// ----------------- private methods -----------------
// make sure that for a single molecule, the way that the hashes are created & folded is consistent with a reference
private static void CheckFP(string molstr, CircularFingerprinterClass classType, int folding, int[] refHash)
{
var strType = classType == CircularFingerprinterClass.ECFP6 ? "ECFP6" : "FCFP6";
WriteLine($"Comparing hash codes for {strType}/folding={folding}");
var mol = new EnumerableSDFReader(new StringReader(molstr), ChemObjectBuilder.Instance).First();
var model = new Bayesian(classType, folding);
model.AddMolecule(mol, false);
var calcHash = model.Training[0];
var same = calcHash.Length == refHash.Length;
if (same)
{
for (int n = 0; n < calcHash.Length; n++)
{
if (calcHash[n] != refHash[n])
{
same = false;
break;
}
}
}
if (!same)
{
WriteLine($" ** calculated: {ArrayStr(calcHash)}");
WriteLine($" ** reference: {ArrayStr(refHash)}");
throw new CDKException("Hashes differ.");
}
}
// compares a series of molecules for folding fingerprints being literally identical
private static void CompareFolding(string sdfile, string fpField, CircularFingerprinterClass classType, int folding)
{
WriteLine($"[{sdfile}] calculation of: {fpField}");
using (var ins = ResourceLoader.GetAsStream($"NCDK.Data.CDD.{sdfile}"))
{
var rdr = new EnumerableSDFReader(ins, ChemObjectBuilder.Instance);
int row = 0;
foreach (var mol in rdr)
{
row++;
var model = new Bayesian(classType, folding);
model.AddMolecule(mol, false);
var hashes = model.Training[0];
var gotHashes = ArrayStr(hashes);
var reqHashes = (string)mol.GetProperties()[fpField];
Assert.AreEqual(reqHashes, gotHashes, $"Folded hashes do not match reference at {row}.");
}
}
}
private void RunTest(string sdfile, string actvField, CircularFingerprinterClass classType, int folding, int xval, string modelFN, bool perceiveStereo)
{
WriteLine("[" + modelFN + "]");
WriteLine(" Loading " + sdfile);
try
{
var model = new Bayesian(classType, folding)
{
PerceiveStereo = perceiveStereo
};
int row = 0, numActives = 0;
using (var rdr = new EnumerableSDFReader(ResourceLoader.GetAsStream($"NCDK.Data.CDD.{sdfile}"), ChemObjectBuilder.Instance))
{
foreach (var mol in rdr)
{
row++;
var stractv = (string)mol.GetProperties()[actvField];
int active = stractv.Equals("true", StringComparison.Ordinal) ? 1 : stractv.Equals("false", StringComparison.Ordinal) ? 0 : int.Parse(stractv, NumberFormatInfo.InvariantInfo);
if (active != 0 && active != 1)
{
throw new CDKException("Activity field not found or invalid");
}
model.AddMolecule(mol, active == 1);
numActives += active;
}
}
WriteLine($" Training with {row} rows, {numActives} actives, {(row - numActives)} inactives");
model.Build();
if (xval == 3)
{
model.ValidateThreeFold();
}
else if (xval == 5)
{
model.ValidateFiveFold();
}
else
{
model.ValidateLeaveOneOut();
}
WriteLine($" Validation: ROC AUC={model.RocAUC}");
WriteLine($" Parsing reference model");
//FileReader frdr=new FileReader(modelFN);
Bayesian reference;
using (var mrdr = new StreamReader(ResourceLoader.GetAsStream($"NCDK.Data.CDD.{modelFN}")))
{
reference = Bayesian.Deserialise(mrdr);
}
// start comparing the details...
bool failed = false;
if (model.Folding != reference.Folding)
{
WriteLine($" ** reference folding size={reference.Folding}");
failed = true;
}
if (model.TrainingSize != reference.TrainingSize)
{
WriteLine($" ** reference training size={reference.TrainingSize}");
failed = true;
}
if (model.TrainingActives != reference.TrainingActives)
{
WriteLine($" ** reference training actives={reference.TrainingActives}");
failed = true;
}
if (model.RocType != reference.RocType)
{
WriteLine($" ** reference ROC type={reference.RocType}");
failed = true;
}
if (!DblEqual(model.RocAUC, reference.RocAUC))
{
WriteLine($" ** reference ROC AUC={reference.RocAUC}");
failed = true;
}
if (Math.Abs(model.LowThreshold - reference.LowThreshold) > 0.00000000000001)
{
WriteLine($" ** reference lowThresh={reference.LowThreshold} different to calculated {model.LowThreshold}");
failed = true;
}
if (Math.Abs(model.HighThreshold - reference.HighThreshold) > 0.00000000000001)
{
WriteLine($" ** reference highThresh={reference.HighThreshold} different to calculated {model.HighThreshold}");
failed = true;
}
// make sure individual hash bit contributions match
var mbits = model.Contributions;
var rbits = reference.Contributions;
if (mbits.Count != rbits.Count)
{
WriteLine($" ** model has {mbits.Count} contribution bits, reference has {rbits.Count}");
failed = true;
}
foreach (var h in mbits.Keys)
{
if (!rbits.ContainsKey(h))
{
WriteLine($" ** model hash bit {h} not found in reference");
failed = true;
break; // one is enough
}
}
foreach (var h in rbits.Keys)
{
if (!mbits.ContainsKey(h))
{
WriteLine($" ** reference hash bit {h} not found in model");
failed = true;
break; // one is enough
}
}
foreach (var h in mbits.Keys)
{
if (rbits.ContainsKey(h))
{
double c1 = mbits[h], c2 = rbits[h];
if (!DblEqual(c1, c2))
{
WriteLine($" ** contribution for bit {h}: model={c1}, reference={c2}");
failed = true;
break; // one is enough
}
}
}
if (failed)
{
throw new CDKException("Comparison to reference failed");
}
}
catch (CDKException)
{
throw;
}
catch (Exception ex)
{
throw new CDKException("Test failed", ex);
}
}
// builds a model and uses the scaled predictions to rack up a confusion matrix, for comparison
private static void ConfirmPredictions(string sdfile, int truePos, int trueNeg, int falsePos, int falseNeg)
{
WriteLine($"[{sdfile}] comparing confusion matrix");
var molecules = new List <IAtomContainer>();
var activities = new List <bool>();
var model = new Bayesian(CircularFingerprinterClass.ECFP6, 1024);
try
{
using (var ins = ResourceLoader.GetAsStream("NCDK.Data.CDD." + sdfile))
{
var rdr = new EnumerableSDFReader(ins, ChemObjectBuilder.Instance);
foreach (var mol in rdr)
{
bool actv = "true" == (string)mol.GetProperties()["Active"];
molecules.Add(mol);
activities.Add(actv);
model.AddMolecule(mol, actv);
}
}
}
catch (CDKException)
{
throw;
}
catch (Exception ex)
{
throw new CDKException("Test failed", ex);
}
model.Build();
model.ValidateLeaveOneOut();
// build the confusion matrix
int gotTP = 0, gotTN = 0, gotFP = 0, gotFN = 0;
for (int n = 0; n < molecules.Count; n++)
{
double pred = model.ScalePredictor(model.Predict(molecules[n]));
bool actv = activities[n];
if (pred >= 0.5)
{
if (actv)
{
gotTP++;
}
else
{
gotFP++;
}
}
else
{
if (actv)
{
gotFN++;
}
else
{
gotTN++;
}
}
}
WriteLine(" True Positives: got=" + gotTP + " require=" + truePos);
WriteLine(" Negatives: got=" + gotTN + " require=" + trueNeg);
WriteLine(" False Positives: got=" + gotFP + " require=" + falsePos);
WriteLine(" Negatives: got=" + gotFN + " require=" + falseNeg);
if (gotTP != truePos || gotTN != trueNeg || gotFP != falsePos || gotFN != falseNeg)
{
throw new CDKException("Confusion matrix mismatch");
}
}