Efficient identification of novel anti-glioma lead compounds by machine learning models

Article abstract of DOI:10.1016/j.ejmech.2019.111981

Glioblastoma multiforme (GBM) is the most devastating and widespread primary central nervous system tumor. Pharmacological treatment of this malignance is limited by the selective permeability of the blood-brain barrier (BBB) and relies on a single drug, temozolomide (TMZ), thus making the discovery of new compounds challenging and urgent. Therefore, aiming to discover new anti-glioma drugs, we developed robust machine learning models for predicting anti-glioma activity and BBB penetration ability of new compounds. Using these models, we prioritized 41 compounds from our in-house library of compounds, for further in vitro testing against three glioma cell lines and astrocytes. Subsequently, the most potent and selective compounds were resynthesized and tested in vivo using an orthotopic glioma model. This approach revealed two lead candidates, 4m and 4n, which efficiently decreased malignant glioma development in mice, probably by inhibiting thioredoxin reductase activity, as shown by our enzymological assays. Moreover, these two compounds did not promote body weight reduction, death of animals, or altered hematological and toxicological markers, making then good candidates for lead optimization as anti-glioma drug candidates.

Full text of DOI:10.1016/j.ejmech.2019.111981

Journal Pre-proof
Efficient identification of novel anti-glioma lead compounds by machine learning
models
Bruno Junior Neves, Jonathan Paulo Agnes, Marcelo do Nascimento Gomes, Marcio
Roberto Henriques Donza, Rosângela Mayer Gonçalves, Marina Delgobo, Lauro
Ribeiro de Souza Neto, Mario Roberto Senger, Floriano Paes Silva-Junior, Sabrina
Baptista Ferreira, Alfeu Zanotto-Filho, Carolina Horta Andrade
PII:
S0223-5234(19)31133-X
DOI:
https://doi.org/10.1016/j.ejmech.2019.111981
EJMECH 111981
Reference:
To appear in: European Journal of Medicinal Chemistry
Received Date: 3 July 2019
Revised Date: 18 November 2019
Accepted Date: 16 December 2019
Please cite this article as: B.J. Neves, J.P. Agnes, M.d.N. Gomes, M.R. Henriques Donza, Rosâ.Mayer.
Gonçalves, M. Delgobo, L. Ribeiro de Souza Neto, M.R. Senger, F.P. Silva-Junior, S.B. Ferreira, A.
Zanotto-Filho, C.H. Andrade, Efficient identification of novel anti-glioma lead compounds by machine
learning models, European Journal of Medicinal Chemistry (2020), doi: https://doi.org/10.1016/
j.ejmech.2019.111981.
This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition
of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of
record. This version will undergo additional copyediting, typesetting and review before it is published
in its final form, but we are providing this version to give early visibility of the article. 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.
© 2019 Published by Elsevier Masson SAS.

1
2
Efficient Identification of Novel Anti-Glioma Lead
Compounds by Machine Learning Models
3
4
Bruno Junior Neves1,2§, Jonathan Paulo Agnes3§, Marcelo do Nascimento Gomes2,4, Marcio
Roberto Henriques Donza5, Rosângela Mayer Gonçalves3, Marina Delgobo3, Lauro Ribeiro de
Souza Neto6, Mario Roberto Senger6, Floriano Paes Silva-Junior6, Sabrina Baptista Ferreira5,
Alfeu Zanotto-Filho3*, and Carolina Horta Andrade2*
5
6
7
8
9
1LabChem – Laboratory of Cheminformatics, Centro Universitário de Anápolis,
UniEVANGÉLICA, Anápolis, GO, 75083-515, Brazil.
10
11
12
13
14
15
16
17
18
19
20
21
22
2 LabMol – Laboratory for Molecular Modeling and Drug Design, Faculdade de Farmácia,
Universidade Federal de Goiás, Goiânia, GO, 74605-510, Brazil.
3LabCancer - Laboratório de Farmacologia e Bioquímica do Câncer, Departamento de
Farmacologia, Centro de Ciências Biológicas, Universidade Federal de Santa Catarina,
Florianópolis, SC, Brazil.
4InSiChem Drug Discovery – Universidade Estadual de Goiás, Anápolis, GO, 74643-090, Brazil.
5LSOPB – Laboratório de Síntese Orgânica e Prospecção Biológica, Instituto de Química,
Universidade Federal do Rio de Janeiro, Rio de Janeiro, RJ, 21949-900, Brazil.
6 LaBECFar – Laboratório de Bioquímica Experimental e Computacional de Fármacos, Instituto
Oswaldo Cruz, Fundação Oswaldo Cruz, Rio de Janeiro, RJ, 21040-900, Brazil.
§
These authors provided equal contributions to this study and should be considered as co-first authors.
23
24
*Corresponding authors: CHA, Tel: +55 62 3209 6451, Fax: +55 62 3209 6037, E-mail: carolina@ufg.br; AZF, Tel:
+55 48 37212474, Fax: +55 48 33375479, E-mail: alfeu.zanotto@ufsc.br
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
ABSTRACT
Glioblastoma multiforme (GBM) is the most devastating and widespread primary central nervous
system tumor. Pharmacological treatment of this malignance is limited by the selective permeability
of the blood-brain barrier (BBB) and relies on a single drug, temozolomide (TMZ), thus making the
discovery of new compounds challenging and urgent. Therefore, aiming to discover new anti-
glioma drugs, we developed robust machine learning models for predicting anti-glioma activity and
BBB penetration ability of new compounds. Using these models, we prioritized 41 compounds from
our in-house library of compounds, for further in vitro testing against three glioma cell lines and
astrocytes. Subsequently, the most potent and selective compounds were resynthesized and tested in
vivo using an orthotopic glioma model. This approach revealed two lead candidates, 4m and 4n,
which efficiently decreased malignant glioma development in mice, probably by inhibiting
thioredoxin reductase activity, as shown by our enzymological assays. Moreover, these two
compounds did not promote body weight reduction, death of animals, or altered hematological and
toxicological markers, making then good candidates for lead optimization as anti-glioma drug
candidates.
Keywords: Cancer; glioblastoma; machine learning; predictive modeling; orthotopic glioma model;
thioredoxin reductase.

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
1. Introduction
Glioblastoma multiforme (GBM) or grade IV glioma, is the most commonly occurring and
aggressive type of primary central nervous system (CNS) tumor [1,2]. Most of GBM occur
particularly in the brain, but they can also appear in cerebellum brainstem and also in spinal cord
[3,4]. This malignance can manifest at any age, but is more frequent in adult men with a median age
of 64 years [5]. The majority symptoms of GBM include intracranial pressure, headache and focal
or progressive neurologic deficits [5]. Consequently, GBM prognosis remains dismal for decades
with a median overall survival of ~14 months, with less than 10% of patients surviving beyond five
years [6,7].
GBM is particularly difficult to treat due to its characteristic of excessive invasiveness and fast-
growing behavior, as well as its particular location and the selective permeability offered by the
blood-brain barrier (BBB)[8,9]. At present, chemotherapy is the main postsurgical and adjuvant
therapy for GBM, and the alkylating agent temozolomide (TMZ) is the first-line drug frequently
combined with radiotherapy [10]. TMZ is absorbed orally and presents favorable toxicity profile
compared with older alkylating agents such as carmustine [11]. However, the emergence of
temozolomide resistance hampers its use in GBM patients [12,13]. Hence, the discovery of new
anti-glioma drugs is urgently needed.
In this context, quantitative machine learning (ML) has exerted profound impact on drug
discovery, making it faster and less expensive [14,15]. ML is a growing field of artificial
intelligence that uses different statistical techniques to enable computers to learn from various data
types without being explicitly [16]. Several ML methods, such as Support Vector Machines [17],
Random Forest [18–20], and more recently Deep Neural Networks [21–24], have been utilized
for drug discovery. Methodologically, ML tools uses pattern recognition algorithms to discern
mathematical relationships between experimental observations of small molecules and extrapolate
them to predict biological properties of novel compounds.[25–27] So, ML represents a helpful tool
for virtual screening (VS) of new chemicals with desired biological properties.

77
78
79
80
In this work, we developed ML models and applied them for predicting the anti-proliferative
activity against glioma cells and the BBB penetration ability of new compounds from our in-house
library. Then, the prioritized compounds were experimentally evaluated in vitro against glioma cells
and astrocytes, and in vivo using an orthotopic glioma model.
81
82
83
84
85
86
87
88
2. Results and discussion
The general study design is presented in Fig 1. Briefly, we followed successive steps as follows:
(i) data collection, curation, and integration of compounds reported in the literature with activity
against C6 glioma cells and BBB penetration; (ii) chemical space analysis of curated datasets; (iii)
development and validation of ML models; (iv) VS of an in-house chemical database (1,250
compounds) using the ML models for prioritization of compounds; (v) in vitro experimental
validation using phenotypic and enzymatic assays; (vi) chemical synthesis of most promising
compounds; and (vii) in vivo investigation using orthotopic glioma model.
89
90
Figure 1. Study design. (i) data collection, curation, and integration of molecules with activity
against C6 glioma cells and BBB penetration; (ii) chemical space analysis of curated datasets; (iii)
development of binary and continuous QSAR models; (iv) VS of an in-house chemical database
(containing 1,250 compounds); (v) in vitro experimental validation using phenotypic and enzymatic
assays; (vi) chemical synthesis of most promising compounds; (vii) in vivo investigation using
orthotopic glioma model, and (viii) identification of novel anti-glioma leads.
91
92
93
94
95

96
97
Initially, two datasets of compounds with anti-glioma (Supplementary File S1) and blood-brain
barrier (File S2) bioactivity data were retrieved from the ChEMBL database (ID: CHEMBL614657
[28]) and scientific literature [29–32], respectively. Both datasets were submitted to a rigorous data
curation protocol. An activity threshold of 10 µM based on half maximal effective concentration
(EC50) against C6 rat glioma cells was defined for discrimination between active and inactive
compounds. Compounds with experimental logBB greater than or equal to −1 were labeled as
BBB+ (penetrating) and those with logBB below −1 as BBB− (not penetrating). Subsequently, the
C6 dataset (97 active and 173 inactive compounds), and BBB dataset (433 BBB− and 1436 BBB+
compounds) were balanced using a linear under-sampling approach [33].
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
2.1. Chemical space analysis
To visualize the structural diversity of our datasets, we performed a principal component
analysis (PCA, see Supplementary Figure S1). The PCA reduces high-dimensional space composed
by Molecular ACCess System (MACCS) keys into a smaller number of orthogonal (non-correlated)
variables called principal components (PCs), thus making it more manageable and comprehensible
by extracting essential information [34,35]. The PCA model with the first two PCs described 34.8%
of total data variance. Projecting variables on the planes defined by a PC1 and PC2 allows an
interesting chemical space analysis, in which most of the active (green dots) and inactive (yellow
dots) compounds from C6 dataset overlap within the same regions of chemical space (defined by
PC1 and PC2) of BBB+ (blue dots) and BBB− (red dots) compounds from the BBB dataset. This
analysis revealed that multiple compounds active against glioma cells can also potentially penetrate
the BBB. Based on these data, we developed predictive computational models for both biological
properties in order to select only compounds predicted as active for C6 cells and BBB+.
118
119
120
2.2. Performance of ML models
ML models were built to distinguish active vs. inactive compounds for C6 cell line (see
Supplementary File S1) and BBB+ vs. BBB− compounds (Supplementary File S2). Statistical

121
122
123
124
125
126
127
128
129
130
characteristics of developed ML models estimated by 5-fold external cross-validation are reported
in Table 1. According to the statistical results, the combination of Morgan and FeatMorgan
fingerprints (radius 2: FeatMorgan_2, Morgan_2; radius 4: FeatMorgan_4, Morgan_4) with
Random Forest algorithm led to predictive binary ML models. Briefly, correct classification rate
(CCR) values were ranging between 0.83–0.87; sensitivity (SE) 0.82–0.87; specificity (SP) 0.82–
0.87; positive predictive value (PPV) 0.84–0.88; negative predictive value (NPV) 0.81–0.88; and a
Cohen’s kappa (κ) 0.66–0.73. The model built using Morgan_2 demonstrated the best performance
among all other models developed for antiproliferative activity against C6 cells (CCR = 0.87; SE =
0.89; SP = 0.85) and BBB (CCR = 0.85; SE = 0.81; and SP = 0.89).
Table 1. Summarized statistical characteristics of ML models.
Models
CCR
SE
SP
PPV
NPV
κ
Coverage
Glioma (C6 cell line) models
Morgan_2
Morgan_4
0.87
0.85
0.85
0.85
0.89
0.84
0.86
0.86
0.85
0.86
0.84
0.84
0.85
0.85
0.84
0.84
0.88
0.84
0.85
0.85
0.73
0.69
0.69
0.69
0.52
0.51
0.59
0.51
FeatMorgan_2
FeatMorgan_4
Blood-brain barrier (BBB) models
Morgan_2
Morgan_4
0.85
0.84
0.83
0.84
0.81
0.81
0.79
0.80
0.89
0.88
0.87
0.88
0.88
0.87
0.86
0.87
0.82
0.82
0.81
0.81
0.70
0.68
0.66
0.67
0.59
0.57
0.61
0.57
FeatMorgan_2
FeatMorgan_4
131
132
CCR: correct classification rate; SE: sensitivity; SP: specificity; PPV: positive predictive value; and NPV: negative
predictive value; κ: Cohen’s kappa; Coverage: percentage of test set compounds within the applicability domain.
133
134
2.3. Virtual screening
The virtual screening (VS) was carried out following the workflow presented in Figure 2.

135
136
Figure 2. Virtual screening workflow used for identifying new anti-glioma hits. Colloidal
aggregation tool was used to filter out molecules that are known to aggregate in experimental
assays; chemical similarity analysis and visual inspection were performed to select a subset of
structurally diverse virtual hits.
137
138
139
140
141
142
143
144
145
146
147
148
149
150
Initially, 1,250 structurally diverse compounds (synthesized or purchased) available in our in-
house database were compiled and standardized for VS. Then, the best ML models developed for
C6 and BBB were used to prioritize potential anti-glioma compounds. The final selection of hits can
be summarized as follows: (i) the compounds predicted as active and BBB+ by the ML models; and
(ii) compounds inside the applicability domain (AD) of the ML models. The AD was determined in
order to set “reliable” and “unreliable” predictions [36,37]. The predictions were considered reliable
when the virtual hits are within the chemical space of the molecules used for training models.
Subsequently, a colloidal aggregation tool was used to filter out molecules that are known to
aggregate in experimental assays [38,39]. Finally, we performed a chemical similarity analysis to
select a subset of structurally diverse virtual hits. At the end of this process, twelve putative hits
with model probability >0.65 (Table 2) were selected for biological evaluation.
151
2.4. In vitro antiproliferative activity

152
153
154
155
156
157
The twelve prioritized hits were primarily evaluated in vitro against three glioma cell lines
(Table 2). The EC50 values (see Table 2) indicated that three compounds, 2-[(E)-2-(2-
nitrophenyl)ethenyl]quinolin-8-yl
propanoate
(1),
(2E)-1-[4-(morpholin-4-yl)phenyl]-3-(5-
nitrofuran-2-yl)prop-2-en-1-one (4), and (2E)-1-[4-(1H-imidazol-1-yl)phenyl]-3-(5-nitrothiophen-2-
yl)prop-2-en-1-one (5) were potent at inhibiting the cell growth, showing activities in
submicromolar range against C6, U251MG, and U87MG cells.

158
159
Table 2. ML probability of selected virtual screening hits, in vitro activity against glioma cell lines (C6, U251MG and U87MG), cytotoxicity on
astrocytes and selectivity index.
ML
EC50 (µM) ± SD
Selectivity indexb
Probability
ID
1
Chemical structure
C6
BBB
C6
U251MG
U87MG
Astrocytes
C6
U251MG
U87MG
0.76
0.67
6.3 ± 0.8
10.1 ± 2.2
10.5 ± 2.8
55.2 ± 5.5
8.7
5.4
5.3
0.81
0.82
0.88
0.94
0.94
0.96
>50
>50
>50
>50
−
−
−
−
−
−
−
−
2
−
−
2
3
4
6.6 ± 2.1
37.8 ± 3.6
9.8 ± 2.9
75.8 ± 8.5
11.5
7.7

0.92
0.94
1.9 ± 1.1
9.4 ± 2.7
10.1 ± 3.4
40.4 ± 3.2
21.3
4.3
4
5
0.68
0.76
0.70
0.72
0.95
0.91
0.96
0.68
48.9 ± 4.4
>50
>50
>50
>50
>50
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
6
7
8
9
>50
>50
0.81
0.80
0.96
0.98
>50
>50
>50
−
−
−
−
−
−
−
−
−
−
10
11
46.6 ± 5.2
0.88
0.89
>50
>50
−
−
−
−
−
12

TMZa
−
−
60.46 ± 3.62
>200
>200
−
−
−
−
b
160
161
aData retrieved from reference [40]; Selectivity index calculated by astrocyte CC50/glioma cell line EC50; The data are expressed as mean ± SD of three independent assays.
Dashed SI values means that cytotoxicity against astrocytes cannot be calculated because compounds did not show activity even at highest concentrations used in the assay.

162
163
164
165
166
Since heteroaryl chalcones 4 (EC50 = 6.6 µM and 9.8 µM for C6 and U87MG, respectively)
and 5 (EC50 = 1.9 µM, 9.4 µM and 10.1 µM for C6, U251MG and U87MG, respectively) were
the most promising anti-glioma hits in experimental assays, a new round of biological assays
were performed with 29 structural analogs available on in-house database (Table 3).

167
168
Table 3. In vitro cytotoxicity of heteroaryl chalcones against glioma cell lines (C6, U251MG, and U87MG) and primary astrocytes, and thioredoxin
reductase activity.
ML
Selectivity indexb
TrxRa
(%)
Chemical structure
EC50 (µM) ± SD
probability
C6 BBB
ID
R1
R2
R3
C6
U251MG
U87MG
Astrocytes
C6 U251MG U87MG
H
H
H
H
H
Br
I
A
A
A
A
A
0.85 0.97
0.86 0.97
0.85 0.91
0.84 0.88
0.65 0.96
2.7 ± 1.1
2.1 ± 0.7
1.6 ± 1.1
28.6 ± 5.7
1.7 ± 0.8
21.4 ± 4.4
12.5 ± 4.1
14.2 ± 3.3
44.5 ± 6.3
3.9 ± 2.1
12.6 ± 3.4
5.2 ± 2.7
3.9 ± 1.7
−
85.4 ± 7.4
68.2 ± 5.2
28.4 ± 3.6
−
−
−
31.6
32.5
17.7
−
3.9
5.4
2
6.7
13.1
7.3
−
4a
4b
4c
4d
4e
SCH3
56.4
−
C(CH3)3
(CH2)3CH3
−
5.8 ± 1.6
41.2 ± 2.8
80.7
24.2
10.6
7.1
H
A
0.86 0.97
3.5± 1.2
22.6 ± 2.5
13.4 ± 2.9
45.3 ± 4.4
−
12.9
2
3.4
4f

H
H
A
A
0.86 0.98
0.76 0.96
16.4 ± 5.1
17.7 ± 4.5
35.9 ± 3.3
24.6 ± 4.4
35.4 ± 7.7
65.8 ± 5.7
−
−
4
1.8
−
1.8
−
4g
4h
−
−
−
H
H
A
A
0.80 0.96
0.87 0.97
5.1± 1.6
4.4 ± 1.5
>50
9.5 ± 3.3
57.2 ± 5.2
−
−
11.2
−
13
−
6
4i
4j
48.7 ± 6.6
−
−
−
H
H
A
A
0.73 0.94
0.92 0.87
1.5 ± 0.1
6.6 ± 2.1
16.7 ± 6.4
37.8 ± 3.6
4.9 ± 2.2
9.8 ± 2.9
44.3 ± 4.2
75.8 ± 8.5
0
29.5
11.5
2.6
2
9
4k
4l
−
7.7
CH3
Br
H
H
H
A
A
B
B
0.81 0.96
0.81 0.97
0.84 0.97
0.92 0.87
2.1 ± 0.5
1.4 ± 0.9
30.5 ± 6.8
7.8 + 3.1
2.3 ± 0.3
5.8 ± 1.3
47.1 ± 6.2
23.3 ± 3.4
3.3 ± 0.8
6.8 ± 2.4
−
52.2 ± 4.8
35.6 ± 3.2
−
95.2
78.7
−
24.8
25.4
−
22.7
6.1
−
15.8
5.2
−
4m
4n
5ª
SCH3
CH3
H
19.2 ± 3.5
>100
−
−
−
−
5b

H
H
H
SCH3
B
B
B
0.90 0.84
0.89 0.79
0.68 0.90
14.4 ± 5.1
18.5± 4.1
8.4 ± 2.1
24 ± 5.6
16.4 ± 5.1
6.4 ± 3.3
45.1 ± 5.2
32.3 ± 4.3
12.3 ± 3.5
98.0 ± 8.9
88.7 ± 8.3
>100
−
−
−
−
−
−
−
−
−
−
5c
5d
5e
C(CH3)3
(CH2)3CH3
3.6
60.4
H
B
0.87 0.88
>50
26.4 ± 3.9
−
−
−
−
−
−
5f
H
H
H
B
D
D
0.92 0.87
0.92 0.98
0.84 0.99
23.6 ± 4.1
14.5 ± 3.3
36.6 ± 4.4
>50
35.1 ± 5.1
>50
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
−
5g
5h
5i
H
E
0.83 0.87
>50
>50
−
−
−
−
−
−
13
Br
I
H
H
E
E
0.87 0.94
0.86 0.93
22.3 ± 5.1
14.6 ± 3.5
32.1 ± 4.2
21.4 ± 3.1
−
−
−
−
−
−
−
14
15
13.4 ± 3.1
>100
>6.8
>4.7
>7.5

H
F
0.88 0.94
32.5 ± 5.2
20.3 ± 3.6
−
−
−
−
−
−
16
H
H
G
H
0.91 0.94
0.86 0.95
>50
>50
>50
−
−
−
−
−
−
−
−
−
−
−
−
17
18
49.5 ± 6.7
169
170
171
aPercent of inhibition of thioredoxin reductase activity at 100 µM; bSelectivity index calculated by astrocyte CC50/glioma lineage EC50; The data are expressed as mean ± SD of
three independent assays. Dashed SI values means that cytotoxicity against astrocytes cannot be calculated because compounds did not show significant cytotoxicity even at
highest concentrations used in the assay.
172

173
174
175
176
177
178
179
180
181
182
183
184
185
186
The most promising compound was the nitrofuran analog (2E)-1-(3-methylphenyl)-3-(5-nitrofuran-2-
yl)prop-2-en-1-one (4m) with EC50 of 2.1 µM, 2.3 µM and 3.3 µM for C6, U251MG and U87MG cells,
respectively. The compounds 4c, 4e, 4n, and 4k were the most active against C6 cells (EC50 ~1.5), but
lost activity against U251MG and U87MG cells in comparison to 4m. The positive control TMZ had
activity against C6 cells with an EC50 value of 60.46 µM, and showed minimal activity in U251MG and
U87MG cells at 200 µM [40], probably because TMZ requires a slightly basic pH to spontaneously
converts to a bioactive methylating agent [41]. Concentration-response curves used for EC50
determination of 4e, 4m, and 4n are shown in Figure 3A. Considering the lipophilicity (LogP calculate
~3.6 for 4m and 4n, and 4.8 for 4e) of test compounds, we performed cytotoxicity curves in C6 cells in
the presence of varying Triton X-100 concentrations (0.001%, and 0.0001%). At 72 h treatment, 4e, 4m
and 4n cytotoxicity (EC50) was similar to C6 cells without detergent thereby excluding a possible
promiscuous colloidal aggregate effect (Figure 3B) as well as indicating that cytotoxicity of the hit
compounds is related to specific inhibition. It is noteworthy that Triton-X100 0.01% killed all C6 cells,
thus making the EC50 estimation unfeasible at this concentration (data not shown).

187
188
Figure 3. (A) Viability concentration-responses curves for compounds 4e, 4m, and 4n against C6,
U251MG, U87MG, and astrocytes after 48h of incubation. (B) Concentration-response curves for test
compounds in the presence of Triton X-100 in C6 cells (72 h treatment, sulforhodamine B (SRB) assay).
The data are expressed as mean ± SD of three independent assays. * Different from untreated cells
(p<0.05, ANOVA).
189
190
191
192
193
194
195
196
197
198
2.5. In vitro Cytotoxicity assays in astrocytes
The hit compounds were also evaluated for their cytotoxicity in vitro against astrocytes, as a control
for normal glial cells. These compounds demonstrated modest to moderate cytotoxicity on this assay
(see Table 3 and Figure 3), with selectivity indexes (SI) ranging between 1.8 and 32.5. The compound
4m showed the most promising cytotoxicity profile, with SI of 24.8, 22.7 and 15.8 for C6, U251MG and
U87MG cells, respectively.

199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
2.6. Inhibition of thioredoxin reductase (TrxR) activity and involvement in thiol homeostasis
Since 5-nitrofuran chalcones display potent antiproliferative activities, we then asked the possible
cellular target of these in glioma cells. Chalcones contain an α,β-unsaturated ketone moiety, a key
structure for many reported TrxR inhibitors [42,43]. Considering this key feature, we made
computational predictions with these molecules using binary ML models developed in-house for TrxR
(data not shown). As a result, all investigated compounds were predicted as inhibitors this enzyme. In
view of this, the most promising compounds were tested in vitro against TrxR. As shown in Table 3,
compound 4m showed the greatest TrxR inhibitory activity (95.2% at 100 µM), followed by 4e (80.7%)
and 4n (78.7%). In addition, compounds 4e and 4m showed dose-dependent decrease of TrxR activity,
with IC50 values ~25 µM.
Compounds with α,β-unsaturated carbonyl system (i.e. chalcones) have been reported to form
covalent bonds with cysteines [44]. Consequently, it has been suggested that chalcones are pan-assay
interference compounds (PAINS) [45] due to reactivity under assay conditions. In order to evaluate the
role of thiol residues in 4n, 4m and 4e cytotoxicity, we pre-incubated C6 glioma cells with 2 mM N-
acetyl-cysteine (NAC) for 1 h prior to test compounds treatments. The classical TrxR inhibitor,
auranofin (Au), was used as a positive control. Cell viability assays showed that NAC abrogated the
cytotoxicity of our test compounds as well as auranofin (Fig. 4A). Using a cell-free in vitro system, we
observed that 4n, 4m and 4e (10 to 500 µM) were not alkylating agents in presence of NAC and bovine
serum albumin (BSA) at physiological conditions (pH 7.4, 37 ºC) (Figures 4B and 4C). The alkylating
agent N-ethylmaleimide (NEM) was used as a positive control for cysteine alkylation, leading to thiol
depletion (R-SH) at 50 µM (Figs. 4B and 4C). These results indicate that 4n, 4m and 4e are PAINS
even though thiol homeostasis and inhibition of TrxR may play a role, at least in part, in 5-nitrofuran
chalcones cytotoxicity.

223
224
225
226
227
228
229
230
Figure 4. (A) Effect of NAC pretreatment (2 mM, 2 h) on 4e, 4m, 4n and auranofin (Au) cytotoxicity in
C6 glioma cells incubated for 72 h as assessed by SRB assay. (B-C) In vitro reactivity of 4e, 4m and 4n
with reduced thiol residues (R-SH) of NAC and BSA in cell-free assay as determined by 5,5'-dithiobis-
2-nitrobenzoic acid (DTNB) assay. In B and C, the thiol-alkylating agent NEM was used as a positive
control for thiol alkylation/depletion. *different from untreated cells or control group (NAC or albumin
alone, in B and C graphs); #different from untreated and test compound-treated cells at equivalent
concentrations (1-way ANOVA, post-hoc Tukey; p<0.05).
231
232
233
234
235
236
237
2.7. Structure-activity relationships (SAR)
Based on the experimental results, we derived structure-activity relationships (SAR) rules to reveal
the molecular substituents favorable and unfavorable for anti-glioma activity (Figure 5). The information
revealed by the SAR allowed us to derive the following rules: (i) halogen atom and methyl in R1
position increases the activity; (ii) aliphatic six-member rings and hydrophobic groups with primary to
tertiary carbons in R2 position increases the activity; (iii) 5-nitrofuran ring in R3 position increases the
activity; (iv) aromatic rings or hydrophobic groups with quaternary carbons in R2 position decreases the

238
239
activity; (v) 5-nitrothiophene, furan, 5-chlorothiophene, and six-member aromatic rings in the R3
position decreases the activity in glioma cells.
240
241
Figure 5. Derived SAR rules for chalcones with antiproliferative activity against glioma cells.
Substituents inside green boxes increase the activity, whereas substituents in red boxes decrease the
activity.
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
2.8. Synthesis of lead candidates
The most promising hit compounds (i.e., 4e, 4m, and 4n) identified in antiproliferative assay were
selected for in vivo testing. Considering the insufficient amount of compound mass for in vivo assays, we
resynthesized them again in larger scale using Claisen-Schmidt condensation using acetic acid as solvent
and sulfuric acid (H2SO4) as a catalyst [46]. For the resynthesis of desired chalcones in large quantities
required for in vivo testing it was necessary an optimization study and improvement of the yields
obtained in the methodology previously described by Dr. Andrade´s research group. Gomes and co-
workers [46] reported on previous study that standard condition under basic medium could not be used
because the starting materials are alkali-sensitive. Thus, the resynthesis of our test chalcones was carried
out in acidic medium (Table 4). Table 4 lists the optimized conditions employed for the resynthesis of
4n, under reflux at 100 ºC, reduced evaporate pressure (as previously reported), isolation with cooled
water, H2SO4 catalyst reduced, and microwave.
Table 4. Claisen-Schmidt condensation under different conditions.

Entry
Method*
Yield (%)
31%
Reaction time (h)
1
2
3
4
Reflux + reduced evaporate pressure
Reflux + isolation with cooled water
Reflux + 0.5 H2SO4 catalyst
Microwave
30
20
30
02
87%
56 %
37 %
257
258
*Conditions we employed for the resynthesis of compound 4n.
The protocol used in entry 1 was used in the previous work where after consumption of all starting
materials (monitored by TLC), the reaction mixture is concentrated under reduced pressure to remove
the acetic acid, thus obtaining the desired chalcone. The problem encountered with the increase in scale
is that a drastic decrease in yield is observed because of the higher amount of sulfuric acid causing the
degradation of the formed product. As we may notice from Table 4, under reflux at 100 ºC and pressure
reduced the reaction pathways had other disadvantages beside low yield (31%), the long reaction time
(30h) Under microwave irradiation (entry 4), the reaction time decreased substantially (2h), but yield
(37%) remained similar. Also explained by the presence of the acid medium is observed a large amount
of degradation of material. The reduction of the amount of sulfuric acid (entry 3) was not satisfactory
because there was no total consumption of the starting materials with the same reaction time as the entry
1. The protocol used in entry 2 was the better choice had the higher yield (87%), where after
consumption of all starting materials (monitored by TLC), the reaction mixture was poured into cooled
water, precipitating the desired chalcone. Taking into consideration the aspects mentioned above,
especially those related to yield, we decided to extend the protocol entry 2 for 4e and 4m resynthesis.
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
2.9. In vivo anti-glioma activity
Many prior in vitro studies have found potential candidates to treat gliomas, but most of them did not
include animal testing or failed in in vivo studies. In addition, most of prior studies use

275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
immunocompromised mice [47] and subcutaneous xenografts [48,49], which lack important components
of tumor microenvironment and blood barrier restrictions, respectively. In this study, we confirmed the
in vivo anti-glioma effects of resynthesized compounds (4e, 4m, and 4n) using a very aggressive model
of glioma in immunocompetent mice orthotopically implanted. The overall design of the study,
treatment groups, route of injection, and short-term/long-term treatment schedules are described in
Figure 6A. Briefly, we first implanted GL261 cells into the right hemisphere of mice brain. Seven days
after implantation, vehicle or test compounds were intraperitoneally or orally administrated every day
for 10 days. The alkylating agent temozolomide (TMZ) was used as a positive control for anti-glioma
activity in mice. During the study, the treated animals did not show any visible toxic effects or mortality
and had no significant difference in body weights compared to the negative control group (Figure 6B).
The treatment with compounds by gavage (Figure 6C) does not significantly decrease (p >0.05) the
tumor volume compared to negative control group (vehicle consisted of saline 0.9%, DMSO 10%, and
Tween-80 0.3%), indicating the poor oral absorption of these molecules. However, the tumor volume of
mice treated intraperitonially with 4m (100 mg/kg) and 4n (100 mg/kg) significantly reduced by 43.8%
and 41.3% (Table S1 and Figure 6D), respectively, compared to vehicle-treated mice. These results
indicate that the compounds are able to cross the BBB, confirming our ML predictions. In addition, no
significant differences in tumor volume were observed between 4m (100 mg/kg) and 4n (100 mg/kg)
treatment and the positive control group (TMZ at 20 mg/kg, p > 0.05).

294
295
296
297
298
299
300
301
302
Figure 6. Preclinical testing of 4e, 4m, and 4n in brain implanted GL261 gliomas. (A) Animal study
design: from treatment day 1 (7th protocol day), the test compounds were administered once a day for a
total 10 days. TMZ was administrated in alternate days (protocol days 7, 9, 11, 13 and 15th. (B) Delta
body weight change (final – initial) in C57BL/6 mice across the different treatments; (C) GL261 tumor
volumes (mm3) quantification of the mice treated by gavage at the 17th day; (D) GL261 tumor volumes
(mm3) quantification of the mice treated by intraperitoneal injection at the 17th day. Asterisks denote
differences from vehicle treated mice (*p<0.05; **p<0.01; ***p<0.001); &different from test compounds
at 10 mg/kg (1-way ANOVA, post-hoc Tukey).
303
304
305
306
307
2.10. Toxicity studies
For toxicity studies, the blood samples were collected at the end of treatment of glioma implanted
mice and various enzyme markers and blood parameters were estimated (Table 5). Treatment with 4m,
4e and 4n at 100 mg/kg did not alter the levels of enzymatic markers for hepatotoxicity, i.e. alanine
aminotransferase (ALT) and alkaline phosphatase (ALP), and nephrotoxicity (creatinine). The glucose

308
309
310
311
levels remained unchanged. In addition, these compounds did not promote hematological toxicity in
white blood cells counts and hematocrit when compared to vehicle-treated mice. These results are
encouraging since TMZ promotes toxicity through the significant reduction in the number of circulating
immune cells and increases in the levels of enzyme markers for hepatotoxicity and nephrotoxicity [50].
312
Table 5. Serum and blood markers of toxicity at the end of treatments in glioma bearing mice.
Concentration ± SD
Parameters
Untreated
7.6 ± 4.2
19.1 ± 8.9
0.4 ± 0.3
108 ± 13
8.6 ± 1.6
44.5 ± 1.8
4e
4m
4n
ALT (U/L)
19.8 ± 18.6
22.2 ± 8.7
0.3 ± 0.2
94 ± 14
6.2 ± 3.5
20.3 ± 7.8
0.4 ± 0.1
108 ± 16
8.4 ± 2.5
46.0 ± 2.5
9.9 ± 4.4
20.6 ± 6.1
0.5 ± 0.4
109 ± 15
9.2 ± 2.6
45.3 ± 2.9
ALP (U/L)
Creatinine (mg/dL)
Glucose (mg/dL)
WBC (x103/µL)
Hematocrit (%)
9.5 ± 1.3
45.5 ± 2.3
313
ALT: alanine aminotransferase; ALP: alkaline phosphatase; WBC: white blood cells.
314
315
316
317
318
319
320
321
322
323
324
3. Conclusions
We have developed robust machine learning models for the identification of new compounds able to
penetrate BBB and active glioma cells. The ML models were applied for virtual screening of our in-
house database of compounds. As a result, forty-one potential anti-glioma hits were prioritized and
tested in vitro against three glioma cell lines and astrocytes. Among them, compounds 4e, 4m, and 4n
were the best candidates from the chalcone series, presenting high potency at submicromolar range
(EC50 of 1.4–6.8 µM) and moderate cytotoxicity against astrocytes. Then, SAR rules revealed that
compounds containing halogen atom and methyl group in R1 position or aliphatic six-member rings and
hydrophobic groups with primary to tertiary carbons in R2; and containing 5-nitrofuran ring in R3
position were the most potent. Enzymatic assays indicated that inhibition of TrxR may be at least one of
the biological targets of 5-nitrofuran chalcones. In addition, orthogonal in vitro assays excluded the

325
326
327
328
329
330
331
332
possible promiscuous colloidal aggregate and alkylating effect of test compounds, indicating that the
cytotoxic effect of the hit compounds is not related to promiscuous assay-interference. Subsequently, we
confirmed the in vivo anti-cancer effects of resynthesized compounds (4e, 4m, and 4n) using mice
orthotopic glioma model. The treatment of mice with 4m and 4n efficiently decreased glioma growth
without promoting body weight reduction, death of animals, or altering hematological and toxicological
markers. To summarize, the machine learning models developed in this study allowed us to discover two
new lead compounds, which are new chemical scaffolds for developing novel anti-glioma drug
candidates.
333
334
4. Experimental section
4.1. Computational
335
336
337
338
339
340
341
342
343
344
4.1.1. Datasets
In this study, a dataset of compounds containing bioactivity data for C6 cell line was extracted from
ChEMBL database (https://www.ebi.ac.uk/chembl/; ID: CHEMBL614657) [28], while a dataset of
compounds with BBB penetration data was selected from a number of publications [29–32]. A brief
description of the datasets is presented below.
•
C6 dataset: 376 compounds with EC50 data. Based on a threshold of 10 µM, it consisted of 144
active compounds (EC50 ≤ 10 µM) and 232 inactive compounds (EC50 > 10 µM);
BBB dataset: 2,053 compounds with LogBB data. Based on a threshold of −1, it consisted of
1,570 BBB+ compounds (if LogBB ≥ −1: penetrate) and 483 BBB− compounds (if LogBB < −1:
not penetrate).
•
345
346
347
348
4.1.2. Data Curation
All chemical structures and correspondent biological information were carefully standardized using
Standardizer v.16.9.5.0 (ChemAxon, Budapest, Hungary, http://www.chemaxon.com) according to the
protocols proposed by Fourches and colleagues [51–53]. Briefly, specific chemotypes such as nitro

349
350
351
352
353
354
355
356
357
groups and aromatic rings were normalized. In addition, explicit hydrogens were added, whereas
organometallic compounds, mixtures, polymers, and salts were removed. Then, we performed the
analysis and exclusion of duplicates as follows: (i) if the reported outcomes of the duplicates were the
same (e.g. active vs active, inactive vs inactive, etc.), one entry would be retained in the dataset and the
other excluded; and (ii) if duplicates presented discordance in biological activity (e.g. active vs inactive,
BBB+ vs BBB-), both entries would be excluded. Consequently, 81 duplicates within the C6 dataset and
102 duplicates within the BBB were identified and removed from original datasets. Furthermore, a high
concordance was observed between duplicate records of C6 dataset (82.7%), and BBB dataset (85.3%),
revealing the high quality of these datasets.
358
359
360
361
362
363
364
365
366
4.1.3.Dataset Balancing and chemical space analysis
The curated datasets (C6: 97 actives and 173 inactives; BBB: 433 BBB− and 1436 BBB+
compounds) were balanced using a linear under-sampling approach [33]. The linear under-sampling
strategy calculates the Euclidean distances between each compound in majority class and whole set of
minority class are measured using k-nearest neighbor (k-NN) algorithm [54]. Then, the samples on
majority classes of C6 and BBB datasets were linearly extracted over the whole set by using k-distances
and used to generate balanced datasets (Supplementary Files S1 and S2, respectively). Finally, a
chemical space analysis of balanced datasets was generated combining PCA and MACCS keys and
employing the KNIME workspace v.3.2 [55,56].
367
368
369
370
371
372
373
4.1.4. Molecular fingerprints
Morgan and FeatMorgan fingerprints were calculated in the open-source cheminformatics software
RDKit (http://www.rdkit.org [57]) executed on Python v.3.6 (https://www.python.org) [58]. Both
fingerprints were generated with radius 2−4 and bit vector of 2,048 bits. Morgan and FeatMorgan are
circular fingerprints built by applying the Morgan algorithm to a set of user-supplied 2D chemical
structures [59,60]. The fingerprint generation process systematically records the neighborhood of each
non-hydrogen atom into multiple circular layers up to a stablished radius. The Morgan captures highly

374
375
376
377
378
specific atomic information enabling the representation of a large set of precisely defined structural
features [59], whereas FeatMorgan uses functional features (i.e., hydrogen-bond donor and acceptors,
aromatic, halogen, basic and acid groups) [61]. Subsequently, these atom-centered substructural features
are interpreted as indexes of bits in a huge virtual bit string. Each position in this bit string accounts for
the presence or absence of a specific fragment feature [59,60].
379
380
381
382
383
384
4.1.5. Machine learning models
ML models were developed using Random Forest algorithm implemented in Scikit-learn v.0.19.2
(http://scikit-learn.org/) package available on Python v.3.6. The grid search was done using 50−500
estimators (intervals of 25 trees), number of features (Morgan or FeatMorgan bits) ranging from 6.6% to
100% along the bit vector of 2,048 bits (Morgan and FeatMorgan). The Cohen’s kappa (
scoring function of the estimator.
κ) was used as
385
386
387
388
389
390
391
4.1.6. 5-fold external cross-validation (5FECV)
The full dataset of compounds was randomly divided into five subsets of equal size; then one of these
subsets (20% of all compounds) is set aside as an external validation set and the remaining four sets
together form the training set (80% of the full set). Models were built using the training set while the
compounds in momentary external set (fold) were employed to evaluation of predictive performance.
ML models were developed five times, allowing each of the five subsets to be used as a momentary
external validation set.
392
393
394
4.1.7. Performance of ML models
The predictive performance of ML models was evaluated using SE, SP, CCR, PPV, NPV and κ.
These metrics were calculated as follows:
TP
SE =
SP =
(1)
(2)
TP + FN
TN
TN + FP