public double GetGenotypeLogLikelihood(Balleles gtObservedCounts, PhasedGenotype gtModelCount)
{
double minLogLikelihood = Math.Log(1.0 / Double.MaxValue);
double currentLogLikelihood = 0;
foreach (var gtCount in gtObservedCounts.GetTruncatedAlleleCounts())
{
int rowId = Math.Min(gtCount.Item1, _maxCoverage - 1);
int colId = Math.Min(gtCount.Item2, _maxCoverage - 1);
int numHapsNonZero = (gtModelCount.CopyNumberA > 0 ? 1 : 0) + (gtModelCount.CopyNumberB > 0 ? 1 : 0);
double likelihoodThisLocus = 0;
// the observations can arise from a het locus, if both copy numbers are positive
if (numHapsNonZero == 2)
{
// Given a variant locus with two haplotypes, we have a roughly 2/3 chance of it being het.
// Alleles have 50:50 chance of being 'A' or 'B'.
// We ignore error terms, as they should have a negligible impact here.
likelihoodThisLocus += 1.0 / 3.0 *
(
_alleleDistribution[gtModelCount.CopyNumberA][gtModelCount.CopyNumberB].Item1[rowId] *
_alleleDistribution[gtModelCount.CopyNumberA][gtModelCount.CopyNumberB].Item2[colId]
+
_alleleDistribution[gtModelCount.CopyNumberA][gtModelCount.CopyNumberB].Item1[colId] *
_alleleDistribution[gtModelCount.CopyNumberA][gtModelCount.CopyNumberB].Item2[rowId]
);
}
// they can also arise from a hom locus in various ways
if (numHapsNonZero > 0)
{
// these should be constants to avoid calling Log over and over.
double logErrorProb = Math.Log(0.01);
double logNoErrorProb = Math.Log(.99);
// If both haplotypes have non-zero depth and the locus is non-ref, a locus has a prior prob of 1/3 of being hom,
// assuming a well-mixed population. We could adjust for observed het:hom, but we do not at this time.
// Of course, if only one haplotype has non-zero depth, it must be hom.
double priorFactorHom = numHapsNonZero == 2 ? 0.5 * (1.0 / 3.0) : 1.0;
// limit ttlReads to maxTotalDepth as that is all we have _readDepth probabilities for
int totalReads = Math.Min(rowId + colId, _maxAlleleCounts);
int totalCN = gtModelCount.CopyNumberA + gtModelCount.CopyNumberB;
// Split the likelihood into two parts:
// First, the probability of getting the observed total number of reads, given the total copy number
double probTotalReadDepth = _totalAlleleCountsDistribution[totalCN][totalReads];
// Second, the probability of the observed per-allele read counts assuming one of the alleles is an error.
// The calculation here is simply binomial, in log space
double logProbCountAErrors = LogCombinations(rowId, colId) + rowId * logErrorProb + colId * logNoErrorProb;
double logProbCountBErrors = LogCombinations(rowId, colId) + colId * logErrorProb + rowId * logNoErrorProb;
likelihoodThisLocus += priorFactorHom * probTotalReadDepth * (
Math.Exp(logProbCountAErrors) + Math.Exp(logProbCountBErrors));
}
else
{
// uses alleleStateZeroCorrector to enable non-zero likelihoods
int totalReads = Math.Min(rowId + colId, _maxAlleleCounts);
likelihoodThisLocus = _totalAlleleCountsDistribution[0][totalReads];
}
likelihoodThisLocus = Math.Max(minLogLikelihood, likelihoodThisLocus);
currentLogLikelihood += Math.Log(likelihoodThisLocus);
}
return(currentLogLikelihood);
}
