Subscriber access provided by Washington & Lee University Library and VIVA (Virtual Library of Virginia)
Article
Modeling compound-target interaction network of Traditional Chinese Medicines for type II diabetes mellitus: insight for polypharmacology and drug design Sheng Tian, Youyong Li, Dan Li, Xiaojie Xu, Junmei Wang, Qian Zhang, and Tingjun Hou J. Chem. Inf. Model., Just Accepted Manuscript • DOI: 10.1021/ci400146u • Publication Date (Web): 16 Jun 2013 Downloaded from http://pubs.acs.org on June 18, 2013
Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a free service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are accessible to all readers and citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.
Journal of Chemical Information and Modeling is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.
Page 1 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
1 2
Modeling compound-target interaction network of
3
Traditional Chinese Medicines for type II diabetes mellitus:
4
insight for polypharmacology and drug design
5 6
Sheng Tiana, Youyong Lia, Dan Li,b Xiaojie Xuc, Junmei Wangd, Qian Zhanga,
7
Tingjun Houa,b*
8 a
9
Institute of Functional Nano & Soft Materials (FUNSOM) and Jiangsu Key Laboratory for
10
Carbon-Based Functional Materials & Devices, Soochow University, Suzhou, Jiangsu 215123,
11
China b
12
College of Pharmaceutical Sciences, Zhejiang University, Hangzhou, Zhejiang 310058, China
13 14 15
c
College of Chemistry and Molecular Engineering, Peking University, Beijing 100871, China d
Department of Biochemistry, The University of Texas Southwestern Medical Center, 5323 Harry Hines Blvd., Dallas, TX 75390
16 17 18 19
Corresponding author:
20
Tingjun Hou
21
E-mail:
[email protected] 22
Phone: +86-512-65882039
or
[email protected] 23 24
Keywords: Type II diabetes mellitus; Traditional Chinese Medicines; Drug-target
25
interaction network; Molecular docking; Pharmacophore mapping; Naïve Bayesian
26
classifier; Polypharmacology
27 1
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
28 29
For Table of Contents Use Only
30
31 32 33 34 35 36 37 38 39 40 41 42 2
ACS Paragon Plus Environment
Page 2 of 46
Page 3 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
43
Abstract
44
In this study, in order to elucidate the action mechanism of Traditional Chinese
45
Medicines (TCM) that exhibit clinical efficacy for type II diabetes mellitus (T2DM),
46
an integrated protocol that combines molecular docking and pharmacophore mapping
47
was employed to find the potential inhibitors from TCM for the T2DM-related targets
48
and establish the compound-target interaction network. First, the prediction
49
capabilities of molecular docking and pharmacophore mapping to distinguish
50
inhibitors from non-inhibitors for the selected T2DM-related targets were evaluated.
51
The results show that molecular docking or pharmacophore mapping can give
52
satisfactory predictions for most targets but the validations are still quite necessary
53
because the prediction accuracies of these two methods are variable across different
54
targets. Then, the Bayesian classifiers by integrating the predictions from molecular
55
docking and pharmacophore mapping were developed, and the well-validated
56
Bayesian classifiers for 15 targets were utilized to find the potential inhibitors from
57
TCM and establish the compound-target interaction network. The analysis of the
58
compound-target network demonstrates that a small portion (18.6%) of the predicted
59
inhibitors can interact with multi-targets. The pharmacological activities for some
60
potential inhibitors have been experimentally confirmed, highlighting the reliability of
61
the Bayesian classifiers. Besides, it is interesting to find that a considerable number of
62
the predicted multi-target inhibitors have free radical scavenging/antioxidant activities,
63
which are closely related to T2DM. It appears that the pharmacological effect of the
64
TCM formulae is determined not only by the compounds that interact directly with
65
one or more T2DM-related targets, but also by the compounds with other
66
supplementary bioactivities important for relieving T2DM, such as free radical
67
scavenging/antioxidant effects. The mechanism uncovered by this study may offer a
68
deep insight for understanding the theory of the classical TCM formulae for
69
combating T2DM. Moreover, the predicted inhibitors for the T2DM-related targets
70
may provide a good source to find new lead compounds against T2DM.
71 3
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
72
Introduction
73
Diabetes mellitus, or simply diabetes, is a group of metabolic disorder diseases.
74
The diabetes patients have high blood glucose, because either the body system cannot
75
produce enough insulin (absolute insulin deficiency) or the related cells are not
76
sensitive to or respond to insulin (insulin resistance or relative insulin deficiency).
77
Diabetes mellitus has become a serious health problem around the world. Diabetes
78
mellitus is classified into three categories: type 1, type 2 and gestational diabetes.
79
Type 1 diabetes mellitus (T1DM) is characterized by the loss of the insulin-producing
80
beta cells of the islets of langerhans in the pancreas, leading to absolute insulin
81
deficiency. Type 2 diabetes mellitus (T2DM) is characterized by insulin resistance,
82
which may be combined with relatively reduced insulin secretion. The last one,
83
gestational diabetes mellitus (GDM), which occurs in about 2~5% of all pregnancies,
84
resembles T2DM in several respects, involving a combination of relatively inadequate
85
secretion and responsiveness. Globally, it is estimated that 285 million people have
86
diabetes mellitus, with T2DM making up about 90% of the cases in 2010.1 Its
87
incidence is increasing rapidly, and by 2030, this number is estimated to be almost
88
double.2
89
T2DM is a chronic disease, associated with a ten-year shorter life expectancy,
90
caused by a number of complications including cardiovascular, eyes, kidney and
91
nervous system diseases.3-6 Most T2DM patients need the medication to reduce their
92
blood glucose to near-normal level.7 To control the high-level blood glucose and
93
prevent the diabetic complications, many drugs, such as alpha-glucosidase inhibitors,
94
sulfonylureas, metformin, thiazolidinediones (TZDs) and insulin injections, have been
95
used to treat T2DM patients.7 To our disappointment, all of these therapeutic agents
96
have limited efficacy and are associated with undesired side effects, including
97
hypoglycemia, weight gain, gastrointestinal disturbances, edema, anemia, lactic
98
acidosis, etc.8, 9 Due to the deficiencies of marketed drugs for T2DM, the philosophy
99
of rational drug design, or more specifically, the “one gene, one drug, one disease”
100
paradigm for treating T2DM was met with controversy. 4
ACS Paragon Plus Environment
Page 4 of 46
Page 5 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
101
Over the past decade, the number of new molecular entities (NME) approved by
102
US Food and Drug Administration (FDA) did not change significantly. The reasons
103
that have been proposed for this discouraging situation may not be technological,
104
environmental or even scientific but philosophical.10 Also, it has been proven that
105
many effective drugs act on multiple rather than one specific disease-related target.11
106
It seems that the “magic bullets” cannot cure some specific diseases as expected.
107
The new drug candidates with low efficacy and undesired ADME/T properties are
108
not the only factor to debate the traditional paradigm for drug design/discovery.12, 13
109
The phenotype usually cannot be disturbed by single-gene knockouts14 due to
110
redundant functions and alternative compensatory signaling pathways.10,
111
intrinsic robustness of living systems against various perturbations is an essential
112
factor that hinders such magic bullet drugs to be successful in the drug market.16
15
The
113
Compared with the traditional single-target-drug paradigm, the new philosophy
114
for drug design known as polypharmacology or network pharmacology has attracted
115
more attentions.10, 17-19 Therefore, it has been realized that designing drug candidates
116
to target a disease-associated network rather than a single target may be a better
117
strategy to find better drug candidates for some complicated diseases, such as T2DM.
118
It is estimated that approximated 34% of new drugs were directly or indirectly
119
from natural products from 1981 to 2010.20 As an important source of natural products
120
in clinical use, the Traditional Chinese Medicines (TCM) has played an important role
121
in healthcare system of Chinese people for more than several thousand years. Because
122
of satisfactory in vivo efficacy and safety of TCM in targeting complicated chronic
123
diseases,21 finding potential drug candidates from TCM may be a good practice.22, 23
124
The main philosophical theories of TCM are based on Yin-Yang, Five Elements,
125
human body meridian systems and Zang Fu theory.24 It is well known that the basic
126
form of TCM for combating diseases is TCM formula (or prescription), which is the
127
mixture of special herbs with suitable dosages. A TCM formula consists of a large
128
number of chemical compounds, which may interact with multi-targets. Therefore, at
129
the molecular level, the mechanisms of TCM formulae may be explained by 5
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
130
polypharmacology or network pharmacology.
131
Compared with the time-consuming and costly experimental methods to uncover
132
the action mechanism of TCM for combating diseases, the use of computational
133
approaches may provide an easier and more practical way to understand the
134
mysterious power of TCM.25-27 Previous studies suggest that molecular modeling
135
techniques, especially molecular docking and pharmacophore mapping, not only can
136
find potentially active compounds from TCM for targeting disease-related proteins,
137
but also can identify potential targets for compounds from TCM.28-34 However, it
138
seems that most previous studies just simply used computational approaches to find
139
potential inhibitors from TCM or even establish the compound-target networks to
140
explore the polypharmacology of TCM. It should be noted that the validations of the
141
computational approaches is quite critical to guarantee the reliability of the
142
predictions for TCM given by the computational approaches.
143
Through thousands of years of medical use for various diseases in the Chinese
144
medical system, a large amount of knowledge has also been accumulated for diabetes
145
therapy that is termed as “XiaoKeZheng”.24, 35 Similar to the combination therapy of
146
multi-component drugs, the multi-components from TCM can interact with
147
multi-targets through an integrated way to lowing blood glucose or relieving
148
diabetes.21, 36 However, we still know little about how the active compounds in the
149
TCM formulae modulate the network for combating diabetes. Therefore, in this study,
150
we tried to establish the compound-target interaction network by theoretical
151
approaches and uncover the underlying action mechanisms of TCM for T2DM. To
152
make our predictions more reliable, the capabilities of molecular docking and
153
pharmacophore mapping to discriminate the known inhibitors from the non-inhibitors
154
for each T2DM-related target were assessed. To integrate the advantages of molecular
155
docking and pharmacophore modeling, the naïve Bayesian classifiers were developed
156
by combining the predictions from molecular docking and pharmacophore mapping
157
together. The combined models show better performance than those only based on the
158
predictions from molecular docking or pharmacophore mapping. Then, the 6
ACS Paragon Plus Environment
Page 6 of 46
Page 7 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
159
well-validated Bayesian classifiers were utilized to find potentially active compounds
160
in TCM that can interact with the important T2DM-related targets. Finally, the
161
compound-target interaction network was established to uncover the action
162
mechanisms of the classical TCM formulae for relieving T2DM.
163 164
Methods and Materials
165
Collection of the T2DM-related targets
166
By searching the public data sources, including Therapeutic Targets Database
167
(TTD)37, Drugbank38, Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway
168
database39 and previous reported literatures31,
169
T2DM were identified and the available crystal structures of the protein-ligand
170
complexes for these targets were retrieved from the RCSB protein data bank41. For
171
each target, the crystal structure with the highest resolution was reserved if multiple
172
structures were available, and finally a non-redundant set of 48 crystal complexes for
173
the T2DM-related targets was obtained. It is necessary to point out that almost all of
174
the T2DM-related targets may have several synonyms because different standards
175
might be used in different TCM-related databases.37 For example, the corticosteroid
176
11-beta-dehydrogenase isozyme 1 also can be named as 11-Beta hydroxysteroid
177
dehydrogenase 1, 11-beta-HSD1, 11HSD1, and so on.37 The specific target with
178
different names should be check carefully. The names, PDB entries and the
179
corresponding crystal resolutions for the selected protein-ligand complexes are
180
summarized in Table S1 in the Supporting Materials.
40
, the important targets related to
181 182
Collection of the compounds in the T2DM-related TCM formulae
183
By extensive data profiling, 52 classical TCM formulae with clinical efficacy for
184
T2DM were compiled. There are 95 Chinese herbs in those 52 TCM formulae. The
185
frequency of each herb was counted, and only the 32 Chinese herbs with the
186
frequency to 5 or higher were considered in our study (Table S2 in the Supporting
187
Materials). The Latin names for the 32 studied Chinese herbs were collected from the 7
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Page 8 of 46
188
latest version of TCMCD developed in our group42-45 and
[email protected] 189
It should be noted that multiple Latin names might be available for the same Chinese
190
herb. The reason is that different standards may be used in different TCM databases.
191
For example, the herb “Cang Zhu” in Chinese name is named as Latin name
192
Atractylodes lancea in TCMCD and Atractylodes lancea (Thunb.) DC. in
193
TCM-Database@Taiwan (Table S2 in the Supporting Materials). A total of 2,479
194
non-duplicated compounds isolated from these 32 Chinese herbs were found from the
195
TCMCD
196
standardized and optimized by using Discovery Studio 3.1 (DS3.1) molecular
197
simulation package.47
and
TCM-Database@Taiwan
databases.
These
compounds
were
198 199
Preparation of the validation datasets
200
The validation dataset for each T2DM-related target was prepared to evaluate the
201
prediction accuracies of molecular docking or pharmacophore mapping to distinguish
202
known inhibitors from non-inhibitors. The known inhibitors were obtained from the
203
BindingDB database.48 The number of the inhibitors for each target is listed in Table
204
S3 in the Supporting Materials. To ensure the statistical significance of the validation
205
results, the targets with the number of the known inhibitors less than 10 are excluded.
206
Besides, in order to achieve acceptable computational efficiency, when the number of
207
the known inhibitors for a target is higher than 500, only 500 randomly-selected
208
inhibitors
209
experimentally-validated non-inhibitors are quite limited, the compounds were
210
randomly chosen from the Chembridge database to serve as non-inhibitors for each
211
target.49 The inhibitors and non-inhibitors were compared and the duplicates were
212
removed from the non-inhibitor subset. To mimic the unbalanced nature of inhibitors
213
versus non-inhibitors, we chose to set the ratio of non-inhibitors versus inhibitors to
214
20. By applying the processing protocol mentioned above, the validation datasets for
215
the 33 T2DM-related targets were generated (Table S3 in the Supporting Materials).
were
included
in
the
validation
dataset.
216 8
ACS Paragon Plus Environment
Then,
because
the
Page 9 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
217
Validation of molecular docking-based virtual screening
218
The crystal structures of the protein-ligand complexes for the 33 T2DM-related
219
targets were used for the docking calculations performed with Glide50 in Schrodinger
220
9.0.51 For each crystal structure, the crystallographic water molecules were removed,
221
the missing hydrogen atoms were added, and the protonated states and partial charges
222
were assigned with the OPLS-2005 force field52. Minimizations were performed until
223
the average root mean square deviation of non-hydrogen atoms reached 0.3 Å.
224
Preparation and refinement were accomplished with the protein preparation wizard in
225
Schrodinger 9.051.
226
The validation datasets were processed using the LigPrep module in
227
Schrodinger51. For each small molecule, the possible ionization states were generated
228
by using Epik at pH=7.0. Because the non-inhibitors randomly chosen from
229
Chembridge do not have 3D structural information, all possible stereoisomers for
230
each molecule were generated. All of the other parameters were kept as the default
231
settings in LigPrep.
232
Then, the bounding box of size 10 Å × 10 Å × 10 Å was defined for each target and
233
centered on the mass center of the co-crystal ligand to confine the mass center of the
234
docked ligand by using the Receptor Grid Generation function of Glide.51 According
235
to our docking results, the size of the bounding box is big enough for almost all of the
236
ligands to be docked into the binding site successfully for each target. The default
237
Glide settings were used for grid generation. All of the compounds were docked into
238
the binding site of the studied targets and scored by the Standard Precision (SP) and
239
Extra Precision (XP) scoring modes of Glide.51 Five thousand poses per ligand were
240
generated during the initial phase of the docking calculation, and out of which the best
241
400 poses were chosen for energy minimization. Dielectric constant of 2.0 and 100
242
steps of conjugate gradient minimizations were used in energy minimization stage.
243
The performances of the SP and XP scoring modes of Glide on the 33 T2DM-related
244
targets were evaluated and compared.
245 9
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
246
Validation of pharmacophore-based virtual screening
247
The low-energy conformations of the molecules in the validation datasets were
248
generated by using the Conformation Generation protocol in DS3.147 with the FAST
249
setting and the number of the maximum conformations per compound was set to 100.
250
Then, the receptor-ligand interaction patterns (or pharmacophore features) for each
251
target was generated by using the Receptor-Ligand Pharmacophore Generation
252
(RLPG) protocol in DS3.1.47 A set of predefined pharmacophore features that include
253
hydrogen bond acceptor/donor (HBA/HBD), hydrophobic (HYD), negative/positive
254
ionizable (PI/NI) and ring aromatic (RA) of the ligand are identified for each
255
receptor-ligand complex firstly and then pruned when not satisfying the protein-ligand
256
interactions using adjustable topological rules.53
257
The RLPG protocol in DS3.1 was applied to the receptor-ligand complexes for the
258
33 T2DM-related targets. For each pharmacophore model, the minimum number of
259
the pharmacophore features was set to 3, and the maximum number of the
260
pharmacophore features was same to the number of the total features that can match
261
the receptor-ligand interactions. Up to ten pharmacophore models per receptor-ligand
262
complex were built and ranked by the selectivity as the fault settings in DS3.1. The
263
selectivity of a pharmacophore model was evaluated by a rule-based score function
264
based on the genetic function approximation (GFA) technique54. The details of the
265
GFA model for pharmacophore modeling were described in the previous study.55
266
Besides, the discrimination power of a pharmacophore model was determined by the
267
capability to distinguish known inhibitors from non-inhibitors. The discrimination
268
capability of a pharmacophore model is more important than the selectivity for virtual
269
screening. An ideal pharmacophore model used in virtual screening must have the
270
ability not only to find the most promising active compounds, but also to abandon the
271
inactive compounds. The discrimination capability of each pharmacophore model for
272
the corresponding validation set was evaluated.
273 274
Generation of the molecule-target interaction network 10
ACS Paragon Plus Environment
Page 10 of 46
Page 11 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
275
As mentioned above, the performance of molecular docking or pharmacophore
276
modeling for each target was validated. The validation process aiming to distinguish
277
the known inhibitors from non-inhibitors for each target is a binary classification
278
problem. Recently, the naïve Bayesian classification technique was successfully
279
applied to different classification problems by our group.45,
280
classification can process large amounts of data, learn fast, and is tolerant of random
281
noise. It only requires a small amount of training data to estimate the parameters
282
(means and variances of the variables) necessary for classification. Moreover, in the
283
scheme of naïve Bayesian classification, the belonging of a compound in which class
284
can be qualitatively evaluated by the posterior probability. The mathematical
285
procedure to train a naive Bayesian classifier was described previously.45, 56, 57 To
286
integrate the advantages of molecular docking and pharmacophore mapping, each
287
compound in the validation datasets was docked into the binding site firstly and then
288
mapped onto the pharmacophore model for each target. The docking scores given by
289
molecular docking and the fit values given by pharmacophore mapping were used as
290
the independent variables to train the naïve Bayesian classifiers.
56, 57
Naïve Bayesian
291
After rigorous validation, the naïve Bayesian classifiers with reliable predictions
292
for 15 targets were utilized to virtually screen the TCM compound dataset. First, The
293
2,479 non-duplicate compounds from 32 Chinese herbs of the TCM formulae for
294
T2DM (Table S2 in the Supporting Materials) were processed using the LigPrep
295
module in Schrodinger.51 Then, those compounds were docked into the binding sites
296
of the T2DM-related targets by using Glide in Schrodinger 9.0 and scored by the SP
297
and XP scoring modes. Next, each compound was mapped onto the pharmacophore
298
model with highest discrimination capability for each target. Based on the calculations
299
from molecular docking and pharmacophore mapping, each compound was predicted
300
by the reliable naïve Bayesian classifiers for 15 T2DM-related targets.
301
By applying the Bayesian classifiers, the potential inhibitors for each target were
302
determined. In order to balance the prediction accuracy and the complexity of
303
interaction network, only the top 10% of the predicted inhibitors ranked by decreasing 11
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
304
the Bayesian scores were extracted to generate the compound-target interaction
305
networks by using the Osprey network visualization program (version 1.2.0).58
306 307
Results and Discussion
308
The performance of molecular docking-based ligand profiling
309
Before applying molecular docking to screen the TCM compound dataset (Table
310
S1 in the Supporting Materials), the performance of molecular docking to distinguish
311
inhibitors from non-inhibitors for 33 T2DM-related targets (Table S2 in the
312
Supporting Materials) was evaluated.
313
First, as a representative measurement for the accuracy and reliability of
314
molecular docking, the “docking power” was examined.59 The co-crystallized ligands
315
were extracted and then re-docked into the binding sites. The root-mean-squared
316
deviation (RMSD) of the docked pose from the original position for each complex
317
was computed to evaluate whether the Glide docking is reliable for virtual screening.
318
The RMSD was used as the criterion for the success of the prediction. If RMSD of the
319
docked pose was less than or equal to 2.0 Å from the experimentally observed
320
conformation, we considered it to be a successful prediction. As shown in Table S3 in
321
the Supporting Materials, the SP and XP scoring modes can successfully recognize
322
the near-native conformations for 25 and 26 of the entire 33 complexes, i.e. the
323
successful rate are 75.8% and 78.8%, respectively. In addition, only four complexes
324
that do not meet the criterion of RMSD ≤ 2.0 Å by using either SP or XP scoring
325
mode (Table S3 in the Supporting Materials). From this data, it is clear that for most
326
complexes the Glide docking can successfully recognize the correct binding
327
conformations.
328
Then, we evaluated the “discrimination power” of molecular docking to
329
distinguish known inhibitors from non-inhibitors for each target. The student’s t-test
330
was employed to evaluate the significance of the difference between the means of the
331
two distributions of the Glide SP or XP scores for the known inhibitor and
332
non-inhibitor classes (Table S3 in the Supporting Materials). On the whole, the 12
ACS Paragon Plus Environment
Page 12 of 46
Page 13 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
333
molecular docking calculations for most targets give satisfactory results, indicated by
334
the quite low P-values associated with the difference in the mean SP or XP scores of
335
the inhibitors versus those of the non-inhibitors. The P-values for 25 of 33 targets
336
(76%) are lower than 10-20, suggesting that molecular docking can successfully
337
discriminate the known inhibitors from non-inhibitors for most targets. In other words,
338
the Glide docking can be used as a reliable tool to identify known inhibitors from
339
diverse chemical space for each target. The distributions of the Glide SP or XP scores
340
for four targets with good predictions and four targets with poor predictions are
341
illustrated in Figures 1a and 1b as examples.
342
Generally speaking, the XP scoring (extra precision) is believed to be better than
343
the SP scoring (standard precision). However, it is interesting to observe that the XP
344
scoring of Glide does not always show better predictions than the SP scoring. As
345
shown in Table S3 in the Supporting Materials, the XP scoring only shows better or
346
equal discrimination power for 15 (45.5%) targets according to the measurement of
347
the P-values. The results suggest that the XP scoring does not always yield better
348
performance than SP. Therefore, the comparison study of the performance between
349
the SP and XP scoring modes of Glide is quite necessary for a specific target. For each
350
target, based on the comparison results, more reliable scoring mode was chosen for
351
the following virtual screening of the TCM compound dataset.
352
In summary, the Glide docking based on the SP or XP scoring mode can
353
successfully recognize the near-native conformations for 25 or 26 of the entire 33
354
complexes. It is indicated that the Glide docking is suitable for our studies. Moreover,
355
it can successfully discriminate the known inhibitors from non-inhibitors for 25 of 33
356
targets, proving the reliability of molecular docking for most targets. Ideally, if the
357
results of docking can meet both of the two requirements: RMSD ≤ 2.0 Å and P-value
358
≤ 10-20, the docking-based virtual screening is reliable.
359
The reason of the failure of the docking calculations for some targets lies in the
360
fact that the performance of scoring function used by Glide is often varied across
361
different protein families.59,
60
For example, for HMG-CoA reductase, six crystal 13
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
362
complexes can be retrieved from PDB database (Table S4 in the Supporting
363
Materials). According to our analysis, there is only one complex (1DQA) can meet the
364
demand of RMSD (≤ 2.0 Å) by using the SP mode of Glide. Meanwhile, the
365
capability of distinguishing the known inhibitors from non-inhibitors for 1DQA
366
complex is not good, indicated by the high P-value (1.13×10-6) associated with the
367
difference in the mean SP scores of the known inhibitors versus that of the
368
non-inhibitors (Figure S1 in the Supporting Materials). That is to say, all of the
369
docking calculations by using the SP and XP scoring modes for HMG-CoA reductase
370
are questionable either duo to unacceptable RMSD or poor discrimination capacity. In
371
most previous application of molecular docking on the TCM studies, the capabilities
372
of molecular docking on the studied systems were not evaluated, and therefore, the
373
reported results may be questionable. According to our calculations, it is strongly
374
recommended that the validation process is necessary before any docking study for
375
TCM.
376 377
The performance of pharmacophore-based ligand profiling
378
Based on the 33 receptor-ligand complexes of the T2DM-related targets (Table S5
379
in the Supporting Materials), 28 complex-based pharmacophore models were
380
generated (Table S5 in the Supporting Materials), and no pharmacophore model can
381
be generated for 5 complexes. There are two possible reasons for explaining the
382
failures for these 5 complexes:(1) the number of extracted pharmacophore features is
383
quite limited (less than 3), and (2) some pharmacophore features extracted from the
384
receptor-ligand complex do not comply with the minimum feature distance criterion
385
(< 1.0 Å).
386
For each complex, the generated pharmacophore models were ranked by the
387
selectivity scores (Table S5 in the Supporting Materials). Then, the prediction power
388
of each pharamcophore model to distinguish the known inhibitors from non-inhibitor
389
in the validate dataset was evaluated. The discrimination power of a pharmacophore
390
model was quantitatively characterized by the AUC area under a receiver operating 14
ACS Paragon Plus Environment
Page 14 of 46
Page 15 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
391
characteristics (ROC) curve. All pharmacophore features derived from the
392
receptor-ligand interactions (Total features), the pharmacophore features for the most
393
selective pharmacophore model (Feature set A) and those for the pharmacophore
394
model with the highest AUC value (Feature set B) are listed in Table S5 in the
395
Supporting Materials. The comparison of the numbers of the pharmacophore features
396
for 25 targets among Total features, Feature set A and Feature set B are shown in the
397
Figure 2a and listed in Table S6 in the Supporting Materials. We can observe that the
398
numbers of Total features are generally more than those of Feature set A, and the
399
numbers of Feature set A are also more than those of Feature set B for the same
400
complex (Table S5 in the Supporting Materials) for most cases. Besides, the
401
percentage of the number of the pharmacophore features that are more than five for
402
Total features, Feature set A and Feature set B are 71.4% (20/28), 64.3% (18/28) and
403
32.1% (9/28), respectively.
404
It is expected that a pharmacophore model with more pharmacophore features
405
should be more selective than that with less pharmacophore features. According to our
406
analysis, the numbers of Feature set A for most selective pharmacophore models are
407
not always equal to those of Total features. If the number of Total features derived
408
from a complex is n, the numbers of Feature set A for most selective pharmacophore
409
models determined by GFA are usually equal to n (14/28) or n-1 (11/28), except n-2
410
for PPAR-γ, PTP1B and estrogen receptor. The results demonstrate that some
411
pharmacophore features derived directly from the protein-ligand interactions have no
412
effect on the identification of the known inhibitors from chemical space. Meanwhile,
413
the numbers of Feature set B of the pharmacophore models that have the highest
414
discrimination capacity determined by the AUC values are usually n-1 (11/28) or n-2
415
(11/28), except n for angiotensin-converting enzyme, insulin-like growth factor 1
416
receptor and mitogen-activated protein kinase 1, n-3 for PPAR-γ, PTP1B and estrogen
417
receptor. By comparing the pharmacophore features between Feature set A and
418
Feature set B, it was found that the numbers of the pharmacophore models with
419
Feature set B equaling to, one less than, and two less than those in Feature set A are 15
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
420
21.4% (6/28), 64.3% (18/28) and 14.3% (4/28), respectively.
421
In summary, the pharmacophore models with Feature set B determined by the
422
validation datasets are more reasonable to be used than those with Feature set A
423
determined by GFA (Table S5 in the Supporting Materials and Figure 2b). If AUC
424
higher than 0.70 was used as the criterion to judge a successful prediction, 5 (18%) of
425
the most selective pharmacophore models yield acceptable predictions while 15 (55%)
426
of the pharmacophore models with the best discrimination capabilities yield
427
acceptable predictions. The most selective pharmacophore model was optimized by
428
GFA, but it should be cautiously used in virtual screening when there are no enough
429
known inhibitors for the validation. The aim of virtual screening is to find more
430
promising compounds (true positives) and avoid false positives from the huge
431
chemical space. The most distinguishing pharmacophore models, which are validated
432
by the validation dataset, are more suitable to be used in structure-based virtual
433
screening. Obviously, the validation process before any pharmacophore modeling
434
study for TCM is quite necessary.
435 436
The Bayesian classifier by combining the predictions from molecular docking
437
and pharmacophore mapping
438
As mentioned above, molecular docking gives good predictions for 21 targets, and
439
pharmacophore mapping gives acceptable predictions for 15 targets. Both of
440
molecular docking (RMSD ≤ 2.0 Å and P-value ≤ 10-20) and pharmacophore mapping
441
(AUC ≥ 0.7) give good predictions for 8 targets, and five targets cannot be well
442
predicted by either molecular docking or pharmacophore mapping. However,
443
molecular docking or pharmacophore mapping can give relatively reliable predictions
444
for 28 T2DM-related targets (28/33=84.8%). It must keep in mind that the virtual
445
screening based on molecular docking or pharmacophore mapping is a binary
446
classification problem. Then, we may raise the following question: could the
447
combination of these two structure-based methods enhance the prediction? For each
448
molecule, the docking score and fit value are the quantitative measurements for 16
ACS Paragon Plus Environment
Page 16 of 46
Page 17 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
449
molecular docking and pharmacophore mapping, respectively. Therefore, the naïve
450
Bayesian classifiers were built by using the docking scores and fit values as the
451
independent variables for 24 targets with reliable predictions of molecular docking
452
and/or phamacophore mapping. Three criteria should be satisfied for the chosen
453
targets: (1). the number of known inhibitors should be equal to or higher than 10; (2).
454
the native-like conformation of the ligand should be reproduced by molecular docking
455
(RMSD ≤ 2.0 Å); (3). reasonable pharmacophore model should be generated. In order
456
to compare the prediction capacity among molecular docking, pharmacophore
457
modeling and the combination of the predictions from both, the AUC values with
458
5-fold cross validation were employed to measure the performance of the naïve
459
Bayesian classifiers (Table 1 and Figure 3). According to statistical results shown in
460
Table 1, we observe that the AUC values of the Bayesian classifiers based on the
461
docking scores and fit values are higher than those only based on the docking scores
462
or fit values for almost all targets (except PTP1B). For example, for insulin-like
463
growth factor 1 receptor (IGF-1R), the AUC values for the Bayesian classifier based
464
on the fit values and that based on the docking scores are 0.729 and 0.788,
465
respectively, while the AUC value for the Bayesian classifier based on both the fit
466
values and docking scores is 0.904 (Figure 4). The reason why the naïve Bayesian
467
classifier based on both fit values and docking scores for PTP1B cannot get the
468
improvement was also investigated. We found that only six known inhibitors could be
469
docked into the binding site of PTP1B and mapped onto the pharmacophore model of
470
the PTP1B complex, and the limited number of the PTP1B inhibitors could not
471
guarantee the statistical reliability of the validation results.
472
Here, the AUC values in the range of 0.5-0.6, 0.6-0.7, 0.7-0.8, 0.8-0.9 and 0.9-1.0
473
represent fail, poor, fair, good and excellent predictions for the Bayesian classifiers
474
(Table S7 in the Supporting Materials and Figure 3). The definition of AUC for
475
indicating the performance of prediction models is consistent with the definition used
476
by the Discovery Studio simulation package. The Bayesian classifiers built based on
477
docking scores or fit values can only give good or excellent predictions for 4 or 9 17
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
478
targets, while the combination of these two measurements can give good or excellent
479
predictions for 15 targets. Clearly, the combination of docking scores and fit values
480
can enhance the discrimination capacity.
481 482
The deep analysis of the potential active compounds for the T2DM-related
483
targets
484
As we discussed in the above section, for 15 targets (15/24=62.5%), the Bayesian
485
classifiers based on both docking scores and fit values give reliable predictions (AUC
486
≥ 0.8), and therefore, only the Bayesian classifier for these 15 targets were employed
487
to virtually screen the TCM compound dataset.
488
The numbers of the predicted inhibitors are listed in Table S8 in the Supporting
489
Materials. For protein-tyrosine kinase erbB-2, no potential inhibitor was found by the
490
Bayesian classifier. For protein-tyrosine kinase erbB-2, there are only ten inhibitors
491
for building and validating the Bayesian classifier. As shown in Figure S2 in the
492
Supporting Materials, these ten inhibitors almost have the same molecular framework
493
(Murcko framework).61 Therefore, the conservation of the target-inhibitor interaction
494
patterns may be unfavorable to generate a general prediction model for finding other
495
dissimilar compounds from the TCM compound dataset. Then, we investigated the
496
correlation between the number of the potential inhibitors and the volume and
497
hydrophobic property of the binding pocket for each target (Table S8 in the
498
Supporting Materials), and did not find obvious relationship between the number of
499
the potential inhibitors for each target and the volume or hydrophobic property of the
500
binding pocket.
501
It is well known that two similar compounds or two compounds that share similar
502
framework usually have similar pharmacological effects.62 Therefore, the similarities
503
between the potential inhibitors from TCM predicted by the Bayesian classifiers and
504
the known inhibitors for each target (Table S8 in the Supporting Materials) were
505
evaluated by the 2-D similarities (Tanimoto Coefficient) based on the MDLPubicKeys
506
fingerprints63 supported in DS3.1. The results are summarized in Table S9 in the 18
ACS Paragon Plus Environment
Page 18 of 46
Page 19 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
507
Supporting Materials. It is encouraging to find that a number of the potential
508
inhibitors are similar to the known inhibitors for 12 of 15 targets when the similarity
509
coefficient higher than 0.6 was used as the threshold. In addition, for insulin receptor
510
and estrogen receptor, 7 and 11 potential inhibitors have the similarity coefficient of
511
1.0 to the known inhibitors. For estrogen receptor, 9 of the 11 potential inhibitors are
512
the known inhibitors (Figure 5a), and the other 2 are quite similar to the known
513
inhibitors of estrogen receptor (Figure 5b).
514
It is interesting to find that all of the 8 compounds shown in Figure 5a were
515
directly isolated from natural products. Be specific, the compounds 2 (Emodin), 6
516
(Genistein) and 7 (Daidzein) were isolated from Polygonum cuspidatum
517
(polygonaceae),64 the compound 8 (Broussonin A) from Broussonetia Kazinoki,65 the
518
compound 4 (8-PN) from Humulus lupulus L.,66 the compound 3 (Coumestrol) from
519
plant-based natural molecules or derivatives,67 and the compounds 1 (Resveratrol)68
520
and 5 (Apigenin)69 also from natural plants. It should be noted that all of the 8
521
compounds are not only found in the natural plants mentioned above, but also found
522
in the Chinese herbs in the classical TCM formulae for treating T2DM. Furthermore,
523
for insulin receptor, we can also observe that 21 predicted inhibitors have the
524
similarity coefficients higher than 0.9 to the known inhibitors of insulin receptor
525
(Table S9 in the Supporting Materials). The structural analysis shows that 18 of the 21
526
molecules are quite similar to the known inhibitor, myricetin, of insulin receptor. The
527
myricetin and the 18 potential inhibitors similar to myricetin are shown in Figure 6.
528
Obviously, these 18 molecules share the same molecular framework (Murcko
529
framework)61 with myricetin and can be seen as the derivatives of myricetin.
530
In summary, our results imply that the Bayesian classifiers by combining the
531
predictions from molecular docking and pharmacophore mapping have exciting
532
capability to find promising inhibitors from TCM for the T2DM-related targets. Many
533
potential inhibitors predicted by the Bayesian classifiers are quite similar or even
534
identical to the known inhibitors. Moreover, by the 2D similarity analysis, we can
535
observe that some potential inhibitors are dissimilar (low similarity) to the known 19
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Page 20 of 46
536
inhibitors for most targets and they may be novel lead compounds for the studied
537
targets. Our observation can provide a deep insight for understanding the action
538
mechanism of TCM formulae for T2DM. The main mechanism of the TCM formulae
539
for treating T2DM may be similar to western drugs: the active compounds or
540
inhibitors in TCM can interact with the important disease-related targets, forming the
541
leading force for combating diabetes.
542 543
Analysis of the compound-target interaction network
544
A TCM formula usually consists of various herbs with numerous compounds to
545
exert potency towards diseases. The action mechanism of TCM is also termed as
546
“cocktails
547
multi-targets. Two possible mechanisms may be proposed for the pharmacological
548
effect of the TCM formulae for treating T2DM: (1). multi-active-components interact
549
with different T2DM-related targets to achieve synergic effect, (2). limited individual
550
components interact with multi-targets and most of them have activities against single
551
specific T2DM-related targets.
of
drugs”,
which
represents multi-components
interacting
with
552
After employing the Bayesian classifiers to virtually screen the TCM compound
553
dataset, 1,996 non-duplicated compounds were determined to be potential inhibitors
554
for 14 targets. Among these 1,996 compounds, 1,590 (74.38%) of them were
555
predicted to be drug-like by using the drug-likeness model developed in our group.45
556
The distributions of eight important molecular physicochemical properties, including
557
MW, PSA, AlogP, logD, logS, the number of hydrogen bond acceptors/donors and the
558
number of rotatable bonds, for these compounds are displayed in Figure S3 in the
559
Supporting Materials. In order to achieve more reliable predictions, only the top 10%
560
of the compounds (693) ranked by decreasing the Bayesian scores for each target
561
were chosen for the following analysis (Table S8 in the Supporting Materials). The
562
numbers of the predicted potential inhibitors for each target are listed in Figure 7.
563
It is interesting to find that there are a number of potential inhibitors that interact
564
with multiple targets (Figure 8 and Table S10 in the Supporting Materials). There are 20
ACS Paragon Plus Environment
Page 21 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
565
519 non-duplicated predicted inhibitors against a single target, 94 against two targets,
566
28 against three targets, 8 against four targets, and only one against five targets. The
567
compound-target interaction networks are shown in Figure 9. As can be seen from
568
Figure 8, the number of the potential inhibitors decreases rapidly with the increase of
569
the number of targets. Apparently, most compounds in Chinese herbs have potential
570
activities against limited targets. As can be seen in Figure 9a, the whole
571
compound-target network has 519 nodes (potential inhibitors) and 696 edges
572
(interactions) with an average of 1.34 interactions per node. About 81.39% (564/693)
573
of the potential inhibitors only interact with one target. According to our analysis, we
574
can conclude that the mechanism of the TCM formulae may belong to the second
575
hypothesis: limited individual components interact with multi-targets and the others
576
only interact with the single specific T2DM-related targets.
577
We then investigated the potential multi-target inhibitors predicted by the
578
Bayesian classifiers. The similarities between the potential multi-target inhibitors and
579
the known inhibitors for each target were calculated. When the minimum similarity
580
coefficient based on the MDLPubicKeys fingerprint was set to 0.6, no similar known
581
inhibitor was found for these potential inhibitors for all of four and five targets, but
582
similar known inhibitors were successfully found for some potential inhibitors for all
583
of two and three targets.
584
For the only potential inhibitor (mirificin) against five targets (carbonic anhydrase
585
1, carbonic anhydrase 2, FBPase, glycogen synthase kinase-3 beta and insulin
586
receptor), similar known inhibitors for three of five targets (carbonic anhydrase 1,
587
carbonic anhydrase 2 and insulin receptor) could be found. For the 8 potential
588
inhibitors against four targets, similar known inhibitors for two or three of four targets
589
could be found. Besides, for the predicted inhibitors against two and three targets,
590
similar known inhibitors for all of two or three targets could be found. For examples,
591
the compound (Pureoside A) that was predicted to be a potential inhibitor of carbonic
592
anhydrase 1, PPAR delta and PPAR gamma, and similar known inhibitors for each
593
target were found and shown in Figure 10; two compounds (Epicatechin and 21
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
594
(2RS)-2-(4-O-β-D-glucopyranosylphenyl) propionyl β-D-glucopyranoside) that were
595
predicted to be potential inhibitors for two targets and similar known inhibitors were
596
found and shown in Figure 11.
597
In summary, our predictions suggest that most of the compounds extracted from
598
the Chinese herbs in TCM formulae can only interact with a single specific target.
599
Although the theory of TCM formulae is multi-compounds for multi-targets,
600
according to our analysis, it is obvious that only limited compounds from the TCM
601
formulae for T2DM show activities against multi-targets. The 2D-similarity analysis
602
shows that some known inhibitors are similar to the predicted potential inhibitors
603
against multi-targets. Furthermore, the low similarities between some of the predicted
604
potential inhibitors and the known inhibitors for each target suggest that some novel
605
potential inhibitors might be found from the predictions given by the Bayesian
606
classifiers. Those novel potential inhibitors may provide good source for
607
pharmaceutical chemists to find promising leads for treating T2DM.
608 609
The possible mechanisms of the classical TCM formulae for treating T2DM
610
As discussed in the above section, it is demonstrated that the 519 non-duplicated
611
potential inhibitors can interact with 14 T2DM-related targets. More than 80% of the
612
potential inhibitors only interact with one target, and the other potential inhibitors
613
interact with two or more targets. The multiple compounds from one herb or different
614
herbs of TCM formulae may form synergistic interactions for combating T2DM. Then,
615
the pharmacological activities of those predicted inhibitors that can interact with two
616
or more targets were examined by extensive literature searching. First, we analyzed
617
the nine compounds that interact with at least four targets (Table S10 in the
618
Supporting Materials). The only compound, mirificin, which was predicted to be an
619
inhibitor of five targets (carbonic anhydrase 1, carbonic anhydrase 2, FBPase,
620
glycogen synthase kinase-3 beta and insulin receptor), has free radical scavenging
621
activity.70 Moreover, the eight compounds that are predicted to be the inhibitors of
622
four targets have reported activities. Interestingly, the reported activities for three 22
ACS Paragon Plus Environment
Page 22 of 46
Page 23 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
623
compounds (Apiopaeonoside, Kakkalide and tectorigenin 7-O-xylosylglucoside) are
624
not related to the studied targets while related to aldose reductase.71,
625
studies have shown the close relationship between aldose reductase and free radical
626
scavenging/antioxidant.73 Aldose reductase might induce the oxidative damage
627
resulted from the competition between aldose reductase and glutathione reductase for
628
NADPH. Furthermore, the inhibitors of aldose reductase not only show activities for
629
aldose reductase, but also have antioxidant activity.74 By checking the top 10% of the
630
predicted inhibitors of aldose reductase (Table S8 in the Supporting Materials), we
631
also found that 6 of the 15 compounds have free radical scavenging/antioxidant
632
activities. According to those observations, the linkage between the three potential
633
inhibitors for four targets and the one potential inhibitor for five targets can be
634
established. Although the predicted potential inhibitors for more than four targets do
635
not have reported activities as predicted by us, all of those four compounds have the
636
reported activities of free radical scavenging or antioxidant. Besides, the previous
637
studies have shown that there were direct relationship between free radical
638
scavenging/antioxidant and T2DM.75 The level of free radical increases with the
639
T2DM disease progressing76, and in addition, some complications of T2DM,
640
including obesity, retinopathy and nephropathy, and some T2DM-related diseases,
641
such as coronary arteriosclerotic heart disease and arteriosclerosis, are also related to
642
the out of control of free radical.77-81 At last, it is necessary to emphasize that some
643
potential inhibitors that can interact only one, two or three targets also possess the
644
reported activities of free radical scavenging/antioxidant, for example, the compound,
645
apigenin-7-glucoside, a potential inhibitor of two targets (insulin receptor and
646
pancreatic alpha-amylase). Actually, the activity of apigenin-7-glucoside against
647
pancreatic alpha-amylase has been reported;82 in addition, this compound also has
648
aldose reductase and free radical scavenging/antioxidant activities.83-85 The compound,
649
4'-O-β-glucopyranosyl-3',5,7-trihydroxyflavone with predicted activities for carbonic
650
anhydrase 1 and pancreatic alpha-amylase targets have aldose reductase and free
651
radical scavenging activities.84, 86 For the compound, taxifolin, a predicted inhibitor 23
ACS Paragon Plus Environment
72
Previous
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
652
against three targets (carbonic anhydrase 2, estrogen receptor and pancreatic
653
alpha-amylase), the experimental activity against estrogen receptor has been
654
reported,87 and this compound also shows free radical scavenging/antioxidant,88, 89
655
PI3K inhibition,90 and PPAR gamma inhibition activities.91 Moreover, we also found
656
that some potential multi-target inhibitors did not have any documented activities for
657
the T2DM-related targets. For example, the compound, lobetyolin, was predicted to be
658
a potential inhibitor of carbonic anhydrase 2, insulin receptor and pancreatic
659
alpha-amylase targets, but the reported activity of lobetyolin is antibacterial.92 It was
660
interesting to find that the known drug of T2DM, glibenclamide, also has antibacterial
661
activity. The antibacterial activity is also found to be valuable in relieving diabetes
662
and the associated secondary disorders.93 According to our observations, we can find
663
that some of the predicted potential inhibitors have reported pharmacological
664
activities related to T2DM for more than one target.
665
In summary, we have examined the pharmacological activities of the multi-targets
666
compounds extensively. It was found that some pharmacological effects of the
667
predicted inhibitors of multi-targets have been reported previously, highlighting the
668
reliable predictions of the Bayesian classifiers. Furthermore, some potential inhibitors
669
predicted by our study need experimental verifications. Then, during the analysis of
670
the pharmacological activities of potential multi-target inhibitors, it was found that
671
most of these potential compounds have free radical scavenging/antioxidant activities
672
that have been proved to be effective to relieve T2DM and the related diseases of
673
T2DM.
674
The theory of TCM formulae is multi-compounds versus multi-targets. The
675
hundreds or even thousands of compounds extracted from the herbs of TCM formulae
676
interact with multi-targets to achieve synergistic effects. How the compounds from the
677
herbs of TCM formulae affect the T2DM-related targets is also unclear and needs to
678
be uncovered. According to our analysis, it was found that most of the predicted
679
potential inhibitors can only interact with a single target. However, a few of them still
680
can interact with more than one target. Some pharmacological effects, such as radical 24
ACS Paragon Plus Environment
Page 24 of 46
Page 25 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
681
scavenging/antioxidant and antibacterial activities, which are not directly related to
682
the studied targets, were found. Most compounds from the herbs in the TCM formulae
683
only interact with an individual target, forming the leading fighting force to combat
684
T2DM. Then, those potential multi-target compounds may influence the
685
T2DM-related targets, forming additional forces to enhance the therapeutic effects. At
686
last, a portion of the compounds are responsible for remedying the other related
687
symptoms that are proved to be related to T2DM, such as abnormal status of free
688
radical and oxidation in the human body. All of these observation found in this study
689
can be seen as a proper way to reveal the classical theory “Monarch, Minster,
690
Assistant and Guide” in TCM prescriptions at the molecular level.
691 692
Conclusion
693
In this study, we employed structure-based and complex-based virtual screening
694
approaches to find potential inhibitors from the Chinese herbs of the classical TCM
695
formulae for T2DM against the important T2DM-related targets. In order to achieve
696
more reliable predictions, the performance of molecular docking and pharmacophore
697
mapping for the 33 T2DM-related targets was evaluated. By combining the
698
advantages of molecular docking and pharmacophore mapping together, the Bayesian
699
classifiers were developed by integrating the predictions from molecular docking and
700
pharmacophore mapping. The validation results show that the integrated models are
701
superior to that based on the prediction from molecular docking or pharmacophore
702
mapping only.
703
The Bayesian classifiers with satisfactory discrimination capabilities for 15
704
targets were employed to screen the 2,479 compounds from the Chinese herbs of the
705
TCM formulae for T2DM. The pharmaceutical activities of some compounds
706
predicted by the Bayesian classifiers are confirmed by the experimental studies. In
707
order to uncover the underlying mechanism of the TCM formulae for combating
708
T2DM, the top 10% of the potential compounds ranked by the Bayesian scores for
709
each target were extracted to build the compound–target interaction network. The 25
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
710
analysis of the compound-target network shows that ~19% of the potential inhibitors
711
interact with more than one target. The examination of the pharmacological activities
712
demonstrates that the biological activities for a small portion of these potential
713
multi-target inhibitors have been reported, but the underlying activities for most
714
potential multi-target inhibitors need experimental verification. Besides, it is
715
important for us to find that a variety of the potential multi-target inhibitors have free
716
radical scavenging/antioxidant activities, which are closely related to T2DM. We can
717
conclude that the pharmacological effects of the TCM formulae for T2DM are not
718
only determined by the compounds that interact directly with one or more
719
T2DM-related targets, but also by the compounds with other supplementary
720
bioactivities important for treating T2DM, such as free radical scavenging/antioxidant
721
activities. The holistic way to express join forces for treating T2DM may be the
722
essence of the TCM formulae to break the robustness of the phenotype of T2DM and
723
restore the human body to become normal physiological condition.
724 725
Acknowledgements
726
This study was supported by the National Science Foundation of China (21173156),
727
the National Basic Research Program of China (973 program, 2012CB932600), and
728
the Priority Academic Program Development of Jiangsu Higher Education Institutions
729
(PAPD)
730 731
Supporting Information
732
Table S1. The names, PDB entries and crystal resolutions for the 48 T2DM-related
733
targets; Table S2. The 32 Chinese herbs in the TCM classical formulae for T2DM
734
with the highest frequency; Table S3. The docking power and discrimination power
735
of the Glide docking for 33 T2DM-related targets; Table S4. The docking power of
736
the Glide docking based on the six crystal structures of the complexes of HMG-CoA
737
reductase; Table S5. The numbers of Total features, Feature set A and Feature set B,
738
and the selectivity and discrimination capabilities of the pharmacophore models based 26
ACS Paragon Plus Environment
Page 26 of 46
Page 27 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
739
on Feature set A and Feature set B; Table S6. The comparisons of the number of
740
pharmacophore features in Total features, Feature set A and Feature set B; Table S7.
741
The comparison of the performance of the naïve Bayesian classifiers built by using
742
docking scores, fit values and the combination of docking scores and fit values; Table
743
S8. The volume and hydrophobic properties of the binding pockets and the numbers
744
of the potential inhibitors predicted by the naïve Bayesian classifiers for 15 targets;
745
Table S9. The number of the potential inhibitors from the TCM formulae that are
746
similar to the known inhibitors for 15 targets by using different thresholds of the
747
similarity based on the MDLPublicKeys fingerprints; Table S10. The number of the
748
predicted inhibitors that can interact with two, three, four and five targets; Figure S1.
749
The distributions of the docking scores of the known inhibitors and non-inhibitors by
750
using the (a) SP and (b) XP scoring modes of Glide for 1DQA of HMG-CoA
751
reductase; Figure S2. Ten known inhibitors of receptor protein-tyrosine kinase erbB-2
752
target with similar Murcko framework (colored in red); Figure S3. The distributions
753
of the eight important physicochemical properties for the 1,996 non-duplicated
754
potential inhibitors for 15 targets. This information is available free of charge via the
755
Internet at http://pubs.acs.org
756 757
References
758 759 760 761 762 763 764 765 766 767 768 769 770 771 772
1.
Smyth, S.; Heron, A. Diabetes and obesity: the twin epidemics. Nature Med. 2006, 12, 75-80.
2. Wild, S.; Roglic, G.; Green, A.; Sicree, R.; King, H. Global prevalence of diabetes - Estimates for the year 2000 and projections for 2030. Diabetes Care 2004, 27, 1047-1053. 3. Florez, H.J.; Sanchez, A.A.; Marks, J.B. Type 2 Diabetes. Diabetes and the Brain 2010, 33-53. 4. Pasquier, F. Diabetes and cognitive impairment: how to evaluate the cognitive status? Diabetes Metab. 2010, 36, S100-S105. 5. Reaven, G.M. Pathophysiology of insulin-resistance in human-disease. Physiol. Rev. 1995, 75, 473-486. 6. Ripsin, C.M.; Kang, H.; Urban, R.J. Management of Blood Glucose in Type 2 Diabetes Mellitus. Am. Fam. Physician 2009, 79, 29-36. 7. Morral, N. Novel targets and therapeutic strategies for type 2 diabetes. Trends Endocrin. Met. 2003, 14, 169-175. 8. Marcus, A.O. Safety of drugs commonly used to treat hypertension, dyslipidemia, and type 2 diabetes (the metabolic syndrome): part 1. Diabetes Technol. Ther. 2000, 2, 101-10. 9. Marcus, A.O. Safety of drugs commonly used to treat hypertension, dyslipidemia, and Type 2 27
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816
diabetes (the metabolic syndrome): part 2. Diabetes Technol. Ther. 2000, 2, 275-81. 10. Hopkins, A.L. Network pharmacology: the next paradigm in drug discovery. Nature Chem. Biol. 2008, 4, 682-690. 11. Roth, B.L.; Sheffler, D.J.; Kroeze, W.K. Magic shotguns versus magic bullets: selectively non-selective drugs for mood disorders and schizophrenia. Nat. Rev. Drug Discovery 2004, 3, 353-359. 12. Hou, T.; Wang, J. Structure - ADME relationship: still a long way to go? Expert Opin. Drug Metab. Toxicol. 2008, 4, 759-770. 13. Hou, T.J.; Xu, X.J. Recent development and application of virtual screening in drug discovery: An overview. Curr. Pharm. Des. 2004, 10, 1011-1033. 14. Winzeler, E.A.; Shoemaker, D.D.; Astromoff, A.; Liang, H.; Anderson, K.; Andre, B.; Bangham, R.; Benito, R.; Boeke, J.D.; Bussey, H.; Chu, A.M.; Connelly, C.; Davis, K.; Dietrich, F.; Dow, S.W.; El Bakkoury, M.; Foury, F.; Friend, S.H.; Gentalen, E.; Giaever, G.; Hegemann, J.H.; Jones, T.; Laub, M.; Liao, H.; Liebundguth, N.; Lockhart, D.J.; Lucau-Danila, A.; Lussier, M.; M'Rabet, N.; Menard, P.; Mittmann, M.; Pai, C.; Rebischung, C.; Revuelta, J.L.; Riles, L.; Roberts, C.J.; Ross-MacDonald, P.; Scherens, B.; Snyder, M.; Sookhai-Mahadeo, S.; Storms, R.K.; Veronneau, S.; Voet, M.; Volckaert, G.; Ward, T.R.; Wysocki, R.; Yen, G.S.; Yu, K.X.; Zimmermann, K.; Philippsen, P.; Johnston, M.; Davis, R.W. Functional characterization of the S-cerevisiae genome by gene deletion and parallel analysis. Science 1999, 285, 901-906. 15. Kitano, H. Towards a theory of biological robustness. Mol. Syst. Biol. 2007, 3. 16. Kitano, H. Innovation - A robustness-based approach to systems-oriented drug design. Nat. Rev. Drug Discovery 2007, 6, 202-210. 17. Berger, S.I.; Iyengar, R. Network analyses in systems pharmacology. Bioinformatics 2009, 25, 2466-2472. 18. Goh, K.-I.; Cusick, M.E.; Valle, D.; Childs, B.; Vidal, M.; Barabasi, A.-L. The human disease network. Proc. Natl. Acad. Sci. U. S. A. 2007, 104, 8685-8690. 19. Yildirim, M.A.; Goh, K.-I.; Cusick, M.E.; Barabasi, A.-L.; Vidal, M. Drug-target network. Nature Biotechnol. 2007, 25, 1119-1126. 20. Newman, D.J.; Cragg, G.M. Natural Products As Sources of New Drugs over the 30 Years from 1981 to 2010. J. Nat. Prod. 2012, 75, 311-335. 21. Zhao, J.; Jiang, P.; Zhang, W. Molecular networks for the study of TCM Pharmacology. Brief. Bioinform. 2009, 11, 417-430. 22. Ji, H.-F.; Li, X.-J.; Zhang, H.-Y. Natural products and drug discovery Can thousands of years of ancient medical knowledge lead us to new and powerful drug combinations in the fight against cancer and dementia? EMBO Rep. 2009, 10, 194-200. 23. Newman, D.J.; Cragg, G.M.; Snader, K.M. The influence of natural products upon drug discovery. Nat. Prod. Rep. 2000, 17, 215-234. 24. Lukman, S.; He, Y.; Hui, S.-C. Computational methods for Traditional Chinese Medicine: A survey. Comput. Methods Programs Biomed. 2007, 88, 283-294. 25. Chang, S.-S.; Huang, H.-J.; Chen, C.Y.-C. Two Birds with One Stone? Possible Dual-Targeting H1N1 Inhibitors from Traditional Chinese Medicine. PLoS Comput. Biol. 2011, 7. 26. Chen, K.-C.; Chang, S.-S.; Tsai, F.-J.; Chen, C.Y.-C. Han ethnicity-specific type 2 diabetic treatment from traditional Chinese medicine? J. Biomol. Struct. Dyn. 2012, 1-17. 27. Tou, W.I.; Chang, S.-S.; Lee, C.-C.; Chen, C.Y.-C. Drug Design for Neuropathic Pain Regulation 28
ACS Paragon Plus Environment
Page 28 of 46
Page 29 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860
from Traditional Chinese Medicine. Sci. Rep. 2013, 3. 28. Chen, K.-C.; Chang, K.-W.; Chen, H.-Y.; Chen, C.Y.-C. Traditional Chinese medicine, a solution for reducing dual stroke risk factors at once? Mol. Biosyst. 2011, 7, 2711-2719. 29. Chen, X.; Ung, C.Y.; Chen, Y.Z. Can an in silico drug-target search method be used to probe potential mechanisms of medicinal plant ingredients? Nat. Prod. Rep. 2003, 20, 432-444. 30. Ehrman, T.M.; Barlow, D.J.; Hylands, P.J. In silico search for multi-target anti-inflammatories in Chinese herbs and formulas. Bioorg. Med. Chem. 2010, 18, 2204-2218. 31. Gu, J.; Zhang, H.; Chen, L.; Xu, S.; Yuan, G.; Xu, X. Drug-target network and polypharmacology studies of a Traditional Chinese Medicine for type II diabetes mellitus. Comput. Biol. Chem. 2011, 35, 293-297. 32. Harvey, A.L.; Clark, R.L.; Mackay, S.P.; Johnston, B.F. Current strategies for drug discovery through natural products. Expert Opin. Drug Discov. 2010, 5, 559-568. 33. Schuster, D.; Wolber, G. Identification of Bioactive Natural Products by Pharmacophore-Based Virtual Screening. Curr. Pharm. Des. 2010, 16, 1666-1681. 34. Tou, W.I.; Chen, C.Y.-C. In Silico Investigation of Potential Src Kinase Ligands from Traditional Chinese Medicine. PLoS One 2012, 7. 35. Li, W.L.; Zheng, H.C.; Bukuru, J.; De Kimpe, N. Natural medicines used in the traditional Chinese medical system for therapy of diabetes mellitus. J. Ethnopharmacol. 2004, 92, 1-21. 36. Herrick, T.M.; Million, R.P. Tapping the potential of fixed-dose combinations. Nat. Rev. Drug Discovery 2007, 6, 513-514. 37. Zhu, F.; Shi, Z.; Qin, C.; Tao, L.; Liu, X.; Xu, F.; Zhang, L.; Song, Y.; Liu, X.; Zhang, J.; Han, B.; Zhang, P.; Chen, Y. Therapeutic target database update 2012: a resource for facilitating target-oriented drug discovery. Nucleic Acids Res. 2012, 40, D1128-D1136. 38. Wishart, D.S.; Knox, C.; Guo, A.C.; Cheng, D.; Shrivastava, S.; Tzur, D.; Gautam, B.; Hassanali, M. DrugBank: a knowledgebase for drugs, drug actions and drug targets. Nucleic Acids Res. 2008, 36, D901-D906. 39. Kanehisa, M.; Goto, S.; Sato, Y.; Furumichi, M.; Tanabe, M. KEGG for integration and interpretation of large-scale molecular data sets. Nucleic Acids Res. 2012, 40, D109-D114. 40. Moller, D.E. New drug targets for type 2 diabetes and the metabolic syndrome. Nature 2001, 414, 821-827. 41. Berman, H.M.; Westbrook, J.; Feng, Z.; Gilliland, G.; Bhat, T.N.; Weissig, H.; Shindyalov, I.N.; Bourne, P.E. The Protein Data Bank. Nucleic Acids Res. 2000, 28, 235-242. 42. Qiao, X.B.; Hou, T.J.; Zhang, W.; Guo, S.L.; Xu, S.J. A 3D structure database of components from Chinese traditional medicinal herbs. J. Chem. Inf. Comput. Sci. 2002, 42, 481-489. 43. Shen, M.; Tian, S.; Li, Y.; Li, Q.; Xu, X.; Wang, J.; Hou, T. Drug-likeness analysis of traditional Chinese medicines: 1. property distributions of drug-like compounds, non-drug-like compounds and natural compounds from traditional Chinese medicines. J. Cheminform. 2012, 4, 1-13. 44. Tian, S.; Li, Y.; Wang, J.; Xu, X.; Xu, L.; Wang, X.; Chen, L.; Hou, T. Drug-likeness analysis of traditional Chinese medicines: 2. Characterization of scaffold architectures for drug-like compounds, non-drug-like compounds, and natural compounds from traditional Chinese medicines. J. Cheminform. 2013, 5, 1-14. 45. Tian, S.; Wang, J.; Li, Y.; Xu, X.; Hou, T. Drug-likeness Analysis of Traditional Chinese Medicines: Prediction of Drug-likeness Using Machine Learning Approaches. Mol. Pharm. 2012, 9, 2875-86. 46. Chen, C.Y.-C. TCM Database@Taiwan: The World's Largest Traditional Chinese Medicine 29
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904
Database for Drug Screening In Silico. PloS One 2011, 6. 47. Discovery Studio 3.1 Guide, Accelrys Inc., San Diego, 2012, http://www.accelrys.com. 48. Liu, T.; Lin, Y.; Wen, X.; Jorissen, R.N.; Gilson, M.K. BindingDB: a web-accessible database of experimentally determined protein-ligand binding affinities. Nucleic Acids Res. 2007, 35, D198-D201. 49. Krueger, D.M.; Evers, A. Comparison of Structure- and Ligand-Based Virtual Screening Protocols Considering Hit List Complementarity and Enrichment Factors. Chemmedchem 2010, 5, 148-158. 50. Friesner, R.A.; Murphy, R.B.; Repasky, M.P.; Frye, L.L.; Greenwood, J.R.; Halgren, T.A.; Sanschagrin, P.C.; Mainz, D.T. Extra precision glide: Docking and scoring incorporating a model of hydrophobic enclosure for protein-ligand complexes. J. Med. Chem. 2006, 49, 6177-6196. 51. Schrödinger, version 9.0, Schrödinger, LLC, New York, NY, 2009, http://www.schrodinger.com. 52. Kaminski, G.A.; Friesner, R.A.; Tirado-Rives, J.; Jorgensen, W.L. Evaluation and reparametrization of the OPLS-AA force field for proteins via comparison with accurate quantum chemical calculations on peptides. J. Phys. Chem. B 2001, 105, 6474-6487. 53. Sutter, J.; Li, J.; Maynard, A.J.; Goupil, A.; Luu, T.; Nadassy, K. New Features that Improve the Pharmacophore Tools from Accelrys. Curr. Comput. Aided Drug Des. 2011, 7, 173-180. 54. Rogers, D.; Hopfinger, A.J. Application of genetic function approximation to quantitative structure-activity-relationships and quantitative structure-property relationships. J. Chem. Inf. Comput. Sci. 1994, 34, 854-866. 55. Meslamani, J.; Li, J.; Sutter, J.; Stevens, A.; Bertrand, H.-O.; Rognan, D. Protein-Ligand-Based Pharmacophores: Generation and Utility Assessment in Computational Ligand Profiling. J. Chem. Inf. Model. 2012, 52, 943-955. 56. Chen, L.; Li, Y.; Zhao, Q.; Peng, H.; Hou, T. ADME Evaluation in Drug Discovery. 10. Predictions of P-Glycoprotein Inhibitors Using Recursive Partitioning and Naive Bayesian Classification Techniques. Mol. Pharm. 2011, 8, 889-900. 57. Wang, S.; Li, Y.; Wang, J.; Chen, L.; Zhang, L.; Yu, H.; Hou, T. ADMET Evaluation in Drug Discovery. 12. Development of Binary Classification Models for Prediction of hERG Potassium Channel Blockage. Mol. Pharm. 2012, 9, 996-1010. 58. Breitkreutz, B.J.; Stark, C.; Tyers, M. Osprey: a network visualization system. Genome Biol. 2003, 4. 59. Cheng, T.; Li, X.; Li, Y.; Liu, Z.; Wang, R. Comparative Assessment of Scoring Functions on a Diverse Test Set. J. Chem. Inf. Model. 2009, 49, 1079-1093. 60. Cross, J.B.; Thompson, D.C.; Rai, B.K.; Baber, J.C.; Fan, K.Y.; Hu, Y.; Humblet, C. Comparison of Several Molecular Docking Programs: Pose Prediction and Virtual Screening Accuracy. J. Chem. Inf. Model. 2009, 49, 1455-1474. 61. Bemis, G.W.; Murcko, M.A. The properties of known drugs .1. Molecular frameworks. J. Med. Chem. 1996, 39, 2887-2893. 62. Yang, L.; Wang, K.; Chen, J.; Jegga, A.G.; Luo, H.; Shi, L.; Wan, C.; Guo, X.; Qin, S.; He, G.; Feng, G.; He, L. Exploring Off-Targets and Off-Systems for Adverse Drug Reactions via Chemical-Protein Interactome - Clozapine-Induced Agranulocytosis as a Case Study. PLoS Comput. Biol. 2011, 7. 63. Durant, J.L.; Leland, B.A.; Henry, D.R.; Nourse, J.G. Reoptimization of MDL keys for use in drug discovery. J. Chem. Inf. Comput. Sci. 2002, 42, 1273-1280. 64. Matsuda, H.; Shimoda, H.; Morikawa, T.; Yoshikawa, M. Phytoestrogens from the roots of 30
ACS Paragon Plus Environment
Page 30 of 46
Page 31 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948
Polygonum cuspidatum (Polygonaceae): Structure-requirement of hydroxyanthraquinones for estrogenic activity. Bioorg. Med. Chem. Lett. 2001, 11, 1839-1842. 65. Lee, D.Y.; Kim, D.H.; Lee, H.J.; Lee, Y.; Ryu, K.H.; Jung, B.-I.; Song, Y.S.; Ryu, J.-H. New estrogenic compounds isolated from Broussonetia kazinoki. Bioorg. Med. Chem. Lett. 2010, 20, 3764-3767. 66. Roelens, F.; Heldring, N.; Dhooge, W.; Bengtsson, M.; Comhaire, F.; Gustafsson, J.-A.; Treuter, E.; De Keukeleire, D. Subtle side-chain modifications of the hop phytoestrogen 8-prenylnaringenin result in distinct agonist/antagonist activity profiles for estrogen receptors alpha and beta. J. Med. Chem. 2006, 49, 7357-7365. 67. Zhao, L.Q.; Brinton, R.D. Structure-based virtual screening for plant-based ER beta-selective ligands as potential preventative therapy against age-related neurodegenerative diseases. J. Med. Chem. 2005, 48, 3463-3466. 68. Lyons, M.M.; Yu, C.W.; Toma, R.B.; Cho, S.Y.; Reiboldt, W.; Lee, J.; Van Breemen, R.B. Resveratrol in raw and baked blueberries and bilberries. J. Agric. Food Chem. 2003, 51, 5867-5870. 69. Ferreira, C.V.; Justo, G.Z.; Souza, A.C.S.; Queiroz, K.C.S.; Zambuzzi, W.F.; Aoyama, H.; Peppelenbosch, M.P. Natural compounds as a source of protein tyrosine phosphatase inhibitors: Application to the rational design of small-molecule derivatives. Biochimie 2006, 88, 1859-1873. 70. Wang, F.-R.; Yang, X.-W.; Zhang, Y.; Liu, J.-X.; Yang, X.-B.; Liu, Y.; Shi, R.-B. Three new isoflavone glycosides from Tongmai granules. J. Asian Nat. Prod. Res. 2011, 13, 319-329. 71. Ha, D.T.; Tran Minh, N.; Lee, I.; Lee, Y.M.; Kim, J.S.; Jung, H.; Lee, S.; Na, M.; Bae, K. Inhibitors of Aldose Reductase and Formation of Advanced Glycation End-Products in Moutan Cortex (Paeonia suffruticosa). J. Nat. Prod. 2009, 72, 1465-1470. 72. Matsuda, H.; Morikawa, T.; Toguchida, I.; Yoshikawa, M. Structural requirements of flavonoids and related compounds for aldose reductase inhibitory activity. Chem. Pharm. Bull. 2002, 50, 788-795. 73. Srivastava, S.K.; Ansari, N.H. Prevention of sugar-induced cataractogenesis in rats by butylated hydroxytoluene. Diabetes 1988, 37, 1505-1508. 74. Wolff, S.P.; Crabbe, M.J.C. Low apparent aldose reductase-activity produced by monosaccharide autoxidation. Biochem. J. 1985, 226, 625-630. 75. Collier, A.; Wilson, R.; Bradley, H.; Thomson, J.A.; Small, M. Free-radical activity in type-2 diabetes. Diabetic Med. 1990, 7, 27-30. 76. Sato, Y.; Hotta, N.; Sakamoto, N.; Matsuoka, S.; Ohishi, N.; Yagi, K. Lipid peroxide level in plasma of diabetic-patients. Biochem. Med. 1979, 21, 104-107. 77. Kannel, W.B.; McGee, D.L. Diabetes and Cardiovascular Risk-Factors - Framingham Study. Circulation 1979, 59, 8-13. 78. Greene, D.A. Acute and chronic complications of diabetes-mellitus in older patients. Am. J. Med. 1986, 80, 39-53. 79. Mohora, M.; Virgolici, B.; Paveliu, F.; Lixandru, D.; Muscurel, C.; Greabu, M. Free radical activity in obese patients with type 2 diabetes mellitus. Rom. J. Intern. Med. 2006, 44, 69-78. 80. Nishimura, C.; Kuriyama, K. Alteration of lipid peroxide and endogenous antioxidant contents in retina of streptozotocin-induced diabetic rats - effect of vitamin-a administration. Jpn. J. Pharmacol. 1985, 37, 365-372. 81. Paller, M.S.; Hoidal, J.R.; Ferris, T.F. Oxygen free-radicals in ischemic acute-renal-failure in the rat. 31
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979
J. Clin. Invest. 1984, 74, 1156-1164. 82. Funke, I.; Melzig, M.F. Effect of different phenolic compounds on alpha-amylase activity: screening by microplate-reader based kinetic assay. Pharmazie 2005, 60, 796-797. 83. He, Z.D.; Lau, K.M.; But, P.P.H.; Jiang, R.W.; Dong, H.; Ma, S.C.; Fung, K.P.; Ye, W.C.; Sun, H.D. Antioxidative glycosides from the leaves of Ligustrum robustum. J. Nat. Prod. 2003, 66, 851-854. 84. Jang, D.S.; Lee, Y.M.; Jeong, I.H.; Kim, J.S. Constituents of the Flowers of Platycodon grandiflorum with Inhibitory Activity on Advanced Glycation End Products and Rat Lens Aldose Reductase In Vitro. Arch. Pharmacal Res. 2010, 33, 875-880. 85. Wang, N.; Yang, X.-W. Two new flavonoid glycosides from the whole herbs of Hyssopus officinalis. J. Asian Nat. Prod. Res. 2010, 12, 1044-1050. 86. Xie, H.H.; Wang, T.; Matsuda, H.; Morikawa, T.; Yoshikawa, M.; Tani, T. Bioactive constituents from Chinese natural medicines. XV.(1)) - Inhibitory effect on aldose reductase and structures of saussureosides A and B from Saussurea medusa. Chem. Pharm. Bull. 2005, 53, 1416-1422. 87. Fang, H.; Tong, W.D.; Shi, L.M.; Blair, R.; Perkins, R.; Branham, W.; Hass, B.S.; Xie, Q.; Dial, S.L.; Moland, C.L.; Sheehan, D.M. Structure-activity relationships for a large diverse set of natural, synthetic, and environmental estrogens. Chem. Res. Toxicol. 2001, 14, 280-294. 88. Dong, Y.; Shi, H.; Yang, H.; Peng, Y.; Wang, M.; Li, X. Antioxidant Phenolic Compounds from the Stems of Entada phaseoloides. Chem. Biodiv. 2012, 9, 68-79. 89. Khairullina, V.R.; Gerchikov, A.Y.; Denisova, S.B. Comparative study of the antioxidant properties of selected flavonols and flavanones. Kinet. Catal. 2010, 51, 219-224. 90. Kong, D.; Zhang, Y.; Yamori, T.; Duan, H.; Jin, M. Inhibitory Activity of Flavonoids against Class I Phosphatidylinositol 3-Kinase Isoforms. Molecules 2011, 16, 5159-5167. 91. Mueller, M.; Lukas, B.; Novak, J.; Simoncini, T.; Genazzani, A.R.; Jungbauert, A. Oregano: A Source for Peroxisome Proliferator-Activated Receptor gamma Antagonists. J. Agric. Food Chem. 2008, 56, 11621-11630. 92. Mei, R.-Q.; Lu, Q.; Hu, Y.-F.; Liu, H.-Y.; Bao, F.-K.; Zhang, Y.; Cheng, Y.-X. Three new polyyne (= polyacetylene) glucosides from the edible roots of Codonopsis cordifolioidea. Helv. Chim. Acta. 2008, 91, 90-96. 93. Ilango, K.; Chitra, V.; Kanimozhi, P.; Balaji, G. Antidiabetic, antioxidant and antibacterial activities of leaf extracts of Adhatoda zeylanica. Medic (Acanthaceae). J. Pharm. Sci. Res. 2009, 1, 67-73.
980
Legend of figures
981
Figure 1. The distributions of the docking scores of the validation datasets for (a) four
982
targets (PPAR-gamma, PPAR-alpha, DPP-4 and insulin receptor) with low p-values,
983
and (b) four targets (carbonic anhydrase 2, pyruvate dehydrogenase [lipoamide]
984
kinase isozyme 1, transthyretin and corticosteroid 11-beta-dehydrogenase, isozyme 1)
985
with high p-values.
986
Figure 2. (a) The comparison of the numbers of the pharmacophore features in Total
987
features, Feature set A and Feature set B, and (b) The comparison of the 32
ACS Paragon Plus Environment
Page 32 of 46
Page 33 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
988
discrimination capacity indicated by the AUC values of the pharmacophore models
989
based on Feature set A and Feature set B.
990
Figure 3. The comparison of the discrimination capability indicated by the AUC
991
values for the naïve Bayesian classifiers based on fit values, docking scores and the
992
combination of fit values and docking scores.
993
Figure 4. The distributions of Bayesian scores of the inhibitors and non-inhibitors and
994
the AUC value of the naïve Bayesian classifiers based on (a) docking scores, (b) fit
995
values and (c) the combination of docking scores and fit values for insulin-like growth
996
factor 1 receptor (IGF-1R).
997
Figure 5. (a) Eight potential inhibitors predicted by the Bayesian classifier are
998
identical to the known inhibitors of estrogen receptor, and (b) three potential
999
inhibitors predicted by the Bayesian classifier are similar to the known inhibitors of
1000
estrogen receptor.
1001
Figure 6. Eighteen potential inhibitors from the TCM formulae predicted by the
1002
Bayesian classifier are similar to the known inhibitor, myricetin, of insulin receptor.
1003
Figure 7. The number of the potential inhibitors ranked by the Bayesian scores for
1004
establishing the compounds-targets network.
1005
Figure 8. The distribution of the number of the potential multi-target inhibitors.
1006
Figure 9. (a) The compound-target network for all potential inhibitors, (b) for those
1007
against no less than targets, (c) for those against no less than three targets, and (d) for
1008
those against no less than four targets.
1009
Figure 10. The molecule, Pueroside A, was predicted to an inhibitor of three targets
1010
(carbonic anhydrase, PPAR delta and PPAR gamma). (a) schematic representation of
1011
the interaction between Pueroside A and three targets, and (b) the known inhibitors
1012
that have the maximum similarity to Pueroside A based on the MDLPulicKeys
1013
fingerprints for three targets, respectively.
1014
Figure 11. The molecules, Epicatechin and (2RS)-2-(4-O-β-D-glucopyranosylphenyl)
1015
propionyl β-D-glucopyranoside were predicted to be the inhibitors of two targets. (a)
1016
schematic representation of the interactions between Epicatechin and carbonic 33
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
2
and
estrogen
Page 34 of 46
1017
anhydrase
receptor,
and
those
between
1018
(2RS)-2-(4-O-β-D-glucopyranosylphenyl)
propionyl
β-D-glucopyranoside
1019
carbonic anhydrase 1 and carbonic anhydrase 2 targets, and (b) the known inhibitors
1020
of carbonic anhydrase 2 and estrogen receptor that have the maximum similarity to
1021
Epicatechin; the known inhibitors of carbonic anhydrase 1 and carbonic anhydrase
1022
that have the maximum similarity to (2RS)-2-(4-O-β-D-glucopyranosylphenyl)
1023
propionyl β-D-glucopyranoside based on the MDLPublicKeys fingerprints.
and
1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038
Table 1. The performances of the naïve Bayesian classifiers by using fit values,
1039
docking scores and the combination of fit values and docking scores for 24 targets No
Target name
Ninh
Nnon
AUC Fit value
Score
Combined
1
Carbonic anhydrase 1
52
1040
0.908
0.841
0.942
2
Carbonic anhydrase 2
91
1820
0.803
0.500
0.867
3
Growth factor receptor-bound protein 2
16
320
0.687
0.597
0.741
4
Peroxisome proliferator-activated receptor gamma
500
10000
0.532
0.795
0.837
5
Glycogen synthase kinase-3 β
69
1380
0.833
0.916
0.971
6
Aldose reductase
145
2900
0.717
0.836
0.892
7
Pancreatic α-amylase
22
440
0.795
0.743
0.865
34
ACS Paragon Plus Environment
Page 35 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
8
Peroxisome proliferator-activated receptor delta
224
4480
0.624
0.781
0.830
59
1180
0.641
0.573
0.672
0.614
0.741
0.814
(PPAR-δ) 9
Pyruvate dehydrogenase [lipoamide] kinase isozyme 1
10
Insulin receptor
370
7400
11
Insulin-like growth factor 1 receptor
177
2540
0.788
0.729
0.904
12
mRNA of Protein-tyrosine phosphatase,
34
680
0.604
0.759
0.596
263
5260
0.533
0.809
0.847
0.856
0.896
non-receptor type 1 (PTP1B) 13
Peroxisome proliferator-activated receptor α (PPAR-α)
14
Glucokinase (Hexokinase D)
314
6280
0.624
15
Protein Kinase C, alpha type
500
10000
0.572
0.600
0.683
16
Mitogen-activated protein kinase 10
500
10000
0.573
0.655
0.720
17
Corticosteroid 11-beta-dehydrogenase, isozyme 1
500
10000
0.526
0.578
0.604
(11-Beta hydroxysteroid dehydrogenase 1) 18
Mitogen-activated protein kinase 8
500
10000
0.551
0.601
0.661
19
Mitogen-activated protein kinase 1
130
2600
0.562
0.668
0.722
20
Receptor protein-tyrosine kinase erbB-2
13
260
0.669
0.937
0.957
21
Estrogen receptor
500
10000
0.676
0.897
0.940
22
Glucocorticoid receptor
115
2300
0.628
0.736
0.788
23
Fructose-1,6-bisphosphatase 1 (FBPase)
89
1780
0.740
0.821
0.856
24
Phosphatidylinositol-4,5-bisphosphate 3-kinase
16
320
0.840
0.898
0.934
catalytic subunit, delta isoform (Phosphoinositide 3-kinase delta)
1040 1041 1042
35
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
1043 1044
Figure 1
1045
36
ACS Paragon Plus Environment
Page 36 of 46
Page 37 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
1046 1047
Figure 2
1048
37
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
1049 1050
Figure 3
1051
38
ACS Paragon Plus Environment
Page 38 of 46
Page 39 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
1052 1053
Figure 4
1054
39
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
1055 1056
Figure 5
40
ACS Paragon Plus Environment
Page 40 of 46
Page 41 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
1057 1058
Figure 6
1059
41
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
1060 1061
Figure 7
1062 1063 1064
42
ACS Paragon Plus Environment
Page 42 of 46
Page 43 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
1065 1066
Figure 8
1067 1068 1069 1070 1071
43
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
1072 1073
Figure 9
1074 1075 1076 1077 1078 1079
44
ACS Paragon Plus Environment
Page 44 of 46
Page 45 of 46
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
Journal of Chemical Information and Modeling
1080 1081
Figure 10
45
ACS Paragon Plus Environment
Journal of Chemical Information and Modeling
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60
1082 1083
Figure 11
46
ACS Paragon Plus Environment
Page 46 of 46