I think I brought this up the other month but I really need some help on fixing this. Egon and I noticed some test regressions in a bond qsar a few months ago. The regressions were due to changing atom order (that is the result was different if the atom order was changed). I’ve managed to track down the problem which resolved the recent regressions but it now means several other tests are failing. I want to verify the expected results are correct but unfortunately the NIST data isn’t linked and so all I have "This class returns the ionization potential of a Bond. It is based on a function which is extracted from Weka(J48) from experimental values (NIST data).” - IPBondLearningDescriptor
. Does anyone know what NIST data set this could be?
Also reading that description might indicate that the bug was present when the machine learning was run (i.e. it will always be wrong) and it should be re-run.
Quick summary of the problem (now fixed for this example):
IPBondLearningDescriptor descriptor = new IPBondLearningDescriptor();
IAtomContainer mol = Smiles.parseString(“C=C(C=CC)C");
// configure molecule
descriptor.calculate(mol.getBond(3), mol); // result=8.85
descriptor.calculate(mol.getBond(3), mol); // result=9.04