public static void SearchHlaAssignmentsRelaxation(string fileName, string header, string solutionFileName, string solutionHeader, double causePrior, double linkProbability, double leakProbability, int howManyBest, string outputFilename, HlaResolution hlaResolution)
{
Debug.Assert(Math.Exp(double.NegativeInfinity) == 0);
//Create the structure of patients and their HLAs
Quickscore <Hla, int> quickscore = HlaAssignmentParams.CreateQuickscore(fileName, header, causePrior, linkProbability, leakProbability, hlaResolution);
// Get the list of patients, Hlas, and patient weights
List <int> patientList = quickscore.EffectList();
IEnumerable <Hla> hlaList = quickscore.CauseList();
//The structure file can contain a "weight" column. If it does, weight the patients, otherwise give them all weight 1.0
Dictionary <int, double> patientWeightTable = HlaAssignmentParams.CreatePatientWeightTable(fileName, header);
//The cause assignment table is a mapping from
// peptide to HLA assignments for that peptide
Dictionary <string, List <Dictionary <string, string> > > causeAssignmentTable = CreateCauseAssignmentTable(solutionFileName, solutionHeader);
//We start the report
using (StreamWriter streamwriterOutputFile = File.CreateText(outputFilename))
{
//The report is the same as the input "solution" except that we now list new HLA, a note, and a p(assignment)
streamwriterOutputFile.WriteLine(SpecialFunctions.CreateTabString(solutionHeader,
"newHLA", "rank", "p(assignment)", "note", "Peptide", "LogLikelihood", "TrueHlas.Count", "TrueHlas", "TrueHlasAndRespondingPatients", "UnexplainedPatients.Count"));
//Now we go through the model file, one peptide at a time. For each peptide, we know the patientID (did) of
// every patient who responded. (It also tells the HLAs of the patient, but we already have that info in the structure)
/*
* peptide did a1 a2 b1 b2 c1 c2
*
* ACQGVGGPGHK 10 2 11 38 44 12 16
* ACQGVGGPGHK 14 2 24 1517 58 3 7
* ACQGVGGPGHK 41 11 33 35 40 6 7
* ACQGVGGPGHK 102 2 11 35 44 4 5
*
* AENLWVTVY 25 24 66 35 39 4 12
* AENLWVTVY 36 23 32 8 44 7 7
* AENLWVTVY 45 3 31 39 44 5 12
* AENLWVTVY 46 2 29 41 52 16 16
* AENLWVTVY 59 2 33 7 44 3 15
*
* [...]
*/
foreach (List <Dictionary <string, string> > tableByPeptide in HlaAssignmentParams.QuickScoreOptimalsGroupByPeptide(fileName, header))
{
Debug.Assert(tableByPeptide.Count > 0); // real assert
string peptide = tableByPeptide[0]["peptide"];
//Create lists of patients who responded and assume everyone else didn't.
List <int> patientsWhoRespond;
List <int> patientsWhoDoNotRespond;
FindPatientsWhoRespondAndWhoDoNot(patientList, tableByPeptide, out patientsWhoRespond, out patientsWhoDoNotRespond);
//If we get a peptide with no row in the solution file, skip it.
if (!causeAssignmentTable.ContainsKey(peptide))
{
streamwriterOutputFile.WriteLine("{0}\t\t{1}", peptide, "explained by noise");
continue;
}
//We assign every HLA mentioned by the solution file with this peptide to TRUE and all others to FALSE
//An assignment of TRUE means that this HLA is a cause of this response.
List <Dictionary <string, string> > rowsOfThisPeptide = causeAssignmentTable[peptide];
HlaAssignmentWithResponses hlaAssignmentBase = CreateHlaAssignment(rowsOfThisPeptide, hlaList, hlaResolution, quickscore, patientsWhoRespond);
//Find the likelihood of this structure, with these patient responses, and the solution's HLA assignment.
double baseLogLikelihood = quickscore.LogLikelihoodOfModelWithCompleteAssignments(patientsWhoRespond, patientsWhoDoNotRespond, hlaAssignmentBase.AsDictionary, patientWeightTable);
//Consider each HLA assigned to TRUE (for this peptide) by the solution
foreach (Dictionary <string, string> row in rowsOfThisPeptide)
{
Hla hla = GetHlaFromRow(row, hlaResolution);
//If we already really know that this HLA is a cause of this reponse, just report that.
if (IsKnown(row))
{
streamwriterOutputFile.WriteLine("{0}\t{1}\t{2}", row[""], hla, "known");
continue;
}
//Set it to false as a way to see measure the probability that it is true.
HlaAssignmentWithResponses hlaAssignmentL0 = CreateHlaAssignment(rowsOfThisPeptide, hlaList, hlaResolution, quickscore, patientsWhoRespond, hla, null);
double logL0 = quickscore.LogLikelihoodOfModelWithCompleteAssignments(patientsWhoRespond, patientsWhoDoNotRespond, hlaAssignmentL0.AsDictionary, patientWeightTable);
Debug.Assert(hlaAssignmentL0.TrueCount + 1 == hlaAssignmentBase.TrueCount); // real assert
string noteNote = "";
if (logL0 > baseLogLikelihood)
{
noteNote = string.Format("\tlogL0 > baseLogLikelihood ({0}>{1})", logL0, baseLogLikelihood);
}
double probability = Math.Exp(baseLogLikelihood - SpecialFunctions.LogSum(baseLogLikelihood, logL0));
streamwriterOutputFile.WriteLine(SpecialFunctions.CreateTabString(
row[""], hla, "1 best", probability, noteNote,
peptide, baseLogLikelihood, hlaAssignmentBase.TrueCount, hlaAssignmentBase.TrueToString(), hlaAssignmentBase.TrueToListString(), hlaAssignmentBase.UnexplainedPatients.Count));
streamwriterOutputFile.WriteLine(SpecialFunctions.CreateTabString(
row[""], hla, "remove 1 best", "", noteNote,
peptide, logL0, hlaAssignmentL0.TrueCount, hlaAssignmentL0.TrueToString(), hlaAssignmentL0.TrueToListString(), hlaAssignmentL0.UnexplainedPatients.Count));
//Also, while we're in the world were it is false, let's measure how well each of the currently-set-to-false HLAs would do instead.
//!!!should this be one dictionary to a new class instead of three dictionaries?
List <KeyValuePair <Hla, double> > listOfhla1AndProbability1 = new List <KeyValuePair <Hla, double> >();
Dictionary <Hla, double> loglikelihoodCollection = new Dictionary <Hla, double>();
Dictionary <Hla, HlaAssignmentWithResponses> hla1ToHlaAssignment = new Dictionary <Hla, HlaAssignmentWithResponses>();
foreach (Hla hla1 in hlaList)
{
if (hla1 == hla || hlaAssignmentBase.AsDictionary[hla1])
{
continue;
}
HlaAssignmentWithResponses hlaAssignmentL1 = CreateHlaAssignment(rowsOfThisPeptide, hlaList, hlaResolution, quickscore, patientsWhoRespond, hla, hla1);
if (hlaAssignmentL1.TrueCount == hlaAssignmentBase.TrueCount)
{
double logL1 = quickscore.LogLikelihoodOfModelWithCompleteAssignments(patientsWhoRespond, patientsWhoDoNotRespond, hlaAssignmentL1.AsDictionary, patientWeightTable);
double probability1 = Math.Exp(logL1 - SpecialFunctions.LogSum(logL1, logL0));
loglikelihoodCollection[hla1] = logL1;
hla1ToHlaAssignment.Add(hla1, hlaAssignmentL1);
listOfhla1AndProbability1.Add(new KeyValuePair <Hla, double>(hla1, probability1));
}
}
//Report on the best of these
listOfhla1AndProbability1.Sort(delegate(KeyValuePair <Hla, double> x, KeyValuePair <Hla, double> y) { return(y.Value.CompareTo(x.Value)); });
for (int i = 0; i < howManyBest - 1; ++i)
{
Hla hla1 = listOfhla1AndProbability1[i].Key;
streamwriterOutputFile.WriteLine(SpecialFunctions.CreateTabString(
row[""], hla1, string.Format("{0} best", i + 2), listOfhla1AndProbability1[i].Value,
"", peptide, loglikelihoodCollection[hla1], hla1ToHlaAssignment[hla1].TrueCount, hla1ToHlaAssignment[hla1].TrueToString(), hla1ToHlaAssignment[hla1].TrueToListString(), hla1ToHlaAssignment[hla1].UnexplainedPatients.Count));
}
}
}
}
}
