Discovery of novel aminopiperidinyl amide CXCR4 modulators through virtual screening and rational drug design

Abstract

The C-X-C chemokine receptor type 4 (CXCR4) is a potential therapeutic target for HIV infection, metastatic cancer, and inflammatory autoimmune diseases. In this study, we screened the ZINC chemical database for novel CXCR4 modulators through a series of in silico guided processes. After evaluating the screened compounds for their binding affinities to CXCR4 and inhibitory activities against the chemoattractant CXCL12, we identified a hit compound (ZINC 72372983) showing 100 nM affinity and 69% chemotaxis inhibition at the same concentration (100 nM). To increase the potency of our hit compound, we explored the protein-ligand interactions at an atomic level using molecular dynamics simulation which enabled us to design and synthesize a novel compound (Z7R) with nanomolar affinity (IC₅₀ = 1.25 nM) with improved chemotaxis inhibition (78.5%). Z7R displays promising anti-inflammatory activity (50%) in a mouse edema model by blocking CXCR4-expressed leukocytes, being supported by our immunohistochemistry study.

Graphical Abstract

1. Introduction

C-X-C chemokine receptor type 4 (CXCR4) is a class A G-protein-coupled receptor (GPCR) that regulates a diverse array of intracellular signaling cascades associated with human diseases. GPCRs, the largest family of cell-surface receptors, represent the target of 30–40% of marketed drugs [1]. However, only a few chemokine GPCR-targeted drugs are currently available despite significant endeavors and investments so far [2]. Primarily, this is the result of the complex biology of chemokine systems [3], poorly predictive animal models [4], and a limited number of safe and efficacious small molecules [2] capable of competing with the chemokine proteins (8–10 KDa) which are larger than most other GPCR ligands such as hormones, neurotransmitters, ions, and odorants.

Among chemokine receptors, CXCR4 has rapidly emerged as an attractive therapeutic target because it is involved in the pathogenesis of various diseases including HIV-infection [5], rheumatoid arthritis [6], WHIM syndrome [7], inflammatory bowel disease [8], pulmonary fibrosis [9], lupus [10], and metastatic cancer [11–13] through the interaction with HIV glycoprotein 120 or endogenous chemokine ligand 12 (CXCL12) that is also known as stromal-derived-factor-1 (SDF-1). Accordingly, an extensive effort has been focused on the discovery and development of small molecular drugs targeting CXCR4 [14–17]. Such an effort was rewarded by the approval of AMD3100 (plerixafor, Mozobil™), the unique CXCR4 antagonist available in the current market, with the indication of hematopoietic stem cell mobilization. However, AMD3100 possesses a metal-chelating bicyclam moiety which could be related to cardiac disturbances reported in anti-HIV clinical trials [18–22]. Moreover, AMD3100 is also reported to potentially cause lung or liver fibrosis [23]. Therefore, even though the limited use of AMD3100 is widely accepted, long-term clinical usage may not be tolerated by patients due to its serious side effects. Several other CXCR4 modulators are under clinical investigation, but none of them have been approved due to their low efficacy and long-term toxicity [11, 12, 24].

The prior approaches to discover GPCR-targeted drugs have depended largely on ligand-based methods owing to the difficulty of obtaining the crystal structures of GPCRs, the basis of structure-based approaches [25]. In this context, the effort to discover new CXCR4 ligands was mainly focused on ligand-based approaches such as modification of known antagonists [26], deconstruction of known inverse agonistic peptides [27], and high-throughput screening (HTS) assays [28]. The recent high-resolution crystal structures of CXCR4 bearing small molecule inhibitors [29] have greatly increased our understanding of receptor inactivation in terms of the binding poses and the interactions between inhibitors and the critical residues of CXCR4 (Fig. 1). Hence, these crystal structures boosted structure-based virtual screening studies and led to the discovery of several active small-molecule CXCR4 modulators whose activities are attested by in vitro assays [30–32].

Figure 1. Small molecule biding sites and poses of CXCR4.

(A) GPCR chemokine receptors have multiple chemokine recognition sites (blue) such as CRS1, and CRS2 surrounded by extracellular loop2 (EL2) and transmembrane domains (TM2, TM3, and TM7). GPCRs themselves are allosteric proteins, illustrating G protein binding site (green) comprising TM3, TM5, TM6, and TM7; major and minor binding pockets (red); allosteric ligand binding site (orange). (B) Two small molecule binding sites of CXCR4 were elucidated by X-ray crystallography. The small molecule antagonist IT1t (green) binds to the minor binding pocket (TMS1) where W94^2.60, D97^2.63,Y116^3.32, R183 ^45.47, D187^45.51, and E288^7.39 are involved in critical protein-ligand interactions, suggesting that it competitively blocks the interactions of the CXCL12 N-terminus with CXCR4, while a cyclic hexadecameric peptide (CVX15, magenta) binds mainly in the major binding pocket (TMS2) encompassed by TM4, TM5, and TM6, where D171^4.60, D187^45.51, D188^45.52, Y190^45.54, D193 ^5.32, and D262^6.58 interact with the ligand by hydrogen bonding and salt bridges. (C) Top view of B. (D) Bottom view of B, which shows the inactive conformation of the receptor where TM3 is ousted from narrowed G protein binding site.

In this study, we report the successful application of our in silico strategy to discover and optimize novel CXCR4-targeted small molecules and demonstrate its potency in an in vivo model (Fig. 2A). First, we identified candidate small molecule CXCR4 modulators through virtual screening of the ZINC chemical database [33] and following two orthogonal in vitro assays. To improve the overall hit-rates, we used a series of sequential computational methods including ligand-based virtual screening (LBVS), structure-based virtual screening (SBVS), knowledge-based in silico ADMET (absorption, distribution, metabolism, excretion, and toxicity) prediction, and pan-assay interference compounds (PAINS) [34] removal. Furthermore, we improved the potency of the hit compound by molecular dynamics (MD) simulation and the corresponding structure modifications, which led us to discover a novel CXCR4 modulator (Z7R) demonstrating promising anti-inflammatory activity in a mouse model. Our approach may provide a valuable in silico guided methodology to discover and optimize novel small molecule drugs which target GPCR chemokine receptors.

Figure 2. Overall workflow of in silico guided drug discovery approach and setup of each computational method.

(A) The workflow describes the overall process of in silico approach which guides identification and optimization of hit compounds. To improve overall hit-rates, a series of sequential virtual screening (VS) methods are introduced, such as ligand-based VS (LBVS), structure-based VS (SBVS), knowledge-based ADMET prediction, and pan-assay interference compounds (PAINS) removal. Further, molecular dynamics (MD) simulation is employed to improve the efficacy of the hit, leading to the development of highly potent novel compounds. (B) Receiver Operating Characteristic (ROC) curves of 36 known CXCR4-modulators upon the test pool composed of active and decoy conformers, showing AMD3329 as the best query structure with the highest early recovery of the active compounds (AUC = 0.794). (C) Docking grid (magenta) of CXCR4 (gray, 3OE0) centered on the Arg2–Nal3 residues of CVX15. The small molecule in the central binding pocket indicates ZINC72372983 (green). (D) Molecular Dynamics (MD) simulation setup of CXCR4-ligand complex. The docked protein-ligand complex was solvated in an orthorhombic box using the simple point-charge water model. POPC membrane model was introduced onto transmembrane-associated residues 3OE0.

2. Materials and Methods

2.1. General

All chemical reagents and virtually screened compounds were obtained from commercial sources (Sigma Aldrich Co., St Louis MO, USA; Mcule Inc., Palo Alto, CA, USA; Princeton BioMolecular Research Inc., Princeton, NJ, USA; Ambinter c/o Greenpharma, Orléans, France) and used without further purification unless otherwise noted. NMR spectra were recorded on a Varian 400 MHz NMR spectrometer or Inova 400 MHz NMR spectrometer in the Emory NMR Center. Mass spectra for small molecules were obtained using an Agilent 1100 LC/MSD VL instrument in the Emory Mass Spectrometry Center. Fluorescence cell images were acquired by Leica SP8 Confocal microscope in the Emory’s Imaging Core. Thin Layer Chromatography (TLC) and Flash Column Chromatography (FCC) were carried out on silica gel 60 (Merck; 230–400 mesh ASTM).

2.2. Cell culture

Dulbecco’s modified Eagle’s medium (DMEM) with glutamine, Penicillin/Streptomycin and 0.5% Trypsin-EDTA were purchased from GIBCO (Grand Island, NY). Fetal bovine serum (FBS) was purchased from HyClone (Logan, UT). MDA-MB-231 (human breast adenocarcinoma) cells were cultured in DMEM supplemented with 10% FBS, 100 IU/mL penicillin, 100 μg/mL streptomycin.

2.3. Ligand-based screening

The 3-D conformers of “Drug-like Now” subset of ZINC database [33] (2013 release; 8,356,438 compounds) were generated by Omega2 (OpenEye Scientific Software, Inc., Santa Fe, NM) as described previously [35] and screened to identify potential small molecule modulators of CXCR4 using the shape overlay program ROCS (ver 3.0.0; OpenEye Scientific Software, Inc., Santa Fe, NM). The screening was performed with the following process. (i) Construct CXCR4-ligand set and decoy set; to avoid biased screening outcomes resulting from large differences in basic molecular properties (molecular weight, number of rotatable bond, number of hydrogen bond donor/acceptor, and polar surface area), a set of 36 CXCR4-ligands was compiled from the literature and decoy set from the DUDE database (http://dude.docking.org/targets/cxcr4). Then, 3-D structures of the selected CXCR4-ligands and decoys were generated by LigPrep (ver 2.4, Schrödinger, LLC, New York, NY) which led to 169 active and 5679 decoy conformations including different protonation states and tautomers at physiological pH (optimized by OPLS3 force field). (ii) Select the best query structure using vROCS (ver 3.0.0; OpenEye Scientific Software) as follows: each active structure was queried on the test set composed of active and decoy compounds. The compounds were ranked by combination of their Shape Tanimoto score and Color (functional group geometry) Tanimoto score, and Receiver Operating Characteristic (ROC) curves were plotted. The area under the ROC curve (AUC) was quantified to evaluate the performance of each query compound structure. (iii) Query the best structure on the ZINC database. Molecules whose TanimotoCombo (Shape + Color) similarity scores for the best structure were at least 0.72 were retained for molecular docking.

2.4. Structure-based screening using molecular docking

The default parameters in the Glide docking module of Schrödinger Suite (ver 9.3, Release 2016–3; Schrödinger, LLC, New York, NY) were used unless otherwise noted. The CXCR4 receptors were prepared using the Protein Preparation Wizard in the Schrödinger Suite that assigns bond orders, adds missing hydrogen atoms, and creates disulfide bonds. The hydrogen-bonding network was optimized at neutral pH 7.0. Receptor grids were constructed with Glide using a 10 × 10 × 10 ų boundary box spanning the entire ligand binding pocket and centered on the centroid of either IT1t (from PDB ID: 3ODU) or the Arg2–Nal3 residues of CVX15 (from PDB ID: 3OE0) [36]. The shape-based-screened ligand poses were docked into the CXCR4 receptor grids using Glide Standard Precision (Glide SP) and/or Extra Precision (Glide XP). The Epik protonation state penalties were added to the Glide scores. To estimate the binding affinity of designed ligands for CXCR4, the energy of the ligand-receptor complex (ΔG_Bind) was calculated using Prime MM-GBSA. Finally, we chose the structures which exhibited high docking scores (< −6.0 for SP docking; <−5.0 for XP docking), high binding energies (−50 kcal/mol from XP docking) and reasonable poses representing the interactions with at least two of the critical residues (Trp94, Asp 97, Asp171, Arg183, Asp187, Arg188, Tyr190, Asp 193, Asp262, and Glu288) [29] of CXCR4.

2.5. ADMET and PAINS evaluation

QikProp (ver 4.4, Schrödinger, LLC, New York, NY) was used to predict the ADMET properties of the ligands. This module was used to estimate pharmaceutically relevant physicochemical properties, including octanol/water partition coefficient (logP_o/w), aqueous solubility (logS), human oral absorption (OA), Caco-2 cell permeability (PCaco), binding to human serum albumin (logK_hsa), IC₅₀ value for blockage of HERG K⁺ channels (logHERG), and Lipinski’s rule of five (Ro5) [37]. The compounds obtained from shape-based screening were selected as input molecules for estimating the ADMET properties. The molecules showing values within the permissible range (−2.0 ≤ logP_o/w ≤ 6.5; −6.5 ≤ logS ≤ 0.5; 25% ≤ OA; 25 ≤ PCaco ≤ 500; −1.5 ≤ logK_hsa ≤ 1.5; Ro5≤1) were further analyzed with a PAINS [34] filter to screen out false positives using the module of CANVAS (ver 1.5, Schrödinger, LLC, New York, NY; PAINS filter 1, 2, and 3).

2.6. Molecular dynamics (MD) simulation

Desmond (ver 4.2, Schrödinger, LLC, New York, NY) was used to perform the molecular dynamics simulations of all the ligand-protein complexes. Using the system-builder module, the docked protein-ligand complex was solvated in an orthorhombic box with a volume of 708,475 ų containing 66581 ~ 66587 atoms using the simple point-charge (SPC) water model. The total systems were neutralized by adding Cl⁻ ions, and a salt concentration of 0.15 M was selected. A 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphocholine (POPC) membrane model was introduced onto transmembrane-associated residues (39–62, 76–98, 107–132, 151–173, 197–220, 241–264, 282–302) of 3OE0. The system was first relaxed using the Desmond relaxation model. The completed equilibration run was followed by a production run performed under normal temperature and pressure conditions (300K and 1.103 bar), using isothermal-isobaric (NPT) ensemble, and particle mesh Ewald (PME) electrostatics with a cutoff of 9 A. Time-step calculations were performed every 2 femtoseconds. The simulation job was carried out over a period of 20 ns. Root-mean-square deviation (RMSD) and root-mean-square fluctuation (RMSF) were calculated by the following formulae and analyzed by using event-analysis and simulation-interaction diagrams of Desmond.

The RMSD for frame x is:

$$RMS{D}_x=\sqrt{\frac{1}{N}\sum _{t=1}^N{\left({r^{\prime }}_i\left(t_x\right)-r_i\left(t_{rel}\right)\right)}^2}$$

where N is the number of atoms in the atom selection; t_ref is the reference time, (typically the first frame is used as the reference and it is regarded as time t=0); and r’ is the position of the selected atoms in frame x after superimposing on the reference frame, where frame x is recorded at time t_x. The procedure is repeated for every frame in the simulation trajectory.

The RMSF for residue i is:

$$RMS{F}_i=\sqrt{\frac{1}{T}\sum _{t=1}^T\langle {\left({r^{\prime }}_i(t)-r_i\left(t_{rel}\right)\right)}^2\rangle }$$

where T is the trajectory time over which the RMSF is calculated, t_ref is the reference time, r_i is the position of residue i; r’ is the position of atoms in residue i after superposition on the reference, and the angle brackets indicate that the average of the square distance is taken over the selection of atoms in the residue.

2.7. Chemical synthesis

Synthesis of compound 1

A 100 mL round bottom flask fitted with magnetic stirrer was charged with 10 mL of DMF. To the stirred solvent was added 6-azaindole (1.01 g, 8.53 mmole) followed by sodium hydride (400 mg, 10.2 mmole) and stirred at RT for 1 h. Then, methyl-3-bromo-propioate (2.13 g, 12.79 mmole) was added and stirred at RT for 24 h. After completion of reaction, the reaction mass was diluted with 30 mL of water and extracted with chloroform (3 × 100 mL). The combined organic layer was washed with brine, dried over Na₂SO₄, and then concentrated under reduced pressure. The crude product was purified by flash column chromatography, eluting with DCM/methanol (20:1, v/v). The purified compound was obtained as colorless oil. (958 mg, yield 55%). ¹H NMR (400MHz, CD3O): δ 8.75 (1H, m), 8.21 (1H, d), 7.47 (1H, d), 7.26 (1H, d), 6.44 (1H, d), 4.51 (2H, t), 3.62 (3H, s), 2.84 (3H, t) ppm.

The purified compound (methyl 3-(1H-pyrrolo[2,3-c]pyridin-1-yl)propanoate, 958 mg, 4.70 mmole) was dissolved in methanol (10 mL) and water (10 mL), then was added lithium hydroxide (LiOHH₂O, 1.0 g) and stirred overnight at RT. The reaction solvent was removed by rotary evaporator and dried reaction mixture was dissolved in DMF 20 mL, then was added t-butyl piperidin-4-ylcarbamate (939 mg, 4.70 mmole), N-(3-dimethylaminopropyl)-N′-ethyl carbodiimide hydrochloride (EDCI, 1.35 g, 7.04 mmole), 1-hydroxybenzotriazole (HOBt, 7.04 mmole), and triethylamine (1.4 mL). The reaction mixture was stirred overnight at RT. After the reaction was completed and diluted with water (5 mL), then extracted with ethyl acetate (3 × 100 mL), washed with water, brine, and dried over MgSO₄. The extract was concentrated, and crude mixture was purified by flash column chromatography, eluting with DCM/methanol (10:1, v/v). The purified compound was obtained as a white powder. (1.49 mg, yield 85%). ¹H NMR (400MHz, CD3O): δ 8.76 (1H, s), 8.18 (1H, d), 7.50 (1H, d), 7.35 (1H, d), 6.45 (1H, d), 4.58 (2H, m), 4.45 (2H, m), 3.54 (2H, d), 2.81 (4H, m), 2.65 (1H, m), 1.80 (2H, m), 1.39 (9H, s) ppm

Synthesis of compound 2 (Z7R)

Compound 1 (1.49 mg, 4.00 mmol) was deprotected by stirring in cocktail solution composed of TFA (13 eqv.) and DCM (32 eqv.) in the presence of triisopropylsilane (TIPS, 2.5 eqv.) at RT with the exclusion of moisture. The deprotection was monitored by TLC. After completion of deprotection, the reaction mixture was dried under reduced pressure. The dried mixture was re-dissolved in DMF, then was added 2.11 mg (1 eqv.) of N-alpha-Boc-N-omega-(2,2,4,6,7-pentamethyl-2,3-dihydrobenzo[b]furan-5-ylsulfonyl)-arginine), N-(3-dimethylaminopropyl)-N′-ethylcarbodiimide hydrochloride (EDCI, 1.5 eqv.), 1-hydroxybenzotriazole (HOBt, 1.5 eqv.), and triethylamine (4 eqv.). The reaction mixture was stirred overnight at RT. After the reaction was completed, diluted with water (5 mL) and extracted with ethyl acetate (3 × 100 mL), then washed with water, brine, and dried over MgSO₄.and concentrated. The extract was concentrated, and crude mixture was purified by flash column chromatography, eluting with DCM/methanol (10:1, v/v). The purified compound was obtained as yellow oil. (2.19 mg, yield 70%). The purified compound was deprotected by stirring in the cocktail solution (90% TFA, 5% TIPS, 5% water) and dried in vacuo. The crude mixture was dissolved in acetonitrile and precipitated with cold ether. The final compound was obtained as white crystalline powder. ¹H NMR (400MHz, D₂O): δ 9.03 (1H, s), 8.11 (1H,d), 8.02 (2H, m), 6.90 (1H, s), 4.68 (2H, t), 4.12 (1H, br), 3.99 (1H, t), 3.90 (1H, m), 3.78 (2H, m), 3.73~3.68 (2H, m), 3.21~3.18 (4H, m), 3.01~2.98 (3H, m), 2.73 (1H, m), 1.98~1.82 (6H, m), 1.76~1.67 (6H, m), 1.18 (2H, m). ¹³C NMR (400MHz, D₂O): δ 170.6, 168.5, 156.6, 141.6, 138.2, 131.3, 127.5, 125.5, 117.3, 103.6, 52.7, 52.6, 46.6, 43.1, 40.1, 32.8, 27.8, 26.9, 23.6, 23.3. HRMS (ESI-TOF) m/z: [M+H]⁺ Calcd. for C₂₁H₃₃N₈O₂ 429.26; Found 429.27.

2.8. In vitro binding assay

Binding affinity assay was performed as described previously [38]. Briefly, MDA-MB-231 cells cultured in an 8-well slide chamber were preincubated with the testing compounds at 1, 10, 100, and 1000 nM. Then the cells were fixed with 4% formaldehyde and incubated with 50 nM biotinylated-TN14003, and followed by streptavidin–rhodamine staining. The pictures of stained cells for each treatment were taken on a Nikon Eclipse E800 microscope. Pictures were analyzed quantitatively with ImageJ. IC50 value of the hit compound was fitted with GraphPad Prism.

2.9. In vitro Matrigel invasion assay

Matrigel invasion assay was performed as described previously [38]. Briefly, MDA-MB-231 cells were cultured on a layer of Matrigel in the upper chamber with testing compounds at 100 nM, while 200 ng/mL CXCL12 was added in the lower chamber as a chemoattractant. Invading cells at the bottom of the Matrigel were fixed in methanol and stained with H&E. The percentage of invading cells was determined by counting the H&E-stained cells.

2.10. In vitro serum stability test

Mouse blood was collected from the orbital venous sinus of nude mouse (Hsd:Athymic Nude-Foxn1<nu>, female, 7 weeks old) as described in the reference [39]. Collected blood was transferred into a 1.5 mL microcentrifuge tube and centrifuged at 13,000 rpm, 6 min, 4 °C. The supernatant was mixed with Z7R aqueous solution (10 mM) (solution : plasma = 1:4, v/v) and incubated at 36.6 °C. The mixture was collected at 0, 5, 15, 30, 60, and 120 min and transferred into a 1.5 mL microcentrifuge tube with an Microcon-10kDa filter and centrifuged at 13,000 rpm, 30 min, 4°C. The filtrate was analyzed by HPLC using a C18 column (250 × 4.6 mm) and eluted by acetonitrile (0.1%TEA) / water (0.1%TEA) (3:7, v/v). The peaks were detected at 310 nm and the areas were analyzed by ‘32 karat’ software. Half-life of Z7R in serum was calculated by Prism software using one phase decay model and nonlinear least squares fit.

2.11. In vivo anti-inflammation study using mouse paw edema model

Paw edema suppression test was conducted as described previously [40]. Concisely, acute inflammation was induced by subcutaneous injection of 50 μL of λ-carrageenan (1% w/v in saline) into the left hind paws of male nude mice (6 weeks, 20 g, Jackson Laboratory, Strain JAZ 007850 J:NU); the other hind paw was used as the non-inflammation control. The compounds were administered intraperitoneally (i.p.), 30 min after carrageenan challenge with the dose of 10 mg/kg, and then once a day following the initial dose. The mice were sacrificed 74 h after edema was induced and 2 h after the last injection of the compounds. The hind paws of the mice were photographed, the thicknesses of the paws were measured by a caliper, and the weights of the paws were measured by a balance after transecting at the level of the malleoli. To quantify the edema, the measured values of the untreated group were subtracted from the treated group. The inflammation suppression percentage was calculated by comparing the compound-treated groups to the control group (n =5). Fixed paw tissue was sliced and stained with H&E for histology study.

2.12. Immunohistology study

The avidin-biotin complex method was used to detect the proteins CXCR4 with anti-CXCR4 antibody (R&D System) at dilution 1:500. Formalin-fixed and paraffin-embedded tissues were deparaffinized and subsequently microwaved in 10 mM of sodium citrate buffer (pH 6.0) to retrieve antigens. After pre-incubation with hydrogen peroxide, avidin/biotin blocking kit (Invitrogen Co, CA), and horse serum (Vector Laboratories), the primary antibodies were applied overnight at 4 °C temperature. After incubation with the secondary antibody (rabbit anti-mouse biotinylated; Invitrogen Co, CA), the avidin-biotin complex was added and the enzyme activity was visualized with diaminobenzidine (ABC kit, Vector laboratories). Counterstaining was performed with hematoxylin.

3. Results

3.1. Virtual screening

To pre-screen the large chemical database efficiently and reduce the final number of candidate compounds, we performed LBVS prior to SBVS as presented in many cases [41] (Fig. 2A). We screened the subset of the ZINC chemical database (http://zinc.docking.org/) which contains ~ 8.4 million “Drug-like Now” compounds ready for 3-D virtual screening as described in the Methods section. For LBVS, we used the shape-based screen package ROCS [42]. First, we generated multiple queries based on the structures of known CXCR4 modulators and evaluated their ability to enrich for active over inactive compounds. The dataset consisted of both active and inactive compounds. The active compound set includes 169 conformers of the 36 known molecules which are representative of the groups of current CXCR4 modulators with different molecular scaffolds (Fig. S1). The inactive compounds set contains 5679 conformers of the 3406 decoy molecules from DUDE CXCR4-decoy library [43] to avoid potential bias of the screened molecules stemming from large differences in physicochemical properties between active and inactive compounds. Analyzing the receiver operating characteristic (ROC) curves of the queries generated from the known CXCR4-modulators using the TanimotoCombo score (the sum of the shape similarity and chemical functionality topology overlay) (Fig. 2B), we determined AMD3329 [44–46] as the query structure with the best early recovery of the active compounds (AUC = 0.794). Further, we validated our LBVS by Güner-Henry scoring method. We evaluated Enrichment Factor (EF = 26.3) and Goodness of hit (GH = 0.6) score using the following formulae:

$$EF=\frac{Ha/Ht}{A/D},GH=\left[\frac{Ha\times (3A+Ht)}{4HtA}\right]\times \left[1-\frac{Ht-Ha}{D-A}\right]$$

(where D = total # molecules; A = # active molecules; Ht = # hits retrieved; Ha = # active molecules in the hit list )

EF is the enrichment of the concentration of active compounds by the model relative to random screening. GH score ranges from 0, which indicates the null model, to 1, which indicates the ideal model. GH scores ranging from 0.6 to 1 is indicative of a good model and gave us confidence in the predictive capability of our model [47]. Next, we screened the ZINC database using our validated ROCS model and ranked them based on the TanimotoCombo score. The top 1002 molecules (TanimotoCombo score > 0.72) were advanced to the SBVS stage. We then performed a molecular docking simulation to the CXCR4 crystal structure (3OE0) using the standard precision (SP) scoring function of Glide [48] (Schrödinger) (Fig. 2C). For our docking calculations, we tested two crystal structures of CXCR4 (3ODU and 3OE0), and used 3OE0 because it showed the better EF and GH scores for our dataset comprising active and decoy compounds. 3ODU has higher resolution (2.5 Å) than 3OE0 (2.9 Å), but 3ODU is often unable to generate docking poses consistent with X-ray structures and interactions with critical residues of CXCR4 as reported in the previous literature [36]. From here we selected 199 potential compounds based on their docking scores (SP docking score < −6.0). In order to select compounds with drug-like properties, we sorted the compounds by calculating knowledge-based in silico ADMET properties. We used QikProp (Schrödinger) to predict and choose the compounds possessing favorable pharmacochemical properties as described in the Methods section. The ADMET calculations eliminated 28 compounds with poor pharmacochemical properties. We subjected these compounds to a PAINS (pan-assay interference compounds) [34] filter to eliminate compounds that have the possibility to interfere with the biochemical assays, using Canvas (Schrödinger) which removed further 32 compounds bearing PAINS moieties. The remaining 139 compounds were docked using the Extra Precision (XP) scoring function. Next, we calculated the binding free energy (ΔG_bind) for all compounds with a docking score less than −5.0 kcal/mol using MM-GBSA (molecular mechanics-generalized Born surface area) calculations with an implicit solvent model (Prime MMGBSA [49], Schrödinger). From which we selected compounds with a binding free energy of less than −50 kcal/mol because our prior CXCR4 model showed a high correlation with in vitro binding affinity and the compounds with high binding energy (−40 ~ −50 kcal/mol) presented submicro ~ nanomolar affinities [36, 50]. As a final stage of VS, we performed a visual inspection of the remaining poses and removed the molecules that did not show a reasonable orientation and conformation relative to the receptor. Finally, we selected compounds that showed molecular interactions such as hydrogen bonds, salt bridges, and hydrophobic interactions with at least two of the critical residues (Trp94, Asp 97, Asp171, Arg183, Asp187, Arg188, Tyr190, Asp 193, Asp262, and Glu288) reported by previous crystallography and mutational studies of CXCR4 [29]. Through our VS campaign, we finally identified 7 compounds from the ZINC chemical database as potential CXCR4 modulators (Table 1).

Table 1.

Properties of compounds

Name	Structure	Binding Affinity (EC*, nM)	Chemotaxis Inhibition** (%)
ZINC72372983		100	68.6 ± 7.3
ZINC57773744		100	15.7 ± 5.5
ZINC08694519		100	40.9 ± 9.5
ZINC09388085		10	9.3 ± 7.6
ZINC48375389		1000	n/a
ZINC62573917		>1000	n/a
ZINC13688671		1000	n/a
Z7R		1	78.5 ± 7.8
AMD3100		1000	55.0 ± 8.7

^*In this paper, EC (effective concentration) is defined as the concentration at which the compound blocks > 50% of CXCR4-targeted peptidomimetic conjugate (rhodamine Red X-TN14003).

^**The inhibition % against the chemoattract, CXCL12 was determined by Matrigel invasion assay using 100 nM of compounds.

3.2. In vitro assay

We tested the 7 compounds by a series of mutually orthogonal in vitro assays. The first binding assay quantifies the affinity of compounds to the receptor by measuring the fluorescence of CXCR4-targeted peptidomimetic conjugate (rhodamine Red X-TN14003) after treating CXCR4-positive with each compound, which presents a lower fluorescent intensity for the compound with higher affinity (Fig. S2). Among the 7 virtually identified compounds, 4 compounds showed a binding affinity (effective concentration: the concentration at which the compound blocks more than 50% of rhodamine Red X-TN14003, EC) less than 100 nM in our assay (Table 1). As a secondary orthogonal functional assay, we performed Matrigel invasion assay which measures the inhibitory efficacies of compounds against the chemotaxis between CXCR4-positive cells and the chemokine, CXCL12 (Fig. S2). Through these consecutive assays, we specified ZINC72372983 as our hit compound which showed 100 nM affinity and 69% chemotaxis inhibition (Table 1). Of note, we used MDA-MB-231cells which express high levels of CXCR4, but insignificant level of CXCR7 [51] to select only CXCR4-specific compounds.

3.3. Chemical improvement of the hit compound using MD simulation and SAR study

To improve the potency of our hit compound, we explored the interactions between our hit compound and the receptors using molecular dynamics (MD) simulations to provide insight into the design of novel compounds with more stable molecular interactions within the protein-ligand complexes. We performed a 20 ns MD simulation of the CXCR4-ZINC72372983 docked complex to determine if a stable ligand receptor complex is formed. The initial docking results showed H-bond interactions between ZINC72372983 and receptor binding site residues Thr117, Asp171, Arg188, Gln200, and Asp262. We monitored protein-ligand interactions in each trajectory frame throughout the simulation to check which interaction is maintained. At the end of the simulations, we found the following interactions occur more than 30% of the simulation time in the whole trajectory (0.00 through 20.02 ns), meaning that in more than 30% frames out of the entire trajectory frames, the interactions (H-bond, hydrophobic, ionic or water bridge) with specific residue are maintained; 92% with Asp171 (H-bond or water bridge interaction), 107% with Arg188 (multiple H-bond interactions with atoms of arginine side chain or water bridges), 57% with Gln200 (water bridge), 97% and 76% with Asp262 (H-bond, ionic or water bridge). Hydrophobic interactions with nonpolar residues (His113, Try116, Leu120, His203, Ile259, His 281, Ile284) were not considered due to their low maintenance (<30%) during simulation.” (Fig. 3B). The root mean square deviations (RMSDs) of the protein backbone (Cα) and the ligand (Lig fit Prot) demonstrate that our simulation has equilibrated and converged. The RMSD of ZINC72372983 in the binding pocket during the 20 ns simulation showed deviations between 1.2 and 3.9 Å. The 3OE0 altered its RMSD from 1.5 Å in the starting frame to a maximum of 4.2 Å RMSD. Higher fluctuations were observed between 12 and 17 ns, followed by a stable RMSD up to the final 20 ns, implying that the protein fluctuations towards the end of simulation are around some thermal average structures. The RMSD values of the protein and the ligand stabilized around 2.5 Å (Fig. 3C). To investigate the fluctuation of the ligand within the binding pocket, we analyzed the root mean square fluctuation (RMSF) of the ligand (Fig. 3D) which showed a higher RMSF of the phenyl group when compared to other parts of the hit compound, suggesting that the phenyl group is the weakest receptor-bound moiety of the ligand throughout the simulation. While RMSDs of receptor and ligand are stable during the simulation period, the phenyl moiety of the hit compound most fluctuates around the binding site (Supplementary 2).

Figure 3. MD simulation of CXCR4 and lead compound (ZINC72372983) complex.

(A) The pose of ZINC72372983 in the binding pocket of 3OE0 (docking pose). (B) Proteinligand (CXCR4-ZINC72372983) interactions during MD simulation. (C) MD simulation trajectory depicting RMSDs of protein Ca atoms (blue) and ligand heavy atoms (red). (D) RMSF of the ligand (ZINC72372983) broken down by atom, corresponding to the 2D structure in the top panel. In the bottom panel, the “Fit Ligand on Protein” line shows the ligand fluctuations, with respect to the protein. Green color indicates the five highest fluctuated atoms during MD simulation; oxygen (red); nitrogen (blue).

To increase the affinity of the lead to CXCR4, we attempt to replace the phenyl moiety with other functional group which enables the hit compound to form additional interactions between the ligand and critical residues of the receptor while it retains existing binding interactions of the hit compound. First, we calculated the distance (6 ~ 12 Å; Supplementary 3) between the phenyl group of the hit compound and the receptor’s crucial amino acid residues (Asp97, Asp187, or Glu288) which did not show interactions with the hit compounds.

Next, we sought to replace the phenyl group with the moieties that have the corresponding sizes (5 ~ 11 covalent bond length) and functionalities capable of interacting with the crucial amino acids (Asp97, Asp187, or Glu288) of the receptor. To this end, we modified structure of our hit compound by replacing the phenyl group with 14 different amino acid side chains (Fig. 4B, Table 2), considering the amido scaffold of the hit compound and the structures of known peptidic CXCR4-antagonists. Subsequently, we ran MD simulations for those modified compounds to check whether those compounds have the desired additional interactions while retaining existing ligand-receptor interactions. Among 14 conceptualized compounds, 3 structures showed additional interactions with the Glu288 more than 30% trajectory frame during the entire MD simulation while maintaining existing interactions (Fig. 4A). Z7R showed the best additional interaction (89%) with Glu288 whereas Z7K displayed less maintained interaction (39%) with Glu288 at the expense of losing the frames of interaction with Asp262. From this we selected compound Z7R as our top candidate to be synthesized (Fig. 4B and 4C). In vitro test showed compound Z7R to have a 100-fold improved binding affinity (IC₅₀ = 1.25 nM) as well as an enhanced chemotaxis inhibition (78.5 %) for the in vitro assays (Table 1, Fig. S2 and S3).

Figure 4. Design and synthesis of a novel compound.

(A) MDS studies of designed derivatives, upper graphs show RMSDs of conceptualized compounds (Z7R, Z7W, and Z7K)-CXCR4 complexes. All protein frames are first aligned on the reference frame backbone, and then the RMSD is calculated based on the position of “Cα” atom. “Lig fit Prot” indicates the RMSD of a ligand when the protein-ligand complex is first aligned on the protein backbone of the reference and then the RMSD of the ligand heavy atoms is measured. “Lig fit Lig” indicates the RMSD of a ligand that is aligned and measured just on its reference conformation. (B) Retrosynthesis of the compound Z7R (C) Synthetic scheme of Z7R; (a) NaH, DMF, 1 h @RT; (b) 24 h @RT, 55%; (c) LiOH, MeOH:H₂O = 1:1, 12 h @RT; (d) HOBt, EDCI, TEA, DMF, 12 @RT, 85%; (e) TFA:TIPS = 95:5, DCM; (f) HOBt, EDCI, TEA, DMF, 12 @RT, 70%; (g) TFA:TIPS:H2O = 90:5:5.

Table 2.

Docking and MD study results of Z7X^† compounds

Molecule	MMGBSA ΔGbind‡ (kcal/mol)	Interactions※ with CXCR4
ZINC72372983	−54.84	A171, R188, Q200, D262
Z7R	−71.67	Y116, D171, R188, H203, D262, E288
Z7W	−65.57	Y116, D171, R188, D262, H281, S285
Z7K	−74.75	Y116, T117, D171, R188, E288
Z7C	−50.16	R188, Y45, E288
Z7Y	−54.79	Y94, R188, D262, H281, S285
Z7T	−57.58	D97, D187, R188
Z7S	−61.69	D171, R188
Z7H	−58.67	Y116, R188, D262, E288
Z7D	−31.86	Y94, H113, Y121, D171, R188, H203
Z7E	−40.43	H113, Y121, D171, R188, D262, Q202
Z7Q	−63.75	W94, D97, H113, C186, D262, E288
Z7M	−68.66	D171, R188, Q200, H203, D262
Z7N	−68.43	Y116, D171, R188, Q200, D262
Z7P	−59.64	W94, D262, E288

^†Z7X: X stands for side chain of amino acid X

^‡MMGBSA ΔG_bind: calculated binding free energy from XP docking pose using molecular mechanics-generalized Born surface area method

^※Interactions (H-bonding, hydrophobic, ionic, or water bridge) between the ligand and CXCR4 protein residues which maintain more than 30% frames out of entire trajectory frames throughout MD simulation

3.4. In vitro serum stability of Z7R

Z7R has a terminal arginine moiety which may affect in vivo function of the compound by exopeptidases. Therefore, we assessed the serum stability of Z7R using fresh mouse serum. The serum-incubated Z7R samples were collected at the specified time points and the amount of Z7R was analyzed by HPLC immediately. The calculated half-life is 77.1 min (Fig. 5) which suggests the acceptable in vivo stability of Z7R.

Figure 5. In vitro serum stability of Z7R.

Z7R aqueous solution was mixed with mouse serum and incubated at 36.6 °C. The mixtures are analyzed by HPLC to estimate the amount of Z7R at 0, 5, 15, 30, 60, and 120 min during incubation. Half-life of Z7R is calculated by Prism software using one phase decay model and nonlinear least squares fit.

3.5. Anti-inflammatory activity of Z7R on mouse paw edema

To validate the physiological effects of compound Z7R which showed promising results in in vitro assays, we conducted carrageenan-induced paw edema (CIPE) assays using our mouse model. Z7R-treated mice (n =5) showed significant reductions of the edema weights (~ 50%) in the lesion, compared with untreated mice (Fig. 6).

Figure 6. Anti-inflammatory activity of Z7R on mouse paw edema.

(A) Z7R-treated mice show the reduction of paw weight (~ 50%) in the carrageenan-induced paw edema lesion, compared with the untreated mice. Acute paw inflammation was induced by subcutaneous injection of 50 μL of λ-carrageenan in left hind paw. The mice in the treatment group were all administered Z7R at 10 mg/kg i.p., while control animals received corresponding i.p. injection of vehicle. Error bars indicate standard errors, **: p < 0.001, n = 5. Comparison of untreated control (B) vs. Z7R-treated (C) mouse paw edema. Anti-CXCR4 immunochemistry studies (upper right) indicate CXCR4-positive leukocytes (brown).

Moreover, Z7R significantly attenuated the mouse paw inflammation in our histological assay. As shown in Fig. 6B and 6C, there was an observable decrease in the number of inflammatory cells. Our immunohistochemistry study revealed a large number of CXCR4-expressing infiltrating cells in the carrageenan-induced inflammatory tissue, compared with a reduced presence of CXCR4s which are mainly expressed in the epidermis of Z7R-treated mouse tissue. This immunohistochemistry study suggests that Z7R can suppress the inflammation by inhibiting chemotaxis of CXCR4-expressing leukocytes.

4. Discussion

In this study, we report a novel CXCR4 modulator (Z7R) that was discovered and optimized by a series of in silico guided methods, and proven to be potent in our in vitro and in vivo assays. Z7R showed promising anti-inflammatory activity in our mouse edema model.

We performed SBVS first to increase the performance of database screening, then followed SBVS to enhance the odds of success as presented previously [41]. From our LB VS we selected 199 compounds from ~ 8.4 million Drug-like Now compounds (ZINC). We selected the top 0.008% (TanimotoCombo score > 0.72) of compounds from the database to save time and cost, but we may increase the number of candidate compounds by adjusting the cutoff value of similarity index (TanimotoCombo score) with the conditions of reasonable EF and GH scores [47].

The crystal structures of CXCR4 bearing a small molecule antagonist (IT1t, PDB ID: 3ODU), a peptidomimetic ligand (CVX15, PDB ID: 3OE0), and a viral CC chemokine (vMIP-II, PDB ID: 4RWS) were solved in the past decade [29, 52], which opened an avenue to the structure-based drug discovery of CXCR4 modulators. Using the CXCR4-crystal structure 3ODU, a few virtual high-throughput screenings (vHTS) were attempted to identify new CXCR4 antagonists or agonists, and several in vitro active hits were reported [30–32, 53].

Like other chemokine receptors, CXCR4 has a major and a minor binding pocket where a small molecule can bind and modulate receptor functions as depicted in Fig. 1. IT1t is a competitive antagonist which binds in the minor pocket (TMS1) surrounded by extracellular loop2 (EL2) and transmembrane domains (TM2, TM3, and TM7) where W94^2.60, D97^2.63,Y116^3.32, R183 ^45.47, D187^45.51, and E288^7.39 are involved in critical protein-ligand interactions, competitively blocking the interactions of the N-terminus of CXCL12 with CXCR4. Unlike IT1t, CVX15 binds mainly in the major binding pocket (TMS2) encompassed by TM4, TM5, and TM6, where D171^4.60, D187^45.51, D18 8^45.52, Y190^45.54, D193 ^5.32, and D262^6.58 interact with the ligand. Previous studies revealed that the binding site of AMD3100, an FDA-approved CXCR4 antagonist, is mainly lined by three acidic TM residues, D171 (TM4), D262 (TM6), and E288 (TM7) [54–57]. Although the binding site of AMD3100 and IT1t do not overlap completely, AMD3100 binds partly to TMS1 where the N-terminus of CXCL12 interacts with the receptor, leading to the antagonism of functional responses [58]. In addition, prior studies of known antagonists such as AMD11070 and GSK812397 suggested the existence of allosteric binding sites on CXCR4 [59]. Considering those findings, we conducted our molecular docking simulation. In principle, docking scores are dependent upon the poses of ligands with respect to one receptor conformation among numerous variants. This protein inflexibility is a main challenge of docking simulation. Further, in many CXCR4-ligand modeling endeavors, several basic moieties of ligands and multiple crucial acidic residues of the chemokine receptor make it more difficult to evaluate the putative binding modes of novel compounds without additional experimental data. MD simulation can be useful to determine the most stable pose of complexes with flexibility, but at the stage of database screening, such a computationally expensive study is not applicable due to the huge size of chemical libraries. In this study, we prioritized compounds presenting the interactions with crucial residues of CXCR4 in our SBVS. To wit, we selected compounds that exhibited the interactions with at least two of the critical residues (Trp94, Asp 97, Asp171, Arg183, Asp187, Arg188, Tyr190, Asp 193, Asp262, and Glu288) recognized by crystallography and mutation studies of CXCR4 [29].

To test the activities of the identified compounds, we performed in vitro binding assays first using a fluorescent conjugate (rhodamine Red X-TN14003) known to bind across a major binding pocket TMS2 and a minor pocket TMS1 of CXCR4 (Fig. 1). Thus, our fluorescent probe can screen diverse compounds binding to different sites of the receptors. We have already reported the performance of this assay to distinguish bound from unbound ligands and specific from non-specific binding using a known antagonist (AMD3100) [51]. The signal-to-background ratio was greater than 10 and the Z’ factor was greater than 0.5. Numerous chemokine receptors are known to have different binding sites which can be associated with the different signaling pathways and functions. Hence, the orthogonal functional assays are critical paths of screening chemokine receptor-targeted ligands not only to remove false positives, but also to find potential selective modulators. CXCL12-induced chemotaxis is one of fundamental functions of CXCR4 and plays a crucial role in developing inflammation by attracting leukocytes from the blood into the tissues in response to stimuli. We tested the chemotaxis inhibition of the compounds with high binding affinity (EC ≤ 100 nM: effective concentration at which the compound blocks more than 50% of the targeted probe) and selected the compound showing chemotaxis inhibition more than 50% at 100 nM, which led us to specify our hit compound (ZINC72372983).

Our MD study started from the visually inspected best docking pose of the complexes. The complex bearing the hit compound displayed a stable MD trajectory of RMSD over simulation time (20 ns, Fig. 3C). This indicated stable and specific interactions with the target, regarding the starting conformation (Fig. 3A, 3B). Bad docking poses tend to generate an unstable MD trajectory, during which the ligand could even dissociate from the binding site. “Lig fit Prot” shows the RMSD of a ligand when the protein-ligand complex is first aligned on the protein backbone of the reference and then the RMSD of the ligand heavy atoms is measured, while “Lig fit Lig” shows the RMSD of a ligand that is aligned and measured just on its reference conformation. “Lig fit Prot” values of the hit compound are observed mostly lower than the RMSD of the protein, suggesting the ligand has bonded well and not diffused away from its initial binding site (Fig. 3C, 4A). Here, we sought to increase the affinity of the hit compound by analyzing the ligand RMSF which shows the ligand’s fluctuations broken down by atom. The ligand RMSF can provide insight on how ligand fragments interact with the protein and their entropic role in the binding event. The “Fit Ligand on Protein” of the hit compound indicated large fluctuations in position of the phenyl group (Fig. 3D, Supplementary 2) which gave us clues about modifications of the ligand for better affinity. We analyzed the trajectory focused on the distance between the phenyl group and the acidic residues (D97, D187, or E288) where we desired to confer additional strong interactions reported in known inhibitors (Supplementary 3). Then, we replaced the phenyl group with moieties having the corresponding length and functionalities (5 ~ 11 covalent bond length and basic functionalities), which led to a more potent novel compound (Z7R) as described in detail in the Results section.

It is well-known that CXCR4 plays a key role in the recruitment of leukocytes to sites of inflammation. Thus, blocking CXCR4 could be a therapeutic strategy in inflammatory diseases. To investigate the anti-inflammatory activity of our novel compound (Z7R), we adopted the carrageenan-induced paw edema model which has long been used to assess anti-inflammatory activity through the suppression of prostaglandin production and other inflammatory mediators. An apparent edema response was observed after injecting Z-carrageenan, and Z7R showed a significant suppression effect (~50%) on edema weight of the lesion, confirming that Z7R has anti-inflammatory activity as anticipated. However, inflammation is a complex physiological condition where a number of different chemokines regulate leukocyte trafficking, which presents a serious hurdle to developing highly potent anti-inflammatory agents. To address this pharmacological complexity, unconventional broad-spectrum chemokine inhibitors were reported [60]. Given this finding, the inclusion of a broad-spectrum chemokine moiety (3-acyaminocaprolactam) under in silico guidance may further increase the anti-inflammatory potency of Z7R.

In summary, we present here the successful application of our in silico guided methodology to discover novel CXCR4 modulators. Through our methodology, we screened ZINC chemical database and identified new CXCR4 ligands with submicromolar affinity and chemotaxis inhibition activity at an excellent hit rate (57%). The best hit compound (ZINC72372983) showed 100 nM affinity and 69% inhibitory activity against the chemoattractant CXCL12. To improve the efficacy of our hit compound, we investigated molecular interactions between the hit compound and CXCR4 using MD simulation, a part of our methodology, which allow us to redesign the structure of the hit compound and to obtain a novel compound (Z7R) with 1 nM binding affinity and 79% chemotaxis inhibition activity. Z7R displayed acceptable stability (t_1/2 = 77.1 min) in mouse serum and 50% anti-inflammatory activity in a mouse edema model. Our in silico guided approach may provide insights into the development of novel small molecule therapeutics targeting chemokine receptors.

Highlights.

• New CXCR4 ligands were identified by in silico screenings and in vitro assays

• A novel compound was rationally designed from the hit compound using MD simulation

• Z7R displayed nanomolar binding affinity (1 nM) and inhibited 79% of chemotaxis

• Z7R demonstrated 50% anti-inflammatory activity in a mouse edema model

Abbreviations

GPCRs

G-protein-coupled receptors

CXCR4

C-X-C chemokine receptor type 4

CXCL12

C-X-C chemokine ligand 12

SDF-1

stromal cell-derived factor-1

LBVS

ligand-based virtual screening

SBVS

structure-based virtual screening

MD

molecular dynamics

HIV

human immunodeficiency virus

PDB

protein databank

EC

effective concentration