Structural insights of SmKDAC8 inhibitors: Targeting Schistosoma epigenetics through a combined structure-based 3D QSAR, in vitro and synthesis strategy

Article abstract of DOI:10.1016/j.bmc.2017.02.020

A predictive structure-based 3D QSAR (COMBINEr 2.0) model of the Schistosoma mansoni lysine deacetylase 8 enzyme (SmKDAC8) was developed, validated and used to perform virtual screening (VS) of the NCI Diversity Set V database (1593 compounds). Three exte

Full text of DOI:10.1016/j.bmc.2017.02.020

Accepted Manuscript
Structural Insights of SmKDAC8 Inhibitors: Targeting Schistosoma Epigenetics
Through a Combined Structure-Based 3D QSAR, in vitro and Synthesis Strategy
Flavio Ballante, D. Rajasekhar Reddy, Nancy J. Zhou, Garland R. Marshall
PII:
S0968-0896(16)31413-4
DOI:
http://dx.doi.org/10.1016/j.bmc.2017.02.020
BMC 13548
Reference:
To appear in:
Bioorganic & Medicinal Chemistry
Received Date:
Revised Date:
Accepted Date:
11 December 2016
6 February 2017
9 February 2017
Please cite this article as: Ballante, F., Rajasekhar Reddy, D., Zhou, N.J., Marshall, G.R., Structural Insights of
SmKDAC8 Inhibitors: Targeting Schistosoma Epigenetics Through a Combined Structure-Based 3D QSAR, in
vitro and Synthesis Strategy, Bioorganic & Medicinal Chemistry (2017), doi: http://dx.doi.org/10.1016/j.bmc.
2
017.02.020
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.

Structural Insights of SmKDAC8 Inhibitors: Targeting Schistosoma Epigenetics Through a
Combined Structure-Based 3D QSAR, in vitro and Synthesis Strategy
§
§
§
§*
Flavio Ballante, D. Rajasekhar Reddy, Nancy J. Zhou, and Garland R. Marshall
§
Department of Biochemistry and Molecular Biophysics, Washington University School of
Medicine, St. Louis, MO 63110
*
Corresponding author E-mail: garlandm@gmail.com (Garland R. Marshall)
ABSTRACT
A predictive structure-based 3D QSAR (COMBINEr 2.0) model of the Schistosoma mansoni
lysine deacetylase 8 enzyme (SmKDAC8) was developed, validated and used to perform virtual
screening (VS) of the NCI Diversity Set V database (1593 compounds). Three external datasets
(with congeneric structures to those experimentally resolved in complexes by X-ray and
previously reported as SmKDAC8 inhibitors) were employed to compose and validate the most
predictive model. Two series characterized by 104 benzodiazepine derivatives (BZDs) and 60
simplified largazole analogs (SLA), recently reported by our group as human KDAC inhibitors,
were tested for their inhibition potency against SmKDAC8 to probe the predictive capability of
the quantitative models against compounds with diverse structures. The SmKDAC8 biochemical
results confirmed: 1) the benzodiazepine moiety as a valuable scaffold to further investigate
when pursuing SmKDAC8 inhibition; 2) the predictive capability of the COMBINEr 2.0 model

towards non-congeneric series of compounds, highlighting the most influencing ligand-protein
interactions and refining the structure-activity relationships. From the VS investigations, the first
4
0 top-ranked compounds were obtained and biologically tested for their inhibition potency
against SmKDAC8 and hKDACs 1, 3, 6 and 8. Among them, a non-hydroxamic acid
benzothiadiazine dioxide derivative (code NSC163639), showed interesting activity and
selectivity against SmKDAC8. To further elucidate the structure-activity relationships of
NSC163639, two analogs (herein reported as compounds 3 and 4) were synthesized and
biologically evaluated. Results suggest the benzothiadiazine dioxide moiety as a promising
scaffold to be used in a next step to derive selective SmKDAC8 inhibitors.
KEYWORDS
Electrophoretic Mobility Shift Assay (EMSA); Epigenetics; Histone Deacetylase (HDAC);
Lysine Deacetylase (KDAC); per-residue 3D QSAR (COMBINEr 2.0); Schistosoma mansoni;
Structure-based drug design (SBDD)
1
. INTRODUCTION
Schistosomiasis, or bilharzia, is the second most common parasitic disease after malaria,
infecting at least 258 million people worldwide and accounting for up to 200,000 deaths per
1
year. The disease is caused by parasites from the genus Schistosoma, with S. mansoni being the
most widely distributed species, leading to acute (e.g. abdominal pain, diarrhea and haematuria)
or chronic conditions (e.g. liver enlargement, kidney damage, bladder cancer and infertility).
Currently, the only recommended treatment involves mass drug administration of praziquantel
(Figure 1), which raises concerns about drug resistance as more than 61.6 million people were

1
treated with the drug in 2014. In fact, praziquantel-resistant strains along with reduced
efficiency of the drug have already been reported and, even if oxamniquine (Figure 1) is listed
for use when praziquantel treatment fails, development of novel therapeutic treatments targeting
S. mansoni is recommended. Parasites typically have many phenotypic forms depending on
environment and stage of their life cycle, and progression through these forms is determined by
epigenetic modulation of gene expression. By inhibiting epigenetic modulation, such as through
inhibition of lysine deacetylases (KDACs), adaptation to a new environment and/or progression
in the life cycle can be impeded/stopped.
Figure 1. Praziquantel and Oxamniquine structures.
1
.1 KDACs-KATs. Lysine deacetylases and acetyltransferases (KDACs and KATs, or generally
referred to as histone deacetylases (HDACs) and histone acetyltransferases (HATs),
respectively) are a class of enzymes found in bacteria, fungi, plants, and animals. Lysine
acetyltransferases (KATs) catalyze the acetylation of pendant lysine side chains of proteins, such
as histones, causing localized relaxation of chromatin and transcriptional activation of nearby
genes as well as many other proteins modulating intracellular signaling. KDACs catalyze the
deacetylation of acetylated histones and other proteins leading to transcriptional repression and
modulation of functional responses, thus both KDACs and KATs influence gene expression and
modulate molecular recognition and signal transduction (epigenetics). To date, a series of KDAC
2
inhibitors (KDACIs) have been identified for the treatment of cancer. The role of KDACs in

pathology is ubiquitous: in fact, besides oncology, KDACIs have shown potential therapeutic
3
4
5, 6
7
roles in addiction, asthma, cardiovascular disease, immunosuppression, neurodegenerative
8
, 9
10
11, 12
diseases,
sepsis and termination of viral latency,
among others. Several studies of
various KDACIs have validated their anti-proliferative and anti-parasitic activity on major
human parasitic diseases such as leishmaniosis, malaria, schistosomiasis, toxoplasmosis, and
1
3
trypanosomiasis. Treatment of schistosomes with generic KDAC inhibitors caused protein
acetylation and a dose-dependent mortality of schistosome larvae, schistosomula, and adult
1
4
worms. S. mansoni has three class I KDACs (SmKDAC 1, 3, and 8), all of which are expressed
throughout the parasite’s life-cycle stages. In particular, SmKDAC8 is specifically targeted
because its transcripts are the most abundant and the enzyme contains six relatively large
1
4
insertions in the protein sequence. This abundance of SmKDAC8 is interesting to note because
normal human KDAC8 transcripts are typically lower than those of hKDAC1 and hKDAC3,
1
5-17
although hKDAC8 expression is often significantly upregulated in the case of some cancers.
In the design of an effective therapeutic targeting parasitic disease, unwanted side effects should
be minimized by avoiding inhibition of the eleven human zinc-based KDAC isozymes.
Therefore, a major goal is to identify compounds with high selectivity for SmKDAC8 over the
human KDAC isoforms.
1
.2 SmKDAC8. The potential of SmKDAC8 as a therapeutic target has been validated through
both biochemical and in vivo assays. RNA interference (RNAi)-mediated down-regulation of
SmKDAC8 expression in schistosomula followed by their intravenous injection into mice and
harvesting of surviving worms 35 days later demonstrated a significantly reduced worm recovery
18
compared to mice infected with schistosomula treated with control dsRNA. The crucial roles of
zinc-dependent KDACs in schistosome biology were also confirmed with the use of small-

molecule KDAC inhibitors. Beside the well-known pan-inhibitors (e.g SAHA or
1
9-21
panobinostat),
combination of virtual screening and in vitro testing as well as structure-based drug design
Figure 2).
small molecule inhibitors of SmKDAC8 have been identified through a
2
2
23
(
2
0, 21
19
Figure 2. KDACs pan-inhibitors panobinostat
and SAHA, with SmKDAC8 co-
1
8, 23, 24
crystallized ligands (PDB codes: 4BZ5, 4BZ6, 4BZ7, 4BZ8, 4BZ9, 4CQF, 5FUE)
recently reported active SmKDAC8 inhibitors (13l and 13za).
and
2
3
A comparison of the active sites of SmKDAC8 and its human counterpart reveals several
differences that can be targeted for drug design. By superimposing the two protein structures it
was possible to highlight, according to the residue and secondary structure conservation, those
protein portions suggested for ligand-protein specificity (Figure 3). As shown in Figure 3B, the
active site is mainly characterized by conserved residues in contrast to the rest of the protein,
thus suggesting significant differences in protein adaptation and flexibility. A deeper analysis of
the binding pocket (Figure 3C) reveals key structural differences that represent a basis of

inhibitor selectivity for SmKDAC8 over hKDAC8. As anticipated above, most of the residues
characterizing the active site, or in the near proximity, are conserved (Figure 4A), suggesting
common ligand-protein recognition: mainly confirmed by the presence of conserved Asp100 and
Tyr341 (previously detected, by mutation experiments, to be critical for deacetylase activity in
1
8
both the human or Schistosoma KDAC8s). Conversely, the main differences are the
replacement of hKDAC8 Leu31, Ala32, Ile34, Lys36, Gly99, Met274, Cys275 and Ser276 by
Ser18, Pro19, Phe21, Asp23, Asn98, His292, Arg293 and Ile294 in SmKDAC8, respectively
(Figure 4B). By comparing the two experimentally resolved KDAC8 proteins complexed with
25
18
SAHA (Figures 4C and D; PDB codes: 1T69 and 4BZ6, from human and Schistosoma
respectively) it was possible to detect how the replacement of Met274 (hKDAC8) with His292
(SmKDAC8) is responsible of an H-bond formation with the SAHA amide NH group, thus
affecting its binding pose (mutation of His292 to Met292 or Ala292 doesn’t affect the
1
8
SmKDAC8 catalytic activity as previously reported). Another H-bond interaction (which is
missing in hKDAC8) involves Lys20 with the SAHA amide C=O group, with a consequential
establishment of a flipped-in versus flipped-out conformations for Lys20 and Phe151,
respectively. Thus, an H-bond interaction with His292 and Lys20 with a same chemical group
(i.e., the amide group in SAHA) could affect the conformation of Lys20 and Phe151, which
cannot occur with hKDAC8 due to the replacement of His292 by Met274.
To quantitatively rationalize the ligand-SmKDAC8 recognition, as well as predict the inhibitory
activity of novel SmKDAC8 inhibitors, several structure-based per-residue 3D QSAR PLS
models have been built and assessed by means of the COMBINEr 2.0 procedure (an enhanced
2
6, 27
version of the previously reported COMBINEr methodology).
The COMBINEr 2.0 protocol
permits a per-residue rationalization by computing ligand-residues interaction energies for each

ligand/enzyme complex, and allows to clarify those which should be magnified or avoided for a
better biological response (through numerical data, 2D and 3D plots) and predict the bioactivity
of novel/untested compounds.

2
5
18
Figure 3. Superimposed hKDAC8 (PDB 1T69) and SmKDAC8 (PDB 4BZ6). The zinc ion is
depicted as a grey ball. A: hKDAC8 (cyan colored), SmKDAC8 (orange colored); B: hKDAC8
and SmKDAC8 color-coded according to the percentage of conserved residues (gradient color
from blue: no residue conservation to yellow: conserved residue) are depicted as ribbons (left
side) and surfaces (right side), respectively. C: highlighted sequences (blue colored) with no
residue conservation in the proximity of the active site. The minimized hKDAC8 (1T69,
2
8
prepared as previously reported) and SmKDAC8 (4BZ6, see section 2.1.1) complexes were
used.

A
B
C
D
2
5
18
Figure 4. hKDAC8 (PDB 1T69, cyan colored) and SmKDAC8 (PDB 4BZ6, orange colored)
complexes. A: superimposed protein structures showing the conserved residues (depicted as
wire); B: superimposed protein structures showing the non-conserved residues (shown as stick);
C: SAHA in the hKDAC8 active site (PDB 1T69, Met274 is highlighted as stick); D: SAHA in
the SmKDAC8 active site (PDB 4BZ6, His292 is highlighted as stick). The zinc ion is depicted
as a grey ball.

Since a limited number of experimentally resolved complexes (a total of 7 SmKDAC8-inhibitor
complexes) were used as training/test sets for model composition and assessment, a protocol was
designed to 1) derive reliable docked inhibitors-SmKDAC8 complexes (already reported in
literature and characterized by congeneric structures compared to ones experimentally
determined by X-ray); 2) generate multiple COMBINEr 2.0 models characterized by different
training/test sets compositions to detect the one with highest predictability. Moreover, 104
2
9
30
benzodiazepine derivatives (BZDs) and 60 simplified largazole analogs (SLAs), recently
reported by our group as hKDACIs, have been biologically tested for SmKDAC8 inhibition and
used as further external test sets to evaluate the predictive capability of the selected COMBINEr
2
.0 model that considers diverse molecules interacting at the same receptor site. The selected 3D
QSAR model allowed rationalization and clarification of the role of the residues characterizing
the active site, as well as a useful tool to predict the activity of novel/untested compounds. Thus,
31
a virtual screening application with 1593 compounds characterizing the NCI Diversity Set V
3
1
database was performed. The NCI Diversity Set V was specially chosen to find, among a set of
diverse structures, novel promising scaffolds to be investigated in a next drug development step.
The first 40 top-ranked predicted compounds were ordered from the NCI DTP repository and
tested towards SmKDAC8 as well as hKDACs1, 3, 6 and 8 (Figure 5). This virtual screening
application, identified a non-hydroxamic acid-based derivative (code NSC163639) with
detectable activity and selectivity (as a racemic mixture) against the SmKDAC8 enzyme, a
promising new scaffold for the design of Schistosoma mansoni lysine deacetylase 8 inhibitors as
well as the benzodiazepine scaffold (as suggested from the biological evaluation of BZDs).
Binding mode analysis of the predicted best docked pose of NSC163639 suggested interactions
with Lys20 His188, His292, Tyr341, Phe216 essential for SmKDAC8 selectivity, even without

the presence of a strong zinc binding group (i.e. hydroxamic acid moiety). To confirm that: 1)
the effective activity and selectivity of NSC163639 is given by established interactions between
the benzothiadiazine moiety and the above mentioned SmKDAC8 residues, and 2) the sulfone
and nitrogen groups of the characterizing scaffold are not interacting with the zinc ion (as a
sulfone analog of an hydroxamic acid), a sulfone hydroxamic acid precursor and a simplified
analog of NSC163639 (herein coded as compounds 3 and 4, respectively) have been synthesized
and biochemically tested as suggested by preliminary docking analyses. Experimental results
confirmed the docked binding modes of 3 and 4 and the quantitative model (COMBINEr 2.0)
evidences, suggesting the benzothiadiazine dioxide moiety of NSC163639 as a promising
scaffold to be further investigated for development of new selective SmKDAC8 inhibitors as
antischistosomal drugs.
Figure 5. General procedure of COMBINEr 2.0 model validation and selection of NCI
compounds for testing.

2
2
2
. RESULTS AND DISCUSSION
.1 Molecular Modeling
.1.1 Preparation of SmKDAC8-Inhibitor Complexes. Seven SmKDAC8-inhibitor co-crystal
1
8
18
18
18
18
24
structures have been released by the PDB: 4BZ5, 4BZ6, 4BZ7, 4BZ8, 4BZ9, 4CQF,
2
3
and 5FUE. All of the relative PDBs have completely resolved sequences characterizing the
active site with non-terminal missing residues outside the active site. Thus, a SmKDAC8 protein
as template structure was modeled by homology using only the incompletely resolved protein
sequences outside the binding pocket, and energetically minimized (see section 3). Among the
available experimental structures, 4BZ5 chain C (4BZ5.C) has the highest number of resolved
residues, thus, it was selected as template structure to model the unresolved residues of the other
co-crystallized complexes (the determined portions were maintained fixed), by applying the
same homology and minimization protocols. In particular, 4BZ6.B, 4BZ7.A, 4BZ8.C, 4BZ9.C,
4
CQF.B, 4CQF.C and 5FUE.B were considered, with the lowest number of missing residues,
and unique conformations of the characterizing ligands (i.e. 4CQF.B and 4CQF.C were both
considered since the same ligand had a different binding pose in the active site). All the final
structures were finally aligned to a hKDAC8 co-crystal complex (1T64 chain A, similarly as
2
8
previously reported) to allow for a comparative analysis between the hKDACs and the
SmKDAC8 structures. Due to the limited number of experimental SmKDAC8-inhibitor
structures for the COMBINEr 2.0 model development (a total of 7), the relative narrow range of
biological response (≈2 log unit range for pIC ) and the presence of only one compound (PDB:
5
0
5
FUE.B) with a pIC (M) value greater than 6, we increased the number of training set
50
complexes employed to improve the domain applicability, while reducing the probability of
having overfitting and chance correlations from the PLS-based COMBINEr 2.0 models. After

rationalizing the effect of SmKDAC8 protein flexibility on the docking precision, a multi-
conformational docking procedure was developed, assessed (see section 2.1.3) and applied to
predict reasonable SmKDAC8-inhibitor complex structures of non-crystallized smKDAC8
inhibitors (reported in the literature) included in the training set and test set of compounds for 3D
QSARs development and validation. Previously validated as a docking program to predict the
28
32
binding conformations of mammalian KDAC inhibitors, Surflex-Dock was used for all
docking calculations.
2
.1.2 Accounting for Protein Flexibility. To quantify the impact of protein flexibility on docking
predictability, and confirm the necessity of implementing a multi-conformational docking
procedure, re-docking and cross-docking assessments (RD and CD, respectively), considering
randomly generated conformations of the input ligands (thus, RCRD and RCCD, respectively),
were first performed using the above reported series of 8 co-crystal complexes (4BZ5.C,
4
BZ6.B, 4BZ7.A, 4BZ8.C, 4BZ9.C, 4CQF.B, 4CQF.C and 5FUE.B). Then, best docked (BD),
best cluster (BC) and best fit (BF) poses where extrapolated and the relative docking accuracies
3
3
28
(
DAs) quantified by means of the Clusterizer and DockAccessor software.
Table 1. Docking accuracy from RCRD (Random Conformation Re-Docking) and RCCD
(
Random Conformation Cross-Docking).
a
b
c
BD
BC
BF
RCRD
RCCD
62.5
33.04
56.25
31.25
100
73.21
a
b
c
Best Docked pose; Best Cluster pose; Best Fit pose
BD, BC and BF poses extrapolation as well as the docking accuracy quantification were
2
8
performed by using the Clusterizer and DockAccessor software.
By comparing the RCRD and RCCD docking accuracy (DA) results (Table 1) the decrement of
the docking predictability was detectable when accounting for protein flexibility: in fact, re-

docking calculations proved a better precision of the applied docking protocol, which
dramatically decreases (lower overall cross-docking accuracy value, Table 1) when applying a
cross-docking procedure (docking each ligand in all the different protein confirmations, but not
the native one). Beside the effect of protein flexibility on the docking predictability, it should be
considered, also, that among the available X-ray complexes, 4BZ5.C is the one characterized by
a L(+)-tartaric acid as ligand (which showed no inhibitory activity) and showed a Tyr341 with a
flipped out conformation from the active site (contrarily to all other Tyr341 characterizing all the
inhibitor-SmKDAC8 X-ray complexes). Therefore, an analysis without considering 4BZ5.C
(
Table 2), was also performed consistent with its exclusion when developing the 3D QSARs
since the requirement of a Tyr341 with a flipped in conformation for the enzyme inhibition); not
(
much difference in results was obtained.
Table 2. Docking accuracy from RCRD (Random Conformation Re-Docking) and RCCD
(
Random Conformation Cross-Docking) without considering the results from 4BZ5.C.
a
b
c
BD
BC
BF
RCRD (w/o 4BZ5.C)
RCCD (w/o 4BZ5.C)
64.29
33.33
50
29.76
100
69.05
a
b
c
Best Docked pose; Best Cluster pose; Best Fit pose
BD, BC and BF poses extrapolation as well as the docking accuracy quantification were
performed by using the Clusterizer and DockAccessor software.
2
8
In such a situation, a multi-conformational docking procedure (docking a certain ligand in an
ensemble of representative protein structures and extrapolating the binding poses with the most
successful protein conformer) was also investigated to account for protein flexibility and to
improve the docking accuracy (see section 2.1.3).
2
.1.3 Multi-Conformational Docking Assessment. Considering the need of predicting a reliable
binding conformation of several crystallographically undetermined ligands (to include in the

training set of compounds during the 3D QSAR COMBINEr 2.0 model development, or used as
external test set for bioactivity prediction during a virtual screening application), and the
impracticability (due to the required time) of using all the crystallized protein structures, it was
decided to simplify the ensemble of protein conformations. To adequately account the effect of
protein flexibility, a set of representative experimental SmKDAC8 conformations (excluding
4
BZ5.C since it was characterized by a small ligand, as L(+)-tartaric acid and by a Tyr341 with a
flipped-out conformation) was selected and used as target structures. A total of three different
SmKDAC8 protein conformations (4BZ8.C, 4BZ9.C, and 5FUE.B) were chosen because of the
dissimilar coordinates characterizing essential residues in the active site. In particular, Lys20,
Asp100, Phe151 and Tyr341 had the highest conformational variability (Figure 6). To test the
effectiveness of that procedure, each co-crystallized ligand was docked in each of these 3
enzyme structures, then the first 100 lowest energy docked conformations (among all of those
derived from the used target structures) were collected, and BD, BC and BF poses were
extrapolated with the relative protein conformer structure. The computed docking accuracy
values confirmed the efficacy of the adopted procedure, i.e., the docking predictive capability
increased compared to that characterizing the overall cross-docking application (compare Table
3
5
with Table 1). Moreover, the three original protein conformations (4BZ8.C, 4BZ9.C, and
FUE.B) characterizing the ensemble were correctly detected as the best ones when docking
relative ligands (Table 4). Similarly as performed for RCRD and RCCD, an analysis was
performed without considering 4BZ5.C. Even in this case the results were not significantly
affected (compare Table 3 with Table 5).

Table 3. Docking Accuracy: RCMD (Random Conformation Multi-conformation Docking).
a
b
c
BD
43.75
BC
43.75
BF
93.75
RCMD
Best Docked pose; Best Cluster pose; Best Fit pose
BD, BC and BF poses extrapolation as well as the docking accuracy quantification were
a
b
c
2
8
performed by using the Clusterizer and DockAccessor software.
Table 4. Best protein conformation (from the ensemble characterized by 4BZ8.C, 4BZ9.C, and
5
FUE.B) predicted for each X-ray SmKDAC8 structure from docking the relative ligand.
Experimental Protein Conformation
Predicted Protein Conformation
4
4
4
BZ5.C
BZ6.B
BZ7.A
5FUE.B
4BZ8.C
5FUE.B
4BZ8.C
4BZ9.C
4BZ8.C
4BZ8.C
5FUE.B
a
4
4
BZ8.C
BZ9.C
a
4
4
CQF.B
CQF.C
a
5
FUE.B
a
SmKDAC8 protein structures composing the ensemble of target conformations.

A
4
BZ8
4BZ9
B
5FUE
Figure 6. SmKDAC8 protein. A: left side, superimposed 4BZ8.C, 4BZ9.C, and 5FUE.B protein
structures with clipped and capped active site surface; right side, focused view of the binding
site. B: Comparative active site view showing Lys20, Asp100, Phe151 and Ty341 residues
+
2
characterizing the three different protein conformations (4BZ8.C, 4BZ9.C, and 5FUE.B). Zn is
depicted as a grey ball.

Table 5. Docking Accuracy without considering the results from 4BZ5.C: RCMD (Random
Conformation Multi-conformation Docking).
a
b
c
BD
42.86
BC
42.86
BF
92.86
RCMD (w/o 4BZ5.C)
a
b
c
Best Docked pose; Best Cluster pose; Best Fit pose
BD, BC and BF poses extrapolation as well as the docking accuracy quantification were
2
8
performed by using the Clusterizer and DockAccessor software.
2
.1.4 Application of Multi-Conformational Docking. Once its effectiveness was verified, the
multi-conformational docking procedure was applied to predict the binding poses of
experimentally unresolved SmKDAC8 inhibitors to be included in the training set or used as the
test set for the COMBINEr 2.0 3D QSAR model development and assessment. Three sets of
previously reported SmKDAC8 inhibitors (for a total of 46 compounds, see Supplementary
2
2-24
material Table S1)
were selected (since they were biologically evaluated using a same assay
protocol by the same research group), and docked. Thus, a total of 46 complexes, each
characterized by the ligand’s best docked (BD) pose with the associate best protein conformer
(among 4BZ8.C, 4BZ9.C, and 5FUE.B) were obtained (all the predicted binding poses show a
distance ≤2Å, between the zinc ion and the closest ligand’s atom, see Supplementary material
Figure S1) and used to implement and test the COMBINEr 2.0 3D QSAR models.
2
.1.5 3D QSAR (COMBINEr 2.0) Model Implementation and Assessment. To implement and
maximize the predictive capability of the 3D QSARs (compared to the model characterized by a
training set of only 7 co-crystal structures and a limited bioactivity range), three different
protocols have been applied to build and assess several COMBINEr 2.0 models using different
training set and test set compositions (Figure 7). Procedure 1 (Figure 7A): a model characterized
by only X-ray structures as the training set (a total of 7, 4BZ5.C was not considered as reported
above) has been developed and assessed by using as external test set the 46 predicted complexes;

this model was then used as a reference (or baseline) model for comparison purpose. Procedure 2
3
4, 35
(Figure 7B): similarly as the maximum dissimilarity sampling method,
the second procedure
adds compounds that are most dissimilar to those already included in the training set with the
3
6
purpose to increase the represented descriptor space. Starting with considering the 7 co-
crystallized ligands (PDB codes: 4BZ6.B, 4BZ7.A, 4BZ8.C, 4BZ9.C, 4CQF.B, 4CQF.C and
5
FUE.B) as training set, for each docked compound (a total of 46 at the beginning, see
Supplementary material Table S1) a Tanimoto index of molecular fingerprint similarity was
computed respect each training set compound, then an average index was derived. The docked
compound showing the lowest average similarity index (thus the most dissimilar one) was then
included in the training set. The next similarity evaluation was then performed between the
remaining docked compounds and the new training set. That protocol has been iterated until the
first 23 most dissimilar docked ligands were included in the training set, while the others 23 were
maintained as external test set for model assessment (see Supplementary material Table S2).
Procedure 3 (Figure 7C): a 2-level (presence-absence) resolution IV fractional factorial design
(FFD) using all the 53 compounds (7 X-ray + 46 docked) as factors has been performed to
compose the training and test sets. Even if a complete foldover of a resolution IV design is often
3
7
unnecessary, the original design matrix (128 runs) was mirrored for a total of 256 runs.
According to its level (presence or absence) in the matrix, a compound was maintained in the
training set or moved to the test set to obtain a total of 256 training/test sets combinations. For all
the 3 procedures, three different kinds of ligand-residue interactions (molecular interaction field,
3
8
MIF) were computed using the AutoGrid module of AutoDock4 (DRY, ELE, STE, see Table 6
2
6
for probe-field specification), thus seven different combination from these fields (DRY, ELE,
STE, DRY+ELE, DRY+STE, ELE+STE, DRY+ELE+STE) were considered when building the

3
2
D-QSARs from each derived training/test sets combination, giving a total of 1806 COMBINEr
.0 models which were investigated. In the case of multiple probes, a Block Unscaled Weights
3
9
(
BUW) scaling has been applied to give each probe the same importance in the model.
Table 6. List of COMBINEr 2.0 energetic fields (probes).
PROBE
DRY
ELE
Energetic Field
Desolvation
Electrostatics
STE
Lennard-Jones (LJ) and hydrogen-bonding (HB)
3
8
Molecular interaction fields (MIFs) were computed with the AutoGrid module of AutoDock4,
as previously reported.
2
6
To perform the assessment and comparative evaluation of all the derived COMBINEr 2.0 models
a series of coefficients, generally used to quantify the inner and the external predictive capability
of a quantitative model, as: 1) the K-fold cross-validation coefficient using 5 random groups and
2
1
00 iterations (q KF5CV); 2) the standard deviation error of prediction coefficient for the K5FCV
2
40
(
SDEPKF5CV); 3) the external q coefficient evaluated by Consonni et al., 4) the predictive
F3
2
correlation coefficient (r pred); 5) the standard deviation error of prediction coefficient of the
4
1
2
2
external test set (SDEPEXT); and the Y-Scrambling (r and q YS) coefficients (Table 7) were
YS
considered for a consensus scoring to identify the most descriptive and predictive model; thus
discarding those characterized by a poor structure-activity correlation or over-fitted equations
2
(
high q values but low predictive capabilities).

A
B
C
Figure 7. Adopted protocols to derive different training/test set combinations. A total of 1806
COMBINEr 2.0 models were evaluated.

Table 7. Statistical coefficients’ formulas used to evaluate the COMBINEr 2.0 models.
Internal Validation
External Validation
ꢏ
ꢐꢑ
ꢏ
ꢘꢙꢐ
ꢆꢇ ꢃꢇ
ꢎꢀ
ꢗꢅ
ꢆꢇꢉꢊꢋꢌꢍꢈ ꢃꢇꢎ ꢚꢛꢜ
ꢀ
ꢅ
ꢈ
ꢉꢊꢋꢌꢍꢈ
ꢈꢒꢓ
ꢈ
ꢝꢞꢟ
ꢀ
ꢈꢒꢓ
ꢀ
ꢁ
ꢂꢃꢄ
ꢁꢂꢃ
ꢏ
ꢕꢖ
ꢏꢐꢑ
ꢐꢑ
ꢆꢇ ꢃꢄ ꢇꢔ ꢎꢀ
ꢀ
ꢅ
ꢗꢅ ꢆꢇ ꢃ ꢇꢔ ꢎ ꢚꢛꢜ
ꢈ
ꢈ
ꢟꢠ
ꢟꢠ
ꢎꢀ
ꢈꢒꢓ
ꢈꢒꢓ
ꢏ
ꢘꢙꢐ
ꢪ
ꢣꢫꢬ
ꢅ
ꢐꢑ
ꢎꢩ
ꢆꢇ ꢃꢇ
ꢅ
ꢆꢢ ꢤꢢ
ꢣ
ꢀ
ꢈꢒꢓ
ꢈ
ꢉꢊꢋꢌꢍꢈ
ꢆꢇ ꢃꢄ ꢇꢔ ꢎꢀ
ꢥꢦꢧꢨꢍꢣ
SDEP=ꢡ
ꢭ
ꢁꢂꢃꢄ
ꢉꢊꢋꢌ
ꢏꢘꢙꢐ
ꢏꢐꢑ
ꢅ
ꢈ
ꢈꢒꢓ
ꢘꢙꢐ
ꢪ
2
2
ꢅ
ꢆꢢ ꢤꢢ
ꢎꢩ
ꢣ
ꢥꢦꢧꢨꢍꢣ
ꢣꢫꢬ
Y-Scrambling: r and q
EXT=ꢡ
SDEP
YS
YS
ꢏꢘꢙꢐ
2
2
2
q : cross-validation coefficient; SDEP: standard deviation error of prediction for the q ; r YS:
average square correlation coefficient obtained from 100 iterations Y-scramblings; q YS: average
2
leave-one-out cross-validation correlation coefficient obtained from 100 iterations Y-
2
2
40
2
scramblings; q : external q coefficient evaluated by Consonni et al.,
r pred: predictive
F3
correlation coefficient; SDEPEXT: standard deviation error of prediction coefficient of the
external test set.
Among the evaluated 1806 models, the one showing the best descriptive/predictive balance was
obtained from the procedure 3, characterized by a combination of three fields (DRY+ELE+STE,
see Supplementary material Table S3 for BUW coefficients) and using two principal components
(2PCs, see Tables 8, 9 and 10 for training set, test set compounds and statistical results,
respectively). Statistical results from the other field combinations with the same training/test sets
composition are reported in Supplementary material Table S4.
Table 8. Selected COMBINEr 2.0 Model (DRY+ELE+STE): Training Set.
Best
Ligand
Structure
PDB
Code
Protein
Conformer
Structure
#
Original ID
pIC (M)
50
1
8
1
4BZ9.C
-
-
5.36
2
4
4
2
3
4CQF.B
4CQF.C
-
-
-
-
4.30
4.30
2

Table 8. Selected COMBINEr 2.0 Model (DRY+ELE+STE): Training Set.
Best
Ligand
Structure
PDB
Code
Protein
Conformer
Structure
#
Original ID
pIC (M)
50
2
4
4
-
04
5FUE.B
4.92
2
3
5
6
7
-
-
-
10a
4BZ8.C
5FUE.B
5FUE.B
5.97
5.44
5.80
2
3
10b
2
3
10c
10d
13b
2
3
8
9
-
4BZ8.C
5.03
2
3
3
-
-
5FUE.B
5FUE.B
6.93
6.72
2
1
0
13c
13d
2
3
1
1
1
1
2
3
-
-
-
5FUE.B
5FUE.B
4BZ8.C
6.75
7.17
6.82
2
3
13e
23
13f

Table 8. Selected COMBINEr 2.0 Model (DRY+ELE+STE): Training Set.
Best
Ligand
Structure
PDB
Code
Protein
Conformer
Structure
#
Original ID
pIC (M)
50
2
3
1
4
5
6
-
-
-
13g
5FUE.B
5FUE.B
5FUE.B
6.86
6.89
7.12
2
3
1
1
13h
23
13l
2
3
1
1
7
8
-
-
13m
5FUE.B
5FUE.B
6.97
6.84
23
13n
2
3
1
2
2
9
0
1
-
-
-
13s
5FUE.B
4BZ8.C
4BZ8.C
6.74
6.83
6.40
2
3
13t
2
3
13y

Table 8. Selected COMBINEr 2.0 Model (DRY+ELE+STE): Training Set.
Best
Ligand
Structure
PDB
Code
Protein
Conformer
Structure
#
Original ID
pIC (M)
50
2
3
2
2
3
-
-
13z
4BZ8.C
6.72
7.04
2
3
2
13za
4BZ8.C
SAHA-
COOH
2
2
4
5
-
-
24
4BZ9.C
5FUE.B
5.72
6.10
24
ST80
Tubastatin
2
2
6
7
-
-
22
5FUE.B
5FUE.B
5.83
5.52
A
2
2
Z59184147
Table 9. Selected COMBINEr 2.0 Model (DRY+ELE+STE): Test Set.
Best
Ligand
Structure
PDB
Code
Protein
Conformer
Structure
#
Original ID
pIC (M)
5
0
1
8
1
4BZ6.B
-
-
5.81

Table 9. Selected COMBINEr 2.0 Model (DRY+ELE+STE): Test Set.
Best
Ligand
Structure
PDB
Code
Protein
Conformer
Structure
#
Original ID
pIC (M)
50
1
8
2
4BZ7.A
-
-
5.50
1
8
3
4
5
4BZ8.C
5FUE.B
-
T5979345
13a
-
5.83
6.33
5.91
2
3
-
2
2
3
4
03
03
5FUE.B
6
7
-
-
5FUE.B
4BZ8.C
5.54
6.36
2
2
3
4
04
05
8
9
-
5FUE.B
5.85
2
3
-
-
10e
5FUE.B
5FUE.B
6.22
6.57
23
1
0
13i

Table 9. Selected COMBINEr 2.0 Model (DRY+ELE+STE): Test Set.
Best
Ligand
Structure
PDB
Code
Protein
Conformer
Structure
#
Original ID
pIC (M)
50
2
3
1
1
-
13j
5FUE.B
6.66
2
3
1
1
2
3
-
-
13k
4BZ8.C
5FUE.B
7.02
6.16
2
3
13o
13p
13q
23
1
4
-
5FUE.B
6.92
23
1
1
5
6
-
-
5FUE.B
4BZ8.C
6.54
6.52
2
3
13r
23
1
7
-
13u
4BZ8.C
6.42
2
3
1
1
8
9
-
-
13v
4BZ8.C
4BZ8.C
6.40
6.01
2
3
13w

Table 9. Selected COMBINEr 2.0 Model (DRY+ELE+STE): Test Set.
Best
Ligand
Structure
PDB
Code
Protein
Conformer
Structure
#
Original ID
pIC (M)
50
2
2
2
3
3
3
2
0
1
2
-
-
-
13x
14a
15a
5FUE.B
5FUE.B
4BZ8.C
6.63
5.09
6.57
2
2
2
3
2
3
-
16a
5FUE.B
6.31
2
4
2
2
4
5
-
-
Crebinostat
5FUE.B
4BZ8.C
6.05
5.60
24
HHA
2
2
2
6
-
K783-3816
5FUE.B
5.25

Table 10. Selected COMBINEr 2.0 model (DRY+ELE+STE): statistical results.
a
b
c
2
d
2
e
f
2
g
YS
2
h
Field
PC
q²K5FCV
SDEP
q
r
SDEPext
r
q
YS
F3
pred
DRY+ELE+STE
2
0.55
0.57
0.35
0.42
0.58
0.38 -0.40
a
b
Number of principal components. K-fold cross-validation coefficient using 5 random groups
and 100 iterations (K5FCV). Standard deviation error of prediction for the K5FCV. external
coefficient evaluated by Consonni et al.. Predictive correlation coefficient. Standard
deviation error of prediction coefficient of the external test set. Average square correlation
c
d
2
40 e
f
q
F3
g
h
coefficient obtained from 100 iterations Y-scramblings. Average leave-one-out cross-validation
correlation coefficient obtained from 100 iterations Y-scramblings.
The selected COMBINEr 2.0 model showed decent cross-validation coefficient (e.g. q²K5FCV and
2
SDEP values, Table 10), while the absence of any chance correlation (the negative q YS value,
Table 10) using a low number of considered principal components (2PCs), thus confirming to be
reliable to describe the structure-activity relationship of the characterizing training-set of
compounds. Being a per-residue 3D QSAR study, a COMBINEr 2.0 model can be inspected by
means of PLS coefficients and activity contribution plots (PLS coefficients x Field), giving
information about the general contribution of each residue to the PLS regression or the actual
contribution (favorable or detrimental) of each ligand-residue interaction to the biological
response (IC50 value in this case), respectively. Furthermore, a ‘standard deviation times
coefficient’ field (PLS x SD, obtained by weighting point-by-point the PLS coefficients with the
standard deviation values of energies) can be derived to better clarify the global importance of
each amino-acid residue. According to the relative field (Table 6) the significance of a PLS
coefficient can differ. The DRY field estimates the desolvation potential which is largely
dependent from the atomic charge; thus, in the case of polar atoms it counterbalances the
hydrogen bonding energy meaning that those polar atoms capable to form hydrogen bonds could
determine a slight contribution to the binding free energy (because of the replacement of an