–Egon Willighagen, Lunteren 2005
Representation of chemical data in
QSAR and crystallography

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
Computer representation of molecular
structures
•Spectra (NMR, IR, ...) •Connection Table
•Schrödinger •Dietz Representation
Equation
•Molecular Invariants •Physical Properties

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
Computer representations of molecular
structures
•Spectra (NMR, IR, ...) •Connection Table
•Schrödinger •Dietz Representation
Equation
•Molecular Invariants •Physical Properties

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
Representing relations between descriptors
•
• Descriptor ontology: explicit definition of descriptor
types and descriptor properties
•
•
•
•
•
•C.Steinbeck, C.Hoppe, S.Kuhn, M.Floris, R.Guha, E.L.Willighagen, Recent Developments of the Chemistry
Development Kit (CDK) - An Open-Source Java Library for Chemo- and Bioinformatics, Current
Pharmaceutical Design, accepted

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
Representation does make a difference
• Use of NMR spectra in Quantitative Structure Activity
Relationship (QSAR) modeling
• Three representations: simulated 1H NMR, 13C NMR
spectra and theoretical descriptors
• Three data sets:
– water solubility of 431 compounds (WS)
– boiling points of 277 compounds (BP)
– LogP values of 154 compounds (LogP)
–
•E.L.Willighagen, H.M.G.W.Denissen, R.Wehrens, L.M.C.Buydens, On the use of 1H and 13C NMR
spectra as QSAR descriptors, submitted

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
How the experiment is performed...
• Partial Least Squares
– 220 NMR bins
– 220 randomly selected theoretical descriptors (Dragon)
–
– five random divisions in training and test sets
–
– number of latent variables chosen with leave-one-out cross
validation

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
Number of latent variables

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
1
H and 13C NMR versus Dragon Descriptors

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
Prediction Errors

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
Model interpretation?

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
What can we conclude?
• Representation has a large effect
• H NMR models have no predictive power
1
• 13
C NMR models have some predictive power
... but no advantages

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
Finding a representation for crystal structures

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
Electronic Radial Distribution Function (ReDF)
•
• Describes patterns in atom interactions in and around the
unit cell
• Å
•E.L.Willighagen, R.Wehrens, P.Verwer, R.de Gelder, L.M.C.Buydens, A Method for the
Computational Comparison of Crystal Structures, Acta.Cryst., 2005, B61, 29-36

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
Quantifying Similarities

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
Quantifying Similarities
ReDF 1
 Weighted
Cross
 Similarity
[0,1]
Correlation
ReDF 2

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
Cephalosporin crystal structures

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
Polymorph Prediction
• Polymorphs: different
crystal structures for the
same molecular
compound
• Polymorph Prediction:
computational method to
predict the polymorphs
given a molecular
structure

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
Polymorphic estrone crystal structures
• Better trend in similarity going from identical to different
structures :
ReDF + WCC Cerius2

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
Conclusions
• It is important to pick a proper representation
• H and 13C NMR spectra are not good representations for
1
QSAR models
• A new crystal structure descriptor for gives chemically
better interpretable similarities

Egon Willighagen, Lunteren 2005
Representation of chemical data
in QSAR and crystallography
Acknowledgments
● Ron Wehrens, Lutgarde Buydens (supervisors)
● René de Gelder, Paul Verwer (crystal structures)
● Harm Denissen (QSAR)
● Peter Murray-Rust (Cambridge University, UK)
Christoph Steinbeck (Cologne University, DE)
● NWO (for financial support)