Structural Basis of Activin Receptor-Like Kinase 2 (R206H) Inhibition by Bis-heteroaryl Pyrazole-Based Inhibitors for the Treatment of Fibrodysplasia Ossificans Progressiva Identified by the Integration of Ligand-Based and Structure-Based Drug Design Approaches.

Fibrodysplasia ossificans progressiva (FOP) is a rare but severe genetic disorder in which acute inflammation elicits progressive heterotopic ossification in the muscles, tendons, and ligaments. Classic FOP is caused by the R206H mutation in ALK2/ACVR1. While several activin receptor-like kinase 2 (ALK2) inhibitors were found to be efficacious in animal models of FOP, most of the ALK2 (R206H) inhibitors lacked sufficient oral bioavailability for efficacy. Previously, the synthesis of a series of novel bis-heteroaryl pyrazole-based ALK2 (R206H) inhibitors that achieved both substantial potency and an improved ADMET profile was reported. In the present study, the detailed procedure of the in silico approach employed to identify the initial bis-heteroaryl pyrazole-based ALK2 (R206H) inhibitor RK-59638 and the analysis of the ALK2 (R206H) RK-59638 complex structure to guide the synthetic optimization of the chemical series, obtaining RK-71807 showing improved potency and metabolic stability, were described. According to the initial in silico screening, the screening efficiencies and chemical diversity of the hit compounds of both ligand-based and structure-based methods were evaluated. Then, X-ray structures of ALK2 (R206H) and the inhibitors were analyzed to assess the structure-activity relationships of the synthesized compounds. The 3D-RISM analysis indicated the existence of the additional hydrogen bond via water molecules restricting the attachment point in the pyrazole scaffold. The quantum mechanics calculation of the newly determined ALK2 (R206H) RK-71807 complex structure using a fragment molecular orbital method and pair interaction energy decomposition analysis was employed to evaluate the interaction energies between the inhibitor and each of the amino acid residues and decompose them to electrostatic, exchange-repulsion, and charge transfer energies. The pattern of decomposed interaction energies was then compared to that formed by RK-59638 and LDN-193189 to investigate the structural basis of ALK2 (R206H) inhibition.

Introduction

Fibrodysplasia ossificans progressiva (FOP) is a rare but severe genetic disorder in which acute inflammation elicits progressive heterotopic ossification in the muscles, tendons, and ligaments. Classic FOP is caused by a gain-of-function mutation (617G > A; R206H) in activin receptor-like kinase 2 (ALK2) encoded by the ACVR1 gene, a subtype of the BMP type-I receptors.[1] The R206H mutation enhances BMP signaling in FOP patients by decreasing the inhibitory activity of the ALK2 interacting protein, FK506 binding protein 12 (FKBP12).[2] The crystal structure of the cytoplasmic domain of ALK2 in complex with FKBP12 and the ALK2 inhibitor dorsomorphin revealed that FOP mutations destabilize the inactive state of the kinase, resulting in structural rearrangements that prevent FKBP12 binding and promote the correct positioning of the glycine–serine-rich loop and α-C helix for kinase activation.[3]

These data indicated that small-molecule BMP type I receptor inhibitors could be useful as both therapeutic agents and chemical tools to investigate cellular signaling.[4,5] Dorsomorphin was discovered as the first small-molecule BMP receptor inhibitor by phenotypic screening using zebrafish embryos.[6] The crystal structure of human ALK2 confirmed the direct binding of dorsomorphin to the ATP-binding pocket in the kinase domain.[3] Further development of the pyrazolo[1,5-a]pyrimidine-containing scaffold has produced a series of derivative compounds, including LDN-193189 and LDN-212854.[7−10] LDN-193189 exhibited improved potency and pharmacokinetic properties as compared to dorsomorphin[7] and inhibited heterotopic ossification in FOP mouse models.[11] LDN-212854 was recently reported as a highly selective inhibitor for ALK3 and ALK2.[10] Another series of pyridine-based, small-molecule ALK2 inhibitors was identified from a biochemical screen against the ALK2 kinase domain. The initial screening hit K02288 showed high selectivity for BMP receptors in a panel of approximately 250 human kinases.[12] A derivative of K02288, LDN-214117, displayed superior potency in cells, with more than 100-fold selectivity for ALK2 over ALK5.[13]

While several ALK2 inhibitors showed efficacy in animal models of FOP, most of the ALK2 (R206H) inhibitors lacked sufficient oral bioavailability for efficacy. For example, LDN-193189 and LDN-212854 are potent inhibitors of ALK2 (R206H) (IC50 = 20 and 11 nM, respectively) but are poorly soluble in aqueous solutions and highly bound to serum proteins.[14]

We previously identified a series of novel bis-heteroaryl pyrazole-based ALK2 (R206H) inhibitors[15] with substantial potency and improved ADMET profiles and reported their synthetic procedures.[14] In this study, the details of the in silico approach employed to detect the initial bis-heteroaryl pyrazole-based ALK2 (R206H) inhibitor, RK-59638, from 142,785 compounds in the Drug Discovery Initiative compound library at The University of Tokyo and the analysis of the ALK2 (R206H) RK-59638 complex structure to guide the synthetic optimization to improve its ALK2 (R206H) potency, aqueous solubility, and metabolic stability were presented. Since the usefulness of ensemble docking using multiple structures to take receptor flexibility into account was reported,[16] various in silico methods including ligand-based drug discovery (LBDD) using molecular fingerprints and structure-based drug discovery (SBDD) by docking simulations using multiple ALK2 structures were combined to obtain diverse hit compounds in this study. The efficiency of each in silico method was assessed based on the high-throughput screening results. Then, the analysis of the X-ray structure of the RK-59638 ALK2 (R206H) complex employed to guide the previously reported derivatization of RK-59638, obtaining improved inhibitors such as RK-71807 (ALK2 (R206H) IC50 = 9.4 nM; metabolic stability in liver microsomes = 86.5% (human) and 79.9% (mouse) remaining after 60 min), was explained. The structure–activity relationships of synthesized inhibitors were well consistent to the insight from the X-ray structure. The structural basis of ALK2 (R206H) inhibition by RK-71807 was then analyzed by quantum mechanics calculation using the fragment molecular orbital method to assess how the additional chemical moieties of RK-71807 contributed to its improved ALK2 (R206H) potency.

Results and Discussion

Ligand-Based Drug Design

To detect novel ALK2 (R206H) inhibitors, 142,785 compounds in the Drug Discovery Initiative compound library at The University of Tokyo were used for the similarity search. After removing non-drug-like compounds, 137,456 compounds were selected for analysis. As query compounds, 16 ALK2 inhibitors and 220 inhibitors of other ALK family proteins (ALK1, ALK3, ALK4, ALK5, and ALK6) were derived from the ChEMBL database with the IC50 ≤ 10 μM criterion. Although some ALK2 selective inhibitors such as LDN-212854 were previously reported, the complex structure of ALK2 and LDN-212854 (PDB ID: 5OXG) revealed that the inhibitor selectivity between ALK2 and other ALK family proteins was achieved by utilizing a slight difference in the position of a water molecule in the hydrophobic pocket. The amino acid residues in the binding site and the majority of the interactions by known inhibitors of the ALK family proteins have been discussed.[10] Thus, the inhibitors of other ALK proteins were also used as the query compounds to discover novel ALK2 inhibitors in this study. To evaluate the molecular similarity from various aspects, a similarity search using four structural fingerprints (MDLPublicKeys, ECFP4, FCFP4, and GpiDAPH3) calculated by Pipeline Pilot,[17] a discrimination model by random forest (RF) using the 299 molecular descriptors calculated by MOE,[18] and a 3D similarity comparison by ROCS[19] were employed. By merging the compounds selected by each method, 976 compounds were selected for the initial screening assay based on an ALK2 (R206H) kinase assay at 4 μM.

Structure-Based Drug Design

As the receptor structures for the SBDD approach, the complex structures of ALK2 and dorsomorphin (PDB ID: 3H9R), ALK2 and the imidazo[1,2-b]pyridazine inhibitor K00507 (PDB ID: 3OOM), and ALK2 and the 2-aminopyridine inhibitor (3MTF) were used for docking simulations by the GLIDE SP mode.[20] The scaling factors of the grid calculation in GLIDE were tuned according to the test docking between each ALK2 structure and known inhibitors. Excluding the compounds already selected by the LBDD method, the Glide SP scores were calculated for 137,456 compounds in the Drug Discovery Initiative compound library. After filtering the docking poses that did not form hydrogen bonds with the hinge region of ALK2, the compounds were clustered based on MDLPublicKeys to obtain structurally diverse compounds. By selecting the high GlideScore compounds from each cluster, 1428 compounds were chosen for the initial screening based on an ALK2 (R206H) kinase assay at 4 μM.

ALK2 (R206H) Kinase Assay

For the 976 compounds selected by the LBDD method and 1428 compounds selected by the SBDD method, the ALK2 (R206H) inhibitory activities were evaluated by an ALK2 (R206H) kinase assay at a 4 μM compound concentration (Figure ). When a criterion of 50% inhibition at 4 μM was applied, 22 of the 973 compounds selected by LBDD and 37 of the 1428 compounds selected by SBDD were identified as the initial hit compounds. The top 14 compounds showing the highest potencies detected by the LBDD and BDD approaches are listed in Figure a,b, respectively. The hit rates in the ALK2 (R206H) kinase assay of 976 compounds selected by LBDD and 1428 compounds selected by SBDD were 2.25 and 2.59%, respectively. Welch’s t-test at a 0.05 significance level concluded that no statistically significant difference existed between the distributions of the inhibitory activities of the two compound sets.

Figure 1

Inhibitory activities of 2404 compounds in the initial screening by the ALK2 (R206H) kinase assay. The red dotted line represents the criteria for the hit compounds (50% inhibition at 4 μM) in this study.

Figure 2

Top 12 ALK2 (R206H) inhibitors detected by (a) the LBDD approach and (b) the SBDD approach. The inhibitory activities for R206H ALK2 were measured at a 4 μM compound concentration.

To evaluate the efficiency of each LBDD method, the hit rates for the respective LBDD methods were assessed, as listed in Table . As expected, the similarity search using ALK2 inhibitors as the queries showed a slightly higher hit rate than that using inhibitors of other ALK family proteins in both 2D structural fingerprint similarity (3.54 to 3.06%) and 3D shape similarity (2.48 to 1.44%). The random forest model was built using all inhibitors and achieved the highest hit rate among the LBDD methods in this study (5.51%). In terms of the relationship between the LBDD methods and potency, there was no obvious trend that a certain method could detect highly potent compounds more efficiently than the others. On the other hand, all of the top five compounds that showed more than 90% inhibition of ALK2 (R206H) were selected by more than two methods, indicating the efficiency of employing multiple methods to evaluate the compound similarity. Despite the slightly lower hit rate observed in this study, the 3D shape comparison was often employed for scaffold hopping in previous studies because of its independence from substructures in the query compound.[21] To assess the reputations, the distributions of the structural similarities between hit compounds obtained by each LBDD method and the query molecules were compared, as shown in Figure . The median Tanimoto coefficient values between hit compounds and query molecules were 0.532 and 0.52 for the 2D fingerprint similarity search and 3D shape comparison using ALK2 inhibitors as the query, respectively, and 0.788 and 0.756 for those using other ALK inhibitors. For both query sets, the 3D shape comparison detected slightly more novel compounds, supporting its usefulness to identify structurally diverse hit compounds for scaffold hopping.

Table 1

Number of Hit Compounds and Selected Compounds and Hit Rates by each LBDD Methoda

	no. of hit compounds/selected compounds (hit rate)
query compounds	2D fingerprint similarity	3D shape similarity	random forest model
ALK2 inhibitors	4/113 (3.54%)	5/202 (2.48%)	13/236 (5.51%)
Inhibitors of other ALK family members	7/229 (3.06%)	11/764 (1.44%)

The number of compounds was redundantly counted if a compound was selected by multiple methods.

Figure 3

Structural similarities between the hit compounds and the query molecules for each LBDD method.

The ALK2 (R206H) kinase assay at 4 μM generated 9 hit compounds out of the 439 selected compounds selected by Glide SP docking for 3H9R (hit rate = 2.05%), 12 out of 477 compounds using 3MTF (2.52%), and 16 out of 499 using 3OOM (3.21%). The slightly lower hit rate for 3H9R could be due to the exclusion of the compounds selected by LBDD methods from the docking calculation. While dorsomorphin, the ligand bound in 3H9R, was used as the query compound for LBDD, the ligands in 3MTF and 3OOM were not included in the query compounds for LBDD because of the different implementation dates of the data preparation between the LBDD and SBDD methods. Thus, dorsomorphin-like inhibitors tend to be removed from the docking simulation more frequently, possibly decreasing the screening efficiency of Glide docking using the 3H9R structure. This assumption was supported by the distribution of the molecular similarities as the Tanimoto coefficient of the MACCS key fingerprints of the bound ligand in each X-ray structure and the hit compounds derived by the corresponding docking simulation, as shown in Figure . While the hit compounds derived from docking with 3MTF and 3OOM included highly similar compounds to the corresponding bound ligands K02288 and K00507, the hit compounds derived from 3H9R showed slightly lower similarities to dorsomorphin. The relationship between the bound ligands in the X-ray structures and the hit compounds derived from the docking simulation also indicated the efficiency of the ensemble docking using multiple receptor structures for higher structural diversity in the hit compounds.

Figure 4

Distribution of molecular similarity between X-ray ligands and hit compounds derived from Glide docking using each structure.

Among the 59 hit compounds, the 6-(1H-pyrazol-5-yl)pyrimidine scaffold was found in five compounds, including one of the most potent inhibitors, RK-59638. Considering the consistently high potency, patentability, and synthetic feasibility of the chemical scaffold, RK-59638 was selected as the seed compound for further development.

As previously reported, the IC50 value for ALK2 (R206H) and the ADMET profile of RK-59638 were measured to assess its potential as a drug candidate for FOP.[14] The IC50 value of RK-59638 to ALK2 (R206H), measured by the in vitro enzymatic assay, was 684 nM (Figure ). In the ADMET profile, RK-59638 showed a good plasma protein binding ratio (85%), membrane permeability (29.2 × 10–6 cm/s), in vitro toxicity (78% viability at 30 μM for HepG2 cells), and hERG potassium channel inhibition (25.4% inhibition at 10 μM). The low cell toxicity and hERG inhibitory activity of RK-59639 are especially attractive for drug safety. However, RK-59638 had issues in terms of aqueous solubility (6.4 μg/mL), CYP inhibition (CYP1A2: 84%, 2C9: 40%, 2C19: 30%, 2D6: 4%, and 3A4: 71% inhibition at 10 μM), and metabolic stability in liver microsomes (human: 40% and mouse: 0% remaining after 60 min). Thus, structural modifications of RK-59638 were planned in order to improve its binding affinity to ALK2 (R206H), aqueous solubility, and metabolic stability.

Figure 5

Dose–response curve of ALK2 inhibition by RK-59638 in the in vitro enzyme assay (n = 2).

Structural Basis for RK-59638 Binding to ALK2 (R206H)

The X-ray complex structure between RK-59638 and the kinase domain of ALK2 (R206H) (PDB ID: 6ACR, traced from amino acid residue Thr203 to Ile498 at a 2.01 Å resolution)[14] is shown in Figure a. The amino pyrimidine of RK-59638 formed hydrogen bonds with the nitrogen atom of the H286 main chain at the hinge region of ALK2. The pyridine ring occupied the hydrophobic pocket and formed a hydrogen-bonding network with Glu248 via a water molecule, which is known to play an important role in ALK2 selectivity of LDN-212854.[22] The pyrazole ring, connecting the pyrimidine and pyridine rings, formed a CH−π interaction with Val222. Since the nitrogen atoms in the pyrazole ring and terminal methoxy phenyl ring did not seem to form obvious interactions with the receptor atoms in the current crystal structure, these atoms were selected as the attachment points for the structural modifications of RK-59638. The hydrophobic and hydrophilic regions around the nitrogen atoms in the pyrazole ring and the methoxy phenyl were calculated using the Contact Preference function implemented in MOE (Figure a,b). In front of the pyrazole ring, a hydrophobic region formed by Val222 and Tyr219 was predicted. Near the methoxy phenyl, a hydrophilic region between the main-chain oxygen of Val214 and the side chain of Asp293 was found. Hydrophilic regions corresponding to the oxygen atom at Tyr219 and Lys235 were also observed in a position slightly distant from the pyrazole ring.

Figure 6

(a) Complex structure of ALK2 (R206H) and RK-59638 (PDB ID: 6ACR). (b) Hydrophobic (green mesh) and hydrophilic (purple) areas around the pyrazole ring. (c) Hydrophobic (green mesh) and hydrophilic (purple) areas around the methoxy benzene.

Structure–Activity Relationship of RK-59638 Derivatives

Based on the analysis of the X-ray structure, structural modifications of RK-59638 at the nitrogen atoms in the pyrazole ring and at the para-position of the phenyl ring described in a previous paper[14] were performed to improve the potency with ALK2 (R206H) and metabolic stability by addition of a hydrophilic chemical group and an alkyl chain to reduce the planarity of the compound.

To assess the consistency between the structure–activity relationship of the synthesized compounds and the binding mode of RK-59638, the potency of the key compounds summarized in Table was assessed. In N-alylation of the pyrazole ring, the ethyl group at the N2 position (2a) showed the highest potency. As compared to RK-59638, the IC50 of 2a was improved by approximately 6.5-fold from 683.7 to 94.3 nM. On the other hand, the same ethyl modification at the N1 position (2b) resulted in the complete loss of inhibitory activity. In the complex structure of RK-59638 and ALK2 (R206H), the distance between the N1 nitrogen and Lys235 was 5.62 Å. Assuming the existence of a hydrogen bond between the N1 nitrogen and Lys235 bridged by a water molecule, a 3D-RISM solvent analysis was performed using MOE.[23] 3D-RISM employs a first-principles theory of solvation based on the density functional theory of liquids and can estimate the probability density distribution of non-obvious water molecules. The calculation predicted the high-probability region for a water molecule between the N1 nitrogen and Lys235, as shown in Figure , supporting the hypothesis that the N1 nitrogen forms a hydrogen-bonding network with Lys235 via a water molecule. This result indicated that the N1 nitrogen should not be modified. In the N2 position, 2a with an ethyl group recorded the highest potency among the derivatives and thus was selected as the starting compound for the second modification at the methoxy phenyl.

Table 2

Potencies of the Synthesized Compounds against ALK2 (R206H)

Figure 7

High-probability region for a water molecule around the N1 nitrogen (PDB ID: 6ACR) calculated by 3D-RISM. The predicted hydrophilic region is depicted as a blue contour.

Among the modifications to the p-methoxy group at the phenyl moiety, 1 with a piperazine group at the corresponding position exhibited the highest potency for ALK2 (R206H) with an IC50 value of 47.9 nM. Considering the existence of an acidic region around Asp293 in 6ACR, the cationic nitrogen in the piperazine group was expected to form an electrostatic interaction with Asp293, as confirmed by the subsequent X-ray crystallographic analysis. The two modifications were then merged as RK-71807 with the ALK2 (R206H) IC50 of 9.4 nM, representing increases of 70-fold over RK-59638 (683.7 nM) and 2-fold over the reference inhibitor LDN-193189 (20 nM).

In the aspect of ADMET profiles, RK-71807 showed improved metabolic stability compared to RK-59638 (from 40.0%/0.0% to 86.5%/79.7% remaining in human/rat liver microsomes after 60 min) as reported in previous study.[14] In addition, the detailed ADMET parameters of RK-71807 were measured and compared to those of RK-59638 and LDN-193189 as shown in Table . Due to the addition of the polar substituent (piperazine ring) and the reduced planarity (ethyl group) indicated by the ALK2 (R206H) RK-59638 complex structure, RK-71807 showed increased aqueous solubility (93.8 μg/mL) to RK-59638 (6.4 μg/mL). The plasma protein binding ratio and the cytochrome P450 inhibition also improved from 85% (RK-59638) to 65.8% (RK-71807) and from 84% (RK-59638 to CYP1A2) to 22% at most (RK-71807 to CYP2C19) at 10 μM. RK-71807 also maintained lower hERG inhibition (27.0% at 10 μM) compared to that of LDN-193189 (81.2% at 10 μM).

Table 3

R206H ALK2 IC50 and ADMET Profiles of LDN-193189, RK-59638, and RK-71807a

conditions	LDN-193189	RK-59638	RK-71807
aqueous solubility at pH 7.4 (μg/mL)	10.6	6.4	93.8
partition coefficient (log D, n-octanol/PBS, pH 7.4)	2.26	2.71	0.5
PPBb human (%) at 10 μM	97.8	85.0	65.8
CYP1A2, 2C9, 2C19, 2D6, 3A4 (% inh. at 10 μM)	28, 30, 34, 30, 82	84, 40, 30, 4, 71	2, 6, 22, 9, 5
hERG % inhibition at 10 μMc	81.2	25.4	27.0
in vitro toxicity HepG2 cell (% viability at 30 μM)	77	78	82

All studies were conducted by Cerep (United States) using in vitro ADME/Tox assays.

Plasma protein binding.

hERG inhibitory activity was measured by an automated patch-clamp assay. HLM = human liver microsomes, MLM = mouse liver microsomes.

Structural Basis of RK-71807 Binding to ALK2 (R206H)

To investigate the interaction between RK-71807 and ALK2 (R206H) proteins, the complex structure was determined by X-ray crystallography (PDB ID: 6JUX), as shown in Figure a. The piperazine group of RK-71807 is located near Asp293, which is widely conserved in protein kinases. The distance between the charged nitrogen in the piperazine group and the oxygen atom of the carboxylic acid in Asp293 is 4.02 Å, allowing the formation of an electrostatic interaction. The N-ethyl group attached to the pyrazole is located near Tyr219. The carbon atoms at the ethyl group and aromatic ring are 3.69 Å apart, and they form a CH−π interaction. In this structure, the electron density of a water molecule was observed between the N1 nitrogen of the pyrazole ring and Lys235, and thus a hydrogen bond network exists between the N1 nitrogen and Lys235, as assumed in the RK-59638 and ALK2 (R206H) complex structure. To compare the binding modes, LDN-193189 was superposed to RK-71807 in the ALK2 (R206H) complex structure (Figure b). Both inhibitors shared a hydrogen bond with Glu248 via water molecules, a hydrogen bond with His286 at the hinge region, and an electrostatic interaction with Asp293. On the other hand, the hydrogen bond with Lys235 via a water molecule and the cation−π interaction with Tyr219 were only observed in the case of RK-71807, possibly contributing to its enhanced potency relative to LDN-193189.

Figure 8

(a) X-ray structure of the RK-71807 and ALK2 (R206H) complex (PDB ID: 6JUX). (b) Superposition of RK-71807 (blue) and LDN-193189 (yellow).

FMO Analysis

To compare the interactions between RK-71807 and ALK2 (R206H) to those of RK-59638 and LDN-193189, the fragment molecular orbital (FMO) method[24−29] and pair energy decomposition analysis (PIEDA)[30,31] were employed. FMO assesses the electronic states of a protein–ligand complex with hundreds of amino acid residues. Furthermore, the contribution of the interaction between each fragment pair can be efficiently assessed by the interfragment interaction energy (IFIE). The calculated IFIE values were then decomposed to electrostatic (ES), exchange-repulsion (EX), dispersion interaction (DI), and charge transfer energies with higher-order mixed terms (CT + mix) by PIEDA. To clarify the effect of the piperazine group and ethyl group in RK-71807, the interaction energies between the three inhibitors and the amino acid residues within 4.5 Å from the substituents were assessed, as shown in Figure . The decomposed energies between the ligands and the residues around the piperazine group are depicted in Figure b. As compared to RK-59638, large improvements in the electrostatic energies with Asp293 were observed in both RK-71807 and LDN-193189 (−77.467 kcal/mol for RK-71807, −5.034 kcal/mol for RK-59638, and −59.189 kcal/mol for LDN-193189), indicating that the addition of the piperazine group successfully accommodated the acidic environment around Asp293. Regarding the ethyl group, the interaction with Tyr219 was improved by the dispersion energy (−3.34 kcal/mol) through the CH−π interaction where both RK-59638 and LDN-193189 formed few interactions. On the other hand, electrostatic repulsions with Lys340 were observed in RK-71807 and LDN-193189. Since the FMO calculation was performed in the gas phase, the electrostatic interactions between remote fragments were often overestimated due to the lack of a water shielding effect. Considering the 11.1 Å distance between the positively charged nitrogen atoms in Lys340 and the piperazine group, this repulsion would not become effective in solution. These results suggested that the additional piperazine and ethyl group both contributed to the binding affinity, mainly by the electrostatic interaction with Asp293 and the dispersion energy by the CH−π interaction with Tyr219, respectively, as expected from the X-ray structure of RK-59638.

Figure 9

PIEDA analyses of RK71807, RK-59638, and LDN-193189. (a) Analyzed residues around the piperazine group (Val214, Met288, Gly289, and Asp293) and those around the ethyl group (Tyr219, Val222, and Lys340) in the RK-71807 and ALK2 (R206H) complex (PDB ID: 6JUX). (b) PIEDA results with the residues around the piperazine group of RK-71807. (c) PIEDA results with the residues around the ethyl group of RK-71807.

Conclusions

In this study, in silico methods applied to obtain novel ALK2 (R206H) inhibitors using both the LBDD and SBDD methods and analysis of X-ray structures to assess the structural basis of the SAR of bis-heteroaryl pyrazole-based inhibitors were presented. Among the LBDD methods, the random forest model based on MOE 2D descriptors recorded the highest hit rate of 5.51%. For the selection of query compounds, a similarity search using the existing ALK2 inhibitors was slightly more successful than that using the inhibitors of other ALK family proteins, as expected. However, the difference in the hit rates between the two query sets was marginal: 3.54% for ALK2 inhibitors versus 3.06% for inhibitors of other ALK family proteins using 2D fingerprint similarity and 2.48% for ALK2 inhibitors over 1.44% for inhibitors of other ALK family proteins using 3D shape similarity. To obtain diverse hit compounds, using the inhibitors of related proteins seemed to be effective, especially when more information about the related protein was available than for the target protein itself (as in the case of ALK5 for ALK2 in this study). The docking simulation by Glide SP using the three ALK2 structures also resulted in comparable hit rates to those obtained by ligand-based methods (2.25% for the LBDD methods and 2.59% for the SBDD methods on average). Although the molecular docking program does not use information about known inhibitors, the docking using the three ALK2 structures revealed a slight correlation between the structures of detected ALK2 inhibitors and those of the bound ligands in the crystal structures, indicating the usefulness of employing multiple structures to detect structurally diverse hit compounds.

The analysis of the RK-59638 ALK2 (R206H) X-ray structure was well consistent to the structure–activity relationship of the chemical series.[14] Among the synthesized compounds, RK-71807, showing the highest potency by combining the ethyl group of the pyrazole ring and replacing the methoxy group with the piperazine group, was further analyzed by X-ray crystallography and FMO calculation. The crystal structure of RK-71807 and FMO calculation revealed that the additional ethyl group and piperazine ring both contributed to the ALK2 (R206H) binding by forming the CH−π interaction with Tyr219 and the electrostatic interaction with Asp293, respectively, as expected from the crystal structure of the initial hit RK-59638. The aspartic acid at the entrance of the ATP binding site is widely conserved among protein kinases, and the electrostatic interactions with cationic nitrogen atoms are often used to improve the potency in various inhibitors, such as ALK2 and LDN-193189[7] and EGFR and gefitinib[32] for example. These two modifications improved the potency to ALK2 (R206H) by about 70-fold from 684 nM (RK-59638) to 9.4 nM (RK-71807). RK-71807 also showed improved aqueous solubility (93.8 μg/mL) and lower CYP inhibition (at most 22% inhibition at 10 μM), maintaining low hERG inhibition (27.0% at 10 μM). These results demonstrated the successful application of the in silico methods to the identification and optimization of the bis-heteroaryl pyrazole-based ALK2 (R206H) inhibitors. Further optimization to achieve sufficient oral bioavailability and efficacy in vivo is to be made in future work.

Experimental Section

The screening compounds were obtained from Drug Discovery Initiative in The University of Tokyo (https://www.ddi.u-tokyo.ac.jp). The library was constructed for the purposes of chemical biology research and drug discovery for the support of the Targeted Proteins Research Program. The compound library was designed for the rapid identification of biologically active compounds for an arbitrary protein or phenotype. As of early 2010, the library contained 142,785 compounds. To remove non-drug-like compounds, the compounds with undesirable substructures such as highly reactive double bonds, metal-containing molecules, compounds violating the loosened Lipinski’s rules of five (the criterion of each rule was loosened from 5 to 6), and potential frequent hitters predicted by PAINS[33] were excluded. As a result, 137,456 compounds were selected as the source for the in silico screening.

As the query compounds for the similarity search, small-molecule inhibitors of ALK2 and other ALK family proteins were derived from the ChEMBL database version 4 in 2010. Using the amino acid sequence of the ALK2 kinase domain, 10 target proteins with e-values lower than e-40 were selected, and their small-molecule inhibitors with IC50 values lower than 10 μM were collected. As the result, 19 inhibitors of ALK2 and 220 inhibitors of other ALK family proteins were selected.

As the LBDD methods, 2D similarity search, a machine learning model based on a 2D structural fingerprint, and 3D molecular shape comparison were employed. To calculate the structural similarity, the Tanimoto coefficients of ECFP4, FCFP4, MDLPublicKeys, and GpiDAPH3 were used as the 2D methods. The similarity search was performed using the query compounds of ALK2 inhibitors and those of other ALK family proteins. The RF model was also built by learning with 239 inhibitors and 5000 decoy compounds randomly selected from the ZINC database. As the explanatory variables to build the RF model, 299 2D molecular descriptors calculated using MOE were employed (Table S1). The 3D shape similarity search was performed using ROCS.[18] The similarity threshold for each method was determined according to the comparison of the similarity values among known inhibitors to the similarities between known inhibitors and decoy compounds randomly selected from ZINC drug-like compounds.

The results of the 2D similarity, machine learning, and 3D shape comparison assessments were merged by a two-step procedure. At first, the primary threshold was defined as the similarity value at which the similarity distribution between known inhibitors was clearly separated from the similarity distribution between known inhibitors and decoy compounds. The compounds with molecular similarity (predicted probability to inhibit ALK2 (R206H) in the case of the RF model) that exceeded the primary threshold were directly selected for the initial ALK2 (R206H) kinase assay. To reduce the structural redundancy, the Tanimoto coefficients of the MACCS key fingerprint between the remaining compounds in the library and the above selected compounds were calculated. The remaining compounds with similarities to the already selected compounds that exceeded 0.8 were excluded in the next procedure. The secondary threshold was then determined as the similarity value at which the distribution of decoy compounds showed a significant decrease and almost disappeared. The compounds that passed the second threshold were selected as the additional compound set to increase the structural diversity and clustered by the k-means method using MDLPublicKeys. The cluster center compounds were merged to the first compound set for the ALK2 (R206H) kinase assay. The details of the distributions of the Tanimoto coefficient for each fingerprint used to determine the threshold are described in the Supporting Information.

Based on the procedure, 232 compounds from the 2D similarity search, 202 compounds from the random forest model, and 193 compounds from the 3D shape similarity satisfied the primary threshold and were merged to 595 unique compounds. Subsequently, 2510 additional compounds satisfying the secondary threshold consisting of 1334 compounds from the 2D similarity, 388 compounds from the random forest model, and 1127 compounds from the 3D shape similarity were clustered by the MDLPublicKeys fingerprint. By selecting the cluster centers, 381 structurally diverse compounds were chosen. In total, 976 compounds were selected from the LBDD approach for the initial screening based on the ALK2 (R206H) kinase assay at 4 μM.

Structure-Based Drug Design

For the docking simulations in the GLIDE SP mode, the complex structures between the ALK2 kinase domain and dorsomorphin (PDB ID: 3H9R), ALK2 and the imidazo[1,2-b]pyridazine inhibitor K00507 (PDB ID: 3OOM), and ALK2 and the 2-aminopyridine inhibitor (3MTF) were used. For the receptor structure preparation, ionization states were generated at pH 7.4, and then the positions of the hydrogen atoms and the whole structures were minimized by an OPLS2005 force field in a sequential order using the Schrödinger Maestro program.[34] In the Glide program, the manipulation of a parameter for the van der Waals radii scaling, the scaling factor, can modify the distance of the Lennard-Jones potential function and avoid overly strict steric repulsion by loosening the potential. For each ALK2 structure, the optimal scaling factor was selected from four settings, 0.7, 0.8, 0.9, and 1.0 (default), to efficiently discriminate the known ALK2 inhibitors from the decoy compounds by GlideScore.

After the parameter optimization, docking simulations were performed for 137,456 compounds in the Drug Discovery Initiative compound library using each of the three receptor structures. To select structurally diverse compounds, the docking poses forming hydrogen bonds with the hinge region of ALK2 were clustered based on MDLPublicKeys. By selecting the high GlideScore compounds from each cluster, 1428 compounds were chosen for the initial screening based on the ALK2 (R206H) kinase assay at 4 μM.

For the structural preparation for the FMO calculation, the X-ray structures of wild-type ALK2 with LDN-193189 (PDB ID: 3U4Q), R206H ALK2 with RK-59638 (PDB ID: 6ACR), and R206H ALK2 with RK-71807 (PDB ID: 6JUX) were processed as follows. For the two R206H ALK2 complexes (6ACR and 6JUX), the missing Thr326 and Gly327 residues were modeled using the wild-type ALK2 structure, 3U4Q, as the template structure. The structures were protonated using MOE at pH 7.0. Several improper ionization states were manually corrected. The heavy atoms of the complemented Thr326 and Gly327 in 6ACR and 6JUX and all hydrogen atoms were then minimized using an AMBER10:EHT force field. Using the three structures as inputs, the automated FMO calculation protocol (“Auto-FMO protocol”)[35] was applied to perform the FMO calculation and PIEDA. The protocol automatically creates input files for the FMO calculation by the ABINIT-MP program.[28,36,37] In this work, all of the FMO calculations were performed at the FMO-MP2/6-31G* level. To characterize the intermolecular interaction energy of the ligand binding site in detail, each complex structure was divided into fragments in the units of each amino acid residue and ligand. The calculated IFIE values were then decomposed into electrostatic (ES), exchange-repulsion (EX), dispersion interaction (DI), and charge transfer energies with higher-order mixed terms (CT + mix) by PIEDA.

Expression and Purification of ALK2 (R206H)

The ALK2 kinase domain, including residues 201 to 499 with the R206H mutation, was expressed with a recombinant baculovirus expression system. Sf9 cells were inoculated at 27 °C and harvested 48 h after infection. The cells were resuspended in 20 mM Tris–HCl, pH 8.0, 500 mM NaCl, 10% glycerol, and 20 mM imidazole and disrupted by sonication. Cell debris and insoluble components were removed by centrifugation. The supernatant was applied to a 5 mL HisTrap HP column (GE Healthcare, Sweden) and eluted with a linear gradient from 20 to 500 mM imidazole in 20 mM Tris–HCl, pH 8.0, 500 mM NaCl, and 10% glycerol. The His tag was cleaved by TEV protease and removed on the HisTrap HP 5 mL column. The buffer was exchanged to 20 mM Tris–HCl, pH 8.0, 500 mM NaCl, and 10% glycerol and 2 mM dithiothreitol (DTT) for further purification. The protein was passed through a HiTrap Q-XL 5 mL column (GE Healthcare) and then purified on a 16/60 HiLoad Superdex 75 prep grade column (GE Healthcare), equilibrated with 50 mM N-(2-hydroxyethyl) piperazine-N′-2-ethanesulfonic acid (HEPES), pH 7.5, 300 mM NaCl, 50 mM l-arginine, 50 mM l-glutamate, and 10 mM DTT.

The initial screening to identify candidate compounds was performed by the Kinase-Glo Luminescent Kinase Assay (Promega, Madison, WI). The reaction buffer contained 10 mM HEPES (pH 7.4), 150 mM NaCl, 20 mM MgCl2, and 0.5% dimethyl sulfoxide (DMSO). The assay mixture contained 1 μM purified ALK2 (R206H) kinase domain (aa. 201-499), 10 μM substrate peptide, 5 μM ATP, and 4 μM each compound. The substrate peptide used was NPISSVS, designed and synthesized to be homologous to the phosphorylated Smad sequence. First, the protein was incubated with each compound for 30 min, and then the kinase reaction was started and continued for 30 min. The efficacies of the compounds were determined by the remaining amounts of ATP, quantified by the luminescent signal after the reaction.

In Vitro Enzyme Assay

To determine the IC50 of the bis-heteroaryl pyrazole-based ALK2 inhibitors, ALK enzyme assays were conducted by the Reaction Biology Corporation using the “HotSpot” assay platform and the kinase assay protocol.[14] This was a 10-point assay, ranging from 100 μM to 0.5 nM, performed in duplicate using the ALK inhibitor LDN-193189 as a control. In the single dose screening, the compound was tested with a single-dose duplicate at a concentration of 0.3 μM. The reaction was performed with a 10 μM ATP concentration and incubated for 2 h.

Chemistry

The compounds were synthesized using previously described methods,[14] as illustrated in Scheme . The nucleophilic displacements of the chloride 3 with anilines afforded 4 and RK-59638. Subsequent Boc deprotection of 4 gave 1. The N-ethylation of the pyrazole RK-59638 afforded 2a as the major product and 2b as a minor product. 2a and its regioisomer 2b were chromatographically separated. As described above, N-alkylation of the pyrazole 1 afforded 5 as the major product. Displacement of the chloride 5 with 1-Boc-4-(4-aminophenyl)piperazine and subsequent deprotection of the Boc-group gave RK-71807. The purities of all tested compounds were determined by LCMS monitoring at 254 nm and confirmed to be more than 95%. LCMS charts of RK-59638 and RK-71807 are shown in the Supporting Information, Figures S2 and S3.

Scheme 1

Reagents and Conditions:[14] (a) THF, rflx.; (b) TFA, CH2Cl2, rt; (c) EtI, K2CO3, DMF, rt; and (d) Amine, MeOH, rlx

All reagents and solvents were obtained from commercial sources and used without purification. 1H NMR spectra were recorded with tetramethylsilane as an internal standard using a JEOL JNM-Ex 270 MHz spectrometer. Automated column chromatographic separations were performed by flash chromatography (Biotage ZIP, Biotage; Isolute SCX-2, Biotage; and amino inject column, Yamazen). LC/MS analysis was performed on a Waters Acquity Ultra Performance LC equipped with a 2.1 mm × 50 mm Waters Acquity UPLC BEH C18 1.7 μm column. The column temperature was 40 °C with a run time of 2 min, flow rate of 0.6 mL/min, and elution solvent composed of a mixture of acetonitrile and water containing 0.1% trifluoroacetic acid with a gradient of 10–90%. The mass spectrometry data were acquired on an SQD2 quadrupole mass spectrometer.

The synthesis of 1 is described in this section. The analytical data for all other compounds (RK-59638, 2a, 2b, 5, and RK-71807) were described in a previous paper.[14]

tert-Butyl-4-((4-((3-pyridin-3-yl)-1H-pyrazol-4-yl)pyrimidin-2-yl)aminophenyl)piperazine-1-carboxylate (4): A solution of compound 1 (212 mg, 0.823 mmol) and 1-Boc-4-(4-aminophenyl)piperazine (269 mg, 0.970 mmol) in THF (1.5 mL) was heated for 24 h in a sealed tube. After cooling to room temperature, the mixture was concentrated in vacuo. The residue was purified twice by column chromatography over amino silica gel (CH2Cl2/MeOH) and silica gel (CH2Cl2/MeOH) to give compound 4 (234.0 mg, 60% yield) as a beige solid. 1H NMR (270 MHz, CDCl3) δ 8.86–8.87 (1H, m), 8.63–8.65 (1H, m), 8.26 (1H, d, J = 5.1 Hz), 8.13 (1H, s), 7.87–7.91 (1H, m), 7.30–7.35 (3H, m), 7.10 (1H, s), 6.83 (2H, d, J = 8.9 Hz), 6.21 (1H, d, J = 5.1 Hz), 3.59 (4H, t), 3.06 (4H, t), 1.49 (9H, s). LCMS (ESI): m/z 499.5 [M + H]+.

N-[4-(Piperazin-1-yl)phenyl]-4-(3-pyridin-3-yl-1H-pyrazol-4-yl)pyrimidin-2-amine (1): To a solution of Boc-protected 2 (70 mg, 0.140 mmol) in CH2Cl2 was added trifluoroacetic acid (53.6 μL, 79.8 mg, 0.700 mmol) in an ice bath. The mixture was warmed to room temperature and stirred for 16 h. The solution was concentrated in vacuo. The residue was purified on an SCX ion-exchange column (MeOH/NH4OH) to give compound 1 (40 mg, 71% yield) as a yellow solid. 1H NMR (270 MHz, CDCl3) δ 9.15 (1H, s), 8.69 (1H, d, J = 1.6 Hz), 8.53–8.56 (1H, m), 8.31 (1H, s), 8.28 (1H, d, J = 3.0 Hz), 7.90–7.95 (1H, m), 7.39–7.43 (H, m), 7.24 (2H, d, J = 8.6 Hz), 6.73 (1H, d, J = 5.1 Hz), 6.63 (1H, d, J = 8.9 Hz), 2.88–2.92 (4H, m), 2.78–2.82 (4H, m). LCMS(ESI): m/z 399.4 [M + H]+. Retention time: 0.60 min. Purity: 95%.

X-ray Crystallography

The best crystals of ALK2 R206H were grown in 100 mM HEPES, pH 7.6–8.0, containing 1.5–1.6 M ammonium sulfate, using the sitting-drop vapor diffusion method at 20 °C. Each drop contained 500 nL of 10 mg/mL protein solution and an equal volume of the crystallization reagent. Streak seeding facilitated the growth of crystals that were tolerant to ligand soaking and cryoprotection. The crystals were soaked in mother liquor supplemented with 1 mM RK-71807 for 2 h. The crystals were then soaked in a solution containing 25% ethylene glycol and 75% mother liquor for cryoprotection and flash-cooled in a nitrogen stream. Diffraction data were collected at the beamline X06DA of the Swiss Light Source, Paul Scherrer Institut (Villigen, Switzerland), at 100 K and a wavelength of 1.000 Å and processed with XDS.[38] For structure determination, CCP4 suite software[39] was used. The structure of ALK2 R206H was determined by molecular replacement using MOLREP.[40] The coordinates of the ALK2 Q207D mutant in complex with an inhibitor (PDB ID: 3MTF) were used as the search model. Structure refinement was performed using REFMAC[41] with iterative manual model inspections using COOT.[42]Table shows a summary of the data collection and structure refinement results. The atomic coordinates and structure factors of ALK2 R206H in complex with RK-71807 have been deposited in the Protein Data Bank with the accession code 6JUX.

Table 4

Crystallographic Data and Refinement Statistics

parameters	ALK2 (R206H) RK-71807
PDB ID	6JUX
data collection
beamline	X06DA, Swiss Light Source
wavelength (Å)	1.000
space group	C2221
unit cell parameters (Å, °)	a = 58.9, b = 86.8, c = 140.5; α = β = γ = 90.0
resolution (Å)a	48.7–1.73 (1.84–1.73)
reflectionsa	486,912 (75284)
unique reflectionsa	36,528 (5793)
redundancya	13.33 (13.00)
completeness (%)a	99.9 (99.6)
Rmergea,b	0.077 (1.69)
Rr.i.m.a,c	0.080 (1.763)
I/σ(I)a	20.64 (1.59)
refinement
resolution (Å)	48.7–1.75
reflections (total)	34,739
reflections (test)	1787
Rcryst	0.1940
Rfree	0.2260
no. of atoms
protein	2376
ligand/ion	42
water	182
average B-factors (protein/ligand) (Å2)	40.837
r.m.s. deviations
bond lengths (Å)	0.014
bond angles (°)	1.648

Values in parentheses are for the outer shell.

Rmerge = ΣΣ|I(hkl) – ⟨I(hkl)⟩|ΣΣ(hkl)

Rr.i.m. = Rmeas = Σ(N/(N – 1))1/2Σ|I(hkl) – ⟨I(hkl)⟩|ΣΣ(hkl)