Dear Greg,
I suspect that it's a precision error or eigen algorithm shift between rdkit
c++ & dragon.
To obtain good value, I suggest to try to implement a test on the eigen values
like i did in gateway.cpp implementation.
JacobiSVD<MatrixXd> getSVD(MatrixXd A) {
JacobiSVD<MatrixXd> mysvd(A, ComputeThinU | ComputeThinV);
return mysvd;
}
// get the A-1 matrix using
MatrixXd GetPinv(MatrixXd A){
JacobiSVD<MatrixXd> svd = getSVD(A);
double pinvtoler=1.e-2; // choose your tolerance wisely!
VectorXd vs=svd.singularValues();
VectorXd vsinv=svd.singularValues();
for (unsigned int i=0; i<A.cols(); ++i) {
if ( vs(i) > pinvtoler )
vsinv(i)=1.0/vs(i);
else vsinv(i)=0.0;
}
MatrixXd S = vsinv.asDiagonal();
MatrixXd Ap = svd.matrixV() * S * svd.matrixU().transpose();
return Ap;
}
If it's not solve the problem, I would like to test it in Matlab. can you
provide me the 3 (3d xyz matrix) of your example please ?
I also have Dragon 6
best regards,
Dr. Guillaume GODIN
Principal Scientist
Chemoinformatic & Datamining
Innovation
CORPORATE R&D DIVISION
DIRECT LINE +41 (0)22 780 3645
MOBILE +41 (0)79 536 1039
Firmenich SA
RUE DES JEUNES 1 | CASE POSTALE 239 | CH-1211 GENEVE 8
________________________________
De : Greg Landrum <greg.land...@gmail.com>
Envoyé : dimanche 15 janvier 2017 11:50
À : Chris Earnshaw; RDKit Discuss
Objet : Re: [Rdkit-discuss] PMI API
I managed to make some time to look into this this weekend and I've found a bug
and something I don't understand. Hopefully the community can help out here.
On Sun, Jan 8, 2017 at 11:17 AM, Chris Earnshaw
<cgearns...@gmail.com<mailto:cgearns...@gmail.com>> wrote:
4) The big one! The returned results look very odd. They appear to relate more
to the dimensions of the molecule than the moments of inertia. For a rod-like
molecule (dimethylacetylene) I'd expect two large and one small PMI (e.g. PMI1:
6.61651 PMI2: 150.434 PMI3: 150.434 NPR1: 0.0439828 NPR2: 0.999998) but
actually get PMI1: 0.061647 PMI2: 0.061652 PMI3: 25.3699 NPR1: 0.002430
NPR2: 0.002430.
For disk-like (benzene) the result should be one large and two medium (e.g.
PMI1: 89.1448 PMI2: 89.1495 PMI3: 178.294 NPR1: 0.499987 NPR2: 0.500013)
but get PMI1: 2.37457e-10 PMI2: 11.0844 PMI3: 11.0851 NPR1: 2.14213e-11
NPR2: 0.999933.
Finally for a roughly spherical molecule (neopentane) the NPR values look
reasonable (no great surprise) but the absolute PMI values may be too small:
old program - PMI1: 114.795 PMI2: 114.797 PMI3: 114.799
NPR1: 0.999966 NPR2: 0.999988, new program - PMI1: 6.59466 PMI2: 6.59488
PMI3: 6.59531 NPR1: 0.999902 NPR2: 0.999935
Your expectations are correct: the current RDKit implementation is wrong. The
corresponding github entry is here: https://github.com/rdkit/rdkit/issues/1262
This is due to a mistake in the way the principal moments are calculated (which
is due to the fact that I don't spend a lot of time working with/thinking about
3D descriptors). Instead of using the eigenvectors/eigenvalues of the inertia
matrix (the tensor of inertia) the RDKit is currently using the covariance
matrix. There's some more on the relationship between these two here:
http://number-none.com/blow/inertia/deriving_i.html
The problem is easy to fix (and I have something working here:
https://github.com/greglandrum/rdkit/tree/fix/github1262), but it screws up the
values of the descriptors that are derived from here:
Todeschini and Consoni "Descriptors from Molecular Geometry" Handbook of
Chemoinformatics http://dx.doi.org/10.1002/9783527618279.ch37
These include the radius of gyration, inertial shape factor, etc.
Within that article they state that Ic = 0 for planar molecules. Ignoring the
inequality on page 1010, which says that Ic is the largest moment and is
contradicted by the rest of the text (particularly the inequalities on page
1011), Ic corresponds to the smallest principal moment : PMI1.
So now I'm confused, but I'm hoping this is obvious to someone versed in the
field: I'd like to reproduce the descriptors described in the Todeschini
article, but I clearly can't do that using the actual moments of inertia. I
could keep using the eigenvalues of the covariance matrix there, but that
doesn't match what's described in the text.
Two things that would be extremely helpful:
1) an explanation of the disconnect here from someone who knows this stuff, I
would guess that it's pretty simple
2) The results of running the files github1262_1.mol, github1262_2.mol, and
github1262_3.mol from here:
https://github.com/greglandrum/rdkit/tree/fix/github1262/Code/GraphMol/MolTransforms/test_data
through Dragon and calculating the radius of gyration, inertial shape factor,
eccentricity, molecular asphericity, and spherocity index.
Best,
-greg
**********************************************************************
DISCLAIMER
This email and any files transmitted with it, including replies and forwarded
copies (which may contain alterations) subsequently transmitted from Firmenich,
are confidential and solely for the use of the intended recipient. The contents
do not represent the opinion of Firmenich except to the extent that it relates
to their official business.
**********************************************************************------------------------------------------------------------------------------
Developer Access Program for Intel Xeon Phi Processors
Access to Intel Xeon Phi processor-based developer platforms.
With one year of Intel Parallel Studio XE.
Training and support from Colfax.
Order your platform today. http://sdm.link/xeonphi
_______________________________________________
Rdkit-discuss mailing list
Rdkit-discuss@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/rdkit-discuss
