Structural insights of the conserved “priming loop” of hepatitis B virus pre-genomic RNA

Abstract

Initiation of protein-primed (−) strand DNA synthesis in hepatitis B virus (HBV) requires interaction of the viral polymerase with a cis-acting regulatory signal, designated epsilon ($\text{ε}$), located at the 5′-end of its pre-genomic RNA (pgRNA). Binding of polymerase to $\text{ε}$ is also necessary for pgRNA encapsidation. While the mechanistic basis of this interaction remains elusive, mutagenesis studies suggest its internal 6-nt “priming loop” provides an important structural contribution. $\text{ε}$ might therefore be considered a promising target for small molecule interventions to complement current nucleoside-analog based anti-HBV therapies. An ideal prerequisite to any RNA-directed small molecule strategy would be a detailed structural description of this important element. Herein, we present a solution NMR structure for HBV$\text{ε}$which, in combination with molecular dynamics and docking simulations, reports on a flexible ligand “pocket”, reminiscent of those observed in proteins. We also demonstrate the binding of the selective estrogen receptor modulators (SERMs) Raloxifene, Bazedoxifene, and a de novo derivative to the priming loop.

1. Introduction

Globally, ~270 million people are chronically infected with hepatitis B virus (HBV) (Stanaway et al., 2016), leading to severe liver diseases, e.g., liver cirrhosis (scarring) and hepatocellular carcinoma, that account for more than 600,000 deaths annually (Stanaway et al., 2016). HBV, a member of the Hepadnaviridae family, is the smallest animal infecting DNA virus, with a ~3.2 Kb genome (Galibert et al., 1979) encoding seven proteins contained in four overlapped open reading frames (Galibert et al., 1979). Of these proteins, HBV polymerase (P) contains four domains: the terminal protein (TP), a spacer, the reverse transcriptase (RT), and a ribonuclease H (RNaseH) domain (Jones & Hu, 2013). The HBV life cycle starts with binding of the virus to the sodium-taurocholate cotransporting polypeptide receptor of the host liver cell (Yan et al., 2012). Following infection, partially double-stranded viral relaxed circular (rc) DNA is repaired in the nucleus, forming covalently closed circular DNA (cccDNA), which provides the transcriptional template for the host RNA polymerase II (Rall et al., 1983). Viral transcripts are transported to the cytoplasm, where pre-genomic RNA (pgRNA) provides the mRNA translated into the nucleocapsid proteins (C) and P, and the template for reverse transcription of (−) strand DNA and primer for (+) strand DNA. Following translation from pgRNA in the cytosol (Summers & Mason, 1982), P specifically binds to the ~61-nucleotide (nt) cis-acting RNA regulatory signal, epsilon ($\text{ε}$) (located at the 5′-end of pgRNA) (Hirsch et al., 1990; Knaus & Nassal, 1993), and uses the first two nucleotides in the internal bulge to initiate synthesis of (−) strand DNA.

Despite a clear demonstration that $\text{ε}$ is central to HBV replication, high-resolution structures of this critical regulatory element have not been forthcoming. Until now, the only $\text{ε}$ structure available is a truncated 27-nt fragment derived from its apical loop (i.e., residues belonging to the upper helix and the pseudo-triloop) (Flodell et al., 2002, 2006). Lack of structural information for full-length $\text{ε}$ presents a major barrier to drug development by preventing structure-enabled design of small molecule antagonists. Indeed, small molecules offer an opportunity to target RNA motifs—such as pseudoknots, bulges, and hairpins—which are often highly conserved and mediate important biological functions (Disney et al., 2014; Shortridge & Varani, 2015). Several recent in vitro and in silico high-throughput screening (HTS) approaches have identified chemotypes that selectively bind RNA motifs, with corresponding physiological effects in cell culture and animal models (Naryshkin et al., 2014; Stelzer et al., 2011).

To address this knowledge gap, we have determined for the first time the full-length structure of HBV $\text{ε}$ RNA using solution nuclear magnetic resonance (NMR) spectroscopy, complemented with molecular dynamics (MD) simulations. These approaches suggest that nucleotides of the “priming loop” constitute a binding pocket that might be targeted via HTS strategies. Our preliminary data indicate the selective estrogen receptor modulators (SERMs) Raloxifene and Bazedoxifene (Valentovic, 2007) are $\text{ε}$-binding ligands. NMR titrations and MD and docking simulations provided further insight on ligand binding to nucleotides of the bulged priming loop, corroborating this structural feature as a binding pocket. Taken together, our studies provide an important starting point for additional high-resolution and computational strategies for identifying ligands to target early steps in HBV DNA synthesis.

2. Materials and methods

2.1. NMR resonance assignment of HBV $\text{ε}$

The 61-nt HBV$\text{ε}$was transcribed in vitro using the p2RZ HDV ribozyme-containing plasmid (Walker et al., 2003). A combination of selectively ¹³C/¹⁵N-labeled nucleotides was utilized for NMR resonance assignment and structure studies (Leblanc et al., 2017). Combining asymmetric isotopic labeling and NOESY filter/edit techniques allowed complete assignment of the C8–H8, C6–H6, C2–H2, C1′–H1′, C2′–H2′, N1–H1 (guanosine), and N3–H3 (uridine) resonances. A through-bond HNN-COSY experiment (Dallmann et al., 2013; Dingley & Grzesiek, 1998) confirmed hydrogen-bonding restraints for the upper and lower helical regions of $\text{ε}$. Assignment of base-pairs in helical regions was confirmed with ¹H-¹H NOESY of the imino protons with RNA-PAIRS (Bahrami et al., 2012). Non-exchangeable protons (H1′, H2′, H2, H6, and H8) were assigned with a combination of through-bond HCCH-COSY-TOCSY (Hu et al., 1998) and TROSY HCN experiments (Brutscher & Simorre, 2001) and 3D ¹³C-edited and ¹³C-filtered ¹H-¹H NOESY experiments (Leblanc et al., 2017; Peterson et al., 2004). Starting points for non-exchangeable assignments were identified with a ¹³C site-selectively labeled sample (¹³C-2′,8-Ade; ¹³C-1′,6-Cyt/Uri; ¹³C-1′,8-Gua) that gave unambiguous contacts between unlabeled Ade-H2 (¹²C-¹H) and labeled ribose-H1′ (¹³C-¹H) resonances, as previously reported (Leblanc et al., 2017) when applying a 3D ¹³C-filter/edit NOESY experiment. Further ambiguity was resolved using modular HBV $\text{ε}$ constructs (priming and apical loop constructs, supplementary material, Figures 1 and 2) and ¹³C site-selectively labeled full-length $\text{ε}$. NMR experiments were performed in buffer A (10 mM sodium phosphate, 0.1 mM EDTA pH 6.7) on either a 600 MHz triple resonance room temperature probe or a Bruker 800 MHz triple resonance cryoprobe. Raw data were processed using Topspin (Bruker) and NMRPipe (Delaglio et al., 1995) and analyzed using NMRFx and NMRViewJ (Johnson, 2018). The NMR chemical shifts have been deposited in the Biological Magnetic Resonance Data Bank under accession number 50136.

2.2. Solvent paramagnetic relaxation enhancement (sPRE)

To assess the solvent accessibility in HBV $\text{ε}$, ¹H-¹³C HSQC spectra were collected in the absence and presence of MRI contrast dye gadolinium-diethylenetriamine pentaacetic acid-bismethylamide (Gd-DTPA-BMA) from (OmniScan—GE Healthcare) (Hartlmüller al., 2017). All measurements were collected with uniformly ¹³C/¹⁵N-labeled 500 µM HBV $\text{ε}$ RNA in buffer A on a Bruker 800 MHz triple resonance cryoprobe. Using a saturation-recovery approach, proton longitudinal (R₁) relaxation rates were measured as a function of increasing dye concentration (0.8 to 4.2 mM). The slope of the best fit linear relationship between dye concentration and proton R₁ gave the sPRE value for each measured proton. Errors in sPRE measurements were determined by propagating the standard deviations of repeated delay points in R₁ experiments.

2.3. Small-angle X-ray scattering (SAXS) measurements

The 61-nt HBV $\text{ε}$ RNA was prepared as described for NMR samples. The main difference in SAXS samples was that, due to strong scattering of phosphates, the traditional sodium phosphate buffer for NMR experiments could not be used. However, 3-morpholinopropane-1-sulfonic acid (MOPS) is considered a good SAXS buffer. To ensure that its use would not cause major structural perturbations in the RNA, we first made a ¹³C site-selectively (¹³C-1′,2′,8-Ade) labeled NMR sample in a deuterated MOPS buffer. The adenine nucleobase nuclei are extremely sensitive to minor changes in pH and therefore the spectrum was directly compared to that of the same sample in sodium phosphate buffer to determine whether bulk changes were observed. The relatively small perturbations observed between the two indicated that MOPS buffer could be used without complications in data analysis (supplementary material, Figure S3A). RNA samples of increasing concentrations (1.0, 2.5 and 5.0 mg/mL) were prepared in buffer B (10 mM MOPS, 0.1 mM EDTA pH 6.7) and SAXS data were collected at the Advanced Photon Source (APS) at the Argonne National Laboratory. Raw data were reduced, and multiple sets were averaged and subtracted from curves of buffer alone. The three concentrations were merged and extrapolated back to q = 0 to use for subsequent analysis. The Kratky plot indicated a well-folded RNA and real-space pair distance distribution function (PDDF) analysis gave an R_G of 29.8 Å and a D_max of 97 Å (supplementary material, Figure 3B and C).

2.4. Structure modeling of HBV $\text{ε}$

After the chemical shifts of HBV $\text{ε}$ NMR resonances were assigned using a combined asymmetric ¹³C-labeling and isotopic filter/edit NOESY strategy (Leblanc et al., 2017), the NOE distance restraints, dihedral angle restraints, sPRE (Hartlmüller et al., 2017), and RDC restraints were combined for use in structure calculations in Xplor-NIH (version 2.48) (Schwieters et al., 2018). Additional restraint tables were taken from SAXS/WAXS data. Python scripts to fold and refine the RNA structure were adapted from the Xplor-NIH website (Schwieters et al., 2018). Using previously reported protocols (Bermejo et al., 2016; Hartlmüller et al., 2017; Schwieters et al., 2018), the extended RNA sequence was subjected to high temperature molecular dynamics (MD) in torsion angle space, simulated annealing, gradient minimization in torsion angle space, and gradient minimization in Cartesian space. These stages folded the RNA according to the restraint files collected from the established torsion database, NMR distances, H-bond distances, and SAXS/WAXS data. Two hundred structures and violation files were generated. The combination of restraints yielding the fewest violations was selected for refinement, and the top 10 lowest energy structures from refinement were checked by MolProbity (Chen et al., 2010) and the atomic coordinates for the reported solution structure ensemble have been deposited with the Protein Data bank under accession number 6VAR.

2.5. NMR cross-correlated relaxation (${\text{η}}_{\text{xy}}$) experiments

All NMR cross-correlated relaxation experiments were performed on a 600 MHz triple resonance room temperature probe. Hahn-Echo TROSY-detected measurements of the fast (anti-TROSY, $\text{R2α}$) and slowly (TROSY, ${\text{R}}_{\text{2β}}$) relaxing TROSY components were adapted from previous pulse sequences (Lakomek et al., 2013). To avoid resonance overlap and signal decay, ¹³C site-selectively (¹³C-2′,8-Ade; ¹³C-1′,6-Cyt/Uri; ¹³C-1′,8-Gua) labeled RNA samples were used. ${\text{R}}_{\text{2α}}$ and ${\text{R}}_{\text{2β}}$ experiments were acquired in an interleaved manner as a pseudo three-dimensional experiment. The cross-correlated relaxation (${\eta }_{\text{xy}}$) rate was obtained by the following relation: ${\eta }_{\text{xy}}$ = (${\text{R}}_{\text{2α}}$ – ${\text{R}}_{\text{2β}}$)/2. The ${\text{R}}_{\text{2α}}$ and ${\text{R}}_{\text{2β}}$ rates were determined by fitting peak intensities to a monoexponential decay. Errors in ${\eta }_{\text{xy}}$ were determined by propagating the standard deviations of repeated delay points in ${\text{R}}_{\text{2α}}$ and ${\text{R}}_{\text{2β}}$ experiments. Since ${\eta }_{\text{xy}}$ is dependent on the chemical shift anisotropy (CSA), purine and pyrimidine values were normalized separately due to their differing CSAs by dividing all values by the maximum corresponding ${\eta }_{\text{xy}}$.

2.6. Small molecule microarray (SMM) screening

To identify e-targeting ligands, a SMM was used (Abulwerdi et al., 2016, 2019; Calabrese et al., 2018; Connelly et al., 2017, 2019; Felsenstein et al., 2016). In brief, $\text{γ}$-aminopropyl silane (GAPS) microscope slides (Corning) were functionalized with a short Fmoc-protected amino polyethylene glycol spacer. After piperidine deprotection, 1,6-diisocyanatohexane was coupled to the surface by urea bond formation to provide functionalized isocyanate-coated microarray slides that react with primary and secondary amines and primary alcohols to create immobilized small molecule libraries. Slides were exposed to pyridine vapor to facilitate covalent attachment, then incubated with a 1:20 polyethylene glycol:DMF (v/v) solution to quench unreacted isocyanate surface. Fluorescently labeled HBV $\text{ε}$ RNA was dissolved in RNase-free distilled water, diluted to 1 or 5 μM in buffer C (25 mM sodium cacodylate, 50 mM KCl, 1 mM MgCl₂, pH 6.9), and annealed by heating to 95 °C for 3 min, snap cooling on ice for 10 min, and slow equilibration to room temperature for 1 h. Following incubation, slides were gently washed twice for 2 min in buffer C with 0.01% Tween-20, once in buffer C, and dried by centrifugation for 2 min at 4000 rpm. Fluorescence intensity (650 nm excitation, 670 nm emission) was measured on an Innopsys Innoscan 1100 AL Microarray Scanner. The scanned image was aligned with the corresponding GenePix Array List (GAL) file to identify individual features. For statistical analysis, hits were defined as [(mean foreground – mean background)/(standard deviation of background)] and Z-score, defined as: Z = (mean SNR₆₃₅ compound – mean SNR₆₃₅ library)/(standard deviation of SNR₆₃₅ library) with the following criteria: (i) SNR > 0, (ii) Z score > 3, (iii) coefficient of variance (CV) of replicate spots < 200, and (iv) [(Z_{RNA incubated} – Z_{buffer incubated})/Z_{buffer incubated}] > 3, and (5) visual inspection and removal of false positives (e.g., dust particulates) (supplementary material, Figure 4A).

2.7. NMR titration experiments

NMR titration experiments to determine chemical shift perturbations (CSPs) were performed on RNA samples transcribed with either uniformly ¹³C/¹⁵N-labeled or ¹³C site-selectively labeled (¹³C-2′,8-Ade; ¹³C-1′,6-Cyt/Uri; and ¹³C-1′,8-Gua) nucleotides. Initial hit compounds were screened by titrating up to 250–600 µM of compound against 50–150 µM ¹³C or ¹³C/¹⁵N labeled RNA samples utilizing ¹H-¹³C sofast-HMQC or HSQC NMR experiments to monitor CSPs of aromatic and ribose resonances (Williamson, 2013). NMR CSPs helped distinguish specific binders, nonspecific binders, aggregators, and non-binders (supplementary material, Figure 5). Priming and apical loop constructs were used to validate specific binders and to rule out allosteric effects (supplementary material, Figures 6 and 7). Data were analyzed using NMRFx and NMRViewJ (Johnson, 2018).

2.8. Dye-displacement assay

A fixed concentration of HBV $\text{ε}$ RNA (500 nM) and SYBRG II (Invitrogen) (4x) was used. 5 μL of compound (in DMSO) and 95 μL of RNA/dye complex in assay buffer D (5 mM sodium cacodylate pH 6.5, 50 mM KCl, 1 mM MgCl₂, 0.1 mM EDTA, 0.01% Triton-X100) were added to black Nunc 96-well plates (Fisher Scientific), incubated at room temperature for 30 min and fluorescence intensity values were measured (485 ± 5 nm excitation, 525 ± 5 nm emission) using a Tecan plate reader. K_d values were determined by normalizing fluorescence intensity of each well to an average value for the fluorescence intensity of RNA/dye complex (GraphPad Prism 7.0 software). Errors in K_d values are reported as the standard error of triplicate experiments.

2.9. Ligand docking predictions

rDock (Ruiz-Carmona et al., 2014) (http://rdock.sourceforge.net/) was used to predict docking poses of Arzoxifene, Raloxifene, and its analog SG92 to HBV $\text{ε}$ RNA. This program offers a dedicated intermolecular scoring function (including van der Waals, polar, and desolvation components) that has been validated against RNA targets. Run as a collection of scripts that can be converted into a swarm of parallel batch jobs, the two main steps in rDock are generation of the docking cavity for the receptor (docking surface interface, generated by the program rbcavity) and actual docking of a ligand accomplished with the program rbdock. rDock treats the docking prediction problem as a search for 3D ligand conformations based on sampling of the exocyclic dihedral angles that yield best docking scores when fit to a rigid target (receptor). The program employs a Genetic Algorithm-based stochastic search algorithm and, as such, has to be run multiple times. Fifty repeated runs of rbdock were sufficient to generate best-scored docking poses, using the 61-nt HBV$\text{ε}$as the potential target. The receptor (HBV $\text{ε}$) input was converted to a MOL2 format, while the ligand conformations were prepared with the quantum mechanics output (see the next section) and converted into the required SDF file format.

2.10. Molecular dynamics (MD) simulations

Based on rDock predictions indicating that the HBV $\text{ε}$ priming loop was the best docking target, NMR conformers 3 and 6 (designated as R3 and R6, respectively) were selected for comparative studies of their unliganded and liganded dynamics, using three independent MD simulations for unliganded and liganded models. Given the consistency between the data, we only illustrate computational docking and MD results for the R3 target. The Amber 16 software package was used to perform MD simulations, employing the force field ff99OL3 (source file leaprc.RNA.OL3), combining parmbsc0 and chi.OL3 parameters for RNA (Case et al., 2005). Partial atomic charges for the ligands tested (Arzoxifene, Raloxifene, and SG92) were computed using the GAMESS (Schmidt et al., 1993) quantum chemistry package. The Density Functional Theory (DFT) (Kohn & Sham, 1965) method utilizing the B3LYP functional (Hertwig & Koch, 1997) and 6–31 G(d) (Hehre et al., 1972) basis set was used to optimize the geometry and subsequent potential derived (PD) atomic charges were computed by fitting to the electrostatic potential via a geodesic point selection scheme (Spackman, 1996). The Amber antechamber package was used to generate a standard MOL2 file for each ligand with 3D coordinates and atom types matched to the general force field – GAFF. Then, ligand partial charges and 3D coordinates generated by quantum mechanics calculations were spliced into that file. Antechamber also generated the ligand library PREP file, while the Amber utility parmchk2 was used to generate an FCRMOD file that contains any force filed parameters not listed in GAFF. The GAFF, PREP, FRCMOD, and the ligand and receptor PDB files were input into the Amber LEaP module which combined them with TIP3P waters, monovalent ions (Na⁺/Cl⁻), and the RNA-specific force field parameters mentioned above to generate the topology and coordinate files.

Explicit solvent molecular particle mesh Ewald (PME) dynamics simulations were utilized (Cheatham et al., 1995). RNA, RNA-Arzoxifene, RNA-Raloxifene, and RNA-SG92 complexes were placed in a cuboid solvent box with TIP3P waters, and the minimum distance between the solute and solvent box boundary was set at 12 Å. The net solute charge was neutralized with Na⁺ ions, and additional Na⁺/Cl⁻ ion pairs were added to simulate the net 0.15 M salt concentration for the entire system in each case. Simulations were run with 2 fs time steps, employing the SHAKE algorithm to constrain all hydrogen bonds in the system. The Berendsen thermostat and Berendsen algorithm were used to maintain the simulation temperature of 300 K and to maintain the pressure at 1.0 Pa in NPT simulations used in all phases of MD (Berendsen et al., 1984). A cutoff of 9 Å for the non-bonded interactions was used. Explicit solvent periodic boundary conditions were employed.

A 12-step equilibration protocol was used in all simulations that started with energy minimization of the solvent (while the RNA or RNA-ligand complex was restrained), followed by multiple short phases of heating to 300 K, dynamics at 300 K, and energy minimizations with gradually decreasing harmonic restraints applied to the solute. The last phase of the equilibration protocol was an unrestrained heating to 300 K, ramped up over 0.2 ns and kept at the steady target temperature for a total time of 2.0 ns. Following equilibration, unrestrained (production) MD simulations were performed for 300 ns. Analyses of the MD trajectories were performed using the Amber cpptraj module. RMSD calculations were performed with cpptraj applied to the MD production trajectories (i.e., excluding the 2 ns-long equilibrations) with the initial NMR conformer used as a reference structure. Phosphorus backbone atoms of the 61-nt HBV $\text{ε}$ fragment and selected priming loop nucleotides and the flanking base pairs (residues 13–20, 48, and 49) were used as RMSD masks. All MD simulations were repeated in triplicate and individual measurements were used to calculate the median and percentiles for each condition (RNA alone and RNA-ligand complex) (SigmaPlot software). Computational docking scores were calculated from independently chosen samples along MD trajectories and used to calculate the mean and standard deviation for each ligand (SigmaPlot software).

2.11. Raloxifene analog synthesis

A small-scale analog library of Raloxifene was synthesized to provide initial structure/activity relationship for Raloxifene binding to the HBV$\text{ε}$priming loop. Synthesis of these analogs is reported in the supplementary material.

3. Results and discussion

3.1. Resonance assignment and structure calculation of unliganded HBV $\text{ε}$

A first step in mapping regions of$\text{ε}$that bind to small molecule ligands requires resonance assignment and 3D structure calculation. To date, the only high-resolution structure of $\text{ε}$ is a truncated 27-nt fragment derived from its apical loop (Flodell et al., 2002, 2006). Herein, we report the first ever structure of full-length$\text{ε}$that includes the previously undetermined priming loop region. To expedite resonance assignment, we sub-divided HBV$\text{ε}$into modular constructs (priming and apical loop, supplementary material, Figure 1A) for solution NMR studies. We observed good agreement between both constructs when comparing the chemical shifts in both the aromatic and ribose regions to the 61-nt construct (supplementary material, Figure 1B and C). The limited changes in chemical shifts suggest that the modular constructs faithfully recapitulate the full-length RNA structure. We therefore used the priming loop construct to simplify unambiguous resonance assignment with various through-bond and through-space multidimensional NMR experiments, which formed the basis for verifying the assignment in the full-length RNA (Figure 1A). To this end, 3D HCP-HCCH total correlation spectroscopy (TOCSY) (Marino et al., 1995) provided sequential resonance assignments for residues 15–18 (Figure 1B), and ¹³C edited nuclear Overhauser effect spectroscopy (NOESY) experiments connected residues 13–20 via inter-strand H1′-H6/8 NOEs (Figure 1C). Additional NMR resonances were assigned using a combined asymmetric ¹³C-labeling and isotopic filter/edit NOESY strategy (Leblanc et al., 2017) on ¹³C site-selectively labeled (¹³C-2′,8-Ade; ¹³C-1′,6-Cyt/Uri; ¹³C-1′,8-Gua) full-length RNA samples (Figure 1D and E). This technique dramatically reduced spectral complexity and crowding, thereby confirming many of the assignments from the priming loop construct. Moreover, these experiments identified additional NOE cross-peaks of Ade-H2 protons for residues A13 and A20 to other protons in the priming loop (Figure 1D and E).

Figure 1.

Resonance assignment of the HBV $\text{ε}$ priming loop. (A) Secondary structure representation of full-length HBV $\text{ε}$ and our priming loop modular construct. The latter contains priming loop residues 14–19, four flanking base-pairs on either side, an additional three base-pairs to stabilize the lower stem and improve transcription yields, and a UUCG tetraloop to close the upper stem. (B) 3D HCP-HCCH TOCSY and (C) ¹³C edited NOESY experiments were carried out on the priming loop construct to initiate resonance assignment. TOCSY data provided unambiguous sequential assignments for residues 15–18. NOESY data connected residues 13–20 via inter-strand H1′-H6/8 NOEs. The peak highlighted in the red box is the H2 resonance of A13. (D) 2D ¹H-¹³C slice from a 3D filter/edit experiment on full-length HBV $\text{ε}$ at the corresponding H2 proton resonance signal. Resonance assignments marked (*) refer to peaks that are folded into the experimental sweep width. (E) Schematic of all NOE contacts of A13 (dashed arrows) and A20 (solid arrows) H2 protons to other protons in the priming loop. White boxes refer to protons attached to a ¹³C nuclei whereas gray boxes correspond to protons attached to a ¹²C nuclei.

To further confirm base-pairing in HBV $\text{ε}$, HNN-COSY experiments were collected on the full-length and priming loop constructs. Using the latter, only 2 A:U and 7 G:C base-pairs [5 A:U and 8 G:C base-pairs are expected from previously reported secondary structure (Flodell et al., 2002, 2006)] were observed. The absence of a G:C base-pair is expected from fraying of terminal base-pairs which are typically not visible by NMR. However, the lack of detectable H-bonds between A13-U49, A20-U48, and A21-U47 in the HNN-correlation spectroscopy (COSY) experiment suggested either solvent exchange or previously incorrect assignment of the secondary structure of HBV $\text{ε}$ (Flodell et al., 2002). When 2-bond (H2)-NN-COSY experiments were collected on the priming loop construct, four of the five possible A:U base-pairs were observed (supplementary material, Figure 2A). A13-U49 was the only undetectable base-pair, again implying U49 does not form an explicit base-pair with A13. U49 was therefore designated unpaired and unrestrained in structure calculations.

NOESY distance restraints were also collected on the full-length priming and apical loop constructs, and normalized against the known H5-H6 distances (2.5 Å) in pyrimidines. Interestingly, an experimental G16-H8 to G16-H1′ NOE cross-peak indicative of a syn nucleobase orientation was observed at short NOE mixing times (50 ms) (supplementary material, Figure 2B). TOCSY data also suggest that G16 has a 2′-endo sugar pucker (supplementary material, Figure 2C).

To complement the NMR observables described above and aid structure interpretation, we measured solvent paramagnetic relaxation enhancement (sPRE) by titrating the full-length RNA sample with an MRI contrast dye and monitoring relaxation enhancements of RNA protons (Hartlmüller et al., 2017). From these sPRE measurements for $\text{ε}$, the slope provided quantitative solvent accessibility and distance-to-surface information (Hartlmüller et al., 2017), with 0 being the least and 1 the most solvent accessible (Figure 2A). Interestingly, these measurements indicated high solvent accessibility for U49 (Figure 2A), implying this residue is either highly solvent accessible or undergoes local motion. Though unexpected, given A13 and U49 are base-paired in the previously predicted secondary structure (Flodell et al., 2002, 2006), the high solvent accessibility of U49 is consistent with undetectable base-pairing revealed by the hydrogen bond experiments. Moreover, A13 does not show high solvent accessibility as measured by sPRE. To better understand local motions within the RNA, NMR relaxation rates were extracted from cross-correlated relaxation (${\text{η}}_{\text{xy}}$) experiments that report on fast internal motions (Lakomek et al., 2013). Normalized values that are significantly above the dashed gray line (mean ${\text{η}}_{\text{xy}}$ value of 0.74) designate residues with reduced ps-ns motions, whereas those below it are flexible (Figure 2A). These data suggest A13 is rigid whereas U49 is flexible, indicating they are unlikely to base-pair, supporting sPRE and H-bond measurements.

Figure 2.

Structure and dynamics of HBV $\text{ε}$. (A) sPRE of full-length HBV $\text{ε}$ was measured for aromatic H2, H5, H6, and H8 as well as ribose H1′ resonances at various Gd(DTPA-BMA) concentrations. Residues in the priming loop (red box), tri-loop (gold box), and U49 show high solvent accessibility, as evidenced by their sPRE values being significantly higher than the dashed gray line (mean value of 0.25). Local dynamics of the aromatic resonances (C6-H6 and C8-H8) of HBV $\text{ε}$ were measured by the normalized exchange-free cross-correlated relaxation rate (${\text{η}}_{\text{xy}}$) (see Materials and methods). Rigid residues are easily identified above the dashed gray line (mean value of 0.74). Flexible residues fall below the dashed gray line and include those of the priming and tri-loop, as well as U43 and U49. Secondary structural elements are labeled with the same abbreviations of Figure 1A. (B) Bundle of the 10 lowest energy NMR structures of HBV $\text{ε}$ RNA obtained from the structure calculation (RMSD of 1.8 Å). (C) Zoomed in view of the priming loop orientation in three of the top 10-ranked NMR conformers (rank 3, 5, and 6). This is shown in two views, highlighting the backbone kink centered at U15 followed by partially stacked G16 and U17. Priming loop (14–19) and flanking residues (13 and 20) are colored red and orange, respectively.

The HBV $\text{ε}$ NMR structure calculation was carried out using the Xplor-NIH 2.48 framework based on the protocol in the gb1_rdc example included in the software package (Schwieters et al., 2018). The protocol included the recently published RNA torsion potential RNA-ff1 (Bermejo et al., 2016). Initial structures were calculated from NOE distance restraints, H-bonding restraints, and dihedral angle restraints. Then, torsion database statistics, residual dipolar coupling (RDC), sPRE, and small/wide-angle X-ray scattering (SAXS/WAXS) data (supplementary material, Figure 3) were used to refine the structure (Figure 2B and Table 1). The nbTargetPot (Hartlmüller et al., 2017) module was used to correlate sPRE data with distance-from-surface refinements. Following standard refinement protocols, the top 10 of 200 NMR structures are reported with an RMSD of 1.8 Å, indicating a well-folded converged structure (Figure 2B). Our reported structure shows good agreement with the existing 27-nt $\text{ε}$ apical loop solution NMR structure (Flodell et al., 2006). Importantly, our structure provides novel insight into the priming loop structure.

Table 1.

NMR and refinement statistics for full-length (61-nt) HBV $\text{ε}$ RNA.

NMR distance and dihedral restraints
Distance restraints
Total NOE	478
Intra-residue	177
Inter-residue	301
Hydrogen bonds	108
RDC	47
Total dihedral-angle restraints	423
Backbone	408
Sugar pucker	15
Other restraints
sPRE	104
Structure statistics
Violations (mean±s.d.)
NOE (Å)	0.08± 0.004
Dihedral-angle (°)	6.64 ± 0.89
RDC (Hz)	0.62 ± 0.09
sPRE (mM−1s−1)	0.57 ± 0.00
Deviations from idealized geometry
Bond length (Å)	0.83 ± 0.01
Bond angles (°)	0.01 ± 0.001
Impropers (°)	0.98 ± 0.02
Heavy atom RMSD from mean structure (°)
Overall (res. 1–61)	1.83 ± 0.14
Lower helix (res. 1–12, 50–61)	1.34 ± 0.08
Priming loop (res. 13–20, 48, 49)	0.85 ± 0.12
Apical loop (res. 21–47)	1.57 ± 0.05

Analysis of the NMR structural ensemble reveals a unique orientation of the 6-nt priming loop bulge. Specifically, residues 15–19 remain well oriented with G16 and U17 partially stacked in three of the top 10-ranked NMR structures (Figure 2C). Moreover, a “kink” is apparent in the backbone of these RNA conformers between residues 14–16 (Figure 2C). Taken together, the tertiary structure of the HBV $\text{ε}$ priming loop bulge resembles a binding pocket typically found in proteins, suggesting that small molecules might be identified that target this structural feature.

3.2. Small molecule microarray (SMM) and binding assays

Our structural analysis of HBV$\text{ε}$suggests that the 6-nt priming loop bulge forms a pocket that would be amenable to small molecule targeting. Previously, an in-house SMM approach was successfully applied to identify a variety of chemotypes targeting both RNA (Abulwerdi et al., 2016, 2019; Connelly et al., 2017, 2019) and DNA (Calabrese et al., 2018; Felsenstein et al., 2016) motifs. Here, fluorescently tagged HBV $\text{ε}$ and a control RNA were used to screen a ~26,000 compound library (supplementary material, Figure 4A). Hit selection required a pipeline comprising statistical analysis, inspection of pharmacophore properties (i.e., amenability of samples to later medicinal chemistry), selectivity, and commercial availability. For each compound, a composite Z-score (see Materials and methods) was calculated based on increased fluorescence at that location on the array in the presence of HBV $\text{ε}$. Compounds with a Z-score >3 were further investigated. Visual inspection of array fluorescence signals and elimination of false positive signals yielded five candidate compounds (supplementary material, Figure 4B). Following identification of these initial hits, NMR titration experiments were used to discriminate specific binders from non-binders, aggregators, and nonspecific binders.

Initial hit compounds were titrated against ¹³C site-selectively labeled (¹³C-2′,8-Ade; ¹³C-1′,6-Cyt/Uri; ¹³C-1′,8-Gua) full-length HBV $\text{ε}$ and ¹H-¹³C sofast-heteronuclear multiple-quantum correlation (HMQC) NMR experiments were utilized to monitor chemical shift perturbations (CSPs) of aromatic resonances (C6-H6 and C8-H8). Site-selectively labeled samples were used to reduce spectral overlap and complexity. This analysis failed to confirm binding of NSC20618 (1,3,-Bis(6-methyl-2-pyridinyl)thiourea) as well as the structurally related NSC20619 (N,N′-Bis(3-methyl-2-pyridinyl)thiourea) (supplementary material, Figure 5). Moreover, at the concentrations tested, NSC300289 (5-nitro-2-(2-pyrrolidon-1-ylethyl)benzo[-de]isoquinoline-1,3-dione, or Pinafide) induced nonspecific peak broadening indicative of HBV $\text{ε}$ aggregation and was not considered further (supplementary material, Figure 5). In contrast to the previous three compounds, titrations of NSC9921 (4,4′-(pentane-1,5-diylbis(oxy))dibenzimidamide, or Pentamidine) did show HBV $\text{ε}$ binding. Specifically, Pentamidine titrations led to CSPs in a large number of aromatic resonances from residues located in both the priming and apical loops, suggestive of nonspecific binding (supplementary material, Figure 5). Finally, we observed that NSC747974 ((5-hydroxy-2-(4-hydroxyphenyl)-1H-inden-1-yl)(4-(2-(piperidin-1-yl)ethoxy)phenyl)methanone, or Raloxifene) titrations provided CSPs of aromatic resonances exclusive to priming loop residues (Figure 3A-C and supplementary material, Figure 6A). Specifically, CSPs were detected for residues 13–15, 18–19, and 48–50 (Figure 3B and C and supplementary material, Figure 6A).

Figure 3.

The SERM Raloxifene recognizes the HBV $\text{ε}$ priming loop. (A) Chemical structure of Raloxifene with a K_d value determined from a dye-displacement assay (see Supplementary Figure 9). (B) Map of all CSPs per HBV $\text{ε}$ residue. Secondary structural elements are labeled with the same abbreviations of Figure 1A. Priming loop residues 13–19 (light red box) and flanking residues 47–50 show the strongest CSPs, suggesting that Raloxifene binding is localized to the priming loop. Average CSP is shown with a dashed gray line (mean value of 0.034). (C) Secondary structure representation of full-length HBV $\text{ε}$ with residues with strong CSPs from B shown in bold and colored red.

To further confirm that Raloxifene targets the HBV $\text{ε}$ priming loop, we repeated our ¹H-¹³C sofast-HMQC NMR titration experiments with Raloxifene and Pentamidine using the apical loop construct. In agreement with previous observations, this analysis revealed that only Pentamidine binds the HBV $\text{ε}$ apical loop (supplementary material, Figure 6B). For completeness, we then titrated Raloxifene against uniformly ¹³C/¹⁵N-labeled priming loop construct and used ¹H-¹³C heteronuclear single-quantum correlation (HSQC) NMR experiments to monitor CSPs of aromatic (C2-H2, C5-H5, C6-H6, and C8-H8) and ribose (C1′-H1′) resonances. These experiments permitted detection of additional resonances that were not probed in the site-selectively labeled samples. In agreement with previous titrations, Raloxifene perturbed resonances exclusive to the priming loop (supplementary material, Figure 7). Notably, Raloxifene induced greater CSPs at C1′-H1′ and Ade-C2-H2 resonances suggesting direct targeting of the minor groove of the priming loop pocket. Collectively, our NMR titration data unambiguously demonstrate that Raloxifene targets the HBV $\text{ε}$ priming loop, supporting our structure-informed pocket hypothesis.

Raloxifene is a benzothiophene belonging to the class of SERMs and is in clinical use for treatment of osteoporosis by mimicking the effects of the hormone estrogen to increase bone density. Raloxifene is also proposed to lower the risk of breast cancer by blocking the effects of estrogen on breast tissue (Valentovic, 2007). The benzothiophene, Arzoxifene, and the phenylindole, Bazedoxifene, are closely-related SERMs that have also been under clinical investigation (Maximov et al., 2013) (supplementary material, Figure 8A). To investigate whether these compounds also bind $\text{ε}$, NMR titration experiments were again performed. This analysis revealed that Bazedoxifene also targets HBV$\text{ε}$at the priming loop, albeit with a smaller subset of CSPs than Raloxifene (supplementary material, Figure 8B). Arzoxifene, on the other hand, showed no binding (supplementary material, Figure 8B). As an orthogonal measure of SERM binding, a dye-displacement assay was used. These data corroborate the NMR titration data and suggest that Raloxifene and Bazedoxifene have K_d values of 69.2 ± 6.7 and 107.0 ± 32.1 μM, respectively, whereas Arzoxifene did not bind (Figure 3A and supplementary material, Figure 9).

This is the first report of selective binding of these SERMs to structured RNA motifs. It is desirable to obtain a Raloxifene-bound HBV$\text{ε}$NMR structure. Unfortunately, saturation of the RNA is not feasible at the NMR concentrations needed for such experiments, due to Raloxifene insolubility. Preliminary NOESY experiments of the HBV $\text{ε}$-Raloxifene complex were hampered by low signal-to-noise and the absence of cross-peaks (data not shown). Alternative computational approaches were therefore explored to provide a detailed understanding of the HBV $\text{ε}$-Raloxifene interaction.

3.3. Computational docking and molecular dynamics (MD) analysis

To gain further insight into the HBV $\text{ε}$-Raloxifene interaction, computational docking and MD analysis were adopted. Raloxifene docking pose predictions generated by rDock (Ruiz-Carmona et al., 2014) indicated that among the top 10-ranked NMR conformers, three (rank 3, 5, and 6) could be targeted directly at the priming loop, with NMR rank 3 scoring best, rank 6 a close second, and rank 5 scoring third. Interestingly, these three conformers share the unique, well defined priming loop orientation shown in Figure 2C. The predicted docking pose reveals that the Raloxifene core is wedged deeply into the priming loop between residues 15–19 (with G16, U18, and C19 rotated away) and is also close to U48 and U49 (Figure 4A and B). The scoring gap between the first two and the third ranked docking pose was 8.5%, and thus we focused on the top two ranked docking targets, designated as R3 and R6. Given that our analysis showed consistency between the R3 and R6 data, we only illustrate computational docking and MD results for the R3 target. It is worthwhile noting that the docking prediction search space was not restricted to the priming loop and therefore our findings of the best-scored docking poses in the priming loop, which are consistent with NMR titration data, were not biased by the input parameters.

Figure 4.

Molecular dynamics (MD) and computational docking of Raloxifene to HBV $\text{ε}$. (A) Top ranked (scored) predicted docking pose of Raloxifene to the HBV $\text{ε}$ NMR conformer R3. Raloxifene preferentially docks to the priming loop, in agreement with NMR titration data. In all structural representations, priming loop (14–19) and flanking residues (13, 20) are colored red and orange, respectively, as in Figure 2B and C. Models in A-D are aligned to the priming loop of R3 shown in A. Overlay of MD conformations with Raloxifene (B) docked or (C) unliganded HBV $\text{ε}$. All MD runs (sampled every 10 ns from one 300 ns-long MD simulation) were performed in triplicate. (D) Same results as in B, but with Raloxifene hidden to best illustrate the relatively stable backbone and nucleotide orientations in the liganded priming loop. (E) Box plots of the cumulative RMSDs measured for the priming loop and flanking base pairs in MD simulations for unbound (red) and Raloxifene docked (cyan) HBV $\text{ε}$. Median values correspond to the horizontal lines inside the boxes, while the upper and lower box boundaries indicate the 25 and 75 percentile values. “Whiskers” indicate 10 and 90 percentile values. Raloxifene binding significantly lowers the RMSD of the HBV $\text{ε}$ priming loop residues. The difference in the median values is statistically significant in Mann-Whitney Rank Sum Test (p ≤ 0.001).

MD simulations (repeated three times for the R3 and R3-Raloxifene complexes) demonstrated higher flexibility of the priming loop in the unliganded RNA simulations. In all three 300 ns-long simulations of the R3-Raloxifene complex, Raloxifene remained stably bound to the RNA target, suggesting a valid docking pose prediction. A sample pair of runs of the unliganded (Figure 4C) and Raloxifene-bound (Figure 4D) RNA demonstrate stabilization of the priming loop upon ligand binding. Box plots illustrate the difference in the cumulative RMSDs measured for the priming loop and flanking base pairs (residues 13–20, 48, and 49) in the MD simulations (Figure 4E). Raloxifene stabilized the conformation of the priming loop region, lowering its mean RMSD value from 5.2 ± 1.7 Å to 3.6 ± 0.8 Å. The difference in the median values between the two groups is statistically significant based on the Mann–Whitney rank sum test (p ≤ 0.001), with the median value of 5.4 Å for R3 and 3.5 Å for the R3-Raloxifene complex. The same trend holds for a comparison of the RMSDs of single-stranded nucleotides of the priming loop (means of 3.2 ± 0.9 Å versus 2.4 ± 0.5 Å for residues 14–19).

MD simulations of unliganded and Raloxifene-bound RNA also indicate that A13 and U49 remain unpaired. Low levels of H-bond occupancy between the two residues were observed for the priming loop construct. Moreover, in the full-length RNA, both A13 and U49 form intermittent H-bond interactions with priming loop residues 14–19 and most often 16–18. This intermittent base-pairing is supported by the previously unassigned imino peaks that appear in NMR experiments at low temperature (supplementary material, Figure 2B). In general, A13 remains stacked with U12 and in the majority of cases it also stacks with C14 (both in R3 and R3-Raloxifene and R6-Raloxifene complexes), extending the helical geometry of the lower stem into the priming loop, even in the absence of the A13-U49 base-pair. In contrast, U49 is most often rotated away from A13 and is more exposed to solvent, while remaining stable in multiple arrangements. These observations offer additional support for the interpretation of the NMR results for A13 and U49 discussed earlier (Figure 2A and supplementary material, Figure 2A).

3.4. Interaction of raloxifene analogs with HBV $\text{ε}$

A detailed view of the docking poses of the HBV $\text{ε}$-Raloxifene complex reveals that the hydroxyethylpiperidine tail occupies the groove of the RNA and rarely interacts with the binding pocket formed by the priming loop (Figure 5A). This observation suggests that medicinal chemistry efforts could be employed to modulate small molecule binding to HBV $\text{ε}$. To this end, several Raloxifene analogs were synthesized and their affinities were determined by a dye-displacement assay (Supplementary material, Figures 10 and 11) For simplicity, Raloxifene was subdivided into three “units”: (i) a hydroxyethylpiperidine tail, (ii) a 3-(carbonyl) position hinge, and (iii) a 6-,4′-substituted phenylbenzothiophene. With respect to the phenylbenzothiophene, replacing the 4′-OH with −Br (SG70) or −OCH3 (SG74) severely compromised affinity (supplementary material, Figures 10 and 11). In the presence of the hydroxyethylpiperidine tail, replacing the 3-carbonyl with -OH (SG102) or removing it entirely (SG113) decreased affinity (supplementary material, Figures 10 and 11). Interestingly, removing the tail (SG92) increased affinity ~2-fold (Figure 5B), suggesting the tail is dispensable, in agreement with docking observations.

Figure 5.

Medicinal chemistry and computational docking of Raloxifene and its analog SG92 to HBV $\text{ε}$. (A) Zoomed view of the best-scored Raloxifene docking pose with HBV $\text{ε}$ NMR conformer R3. The Raloxifene hydroxyethylpiperidine tail rarely participates in binding to HBV $\text{ε}$. (B) Chemical structure of SG92 with a K_d value determined from a dye-displacement assay. In agreement with observations from docking, removing the tail improved binding. (C) The best predicted docking pose of SG92 to R3. In all structural representations, priming loop (14–19) and flanking residues (13, 20) are colored red and orange, respectively, as in Figure 2B and C. (D) In agreement with K_d data, SG92 shows a higher mean rDock score than Raloxifene and Arzoxifene for the 100 best-scored docks predicted for 600 samples of each HBV $\text{ε}$-ligand 300 ns-long MD trajectory. These are intermolecular docking scores normalized by the size (heavy atoms) of each ligand. The difference in the median values is statistically significant in Mann-Whitney Rank Sum Test (p ≤ 0.001).

The docking pose predictions generated by rDock also demonstrate that SG92 directly targets the priming loop with an orientation nearly identical to Raloxifene, especially in the core region (i.e., the 3-(carbonyl) position hinge and the 6,4′-substituted phenylbenzothiophene) (Figure 5A and C and supplementary material, Figure 12). Indeed, SG92 binding was confirmed by NMR titration experiments with full-length HBV $\text{ε}$. As with Raloxifene, SG92 targeted the priming loop (supplementary material, Figure 13).

In addition, comparative docking was used to complement our experimental binding data. That is, Arzoxifene, Raloxifene, and SG92 were all docked to HBV $\text{ε}$ NMR conformer R3 to represent a non-binder, a baseline binder, and a good binder, respectively. However, several published studies have previously demonstrated bias in the computational docking scores correlated with the size of the ligand, and suggested various normalizations based on the number of heavy atoms in the ligand or the molecular weight of the ligand (Carta et al., 2007; Jacobsson & Karlén, 2006; Pan et al., 2003). Comparisons of our experimental results to the absolute values of the rDock docking scores of the three ligands also indicated bias toward larger Raloxifene and Arzoxifene over a much smaller and better binder SG92. To better compare and rank docking predictions for the three ligands, we used the intermolecular scores normalized on the number of heavy atoms in the ligands, as listed by rDock for all the generated poses and the 100 best predicted poses among 600 conformer targets extracted every 0.5 ns from 300 ns-long HBV $\text{ε}$ MD trajectory. These results correctly rank SG92 as the best binder to the sample of MD targets, superior to Raloxifene and Arzoxifene. Figure 5D illustrates the normalized-score ranking for the best 100 scores per ligand, with 1.23 ± 0.07 for Arzoxifene, 1.29 ± 0.07 for Raloxifene, and 1.45 ± 0.06 for SG92. The difference in mean rDock scores is statistically significant based on the Mann-Whitney Rank Sum Test (p ≤ 0.001). In addition, MD simulations of the HBV $\text{ε}$-Arzoxifene complex for the best predicted docking poses of Arzoxifene indicated dissociation from the RNA target within 100 ns, consistent with NMR titration and dye-displacement data that indicated no binding of Arzoxifene to HBV$\text{ε}$(supplementary material, Figures 8B and 9).

5. Conclusions

The critical role of HBV $\text{ε}$ RNA in both pgRNA encapsidation and initiation of HBV (−) strand DNA synthesis provides an attractive target for small molecule RNA antagonists, for which several precedents have been documented for regulatory RNAs (Abulwerdi et al., 2016; Connelly et al., 2017; Disney et al., 2014; Stelzer et al., 2011). Success in this direction would be aided by a detailed understanding of $\text{ε}$ structure and dynamics, for which only limited information exists (Flodell et al., 2002, 2006; Petzold et al., 2007), and characterizations of potential ligands regarding their interactions with its dynamic 6-nt loop. In this communication, we address both issues, providing the first high-resolution solution NMR structure of full-length$\text{ε}$and demonstrating the binding characteristics of three documented SERMs and a fourth, in-house synthesized analog, to the 6-nt priming loop.

Experimental NMR was used to determine the average solution structure of the 61-nt HBV $\text{ε}$. Xplor-NIH calculations generated structures with a low RMSD (1.8 Å) indicative of a well folded, converged structure (Table 1). Of particular interest with respect to the priming loop structure is the backbone kink centered at U15 and followed by the partially stacked G16 and U17 (Figure 2E). In vitro biochemical studies show stalling of the HBV replication complex following reverse transcription of these nucleotides, potentially as a prerequisite to (−) strand DNA transfer (Beck & Nassal, 2007). Although a high-resolution structure model of P is unavailable, the topology of all crystallized nucleic acid polymerases indicates a common architecture, assembling into a shape likened to a right hand, comprising “fingers,” “palm,” and “thumb” domains (Flodell et al., 2002, 2006). While the palm is involved in catalysis of phosphoryl transfer, the fingers domain interacts with the incoming dNTP and the template nucleotide to which it is base-paired. Lastly, the thumb participates in positioning the nucleic acid duplex, processivity and translocation, potentially serving as a “sensor” of nucleic acid configuration (Flodell et al., 2002, 2006). Thus, one might imagine a scenario where the stacking and unstacking of G16 and U17 could facilitate initiation of DNA synthesis, while sensing of the kink turn at U15 by the thumb of P arrests translocation and, consequently, initiates the first of three strand transfer events. If this hypothesis were true, then small molecules that recognize the priming loop would be attractive therapeutic targets. While this notion is currently speculative, data presented herein suggest this possibility.

The multitude of NMR data used for structure modeling, including sPRE and NMR relaxation measurements, suggest the priming loop is flexible. Given that the priming loop residues are well conserved, their flexibility is likely critical for function. Indeed, considerations of RNA dynamics in small molecule targeting has shown promising results in RNA-targeted drug discovery (Stelzer et al., 2011). From our 3 D structure, the tertiary architecture of the 6-nt priming loop bulge resembles a binding pocket typically found in proteins, suggesting that small molecules binding to this structural element can be identified. HTS and NMR titration experiments indicate modest interactions between Raloxifene and Bazedoxifene and the HBV$\text{ε}$priming loop with low micromolar affinity (Figure 3 and supplementary material, Figures 8B and 9). NMR titrations with apical (supplementary material, Figure 6B) and priming loop (supplementary material, Figure 7) constructs further support the notion that Raloxifene exclusively targets the priming loop, in agreement with our structure-informed pocket hypothesis.

To better understand the priming loop dynamics and visualize its interactions with SERMs, computational docking and MD simulations have proven useful. Docking revealed that three of the top 10-ranked NMR conformers correlated well with predictions of the best docking poses. Each conformer could be directly targeted at the priming loop and show strong agreement with NMR titration experiments. The best-scored NMR conformer R3 was subjected to MD simulations in unliganded and Raloxifene-bound forms (Figure 4). The priming loop was significantly more flexible in the absence of a small molecule (Figure 4C and D). In particular, when bound to the priming loop, Raloxifene stabilized its conformation and reduced the RMSD of priming loop nucleotides. This again suggests that RNA dynamics must be taken into account when considering small molecule antagonists.

As a preliminary medicinal chemistry exercise, we synthesized a Raloxifene analog SG92 lacking the hydroxyethylpiperidine tail, which our docked models indicated was not necessary for binding (Figure 5A–C). Indeed, SG92 bound the HBV $\text{ε}$ priming loop with a ~2-fold increase in affinity, further supporting our MD docked models (Figure 5). Taken together, our initial solution NMR structural model of the full-length HBV $\text{ε}$ RNA element, coupled with MD simulations, provide a platform for structure-based rational discovery of small molecule binders. In the future, it will be of great interest to test Raloxifene and its derivatives in a cell-based assay.