Structure-Based Target-Specific Screening Leads to Small-Molecule CaMKII Inhibitors

ChemMedChem

10.1002/cmdc.201600636

Structure-Based Target-Specific Screening Leads to Small-Molecule CaMKIIInhibitors

2,5

1

1,3

1,2,3,4

David Xu , Liwei Li , Donghui Zhou , Degang Liu , Andy Hudmon , and Samy O. Meroueh

1

2

Department of Biochemistry and Molecular Biology, Center for Computational Biology and

3

4

Bioinformatics, Stark Neurosciences Research Institute, Indiana University Simon Cancer

Center

Indiana University School of Medicine, Indianapolis, Indiana, 46202

⁵Department of BioHealth Informatics, Indiana University School of Informatics and Computing,

Indianapolis, Indiana, 46202

Corresponding Author:

Samy Meroueh

Department of Biochemistry and Molecular Biology

Indiana University School of Medicine

Van Nuys Medical Science Building, MS 4023

635 Barnhill Drive

Indianapolis, Indiana 46202-5122

E-mail: smeroueh@iu.edu

Disclosure of Potential Conflicts of Interest: No potential conflicts of interest were disclosed.

ChemMedChem

10.1002/cmdc.201600636

ABSTRACT

Target-specific scoring methods are more commonly used to identify small-molecule inhibitors

among compounds docked to a target of interest. Top candidates that emerge from these

methods have rarely been tested for activity and specificity across a family of proteins. Here, we

2+

dock a chemical library to CaMKII, a member of the Ca /calmodulin (CaM)-dependent protein

kinase (CaMK) family, and rescore the resulting protein-compound structures using SVMSP, a

target-specific method that we previously developed. Among the 35 selected candidates, three

hits were identified, such as quinazoline compound 1 (KIN-1), which inhibited CaMKII kinase

activity with single-digit micromolar IC50. Activity across the kinome was assessed by profiling

analogs of 1, namely 6 (KIN-236), and an analog of hit compound 2 (KIN-15), namely 14 (KIN-

3

32), against 337 kinases. Interestingly, for 6 (KIN-236), CaMKII and homolog CaMKIIγ were

among the top 10 targets. Among the top 25 targets of 6 (KIN-236), IC50 values ranged from 5 to

2 M. Compound 14 (KIN-332) was not specific towards CaMKII kinases, but the compound

2

inhibited two kinases with sub-micromolar IC50s among the top 25. Derivatives of 1 were tested

against several kinases including several members of the CaMK family. The data afforded a

limited structure-activity relationship study. Molecular dynamics simulations with explicit-solvent

followed by end-point MM-GBSA free energy calculations revealed strong engagement of specific

residues within the ATP-binding pocket, and also changes in the dynamics as a result of binding.

This work suggests that target-specific scoring approaches like SVMSP may hold promise for the

identification of small-molecule kinase inhibitors that exhibit some level of specificity towards the

target of interest across a large number of proteins.

ChemMedChem

10.1002/cmdc.201600636

INTRODUCTION

Despite the large number of kinase inhibitors that have been discovered to date, a large

1

2+

portion of the human kinome remains unexplored. One example is the Ca /calmodulin (CaM)-

dependent protein kinase (CaMK) superfamily of kinases.²^,³Despite its nomenclature, the

2+

majority of kinases in the CaMK family do not possess the characteristic Ca /CaM-binding

regulatory domain. Over the years, several members of the CaMK family have been studied in

cancer, including members of the AMP-activated protein kinase (AMPK)-related kinase family

such as maternal embryonic leucine zipper kinase (MELK) or Novel (nua) kinase family 1

(

NUAK1).⁴^-⁸MELK has been found to be overexpressed in basal-like breast cancer and several

studies have implicated the enzyme with tumor growth and metastasis.⁹^,¹⁰NUAK1 a member of

the AMPK-like family whose function has yet to be elucidated despite strong data associating it

with promoting tumor growth and metastasis.¹¹^-¹³CaMKII is another CaMK family member that

has been studied in cancer.¹⁴^-¹⁷CaMKII was recently discovered to be the target of the potent

anti-leukemia drug berbamine.¹⁸The compound has been previously shown to be a potent

inhibitor of MDA-MB-231 cancer cell proliferation.¹⁹Additional examples of non-calmodulin

dependent members of the CaMK are the PIM group of kinases, which have been explored for

their role in cancer.²⁰

Structure-based computational methods such as virtual screening have been extensively

used to identify small-molecule inhibitors of kinases and other targets. These methods consist of

docking a chemical library to the target of interest followed by re-scoring protein-compound

complexes and rank-ordering compounds for experimental validation. Several docking methods

have been implemented in various computer programs such as AutoDock,²¹^,²²Glide,²³^,²⁴and

Gold.²⁵Methods to predict the binding mode of small molecules have matured significantly, but

26

there is a need for better scoring methods to rank-order protein-compound structures. Scoring

methods for the rank-ordering of protein-compound complexes range from methods that attempt

to predict the Gibbs free energy of binding to those that make use of physico-chemical or structural

information. The methods that use free energy include molecular dynamics-based approaches,

such as free energy perturbation, thermodynamic integration, and end-point methods (e.g. MM-

PBSA and LIE). These methods use molecular dynamics simulations that require substantial

computational resources. Scoring methods are less demanding, but they are generally used for

library enrichment in virtual screening. Several scoring approaches have been developed ranging

from empirical,²⁴^,²⁷^,²⁸force field,²⁵^,²⁹to knowledge-based.³⁰^-³²Increasingly, scoring methods use

machine learning techniques to improve database enrichment and rank-ordering.³³^-³⁷

ChemMedChem

10.1002/cmdc.201600636

The enrichment power of scoring methods is generally dependent on the target. As a

potential solution to this problem, we introduced a target-specific scoring approach (SVMSP,

Support Vector Machine SPecific) that uses Support Vector Machine (SVM) trained on pair

potentials from statistical energy functions.³³^,³⁸SVMSP is trained using a positive and negative

set. The positive set consists of crystal structures of the target in complex with small molecules.

The negative set is composed of decoy molecules docked to the target protein. Since SVMSP is

developed using the structure of the target of interest, we stipulated that the re-scoring method

may favor small molecules that inhibit and potentially show specificity towards the target. Our

previous work showed that SVMSP generally performed as well as or better than existing

methods, and that the unique strength of the method is its ability to show uniform performance

19

across a set of targets. SVMSP was compared to several other scoring methods including Glide.

In particular, we found that SVMSP performed well in the enrichment of chemical libraries docked

to kinases. The method was applied on several occasions to identify small-molecule inhibitors of

kinases and more recently aldehyde dehydrogenases.³³^,³⁹

Here, we use virtual screening with SVMSP re-scoring to identify small-molecule inhibitors

of CaMKII, a representative member of the CaMKII sub-group of the CaMK family implicated in

both cancer progression⁴⁰and cardiovascular disease.⁴¹To that end, we docked 149,479

compounds to its ATP-binding pocket. SVMSP was used to rank-order these compounds and

select top candidates for validation. We acquired the top 35 candidates and screened them for

activity in-house using a FRET-based assay. A quinazoline compound 1 (KIN-1) inhibitor was

identified and, along with an analog 6 (KIN-236), further evaluated for activity and activity across

the kinome using a radiometric assay at Reaction Biology Corp (http://www.reactionbiology.com,

Malvern, PA). We explored the activity profile of the derivative compound using the KinomeScan

service and tested it for activity against 337 kinases. The top 25 targets from the kinome profiling

were selected for a concentration-dependent study. To further explore the activity of 1, we

acquired an additional 8 derivatives and tested them for activity against 11 kinases that includes

three members of the CaMK family. Explicit-solvent molecular dynamics simulations followed by

end-point free energy calculations were used to gain further insight into the activity and binding

profile of the compound.

RESULTS

Virtual Screening Identifies Active Compounds. In an effort to identify small molecules

that inhibit members of the CaMK family, we resorted to virtual screening using SVMSP, a target-

specific re-scoring method. The workflow is summarized in Fig. 1a. A library of 149,979 small

ChemMedChem

10.1002/cmdc.201600636

22

molecules was docked to the ATP-binding pocket of CaMKII using AutoDock Vina. The

resulting protein-compound complexes were re-scored and rank-ordered using SVMSP, which

uses machine learning trained on a large number of kinase three-dimensional structures from the

PDB.³³^,³⁴^,³⁸^,³⁹The Support Vector Machine (SVM) algorithm requires a positive and negative

training set. In our case, the positive set consisted of 389 co-crystal structures of kinase-

compound complexes obtained from the PDB (Table S1), and the negative set consisted of

randomly selected decoy compounds from ZINC⁴²docked to the target (in this case CaMKII)

(

Table S2).

We compare the performance of the SVMSP model to the Glide scoring function using

Vina-docked and Glide-docked poses. We identified a set of 33 active and 58 inactive compounds

43

against CaMKII from ChEMBL (Table S3). We assessed the ability of the four different docking

and scoring combinations in their ability to enrich for active compounds against inactive ones (Fig.

S1a). Performance was assessed using the area under the ROC curve (ROC-AUC). A ROC-AUC

of 1 corresponds to a perfect scoring function that ranks all active compounds before inactive

compounds, while a ROC-AUC of 0.5 represents one that ranks compounds randomly. We find

that the Vina-docked and Glide-docked poses rescored with SVMSP perform the best (ROC-AUC

of 0.72 and 0.64, respective). Glide-docked poses scored with GlideScore and Vina-docked poses

rescored with GlideScore had ROC-AUCs of 0.61 and 0.56, respectively. In virtual screening, only

the top candidates are selected for validation. At low false discovery rates (FDR ≤ 5%), we see

that Glide/SVMSP shows better performance than Vina/SVMSP. At a FDR of 5%, the true positive

rate of Glide/SVMSP is 18% versus 9% in Vina/SVMSP. We then compared the performance of

docked poses with and without energy minimization. Docked poses from Vina and Glide were

refined using a local optimization algorithm. Minimization was limited to residues within 5 Å of the

docked compound. We again assessed performance using ROC-AUC. We find that the ROC-

AUCs of minimized and unminimized poses are generally similar (Fig. S1b). Again, Glide-docked

and Vina-docked poses scored with SVMSP resulted in the best performance, with both having

ROC-AUCs of 0.68. Similarly, Glide-docked and Vina-docked poses scored with GlideScore have

ROC-AUCs of 0.56 and 0.57, respectively. At low FDRs, we again see the Glide/SVMSP

outperforming Vina/SVMSP. We next examined how the difference in the Vina-docked and Glide-

docked poses affected the subsequent SVMSP and Glide scores (Fig. S1c). We find that an

increase in the absolute difference between the GlideScore of Vina-docked and Glide-docked

poses does not correlate with a change in the absolute difference on the SVMSP score.

ChemMedChem

10.1002/cmdc.201600636

Additionally, there is only a weak correlation between the RMSD of the Vina-docked and Glide-

docked poses and the absolute difference in GlideScore (Fig. S1d) and SVMSP (Fig. S1e).

SVMSP was used to rank-order the 149,979 compounds. The top 500 candidates were

selected and clustered into 50 sets. We selected a representative structure from each of the

clusters. Among them, 35 compounds were available for purchase. The SVMSP scores for these

compounds are provided in Fig. 1b.

Figure 1. (a) Virtual screening workflow used to identify compounds that inhibited CaMKII. Approximately 150,000

compounds were docked in AutoDock Vina and rescored using SVMSP. The top 500 compounds were identified and

clustered to 50. Of the 50 compounds, 35 were experimentally validated. (b) SVMSP scores of the 35 experimentally

validated compounds identified from virtual screening against CaMKII. Three hits, 1 (orange), 2 (red), and 3 (red) were

discovered from the initial virtual screening. 2 and 3 were later abandoned for possessing moieties commonly found in

frequent hitters of pan-assay interference compounds.

The top 35 compounds selected from virtual screening were acquired and tested for

inhibition of CaMKII kinase activity using a FRET assay in-house. The physiochemical properties

of these 35 compounds are provided in Table S4. We found three compounds inhibited CaMKII

ChemMedChem

10.1002/cmdc.201600636

enzyme activity by 50% or more at 50 µM, namely 1, 2 (KIN-7), and 3 (KIN-15) (Scheme 1). It is

worth mentioning the fact that 2 contains an acylhydrazone moiety, which is known to be unstable

at low pH, so this compound would not be suitable in future hit-to-lead and lead optimization

efforts unless the acylhydrazone moiety is replaced with a more stable group. We profiled an

analog of 2 to compare the inhibition profile across the kinome of the active compounds that

emerged from this work. We did not test 3 since it was flagged as a pan-assay interference

(

PAINS) compound.⁴⁴

We compared the structure of compound 1 to the approximately 2,300 compounds tested

4

3

against CaMKII in the literature using ChEMBL. The most chemically similar compound is an

inhibitor of CaMKII that was discovered using pharmacophore search, which features a similar

quinazoline core as compound 1 (ECFP4 fingerprints, Tanimoto coefficient of 0.154).⁴⁵In their

compound, the alkene linker of the benzene substituent is shortened with a secondary amine,

and the alkylamine tail of 1 is replaced with a methylpiperidine group. Additionally, we compared

the structure of 1 against approximately 55,000 compounds in ChEMBL’s Kinase SARfari

database. The most similar compound, chloroquine, has a Tanimoto coefficient of 0.256 to 1.

Chloroquine is used to treat malaria, and features a similar core and alkylamine tail as 1. The

quinazoline core is replaced by a quinoline ring while the bezene substituent and alkene linker of

1

are omitted in the compound.

A concentration-dependent study for prepared 1 (Scheme 2) and one of its analogs, 6

(

Scheme 3), using a radiometric assay done at Reaction Biology Corp showed that the

compounds inhibited enzyme activity with an IC50 of 7.8 ± 0.4 µM (Fig. 2). The ADMET of 1 and

46

6

were computationally predicted using QikProp and included in Table S5.

ChemMedChem

10.1002/cmdc.201600636

Scheme 1. Chemical structure of active compounds that emerged from virtual screening. Problematic moieties are

colored in red.

Scheme 2. Reagents and conditions: (a) (CH3CO)2O, 110 °C, 5 h; (b) 28.0 - 30 % NH3 (aq), 50-60 °C, 10 h, 70% (two

steps); (c) PheCHO, NaOAc, HOAc, 110 °C, 30 h; (d) POCl3, N,N-dimethylaniline, toluene, 110 °C, 8 h; (e) 2-amino-

5

-diethylaminopentane, DMAP, Et3N, toluene, 110 °C, 15 h.

ChemMedChem

10.1002/cmdc.201600636

Scheme 3. Chemical structure of derivatives of 1.

ChemMedChem

10.1002/cmdc.201600636

Figure 2. Concentration-dependent study for 1 and one of its analogs, 6, using a radiometric assay at Reaction Biology

Corp showed that the compounds inhibited enzyme activity by an IC50 of about 5 µM.

Kinome-Wide Profiling. To explore the activity of 1 across the human kinome, we

resorted to kinome profiling using the KinomeScan facility at Reaction Biology Corp, which

included 337 unique kinases. We selected 6 for the profiling study. Considering the similarity in

structure between 1 and 6 and their nearly identical activity for inhibition of CaMKII (Fig. 2), it

was expected that both compounds will share similar targets and inhibition profile. 6 was profiled

against 337 kinases at 20 µM (Fig. 3a). The data show that the compound inhibited 16 kinases

by 50% or more at 20 µM. To confirm these results, the top 25 targets of 6 were selected for a

concentration-dependent study (Fig. 3b). There were 10 kinases that were inhibited by 6 with an

IC50 of 10 µM or less, further highlighting the selective nature of the compound series.

Interestingly, there were only 4 serine/threonine kinases among the top 25, and all belong to the

CaMK superfamily, namely CaMKII, MELK, CaMKII and CaMK1. Two (CaMKII and CaMKII)

belong to the CaMKII sub-group, one belongs to the CaMK1 sub-group (CaMK1), and another

belongs to the CaMKL sub-group (MELK). The remaining targets among the top 25 were all

tyrosine kinases.

The inhibition profile of 6 was compared to that of a derivative of 2 (compound 14) (Fig.

S2a). We profiled compound 14 against the same 337 kinases using the KinomeScan facility at

Reaction Biology Corp. 14 was profiled against the 337 kinases at 20 µM (Fig. S2b). The data

show that the compound inhibited 46 kinases by 50% or more at 20 µM. This is in contrast to 16

ChemMedChem

10.1002/cmdc.201600636

kinases that were inhibited by 50% or more by 6. The greater number of targets inhibited 14 was

reflected in follow-up concentration-dependent studies for the top 25 kinases (Fig. S2c). At 20

µM, the compound inhibited CaMKII by 42%, suggesting an IC50 that is lower than that of 1 and

6. CaMKII was at position 46 among the 337 kinases when rank-ordered by percent inhibition,

in contrast to 6 where CaMKII was among the top 10.

Figure 3. (a) Kinome-wide profiling of 6 against 337 kinases at 20 µM. 6 inhibited 16 kinases by 50% or more at 20

µM. (b) Concentration-dependent study of 6 against the top 25 targets.

Binding Mode. An important question is whether the binding mode of 1 that emerged

from virtual screening (Fig. 4a) could be adopted by its derivatives in multiple members of the

CaMK subfamily. To explore this question, we first re-docked the compound with another program

(

Glide) to determine if the binding mode could be reproduced by the re-docking. We re-docked 1

and its derivatives against CaMKII, CaMKII, and MELK. We stipulated that a binding mode that

is found in all three kinases will provide additional confidence for the predicted pose. Interestingly,

in all three cases, we identified a binding mode for the Glide binding pose that adopted a similar

arrangement in three-dimensional space to the Vina binding mode (Fig. 4b). However, the binding

ChemMedChem

10.1002/cmdc.201600636

mode in Fig. 4a shows a shift in the position of the compound to better occupy the two

hydrophobic pockets of the ATP binding domain. The RMSD between the binding pose from

virtual screening and this new binding pose is 4 Å. It is well-known that among co-crystallized

inhibitors bound to the ATP binding site, the majority use one or both of these hydrophobic

pockets.⁴⁷The predicted binding mode of 1 was superimposed with the binding mode of ATP

(

Fig. 4c). The quinazoline core of 1 occupies the same hydrophobic pocket of the adenine of ATP,

while the N-akylamine tail of 1 overlaps with the ribose and triphosphate tail of ATP.

Figure 4. Characterization of the binding mode of 1 to CaMKII. (a) Stereoview of the binding mode of 1 from the

original virtual screening against CaMKII in the R configuration. (b) Stereoview of the predicted binding mode of 1

after analysis of the compound series against CaMKII, CaMKII, and MELK. (c) Stereoview of ATP (PDB ID: 1ATP,

2

.20 Å) is superimposed on the predicted binding mode of 1 (yellow). (d) Stereoview of an experimentally crystallized

inhibitor featuring a quinazoline ring structure in a similar conformation (PDB ID: 3QKM, 2.20 Å, magenta) as the

predicted binding mode of 1 (yellow). In (a-d), a crystal structure of CaMKII (PDB ID: 2WEL, 1.90 Å) is shown as

cartoon (gray) with the hinge loop (red), DFG-motif (green), and αC helix (orange) highlighted. The protein is shown

with the glycine-rich loop removed (Residues Glu-19 to Arg-29) to expose the ATP binding pocket. Key residues

between 1 and the protein kinase are shown in stick format. Two-dimensional representation of the (e) virtual screening

and (f) Glide-predicted binding modes of 1 against CaMKII, showing the key interactions between the atoms of the

ligand and the ATP binding pocket of the receptor. Generated by PoseView (http://poseview.zbh.uni-

hamburg.de/poseview). (g) A multiple sequence alignment of the top candidates of 6 was generated in Cluster Omega

and the residues of CaMKII within 4 Å of the original virtual screening binding mode of 1 were identified (in color).

These residues were annotated based on the hydrophobicity of the residue, ranging from red (most hydrophobic) to

blue (most hydrophilic). Residues in the hinge loop and activation loop are highlighted red and green, respectively, on

the residue numbers.

To further confirm the binding pose, we searched the PDB for small-molecule kinase

inhibitors that share substituents with 1. Since there is no crystal structure of 1 or any of its analogs

bound to a kinase in the PDB, we resorted to crystal structures of compounds that possessed

ChemMedChem

10.1002/cmdc.201600636

similar moieties. The quinazoline moiety of 1 is commonly found on kinase small-molecule

inhibitors. We compared the orientation of the quinazoline scaffold of 1 to existing co-crystal

structures of kinase inhibitors that also featured quinazoline rings in their scaffold. Generally, the

quinazoline ring that forms the core scaffold of these compounds are positioned into one of three

general orientations (Fig. S3). The first orientation features compounds that mimic the nucleobase

of ATP with a benzene-substituent at the imidazole position (PDBID: 4OBO, 2.10 Å; 4OBQ, 2.19

Å) (Fig. S3a). The second group features a quinazoline orientation that mirrors the one found in

our compound, whereby the benzene and pyrimidine rings of the quinazoline core occupies the

other’s position (PDBID: 1M17, 2.60 Å; 3PRZ, 2.60 Å; 4AA5, 2.38 Å) (Fig. S3b). The third group

includes four compounds (PDB ID: 3QKM, 2.20 Å; 4BB4, 1.65 Å; 4PP7, 3.40 Å; 4F7S, 2.20 Å)

possessing a quinazoline moiety in a similar orientation to the one found in our predicted binding

mode of 1 (Fig. S3c). One example is a spriocyclic sulfonamide inhibitor of AKT1 (Fig. 4d), whose

quinazoline superimposes nearly perfectly with the predicted binding pose (Fig. 4b). This

suggests that the position and orientation of the quinazoline of 1 is likely in the correct position in

the predicted binding modes.

A comparison of the two-dimensional representations of the virtual screening and Glide-

predicted binding modes reveal similarities in the binding interactions between them. From the

two-dimensional representation of the virtual screening (Fig. 4e) and Glide-predicted (Fig. 4f)

binding modes, the quinazoline ring occupies the hydrophobic pocket normally occupied by the

adenine of ATP, while the benzene substituent of 1 occupies the hydrophobic pocket formed by

residues in the hinge loop. In both binding modes, the alkylamine substituent of the compound

forms a salt-bridge interaction with the aspartic acid on the DFG motif of the activation loop. In

the original pose, the quinazoline ring of the compound forms a π-π interaction with Phe-90 of

the hinge loop that is not seen in the predicted pose.

A multiple sequence alignment using Clustal Omega⁴⁸of the kinase domains of the top

candidates was carried out. We identified residues in the ATP binding pocket of CaMKII that

49

formed interactions with the compound and color-coded them based on hydrophobicity, ranging

from the most hydrophobic residues in red to the most hydrophilic in blue (Fig. 4g). Across the

majority of these residues, either the residue was conserved across all the kinases (e.g. Lys-43

and Asn-141) or the alternate residue at the position did not result in a large difference in the

hydrophobicity (e.g. Leu-20 to Ile and Val-74 to Ile). Leu-92 and Val-93 are replaced by less

hydrophobic residues in the majority of the top ranked tyrosine kinases of 6. At residue 92, leucine

is replaced by tyrosine or histidine in 7 of the 9 tyrosine kinases, and at residue 93, valine is

ChemMedChem

10.1002/cmdc.201600636

replaced by methionine or alanine in 8 of 9 tyrosine kinases. These residues, along with Leu-20,

border the benzene ring substituent site that is attached to the core quinazoline of 1.

The quinazoline core in compound 1 does not form hydrogen bonds to the hinge region in

either the Vina or Glide pose. This hydrogen bonding pattern to the hinge region is frequently

seen in existing kinase ATP inhibitors.⁵⁰The residues on the hinge loop that commonly form

hydrogen bonds with inhibitors corresponds to the residues from position 90 to 94 in CaMKII

(

amino acids FDLVT) in Fig. 4g. The quinazoline core of 1 forms hydrophobic contacts with Phe-

0, Leu-92, and Val-93 of the hinge loop on CaMKII. The sidechains of these residues do not

9

have moieties that are amenable for hydrogen bonding. The most common sites for hydrogen

bonding are at residues that correspond with positions 91 and 93 on CaMKII, which are referred

to as the hinge acceptor and hinge donor, respectively. On CaMKII, these two residues are Asp-

91 and Val-93. However, they also do not form any hydrogen bonding interaction with our

compounds. While Asp-91 has a moiety amendable for hydrogen bonding, the sidechain in the

crystal structure points away from the binding pocket, and is unlikely to form hydrogen bonds to

the core of the compound in the crystal structure.

Structure-Activity Relationships. Compound 1 and 8 of its derivatives were tested for

activity against 11 kinases selected from the list of the top targets of 6 (Fig. 5). These studies

were done at 20 µM concentration of compound. Measurements at single concentrations can be

prone to error, so these data are used qualitatively. First, an overall inspection of the data shown

in Fig. 5 shows a striking preference for CaMK by the compounds as evidenced by the color-

coding used to represent percent activity of the enzyme. Overall, 1 and several derivatives [4

(

KIN-234), 5 (KIN-235), 6 (KIN-236), 7 (KIN-281), and 8 (KIN-284)] showed strong preference for

the three CaMK members considered in the list. 9 (KIN-286), 10 (KIN-288), and 11 (KIN-289),

however, showed much weaker activity towards the CaMKs. This is explained by the fact that 9,

10, and 11 lack the N-alkylamine moiety in the quinazoline ring. Instead, these compounds

possess piperidine, azepane, and methyl-substituted piperidine moieties, respectively. 9 and 10

did not inhibit any of the other kinases with a few exceptions: 10 inhibited FGFR1, BTK, TIE2/TEK,

and BLK by 60, 35, 25 and 50% respectively.

ChemMedChem

10.1002/cmdc.201600636

Figure 5. Profiling of 1 and 8 derivatives for activity against a panel of 11 kinases at 20 µM concentration of compound.

Numbers represent percent activity ± standard deviation. Values are omitted when the kinase profiling for the compound

exceeds 110%, as it is unlikely that these compounds are activating the respective kinase.

Molecular Dynamics Simulations and Free Energy Calculations. To further gain

insight into the interaction between 1 and its derivatives with CaMKII, we ran explicit-solvent

molecular dynamics simulations followed by free energy calculations. Using the Glide-predicted

binding mode of 1, we generated binding modes for each of the derivatives of 1. We then

subjected these protein-compound structures to 100-nanosecond (ns) explicit-solvent molecular

dynamics simulations. Although the original docked compound of 1 is in its R configuration, the

subsequent synthesis produces a racemic mixture of the compound that was not further

separated. Given the presence of a chiral center in the compounds with the N-alkylamine moiety

(

1, 5, 6, 7, and 8), we considered both the R and S configurations for these 5 compounds. The

components of the MM-GBSA free energy are listed in Table 1 and the RMSD measurements of

each simulation is listed in Table S6. It is important to note that MM-GBSA calculations are not

meant to reproduce the absolute value of the free energy of binding. Instead, these calculations

are meant to rank-order compounds by their binding affinity. In addition, the multiple components

of the MM-GBSA free energy provide deeper understanding of the contributions of the

components of the free energy such as the non-polar, electrostatics and entropy. Pearson

correlation coefficients between the experimental percent activity and the individual MM-GBSA

components are similar for both R and S configurations (Table 2). Considering the bi-modal

nature of the data, Spearman and Kendall coefficients may be more appropriate. Spearman’s ρ

is 0.68 and 0.85 for the correlation between the MM-GBSA free energy and the percent activity

of the compounds for the R and S enantiomers, respectively. The higher correlation for the S

enantiomer suggests that this configuration is likely the one that binds to the kinase.

ChemMedChem

10.1002/cmdc.201600636

-

1

Table 1. Calculated Free Energies (± Standard Error) of Compounds Against CaMKII (kcal·mol )

Analysis of the MM-GBSA components in Table 1 shows that compounds with the N-

alkylamine moiety exhibit substantially more favorable electrostatic potential energies (ΔEELE

)

than compounds that lack the moiety. This suggests that the critical component of the N-

alkylamine moiety is the charged tertiary amine that likely forms favorable electrostatic

interactions with amino acid residues on the kinases. These interactions likely occur with the

charged sidechains of Asp-157 (Fig. 4d). Overall the non-polar interaction van der Waals energy

(

ΔEVDW) is also more favorable among compounds with the N-alkylamine group (Table 1)

consistent with the larger number of atoms for compounds with the N-alkylamine moiety. The

electrostatic component of the solvation energy (ΔEGB), however, was dramatically more

favorable for the compounds lacking the N-alkylamine, consistent with the fact that desolvation of

a charged group results in significant penalty in the free energy. Finally, the flexible nature of the

ChemMedChem

10.1002/cmdc.201600636

N-alkylamine moiety is likely to result in an entropy penalty for binding. We further explored the

MM-GBSA data to determine whether the components of the MM-GBSA free energy correlate

with compound activity data from Fig. 5. Using Spearman’s ρ, activity is highly correlated to the

van der Waals energy (ΔEVDW), electrostatic energy (ΔEELE), non-polar solvation energy (ΔESURF),

and combined internal and solvation terms (ΔEGBTOT). Correlations ranging from 0.64 to 0.69 in

the R configuration and 0.54 to 0.89 in the S configuration. The polar solvation term (ΔEGB) is anti-

correlated with compound activity (ρ = -0.73 and -0.56 for R and S, respectively).

To gain insight into the intermolecular interactions between compound and residues within

the ATP-binding site of CaMKII, we calculated MM-GBSA decomposition energies of each

compound (Fig. 6a). Decomposition energies correspond to intermolecular energies between the

bound compound and individual residues within the target protein. The decomposition energies

are calculated in the same manner as the MM-GBSA energy except that they do not include the

entropy term. Generally, both enantiomers share similar decomposition energies at individual

residues. Highly favorable interaction energies were identified between compound and three

hydrophobic residues, namely Leu-20, Val-28, and Leu-143 (Fig. 6b). Leu-20 forms hydrophobic

interactions with the benzene ring substituent that is attached to the core quinazoline, while Val-

2

8 and Leu-143 form hydrophobic interactions with the quinazoline core of the compound. There

is a weak interaction near Asn-141 and Ala-156, which interacts with the N-alkylamine moiety of

. These interactions disappear in 9, 10, and 11, which lack the N-alkylamine moiety. It is worth

1

noting that for 9, 10 and 11, there is greater interaction with Leu-92, Val-93, and Thr-94 of the

hinge loop.

ChemMedChem

10.1002/cmdc.201600636

Figure 6. (a) Decomposition energies of key residues of binding for 1 and derivatives against CaMKII. For compounds

with a chiral center in the N-akylamine moiety, only the S configuration shown. Each row is a compound while each

column is an individual residue. (b) A snapshot from the simulation of (S)-1 bound to CaMKII, showing the sidechains

of key residues from the interaction in line format. The crystal structure of CaMKII (PDB ID: 2WEL, 1.90 Å) is shown

as cartoon (gray) with the hinge loop (red), DFG-motif (green), and αC helix (orange) highlighted. The protein is shown

with part of the glycine-rich loop removed (Residues Gly-21 to Ser-26) to expose the ATP binding pocket. (c) Plot of

the RMSD of the compound (blue) and the distance from the positively charged amine hydrogen on the N-akylamine

moiety of the compound from the negatively charged epsilon and delta oxygen atoms of Glu-97 (orange) and Asp-157

(gray), respectively, across 1000 snapshots for the R and S configurations of 1 and 7. (d) A snapshot from the simulation

of (S)-7 (KIN-281) bound to CaMKII, showing the additional hydrogen bond created by the amine group on the

quinazoline core of the compound. The crystal structure of CaMKII (PDB ID: 2WEL, 1.90 Å) is shown as cartoon (gray)

with the hinge loop (red), DFG-motif (green), and αC helix (orange) highlighted. The protein is shown with the glycine-

rich loop removed (residues Glu-19 to Arg-29) to expose the ATP binding pocket. Asp-157 and Glu-97 are also shown

as element-colored stick models, with carbons colored as green and cyan, respectively. (e) Dynamic cross-correlation

matrix (DCCM) cross section of 1 and 8 derivatives against the protein kinase domain of CaMKII.

From the decomposition energies, there are unfavorable interactions between the

compound and residues Glu-97 and Asp-157. Although the initial models feature a salt bridge

between the positively charged protonated amine on the N-alkylamine substituent and the

negatively charged aspartic acid of the DFG motif. We observe that this interaction is not stable

in our simulations. Visualization of the individual simulations reveal that the nearby Glu-97

sidechain attracts the flexible N-akylamine and causes it to adopt multiple conformations over the

course of the simulation (Fig. 6c). In (R)-1, the salt bridge with Asp-157 is broken in all but one of

the individual trajectories as the N-alkylamine interacts with Glu-97. The dissociation of the salt

bridge and subsequent flexibility is also observed in (S)-1, although in only approximately half of

the trajectories. In 7, which features an amine substituent on the quinazoline ring instead of a

chlorine, we see stabilization of both the salt bridge and compound as the additional amine

ChemMedChem

10.1002/cmdc.201600636

strengthens the interaction with Asp-157 (Fig. 6d). This is consistent with the higher inhibitory

activity of 7 compared with the parent compound (Fig. 5).

The relative flexibility of the different binding poses of compound 1 is explored in Fig. S4.

Three binding poses of compound 1 emerged from this study: the Vina-docked pose from the

original virtual screen of 1 in its R configuration, and the Glide-docked poses of 1 in both its R and

S configurations. We compared the atomic fluctuations on a per-residue basis for each of these

three binding poses against the apo protein. This is measured as a ratio between the B-factors of

a set of atoms on a residue in the bound and unbound states. A normalized flexibility of 1

corresponds to similar levels of dynamics between the bound and unbound protein. A measure

above 1 represents greater fluctuation in the bound state than the unbound state, while a measure

below 1 represents a more stable bound state compared to the unbound state. This measure was

compared for all atoms of a residue (Fig. S4a), only the backbone C, Cα, and N atoms (Fig. S4b),

and only the sidechain atoms (Fig. S4c). We find that generally, the relative flexibility of the three

binding modes are similar. Compound binding generally stabilizes residues that line the ATP

binding pocket. However, there are exceptions where residues become more flexibility upon

ligand binding. In the Vina-docked pose, Phe-90 of the hinge loop and His-126 show the largest

difference in flexibility compared to the Glide-docked pose. While His-126 is not near the ATP

binding pocket, Phe-90 forms a π-π interaction with the compound in the Vina-docked pose. The

majority of the difference in flexibility at these two residues is attributed to dynamics in the

backbone of these residues. Ala-175 is among the most flexible residue in the protein, but this

change in flexibility is only prevalent in the R configuration of the Vina-docked and Glide-docked

poses.

We next explored the dynamics of the compounds bound to CaMKII using dynamic cross-

correlation matrices (DCCM). The correlation of the motion between each of the compounds and

CaMKII is shown in Fig. 6e. Visual inspection reveals similar motions for 1 and its derivatives.

Generally, there are positive correlations between the residues of the compounds and the

residues of the binding pocket. This is reflected most prominently in the residues near the three

substituents of 1. The alkylamine moiety that forms a salt-bridge with the DFG motif in 1 near Asp-

157 features a largely positively correlated region in each of the compounds. However, this region

is noticeably smaller in the residues following Asp-157 in 4, 9, and 11, which feature shorter

moieties at this position. The benzene ring substituent site that is attached to the core quinazoline

of 1 occupies a region following Val-93 of the hinge loop and a region near Leu-20. The core

quinazoline not only occupies the region near Ala-41 and Lys-43, but also Phe-90 of the hinge

ChemMedChem

10.1002/cmdc.201600636

loop. In most cases, both regions are positively correlated, except in (S)-8, which follows an

opposite trend in the Leu-20 region. Overall, the dynamics of the compounds reveal similar

patterns of correlation in CaMKII.

Molecular Dynamics Simulations to Explore Inhibition Profile Among Kinases. To

gain deeper insight into the basis for the inhibition of compounds, we again resorted to explicit-

solvent molecular dynamics simulations. We carried out simulations of (S)-6 bound to 6 of the

kinases identified from the kinome profiling of the top 25 targets of 6 in Fig. 3b. We picked the S

enantiomer as it featured both higher correlation coefficients in the SAR and increased stability

with the aspartic acid of the DFG motif. We selected kinases with the highest (CaMKII, CaMKII,

BTK, and CSK) and lowest (FGFR4 and EPHB1) affinities to 6. The differences in the IC50 values

between highest and lowest affinity kinases is 5-fold (5 µM in CamKII versus 22 µM in EPHB1).

We docked (S)-6 against each of the 6 kinase structures and selected the pose that best

resembled the predicted binding pose against CaMKIIusing Glide. We ran 100-ns explicit-

solvent molecular dynamics simulations and free energy calculations for these complexes (Table

S7). These extensive simulations will allow the compound to adopt its most favorable binding

pose within the ATP-binding pocket. We next explored the interaction of (S)-6 with individual

residues on the protein kinases by using decomposition energy calculations. We focused on the

interaction energies of these key residues in the binding pocket (Fig. 7a). Like the SAR, three

residues were strongly engaged by the compound across the 6 protein kinases, namely Leu-20,

Val-28, and Leu-143. When comparing the enzymes that are strongly inhibited by (S)-6 (i.e.

CaMKII, CaMKII, BTK, and CSK), with those that are more weakly inhibited by the compound

(

i.e. EPHB1 and FGFR4), we see little differences in the decomposition energies at specific

residues.

ChemMedChem

10.1002/cmdc.201600636

Figure 7. (a) Total decomposition energy (ΔEGBTOT) of key residues of binding for (S)-6 against 6 kinase targets. Each

row is a kinase target while each column is an aligned residue. (b) Summation of the van der Waals (ΔEVDW) and

electrostatic (ΔEELE) terms from the total decomposition energy of key residues of binding for (S)-6 against the same 6

kinase targets. (c) Dynamic cross-correlation matrix (DCCM) cross section of (S)-6 against the protein kinase domain

of 6 kinase targets. Gaps in the multiple sequence alignment are shown in gray. Cross sections were converted into

vectors and hierarchically clustered using Euclidean distance and average linkage. (d) The correlation of motion of 6S

bound to CaMKII and EPHB1 were mapped onto the structures of the protein kinase domains. Correlation coefficients

were colored according to the scheme in panel c. (e) Per-residue DCCM of the protein kinase domain for 6S bound to

CaMKII and EPHB1.

In CaMKII and CaMKII, the interaction energies of the compound with two negatively

charged glutamic acid residues at positions 97 and 140 are favorable. However, the polar and

non-polar solvation terms are unfavorable. At position 97, this residue is replaced by an aliphatic

alanine and an uncharged polar asparagine in EPHB1 and FGFR4, respectively. Similarly, the

negatively charged glutamic acid is replaced by positively charged arginine residues in the 4

tyrosine kinases at position 140. Two residues appear to interact with (S)-6 among these two

groups: Glu-97 and Asp-157. These two negatively charged residues interact with the positively

charged amine group of the N-alkylamine moiety to form hydrogen bonds with the compound

depending on the orientation of the flexible N-alkylamine moiety. In the CaMKs, the interaction

energy appears to be highly unfavorable with Glu-97 and slightly unfavorable with Asp-157. This

can be attributed to the penalty of desolvation of the N-alkylamine. In fact, when the solvation

energies are not considered, the interaction energy is highly favorable between the compound

and these two amino acids (Fig. 7b).

ChemMedChem

10.1002/cmdc.201600636

In addition to interaction energies, we explored the effect of (S)-6 on the dynamics of the

6

kinases considered in the calculations. This was accomplished by measuring the degree of

correlation in the motion of (S)-6 with the motion of individual residues on the kinases to which

the compound is bound (Fig. 7c). Positive correlations suggest that the compound and residue

tend to move in the same direction over the course of the trajectory. Overall inspection of the map

in Fig. 7c reveals similar patterns of interaction between the enzymes that are strongly inhibited

by 6 (i.e. CaMKII, CaMKII, BTK, and CSK) and those that are less strongly inhibited (i.e. EPHB1

and FGFR4). The map reveals strongly correlated motions between compound (S)-6 and residues

in the protein kinases; these regions of strong correlated motions occur between (i) residues 14

and 30; (ii) 40 and 80; and (iii) 110 and 170. Strong anti-correlated motion with (S)-6 occurs in the

other regions, such as those from residues (i) 9 to 14; (ii) 30 to 40; (iii) 95 to 110; and (iv) 170 to

275.

Since the patterns of motion are similar by visual inspection, we wondered if we could

differentiate the two groups of kinases by clustering the protein-compound complexes in a

quantitative manner. We stipulated that kinases with similar dynamics as a result of (S)-6 were

more likely to also have similar inhibition profiles. For each kinase, we defined a fingerprint by

taking a cross-section of the DCCM map and mapping the correlations to the multiple sequence

alignment. This fingerprint consists of correlation coefficients that correspond to the correlation of

motion between compound and each residue on the protein. To compare these fingerprints, we

calculated the Euclidean distance between each pair of kinases and hierarchically cluster these

fingerprints using average linkage (Fig. 7c). Interestingly, CaMKIIexhibited dynamics that was

most dissimilar to the other kinases. Among the other CaMK, CaMKII is clustered with BTK. The

two kinases with the highest IC50, EPHB1 and FGFR4, are clustered together. We visualized the

dynamics of (S)-6 on CaMKII and EPHB1 by applying the correlation coefficients from the DCCM

cross-sections on structures of their kinase domains (Fig. 7d). Generally, the patterns of motion

are similar between the two kinase targets, with the strongest correlations occurring adjacent to

the binding pocket and strongest anti-correlated motion occurring distal to the binding pocket. In

EPHB1, there is less correlation inside the binding pocket than in CaMKII. The DCCMs of the

CaMKII and EPHB1 protein in Fig. 7e reveal similar trends when bound to (S)-6. The largest and

most correlated of these regions occurs in the center of the DCCMs. In the orientation of the

protein kinase domains in Fig. 7d, this region occurs near the ATP binding pocket.

DISCUSSION

ChemMedChem

10.1002/cmdc.201600636

Here, in an effort to identify small-molecules inhibitors of members of the CaMK family,

we resorted to virtual screening using a target-specific re-scoring approach that we had previously

developed. The method, known as SVMSP, uses Support Vector Machine (SVM) to train an

algorithm using statistical pair potentials as features obtained from three-dimensional protein-

compound structures. The positive set consists of protein-compound co-crystal structures

obtained from the PDB, while the negative set consists of a randomly selected set of decoy

molecules docked to the target of interest. The SVMSP scoring approach will favor docked

compounds that adopt a binding mode that is similar to those found in the positive set while

rejecting compounds that are similar to those found in the decoy set. The scoring approach does

not take binding affinity in consideration, so its ability to identify active compounds is based solely

on binding mode. The use of a target of interest (in this case CaMKII) to create the negative set

trains the algorithm to reject compounds that are decoy-like. We believe that the use of CaMKII

enables the algorithm to learn more about the features of the CaMKII binding site, making it

possible for the scoring approach to favor compounds that are more specific to CaMKII. The

performance of SVMSP was previously evaluated using the Directory of Useful Decoys (DUD),⁵¹

which provides a list of targets along with a library containing a set of known ligands and decoy

molecules for validation of scoring methods. The performance of SVMSP was compared to

several other scoring methods, which includes Glide,²³^,²⁴GoldScore, ChemScore, X-Score,

25

26,

52

and PMF.⁵³We found that overall, SVMSP performed better than these scoring functions. More

importantly, SVMSP showed consistently good performance across several protein families that

included kinases, proteases, metalloenzymes and nuclear receptors.

We applied SVMSP re-scoring towards the search for small molecules that inhibit

members of the CaMK family. Our long-term objective is to identify selective and potent inhibitors

of individual members of the CaMK family that play an important part in cancer initiation,

progression and metastasis. Here, a member of the CaMKII sub-group of the large CaMK family

(

CaMKII) was used. First, 149,979 compounds were docked to the ATP-binding pocket of

CaMKIIfollowed by re-scoring of the resulting protein-compound structures with SVMSP. We

identified the top 500 candidates and further clustered them into a set of 50 compounds. We

tested 35 of these compounds that were commercially-available. Three active compounds were

found, 1, 2, and 3. Compound 3’s activity is unlikely to be due to specific inhibition of CaMKII

44

since it is considered a pan-assay interference compound (PAINS). The fact that this compound

emerged among the top candidates confirms that virtual screening is not immune to these

compounds. Compound 2, while not considered a PAINS compound, contains an acylhydrazone

ChemMedChem

10.1002/cmdc.201600636

moiety. This moiety, while robust at higher pH, is considered unstable at lower pH. A recent study

revealed that these compounds can be successfully modified to obtain more stable derivatives.⁵⁴

A question of interest is whether the binding mode of 1 that emerged from virtual screening

is close to the real binding mode of the compound. Re-docking of the compounds using a different

and robust docking program (Glide) led to a similar binding mode. The position of the quinazoline

ring on this Glide-predicted binding pose was nearly exactly the position of quinazoline moieties

found in the crystal structure of other small molecules bound to kinases. This suggests that the

Glide binding mode is likely closer to the real binding mode of the compound. The fact that SVMSP

identified 1 as an active compound despite the fact that it adopts the correct arrangement but

does not likely adopt the binding mode of the compound within 1-2 Å may actually be a positive

aspect of these scoring methods. Virtual screening does not always lead to binding poses that

are in perfect agreement with what is found in the co-crystallized complex.²²^,⁵⁵Since SVMSP is

not as sensitive to the binding mode of a compound, this suggests that the re-scoring method can

identify active compounds that would otherwise be ignored by other scoring functions.

To explore the inhibition profile of 1 across the human kinome, we pursued 1 and profiled

an analog 6 for activity against 337 kinases initially at a single concentration of 20 M.

Remarkably, CaMKIIthe kinase that was used in the virtual screening, was among the top 10

most potently inhibited by 6 among 337 kinases. In addition, the kinase that was most potently

inhibited by 6 was CaMKIIa close homolog of CaMKII. In fact, all four of the serine/threonine

kinases among the top 25 kinases were members of the CaMK family. One belongs to the CaMK1

sub-group (CaMK1), and another belongs to the CaMKL sub-group (maternal leucine embryonic

zipper kinase or MELK). The remaining targets among the top 25 were all tyrosine kinases.

Analysis of the ATP binding site among the top targets of 6 revealed that the majority of residues

were conserved across all the kinases (e.g. Lys-43 and Asn-141) or the alternate residue at the

position did not result in a large difference in the hydrophobicity (e.g. Leu-20 to Ile and Val-74 to

Ile). In some cases, hydrophobic residues in the CaMKIIs are replaced by less hydrophobic

residues in the tyrosine kinases (e.g. Leu-92 and Val-93). The fact that 6 exhibited some

specificity towards CaMKII may be attributed to the target-specific nature of the SVMSP scoring

method. The use of CaMKII in the negative training set may have led to an algorithm that

enriches for compounds that bind to CaMKII

Comparatively, two CaMKII inhibitors found in the literature were among 178 kinase

56

57

inhibitors recently profiled at Reaction Biology Corporation. Both KN-62 (K = 0.9 µM) and KN-

i

ChemMedChem

10.1002/cmdc.201600636

9

3⁵⁸^-⁶⁰(K

i

= 0.37 µM, IC50 ~ 1-4 µM) were systematically tested against 300 kinases at lower

concentrations for the compounds (20 µM versus 0.5 µM) than our kinome-wide screening of 6

and 14. Both CaMKII inhibitors showed 90% kinase activity to the α, β, γ, and δ subtypes of

CaMKII. A more recent screening effort from EMD Millipore profiled these two compounds at both

1

and 10 µM against 234 kinases, including the β, γ, and δ subtypes of CaMKII.⁶¹Neither

compound showed activity at 1 µM against the 3 tested subtypes. At 10 µM, KN-62 inhibits the

β, γ, and δ subtypes by 22, 13, and 9 percent, respectively, while KN-93 inhibits the same

subtypes by 80, 33, and 53 percent, respectively. Both of these inhibitors are non-competitive

ATP inhibitors and bind allosterically to CaMKII.⁶²Two additional inhibitors have been co-

crystallized with CaMKIIδ at the ATP binding site: SU6656⁶³and K00606.⁶⁴SU6656 (PDB ID:

2WEL, 1.90 Å) inhibits kinase activity to the α, β, and δ subtypes of CaMKII by ~30% but not

CaMKIIγ at 0.5 µM in the RBC study. In addition, SU6656 also shows more than 60 and 80 percent

inhibition to 38 and 19 different kinases at 0.5 µM, respectively. The other co-crystallized inhibitor

K00606 (PDB ID: 2VN9, 2.30 Å) was not profiled in either of the aforementioned studies.

In addition to 1, we also tested a derivative of compound 2, another hit that emerged from

the virtual screen. Although 2 may not be suitable for studies in cell culture considering the

presence of an acylhydrazone moiety in its structure, the compound is suitable for binding and

inhibition studies at the near neutral pH values used to conduct the enzyme activity studies for

this work. We selected an analog of 2, namely 14, and tested it for activity against 337 kinases.

The compound inhibited 46 kinases by 50% or more at 20 M. Concentration-dependent studies

of the top 25 kinases most potently inhibited by 14 show that all 25 kinases were inhibited with

IC50s of 10 M of less. Interestingly, when the 337 kinases are ranked by percent inhibition at

20 M, CaMKII is rank 46 on the list. It is interesting to note that 14 inhibited two kinases with

sub-micromolar IC50s, BLK (IC50 = 0.51 M) and BMX/ETK (IC50 = 0.98 M).

To further gain insight into the basis for the preference of 6 towards CaMKIIs, we resorted

to explicit-solvent molecular dynamics simulations followed by end-point free energy calculations

of 1 bound to 6 kinases selected from among the top 25 targets of its close analog 6. We find that

among the most highly favorable interactions of the compound are Val-93 and Glu-97. A valine

residue at position 93 is only present in the three CaMKs. In the other tyrosine kinases, this

position is replaced with less hydrophobic residues and the interaction energy is less favorable.

Position 97 also appears to be unique among CaMKs. A glutamic acid occupies this position in

contrast to the overwhelming majority of the other kinases that possess other neutral residues at

that position (IR is the exception). Glu-97 interacts with the flexible N-alkylamine moiety of 1.

ChemMedChem

10.1002/cmdc.201600636

To explore the moieties on the compound that may be responsible for activity, we

investigated the activity of 1 and its derivatives against 11 of the top kinases that were inhibited

by 6. The data overall showed striking specificity of compounds towards members of the CaMK

family, consistent with the kinome profiling above. Compounds lacking the alkylamine moiety of

6

and its parent structure 1 had significantly less activity against members of the CaMK family.

This suggests that this moiety, which forms a salt-bridge with the conserved DFG motif in our

modeled binding pose, is essential for activity. Molecular dynamics simulations followed by end-

point free energy calculations confirmed these results. The excellent correlation coefficient

between the experimental IC50 and predicted binding affinities provides additional support for the

accuracy of the predicted binding mode. The calculations show that the electrostatic component

of the free energy of binding becomes dramatically more favorable when the salt-bridge of the

alkylamine moiety is present. Generally, there is a large penalty associated with desolvation of

charged groups from solution. However, the fact that the salt bridge is nearly completely solvent

exposed means that the binding of 1 will result in minimal loss of solvation energy and a significant

gain in free energy as a result of the favorable intermolecular interaction of the salt bridge.

Decomposition energies, which measure the internal and solvation components between

the compound and individual residues, reveal the importance of three hydrophobic residues: Leu-

20, Val-28, and Leu-143. Leu-20 interacts with the benzene ring substituent that is attached to

the core quinazoline, while Val-28 and Leu-143 form hydrophobic interactions with the quinazoline

core. Examination of the dynamics of the compounds on CaMKII reveal relatively similar

dynamics across 1 and its derivatives. Positive correlations between the motion of the protein and

the compound are observed at residues surrounding the binding pocket. At the DFG motif of the

activation loop where the salt-bridge is formed, a region of positive correlation becomes smaller

when the N-alkylamine moiety is removed.

In addition to free energy calculations, we also explored the effect of 6 on the dynamics of

the 6 kinases. This was done by determining the correlation of the motion of 6 with each residue

of the target. Correlation coefficients provide a measure of global dynamical changes. Every

protein exhibits unique collection of global dynamics that are driven by the cross-correlation of all

residues within the protein. Even proteins that possess different sequences but adopt the same

fold are expected to exhibit different global dynamics. Interestingly, it appeared that CaMKII

showed the most distinct pattern of correlation. Hence, it is likely that the specificity observed for

6

towards CaMKII across the kinases tested may be partly due to the differences in the global

dynamical changes of the target proteins.

ChemMedChem

10.1002/cmdc.201600636

MATERIALS AND METHODS

Virtual Screening. The virtual screening workflow is described in Fig. 1a. The Protein

Preparation Wizard workflow from Schrödinger (Schrödinger LLC, New York, NY) was employed

to prepare protein structures. The crystal structure of CaMKII (PDB ID: 2WEL, 1.90 Å) was

imported into the workspace from the Protein Data Bank (PDB).⁶⁵The auto-inhibitory and

calmodulin-binding domains of CaMKII (residues 276-334), calmodulin (chain D), solvent

molecules, and inhibitors were removed. The structure was protonated, neutralized, and

minimized using Impref. Both the protein and ligand structures were loaded into AutoDockTools,²¹

where the structures were converted to united-atom models and saved in PDBQT format. The

docking box was centered on the co-crystallized ligand in the ATP binding site with dimensions

of 21 Å × 17.25 Å × 18.75 Å. All other parameters of the docking program were set to default

22

values. The docked conformations from AutoDock Vina (version 1.1.2) were converted with in-

house Python scripts into mol2 format for additional analysis. Approximately 150,000 compounds

from the ChemDiv commercial library⁴²were docked to the ATP-binding site of CaMKII using

Vina and the above parameters. This library was generated by Chemdiv by clustering the entire

ChemDiv library using the SUBSET⁶⁶algorithm at 90% Tanimoto similarity. The library can be

accessed by clicking on the “Distribution & Clustering” tab at the following URL:

http://zinc.docking.org/catalogs/cdiv .

Docked receptor-ligand complexes were rescored using the previously-described 76-

feature SVMSP scoring approach.³³^,³⁴^,³⁸Briefly, this machine learning approach uses predictive

models to rank-order compounds by using knowledge-based pairwise potentials from various

atom types.³³^,³⁹Using these pairwise potentials as descriptors, SVMSP classification model was

trained to discriminate active compounds from inactive compounds based on its 3D complex

structure. The outcome of the SVMSP scoring function is the distance from the hyperplane in the

SVM model, which is used to rank-order compounds. This model has demonstrated promising

performance on the DUD validation set³⁴and was validated by in vitro assays.³³389 co-

crystallized kinase complexes from the PDB (Table S1) and 4,992 randomly selected ligands

docked to the CaMKII binding pocket (Table S2) were used for the positive and negative training

sets of the SVM model, respectively. The atom triplet fingerprints of the top 500 compounds from

the SVM predictions were hierarchically clustered in Canvas⁶⁷to 50 clusters. The compound

corresponding to each cluster center was selected as the representative compound for that

cluster.

ChemMedChem

10.1002/cmdc.201600636

Validation of Scoring Functions Against CaMKII. A set of approximately 2300 compounds

43

were identified to contain biological data against CaMKII in ChEMBL. Compounds were filtered

by the type of activity data and limited to those with direct binding studies (Assay Type = ‘B’) that

reported results as IC50 or K values. Compounds with activities ≤ 1 µM were classified as active

d

and compounds with activities ≥ 10 µM were classified as inactive. Compounds containing pan-

assay interference (PAINS)⁴⁴moieties were discarded. In total, we identified a set of 33 active

and 58 inactive compounds against CaMKII as SMILES strings (Table S3). Compounds were

prepared using LigPrep in Schrödinger at a pH of 7.0. All stereoisomers were generated for each

SMILES string, regardless of existing stereochemistry already present in the string. Compounds

22

were docked to the prepared CaMKII structure using AutoDock Vina and Glide in Standard

Precision (SP) mode.⁶⁸Docked poses were rescored using Glide and the previously generated

SVMSP model. Performance was assessed using receiver operator characteristic (ROC) curves

and the area under the curve (ROC-AUC) as implemented in the Python scikit-learn package.⁶⁹

Performance was assessed for docked poses with and without energy minimization. Energy

minimization of docked poses was carried out using the “Refine Protein-Ligand Complex”

workflow in the Prime module in Schrödinger.⁷⁰The VSGB solvation model⁷¹and OPLS3 force

72

field were used for energy minimization. Only residues within 5 Å of the docked compound were

included. The local optimization algorithm was used to sample the docked structure.⁷³^,⁷⁴

Screening Top Candidates for Activity. Initial high-throughput screens for inhibitors

against CaMKII was carried out using the LANCE Ultra ULight® kinase assays from Perkin Elmer

(

Perkin Elmer Inc., Waltham, MA) according to the manufacturer’s recommendation. This is a

time-resolved fluorescence energy transfer (TR-FRET) assay with a Europium labeled antibody

that recognizes a phosphorylated Ulight-labeled peptide (ULight is a red-shifted dye); TR-FRET

is measured by excitation at 320 nm of Europium and emission at 665 nm from the ULight

acceptor dye using an Envision plate reader. Purified kinases were obtained from Carna

Biosciences (CaMKII) and were used at a final concentration of 20 nM. Testing for compound

activity

against

all

other

kinases

was

done

at

Reaction

Biology

Corp

(

http://www.reactionbiology.com, Malvern, PA). In order to determine which peptide to use for the

different kinases, the LANCE discovery kit (Perkin Elmer) was employed, which includes five

different peptide substrates and covers about 80% of all Ser/Thr kinases. Based on the initial

screen using the LANCE discovery kit of all five substrates against all four kinases, we settled

with the ULight-CREBtide peptide for all kinases. Kinase assay buffer consisted of 50 mM Tris-

HCl, pH 7.5, 10 mM MgCl , 1 mM EGTA, 2 mM DTT and 0.01% Tween-20. The screen with

2

ChemMedChem

10.1002/cmdc.201600636

inhibitors was done using 384-well plates for all kinases against compounds at a final

concentration of 20 M, using 400 M ATP and 100 nM peptide concentration. Kinase reactions

were incubated for 60 min at 23 C and stopped by the addition of 10 mM EDTA. The Eu-anti-

phospho-MBP antibody diluted in detection buffer was then added to a final concentration of 2

nM. The addition of all components for the high-throughput screen was done using a Freedom

Evo robot from Tecan. All curves were fitted with the program SigmaPlot 11 (Systat Software,

Inc, San Jose, CA, 2008), which provides an estimate of the IC50 values.

42

Identification of Compound Analogs. The ZINC (http://zinc.docking.org) website was

used to identify analogs of active compounds. The ZINC website uses Extended Connectivity

Fingerprints (ECFP4) and Tanimoto coefficient to identify analogs. A similarity search was

conducted using compound SMILES strings and an 80% similarity threshold. Several analogs

were selected and purchased.

Binding Mode Prediction. The program AutoDock Vina was used as the docking

program for all the virtual screening carried out in this work. To further validate the binding modes

that emerged from AutoDock Vina during the virtual screening, we used the Glide docking

program. Glide is a robust docking program that has been extensively validated in previous

studies. The structures of 1 (Scheme 1) and its 8 derivatives (Scheme 3) were prepared using

LigPrep in Schrödinger. Compounds were protonated at pH 7. Additionally, the structures of

CaMKII (PDB ID: 2V7O, 2.25 Å), and MELK (PDB ID: 4CQG, 2.57 Å) were prepared following

the protocol used to prepare CaMKII for virtual screening. Grid boxes were generated for the

three proteins for docking in Glide,⁶⁸with a box size of 21 Å and an inner box of 14 Å. All other

parameters were set to default values. A series of binding poses were generated by docking 1

and its derivatives against the ATP binding site of each of the three proteins using Glide in

standard precision (SP) mode. All other parameters were left to default values. A consensus

binding mode was identified by visual inspection. 43 co-crystallized quinazoline-containing

compounds that bound to the ATP site of kinases were identified using a substructure search

against the PDB. Structures were retrieved and aligned against the CaMKII structure using the

built-in align function in PyMOL (The PyMOL Molecular Graphics System, Version 1.5

Schrödinger, LLC.). The orientation of the quinazoline ring of each co-crystallized complex was

visually inspected and compared against the Glide-predicted binding mode of 1.

Analysis of Binding Mode Across Multiple Kinases. Structural representatives of the

protein kinase domain of each kinase was identified from the PDB. Each structure was aligned to

Article Doi

Structure-Based Target-Specific Screening Leads to Small-Molecule CaMKII Inhibitors

DOI: 10.1002/cmdc.201600636

Source and publish data:

Authors:

Article abstract of DOI:10.1002/cmdc.201600636

Full text of DOI:10.1002/cmdc.201600636

Products guided by the article

R&D Labs maybe for 127033-30-1

Relevant to this article

Hot Product