Discovery of novel indoleaminopyrimidine NIK inhibitors based on molecular docking-based support vector regression (SVR) model

Article abstract of DOI:10.1016/j.cplett.2019.01.031

A set of NF-κB-inducing kinase (NIK) inhibitors was used to develop a molecular docking-based QSAR model by using nonlinear regression method. The accuracy of the QSAR model was remarkably improved by integrating the docking scores and key interaction pro

Full text of DOI:10.1016/j.cplett.2019.01.031

Accepted Manuscript
Research paper
Discovery of Novel Indoleaminopyrimidine NIK Inhibitors based on Molecular
Docking-based Support Vector Regression (SVR) Model
Qing Ye, Qiu Li, Anhui Gao, Huazhou Ying, Gang Cheng, Jing Chen, Jinxin
Che, Jia Li, Xiaowu Dong, Yubo Zhou
PII:
S0009-2614(19)30063-6
DOI:
https://doi.org/10.1016/j.cplett.2019.01.031
CPLETT 36213
Reference:
To appear in:
Chemical Physics Letters
Received Date:
Revised Date:
Accepted Date:
19 October 2018
11 January 2019
14 January 2019
Please cite this article as: Q. Ye, Q. Li, A. Gao, H. Ying, G. Cheng, J. Chen, J. Che, J. Li, X. Dong, Y. Zhou,
Discovery of Novel Indoleaminopyrimidine NIK Inhibitors based on Molecular Docking-based Support Vector
Regression (SVR) Model, Chemical Physics Letters (2019), doi: https://doi.org/10.1016/j.cplett.2019.01.031
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Discovery of Novel Indoleaminopyrimidine NIK Inhibitors based on
Molecular Docking-based Support Vector Regression (SVR) Model
b, +
a, b, +
c
a
d
d
a
c
*, a
*, c
Qing Ye, Qiu Li,
Anhui Gao, Huazhou Ying, Gang Cheng, Jing Chen, Jinxin Che, Jia Li, Xiaowu Dong Yubo Zhou
a
Zhejiang Province Key Laboratory of Anti-Cancer Drug Research, College of Pharmaceutical Sciences, Zhejiang University Hangzhou,
310058, P. R. China Email: dongxw@zju.edu.cn
b
State Key Laboratory Breeding Base of Green Chemistry-Synthesis Technology, Zhejiang University of Technology, Hangzhou 310032,
P. R. China
c
National Center for Drug Screening, State Key Laboratory of Drug Research, Shanghai Institute of Materia Medica Chinese Academy
of Sciences, Shanghai, 201203, P.R. China Email: ybzhou@simm.ac.cn.
d
College of Pharmaceutical Science Zhejiang Chinese Medical University, Hangzhou, 311402, P.R. China
+
These authors contributed equally.
Abstract: A set of NF-κB-inducing kinase (NIK) inhibitors was used to develop a molecular docking-based QSAR
model by using nonlinear regression method. The accuracy of the QSAR model was remarkably improved by
integrating the docking scores and key interaction profiles. Two indole-aminopyrimidine derivatives 32a and 32b
predicted as NIK inhibitors were synthesized and biologically evaluated. The significant correlationship between
experimental data and MD-SVR model-predicted results were observed. The binding mode of 32a and 32b with
NIK were further investigated by dynamic simulations. Compound 32b was proposed as a promising lead for the
findings of highly potent inhibitors.

Introduction
The nuclear factor (NF)-κB transcription complexes in malignant cells are critical mediators of inflammatory and
1
immune responses. NF-κB can be activated by a wide range of extracellular stimuli, including pro-inflammatory
2
cytokines, mitogens, DNA-damaging agents, and microbial agents, acting via two major pathways: canonical (or
3
-5
classical) and non-canonical (or alternative). The canonical NF-kB pathway is activated downstream of multiple
6
pro-inflammatory receptors, such as T cell receptor, B cell receptor, Toll-like receptor, and TNF receptor. The
non-canonical NF-kB pathway is activated downstream of a subset of the TNF super family of receptors, including
7
,8
BR3, CD40, TWEAK, LT-bR, and RANK, which primarily regulates lymphoid organogenesis, B cell proliferation and
survival. Dysregulation of both NF-kB pathways is remarkably related with autoimmune disorders and hematologic
9
-14
malignancies. In particular, the inappropriate accumulation of NIK, which is a serine/threonine kinase, also
1
5
known as MAPK kinase 14 (MAP3K14), is significantly associated with the development of multiple myeloma and
1
6
pancreatic cancer. Therefore, intervention in the non-canonical pathway can be of clinical benefit in treating
cancer and autoimmune diseases. The development of specific small-molecule inhibitors for NIK is a promising
approach, however, so far, relatively few inhibitors have been reported, involving with limited structural skeletons
1
7
18
19
such as pyrazolo isoquinoline,
pyrazolo[4,3-c] isoquinoline,
4H-isoquinoline-1,3-dione
and
1
2,7-naphthydrine-1,3,6,8-tetrone.
2
0, 21
Virtual screening-based molecular docking is commonly used during the identification of lead compounds.
However, all of these docking models cannot sort or select small molecules by calculating binding affinities. Thus,
a more accurate prediction of the binding affinity in both virtual screening and structural optimization is needed.
In this field, Deng et al. were the pioneer in reporting the integration of Kernel partial least squares (K-PLS)
2
2
method to improve the accuracy of the scoring function of molecular docking. Recently, there is an increasing
effort in the development of machine-learning scoring function of molecular docking in a wide range of drug
2
3
design, utilizing a variety of machine-learning techniques such as artificial neural network (ANN) and support
2
4
vector machine (SVM). In fact, we also have demonstrated the application of SVR regression algorithm to
2
5-27
improve the prediction for binding affinity, providing a reliable and accurate tool in rational drug design.
To the
best of our knowledge, the development of machine-learning scoring function of molecular docking model for NIK
inhibitors has not been reported yet. Therefore, to meet the imperious demands of discovering novel NIK
inhibitors, the build of a robust and accurate machine-learning molecular docking model would be of great
interest. Here, we report a support vector regression (MD-SVR) model established based on molecular docking,

key interaction profiles and molecular descriptors for the discovery of new NIK inhibitors. Furthermore, this QSAR
model, for its prediction ability, is also successfully applied in rational design of new NIK inhibitors (Scheme 1),
which is significantly preferable than the molecular docking alone.
Scheme 1. The evolution of the establishment of MD-SVR model and its application in development of novel
indoleaminopyrimidine derivatives as NIK inhibitors.
Results and discussion
The development of MD-SVR model
2
8
A total of 25 molecules (Figure 1) with different NIK inhibitory efficiency were selected from recent literatures,
and randomly divided into training set (20 compounds) and test set (5 compounds). At first, all of NIK inhibitors
1-25 were docked into ATP-binding pocket of NIK using LigandFit program embedded in Discovery Studio 2.5. A set
of scoring functions (i.e. Dock_score, LigScore1, LigScore 2, Jain, and PMF) were automatically generated by the
software as regression variables. Feeble correlation between docking scores and the NIK inhibitory activities (IC50
)
of these compounds were observed (Table S1, see supporting information), indicated that these individual scoring
functions could not accurately predict the NIK inhibitory activities of these compounds (Figure 2), with no linear
correlation between score value and experimental biological data. Thus, in order to improve the accuracy and
robustness of QSAR model for the NIK inhibitors, a consensus of docking scores, key interaction profiles and
molecular descriptors were employed to generate nonlinear regression quantitative models. SVR regression,
2
5-27
which has been demonstrated as a preferable methodology in our previous works,
was used to establish
2
9
MD-SVR model. At first, feature selection tool was used to select the most relevant descriptors, providing an

optimal set of molecular descriptors (Table S1, see supporting information). Then, the RBF kernel was used as
kernel function, therefore the capacity parameter C, the ε of ε-insensitive loss functions and the corresponding
parameters γ of RBF kernel needed to be optimized. Initially, intensive grid search method was applied in the
process of leave-one-out (LOO) cross-validation to find the optimal parameters of ε, C and γ. Results showed the
best parameters for ε, C and γ for this MD-SVR model were 0.15, 16, and 0.1, respectively, with the RMSE value
(root mean squared error) of 0.467 (Table S2, see supporting information). The low RMSE confirmed the high
predictability of the model. The pIC50 values estimated by various modelling methods versus experimental values
for the training and test sets are shown in Figure 3A. Furthermore, the plots of residuals versus experimental
factors for each method are shown in Figure 3B.
Figure 1. The 2D structures of 25 NIK inhibitors used to generate the MD-SVR model.
Figure 2. Output docking_scores (A), Ligscore1 (B), Ligscore2 (C), PMF (D), Jain (E) vs. experimental NIK inhibitory
activities (-pIC50), respectively.

Figure 3. pIC50 estimated by various modeling methods versus experimental values for training and test set (A) and
Plots of residuals versus experimental factors (B).
As an example, for illustration, the binding modes of compounds 3 and 25 (Figure 4), as well as their docking
scores were compared. Although compounds 3 and 25 exhibited very different inhibitory potency for NIK (IC50 of 3
and 25: 42000 and 40 nM), they showed a similar interaction with NIK based on molecular docking study. Thus,
rough output docking scores (Dock_scores, Ligscores1, Ligscores2, PMF and Jain) directly from the software
cannot quantitively distinguish the potency of the NIK inhibitors. The QSAR modelling was therefore performed,
introducing other complementary descriptors to enhance the prediction ability of docking scores. Insight into the
interaction modes of compounds 3 and 25 with NIK, two important hydrogen bonds between amino pyrimidine
moieties of compound 3 and 25 and Glu470 and Leu472 of NIK hinge region were observed and recorded. Two
additional hydrogen interactions between the alkynyl alcohol and the 2-methoxyethoxy group of 25 and the
Glu440 and Ser476 of the NIK were also considered. In addition, considering the importance of hydrophilic and
hydrophobic properties, molecular descriptors (ClogP and tPSA) were introduced in the modelling process. After
feature selection, various docking scores (Docking scores, LigScore1, LigScore2, PMF, Jain), key interaction profiles
(Dis2, the distance of the ligand to amino acid LEU472) were integrated for constructing MD-SVR model (Table S2,
2
see supporting information), which can accurately predict the potency of most NIK inhibitors (R training set= 0.938,
2
2
Q
cv= 0.673, R test set= 0.701, Table 1). For compounds 3 and 25, the predicted NIK inhibitory activities (-log (IC50))
were 4.56 and 7.22, respectively, which is very close to the experimental data (-log (IC50) of compounds 3 and 25,
which were 4.38 and 7.4, respectively.

Figure 4. The mode of interaction between 3, 25 and the NIC by molecular docking. (A) The mode of interaction
between NIK and compound 3; (B) The mode of interaction between NIK and compound 25. Hydrogen bonds were
highlighted by yellow dash line.
Table 1 Compounds1-25 in training set and test set for MD-SVR model, and their corresponding experimental and
estimated NIK inhibitory activities.
Experimental
IC50(nM)
Predictive
Experimental
Predictive
Cpd.
1
Res.
Cpd.
14
Res.
b
b
b
pIC50
pIC50
IC50
pIC50
pIC50
4000
190
5.4
6.72
4.38
4.1
5.58
5.88
4.56
4.28
3.92
4.55
6.64
6.63
-0.18
0.83
89
63
7.05
7.2
6.04
6.76
6.08
6.09
3.67
6.82
7.61
4.57
1.01
0.44
0.18
-0.16
1.01
0.18
-0.22
-0.18
[a]
[a]
2
15
3
4
5
6
7
8
42000
80000
80000
80000
150
-0.18
-0.18
0.18
16
17
550
6.26
5.92
4.68
7
1200
21000
100
[a]
4.1
18
19
4.1
-0.44
0.18
[a]
6.82
6.52
20
41
7.39
4.39
300
-0.10
21
41000
[a]
9
110
6.96
6.78
0.18
22
1800
5.75
5.57
0.18
1
1
1
1
0
1
2
3
30000
70
4.52
7.16
5.59
7.2
5.13
6.98
5.55
7.02
-0.60
0.18
0.04
0.18
23
24
25
24000
63
4.62
7.2
4.80
7.38
7.22
-0.18
-0.18
0.18
2600
76
40
7.4
[
a]
The compounds were used as test set. b The -Log IC50 was predicted by MD-SVR model.
The prediction ability of MD-SVR model for designing novel NIK inhibitors
To validate the practical application ability of established MD-SVR models in the actual development of novel
NIK inhibitors, the MD-SVR model was used to predict a novel series of target compounds which were designed by
the shift of side-chain of aminopyrimidine to indole moiety based on the skeleton of compound 25. Compound
32a and 32b exhibiting moderate to potent NIK inhibitory activities (MD-SVR pred. IC50 of 32a and 32b: 5.23 and
4.78, respectively) were selected. As shown in Figure 5, three hydrogen-bonding interactions with Glu470, Leu472
and Gly409 were observed between the aminopyrimidine moiety and the acetylenic alcohol of 32a and 32b,

respectively. For quantitative analysis, five docking scores and key interaction profiles were recorded (Table S4, see
supporting information). In contrast, the biological activity could not be defined by molecular docking model alone,
with the values of Dock_scores, Ligscores1, Ligscores2, PMF and Jain from molecular docking (Table S4) of 32a and
32b were 113.93, 4.61, 5.03, 76.75, 6.00 and 123.51, 3.79, 3.01, 79.78, 9.11, respectively. Based on the different
algorithm, two docking scores (Ligscores1 and Ligscores2) of 32a were superior to that of 32b, while the results
from other three docking score (Dock_scores, PMF and Jain) were opposite. Thus, the discriminant results from
docking scores alone was not clear enough for prediction.
Figure 5. The mode of interaction between 32a, b and NIC by molecular docking. (A) The mode of interaction
between NIK and compound 32a; (B) The mode of interaction between NIK and compound 32b. Hydrogen bonds
were highlighted by yellow dash line.
The synthesis of indole aminopyrimidine alkynyl-alcohol derivatives 32a/b
The synthesis of compounds 32b is outlined in Scheme 2. 5-Bromoindole 26 was N-protected by the tosyl group
and then acylated with acetyl chloride, using aluminum chloride as a catalyst to obtain compound 28. In addition,
o
compound 29 was obtained by treatment of 28 with dimethyl formamide dimethylacetal in DMF at 110 C.
Compound 30 was achieved by the reaction of compound 29 with guanidine hydrochloride in the presence of
3 4
anhydrous potassium carbonate. Sonogashira coupling of 30 with propargyl alcohol in DMF using Pd(PPh ) and
CuI as catalysts resulted in the key intermediate 31. The final product 32a was then obtained by treatment of
o
compound 31 in DMF in the presence of 1-chloro-2-methoxyethane and Cs
2
CO
3
in DMF at 120 C. The target
compounds 32b were obtained by treatmentof compound 31 in presence of NaH and
o
1
-(3-chloropropyl)-1H-imidazole in DMF at 60 C.

Scheme 2. The synthetic route for compounds 32a and 32b. Reaction conditions: (a) p-toluenesulfonylchloride,
o
NaH, DMF; (b)Ac
2
O, AlCl
3
, CH
2
Cl
2
; (c) DMF-DMA, DMF, 110 C; (d) H
2
N(=NH)NH
2
HCl, K
2
3
CO , 2-methoxyethanol,
o
o
reflux; (e) propargylalcohol, Pd(PPh
3
)
4
, CuI, Et
3
N, DMF, 90 C; (f) Cs
2
CO
3
, DMF, 120 C; (g)NaH, chloromethyl propyl,
o
DMF, 60 C.
Pharmacological assay
NIK inhibitory activities of 32a and 32b were determined by using the NIK ATP consumption assay. Compound 8
was used as positive control, showing an IC50 value consistent with the one obtained in previous works. Table 2
presents the pharmacological assay results of 32a and 32b, which showed moderate NIK inhibitory activities (IC50
of 32a and 32b were 250 nM and 1600 nM respectively), suggesting that the position of side chain of compound
25 is essential for NIK inhibitory activities. Compound 32b, bearing imidazole ring as the terminal substituent of
the side chain, showed better activities (2.81-folds more potent) than compound 32a which presents methoxy
group as terminal substituent of the side chain. When comparing predictive and experimental NIK inhibitory
activities of 32a and 32b, a good correlation was observed, which was an accurate and reasonable estimation if
compared with the traditional output docking scores (Dock_scores, Ligscores1, Ligscores2, PMF and Jain) observed
by molecular docking alone.
Table 2 The predictive and experimental NIK inhibitory of compounds 32a-b.
Experimental activities
Predictive
Compd.
Residue
[
a]
IC50 (nM)
250
pIC50
6.60
5.79
/
pIC50
7.08
5.23
/
3
3
1
2a
-0.48
0.56
/
2b
1600
348
[b]
3

[
[
a]
b]
The -Log IC50 was predicted by MD-SVR model.
Compound 13 was synthesized and tested its NIK inhibitory activity as positive control.
Analysis of MD simulations and interaction mechanism
To investigate the interaction mode between the compound 32a and NIK protein, molecular dynamics
(
MD) simulations and interaction decomposition analysis were used. The RMSD value of NIK reached
equilibration and swings around an average value after 100 ps simulation time (Figure 6A). The RMSD
value of the protein backbone was calculated from a 0 to 1500 ps trajectory. After 1500 ps of simulation,
the distance between inhibitors (32a) and key amino acids Leu472 and Gly409 of NIK started to converge,
indicating that the system reached equilibrium conditions (see Figure 6B, C). Based on the stable
conformation obtained from MD simulation, the interactions involved in the protein/ligand complex were
analysed. As shown in Figure 6B, the H-bonds of 32a could form H-bonds with NH and carbonyl of NIK
residue GLY409 in L-shaped geometry. In addition, the aminopyrimidine group on the scaffold also
formed two H-bonds with LEU472; the methoxy group formed one H-bond with LYS429. In accordance
with the interaction energy decomposition, the contribution of LEU472 and LYS429 for 32a were -5.58
kcal/mol and -12.04 kcal/mol respectively (Figure 6D). Therefore, according to the binding pattern and
chemical structure of 32a, we can speculate that the aminopyrimidine moiety might be an important
pharmacophore feature of NIK inhibitors (Table S3, see supporting information).
Figure 6. MD simulations and interaction mechanism analysis for the most potent inhibitor 32a: (A) RMSD of NIK

backbone during the 1500 ps simulation time. (B) plot the binding pattern of 32a. (C) plot the distances of H-bond
between LEU472(C=O), LEU(N), and LYS429 and 32a during 1500 ps of MD simulation; (D) plot the inhibitor–
residue interaction spectrum for 32a.
Conclusion
In this work, we described the development and validation of a molecular docking-based QSAR model by
integrating docking scores and key interaction profiles to improve the accuracy of molecular docking. The optimal
established MD-SVR model was further validated by the application of new NIK inhibitors. Two newly designed
indoleaminopyrimidine derivatives, 32a and 32b, were synthesized and biologically evaluated for their NIK
inhibitory activities, suggesting that the results obtained from the MD-SVR model were more accurate compared
to the traditional output docking scores (Dock_scores, Ligscores1, Ligscores2, PMF and Jain) generated by
molecular docking alone. In addition, the binding mode of compound 32a with NIK protein was investigated by
dynamic simulations, demonstrating the model is a very useful tool for finding highly potent NIK inhibitors.
Experimental section
Dataset
The biological data of NIK inhibitory activities [IC50(nM)] were converted to logarithmic scale [-LogIC50] and then
used for subsequent QSAR analyses as the dependent variable. In general, the 3D-strucures of all compounds
were generated and optimized in Discovery Studio 2.0 software (Accelrys, Inc. San Diego, CA) using the
3
0
CHARMm-like force field. Furthermore, molecular descriptors, such as ClogP and tPSA, were calculated and
introduced as independent variables for modeling, representing the hydrophilicity and hydrophobicity of the
compounds, which are important properties of the binding affinity of the ligand to NIK enzyme.
3
1
Molecular docking with LigandFit
The molecular docking was performed using the LigandFit module embedded in Discovery Studio 2.5. Firstly, the
crystal structure of NIK was obtained from PDB bank (entry code: 4IDV), and then water molecules were removed
and charged by CHARMm force field. The active site was derived from the volume of co-crystal ligand. For
generation of the ligands' conformations, variable numbers of Monte Carlo simulations were employed. All the
1
7
calculations during the docking steps were performed under the PLP energy grid. A short rigid body minimization
was then performed and 50 preferable poses were saved according to their dock scores. Based on the dock scores
and visual inspection, the most possible pose was selected for further analysis. Scoring was performed with a set

of scoring functions (i.e. Dock_score, LigScore2, PLP1, Jain and PMF) implemented in LigandFit module. In addition,
several key interactions have been demonstrated in X-ray diffraction and molecular locking studies, such as amino
acid residues Glu470, Leu472 and Phe535.
Support vector regression (SVR) model construction
QSAR analysis typically generates more descriptors than required for constructing the prediction model, therefore,
30
to selected the most relevant descriptors, SVM feature selection tool were applied. Consequently, support vector
regression (SVR), an extension of a pioneering classification algorithm called the support vector machine, was
performed to construct the QSAR model. The basic idea of SVR is to map the data X into a higher-dimensional
feature space F via a nonlinear mapping and then to do linear regression in this space. The major advantage of
SVR is that it adopts the structure risk minimization (SRM) principle. In the present study, SVR was implemented
3
2
by the SVM function in the e1071 R package which supports several kernel functions (radial basis (or Gaussian),
polynomial, sigmoid, and linear). In present study, we employed radial basis function (RBF) kernels in the modeling.
During the QSAR modeling, the leave-one-out cross (LOO) validation was employed to identify the promising QSAR
model. Finally, cross validation correlation coefficient, a measure of the model generalization capability, for all
candidate models was obtained.
Molecular dynamics
The docked structures of compound 32a complexed within the active pocket of NIK were used as the initial
structures for MD calculations. A CHARMm force field was applied to the complex and the resulting system was
subjected to double-fold minimization (10,000 cycles of steepest descent minimization and 10,000 cycles of
conjugate gradient minimization). The system was heated from 50 K to 300 K over a period of 2 ps and
subsequently equilibrated for 200 ps. Starting from the last frame of the equilibration, a production simulation
was performed for 1500 ps using the NPT ensemble under a constant temperature of 300 K and pressure of 1 atm.
Other parameters of MD simulation were maintained at the default Discovery Studio configuration.
Interaction energy analysis
In order to explore the mode of action of the inhibitor-protein interaction and to highlight the key residues
responsible for the binding affinity, interaction energies for each inhibitor were further decomposed into
individual residue contributions using the “Calculate Interaction Energy” protocol in Discovery Studio. The stable
coordinates extracted in the last 500 ps MD simulation were employed here for the interaction energy
decomposition analysis.
Chemistry

Melting points were determined with a BÜCHI Melting Point B-450 apparatus (Büchi Labortechnik, Flawil,
Switzerland). The 1HNMR spectra were recorded on a Bruker Avance DMX 500 at 500 MHz (chemical shifts are
expressed as d values relative to TMS as internal standard). ESI (positive) was recorded on an Esquire-LC-00075
spectrometer. High resolution mass spectra were measured on an Agilent 1290 HPLC-6224 Time of Fight Mass
Spectrometer. All reactions were monitored by thin-layer chromatography (TLC). All reagents were obtained from
commercial sources and used without further purification unless stated. Anhydrous DMF was distilled from
calcium hydride.
5-Bromo-1-tosyl-1H-indole (27)
To a solution of 5-bromoindole 26 (1 g, 5.1 mmol) in dry DMF (20 mL), sodium hydride (0.209 g, 6.1 mmol) was
added at 0 ℃. The mixture was stirred for 1 h at room temperature and then cooled at 0 ℃. A solution of
p-toluene sulfonyl chloride (1.16 g, 6.1 mmol) in dry DMF (5mL) was added and the mixture was stirred at that
temperature for 2 h. The solution was poured into H
2
O (150 mL) and extracted with EtOAc (3 × 30 mL). The
combined organic layers were dried (Na SO ) and concentrated to dryness. The remaining solid was
2
4
chromatographed on a silica gel column using EtOAc/petroleum ether (1:13) as eluent to afford 27 (1.48 g, 93 %)
1
as a colorless prism, mp: 139-141 ℃. H NMR (500 MHz, CDCl
3
) δ 7.87 (d, J = 8.8 Hz, 1H), 7.75 (d, J = 8.3 Hz, 2H),
7.67 (d, J = 1.7 Hz, 1H), 7.57 (d, J = 3.6 Hz, 1H), 7.41 (dd, J = 8.8, 1.7 Hz, 1H), 7.14 (d, J = 7.9 Hz, 1H), 6.83 (d, J = 8.0
Hz, 1H), 6.60 (d, J = 3.6 Hz, 1H), 2.36 (s, 3H).
-(5-Bromo-1-tosyl-1H-indol-3-yl)ethan-1-one (28)
1
To a suspension solution of AlCl
at ambient temperature for 30 min. A solution of N-tosyl-5-bromoindole (8 g, 23 mmol) in CH
added dropwise at 0 ℃. The mixture was stirred at ambient temperature for 1 h and poured into ice saturated
NH Cl suspension. The aqueous layer was extracted with CH Cl . The organic extracts were washed with brine and
dried over Na
3
(8 g, 114 mmol) in CH
2
Cl
2
(60 mL) was added acetic anhydride (6.99 g, 66 mmol)
2
2
Cl (20 mL) was
4
2
2
2
SO
4
. After removal of the solvent, the crude product was recrystallized with MeOH to give 28 (7.53
1
g, 85 %) as a colorless prism. mp: 153-155 ℃. H NMR (500 MHz, DMSO-d6): δ 8.86 (s, 1H), 8.31 (d, J = 2.0 Hz, 1H),
8
.04 (d, J = 8.4 Hz, 2H), 7.92 (d, J = 8.9 Hz, 1H), 7.57 (dd, J = 8.9, 2.1 Hz, 1H), 7.46 (d, J = 8.1 Hz, 2H), 2.59 (s, 3H),
.35 (s, 3H).
2
(
E)-1-(5-Bromo-1-tosyl-1H-indol-3-yl)-3 (dimethylamino)prop-2-en-1-one (29)
To a solution of N-tosyl-3-Acetyl-5-bromoindole 28 (0.78 g, 1.99 mmol) in dry DMF (20 mL) was added a solution
of dimethyl formamide dimethylacetal (0.64 g, 5.34 mmol). The resultant solution was heated at 110 ℃ for 5 h
2 2
under N . After cooling, the solution was poured into H O (300 mL) and then extracted with EtOAc. The organic

extracts were washed with brine and dried over Na
2
SO
4
. After removal of the solvent, the crude product was
1
recrystallized with EtOAc to give 29 (0.63 g 70.83%) as a yellow prism. mp:175-178 ℃. H NMR (500 MHz,
DMSO-d6) δ8.53 (d, J=1.9 Hz, 1H),8.08(s, 1H), 7.78 (d, J = 8.8 Hz, 1H), 7.76(d, J = 12.2 Hz, 1H), 7.76 (d, J = 8.5 Hz,
2
3
4
H), 7.40 (dd, J = 8.8, 1.9 Hz, 1H), 7.22 (d, J = 8.5 Hz, 2H), 5.56 (d,J = 12.2 Hz, 1H), 3.09 (s, 3H), 3.00(s, 3H), 2.32 (s,
H).
-(5-Bromo-1H-indol-3-yl)pyrimidin-2-amine (30)
A mixture of enaminone 29 (3.45 g, 7.71 mmol), guanidine hydrochloride (1.84 g, 19.28 mmol), anhydrous K
3.72 g, 26.99 mmol) and 2-methoxyethanol (60 mL) was heated at reflux temperature for 4 h under N . After
cooling, the solution was poured into H O (300 mL) and then extracted with EtOAc. The organic extracts were
washed with brine and dried over Na SO and concentrated to dryness. The remaining solid was chromatographed
/MeOH (30:1) as eluent to afford 30 (1.9 g, 95%) as an off-white prism. mp:
2 3
CO
(
2
2
2
4
2 2
on a silica gel column using CH Cl
1
2
=
3
42-243 ℃. H NMR (500 MHz, DMSO-d6) δ 11.86 (s, 1H), 8.63 (d, J = 2.1 Hz, 1H), 8.27 (d, J = 2.9 Hz, 1H), 8.11 (d, J
5.4 Hz, 1H), 7.46 (d, J = 8.6 Hz, 1H), 7.18 (dd, J = 8.6, 2.1 Hz, 1H), 7.01 (d, J = 5.3 Hz, 1H), 6.48 (s, 2H).
-(Pyrimidinamine)-indole-5-methylbut-3-yn-2-ol (31)
A solution of Meridianin C (2 g, 6.99 mmol), 2-methylbut-3-yn-2-ol (2 mg, 24.47 mmol), CuI (533 mg, 2.79 mmol)
and Pd(PPh (82 mg, 1.39 mmol) in 20 mL Et N and 20mL DMF for 10 h at 90 ℃ under N . After cooling, the
solution was poured into H
brine and dried over Na SO
gel column using CH Cl
3
)
4
3
2
2
O (100 mL) and then extracted with EtOAc. The organic extracts were washed with
2
4
. After removal of the solvent, the remaining solid was chromatographed on a silica
1
2
2
/MeOH (15:1) as eluent to afford 31(1.1g, 53%) as a yellow solid. mp: 165-167 ℃. H NMR
(500 MHz, DMSO-d6) δ 11.83 (s, 1H), 8.56 (s, 1H), 8.22 (d, J = 5.22 Hz, 1H), 8.11 (d, J = 5.3 Hz, 1H), 7.43 (d, J = 8.4
Hz, 1H), 7.18 (dd, J = 8.4, 1.5 Hz, 1H), 7.01 (d, J = 5.3 Hz, 1H), 6.44 (s, 2H), 5.46 (s, 1H), 1.50 (s, 6H).
-(3-(2-Aminopyrimidin-4-yl)-1-(2-methoxyethyl)-1H-indol-5-yl)-2-methylbut- 3-yn-2-ol (32a)
4
A solution of compound 31 (50 mg, 0.17 mmol), 1-chloro-2-methoxyethane (48 mg, 0.51 mmol), and Cs
mg, 0.34 mol) in 15 mL DMF were stirred for 20 h at 120 ℃under N . After cooling, the solution was poured into
O (65 mL) and then extracted with EtOAc. The organic extracts were washed with brine and dried over Na SO
/MeOH
2 3
CO (110
2
H
2
2
4
.
2 2
After removal of the solvent, the remaining solid was chromatographed on a silica gel column using CH Cl
1
(
15:1) as eluent to afford 32a (40 mg, 66%) as a yellow solid. mp: 167-169 ℃. H NMR (500 MHz, DMSO-d6) δ 8.56
(d, J = 1.1 Hz, 1H), 8.25 (s, 1H), 8.14 (d, J = 5.3 Hz, 1H), 7.57 (d, J = 8.5 Hz, 1H), 7.23 (dd, J = 8.5, 1.5 Hz, 1H), 6.96 (d,
1
3
J = 5.3 Hz, 1H),6.52 (s, 2H), 5.42 (s, 1H), 4.41 (t, J = 5.1 Hz, 2H), 3.70 (t, J = 5.1 Hz, 2H), 3.34 (s, 3H), 1.51 (s, 6H). C
NMR (126 MHz, DMSO-d6) δ 163.48 (s), 161.93 (s), 157.16 (s), 136.48 (s), 132.68 (s), 125.42 (2s), 114.79 (s),

113.03 (s), 110.85 (s), 105.44 (s), 93.89 (s), 81.92 (s), 70.58 (s), 63.70 (s), 58.08 (s), 45.88 (s), 31.86 (2s). HRMS m/z
+
+
+
(
ESI ): Calculated for C20
22 4 2
H N O H ([M+H] ): 351.1821, Found 351.1828.
4-(1-(3-(1H-imidazol-1-yl)propyl)-3-(2-aminopyrimidin-4-yl)-1H-indol-5-yl)-2-methylbut-3-yn-2-ol (32b)
Sodium hydride (15 mg, 0.37 mmol) was added to a solution of compound 31 (90 mg, 0.31 mmol) in dry 10 mL
DMF at 0 ℃, after 0.5 h, 1-(3-chloropropyl)-1H-imidazole (66 mg, 0.47 mmol) was added. The resultant solution
was heated at 65 ℃ for 10 h. After cooling, the solution was poured into H
2
O (50 mL) and then extracted with
EtOAc. The organic extracts were washed with brine and dried over Na SO
2
4
. After removal of the solvent, the
2 2
remaining solid was chromatographed on a silica gel column using CH Cl /MeOH (15:1) as eluent to afford as a
1
yellow solid (0.97g, 78%), mp: 164-166 ℃. H NMR (500 MHz, DMSO-d6) δ 8.76 (s, 1H), 8.59 (s, 1H), 8.38 (s, 1H),
8
.16 (d, J = 5.2 Hz, 1H), 7.65 (s, 1H), 7.55 (d, J = 8.5Hz, 1H), 7.48 (s, 1H), 7.26 (d, J = 8.2 Hz, 1H), 7.03 (d, J = 5.3 Hz,
H), 6.86 (s, 2H), 5.43 (s, 1H), 4.30 (t, J = 6.5 Hz, 2H), 4.22 (t, J = 6.5 Hz, 2H), 2.42 (m, 2H), 1.51 (s, 6H).HRMS m/z
1
+
+
+
(
ESI ): Calculated for C23
24 6
H N OH ([M+H] ): 401.2090. Found 401.2093.
NIK inhibitory assay
Recombinant human NIK (318-947 aa) was purchased from BPS Bioscince. NIK ATP consumption assay was
performed with ADP-GloTM Kinase Assay (Promega) according to the manufacturer's protocol. Briefly, NIK assay
2
components were diluted in HEPES buffer (50 mM HEPES, 10 mM MgCl , 0.01% Triton X-100, 2 mM DTT, pH7.2),
and cultured for 1 hour at room temperature, which contained 2 μl enzyme solution (30 nM NIK), 2 μl substrate
solution (5 μM Ultra Pure ATP), and 1 μL inhibitor in 384-well microplate. And then, firstly, 5 μl ADP-Glo™ Reagent
was added to terminate the kinase reaction and deplete the remaining ATP; secondly, the Kinase Detection
Reagent was added to simultaneously convert ADP to ATP and allow the newly synthesized ATP to be measured
using a luciferase/luciferin reaction with Envision (PerkinElmer). The test was carried out in triplicate. And the IC50
data was calculated using the software GraphPad Prism 5 and chosen the equation “sigmoidal dose-response
(variable slope)” for curve fitting.
Acknowledgements
The Postdoctoral Science Foundation of China (2014M550256, 2017M612017), the State Key Laboratory of Drug
Research (SIMM1403KF-02, SIMM1705KF-08), Natural Science Foundation of China (81502926), Science and
Technology Planning Project of Zhejiang Province (2014C37050), and National Science & Technology Major Project
“
Key New Drug Creation and Manufacturing Program ”， China (2018ZX09711002-007).

Conflict of Interest
We declare we have no financial/commercial conflict of interest.
References
[
[
[
[
[
1] M. S. Hayden and S. Ghosh, Cell, 2008, 132, 344-362.
2] F. Martinon and J. Tschopp, Trends Immunol., 2005, 26, 447-454.
3] S. Vallabhapurapu and M. Karin, Annu. Rev. Immunol., 2009, 27, 693-733.
4] cG. Bonizzi, M. Karin, Trends. Immunol., 2004, 25, 280-288.
5] U. Senftleben, Y. Cao, G. Xiao, F. R. Greten, G. Krahn, G. Bonizzi, Y. Chen, Y. Hu, A. Fong, S. C. Sun and M. Karin,
Science, 2001, 293, 1495-1499.
[
6] O. H. Negm, H. A. Mannsperger, E. M. McDermott, E. Drewe, R. J. Powell, I. Todd, L. C. Fairclough and P. J.
Tighe, Eur. J. Immunol., 2014, 44, 2096-2110.
[
[
[
[
[
7] S. Yuan, Y. Yu, S. Huang, T. Liu, T. Wu, M. Dong, S. Chen, Y. Yu, A. Xu, Mol. Immunol., 2007, 44, 756-762
8] B. Razani, A. D. Reichardt and G. Cheng, Immunol. Rev., 2011, 244, 44-54.
9] Y. N. Demchenko and W. M. Kuehl, Oncotarget, 2010, 1, 59-68.
10] S. G. Hymowitz and I. E. Wertz, Nature reviews. Cancer, 2010, 10, 332-341.
11] W. Lim, J. Kim, S. Kim, S. Karna, J. Won, S. M. Jeon, S. Y. Kim, Y. Choi, H. Choi and O. Kim, Photochem.
Photobiol., 2013, 89, 199-207.
[
[
12] Y.M. Thu, A. Richmond, Cytokine Growth Factor Rev., 2010, 21, 213-226.
13] T. Nishiuchi, Y. Okutani, Y. Yamagishi, T. Fujita, O. Imataki, H. Ohnishi, T. Matsunaga, T. Shimizu and H. Shimizu,
J. Altern. Complement Med., 2013, 19, 593-597.
[
14] J. B. Perkins, J. Kim, C. Anasetti, H. F. Fernandez, L. E. Perez, E. Ayala, M. A. Kharfan-Dabaja, M. R. Tomblyn, D.
M. Sullivan, J. A. Pidala and T. L. Field, Biol. Blood Marrow Transplant, 2012, 18, 1099-1107.
15] J.H. Je, J.Y. Lee, K.J. Jung, B. Sung, E.K. Go, B.P. Yu, H.Y. Chung, FEBS Lett., 2004, 566, 183-189.
16] M. B. Qin, J. X. Zhang, C. Y. Xu, P. Peng, L. Tan, S. Q. Liu, J. A. Huang, Plos. One. 2017, 12.
17] M. Montes, M. A. Miteva and B. O. Villoutreix, Proteins, 2007, 68, 712-725.
18] J. Mortier, R. Frederick, C. Ganeff, C. Remouchamps, P. Talaga, L. Pochet, J. Wouters, J. Piette, E. Dejardin, B.
Masereel, Biochem. Pharmacol. 2010, 79, 1462-1472.
[
[
[
[
[
19] J. Mortier, B. Masereel, C. Remouchamps, C. Ganeff, J. Piette, R. Frederick, Bioorg. Med. Chem. Lett., 2010, 20,

4515-4520.
[
[
20] T. T. Yao, J. F. Xie, X. G. Liu, J. L. Cheng, C. Y. Zhu, J. H. Zhao, X. W. Dong, Rsc. Adv., 2017, 7, 10353-10360.
21] W. L. Chen, Z. H. Wang, T. T. Feng, D. D. Li, C. H. Wang, X. L. Xu, X. J. Zhang, Q. D. You and X. K. Guo, Bioorg.
Med. Chem. Lett., 2016, 24, 6102-6108.
[
[
[
[
[
[
[
22] W. Deng, C. Breneman and M. J. Embrechts, J. Chem. Inf. Comput. Sci., 2004, 44, 699-703.
23] K. Z. Myint, L. R. Wang, Q. Tong and X. Q. Xie, Mol. Pharm., 2012, 9, 2912-2923.
24] C. Cortes, V. Vapnik, Mach. Learn. 1995, 20, 273-297.
25] C. Q. Hu, K. Li, T. T. Yao, Y. Z. Hu, H. Z. Ying, X. W. Dong, MedChemComm. 2017, 8, 1835-1844
26] W. Zhan, D. Li, J. Che, L. Zhang, B. Yang, Y. Hu, T. Liu and X. Dong, Eur. J. Med. Chem., 2014, 75, 11-20.
27] T. T. Yao, J. L. Cheng, B. R. Xu, M. Z. Zhang, Y. Z. Hu, J. H. Zhao, X. W. Dong, Rsc Adv., 2015, 5, 49195-49203.
28] K. Li, L. R. McGee, B. Fisher, A. Sudom, J. Liu, S. M. Rubenstein, M. K. Anwer, T. D. Cushing, Y. Shin, M. Ayres, F.
Lee, J. Eksterowicz, P. Faulder, B. Waszkowycz, O. Plotnikova, E. Farrelly, S. H. Xiao, G. Chen and Z. Wang,
Bioorg. Med. Chem. Lett., 2013, 23, 1238-1244.
[
[
29] J. E. Stenger, H. Xu, C. Haynes, E. R. Hauser, M. Pericak-Vance, P. J. Goldschmidt-Clermont, J. M. Vance, Bmc
Bioinformatics, 2005, 6, 95.
30] B.R. Brooks, R.E. Bruccoleri, B.D. Olafson, D.J. States, S. Swaminathan, M. Karplus, J. Comput. Chem., 1983, 4,
187-217.
[
[
31] C. M. Venkatachalam, X. Jiang, T. Oldfield and M. Waldman, J. Mol. Graph. Model., 2003, 21, 289-307.
32] Q. Wang and X. D. Liu, Oncotargets Ther., 2015, 8, 2311-2317.

Highlights



The integration of docking scores, key interaction profiles remarkably improved the accuracy of
the QSAR models.
The established MD-SVR (Molecular Docking Based SVM regression) model was applied on
designing new NIK inhibitors.
The identified compounds would be promising leads for finding highly potent NIK inhibitors.

Graphical abstract