3D-QSAR of PET Agents For Imaging - Amyloid in Alzheimer's Disease
3D-QSAR of PET Agents For Imaging - Amyloid in Alzheimer's Disease
3D-QSAR of PET Agents For Imaging - Amyloid in Alzheimer's Disease
7 1231
Department of Bioscience and Biotechnology, Konkuk University, Seoul 143-701, Korea. *E-mail: chongy@konkuk.ac.kr
†
Department of Neuropsychiatry, Seoul National University College of Medicine, Seoul 110-744, Korea
Received May 19, 2007
Figure 1. ThioT and stilbene derivatives used for 3D-QSAR study. Fragments used for structural alignment are represented in bold lines.
1232 Bull. Korean Chem. Soc. 2007, Vol. 28, No. 7 Notes
Figure 2. Structural alignment of ThioT and stilbene derivatives around the common substructure PhCH=CH.
QSAR model. Finally, the contour plots of CoMFA and lated, the optimum number of components was then given,
CoMSIA were analyzed to provide helpful information on and CoMFA model was finally derived corresponding to the
how to improve the binding affinity of ThioT derivatives by optimum number. The column filtering box was kept
structural modifications. unchecked during all operations (Table 1).
All calculations were carried out on a linux enterprise The basic principle of CoMSIA is the same as that of
operating system using molecular modeling software pack- CoMFA, but CoMSIA includes some additional descriptors
age SYBYL v 7.2. All compounds were constructed by the such as hydrophobicity, hydrogen bond donor and hydrogen
Sketch module in SYBYL base and assigned with MMFF94s bond acceptor. As in the CoMFA model, the same regression
charges. For more flexible compounds, systematic searches analysis was used for CoMSIA field energies, and the results
were performed with an interval of 10o on every rotatable obtained from the PLS analysis are also summarized in
bond to ensure their lowest energy conformations. Finally, Table 1.
they were minimized with MMFF94s force field. The most The CoMFA model with 63 molecules in the training set is
crucial step in performing 3D-QSAR is to determine the consequently a clearly statistically significant model show-
bioactive conformations of the compounds so that all ing a cross-validated q2-value of 0.631 at 5 components
compounds could be aligned together. In this study, the (Table 1). This analysis was used for the final non-cross-
styrene moiety [PhC=X (X = C or N)] commonly found in validated run, giving a good correlation coefficient (r2 value
the ThioT and stilbene derivatives was used as the sub- of 0.926) (Table 1). Using steric, electrostatic, hydrophobic,
structure for alignment. The compound PIB (54) was used as and hydrogen bond donor and acceptor properties as
a template for structural alignment from the alignment descriptors, CoMSIA analysis was performed and the results
facility in SYBYL, and 63 training set molecules and 10 test are also listed in Table 1. The best q2 was found using all five
set molecules were all aligned together (Fig. 2). different descriptor variables, which demonstrates that these
With the structure-based ligand alignment in hand, we variables are necessary to describe the interaction mode of
attempted 3D-QSAR study by using CoMFA7,8 as well as the PET ligands with Aβ -plaques. The CoMSIA model with
CoMSIA9,10 method with grid spacing of 2.0 Å. As usual, a cross-validated q2 of 0.654 for 4 components and a
PLS (partial least squares) method was used to establish and conventional r2 of 0.900 was obtained (Table 1). These data
validate 3D-QSAR. The Ki values were converted into pKi demonstrate that the CoMSIA model is stable and statisti-
(−logKi) to describe the binding affinities. CoMFA was set at cally significant and it should be noted that, due to the use of
standard values, with a sp3 carbon atom with one positive less components than CoMFA model, the CoMSIA model
charge used to probe steric and electrostatic fields. The has reduced risk of overtraining the system. The correspond-
standard cutoff value was set to 30 kcal/mol. LOO (leave- ing field distributions of these five descriptor variables were
one-out) cross-validation method was used to evaluate the 20.4, 5.1, 25.9, 14.9, and 33.7%, respectively, which indi-
initial model. The cross-validated coefficient q2 was calcu- cates that steric rather than electrostatic and H-bond acceptor
Notes Bull. Korean Chem. Soc. 2007, Vol. 28, No. 7 1233
Table 2. Predicted binding affinities (PBA) versus experimental binding affinities (EBA, pKi) and residuals (d) by CoMFA and CoMSIA
CoMFA CoMSIA CoMFA CoMSIA CoMFA CoMSIA
EBA EBA EBA
PBA d PBA d PBA d PBA d PBA d PBA d
1 7.82 7.63 0.19 7.28 0.54 26 9.17 8.94 0.23 9.06 0.12 51 8.31 8.57 −0.26 8.19 0.12
2 7.57 7.19 0.38 7.33 0.24 27 9.19 9.15 0.04 9.11 0.08 52 8.72 8.86 −0.14 8.49 0.23
3 7.40 7.36 0.04 7.27 0.13 28* 8.37 8.75 −0.38 8.74 −0.37 53* 7.34 7.80 −0.46 7.61 −0.27
4* 6.62 7.43 −0.81 6.98 −0.37 29* 9.55 8.89 0.66 9.01 0.54 54 8.37 8.22 0.15 7.96 0.40
5 6.20 6.30 −0.10 6.45 −0.26 30 8.97 9.21 −0.23 9.17 −0.20 55 8.36 8.58 −0.22 8.26 0.10
6 6.47 6.58 −0.11 6.41 0.06 31 8.81 9.05 −0.25 9.13 −0.32 56 7.19 7.29 −0.10 7.37 −0.18
7 5.70 5.66 0.04 5.93 −0.23 32 9.14 9.21 −0.07 9.16 −0.02 57 8.07 7.68 0.39 7.72 0.34
8 5.70 5.75 −0.05 5.90 −0.20 33* 9.10 9.06 0.04 8.98 0.12 58 7.96 8.02 −0.06 8.02 −0.06
9 6.75 6.93 −0.18 6.93 −0.17 34 8.64 8.86 −0.22 8.92 −0.28 59 8.14 7.82 0.32 8.19 −0.05
10 7.32 7.46 −0.15 7.13 0.18 35 8.89 8.67 0.22 8.37 0.52 60* 8.77 8.21 0.56 8.54 0.23
11 8.64 8.93 −0.29 8.63 0.01 36 8.36 8.41 −0.05 8.32 0.04 61 8.54 8.58 −0.04 8.84 −0.31
12 7.94 8.16 −0.23 8.45 −0.51 37 8.17 7.98 0.19 8.21 −0.04 62 9.05 8.70 0.34 8.85 0.20
13* 8.89 8.92 −0.04 8.81 0.08 38* 8.08 8.67 −0.59 8.34 −0.27 63 6.82 6.94 −0.12 6.71 0.12
14 8.55 8.78 −0.22 8.66 −0.11 39 8.47 8.26 0.21 8.20 0.27 64 8.92 8.64 0.28 8.59 0.33
15 8.80 8.93 −0.14 8.89 −0.10 40 8.42 8.15 0.27 8.04 0.38 65* 8.22 8.61 −0.39 8.27 −0.04
16 8.57 8.65 −0.08 8.88 −0.31 41 7.91 8.33 −0.43 8.34 −0.43 66 8.89 8.90 −0.01 8.69 0.19
17 8.96 8.54 0.42 8.94 0.02 42 8.72 8.81 −0.09 8.69 0.04 67 8.66 8.91 −0.25 8.45 0.20
18 9.22 9.37 −0.15 9.18 0.04 43 8.80 8.61 0.18 8.66 0.14 68 8.60 8.38 0.22 8.54 0.06
19 9.40 9.37 0.03 9.37 0.03 44 8.27 8.47 −0.20 8.82 −0.55 69 8.00 8.26 −0.26 8.27 −0.27
20 8.89 8.51 0.38 8.66 0.23 45 7.43 7.69 −0.26 7.86 −0.43 70 7.04 7.61 −0.57 7.50 −0.45
21 8.38 8.41 −0.04 8.71 −0.34 46 7.96 8.01 −0.05 8.21 −0.25 71 8.66 8.16 0.50 8.29 0.37
22 8.05 8.17 −0.12 8.32 −0.28 47 8.40 8.35 0.04 8.50 −0.11 72 8.30 8.24 0.07 8.15 0.15
23 8.19 8.07 0.12 8.38 −0.19 48 8.02 7.98 0.04 7.98 0.04 73* 8.17 8.12 0.05 8.17 0.00
24 9.40 8.95 0.44 8.93 0.47 49 8.00 8.24 −0.24 8.31 −0.31
25 8.52 8.60 −0.08 8.56 −0.05 50 8.15 8.12 0.04 7.83 0.33
*
test set compounds
rather than H-bond donor contribute more to the final 60, 65, 73) that were not included in the construction of
CoMSIA model, and thus, these fields play crucial roles in CoMFA and CoMSIA models are selected as the test set.
determining the binding affinity of the PET ligands to the The binding affinities of the test set molecules were predict-
target protein. Predicted binding affinities, given as pKi ed reasonably well (residuals from 0.04 to 0.81 for CoMFA
values, and the residuals of the final non cross-validated model and 0.04 to 0.54 for CoMSIAall model) and the results
CoMFA as well as CoMSIA models are shown in Table 2, were also summarized in Table 2 and Figure 3. The predic-
and the plot of actual pKi values versus predicted pKi is tive performance of models on the test set was estimated by
shown in Figure 3. predictive r2 value (r2pred). The predictive performances of
The ultimate test for the usefulness of a 3D-QSAR model CoMFA and CoMSIAall model on the test were r2pred = 0.74
is predicting the activity of new compounds that are not and r2pred = 0.93, respectively, which indicated that the
included in the dataset that is used to obtain the model. To CoMSIA model was more reliable and able to predict bio-
validate the stability and predictive ability of our 3D-QSAR logical activity of new derivatives more accurately.
model, 10 PET ligands (compounds 4, 13, 28, 29, 33, 38, 53, Graphical representations of CoMSIA maps obtained by
Figure 3. Plot of observed pKi versus conventional fit predictions (predicted activity) of training set (a: CoMFA model, b: CoMSIA model)
and test set (c: CoMFA model, d: CoMSIA model).
1234 Bull. Korean Chem. Soc. 2007, Vol. 28, No. 7 Notes
Figure 4. CoMSIA Contour plots: PIB (54) is depicted as a reference molecule: (a) CoMSIAse: green and blue contours predict that the
binding affinity is enhanced by bulky substituent and positive charge, respectively, whereas yellow contours predict less bulky group is
favored for activity; (b) CoMSIAda: orange contours predict H-bond acceptors enhance activity and cyan contours predict H-bond donor
enhance activity; (c) CoMSIAh: Yellow contours predict hydrophobic substituents enhance activity whereas purple contours predict
hydrophobic substituents decrease activity.