From Cancer to Pain Target by Automated Selectivity Inversion of a Clinical Candidate

Article abstract of DOI:10.1021/acs.jmedchem.8b00140

Elimination of inadvertent binding is crucial for inhibitor design targeting conserved protein classes like kinases. Compounds in clinical trials provide a rich source for initiating drug design efforts by exploiting such secondary binding events. Considering both aspects, we shifted the selectivity of tozasertib, originally developed against AurA as cancer target, toward the pain target TrkA. First, selectivity-determining features in binding pockets were identified by fusing interaction grids of several key and off-target conformations. A focused library was subsequently created and prioritized using a multiobjective selection scheme that filters for selective and highly active compounds based on orthogonal methods grounded in computational chemistry and machine learning. Eighteen high-ranking compounds were synthesized and experimentally tested. The top-ranked compound has 10000-fold improved selectivity versus AurA, nanomolar cellular activity, and is highly selective in a kinase panel. This was achieved in a single round of automated in silico optimization, highlighting the power of recent advances in computer-aided drug design to automate design and selection processes.

Full text of DOI:10.1021/acs.jmedchem.8b00140

Subscriber access provided by Washington University | Libraries
Article
From Cancer To Pain Target By Automated
Selectivity Inversion Of A Clinical Candidate
Samo Turk, Benjamin Merget, Sameh Eid, and Simone Fulle
J. Med. Chem., Just Accepted Manuscript • DOI: 10.1021/acs.jmedchem.8b00140 • Publication Date (Web): 10 May 2018
Downloaded from http://pubs.acs.org on May 13, 2018
Just Accepted
“Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted
online prior to technical editing, formatting for publication and author proofing. The American Chemical
Society provides “Just Accepted” as a service to the research community to expedite the dissemination
of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in
full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully
peer reviewed, but should not be considered the official version of record. They are citable by the
Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore,
the “Just Accepted” Web site may not include all articles that will be published in the journal. After
a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web
site and published as an ASAP article. Note that technical editing may introduce minor changes
to the manuscript text and/or graphics which could affect content, and all legal disclaimers and
ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or
consequences arising from the use of information contained in these “Just Accepted” manuscripts.
is published by the American Chemical Society. 1155 Sixteenth Street N.W.,
Washington, DC 20036
Published by American Chemical Society. Copyright © American Chemical Society.
However, no copyright claim is made to original U.S. Government works, or works
produced by employees of any Commonwealth realm Crown government in the course
of their duties.

Page 1 of 37
Journal of Medicinal Chemistry
1
2
3
4
5
6
7
8
9
From Cancer To Pain Target By Automated
Selectivity Inversion Of A Clinical Candidate
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Samo Turk ‡, Benjamin Merget ‡, Sameh Eid ‡, Simone Fulle*
BioMed X Innovation Center, Im Neuenheimer Feld 515, 69120 Heidelberg, Germany
ACS Paragon Plus Environment

Journal of Medicinal Chemistry
Page 2 of 37
1
2
3
ABSTRACT
4
5
6
7
8
Elimination of inadvertent binding is crucial for inhibitor design targeting conserved protein
classes like kinases. Compounds in clinical trials provide a rich source for initiating drug
design efforts by exploiting such secondary binding events. Considering both aspects, we
shifted the selectivity of tozasertib, originally developed against AurA as cancer target,
towards the pain target TrkA. First, selectivityꢀdetermining features in binding pockets were
identified by fusing interactionꢀgrids of several key and offꢀtarget conformations. A focused
library was subsequently created and prioritized using a multiꢀobjective selection scheme that
filters for selective and highly active compounds based on orthogonal methods grounded in
computational chemistry and machine learning. 18 highꢀranking compounds were synthesized
and experimentally tested. The topꢀranked compound has 10,000ꢀfold improved selectivity
versus AurA, nanomolar cellular activity and is highly selective in a kinase panel. This was
achieved in a single round of automated in silico optimization, highlighting the power of
recent advances in computerꢀaided drug design to automate design and selection processes.
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Keywords: automated compound optimization, machine learning, kinases, selectivity, drug
design, small molecules
2
ACS Paragon Plus Environment

Page 3 of 37
Journal of Medicinal Chemistry
1
2
3
INTRODUCION
4
5
6
7
8
The focus of R&D efforts on kinases for more than two decades has resulted in many kinase
inhibitors with known bioactivity and safety profiles. These compounds provide a great
source for initiating new compound design efforts, as for instance many compounds inhibit
more than one molecular target and can have a therapeutic impact via so far unknown or
ignored mechanisms.1,2 The latter is particularly true for inhibitors that bind to the highly
conserved ATPꢀbinding pocket of kinases and thus, often bind unintentionally also to
secondary targets. Inhibiting multiple targets can result in beneficial synergistic therapeutic
effects, but also could lead to unwanted side effects.3
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Hence, specifically switching off the activity against undesired targets, while maintaining or
even improving the affinity to the key target, is of great importance, especially when treating
nonꢀlifeꢀthreatening diseases such as inflammation or pain. Here, we report the automated
design of improved inhibitors of the tropomyosin receptor kinase A (TrkA) by jointly
employing a de novo design platform and a multiꢀobjective selection scheme that considers
selectivity and activity aspects as predicted by novel in silico tools. The selection strategy
aimed to shift the selectivity of a particular kinase inhibitor from its original key target
towards another validated kinase and takes in addition the selectivity profile against the
kinome into account. The initial target−compound pair (i.e., TrkA and tozasertib) was chosen
by mining the available kinase profiling data 4–6 for kinases inhibitors that had entered clinical
trials but also inhibit nonꢀcancer targets. The corresponding protein binding sites were
subsequently analyzed with respect to selectivityꢀdetermining features to further prioritize
targetꢀcompound pairs (see Experimental Section for details). The original key target of
tozasertib is Aurora kinase A (AurA), whose inhibition has common adverse effects such as
neutropenia and hematological toxicities.7 In turn, the profiling data revealed that tozasertib
also inhibits TrkA, which is a validated drug target for cancer and pain.8 Thus, inverting the
3
ACS Paragon Plus Environment

Journal of Medicinal Chemistry
Page 4 of 37
1
2
3
4
5
selectivity of tozasertib from AurA towards TrkA will likely result in a new compound series
with reduced side effects and, hence is of potential interest for pain treatment.
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
RESULTS AND DISCUSSION
Identification of selectivity-determining features in TrkA
The selectivity hotꢀspots in the TrkA binding site were initially identified by fusing atomꢀ
based interaction energy grids of several TrkA, Aurora A and B structures into one contentꢀ
rich representation of target specific subꢀpockets (i.e., energy grids representing the
conformational flexibility of the key and offꢀtargets were calculated, respectively, and finally
fused to one representation via difference rules; see Experimental Section and Supplementary
Information for details including validation results). The resulting ‘selectivity grids’ highlight
three areas of interest for compound optimization: Two are favorable hydrophobic subꢀ
pockets for TrkAꢀselectivity adjacent to the gatekeeper residue and enclosed between the Asp
residue of the DFGꢀmotif and a Phe residue of the glycineꢀrich loop (Gꢀloop), respectively,
and one is an unfavorable pocket for TrkAꢀselectivity that overlaps with the cyclopropyl
moiety of tozasertib (Figure 1).
Virtual compound library design
The identified selectivityꢀdetermining areas subsequently guided the compound library design
where two parts of tozasertib were enumerated using commercially available drugꢀlike
fragments and retrosynthetic rules (Figure 2). The resulting compound set A contains 7404
possible modifications of the aminoꢀ5ꢀmethylpyrazole that forms hydrogen bonds with the
hinge and extends towards the gatekeeper residue, while set B contains 6326 possible
modifications of the Nꢀ(4ꢀaminothiophene)cyclopropylcarboxamide. It should be noted that in
set A the cyclopropylcarboxyamide moiety was removed to avoid potential intramolecular
4
ACS Paragon Plus Environment

Page 5 of 37
Journal of Medicinal Chemistry
1
2
3
4
5
6
7
8
9
steric clashes, and that in set B modifications extend towards the DFGꢀmotif but can
potentially also form interactions with the gatekeeper residue. The resulting library was
prioritized using a multiꢀobjective compound selection scheme that filters for selective and
highly active compounds (Figure 3). All employed prediction tools showed good prediction
power in an initial validation on projectꢀspecific data (see Supplementary Information for
details).
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Selectivity and activity optimization
Initially, binding poses of compounds in sets A and B were generated with Glide SP.9
Compounds were removed for reasons of either poor docking scores (> –7.5 kcal/mol), wrong
orientation, or lack of key interactions (Experimental Section). In the next step (Figure 3),
compounds with an unfavorable selectivity profile were filtered out. This was accomplished
via machine learningꢀbased activity prediction models10 that were used I) to remove
promiscuous compounds (i.e., predicted to be active at IC50 of 500 nM on ≥ 20 kinases) and
II) those that are predicted to be highly active (IC50 < 10 nM) on Aurora A, B, or C kinases.
The selectivity filtering was complemented by a structureꢀbased procedure employing the
TrkA−Aurora ‘selectivity grids’ for rescoring of docking solutions. The remaining
compounds in both sets (A: 592; B: 1145) were finally prioritized for highly active
compounds using two complementary machine learning (ML) technologies. All compounds
were evaluated by an ‘MMP/ML’ approach,11 which is trained on fragmentꢀbased Matched
Molecular Pairs (MMPs), and quantifies compound activity differences. In addition,
compounds in the A set interacting with the Phe gatekeeper were additionally evaluated by a
hybrid QM/ML pipeline. This pipeline is trained on highꢀlevel quantum mechanical (QM)
calculations to quantify ligand−gatekeeper interactions and rescores the top hits in a second
step with fragment molecular orbital calculations, taking the entire binding pocket into
consideration.12,13
5
ACS Paragon Plus Environment

Journal of Medicinal Chemistry
Page 6 of 37
1
2
3
4
5
6
7
8
9
The final selection of compounds was obtained in an automated fashion based on the scoring
ranks (i.e., without any manual selection beside synthesizability criteria). 18 of the top
prioritized compounds were synthesized and tested for inhibition on TrkA and AurA
(Supplementary Table S1-S2), nine each from sets A (a1ꢀa9) and B (b1ꢀb9). The synthesized
compounds were from the top ~50 compounds from sets A and B, respectively. This yield
was achieved by considering synthetic pathway and building block availability in the library
design.
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Experimental testing on TrkA and AurA
An initial screening via binding assays revealed that 4 out of 18 tested compounds (Table 1)
have an improved activity/selectivity profile compared to the starting compound tozasertib
(i.e. a1, a4, b7 and b8; Figure 3) and originate from both sets A and B. Compound a1 shows
the best overall improvement compared to tozasertib with slightly higher TrkA affinity and
almost complete loss of the AurA activity (Table 2). Compound a1 was highly scored with
both orthogonal ML methods used for the final ranking (ranked 1 with MMP/ML and 3 with
QM/ML). Predicted binding modes (Figure 1) indicate that, in contrast to tozasertib, a1 forms
an additional π−πꢀinteraction with the gatekeeper residue Phe in TrkA via the fluorophenyl
moiety and does not occupy the adjacent unfavorable area (occupied by the
cyclopropylcarboxamide group of tozasertib). This trend is consistent with reference 2 (i.e.,
tozasertib without cyclopropylcarboxamide), which shows improved Trka−AurA selectivity
(∆∆pKd = 0.9) and thus underscores the importance of avoiding this unfavorable area (Table
2). Compound a4 also has a slightly improved TrkA affinity but retains AurA binding. In
contrast to the cyclopentyl of a4, the polarized fluorophenyl of a1 seems to make less
favorable interactions with the Leu gatekeeper in AurA, possibly explaining the higher
selectivity of compound a1 compared to a4. Based on the primary screening results it is
apparent that in the A set all triazoleꢀcontaining compounds are inactive on both TrkA and
6
ACS Paragon Plus Environment

Page 7 of 37
Journal of Medicinal Chemistry
1
2
3
4
5
6
7
8
9
AurA (Supplementary Table S1), whereas modifications containing a pyrazole hinge binder
in general show activity against TrkA as well as TrkA−AurA selectivity at 100 nM screening
concentration. Interestingly, even a small extension toward the gatekeeper, such as the tertꢀ
butyl moiety in compound a8 (Table S1), leads to a complete loss of affinity to AurA. This
underlines that bulky moieties are less tolerated by the AurA binding pocket and confirms the
vicinity of the gatekeeper as selectivityꢀdetermining region.
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
The two compounds b7 and b8 are highly similar and both introduce an additional methyleneꢀ
linker between a sulfur and a paraꢀsubstituted phenyl ring (paraꢀfluoro in b7 and paraꢀmethyl
in b8). The phenyl rings in both compounds are predicted to stack with Phe521 from the Gꢀ
loop, potentially benefiting from a more favorable orientation provided by the additional
methyleneꢀlinker group compared to tozasertib. The higher activity of b7 against TrkA
indicates that the electronꢀwithdrawing fluoroꢀsubstitution of b7 increases the strength of the
stacking interaction, compared to the electronꢀdonating methylꢀsubstitution of b8. The
primary screening results of the B set overall indicate a positive effect of a methyleneꢀlinker
between the ligand scaffold and the terminal aryl as well as a paraꢀsubstitution on the phenyl
ring for TrkA−AurA selectivity (Table S2).
The best compound from each set was further characterized in a cellꢀbased functional assay
revealing low nanomolar cellular potency for both compounds (a1: IC50 = 26 nM; b7: IC50
=
23 nM). This indicates that the compounds are able to reach the intracellular kinase domain
and thus permeate through the cellular membrane.
7
ACS Paragon Plus Environment

Journal of Medicinal Chemistry
Page 8 of 37
1
2
3
Kinase profiling
4
5
6
7
8
To evaluate the overall selectivity profile of all four hit compounds, profiling on a panel of 92
wildꢀtype kinases and 5 additional variants was conducted at 100 nM screening concentration.
The kinase selectivity of compound a1 is very high, with only six kinases (excluding TrkA)
inhibited, leading to a selectivity score of S(35%) = 0.08 and S(10%) = 0.04 (Figure 4; Table
S3). The remaining three hit compounds show slightly lower kinome selectivity (S(35%) ≈
0.2; Figure S1). At a concentration of 1 µM, a1 still inhibits only 19 kinases, which is in line
with the initial filtering step via the in silico kinase profiling platform that has excluded
promiscuous compounds (Figure S2). It is also noteworthy that all four hit compounds are
inactive on AurB, which was part of the offꢀtarget ensemble in the selectivity optimization
scoring via the ‘selectivity grids’. Compound a1 displays the most unique binding profile and
is even more selective than the starting compound tozasertib.
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
The kinases inhibited by compound a1 include, like for tozasertib, FLT3 and RET (both
common cancer targets), but also the two other Trk isoforms TrkB and TrkC, rendering a1 a
panꢀTrk inhibitor with excellent selectivity against other kinases. Compared to a1, b8 is less
active on TrkB and TrkC and inactive on PDGFRb, pointing to the potential value of fusing
both modifications in the future. Finally, hematological toxicities such as thrombocytopenia,
anemia and neutropenia are typically associated with ABL and JAK2 inhibitors.14,15 Thus, it is
encouraging that compared to tozasertib, compound a1 is also inactive against all tested
ABL1 variants (i.e. two mutants and the (un)phosphorylated wild type) as well as Janus
kinases (Figure 4). Overall, the high kinome selectivity of a1, together with the fact that a1
is, compared to tozasertib, inhibiting fewer kinases with targetꢀassociated toxicity (i.e. AurA
and ABL1) makes a1 an interesting candidate for further studies. Many Trk inhibitor
scaffolds have been reported in publications and patent applications and are summarized in
excellent reviews on Trka inhibitors.8,16–19 It is noteworthy that the aminopyrazolylꢀ
8
ACS Paragon Plus Environment

Page 9 of 37
Journal of Medicinal Chemistry
1
2
3
4
5
6
7
8
pyrimidine scaffold in the presented compound series has been already reported previously in
the panꢀTrk inhibitor AZꢀ23.20 However, AZꢀ23 has a significantly different substitution
pattern on the pyrimidine ring (i.e. amino linker at position 2 and unsubstituted position 6).
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Role of TrkA for pain treatment and current status of TrkA inhibitor design
TrkA is a high affinity receptor for the Nerve Growth Factor (NGF) and widely expressed on
peripheral pain sensing neurons. Activation of TrkA can lead to pathogenesis of many
difficult to treat human pain conditions such as osteoarthritis and cancerꢀrelated types of pain.
Lossꢀofꢀfunction TrkA variants are associated with congenital insensitivity to pain.21
Furthermore, in vivo TrkA inhibition studies indicate pain suppression effects 22,23 underlining
the promise of TrkA as a pain target. Other selective panꢀTrk inhibitors already exist 24–27 and
show promise for the treatment of acute and chronic pain. Studies with an allosteric TrkA
inhibitor indicate that TrkA specific inhibitors can be expected to have an improved safety
profile since panꢀTrk inhibitors have a hyperphagic effect.8 Reaching isoform selectivity vs
TrkB and TrkC in the ATPꢀbinding site is difficult as only two residues differ among the three
isoforms (i.e. two Arg in TrkA are replaced by Lys in TrkB and TrkC)8 but might be tackled
in future improvements with modifications of the so far unexplored Nꢀmethylpiperazine
moiety that form e.g. cation−π interactions with one of the arginines.28
9
ACS Paragon Plus Environment

Journal of Medicinal Chemistry
Page 10 of 37
1
2
3
CONCLUSION
4
5
6
7
8
Starting from a compound that reached clinical trials for cancer indications, we have designed
a new series of TrkA inhibitors by employing a multiꢀobjective selection scheme that filters
for selective and highly active compounds. In agreement with the design objectives, the top
compound a1 is highly active on TrkA, overall highly selective in a kinase panel, and − most
importantly from a design perspective − has a 10,000ꢀfold improved selectivity against the
selected offꢀtarget AurA. Nanomolar cellular potency against TrkA further underlines the
potential value of a1 as an advanced hit compound for the treatment of acute and chronic pain
or other conditions resulting from abnormal TrkA activity such as inflammation and cancer.
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Promiscuous interactions between drugs and proteins can cause adverse effects.29 Thus, being
able to switch off undesired targets in a single round of automated in silico optimization
underlines the power of recent advances in computerꢀaided drug design technologies,
especially those employing machine learning techniques such as the virtual kinase profiling
platform10 or the MMP/ML approach.11 Noteworthily, compound a1 was the highestꢀranking
compound in the filtering pipeline (Supplementary Table S1). The initially employed
‘selectivity grids’ pointed to selectivityꢀdetermining features in the TrkA structure and
thereby guided the subsequent compound library design. Furthermore, they were used to filter
for TrkA−Aurora selective compounds. Other successful in silico-driven compound design
efforts considering pharmacological profiles include those that identified new targets30,31 or
guided compound optimization.29 The uniqueness of our study is that we navigated through
the highly conserved and large target class of kinases instead of seeking predictions across
more diverse protein classes. Overall, the discovery platform described herein is widely
applicable for compound optimization for target classes with common selectivity issues but
can be also generally employed for switching off undesired activities in order to reduce
adverse effects.32
10
ACS Paragon Plus Environment

Page 11 of 37
Journal of Medicinal Chemistry
1
2
3
4
5
EXPERIMENTAL SECTION
Computational Methods
6
7
8
9
Details about the employed computational methods are described below and related validation
results in the Supplementary Information. Kinome tree figures were created using KinMap.33
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Selection of initial target-compound pair. Mining kinase inhibitors in clinical trials
(obtained from the data sources ChEMBL and clinicaltials.gov) pointed to TrkA as drug
4
target for pain treatment, while mining bioactivity data of a curated ChEMBL v22 kinase
3,5,6
inhibitor panel and other kinase profiling data sets
tozasertib as promising starting compound.
pointed to the Aurora inhibitor
Identification of selectivity determining features in TrkA. PDB structures of TrkA, AurA,
and AurB kinases in the DFGꢀin state were used for the generation of the selectivity grids
(Table 3; i.e. 4 PDB structures of TrkA were used as the key target set, while 67 AurA and 2
AurB PDB structures constituted the offꢀtarget set). All kinase structures were aligned to
4pmt using PyMOL, considering only binding site residues (defined as all residues within 5 Å
of any bound ligand). Atomꢀbased energy grids were calculated for the aligned structures
using AutoGrid (version 4.2.5),34 a grid spacing of 0.375 Å, and covering all ligand atoms in
the PDB structures. Subsequently, energy grids were fused into single key and offꢀtarget grids
by taking the minimum (median) interaction energy value at each point across the respective
grid ensemble for scoring (for visualization). Extreme energy values from steric clashes were
capped at a maximum value corresponding to the absolute value of the most favorable
interaction energy. Finally, the key and offꢀtarget fused grids were combined using fusion
rules.35 An aromatic carbon grid representation (A probe) was used to identify selective
hydrophobic subꢀpockets, while the acceptor oxygen (OA), acceptor nitrogen (NA), and
donor hydrogen (HD) probes were used to identify polar subꢀpockets.
11
ACS Paragon Plus Environment

Journal of Medicinal Chemistry
Page 12 of 37
1
2
3
4
5
6
7
8
9
Virtual compound library design. A virtual compound library was designed by modification
of tozasertib via systematic replacements of 3ꢀmethylpyrazole (compound set A) and Nꢀ(4ꢀ
aminothiophene)cyclopropylcarboxamide (compound set B), using the MolPort building
block database as source of fragments. For set A, all primary amines connected to an aromatic
N with one further connection were extracted and further narrowed down to fragments with ≤
15 heavy atoms and exactly one amino group. For set B, aromatic thiole containing fragments
were combined with acyl chloride containing fragments. Here, aromatic thiols were initially
filtered to keep only those with ≤ 15 heavy atoms and one amino group while acyl chlorides
were filtered to keep only those with ≤ 10 heavy atoms and no aromatic rings. The resulting
fragments in set A were combined with the remaining part of tozasertib without the
cyclopropylcarboxamide moiety to avoid potential intramolecular steric clashes. All virtual
compounds were finally filtered using the following cutoffs: Molecular weight ≤ 600 Da, logP
≤ 5, number of hydrogen bond donors ≤ 5 and acceptors ≤ 12.
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Prediction of binding modes and structure-based screening via ‘selectivity grids’. The
four preꢀaligned TrkA PDB structures (Table 3) were processed using the Protein Preparation
Wizard in the Schrödinger Suite (release 2016ꢀ4, Schrödinger, LLC, New York, NY, 2016)
by adjusting protonation, optimizing hydrogens, and restricted geometry optimization using
the OPLS3 force field. Docking grids were generated for the four protein models using the
Glide program,9 including a hydrogen bond constraint to the backbone NH of the Met592
hinge residue. The ensemble docking run employed the standard precision protocol (Glide SP)
with enhanced conformational sampling and increasing the number of initial poses to 50000.
For each virtual ligand, only the best pose as judged by the Glide SP score, was kept for
subsequent analysis along with the corresponding protein configuration. To filter the virtual
library further, compounds were excluded if: (1) the top pose had a docking score > –7.5
kcal/mol, (2) it had a wrong binding orientation, e.g. the piperazine moiety is buried rather
12
ACS Paragon Plus Environment

Page 13 of 37
Journal of Medicinal Chemistry
1
2
3
4
5
6
7
8
9
than being solventꢀexposed as in tozasertib, or (3) it lacked key tozasertibꢀTrkA interactions;
namely, Phe589 contact, Hꢀbond to Met592, contact to Asp668 in the DFG motif, and a
favorable contact to either Arg654 or Arg673. Finally, the top poses of the remaining
compounds were reꢀscored by interpolating the energy values from the corresponding TrkAꢀ
Aurora selectivity grids to estimate their selectivity propensity.
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Filtering of selective compound via in silico kinase profiling platform. The employed
platform contains machine learningꢀbased activity prediction models for small molecules
which were trained on chemical fingerprints and a large and diverse data set of kinase
inhibitor data.10 The technology was used in a twoꢀstep filtering procedure: (1) to remove
promiscuous compounds and (2) to filter with respect to selected offꢀtarget activity.10 In the
first step, compounds were removed that were predicted to be active on ≥ 20 kinases with a
probability of ≥ 0.7 in highꢀquality models (132 models with AUC values ≥ 0.8; activity
threshold: IC50 = 500 nM). In the second step, prediction models were trained for Aurora
kinases A, B, and C employing an activity threshold of IC50 = 10 nM to remove compounds
that are highly active on these offꢀtargets. Compounds were finally removed that were
predicted to be active with a probability of ≥ 0.7 on either of the three Aurora models.
Prioritization of active compounds via MMP/ML and a QM/ML pipeline. Two
complementary machine learning technologies were employed for prioritizing active
compounds. (1) a MMP/ML approach11 which is trained on Matched Molecular Pairs (MMPs)
as descriptors, and considers the entire compound for ranking of compound in sets A and B
and (2) a hybrid QM/ML pipeline which is trained on quantum mechanical calculations and
used to optimize the Pheꢀgatekeeper interaction of compounds in set A.
MMP/ML was trained by obtaining all compounds with reported IC50, Ki or Kd measurements
against TrkA from ChEMBL.4,11 MMPs were extracted using an inꢀhouse implementation
13
ACS Paragon Plus Environment

Journal of Medicinal Chemistry
Page 14 of 37
1
2
3
37
based on retrosynthetic rules (BRICS36) and encoded as concatenated Morgan fingerprint
4
5
6
7
8
9
consisting of the static core that is shared between the two molecules forming an MMP and
both fragments F1 and F2 (describing the transformation).11 A regression model was trained
on 23,000 fragmentꢀbased MMPs using a Deep Neural Network to predict the change in
activity (ꢁpIC50) associated with each MMP. Model validation was done using a “new
fragments” scenario, where fragments in the individual test sets inside a 5ꢀfold cross
validation were new to the machine learning model (validation results see Supplementary
Information). The final prediction of change in activity was estimated by employing the
median ꢁpIC50 of all five prediction models.
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
For the hybrid QM/ML pipeline, highꢀlevel QM calculations at the B3LYPꢀD3/6ꢀ31+G**
level were used to calculate pair interaction energies (PIE) between the gatekeeper residue
and a training set of compounds consisting of (1) 30 compounds selected from set A and (2)
50 benzylꢀderivatives of tozasertib (Table S4). The benzylꢀderivatives cover several
substituents and heterocycles (e.g., halogens, alkyls, and pyridyls) to model various
substituent effects. For each compound, the PIE was calculated from the gatekeeperꢀ
compound dimer and the respective monomeric systems: ꢀꢁꢂ = ꢂꢃꢄꢅꢆꢇ − ꢈꢂꢉꢊꢋꢆꢌꢆꢆꢍꢆꢇ
+
ꢂꢎꢏꢅꢍꢏꢐꢑꢃꢒ. Binding modes for set A were obtained via docking calculation, while the
benzylꢀderivatives were modeled manually in the binding pocket starting from the tozasertib
binding mode and each minimized in the OPLS3 force field using Maestro from the
Schrödinger suite.38 QM calculations were conducted using GAMESSꢀUS.13 Fast Hartree–
Fock calculations at the HF/6ꢀ31+G** level were used to derive electronic features (dipole
moment, quadrupole moment, total selfꢀconsistent field energy) of the 80 ligands in the
training set and the compounds in set A. A Random Forest regression model predicting the
PIE from electronic ligand features was generated based on the training set using scikitꢀlearn39
and employed to prioritize the compounds from set A with respect to PIEs with the gatekeeper
14
ACS Paragon Plus Environment

Page 15 of 37
Journal of Medicinal Chemistry
1
2
3
4
5
residue. Hartree–Fock calculations and PIE predictions were applied on the interacting
fragment only, i.e., not the entire ligand. The top 50 compounds were further characterized
6
7
8
12
considering the entire binding pocket by fragment molecular orbital (FMO) calculations at
9
the MP2/6ꢀ31(+)G* level (using docked binding poses and a polarizable continuum solvation
water model) and sorted according to the sum of PIEs. FMO calculations were set up using
Facio 40 and calculated using GAMESSꢀUS.13
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Chemistry
General synthetic route for tozasertib (aꢀd) and the compound sets A (a, iꢀj) and B (a, eꢀg) is
described in Scheme S1.a
a Reagents and conditions: (a) 3ꢀchloroperoxybenzoic acid, DCM, r.t., 3 h, yield: 95%; (b) Nꢀ
(4ꢀmercaptophenyl)cyclopropanecarboxamide, TEA, CH3CN, 80 °C, 3ꢀ10 h, yield: 70%; (c)
5ꢀmethylꢀ1Hꢀpyrazolꢀ3ꢀamine, DIPEA, DMF, 95 °C, 16 h, yield: 44%; (d) amine (1ꢀ
15
ACS Paragon Plus Environment

Journal of Medicinal Chemistry
Page 16 of 37
1
2
3
4
5
6
7
8
9
methylpiperazine or morpholine), DMF, DIPEA, 90 °C, 6ꢀ12 h, yield 43%; (e) corresponding
thiol, TEA, CH3CN, 80 °C, 3ꢀ10 h, yields: 20ꢀ85%; (f) 5ꢀmethylꢀ1Hꢀpyrazolꢀ3ꢀamine,
DIPEA, dioxane, 95 °C, 3ꢀ6 h, yields 50ꢀ75%; (g) amine (1ꢀmethylpiperazine or morpholine),
DMF, DIPEA, 90 °C, 6ꢀ12 h, yields: 30ꢀ70%; (h) thiophenol, TEA, THF, 50 °C, yields: 20ꢀ
85%; (i) corresponding amine, DIPEA, dioxane, 95 °C, 3ꢀ6 h, yields: 50 – 75%; (j) amine (1ꢀ
methylpiperazine or morpholine), DMF, DIPEA, 90 °C, 6ꢀ12 h, yields: 30 – 70%.
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Final yields: tozasertib: 31%, reference 2: 41%, a1: 49%, a2: 37%, a3: 51%, a4: 31%, a5:
38%, a6: 51%, a7: 64%, a8: 34%, a9: 48%, b1: 58%, b2: 41%, b3: 53%, b4: 48%, b5: 51%,
b6: 32%, b7: 39%, b8: 41%, b9: 49%. All compounds were synthesized under contract by
Enamine Ltd. Purification was performed using HPLC (H2O – MeOH; Agilent 1260 Infinity
systems equipped with DAD and massꢀdetectors. Waters Sunfire C18 OBD Prep Column,
100Å, 5 µm, 19 mm X 100 mm with SunFire C18 Prep Guard Cartridge, 100Å, 10 µm, 19
mm X 10 mm).
1
1
Compound characterization was done with H NMR and LC/MS. H NMR was performed
using Bruker AVANCE DRX 500 and Varian UNITYplus 400. LC/MS was performed using
Agilent 1100 Series LC/MSD system with DAD\ELSD and Agilent LC\MSD VL (G1956A),
SL (G1956B) massꢀspectrometer and Agilent 1200 Series LC/MSD system with DAD\ELSD
and Agilent LC\MSD SL (G6130A), SL (G6140A) massꢀspectrometer. Both systems used
Zorbax SBꢀC18 1.8 µm 4.6x15mm Rapid Resolution cartridge. Mobile phase was А –
acetonitrile, 0.1% formic acid, and В – water (0.1% formic acid) with gradient (0 min – 100%
B, 0.01 min – 100% B, 1.5 min ꢀ 0% B, 1.8 min ꢀ 0% B, 1.81 min ꢀ 100% B) and flow rate of
3 ml/min. Detection was performed using APCI ionization mode and scan range of m/z 80 ꢀ
1000. The purities of all final products of the two reference compounds (tozasertib, reference
2) and of 14 out of 18 reported compounds were found to be >95%. The remaining 4
16
ACS Paragon Plus Environment

Page 17 of 37
Journal of Medicinal Chemistry
1
2
3
compounds (i.e. a2, b2, b4, b8) have a purity between 90 and 95%. All reported compounds
4
5
passed common PAINS flags,41 as calculated using RDKit.
6
7
8
9
Experimental Assays
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
The primary screen (TrkA at 10 and 100 nM; AurA at 100 nM and 1 µM concentration), Kd
measurements, profiling over a panel of 97 kinases (tozasertib, a1, a4, b7, b8 at 100 nM; a1
additionally at 1 µM concentration), and cellular assay (a1 and b7) were done using the
42
KINOMEscan and PathHunter technologies from DiscoverX, respectively. Results of the
primary screen were reported as % of control (% Ctrl. = (test compound signal − positive
control signal) / (DMSO signal – positive control signal). The profiling was done via the
scanEDGE assay panel whereat KIT(D816V) and KIT (V559D, T670I) were replaced by
TrkB and TrkC. See www.discoverx.com for details on the employed profiling assays.
Selectivity scores (number of inhibited kinases divided by the total number of tested wildꢀtype
kinases) were calculated at activity cutꢀoffs of 35% and 10% of control, respectively.
17
ACS Paragon Plus Environment

Journal of Medicinal Chemistry
Page 18 of 37
1
2
3
ASSOCIATED CONTENT
4
5
6
Supporting Information
7
8
9
The Supporting Information is available free of charge via the Internet at http://pubs.acs.org.
It contains Figures S1ꢀS2 with profiling results, Table S1ꢀS4 with results of primary screens,
selectivity scores of top hits, the data set used to train the QM/ML approach as well as
additional information about the validation results of the employed prediction tools, about the
biological assays and assay results, and analytic data. Molecular formula strings for all
chemical structures mentioned (CSV) and the predicted binding pose of a1 (PDB) are
provided as data files.
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
AUTHOR INFORMATION
Corresponding Author
*
Eꢀmail: fulle@bio.mx
ORICID
Samo Turk: 0000ꢀ0003ꢀ2044ꢀ7670
Benjamin Merget: 0000ꢀ0002ꢀ3882ꢀ596X
Sameh Eid: 0000ꢀ0002ꢀ4689ꢀ6354
Simone Fulle: 0000ꢀ0002ꢀ7646ꢀ5889
Author Contributions
The manuscript was written through contributions of all authors. All authors have given
approval to the final version of the manuscript. ‡The authors ST, BM, and SE contributed
equally to this work as a first author.
Funding Sources
The project was supported by the BioMed X Innovation Center, Heidelberg.
Notes
18
ACS Paragon Plus Environment

Page 19 of 37
Journal of Medicinal Chemistry
1
2
3
The authors declare no competing financial interest.
4
5
ACKNOWLEDGMENT
6
7
8
We thank Christian Tidona and Ann De Beuckelaer (BioMed X) for fruitful discussions.
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
ABBREVIATIONS
TrkA: Tropomyosin receptor kinase A; AurA: Aurora kinase A; ML: Machine Learning;
MMPs: Matched Molecular Pairs; QM: Quantum Mechanics
REFERENCES
(1)
Duongꢀly, K. C.; Devarajan, K.; Liang, S.; Horiuchi, K. Y.; Wang, Y.; Ma, H.;
Peterson, J. R.; Duongꢀly, K. C.; Devarajan, K.; Liang, S.; Horiuchi, K. Y.; Wang, Y.;
Ma, H. Kinase Inhibitor Profiling Reveals Unexpected Opportunities to Inhibit
DiseaseꢀAssociated Mutant Kinases. CellReports 2016, 14 (4), 772–781.
(2)
(3)
(4)
Oprea, T. I.; Overington, J. P. Computational and Practical Aspects of Drug
Repositioning. Assay Drug Dev. Technol. 2015,
Metz, J. T.; Johnson, E. F.; Soni, N. B.; Merta, P. J.; Kifle, L.; Hajduk, P. J. Navigating
the Kinome. Nat. Chem. Biol. 2011, (4), 200–202.
3 (6), 299–306.
7
Bento, A. P.; Gaulton, A.; Hersey, A.; Bellis, L. J.; Chambers, J.; Davies, M.; Krüger,
F. A.; Light, Y.; Mak, L.; McGlinchey, S.; Nowotka, M.; Papadatos, G.; Santos, R.;
Overington, J. P. The ChEMBL Bioactivity Database: An Update. Nucleic Acids Res.
2014, 42 (D1), D1083ꢀ1090.
(5)
(6)
Davis, M. I.; Hunt, J. P.; Herrgard, S.; Ciceri, P.; Wodicka, L. M.; Pallares, G.; Hocker,
M.; Treiber, D. K.; Zarrinkar, P. P. Comprehensive Analysis of Kinase Inhibitor
Selectivity. Nat. Biotechnol. 2011, 29 (11), 1046–1051.
Anastassiadis, T.; Deacon, S. W.; Devarajan, K.; Ma, H.; Peterson, J. R.
Comprehensive Assay of Kinase Catalytic Activity Reveals Features of Kinase
19
ACS Paragon Plus Environment

Journal of Medicinal Chemistry
Page 20 of 37
1
2
3
Inhibitor Selectivity. Nat. Biotechnol. 2011, 29 (11), 1039–1045.
4
5
6
7
8
9
(7)
(8)
Falchook, G. S.; Bastida, C. C.; Kurzrock, R. Aurora Kinase Inhibitors in Oncology
Clinical Trials: Current State of the Progress. Semin. Oncol. 2015, 42 (6), 832–848.
Norman, B. H.; McDermott, J. S. Targeting the Nerve Growth Factor (NGF) Pathway
in Drug Discovery. Potential Applications to New Therapies for Chronic Pain. J. Med.
Chem. 2017, 60 (1), 66–88.
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
(9)
Friesner, R. A.; Banks, J. L.; Murphy, R. B.; Halgren, T. A.; Klicic, J. J.; Mainz, D. T.;
Repasky, M. P.; Knoll, E. H.; Shaw, D. E.; Shelley, M.; Perry, J. K.; Francis, P.;
Shenkin, P. S. Glide: A New Approach for Rapid, Accurate Docking and Scoring. 1.
Method and Assessment of Docking Accuracy. J. Med. Chem. 2004, 47, 1739–1749.
(10) Merget, B.; Turk, S.; Eid, S.; Rippmann, F.; Fulle, S. Profiling Prediction of Kinase
Inhibitors: Toward the Virtual Assay. J. Med. Chem. 2017, 60 (1), 474–485.
(11) Turk, S.; Merget, B.; Rippmann, F.; Fulle, S. Coupling Matched Molecular Pairs with
Machine Learning for Virtual Compound Optimization. J. Chem. Inf. Model. 2017, 57
(12), 3079–3085.
(12) Kitaura, K.; Ikeo, E.; Asada, T.; Nakano, T.; Uebayasi, M. Fragment Molecular Orbital
Method: An Approximate Computational Method for Large Molecules. Chem. Phys.
Lett. 1999, 313 (3), 701–706.
(13) Schmidt, M. W.; Baldridge, K. K.; Boatz, J. A.; Elbert, S. T.; Gordon, M. S.; Jensen, J.
H.; Koseki, S.; Matsunaga, N.; Nguyen, K. A.; Su, S.; Windus, T. L.; Dupuis, M.; Jr, J.
A. M. General Atomic and Molecular Electronic Structure System. J. Comput. Chem.
1993, 14 (11), 1347–1363.
(14) Oliveira, M. Managing Common Toxicities with New Tyrosine Kinase Inhibitors.
CancerWorld 2015, No. December, 31–37.
(15) O’shea, J. J.; Kontzias, A.; Yamaoka, K.; Tanaka, Y.; Laurence, A. Janus Kinase
Inhibitors in Autoimmune Diseases. Ann. Rheum. Dis. 2013, 72 (2), 1–11.
20
ACS Paragon Plus Environment

Page 21 of 37
Journal of Medicinal Chemistry
1
2
3
4
5
(16) Wang, T.; Yu, D.; Lamb, M. L. Trk Kinase Inhibitors as New Treatments for Cancer
and Pain. Expert Opin. Ther. Pat. 2009, 19 (3), 305–319.
6
7
8
(17) McCarthy, C.; Walker, E. Tropomyosin Receptor Kinase Inhibitors: A Patent Update
9
2009 ꢀ 2013. Expert Opin. Ther. Pat. 2014, 7, 731–744.
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
(18) Bailey, J.; Schirrmacher, R.; Farrell, K.; BernardꢀGauthier, V. Tropomyosin Receptor
Kinase Inhibitors: An Updated Patent Review for 2010ꢀ2016 ꢀ Part I. Expert Opin.
Ther. Pat. 2017, 6, 733–751.
(19) Bailey, J.; Schirrmacher, R.; Farrell, K.; BernardꢀGauthier, V. Tropomyosin Receptor
Kinase Inhibitors: An Updated Patent Review for 2010ꢀ2016 ꢀ Part II. Expert Opin.
Ther. Pat. 2017, 7, 831–849.
(20) Thress, K.; MacIntyre, T.; Wang, H.; Whitston, D.; Liu, Z.ꢀY.; Hoffmann, E.; Wang,
T.; Brown, J. L.; Webster, K.; Omer, C.; Zage, P. E.; Zeng, L.; ZweidlerꢀMcKay, P. A.
Identification and Preclinical Characterization of AZꢀ23, a Novel, Selective, and Orally
Bioavailable Inhibitor of the Trk Kinase Pathway. Mol. Cancer Ther. 2009,
8 (7),
1818–1827.
(21) Bakri, F. G.; Wahbeh, A.; Abu Sneina, A.; Al Khader, A.; Obeidat, F.; AlAwwa, I.;
Buni, M.; Ki, C.ꢀS.; Masri, A. Congenital Insensitivity to Pain and Anhydrosis due to a
Rare Mutation and That Is Complicated by Inflammatory Bowel Disease and
Amyloidosis: A Case Report. Clin. Case Reports 2016,
4 (10), 997–1000.
(22) Nwosu, L. N.; Mapp, P. I.; Chapman, V.; Walsh, D. A. Blocking the Tropomyosin
Receptor Kinase A (TrkA) Receptor Inhibits Pain Behaviour in Two Rat Models of
Osteoarthritis. Ann. Rheum. Dis. 2016, 57 (6), 1246–1254.
(23) Andrews, S. W. Allosteric Small Molecule Inhibitors of the NGF/TrkA Pathway, a
New Approach to Treating Inflammatory Pain. In International Association for the
Study of Pain World Congress on Pain; 2012.
(24) Wang, T.; Lamb, M. L.; Scott, D. A.; Wang, H.; Block, M. H.; Lyne, P. D.; Lee, J. W.;
21
ACS Paragon Plus Environment

Journal of Medicinal Chemistry
Page 22 of 37
1
2
3
4
5
6
7
8
9
Davies, A. M.; Zhang, H.ꢀJ.; Zhu, Y.; Gu, F.; Han, Y.; Wang, B.; Mohr, P. J.; Kaus, R.
J.; Josey, J. A.; Hoffmann, E.; Thress, K.; MacIntyre, T.; Wang, H.; Omer, C. A.; Yu,
D. Identification of 4ꢀAminopyrazolylpyrimidines as Potent Inhibitors of Trk Kinases.
J. Med. Chem. 2008, 51 (15), 4672–4684.
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
(25) Stachel, S. J.; Sanders, J. M.; Henze, D. A.; Rudd, M. T.; Su, H.ꢀP.; Li, Y.; Nanda, K.
K.; Egbertson, M. S.; Manley, P. J.; Jones, K. L. G.; Brnardic, E. J.; Green, A.;
Grobler, J. A.; Hanney, B.; Leitl, M.; Lai, M.ꢀT.; Munshi, V.; Murphy, D.; Rickert, K.;
Riley, D.; KrasowskaꢀZoladek, A.; Daley, C.; Zuck, P.; Kane, S. A.; Bilodeau, M. T.
Maximizing Diversity from a Kinase Screen: Identification of Novel and Selective Panꢀ
Trk Inhibitors for Chronic Pain. J. Med. Chem. 2014, 57 (13), 5800–5816.
(26) Burris, H. A.; Brose, M. S.; Shaw, A. T.; Bauer, T. M.; Farago, A. F.; Doebele, R. C.;
Smith, S.; Fernandes, M.; Cruickshank, S.; Low, J. A. A FirstꢀinꢀHuman Study of
LOXOꢀ101, a Highly Selective Inhibitor of the Tropomyosin Receptor Kinase (TRK)
Family. J. Clin. Oncol. 2015, 33 (15\_suppl), TPS2624ꢀTPS2624.
(27) Skerratt, S. E.; Andrews, M.; Bagal, S. K.; Bilsland, J.; Brown, D.; Bungay, P. J.; Cole,
S.; Gibson, K. R.; Jones, R.; Morao, I.; Nedderman, A.; Omoto, K.; Robinson, C.;
Ryckmans, T.; Skinner, K.; Stupple, P.; Waldron, G. The Discovery of a Potent,
Selective, and Peripherally Restricted PanꢀTrk Inhibitor (PFꢀ06273340) for the
Treatment of Pain. J. Med. Chem. 2016, 59 (22), 10084–10099.
(28) Bissantz, C.; Kuhn, B.; Stahl, M. A Medicinal Chemist’s Guide to Molecular
Interactions. J. Med. Chem. 2010, 53 (14), 5061–5084.
(29) Besnard, J.; Ruda, G. F.; Setola, V.; Abecassis, K.; Rodriguiz, R. M.; Huang, X.ꢀP.;
Norval, S.; Sassano, M. F.; Shin, A. I.; Webster, L. a; Simeons, F. R. C.; Stojanovski,
L.; Prat, A.; Seidah, N. G.; Constam, D. B.; Bickerton, G. R.; Read, K. D.; Wetsel, W.
C.; Gilbert, I. H.; Roth, B. L.; Hopkins, A. L. Automated Design of Ligands to
Polypharmacological Profiles. Nature 2012, 492 (7428), 215–220.
22
ACS Paragon Plus Environment

Page 23 of 37
Journal of Medicinal Chemistry
1
2
3
4
5
6
7
8
9
(30) Lounkine, E.; Keiser, M. J.; Whitebread, S.; Mikhailov, D.; Hamon, J.; Jenkins, J. L.;
Lavan, P.; Weber, E.; Doak, A. K.; Côté, S.; Shoichet, B. K.; Urban, L. LargeꢀScale
Prediction and Testing of Drug Activity on SideꢀEffect Targets. Nature 2012, 486
(7403), 361–367.
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
(31) Schneider, P.; Schneider, G. A Computational Method for Unveiling the Target
Promiscuity of Pharmacologically Active Compounds. Angew. Chemie Int. Ed. 2017,
56 (38), 11520–11524.
(32) Bowes, J.; Brown, A. J.; Hamon, J.; Jarolimek, W.; Sridhar, A.; Waldron, G.;
Whitebread, S. Reducing SafetyꢀRelated Drug Attrition: The Use of in Vitro
Pharmacological Profiling. Nat. Rev. Drug Discov. 2012, 11 (12), 909–922.
(33) Eid, S.; Turk, S.; Volkamer, A.; Rippmann, F.; Fulle, S. KinMap: A WebꢀBased Tool
for Interactive Navigation through Human Kinome Data. BMC Bioinformatics 2017,
18, 16.
(34) Morris, G. M.; Huey, R.; Lindstrom, W.; Sanner, M. F.; Belew, R. K.; Goodsell, D. S.;
Olson, A. J. AutoDock4 and AutoDockTools4: Automated Docking with Selective
Receptor Flexibility. J. Comput. Chem. 2009, 30 (16), 2785–2791.
(35) Reynolds, C. A.; Wade, R. C.; Goodford, P. J. Identifying Targets for Bioreductive
Agents: Using GRID to Predict Selective Binding Regions of Proteins. J. Mol. Graph.
1989,
7 (2), 103–108.
(36) Degen, J.; WegscheidꢀGerlach, C.; Zaliani, A.; Rarey, M. On the Art of Compiling and
Using “DrugꢀLike” Chemical Fragment Spaces. ChemMedChem 2008,
3 (10), 1503–
1507.
(37) Rogers, D.; Hahn, M. ExtendedꢀConnectivity Fingerprints. J. Chem. Inf. Model. 2010,
50 (5), 742–754.
(38) Schrödinger Release 2017ꢀ1: Maestro, Schrödinger, LLC, New York, NY, 2017.
(39) Buitinck, L.; Louppe, G.; Blondel, M.; Pedregosa, F.; Mueller, A.; Grisel, O.; Niculae,
23
ACS Paragon Plus Environment

Journal of Medicinal Chemistry
Page 24 of 37
1
2
3
4
5
6
7
8
9
V.; Prettenhofer, P.; Gramfort, A.; Grobler, J.; Layton, R.; VanderPlas, J.; Joly, A.;
Holt, B.; Varoquaux, G. API Design for Machine Learning Software: Experiences from
the ScikitꢀLearn Project. In ECML PKDD Workshop: Languages for Data Mining and
Machine Learning; 2013; pp 108–122.
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
(40) Masahiko Suenaga. Development of GUI for GAMESS / FMO Calculation. J. Comput.
Chem. Japan 2008,
7 (1), 33–54.
(41) Baell, J. B.; Holloway, G. A. New Substructure Filters for Removal of Pan Assay
Interference Compounds (PAINS) from Screening Libraries and for Their Exclusion in
Bioassays. J. Med. Chem. 2010, 53 (7), 2719–2740.
(42) Fabian, M. A.; Biggs, W. H.; Treiber, D. K.; Atteridge, C. E.; Azimioara, M. D.;
Benedetti, M. G.; Carter, T. A.; Ciceri, P.; Edeen, P. T.; Floyd, M.; Ford, J. M.; Galvin,
M.; Gerlach, J. L.; Grotzfeld, R. M.; Herrgard, S.; Insko, D. E.; Insko, M. A.; Lai, A.
G.; Lélias, J.ꢀM.; Mehta, S. A.; Milanov, Z. V; Velasco, A. M.; Wodicka, L. M.; Patel,
H. K.; Zarrinkar, P. P.; Lockhart, D. J. A Small MoleculeꢀKinase Interaction Map for
Clinical Kinase Inhibitors. Nat. Biotechnol. 2005, 23 (3), 329–336.
24
ACS Paragon Plus Environment

Page 25 of 37
Journal of Medicinal Chemistry
1
2
3
4
5
TABLES
Table 1: Summary of compound sets A and B.
6
Number of experimentally verified hits 1
# more selective & TrkA < 10 nM
7
8
9
Number of compounds
Set
# calculated
7404
# synthesized
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
9
9
4
7
2
2
A
B
6326
1 Number of compounds which are TrkAꢀselective vs. AurA [& which have in addition TrkA
activity < 10 nM in primary screen].
Table 2: Experimental affinity values of top hits.
TrkA
[pKd]
8.5
AurA
[pKd]
9.3
7.4
5.3
7.0
7.6
7.3
Selectivity
pKd]
Improved
selectivity [ꢀꢀpKd]
ID
[
ꢀ
Tozasertib
ꢀ0.8
0.1
3.3
1.8
1.4
1.1
ꢀꢀ
Reference 2 1
7.5
0.9
4.1
2.6
2.2
1.9
8.6
8.8
9.0
8.4
a1
a4
b7
b8
1 Tozasertib without cyclopropylcarboxamide. Reference 2 has an improved selectivity versus
tozasertib which is in line with the unoccupied unfavorable TrkAꢀselectivity area. Hit
compounds from sets A and B restore (or even improve) activity on TrkA without restoring
AurA activity.
25
ACS Paragon Plus Environment

Journal of Medicinal Chemistry
Page 26 of 37
1
2
3
Table 3: PDB structures used to generate ‘selectivity grids’.
4
5
6
Kinase PDB codes
# modelsa
7
8
9
TrkA
AurA
4aoj, 4pmtb, 4yne, 4yps
6
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
1mq4, 1ol5, 1ol6, 1ol7, 2c6d, 2dwb, 2np8, 2w1c, 2w1d, 2w1e, 2w1f,
2w1g, 2wtw, 2x6d, 2x6e, 2xne, 2xng, 2xru, 3d14, 3d15, 3d2i, 3d2k,
3dj7, 3e5ac, 3efw, 3fdn, 3h0z, 3ha6, 3k5u, 3m11, 3myg, 3nrm, 3o50,
3p9j, 3r21, 3r22, 3uo4, 3uo5, 3uod, 3up2, 3up7, 3vap, 4bn1, 4byi,
4c3p, 4c3r, 4ceg, 4dea, 4deb, 4ded, 4dee, 4dhf, 4j8m, 4j8n, 4jaj, 4jbo,
4jbp, 4o0s, 4o0u, 4o0w, 4prj, 4utd, 5aad, 5aae, 5aaf, 5aag, 5ew9
76
AurB
4af3, 4b8m
3
a Total number of models includes individual chains and alternate models present in the PDB codes.
b Used as reference structure for alignment.
c PDB of coꢀcrystallized tozasertib.
26
ACS Paragon Plus Environment

Page 27 of 37
Journal of Medicinal Chemistry
1
2
3
FIGURES
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Figure 1: Selectivity hot-spot areas in TrkA binding site. Predicted binding modes of (a)
tozasertib and (b) compound a1 to TrkA (PDB code 4YNE). Compound a1 occupies the
favorable hydrophobic subꢀpockets adjacent to the gatekeeper residue (green) but does not
reach into the unfavorable area (red) occupied by the cyclopropyl moiety of tozasertib. The
favorable subꢀpockets adjacent to the Phe residue of the GꢀLoop is occupied by substituents in
the B set. Note that Phe521 of the Gꢀloop samples differing conformations in other TrkA
structures and that the PDB structure 4YNE was chosen for binding mode prediction and
structureꢀactivity analysis because it resulted into the most favorable docking scores (i.e.
when bound to compounds of the B series).
27
ACS Paragon Plus Environment

Journal of Medicinal Chemistry
Page 28 of 37
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Figure 2: Summary of computationally designed hit compounds. Affinity values for
tozasertib and the top hits are listed for the key target (TrkA; green) and offꢀtarget (AurA;
red). Modifications compared to the starting compound are highlighted in bold lines.
28
ACS Paragon Plus Environment

Page 29 of 37
Journal of Medicinal Chemistry
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Figure 3: Multi-objective compound selection scheme. After initial binding pose generation
via docking calculations, the compounds are prioritized via in silico methods that filter first
with respect to selectivity criteria (i.e. the ligandꢀbased ‘kinase profiling platform’ approach
filters out promiscuous compounds, while the structureꢀbased ‘selectivity grids’ approach
filters for TrkA−Aurora selective compounds). Finally, highly active compounds on the key
target are prioritized via two complementary machine learning technologies (i.e. using
fragmentꢀbased Matched Molecular Pair descriptors and based on quantum mechanical
calculations, respectively). The final selection does not rely on any manual filtering and
includes in the case of the compound set A top ranking compounds from the MMP/ML and
QM/ML technologies and in the case of the set B top ranking compounds from MMP/ML.
29
ACS Paragon Plus Environment