-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathModel.html
340 lines (303 loc) · 28 KB
/
Model.html
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
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
<!DOCTYPE html>
<html lang="zxx" class="no-js">
<head>
<!-- Mobile Specific Meta -->
<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
<!-- Favicon-->
<link rel="shortcut icon" href="img/elements/fav.png">
<!-- Author Meta -->
<meta name="author" content="colorlib">
<!-- Meta Description -->
<meta name="description" content="">
<!-- Meta Keyword -->
<meta name="keywords" content="">
<!-- meta character set -->
<meta charset="UTF-8">
<!-- Site Title -->
<title>Seo</title>
<link href="https://fonts.googleapis.com/css?family=Poppins:100,200,400,300,500,600,700" rel="stylesheet">
<!--
CSS
============================================= -->
<link rel="stylesheet" href="css/linearicons.css">
<link rel="stylesheet" href="css/owl.carousel.css">
<link rel="stylesheet" href="css/font-awesome.min.css">
<link rel="stylesheet" href="css/nice-select.css">
<link rel="stylesheet" href="css/magnific-popup.css">
<link rel="stylesheet" href="css/bootstrap.css">
<link rel="stylesheet" href="css/main.css">
</head>
<body>
<!-- 头部列表 -->
<header class="default-header">
<nav class="navbar navbar-expand-lg navbar-light">
<div class="container">
<a class="navbar-brand" href="index.html">
<img src="img/logo.png" width = 100px alt="">
</a>
<button class="navbar-toggler" type="button" data-toggle="collapse" data-target="#navbarSupportedContent" aria-controls="navbarSupportedContent" aria-expanded="false" aria-label="Toggle navigation">
<span class="navbar-toggler-icon"></span>
</button>
<!-- 列表 -->
<div class="collapse navbar-collapse justify-content-end align-items-center" id="navbarSupportedContent">
<ul class="navbar-nav">
<li class="Home_page" href="index.html">
<a class="Home_page_text" href="index.html">
Home
</a>
</li>
<li class="dropdown">
<a class="dropdown-toggle" href="#" id="navbardrop" data-toggle="dropdown">
Project
</a>
<div class="dropdown-menu">
<a class="Background" href="Background.html">Background</a>
<a class="Model" href="Model.html">Model</a>
<a class="Supplementary" href="Supplementary.html">Supplementary</a>
<a class="Report" href="Report.html">Report</a>
</div>
</li>
<li class="dropdown">
<a class="dropdown-toggle" href="#" id="navbardrop" data-toggle="dropdown">
Documentation
</a>
<div class="dropdown-menu">
<a class="dropdown-item" href="Protocols.html">Protocols</a>
</div>
<li class="dropdown">
<a class="dropdown-toggle" href="#" id="navbardrop" data-toggle="dropdown">
Team
</a>
<div class="dropdown-menu">
<a class="dropdown-item" href="member.html">Members</a>
</div>
</li>
</li>
</ul>
</div>
</div>
</nav>
</header>
<!-- 头部列表结束 -->
<!-- Start banner Area -->
<section class="generic-banner relative">
<div class="container">
<div class="row height align-items-center justify-content-center">
<div class="col-lg-10">
<div class="generic-banner-content">
<h2 class="text-white">Model</h2>
</div>
</div>
</div>
</div>
</section>
<!-- End banner Area -->
<!-- Start Sample Area -->
<section class="sample-text-area">
<div class="container">
<h3 class="text-heading">Introduction</h3>
<p class="sample-text" style="font-size:25px;color:black">
Though the concept of evolution is typically about organisms, the evolutionary process started long before the advent of life. Simple chemicals formed on the primitive earth, then the competition begun. Driven by thermodynamic and kinetic factors, some chemical structures were conserved and assembled into more complex systems, including efficient catalyze systems, self-replicating information storage units and different kinds of structural compositions. Those successfully withstood the tests of time turned into today’s colorful world of life.
<br>
<br>
While biologists immerse themselves in the wonder of biomacromolecules, which reflects changes of history and contains hopes of future, chemists stay focus on the behaviors of simple molecules, as the evolution on this size scale never stops its pace. With the help of molecular simulation technology, chemists can explore wider chemical space more in a cheaper and quicker way. Thus, it is more possible for us to get benefit from irrational design, which is the strategy taken by nature evolution, within a much shorter time. In this perspective, discovery and development of functional small molecules in modern chemistry researches can be seen as analogs of nature evolution, with a artificially prescribed direction for selection, just like the directed evolution engineering of enzymes.
<br>
<br>
Today, a typical process of chemical drug discovery includes multiple rounds of modification and selection, which consumes a large amount of time and resources. Like other brunches of chemistry, the process is accelerating by In Silico methods. Herein, we attempt to make a de novo molecule generating machine based on structure of target biomacromolecule, in order to specific lead compound. Geometry features of molecules are used to analogize the genetic information in conventional concept of directed evolutionary engineering. By defining the binding site of target macromolecule, we set a environment for diverse molecule to generate. With some other parameters as initial selection stress, molecular dynamic simulations are conveyed to find out the fittest.
<br>
<br>
In this project, we choose DNA polymerase theta (Pol θ) as our target macromolecule. Pol θ is a critical enzyme in microhomology mediated end-joining(MMEJ), which is an alternative pathway of DNA double strand breaking repairment. It is up-regulated in several cancers, causing the resistance of radiotherapy and chemotherapy. Limited by the microhomology mediated mechanism, MMEJ does not completely repair the damaged genome, which means unpredictable mutations will be introduced during end-joining. Besides, cancer cells are actively proliferating, giving them probabilities to evolve into more deadly existence to patient. Thus, it has been a promising target for antitumor drug development, especially for cell lines with homologous reorganization (HR) deficient, such as BRCA mutation. Given that, we attempt to discover a small molecule specifically binding to pol θ and inhibit its function, in order to restrain the disordered evolutionary process of tumor cells and save human lives.
<br>
<br>
Our model, MolEV, is still under testing and modifying. We hope it can be an inspiration of lead compound discovery in the near future, grow from a virtual evolution model to a generator of in vivo effective molecules.
</p>
</div>
</section>
<!-- End Sample Area -->
<!-- Start Sample Area -->
<section class="sample-text-area">
<div class="container">
<h3 class="text-heading">Part1 : Protein optimization and ligand preparation</h3>
<p class="sample-text" style="font-size:25px;color:black">
Before extracting the structural information features of protein and inhibitors, we needed to optimize the molecular structure, so as to improve the reliability of the extracted molecular 3D structural information. Otherwise, errors and inaccurate results will be reported in the subsequent simulation process. For the structural optimization of polymerase theta enzyme, we found that it lacks many loops, so it is important to repair the loop region of the protein. After sequence observation, we found that most of the missing loops are within the range of 5-10. Therefore, after the rough structure containing loops is repaired by the cyclic coordinate center (CCD) [2], we further rationally optimize the generated loop structure through the closed-loop algorithm KIC, Set a loop with four node amino acids, and then cut it from the middle node to form a cut. Then, use the fastelax module in rosetta [3] to eliminate the unreasonable conformation in the protein crystal structure and compare the energy changes between multiple structures. Among them, we search for the optimal conformation at the local energy barrier of a given three-dimensional structure through five iterations (single protein) of amino acid side chain rearrangement and energy minimization calculation, The side chain can enter a new energy minimum point. After that, the <i>clique </i> algorithm is used to cluster the 600 protein structures screened, so as to screen out a more reasonable conformation, after that, foldx was used to conduct annealing optimization and self mutation optimization for the obtained structure in the environment with ionic strength of 0.05 and PH=7. Finally, pdb4amber tool in amber22 was used with reasonable Hydrogens, so far, the protein structure optimization was completely completed, as shown in the following Fig.1:
<br>
<img src="img/model/fig1.png" style="text-align:center ;margin-left: 20%"alt="picture">
<p style="text-align:center;font-size:25px;color:black;margin-top: 0px;" class="dd"><b>Figure1.</b>the final optimized DNA polymerase theta structure</p>
<br>
<p class="sample-text" style="font-size:25px;color:black">
After optimizing the protein, we need to prepare the ligand file for the subsequent extraction of structural information. Because we have generated about 300 small molecules with high druglike properties for the POLQ target according to our model, due to the computational power, we used the <i>GFN2-xtb</i>[4] algorithm to optimize these structures separately at the beginning. After hydrogenated docking, The virtual screened molecules are then optimized and calculated using the
<i>B3LYP/B97-3c</i>[5] in combination with the <i>def2-TZVP</i> basis set. At the same time, the <i>RESP2</i> [6] charge of the molecules is calculated using Multiwfn [11], which further replace the <i>am1-bcc</i> charge used in molecular dynamics simulation. Because the RESP2 charge can greatly reduce the following two problems in using <i>MK/CHELPG</i> charges, (1) the results are highly conformational, (2) Inaccurate fitting of embedded atomic charge.
</p>
<br>
<img src="img/model/fig2.png" style="text-align:center;margin-left: 20% "alt="picture">
<p style="text-align:center;font-size:25px;color:black;margin-top: 0px;" class="dd"><b>Figure2.</b>The optimized molecule (example) calculated by ORCA with
B3LYP/def2-TZVP level</p>
<br>
<p class="sample-text" style="font-size:25px;color:black">
The structures in this model have all been optimized.
After checking the protein structure according to the tleap module in amber22 [7][8] ,
it is found that there is no problem, and no virtual frequency (vibration)
is found in the output files of the xtb and ORCA [9] .
</p>
<h3 class="text-heading">Part2: Construction and principle of the Model (Presite)</h3>
<p class="sample-text" style="font-size:25px;color:black">
Before building the molecular generative model, we built the neural network
prestige with the CNN as the main body of predicted protein binding pockets, and its main architecture is described below.
A four-channel grid protein descriptor is constructed based on ligand site (ligsite), L-J potential,
and Coulomb force, which can be used for protein binding site prediction. By using 12X12X12 sampling to classify and cluster the blocks, the binding sites of the proteins are finally determined. At the same time, in order to obtain protein feature, the pdbqt files of the protein should be included in the dataset, and we suggest that you can get these files through openbabel or ADFRsuites. The detailed processing is shown in the figure3 below and our training dataset is based on the <i>scPDB </i> [10] database.
We define a geometric vector MLP as a GVP block followed by a GVL block, denoted as Gm
The frontier predictor takes the features of atom i as input and utilizes one MLP layer to predict the probability of being a frontier pf as follows:
</p>
<img src="img/model/eq1.png" style="text-align:center ; font-size:100px; margin-left: 40%;"alt="picture">
<br>
<br>
<h3 class="text-heading">Part3: Construction and principle of the model (MolEV)</h3>
<br>
<img src="img/model/fig3.png" style="text-align:center;margin-left: 20% "alt="picture">
<p style="text-align:center;font-size:25px;color:black;margin-top: 0px;" class="dd"><b>Figure3.</b>Framework about predicting the protein pocke</p>
<br>
<br>
<img src="img/model/fig4.png" style="text-align:center;margin-left: 5% "alt="picture">
<p style="text-align:center;font-size:25px;color:black;margin-top: 0px;" class="dd"><b>Figure4.</b> Main part and details of our MolEV Model</p>
<br>
<p class="sample-text" style="font-size:25px;color:black">
In the generation of the molecule, we divided it mainly into two parts called molgraph and prograph, respectively,
in Molgraph, we extracted the dihedral angle of the molecule γ With two geometric angles α,
β we set four atoms of the molecule into one set α,β,γ arrays and mapped to RGB triple primaries:
</p>
<img src="img/model/eq2.png" style="text-align:center ; font-size:100px; margin-left: 35%;"alt="picture">
<br>
<p class="sample-text" style="font-size:25px;color:black">
After this, the (α,β,γ) array is transformed into a picture and the CVAE\_GAN neural network model is constructed for data training to obtain a new picture, where C represents the information about the geometry of the protein, is a restriction for molecular generation, after which the picture is transposed to obtain a suitable fit for a specific protein structure or target pair small molecule inhibitors with ligands.
In order to make our extracted molecular arrays of some physical significance, we propose Bessel functions applicable to biological molecular geometry systems as follows:
</p>
<img src="img/model/eq3.png" style="text-align:center ; font-size:100px; margin-left: 35%;"alt="picture">
<br>
<p class="sample-text" style="font-size:25px;color:black">
Where jl() is the spherical Bessel function of order l, C represents cutoff. In this way, the triple (α,β,γ) can represent the relative position of any atom in a 3D molecular graph through a spherical message passing network, and a more accurate and physically meaningful representation can be generated by combining the positional information in the spherical coordinates, which avoids the problem of missing physical meaning of most molecular descriptors generated through machine learning. Our main model MolEV details are shown in Figure4.
For the part of Prograph, we used this year's improved framework dmasif for feature extraction from the protein surface and input the extracted geometry information into our model as a restrictive condition .
Based on the eMolFrag repository, we develop a classical way to recombine the previously cleaved groups to generate new small molecules from mol2 structure format and make some fixes for the previous bug in eMolFrag to make it work more properly.
</p>
<h3 class="text-heading">Part4: Model training</h3>
<p class="sample-text" style="font-size:25px;color:black">
In the training stage, we randomly mask atoms of molecules and train the model to recover the masked atoms. Specifically, for each pocket-ligand pair, we sample a mask ratio from the uniform distribution U[0, 1] and mask corresponding number of molecular atoms. The remaining molecular atoms that have valence bonds to the masked atoms are
defined as frontiers. Then the position predictor and the element-and-bond predictor try to recover the masked atoms that have valence bonds to the frontiers by predicting their
positions towards corresponding frontiers, the element types and the bonds with remaining molecular atoms. If all molecular atoms are masked, the frontiers are defined as protein atoms that have masked atoms within 4A˚ and the masked atoms around the frontiers are to be recovered. For the element type prediction, we add one more element type representing Nothing at the query position. During the training process, we sample not only the positions of masked atoms for element type predictions but also negative positions from the ambient space and assign their labels as Nothing.
The loss of the frontier prediction, ψf, is the binary cross entropy loss of predicted frontiers. The loss of the position predictor, ψp, is the negative log likelihood of the masked atom positions. For the element type and bond type prediction, we used cross entropy losses for the classification, denoted as ψe and ψb respectively. The overall loss function is the summation of the above four loss functions:
</p>
<img src="img/model/eq4.png" style="text-align:center ; font-size:100px; margin-left: 35%;"alt="picture">
<br>
<h3 class="text-heading">Molcular Dynamic Simulation (MMGBSA And FEP)</h3>
<p class="sample-text" style="font-size:25px;color:black">
After a certain amount of sampling, we first use ledock\_ Omega [12] to obtain a series of molecular conformations through mass molecular docking and related property evaluation, including drug like properties, vdw, hbond quantity, etc. Some results are preseneted as Figure5 follows:
</p>
<br>
<img src="img/model/fig9.png" style="text-align:center ;margin-left: 20%"alt="picture">
<p style="text-align:center;font-size:25px;color:black;margin-top: 0px;" class="dd"><b>Figure5.</b> General docking score range which is screened by ledock</p>
<br>
<p class="sample-text" style="font-size:25px;color:black">
We subsequently conducted the hydrated docking after the previous round of primary screening, that is, the incorporation of explicit water model in the process of docking, improved the precision of docking, because this docking relies on reasonable protein conformation with correct molecular structure, reasonable molecular optimization steps before is highly necessary, after passing the less selections, We selected three molecules with extremely close molecular structures for the 5ns FEP calculations (constrained by computational conditions) and then verified the molecular activities (obtaining the relative free energy between ligand molecules) and performed MMGBSA binding free energy calculations after finally picking the optimal result molecule.
The results showed that the relative free energies of molecular conversion of ART558 and ART615 into ART4215 molecules were -0.2kal/mol and -0.4 kcal/mol, respectively, well demonstrating the stronger activity of ART4215 relative to the other two molecules, and according to the display of the following three graphs, it can be concluded that our simulation reached the level of convergence, After that we subjected the complex of art4215 with POLQ enzyme for up to 200 ns simulation and took the last two thousand frames in the trajectory to calculate the mmgbsa binding free energy, which resulted -27.3786 kcal/mol. The relative RMSD curves generated during the above simulations are provided in thesupplementary files.
</p>
<br>
<img src="img/model/fig7.png" style="text-align:center ;margin-left: 30%"alt="picture">
<p style="text-align:center;font-size:25px;color:black;margin-top: 0px;" class="dd"><b>Figure6.</b> Overlap matrix of free energy perturbation calculation</p>
<br>
<br>
<img src="img/model/fig6.png" style="text-align:center ;margin-left: 30%"alt="picture">
<p style="text-align:center;font-size:25px;color:black;margin-top: 0px;" class="dd"><b>Figure7.</b> Convergence degree of free energy perturbation calculation</p>
<br>
<h3 class="text-heading">Model Results and Evalution</h3>
<p class="sample-text" style="font-size:25px;color:black">
Our model is to generate molecules in the pocket after a good protein pocket is determined, so the process of molecular docking can be omitted to some extent, the most powerful demonstration is shown in Fig. 7. As shown in Figure7, using the hydrated docking of vina, the obtained conformations and models generated by the molecules do not differ much with RMSD ranging from 0.05 to 0.15, but the verification of scoring function is missing, experimentally, For reasons of time there is no way to implement relevant synthetic experiments and in vitro drug trials and there has not been a certain comparison and validation with other VAE models, such as the JT-VAE and so on.
</p>
<img src="img/model/fig8.png" style="text-align:center ;margin-left: 30%"alt="picture">
<p style="text-align:center;font-size:25px;color:black;margin-top: 0px;" class="dd"><b>Figure8.</b> RMSD deviation between gernerated molecule in the pocket and hydrated docking by vina</p>
<br>
<div class="section-top-border">
<h3 class="mb-30"><b>Table</b>: Molecules generated by our models and their properties</h3>
<div class="progress-table-wrap">
<div class="progress-table">
<div class="table-head">
<div class="country">Molecular</div>
<div class="visit">Score</div>
<div class="visit">VDW</div>
<div class="visit">Hbond</div>
</div>
<div class="table-row">
<div class="country"> <img src="img/mol/mol1.jpg" alt="flag"></div>
<div class="visit">-5.676</div>
<div class="visit">-82.44</div>
<div class="visit">-29.54</div>
</div>
<div class="table-row">
<div class="country"> <img src="img/mol/mol2.jpg" alt="flag"></div>
<div class="visit">-8.408</div>
<div class="visit">-32.16</div>
<div class="visit">-33.73</div>
</div>
<div class="table-row">
<div class="country"> <img src="img/mol/mol3.jpg" alt="flag"></div>
<div class="visit">-8.779</div>
<div class="visit">-71.48</div>
<div class="visit">-32.56</div>
</div>
<div class="table-row">
<div class="country"> <img src="img/mol/mol4.jpg" alt="flag"></div>
<div class="visit">-9.923</div>
<div class="visit">-38.6</div>
<div class="visit">-39.56</div>
</div>
<div class="table-row">
<div class="country"> <img src="img/mol/mol5.jpg" alt="flag"></div>
<div class="visit">-8.091</div>
<div class="visit">-41.36</div>
<div class="visit">-20.47</div>
</div>
<div class="table-row">
<div class="country"> <img src="img/mol/mol6.jpg" alt="flag"></div>
<div class="visit">-8.727</div>
<div class="visit">-36.76</div>
<div class="visit">-22.16</div>
</div>
<div class="table-row">
<div class="country"> <img src="img/mol/mol7.jpg" alt="flag"></div>
<div class="visit">-8.311</div>
<div class="visit">-40.21</div>
<div class="visit">-15.77</div>
</div>
</div>
</div>
</div>
<h3 class="text-heading">Conclusion</h3>
<p class="sample-text" style="font-size:25px;color:black">
In general, we calculated the druglike fraction and novelty of the 400 molecules generated by the model, and found that the molecules with excellent druglike properties generated by MolEV often have good novelty. We introduced geometric structure features in the model construction, which also improved the performance of the model. At the same time, when we use metalloenzymes to generate molecules, we often cannot get good structures, The resulting conformation tends to be biased towards the generation of phosphate ions, or even the generation of unstable sulfone hydrolysates. Therefore, in the presence of metal ions, our model is not applicable, and there are significant defects, which also need to be improved. Furthermore, our training set does not use PDBbind, because we found that PDBbind dataset is used in the process of molecular generation, The conformation of the molecule is strongly dependent on the structure of the ligand molecule contained in PDBbind itself. The molecules generated are relatively similar, and there is still a big gap with our expected goal, namely novelty. The following table shows whether some of the molecules (inhibitors) generated for the POLQ target can really play the role of inhibitors, which needs to be verified by synthesis and activity in subsequent experiments.
</p>
<h3 class="text-heading">Reference</h3>
<p class="sample-text" style="font-size:25px;color:black">
[1] Newman JA, Cooper CDO, Aitkenhead H, Gileadi O. Structure of the Helicase Domain of DNA Polymerase Theta Reveals a Possible Role in the Microhomology-Mediated End-Joining Pathway. Structure. 2015 Dec 1;23(12):2319-2330. doi: 10.1016/j.str.2015.10.014. PMID: 26636256; PMCID: PMC4671958.<br>
[2] Kaufmann, K.W., et al., Practically useful: what the Rosetta protein modeling suite can do for you. Biochemistry, 2010. 49(14): p. 2987-98.<br>
[3] Leaver-Fay, A., et al., ROSETTA3: an object-oriented software suite for the simulation and design of macromolecules. Methods in enzymology, 2011. 487: p. 545.<br>
[4] GFN2-xTB: Bannwarth, C.; Ehlert, S.; Grimme, S. GFN2-xTB—An Accurate and Broadly Parametrized Self-Consistent Tight-Binding Quantum Chemical Method with Multipole Electrostatics and Density-Dependent Dispersion Contributions J. Chem. Theory Comput. 2019, 15 (3), 1652–1671 DOI: 10.1021/acs.jctc.8b01176<br>
[5] Tirado-Rives J, Jorgensen WL. Performance of B3LYP Density Functional Methods for a Large Set of Organic Molecules. J Chem Theory Comput. 2008 Feb;4(2):297-306. doi: 10.1021/ct700248k. PMID: 26620661.<br>
[6] LU Tian, CHEN Fei-Wu. Comparison of Computational Methods for Atomic Charges[J]. Acta Phys. -Chim. Sin., 2012, 28(01): 1-18.<br>
[7] D.A. Case, T.E. Cheatham, III, T. Darden, H. Gohlke, R. Luo, K.M. Merz, Jr., A. Onufriev, C. Simmerling, B. Wang and R. Woods. The Amber biomolecular simulation programs. J. Computat. Chem. 26, 1668-1688 (2005).<br>
[8] R. Salomon-Ferrer, D.A. Case, R.C. Walker. An overview of the Amber biomolecular simulation package. WIREs Comput. Mol. Sci. 3, 198-210 (2013).<br>
[9] Neese, F. (2012) The ORCA program system, Wiley Interdiscip. Rev.: Comput. Mol. Sci., 2, 73–78.Neese, F. (2017) Software update: the ORCA program system, version 4.0, Wiley Interdiscip. Rev.:Comput. Mol. Sci., 8, e1327.<br>
[10] Jérémy Desaphy, Guillaume Bret, Didier Rognan, Esther Kellenberger, sc-PDB: a 3D-database of ligandable binding sites—10 years on, Nucleic Acids Research, Volume 43, Issue D1, 28 January 2015, Pages D399–D404, https://doi.org/10.1093/nar/gku928<br>
[11] Tian Lu, Feiwu Chen, J. Comput. Chem., 33, 580-592 (2012).<br>
[12] Phys. Chem. Chem. Phys., 2016,18, 12964-12975</p>
</div>
</section>
<!-- End Sample Area -->
<script src="js/vendor/jquery-2.2.4.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/popper.js/1.11.0/umd/popper.min.js" integrity="sha384-b/U6ypiBEHpOf/4+1nzFpr53nxSS+GLCkfwBdFNTxtclqqenISfwAzpKaMNFNmj4" crossorigin="anonymous"></script>
<script src="js/vendor/bootstrap.min.js"></script>
<script src="js/owl.carousel.min.js"></script>
<script src="js/jquery.sticky.js"></script>
<script src="js/jquery.nice-select.min.js"></script>
<script src="js/parallax.min.js"></script>
<script src="js/jquery.magnific-popup.min.js"></script>
<script src="js/main.js"></script>
</body>
</html>