Virtual Screening for RNA-Interacting Small Molecules

Abstract

Computational virtual screening is useful and powerful strategy for rapid discovery of small biologically active molecules in the early stage of drug discovery. The discovery of a broad range of important biological processes involved by RNA has increased the importance of RNA as a new drug target. To apply structure-based virtual screening methods to the discovery of RNA-binding ligands, many RNA 3D structure prediction programs and optimized docking algorithms have been developed. In this chapter, a number of successful cases of virtual screening targeting RNA will be introduced.

Virtual Screening and Docking Tools

Computational virtual screening strategy is useful in the very early stage of the drug discovery pipeline and provides a powerful tool for rapid discovery of small biologically active molecules. Such strategy can decrease the number of candidate compounds providing a good starting point for chemical synthesis and biological screening. Therefore, in many cases, virtual screening can be used prior to expensive experimental highthroughput screening. For this reason, since the terms of computational virtual screening came out in the late 1990s, it has been considered as a novel and essential technology in drug discovery. Ligand- and structure-based virtual screenings have been successfully applied to drug discovery programs in various disease areas. After prospective results for various protein targets were obtained, the utility of virtual screening to identify compounds for nucleic acid-based receptors has been the focus of much attention.

Many computational docking programs that can automatically dock small molecules into a binding site of a target receptor with minimum input from an operator have been developed and their applicability has been proven (McInnes 2007). In addition, the development of various free or commercially available databases allows for easy use of virtual screening. Each year, many successful cases of virtual screening against various targets including protein and nucleotides have been reported in several major review papers (Whitty and Kumaravel 2006; Seifert and Lang 2008; Villoutreix et al. 2009).

The docking process concerns the prediction of ligand conformation and orientation within a targeted binding site (active site) (Kitchen et al. 2004). In order to carry out docking calculations, it is necessary to know the 3-dimensional (3D) structure of a target and the nature of the binding site. 3D structures are identified by X-ray crystallography or NMR experiments and can be downloaded from Protein Data Bank (PDB) (Berman et al. 2000) or predicted by homology modeling using various programs. The next step is to define the binding site by known information or prediction. When input is prepared, chemical compounds present in the database are docked into the defined binding site of the selected target receptor. There are two purposes of docking studies. One is to predict accurate ligand binding orientation referred as “molecular modeling” and the other is to predict activity or binding affinity referred as “scoring.” A docking result is evaluated by ligand binding orientation through visual inspection and by binding affinity using the scoring function. Scoring function is designed to predict the biological activity or binding affinity through the evaluation of interactions between ligands and receptor (Halperin et al. 2002). Among various docking programs (Table 12.1), the most widely used programs are AutoDock (Morris et al. 1998; Huey et al. 2007), DOCK (Ewing et al. 2001), FlexX (Kramer et al. 1999; Stahl 2000), Gold (Jones et al. 1995), Glide (Zhou et al. 2001), Internal Coordinate Mechanics (ICM) (Abagyan and Totrov 1994), and Surflex-Dock (Jain 2003; Kellenberger et al. 2004). AutoDock is an automated flexible docking program designed to predict how small molecules (ligands) bind into the receptor structure. AutoDock is performed with an empirical free energy force field based on a Lamarckian genetic algorithm, to bring about speedy prediction of conformation with calculated free energies of association (Morris et al. 1998). This program has application in X-ray crystallography, structure-based drug design, virtual screening, and protein–protein interaction studies. DOCK was introduced by the Kuntz group at UCSF and uses a rigid body docking algorithm and flexible ligand docking algorithm to dock the ligand into a negative image of the binding pocket (Ewing et al. 2001). FlexX is a fully automatic computer program for predicting protein–ligand interaction. FlexX can predict not only the lowest energy geometry of the complex of ligand with protein but also the binding affinities using an empirical scoring function (Böhm 1994). The descriptions of other widely used docking programs including commercial and free softwares are listed in Table 12.1.

Table 12.1.

Widely used docking algorithms

Name	URL	Short summary
AutoDock	http://autodock.scripps.edu/	Free software of automated small-molecules docking tools (rigid receptor, flexible ligand)
DOCK	http://dock.compbio.ucsf.edu/	Free software, small molecules-various receptor docking (protein, DNA, and RNA), protein–protein interaction
FlexX	http://www.biosolveit.de/flexx/	Fast computer program for predicting protein–ligand interactions
GOLD	http://www.ccdc.cam.ac.uk/products/life_sciences/gold/	Calculating the docking modes of small molecules in protein binding sites
Glide	http://www.schrodinger.com/	Fast flexible ligand docking program (small molecule-protein)
ICM	http://www.molsoft.com/	Automatic incorporation of flexibility into the ligand and receptor docking (protein, DNA, and RNA), protein–protein interaction
MORDOR	http://mondale.ucsf.edu/index_main_frame.html	Docking program using algorithm considering flexibility of both nucleic acid receptor and ligand
Surflex-Dock	http://www.tripos.com/	“Protomol”-guided flexible molecular docking program

Computational Programs to Predict 2D and 3D RNA Structures

In recent years, while the number of identified RNA sequences has rapidly increased, the number of known 3D structures has not kept pace with it. For this reason, there is a large gap between the number of known RNA sequences and 3D structures. For example, tRNA is one of the most structurally well characterized RNAs and its 1,101,833 characterized sequences are reported in the Rfam (Gardner et al. 2009), a database of sequence families of structural RNAs; however, only 170 structures are reported. To apply structure-based drug design approaches to the identification of RNA binding ligands, computational programs are required for prediction of RNA structures. Several computer programs have been developed for folding of RNA secondary structures, and modeling of RNA 2D and 3D structures. Those computational tools are summarized in Tables 12.2 and 12.3.

Table 12.2.

RNA secondary structure viewers/editors programs

Name	URL	Description
PseudoViewer (Han et al. 1999; Byun and Han 2009)	http://wilab.inha.ac.kr/pseudoviewer/	visualize RNA pseudoknot 2D structure automatically
RNAdraw (Matzura and Wennborg 1996)	http://www.rnadraw.com/	RNA 2D structure prediction, analysis, and visualization
RNA Movies (Evers and Giegerich 1999)	http://bibiserv.techfak.uni-bielefeld.de/rnamovies/	System for the visualization of RNA secondary structure spaces
RNAView/RnamlView (Yang et al. 2003)	http://ndbserver.rutgers.edu	Automatically generate 2D displays of RNA/DNA secondary structures with tertiary interactions
VARNA (Darty et al. 2009)	http://varna.lri.fr	Automated drawing, visualization, and annotation of the secondary structure of RNA
RNA Designer (Andronescu et al. 2004)	http://www.rnasoft.ca/cgi-bin/RNAsoft/RNAdesigner	Design de novo RNA structures with certain structural properties
RnaViz 2 (De Rijk et al. 2003)	http://rnaviz.sourceforge.net/	Drawings of RNA secondary structure with portability and structure annotation
Vienna RNA (Hofacker 2003)	http://www.tbi.univie.ac.at/~ivo/RNA/	Program for the prediction and comparison of RNA secondary structures

Table 12.3.

Computer program for predicting RNA 3D structure

Name	URL	Short summary
Mode RNA (Rother et al. 2011)	http://genesilico.pl/moderna/	Automatic program for comparative modeling of RNA 3D structures. Require sequence alignment and structural template
MMB (formerly RNA Builder) (Flores et al. 2010)	https://simtk.org/home/rnatoolbox	Automatic program for generating model RNA structures from 2D and template structure by simulating in parallel at multiple levels of details
PARADISE (Assemble and S2S) (Jossinet et al. 2010; Jossinet and Westhof 2005)	http://paradise-ibmc.u-strasbg.fr/	Assemble: automatic program for intuitive graphical interface to study and construct complex 3D RNA structures. S2S: graphical system to easily display, manipulate, and interconnect heterogeneous RNA data like multiple sequence alignments, 2D and 3D structures
RNA2D3D (Martinez et al. 2008)	http://www-lmmb.ncifcrf.gov/	Manual manipulation program for RNA 3D modeling with conversion of RNA 2D structures to 3D
ERNA-3D (Zwieb and Müller 1997)	http://www.erna-3d.de/	Molecular Modeling Expert System to develop for the generation of models of RNA and protein molecules. Need to manual manipulation
MC-Fold/MC-Sym (Parisien and Major 2008)	http://www.major.iric.ca/MC-Pipeline/	A web-hosted service for prediction of RNA 3D structure with input 2D secondary
NAST (Jonikas et al. 2009)	https://simtk.org/home/nast	Predicting RNA 3D model from 2D structure
DMD/iFoldRNA (Sharma et al. 2008)	http://danger.med.unc.edu/tools.php	Web portal for interactive RNA folding simulations. Enable to perform molecular dynamics simulations of RNA using coarse-grained structural models (two-beads/residue)
YUP (Tan et al. 2006)	http://rumour.biology.gatech.edu/YammpWeb/	RNA molecular modeling program based on PYTHON. Offer such methods as Monte Carlo, Molecular Mechanics and Energy Minimization

RNA-Targeted Virtual Screening

Many clinical antibiotics including macrolides, aminoglycosides, and others targeting bacterial ribosomal RNA (rRNA) reveal that RNA is the important target for drug development (Knowles et al. 2002; Hermann 2005). The appearance of drug resistance is the most critical problem in treating bacterial (Neu 1992) and viral infections (Perrin and Telenti 1998). RNAs contain highly conserved structural and functional motifs that may serve as drug targets, so the development of resistance to drugs targeting RNA can be slower than that to drugs targeting protein (Gallego and Varani 2001). In contrast to DNA, which mostly has a double-stranded helix structure, RNA is generally single-stranded and folds into complex 3D structures that provide unique pockets for small molecules (Foloppe et al. 2006), thus making RNA an attractive drug target.

In spite of these advantages, RNA has not been focus of structure-based drug design, not only due to lack of information of RNA 3D structures, but also due to the sequence-specific unique features of the binding pockets in RNA. The binding pocket of protein usually lies deep in an internal region, separated from solvent. In RNA targets, the binding pockets are large and flat, located along the surface, and relatively exposed to solvent. Therefore, in using docking algorithms to discover RNA-binding drugs, the physicochemical properties of RNA, such as conformational flexibility, high negative charge, and solvation, should be taken into account more accurately than those for proteins. Despite such differences between protein and RNA targets, classical protein-ligand docking programs have sometimes successfully performed in RNA-targeted virtual screenings. For example, Kuntz first reported a successful virtual screening using DOCK 3.5 program to identify small molecules in the Available Chemicals Directory (ACD) that targeted an RNA double helix (Chen et al. 1997). Following the first study, many research groups have reported successful studies of virtual screening targeting RNA through protein-based docking methods (Filikov et al. 2000; Lind et al. 2002; Kang et al. 2004; Park et al. 2008, 2011). Meanwhile, since most of the available docking methods were developed for protein targets, their compliance with RNA targets has been evaluated extensively, and recently Li et al. demonstrated that two widely-used protein docking programs, GOLD 4.0 and Glide 5.0, are appropriate for structure-based drug design and virtual screening for RNA targets (Detering and Varani 2004; Li et al. 2010).

Several modified algorithms were established for RNA-targeted virtual screening. For example, Morley and Afshar established an empirical scoring function that appropriately describes steric, polar, and charged interactions in RNA-ligand complexes (Morley and Afshar 2004). This scoring function was implanted in RiboDock program and validated for docking screening of a large-size chemical database. Virtual screening with this program successfully identified novel ligands for the bacterial ribosomal A-site (Foloppe et al. 2004). Moitessier et al. applied a unique method to the AutoDock docking process, taking into account inherent RNA flexibility and key water molecules, and this modified docking tool was validated by the docking of aminoglycosides to the ribosomal RNA A-site (Moitessier et al. 2006). In Kuntz et al.’s study, 70 experimental RNA-ligand complexes from PDB were re-docked using the DOCK 6 program, and the resulting docked conformations were rescored with AMBER generalized Born/solvent-accessible surface area (GB/SA) and Poisson–Blotzmann/SA (PB/SA) scoring functions in combination with explicit water molecules and sodium counterions. The success rate for reproducing experimental binding modes was significantly improved by using AMBER GB/SA or PB/SA (Lang et al. 2009).

As small molecules induce RNA conformational changes by binding to structures from preexisting dynamic ensembles (Puglisi et al. 1992; Zhang et al. 2007; Frank et al. 2009; Cruz and Westhof 2009; Fulle and Gohlke 2010), large conformational changes in RNA receptors after binding with small molecules during virtual screening must be accounted for. However, current protocols do not consider this point, limiting the range of target structures for the discovery of small molecules. Guilbert and James therefore developed a flexible docking program called MORDOR, which supports flexibility in the ligand and limited flexibility in the RNA for induced-fit binding. MORDOR performed well not only on 57 test sets of RNA-ligand complexes by retrieving experimental poses within 2.5 Å with a 74% success rate, but also in discovering ligands for novel targets such as human telomerase RNA (Guilbert and James 2008; Pinto et al. 2008). When comparing the practicality of DOCK 6 and MORDOR, DOCK 6 screened about 3–10 times faster than MORDOR, while MORDOR performed better on ligands with a large number of rotatable bonds (Lang et al. 2009). Most recently, Al-Hashimi’s group reported a new strategy for virtual screening targeting an RNA dynamic ensemble constructed by combining NMR spectroscopy and computational molecular dynamics (MD) (Stelzer et al. 2011). This strategy takes into account large degrees of RNA conformational adaptation during virtual screening. This approach was applied to a search for small molecules for HIV-1 TAR (transactivation response element) RNA (vide infra).

Successful Application of Virtual Screening to RNA Receptors

Case 1: HIV-1 TAR RNA Hairpin

As one of the best characterized RNA-based regulatory machineries, the interaction of TAR with Tat protein has been focused upon as a target to inhibit HIV-1 replication (Yang 2005). Ligands that inhibit the TAR-Tat interaction can be developed as anti-AIDS drugs. Various molecules for TAR RNA that inhibit Tat binding have been identified, and many of those molecules were discovered by virtual screening of chemical libraries. In earlier studies, James used automatic docking methods (DOCK and ICM), employing flexible docking with Monte Carlo simulation and optimized scoring function, identifying phenothiazine compounds from ACD as TAR RNA ligands (Filikov et al. 2000; Lind et al. 2002). To circumvent the limitation of incorporating RNA flexibility for structure-based virtual screening, and to find novel scaffolds for TAR-Tat inhibitors, ligand-based virtual screening was conducted. The SQUID fuzzy pharmacophore search method successfully identified a novel heterocyclic compound with an order of magnitude improved activity compared to known phenothiazine compounds (Renner et al. 2005).

In 2011, Al-Hashimi’s group developed the most outstanding technology for RNA-targeted virtual screening through intensive generation of an ensemble of TAR RNA conformers and a robust re-docking validation test. Their strategy was effectively applied to virtual screening of a relatively small-sized chemical library containing 51,000 compounds, and they identified netilmicin, a selective HIV-1 TAR RNA binder, that inhibited HIV-1 replication in vivo (Stelzer et al. 2011). Details of their study are described below.

Validation of Docking Program

To test the accuracy of docking, a total of 96 small molecule-bound RNA structures downloaded from the PDB were assessed for docking performance. All docking performances were carried out using the ICM docking program (Abagyan and Totrov 1994) and results were evaluated by ICM Score and RMSD between native ligand (extracted from RNA structures) and predicted orientation after docking. The binding energies based on the ICM Score were predicted with high accuracy (R  =  0.71). In more than half of cases (53%), the predicted conformations matched the X-ray or NMR structures within 2.5 Å RMSD.

Preparation of HIV-1 TAR RNA Ensemble

To decide on a suitable RNA ensemble, the accuracies of docking with two sets of RNA ensembles of HIV-1 TAR were compared. One ensemble, named TAR^NMR-MD, consisted of 20 conformers generated by SAS selection (select-and-sample strategy, Frank et al. 2009). To construct TAR^NMR-MD, HIV-1 TAR structure (PDB id: 1ANR, Aboul-ela et al. 1996) was downloaded and molecular dynamics simulations were performed by measuring NMR residual dipolar couplings (RDC) in elongated RNA. The other RNA ensemble, named TAR^MD, consisted of 20 randomly selected snapshots from an 80-ns MD simulation of apo-TAR with backbone RMSDs ranging from 3 to 80 Å. The X-ray structure (PDB id: 397D, Ippolito and Steitz 1998) and 20 NMR structures (PDB id: 1ANR, Aboul-ela et al. 1996) of apo-TAR were downloaded from the PDB. A test set ligand library containing 38 known ligands for TAR RNA was obtained from the published literatures. To test the accuracy of the RNA ensemble, virtual screening of the test set ligands targeting the RNA ensembles, TAR^NMR-MD and TAR^MD, was conducted, respectively. Virtual screening was conducted using ICM after binding pockets were predicted by the ICM PocketFinder module based on the calculated surface area and volume of cavities on the receptor surface. The results suggested that TAR^NMR-MD improved the accuracy of docking compared with TAR^MD. From this result TAR^NMR-MD was chosen for further virtual screening to identify small molecule TAR-Tat inhibitors.

Virtual Screening Against TAR Dynamic Ensemble

A chemical database consisting of 49,166 compounds was obtained from the Center for Chemical Genomics at the University of Michigan and 2,060 compounds from the author’s in-house library. Based on the ICM Score, the top 57 commercially available hit compounds were selected and their binding activities were tested by fluorescence-based assays. This identified six small molecules that bound to TAR with high affinity (K _d  =  55 nM to 122 μM) and inhibited TAR interaction with Tat (K _i  =  710 nM to 169 μM) in vitro. Among them, netilmicin (K _d  =  ∼1.4 μΜ) bound to HIV-1 TAR with the highest selectivity over HIV-2 TAR. Netilmicin repressed Tat-mediated transactivation of the HIV-1 promoter through its interaction with TAR in live human T-cells and inhibited HIV-1 replication in the HIV-1 indicator cell line TZM-bl.

Case 2: RNA Pseudoknots

Ribosomal −1 frameshifting (−1 FS) is an essential event during translation for the synthesis of two or more proteins encoded by overlapping reading frames on a single mRNA (Dinman and Berry 2007) in many RNA viruses such as retroviruses, coronaviruses, yeast, plant virus, and even bacteria (Jacks and Varmus 1985; Brierley et al. 1991; Chamorro et al. 1992; Tzeng et al. 1992; ten Dam et al. 1994; Kang and Tinoco 1997; Jacobs et al. 2007). Two cis-acting elements are required to regulate −1 FS. One is a slippery sequence where ribosome-associated tRNAs slip, and the other is RNA secondary structure such as a hairpin or pseudoknot that promotes ribosome pausing. Thermodynamic or kinetic control of RNA secondary structure folding may be important in regulating the efficiency of −1 FS. Human immunodeficiency virus type 1 (HIV-1) utilizes −1 FS to regulate the expression ratio of Gag to Gag-Pol, which is critical for the production of infectious virion particles (Paulus et al. 1999). The RNA stem-loop sequence that is involved in −1 FS of HIV-1 is highly conserved in the main subtypes of HIV-1 (Gareiss and Miller 2009). Mutations of this sequence reduce −1 FS and decrease viral infectivity and replication (Baril et al. 2003; Dulude et al. 2006). In severe acute respiratory syndrome coronavirus (SARS-CoV), replicase genes mainly encode two large replicative polyproteins (pp1a and pp1ab) which are expressed by two partially overlapped open reading frames ORF 1a and ORF 1b. As ORF 1b has no independent translation initiation site, polyprotein pp1ab encoded by ORF 1b is only translated as a fused protein form with ORF 1a through −1 FS. As pp1ab includes RNA-dependent RNA polymerase (RdRp) and other replication components which are important proteins for viral replication, −1 FS is essential for the synthesis of enzymatic proteins. The stability of the RNA pseudoknot that induces −1 FS in the SARS-CoV (SARS-pseudoknot) also has a dramatic effect on −1 FS efficiency. Therefore, the RNA secondary structure in the −1 FS site has emerged as an attractive target for drug development (Baranov et al 2005; Plant et al 2005; Su et al. 2005).

Park first conducted virtual screening against the RNA pseudoknot in the −1 FS site to discover ligands that change −1 FS efficiency (Park et al. 2008). In this pilot study, they used the −1 FS system containing biotin aptamer RNA pseudoknot as the RNA secondary structure element. Biotin RNA aptamer was the only ligand-bound RNA pseudoknot structure (PDB id: 1F27) determined by X-crystallography (Nix et al. 2000). Park et al. used the conventional 2D and 3D pharmacophore search program Unity (Martin 1992) for primary database filtering and the FlexX docking program for final docking screening. RNA flexibility was not considered and ligand flexibility was given during FlexX run. Out of about 80,000 compounds in the chemical DB, they obtained 37 hits which increased −1 FS. Compound h4 showed the highest activity in the in vitro transcription and translation coupled assay (Fig. 12.1). The FlexX-docked pose of h4 is shown in Fig. 12.2. The docking mode of h4 is similar to that of biotin; however, h4 forms a stronger interaction with the receptor RNA (Fig. 12.2a). Compound h4 forms hydrogen bonds with O4′ and the 2-carbonyl oxygen of uracil ring of U7 which is one of the critical residues for interaction with biotin in the X-ray structure, and an additional hydrogen bond with ribose O2′ atom of A16 (Fig. 12.2b). These interactions may alter the stability of the RNA pseudoknot and increase the stalled time of the ribosome on the slippery site, thus increasing the rates of −1 FS.

Fig. 12.1.

−1 frameshifting efficiencies induced by biotin aptamer RNA pseudoknot in the presence of compounds (250 μM) determined by SDS–PAGE (a) and dual luciferase assay (b). −1 FS % values are shown at the bottom of the autoradiogram of SDS–PAGE. Each −1 FS % value from dual luciferase assay is the average of triplicate experiments

Fig. 12.2.

(a) Overlay of FlexX-docked pose of h4 and X-ray pose of biotin in the aptamer RNA pseudoknot. The ligand is rendered in capped stick. Carbon atoms of h4 are magenta and those of biotin are white. Cyan lined-ribbon represents the backbone of RNA. (b) Docked model of h4 in complex with biotin-pseudoknot complex. The residues in the active site are rendered in stick. Carbon atoms of pseudoknot are green, oxygen red, nitrogen blue, and phosphorus orange. Yellow dashed lines are hydrogen bonds

Based on these successful results, Park applied the virtual screening strategy to discover ligands for the SARS-pseudoknot, even though its 3D structure was not completely determined. It was known that the SARS-pseudoknot has a unique 3 stem-3 loop structure. They built a 3D model using the RNA pseudoknot predicting program (PSEUDOVIEWER) and Sybyl molecular modeling software, and then optimized the RNA structure by AMBER molecular dynamics simulation (Case et al. 2005) (Fig. 12.3). Using the DOCK 4.0 program, flexible docking of a commercially-available chemical DB (Leadquest) was conducted. The chemical DB was the same as that previously used for screening against the biotin aptamer RNA pseudoknot. A total of 35 compounds inhibited −1 FS, and three structurally analogous compounds (43, 21, and 10 in Fig. 12.4) reduced −1 FS selectively. Compound 43 was the most active, decreasing −1 FS efficiency by 80%. In HEK 293 cells, 43 inhibited −1 FS in a concentration-dependent manner with an IC₅₀ value of approximately 0.45 μM (Park et al. 2011). The docking model of 43 in complex with the SARS-pseudoknot is shown in Fig. 12.5, and reveals that 43 interacts with various residues including key residues to maintain −1 frameshifting efficiency by hydrogen bonds. Compound 43 interacts with the carbonyl oxygen atom (O₂) of the U58 uracil base by hydrogen bond, and the nitrogen atom in the thiazole ring of 43 forms a hydrogen bond (2.8 Å) with the 2′-OH group of ribose of C62. The hydrogen bond between the thiazole moiety and the receptor pseudoknot was identified as one of the key intermolecular interactions.

Fig. 12.3.

(a) Two-dimensional model of SARS-pseudoknot generated by the PSEUDOVIEWER program. (b) Three-dimensional structural model of the SARS-pseudoknot used in this study. It was optimized by molecular dynamics simulation using the Amber 8.0 program. Brown ribbon renders the phosphate backbone of the RNA pseudoknot

Fig. 12.4.

Measurements of −1 FS efficiencies by in vitro TNT assay. (a) The −1 FS efficiencies (%) in the presence of 43, 21, and 10 obtained from SDS-PAGE analysis. The nonframeshifting product (NRF) is the renilla luciferase protein, and the frameshifting product (RF) is a firefly luciferase-renilla luciferase fusion protein. (b) The −1 FS efficiencies obtained from dual luciferase assays

Fig. 12.5.

Docked model of 43: SARS-psuedoknot complex generated by DOCK 4.0. Several residues in the binding site are rendered in capped stick (brown carbon), 43 in ball and stick (green carbon), and red dashed lines indicate hydrogen bonds

Case 3: Riboswitches (Metabolite-Sensing mRNAs)

Riboswitches are highly organized domains within 5′-UTRs of mRNAs and undergo alternate conformational switches. Riboswitches consist of two domains, an aptamer domain that is a binding site for an effector metabolite, and an expression platform that prompts changes in gene expression. Upon metabolite binding, one of the riboswitch conformers is stabilized, and is capable of controlling expression of a downstream gene either at the transcriptional or translational level (Miranda-Ríos et al. 2001; Mironov et al. 2002; Gilbert and Batey 2005) in bacteria (Winkler et al. 2002; Mandal et al. 2003) as well as in some plants and fungi (Sudarsan et al. 2003; Bocobza and Aharoni 2008). Therefore, the aptamer domain is an essential element of gene regulation. Riboswitches were originally discovered as an antibiotic target (Blount and Breaker 2006; Lee et al. 2009; Kim et al. 2009) and also have the potential to be developed into designer riboswitches for genetics studies (Suess and Weigand 2008). About 20 classes of riboswitches have been reported and a large number (probably up to 100) of new classes may await discovery. Among them, guanine riboswitches are one of the validated targets for development of new antibacterial drugs. X-ray crystal structures of more than 10 riboswitches have been determined, and thus structure-based drug design approaches can be applied to this target (Serganov 2010). Daldrop et al. tried virtual screening to discover novel ligands targeting the purine riboswitch using the program DOCK3.5.54 with minor modification of the scoring function (Daldrop et al. 2011). The Bacillus subtilis xpt-pbuX guanine riboswitch carrying a C74U mutation (called GRA, PDB id: 2G9C) (Gilbert et al. 2006) in complex with pyrimidine-2,4,6-triamine was used as a target receptor for docking.

Validation of Docking Program

All docking and virtual screening studies were carried out using a slightly modified version of the DOCK3.5.54 program that incorporated RNA-specific parameters to calculate van der Waals and electrostatic energies. This group first tested whether native ligand was correctly reproduced in terms of its binding geometry after self-docking. The RMSD between native ligand and the docking result was measured. RMSD was less than 0.34 Å, which showed that the docked model was close to native. Secondly, the accuracy of the prediction was tested by docking with known ligands and decoys. The test set consisted of eight known ligands, with binding affinities ranged from 0.01 to 100 μM. For all 15 decoys, no binding was detected at up to 300 μM; except for guanine that was tested up to its solubility limit. All compounds in the test set docked into the active site and the results were analyzed by sorting the docking scores. Seven out of the eight top-scoring compounds were true ligands and all eight compounds with lowest docking energy scores were decoys.

Virtual Screening

They used their own in-house database, which included a commercially available 2,592 unique compounds. To evaluate the accuracy of binding prediction, all compounds in the database were docked into the active site and ordered by docking score. According to this list, the true positive rate (fraction of known compounds, ligands, and decoy) was plotted against the false positive (fraction of unassigned database compounds) to get a receiver operation characteristic (ROC) curve. Accuracy was calculated by measuring the area under curve (AUC) of a ROC curve value. The AUC value of ligands was 0.98 and that of decoy was 0.75, suggesting that known ligands were perfectly predicted by virtual screening and decoys were also enriched compared to random (AUC 0.5). From this result, five compounds were selected to examine their binding affinity and modes. Three compounds out of five were analogs of known ligands and two compounds were novel scaffolds. To determine the binding affinities, fluorescence assays and isothermal titration calorimetry (ITC) were used. Four out of five chosen compounds bound to GRA with affinities in the micromolar range.

Concluding Remarks

The rapidly increasing number of RNA crystal structures in the PDB, and the biological function studies on a variety of RNA structures have provided a basis for structure-based virtual screening targeting RNA. Due to several features of RNA (conformational flexibility, high negative charge, and solvation) that differ from those of proteins, researchers have observed that the conventional protein-ligand docking programs have limitations in accurately predicting RNA-ligand interactions. In some studies, active molecules have been fortuitously obtained by protein-targeted docking programs. However, continuing progress has been made in the development of docking algorithms or scoring functions optimized for RNA receptors, which makes various RNA targets more amenable for structure-based drug design. As an example, consideration of receptor flexibility is important not only for RNA-based receptors, but has also been a critical issue for protein receptors for a long time. Massive efforts have been made to incorporate protein flexibility into docking process, but it exponentially inflates the potential search space and became impractical. Al-Hashimi’s approaches to use RNA dynamic ensemble (Stelzer et al. 2011) can be practically applied to other RNA targets. RNA receptors are usually smaller than protein receptors in size, so the process to generate RNA ensembles is not extremely computationally expensive.

Structure- or ligand-based virtual screening has identified a plethora of RNA binding ligands from in-house and commercially available chemical databases, originally designed and prepared for protein targets. Actually, some commercial chemical databases possess compounds which only satisfy “Lipinski’s rule of five (Lipinski et al. 1997)”. In comparing the activities of small molecule RNA binders identified by virtual screening, we realized that hits with nanomolar activity were very rare, and overall their activities are in the high micromolar range. These results were probably caused by limitations of currently available chemical database, not only by those of protein-friendly docking algorithms. In protein-targeted virtual screening of chemical databases, many hits with nanomolar or submicromolar activity were identified. Thus, another possible way to improve the hit rate of RNA-targeted virtual screening is to prepare an RNA-focused chemical database. Considering the physicochemical properties of RNA binders, a modified “Lipinski’s rule of five” needs to be applied to database filtering during virtual screening (Aboul-ela 2010).