//it could happen that this is still real
Assert.Equal(stats.ChanceFalsePos, 0.049, 3); //chance this is due to noise ( a false pos if we left it in). not very likely
Assert.Equal(stats.ChanceVarFreqGreaterThanZero, 0.129, 3);
stats = new StrandBiasStats(10, 100); //#observations, coverage
StrandBiasCalculator.PopulateDiploidStats(stats, noiseFreq, diploidThreshold);
Assert.Equal(stats.ChanceFalseNeg, 0.006, 3); //Chance this is a real variant ( a false neg if we filtered it)//it could happen that this is still real
Assert.Equal(stats.ChanceFalsePos, 0.417, 3); //chance this is due to noise ( a false pos if we left it in). not very likely
Assert.Equal(stats.ChanceVarFreqGreaterThanZero, 0.006, 3);
stats = new StrandBiasStats(1, 100); //#observations, coverage
StrandBiasCalculator.PopulateDiploidStats(stats, noiseFreq, diploidThreshold);
Assert.Equal(stats.ChanceFalseNeg, 0, 3); //Chance this is a real variant ( a false neg if we filtered it)//it could happen that this is still real
Assert.Equal(stats.ChanceFalsePos, 1, 3); //chance this is due to noise ( a false pos if we left it in). not very likely
Assert.Equal(stats.ChanceVarFreqGreaterThanZero, 0, 3);
//a few pathological cases
stats = new StrandBiasStats(0, 100); //#observations, coverage
StrandBiasCalculator.PopulateDiploidStats(stats, noiseFreq, diploidThreshold);
Assert.Equal(stats.ChanceFalseNeg, 0, 3);
Assert.Equal(stats.ChanceFalsePos, 1, 3);
Assert.Equal(stats.ChanceVarFreqGreaterThanZero, 0, 3);
stats = new StrandBiasStats(10, 0); //#observations, coverage
StrandBiasCalculator.PopulateDiploidStats(stats, noiseFreq, diploidThreshold);
Assert.Equal(stats.ChanceFalseNeg, 1, 3);
Assert.Equal(stats.ChanceFalsePos, 0, 3);
Assert.Equal(stats.ChanceVarFreqGreaterThanZero, 1, 3);
stats = new StrandBiasStats(0, 0); //#observations, coverage
StrandBiasCalculator.PopulateDiploidStats(stats, noiseFreq, diploidThreshold);
Assert.Equal(stats.ChanceFalseNeg, 1, 3); //not a meaningful answer, but at least nothing explodes.
Assert.Equal(stats.ChanceFalsePos, 0, 3);
Assert.Equal(stats.ChanceVarFreqGreaterThanZero, 1, 3);
stats = new StrandBiasStats(101, 100); //#observations, coverage
StrandBiasCalculator.PopulateDiploidStats(stats, noiseFreq, diploidThreshold);
Assert.Equal(stats.ChanceFalseNeg, 1, 3);
Assert.Equal(stats.ChanceFalsePos, 0, 3);
Assert.Equal(stats.ChanceVarFreqGreaterThanZero, 1, 3);
//check it reacts properly to depth. Ie, a 15% variant in N of 20 isnt a big deal,
//but a 15% varaint in N of 100000 seems rather low.
stats = new StrandBiasStats((20.0 * 0.15), 20); //#observations, coverage
StrandBiasCalculator.PopulateDiploidStats(stats, noiseFreq, diploidThreshold);
Assert.Equal(stats.ChanceFalseNeg, 0.411, 3); //note, the believability of this variant goes up from 0.129
Assert.Equal(stats.ChanceFalsePos, 0.143, 3); //but its also more possible to be noise. Basically, the whole picture is more murky
Assert.Equal(stats.ChanceVarFreqGreaterThanZero, 0.411, 3);
stats = new StrandBiasStats(15, 100); //#observations, coverage
StrandBiasCalculator.PopulateDiploidStats(stats, noiseFreq, diploidThreshold);
Assert.Equal(stats.ChanceFalseNeg, 0.129, 3);
Assert.Equal(stats.ChanceFalsePos, 0.049, 3);
Assert.Equal(stats.ChanceVarFreqGreaterThanZero, 0.129, 3);
//slightly more lilkey to be a variant than noise, but neither hypothesis fits.
stats = new StrandBiasStats((500.0 * 0.15), 500); //#observations, coverage
StrandBiasCalculator.PopulateDiploidStats(stats, noiseFreq, diploidThreshold);
Assert.Equal(stats.ChanceFalseNeg, 0.002, 3);
Assert.Equal(stats.ChanceFalsePos, 0, 3);
Assert.Equal(stats.ChanceVarFreqGreaterThanZero, 0.002, 3);
//it doesnt look like noise or a varaint. no hypothesis is reasonable.
stats = new StrandBiasStats((100000.0 * 0.15), 100000); //#observations, coverage
StrandBiasCalculator.PopulateDiploidStats(stats, noiseFreq, diploidThreshold);
Assert.Equal(stats.ChanceFalseNeg, 0, 3);
Assert.Equal(stats.ChanceFalsePos, 0, 3);
Assert.Equal(stats.ChanceVarFreqGreaterThanZero, 0, 3);
}
