Discovery of RNA-targeted small molecules through the merging of experimental and computational technologies

ABSTRACT

Introduction

The field of RNA-targeted small molecules is rapidly evolving, owing to the advances in experimental and computational technologies. With the identification of several bioactive small molecules that target RNA, including the FDA-approved risdiplam, the biopharmaceutical industry is gaining confidence in the field. This review, based on the literature obtained from PubMed, aims to disseminate information about the various technologies developed for targeting RNA with small molecules and propose areas for improvement to develop drugs more efficiently, particularly those linked to diseases with unmet medical needs.

Areas covered

The technologies for the identification of RNA targets, screening of chemical libraries against RNA, assessing the bioactivity and target engagement of the hit compounds, structure determination, and hit-to-lead optimization are reviewed. Along with the description of the technologies, their strengths, limitations, and examples of how they can impact drug discovery are provided.

Expert opinion

Many existing technologies employed for protein targets have been repurposed for use in the discovery of RNA-targeted small molecules. In addition, technologies tailored for RNA targets have been developed. Nevertheless, more improvements are necessary, such as artificial intelligence to dissect important RNA structures and RNA–small-molecule interactions and more powerful chemical probing and structure prediction techniques.

KEYWORDS:

• RNA-targeted small molecules
• RNA structure
• target identification
• high-throughput screening
• hit validation
• target engagement
• three-dimensional structure determination
• hit-to-lead optimization

1. Introduction

Historically, drug discovery programs have focused on developing small molecules that target proteins to alter the outcome of biological processes. However, according to a recent survey, less than 1% of the total number of human proteins remain as viable new targets for small-molecule drugs [1]. The survey noted that a large portion of disease-related proteins (roughly 60%) are considered ‘undruggable’ because they lack distinctive clefts or pockets capable of binding small-molecule drugs. Given this diminishing pool of druggable proteins, many are looking into the possibility of targeting RNA with small molecules.

Figure 1. Why target RNA with small molecules? (a) RNAs perform a myriad of functions, which are often mediated by structurally conserved motifs. The HIV-1 transactivation response (TAR), severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) stem loop 5 (SL5), and the flavin mononucleotide (FMN) riboswitch, whose secondary structures are shown, are a few examples of functional RNA motifs that are structurally conserved. (b) Many bioactive small molecules that act on RNA targets have now been identified. The equilibrium dissociation constant (K_d), 1.5 maximal effective concentration (EC_1.5) or median inhibitory concentration (IC₅₀) are provided. (c) Like the proteome, the transcriptome is likely to contain many druggable sites [2], as the ones depicted in red spheres. The druggable sites on the FMN riboswitch (PDB ID: 3 f4g) [3], add adenine riboswitch (PDB ID: 5swe) [4], simian retrovirus type-1 (SRV-1) pseudoknot (PDB ID: 1e95) [5], enterovirus 71 internal ribosomal entry sites (IRES) (PDB ID: 6xb7) [6], HIV-1 TAR (PDB ID: 1qd3) [7], UGGAA/UGGAA pentad in spinocerebellar ataxia type 31 (SCA31) [8] were calculated using Fpocket [9].

RNA molecules were previously viewed as mere messengers between DNA and protein synthesis. Today, we know that RNA has catalytic and regulatory functions in diverse biological events (Figure 1(a)). As such, RNA presents boundless opportunities for intervention with small molecules. For example, small molecules that target mRNAs could control the production of disease-related proteins, even those that were heretofore deemed undruggable. In addition, the roles of non-coding RNAs in disease are gradually being substantiated [10]. Since non-coding RNAs occupy a significant fraction of the human genome [11], targeting them would broaden the reach of small-molecule drugs.

The concept of targeting RNAs with small molecules is not a novel one. In the late 1980s, it was discovered that the target of aminoglycoside antibiotics are bacterial ribosomal RNAs [12]. Thereafter, aminoglycosides were revealed to promiscuously bind RNA [13,14], and while some viewed this promiscuity as an opportunity to optimize aminoglycosides for targeting other RNAs, many opted to divert to more drug-like small molecules. Thus far, more than 200 bioactive small molecules were reported to bind RNA molecules with diverse structural motifs (Figure 1(b)) [15,16]. The RNA binders can be classified into two: (1) ‘traditional’ small molecules with properties satisfying Lipinski’s rules (MW ~ 500 Da) and (2) larger, multivalent ligands (~2000 Da). The consensus is that drug-like small molecules would bind structurally complex RNAs with protein-like binding pockets, while multivalent ligands would bind low-complexity RNAs (Figure 1(c)) [17,18].

This review will focus on the early stages of RNA-targeted small-molecule drug discovery and describe the experimental and computational tools that can be used in each stage. First, the strategies for identifying potentially functional and druggable RNA motifs will be delineated. Next, several high-throughput screening (HTS) approaches and biophysical techniques for hit validation will be highlighted, and the factors to consider for each of them will be discussed. Then, the methods for assessing the functional effects and target engagement of RNA-binding small molecules will be described. Finally, an overview of the techniques for determining the three-dimensional (3D) structures of RNA–small-molecule complexes, along with ways by which such structural information can be exploited to accelerate the drug discovery process, will be given. The references for this review were obtained through searches of the PubMed database (www.pubmed.gov) through July 2022.

2. Identification of RNA targets

Most efforts to discover RNA-targeted small molecules have been directed against already-known functional RNA motifs, such as HIV-1 transactivation response (TAR) [19], the flavin mononucleotide (FMN) riboswitch [20–22], the severe acute respiratory syndrome coronavirus frameshifting pseudoknot [23], viral internal ribosomal entry sites (IRES) [8,24], and disease-associated RNA repeat sequences [25,26]. However, there remains a plethora of therapeutically valuable RNA motifs in the transcriptome waiting to be unveiled. The identification of target motifs on RNA molecules can be approached using the following strategy: (1) extraction of motifs on an RNA molecule, (2) analysis of RNA secondary structures, and (3) assessment of RNA motifs to select the most suitable target(s).

2.1. Extraction of RNA motifs

Motifs on an RNA can be extracted using secondary structure prediction (Table 1) and chemical-probing approaches. The mainstream method in RNA secondary structure prediction uses dynamic programming to calculate an optimal secondary structure with the minimum free energy (MFE) among all possible outcomes [39,57]. The free energy is typically estimated from the thermodynamic free energies for the base-pairing stems and the conformational entropies for the unpaired loop regions [58,59]. The accuracy of secondary structure predictions using free energy minimization can vary based on the length of the RNA; for RNA sequences shorter than 700 nucleotides, the prediction accuracy is approximately 70%, but for longer RNAs, the accuracy decreases [60].

Table 1. List of programs for RNA secondary structure prediction.

Program	Approach	Description	References
CentroidFold	Knowledge based	Based on generalized centroid estimator	[27,28]
CentroidHomfold	Knowledge based	Employs automatically collected homologous sequence information	[29]
Context Fold	Machine learning	Based on feature-rich trained scoring models	[30]
CONTRAfold	Machine learning	Based on conditional log-linear models, a flexible class of probabilistic models, which generalize upon stochastic context-free grammars by using discriminative training and feature-rich scoring	[31]
Crumple	Knowledge based, ring graph approach	Generates all possible non-pseudoknotted folds for an RNA sequence, given optional constraints	[32]
CyloFold	Free energy minimization	Based on simulating a folding process in a coarse-grained manner by choosing helices based on established energy rules; Allows complex pseudoknots	[33]
GTFold	Free energy minimization	Fast and scalable RNA secondary structure prediction by thermodynamic optimization for modern multi-core computers	[34]
IPknot	Competitive prediction	Predicts RNA secondary structures with pseudoknots based on maximizing expected accuracy of a predicted structure; Also gives consensus secondary structure with pseudoknots when a multiple sequence alignment is given	[35]
KineFold	Exact stochastic clustering	Predicts RNA secondary structures including pseudoknots through long-time-scale RNA folding simulations, which follow the stochastic closing and opening of individual RNA helices	[36,37]
MC-Fold	Knowledge based	Based on the nucleotide cyclic motif, a first-order object representing nucleotide relationships in structured RNAs; Incorporates non-Watson–Crick base pairs and suites of such into a unified energetic framework	[38]
Mfold	Free energy minimization	Oldest and most well-known software; Determines the set of base pairs that gives the minimum free energy using dynamic programming	[39]
Pknots and Pknots RG	Free energy minimization	Generates the optimal minimum energy structure using standard RNA folding thermodynamic parameters and a few parameters for pseudoknots	[40,41]
RNAfold	Free energy minimization, circular folding	Predicts the minimum free energy structure using thermodynamic parameters; Adapts experimental constraints	[42,43]
RNAshapes	Abstract shape analysis	Searches the folding space using RNA abstract shape analysis and computes the probabilities of structures that share the same shape	[44,45]
RNAstructure	Free energy minimization	Predicts the minimum free energy structure and maximum expected accuracy structures, including pseudoknots; Adapts experimental constraints	[46,47]
SARNA-Predict	Simulated annealing	Predicts RNA secondary structures including pseudoknots based on simulated annealing	[48]
Sfold	Statistical sampling	Based on a sampling algorithm that incorporates comprehensive structural features and thermodynamic parameters	[49]
Sliding Windows & Assembly	Sliding window	Generates all possible helix structures in windows containing short pieces of the RNA sequence and assembles the helices using a dynamic programming algorithm	[50]
SPOT-RNA	Machine learning	Uses an ensemble of two-dimensional deep neural networks and transfer learning for base-pair prediction including noncanonical and non-nested base pairs stabilized by tertiary interactions	[51]
SwiSpot	Switching sequence	Predicts alternative secondary structures of riboswitches	[52]
UNAFold	Free energy minimization	Integrated collection of programs that simulate folding, hybridization and melting pathways	[53]
vsfold and vs subopt	Heuristic modeling with mapping and sequential folding	vsfold: Predicts RNA secondary structure including pseudoknots using an entropy model derived from polymer physics; vs_subopt: Computes suboptimal structures based on the free energy landscape derived from vsfold	[54–56]

Another common method for RNA secondary structure prediction is comparative sequence analysis. It is generally believed that RNA motifs whose biological functions are dictated by structure show high structural conservation when compared to the same RNA of related species [61]. Thus, in comparative sequence analysis, base-pairing stems and loops are identified based on their occurrence in identical or similar locations on homologous RNAs. While comparative sequence analysis is exceptionally accurate at predicting secondary structures (>95% accuracy), it may not be applicable to all RNAs and requires careful multiple sequence alignments [62].

The most widely used chemical probing methods employ dimethyl sulfate (DMS), which modifies the bases of non-Watson–Crick paired or single-stranded nucleotides [63], and selective 2’-hydroxyl acylation analyzed by primer extension (SHAPE) reagents, which modifies the ribose groups of dynamic nucleotides [64,65]. The results of DMS and SHAPE chemical probing are read out using primer extension-truncation and adapter-ligation to create libraries to be analyzed by massive parallel sequencing. However, the reactivities obtained from these methods may be inaccurate due to: (1) low signal levels, (2) biases introduced by the truncation–adapter-ligation steps [66,67], (3) the presence of naturally occurring modified nucleotides and stable secondary structures that result in unwanted truncation, and (4) improper statistical analysis. Thus, DMS mutational profiling with sequencing (DMS-MaPseq) [68] and SHAPE-MaP (Figure 2(a)) were developed to overcome these limitations.

Figure 2. Methods for identification of RNA targets. (a) SHAPE-MaP can be used to probe RNA secondary structures on a massively parallel scale. (b) MobyDick enables the analysis of RNA secondary structure predictions. When the analysis is completed, the MobyDick interface displays the locations, secondary structure, global population (P_Global), and stability (∆G) of the motifs.

In the MaP strategies, the modified nucleotides are detected by mutational profiling. MaP takes advantage of conditions where reverse transcriptases misread modified nucleotides and incorporate non-complementary nucleotides into the original sequence in the newly generated cDNA. These cDNAs are amplified using methods that are easier to execute and produce less bias compared to truncation–adapter-ligation strategies [69,70]. Then, the amplified cDNA sequences are subjected to massively parallel sequencing, and the MaP data are analyzed [71] to produce a reactivity profile.

The chemical reactivities from any of the above probing experiments can be incorporated as constraints in several RNA secondary structure prediction software, such as RNAstructure [46,47] and RNAfold in the Vienna RNA package [42,43]. Mathews et al. reported an average 9% increase in accuracy (from 67% to 76%) of secondary structure prediction with the incorporation of chemical reactivity constraints [34].

2.2. Analysis of RNA secondary structures

After extracting the RNA motifs, the next step would be analyzing the RNA secondary structures to narrow down the list of candidate target RNA motifs. We created a software, MobyDick, which is a part of our ibVIS platform technology, for analyzing the results of secondary structure predictions from any software using a scanning window (Figure 2(b)) [72]. First, the results of the secondary structure prediction are input into MobyDick, and the set of predicted secondary structures for each user-defined window, moved at a user-defined resolution from the 5’- to the 3’-end of the RNA sequence, are gathered. Next, the population of each motif found within the predicted structures for every window (local population or $P_{\mathrm{L}\mathrm{o}\mathrm{c}\mathrm{a}\mathrm{l}}$) is calculated using Maxwell-Boltzmann statistics. Finally, the motif’s global population ($P_{\mathrm{G}\mathrm{l}\mathrm{o}\mathrm{b}\mathrm{a}\mathrm{l}}$) value, which indicates the probability that the motif exists, is generated using the following equation:

$P_{\text{Global}}={\displaystyle \sum _{n=0}^{N-1}\text{W}}\left(n\right)\cdot P_{\text{Local}}\left(n\right)$

where $W\left(n\right)$ is the weight function for the window $n$. Assuming that the ribosomal speed along the mRNA is constant, $W\left(n\right)$ is equal to $\frac{1}{N}$. To date, we have used MobyDick for the analysis of roughly 100 transcripts, and motifs with $P_{\mathrm{G}\mathrm{l}\mathrm{o}\mathrm{b}\mathrm{a}\mathrm{l}}$ values greater than 50% showed high agreement with experimentally derived secondary structures.

The Moss group recently developed ScanFold [73], another RNA secondary structure analytic tool that uses the scanning window approach. ScanFold gives a z-score, which compares the MFE of a sequence within an RNA to the average MFE of a set of randomized RNA sequences normalized by the standard deviation of the MFE. Motifs with more negative z-scores (< −2) are considered very stable and likely to perform a biological function. ScanFold and MobyDick differ not only in terms of the kind of output they provide (i.e. z-score versus P_Global) but also in that the former utilizes only RNAFold while the latter can accept the results any secondary structure prediction software. This confers MobyDick an advantage in that the use of several software in parallel can compensate for the computational bias of each software and increase one’s confidence in the motifs predicted should they be identified by multiple software.

2.3. Target RNA motif selection

Once the number of candidate target RNA motifs is pared down to the most promising, one can assess individual motifs more thoroughly. The location of the motif on an mRNA or pre-mRNA provides clues to its function. In general, stable motifs on the 5’ untranslated region (UTR) may regulate translation [74–79], those on the junctions between the exons and introns may be involved splicing [80–82], and those on the 3’ UTR may control degradation and stabilization processes [83–86]. From a survey of the complete genomes of 340 species, Gu et al. [87] posited that stable RNA motifs near the start codon regulate translation by interfering with start codon recognition. It was also reported that highly stable motifs on the coding region may stall the ribosome and activate the ribosome-associated protein pathway, which targets newly synthesized proteins for proteasomal degradation [88]. As a rule of thumb, we select stable RNA motifs on the 5’ UTR and the coding region of the mRNA and splicing relevant RNA motifs on exon-intron junctions of the pre-mRNA as targets of small-molecule drug discovery. It is also worth considering motifs on the 3’ UTR and non-coding RNAs after verifying their therapeutic implications with experimental and computational approaches. In addition, we select RNA motifs that bind proteins, which are components of cellular metabolism or regulatory networks, as suggested by Warner et al. [17].

3. High-throughput screening of compounds that bind to RNA motifs

Several HTS approaches have been used to identify RNA-binding small molecules, and they can be characterized as target-based or phenotypic (Figure 3, Table 2). Target-based screening approaches usually employ biophysical techniques with the goal of identifying small molecules that bind to a known target. In contrast, in phenotypic screening, the target is unknown, and the goal is to screen for compounds that alters the phenotype of a cell or organism.

Figure 3. High-throughput screening approaches for identifying small molecule compounds that bind to RNA. (a) In fluorescence-based screening, a fluorescently labeled molecule – either the RNA target or an indicator – is used. Fluorescence resonance energy transfer (FRET) employs an RNA probe that is labeled with a fluorophore and a quencher. The FRET RNA probe is combined with the screening compounds and subjected to heating. The melting temperatures (Tm) of the RNA are measured using qPCR. Compounds that increase the Tm of the RNA bind to and stabilize the RNA. (b) Affinity-based screening methods do not require labeling of the RNA target. The automated ligand identification system (ALIS) is an indirect affinity-based screening method that involves the equilibration of the target RNA with a mixture of the screening compounds, size exclusion chromatography to purify the RNA–small molecule complexes, reverse-phase HPLC to dissociate the small molecules from the RNA, and mass spectrometry to identify the small molecule hits. (c) In small molecule microarray screening, the small molecules are immobilized onto microarray plates and used for screening against RNA targets. The hits are analyzed, e.g. by detecting increase in fluorescence if the RNA target is fluorescently labeled. (d) Fragment-based screening employs biophysical techniques, usually NMR, to identify low molecular weight compounds that bind to RNA targets. In ¹⁹F-NMR-based fragment screening, the 1D NMR spectra of mixtures of fluorinated compounds, before and after the addition of the RNA target, are taken. The hits are identified by looking for changes in the NMR signals of the fluorinated compounds; RNA-binding compounds exhibit more pronounced NMR signals after the addition of RNA. (e) In a typical DNA-encoded library (DEL) screening against RNA targets, the RNA is immobilized and incubated with small molecule compounds conjugated to a DNA barcode. The DEL members that bind to the RNA are eluted, and their identities are revealed after DNA amplification and sequencing. (f) Docking is the most widely used virtual screening approach. It requires that the three-dimensional structure of the RNA target be known so that the binding energies between the RNA and the screening compounds can be predicted in silico. (g) Most of the well-known RNA-targeted small molecules have been identified using phenotypic screening. In this method, the small molecules are evaluated for their ability to produce the desired phenotype within a cellular environment. The identity of RNA target is only revealed after molecular target identification and mechanistic studies.

Table 2. Selected screens conducted against RNA targets*.

Screening method	Target	Library size	Hit rate (%)	Affinity/potency	References
Fluorescence-based screening
FRET	HIV-1 Psi Stem Loop 3 (SL3)	2,643	3.0	Hit: IC50 = 4.5 μM (NSC260594)	[89]
FRET	TERRA	56	N/A	Hit: KD = 2.56 μM (CK1-14)	[90]
FRET (modified)	Group II Intron	10,000	0.16	Lead: Ki = 360 nM (intronistat B)	[91]
FRET	C9orf72 (G4C2)4 G-quadruplex	138	8.7	Hit: KD = 200–400 nM	[92]
TR-FRET	(CUG)12 RNA–MBNL1	320	9.0	Hit: IC50 = ~30–~150 μM	[93]
FID	C9orf72 (G4C2) repeat	132	2.3	Lead: Kd = 9.7–16 μM	[94]
FID	SMN2 exon 7 splice site stem-loop 2	304	54	Hit: EC50 = ~25 μM (PK4C9)	[95]
FID	C9orf72 (G4C2) repeat	40	25	Hit: IC50 = 0.9–12 μM	[96]
FID	EV71 IRES SLII	27	7.4	Hit: KD = 520 nM (DMA-135)	[6]
Affinity-based screening
ESI-MS	HCV IRES IIA	180,000	N/A	Hit: KD = ~100 μM (benzimidazole 1)	[97]
ALIS	FMN riboswitch	~53,000	0.12	Hit: KD = 130 nM (WG-3)	[21]
ALIS	Xist non-coding RNA	50,000	N/A	Hit: KD = 0.4 μM (X1)	[98]
Small molecule microarray (SMM)
SMM	HIV-1 TAR RNA	20,000	0.02	Hit: KD = 2.4 μM (analogue)	[99]
SMM	MALAT1 triple helix-forming sequence	26,229	0.7	Hit: KD = 2.3–6.1 μM	[100]
2DCS	HCV stem loops I–III	1,987	12	Hit: KD = ~400–~700 nM	[101]
Fragment-based screening
^1H NMR	Influenza A virus promoter	4,279	0.2	Hit: KD = 50.5 μM (DPQ)	[102]
^19F NMR	TERRA	355	5.6	Hit: KD = 120–1900 μM	[103]
DEL screening
DEL+2DCS	4,096 targets including pre-miRNA-27a	73,728	0.01	Hit: KD = 90 nM (compound 9, MST)	[104]
DEL	FMN riboswitch	46.3 billion	N/A	Hit: KD = 11.38 nM (HGC-1, 15.03 nM (HGC-2)	[105]
Virtual screening
Docking	HIV-1 TAR	181,000	N/A	Hit: KD<1 μM (EMSA)	[106]
Docking	Guanine riboswitch	2,615	N/A	Hit: KD = 80 μM (compound 27)	[107]
Docking	SARS-CoV Pseudoknot	80,000	N/A	Hit: IC50 = 0.45 μM (compound 43)	[23]
NMR+docking	HIV-1 TAR	~51,000	N/A	Hit: KD = 55 nM–122 μM	[108]
Docking	HIV-1 Rev response element stem loop IIB	N/A	N/A	Hit: IC50 = 6.8 μM (compound 6b)	[109]
Docking	DMPK CUG repeats	5	60	Hit: KD = 430 nM	[110]
Docking	DMPK CUG repeats	23	35	Hit: reduced nuclear foci at 40 μM concentration (compound 2–5)	[111]
Docking	HOTAIR	>200,000	N/A	N/A	[112]
Phenotypic screening
Luciferase reporter assay	SMN2	~200,000	1.0	Lead: EC1.5x = 4 nM (risdiplam)	[113,114]
Luciferase reporter assay	SMN2	1,400,000	<1.0	Hit: EC50 = 20 nM (branaplam)	[115]
Whole-cell phenotypic screening	FMN riboswitch	~57,000	N/A	Hit: KD = 16 nM (ribocil)	[20]

*Abbreviations used: ALIS, automated ligand identification system; 2DCS, two-dimensional combinatorial screening; DEL, DNA-encoded library; DMPK, Dystrophia myotonica protein kinase, DPQ, 6,7-dimethoxy-2-(1-piperazinyl)-4-quinazolinamine; EMSA, electrophoretic mobility shift assay; ESI-MS, electrospray ionization mass spectrometry; EV71 IRES SLII, enterovirus 71 internal ribosome entry site stem loop II; FID, fluorescence indicator displacement; FMN, flavin mononucleotide; FRET, fluorescence resonance energy transfer; HCV, hepatitis C virus; HIV-1 TAR, human immunodeficiency virus type 1 transactivation response; HOTAIR HOX antisense intergenic RNA; K_D: equilibrium dissociation constant; MALAT1, Metastasis-Associated Lung Adenocarcinoma Transcript 1 (MALAT1), MBNL1, muscleblind-like splicing regulator 1; NMR, nuclear magnetic resonance; pre-miRNA, precursor microRNA; SARS-CoV, severe acute respiratory syndrome coronavirus; SMM, small molecule microarray; SMN2, survival of motor neuron 2; TERRA, telomeric repeat-containing RNA; TR-FRET, time-resolved fluorescence resonance energy transfer

Some measures that are prerequisite to successful HTS campaigns include: (1) the rigorous design and development of the assay, (2) the use of positive and negative controls for checking the quality of the assay, and (3) the use of compound libraries with wide chemical diversity. In addition, the merits of a properly designed and executed HTS assay would inevitably be wasted if appropriate statistical methods for hit selection are not utilized. Furthermore, the statistical methods for analysis of HTS data must be adapted to the characteristics of the HTS approach and to the responses observed.

Regardless of whether target-based or phenotypic screening approaches are used, it is advisable to perform a cascade of experiments to select the specific and high-quality hits from the HTS active compounds. Since HTS is typically done at a single compound concentration, subsequent dose-dependence experiments using a broad concentration range are ideal. The dose–response curves would not only give the IC₅₀ values but also provide information about the compound’s toxicity and solubility. In addition, counter screens can be used to assess the specificity of the HTS active compounds.

3.1. Fluorescence-based screening

The three fluorescence-based assays that are often employed for screening of RNA binders are: (1) fluorescence resonance energy transfer (FRET)-based assay (Figure 3(a)), (2) time-resolved FRET (TR-FRET) assay, and fluorescent indicator displacement (FID). FRET refers to the transfer of energy from a donor fluorophore to an acceptor fluorophore conjugated to the target biomolecule. FRET-based assays are convenient and extremely sensitive and have become popular for screening small-molecule libraries against RNA targets. Simone et al. [92] used a FRET-based assay to screen for small-molecule stabilizers of the C9orf72 (G₄C₂)₄ G-quadruplex RNA, which is a known cause of frontotemporal dementia and amyotrophic lateral sclerosis [116]. The authors monitored the changes in melting temperature of the 5’-FAM and 3’-TAMRA labeled (G₄C₂)₄ RNA upon heating in the presence of small molecules and identified three structurally similar small molecules that stabilize the RNA. The small molecules were subsequently shown to reduce the frequency of RNA foci and the levels of dipeptide repeat protein in C9orf72 patient neurons. Furthermore, the most effective small molecule, DB1273, was found to improve survival and reduce levels of toxic poly-(glycine-arginine) in C9orf72 flies.

The TR-FRET assay combines the low background aspect of time-resolved fluorimetry with the homogenous format of FRET. Thus, TR-FRET provides increased flexibility, reliability, sensitivity, and throughput. TR-FRET was used to identify small molecules that inhibit the formation of the complex between (CUG)_n expansion repeat RNAs with muscleblind-like splicing regulator 1 (MBNL1) [93,117,118] that causes myotonic dystrophy type 1 (DM1). After combining biotinylated (CUG)₁₂ RNA and MBNL1-His₆ protein, fluorescently labeled streptavidin and anti-His₆ antibody were added to the solution. Streptavidin binds to the biotinylated (CUG)₁₂ RNA whereas the anti-His₆ antibody binds to MBNL1. FRET occurred when the complex between (CUG)₁₂ RNA and MBNL1 formed, since the fluorophores are near each other. On the other hand, FRET does not occur in the presence of small-molecule inhibitors since the (CUG)₁₂–MBNL1 interaction is disrupted. The TR-FRET screens identified inhibitors of the (CUG)₁₂ RNA-MBNL1 protein complex exhibiting IC₅₀ values at low micromolar level. One of the inhibitors improved DM1-associated pre-mRNA splicing defects in cell and mouse models of DM1 [118].

FID is an assay that utilizes a fluorescently labeled indicator, which can be displaced upon ligand binding to the target molecule. FID has been used in primary HTS campaigns by many groups [119]. The Hargrove group recently showed FID to be useful in identifying inhibitors of enterovirus 71 (EV71) viral translation and replication [6]. In their study, the authors used a fluorescently labeled, highly basic peptide fragment as an indicator that can be displaced upon binding of small molecules to the EV71 stem loop II (SLII) IRES structure. The strongest binder to EV71 SLII, DMA-135, induces a conformational change in the RNA that stabilizes a ternary structure with the AU- rich element RNA-binding protein, thus inhibiting viral translation.

3.2. Affinity-based screening

A direct method for detecting RNA–small molecule complexes is electrospray ionization mass spectrometry (ESI-MS) [120]. In an ESI-MS assay, the shift in the molecular weight of the unbound and unlabeled target upon the formation of the target–ligand complex is measured. ESI-MS was applied by Seth et al. [97] to search for small molecules that bind to the IRES IIa subdomain of the hepatitis C virus (HCV). The authors identified benzimidazole inhibitors, which reduce viral RNA levels in the HCV replicon at low micromolar concentrations.

In indirect affinity-selection mass spectrometry (AS-MS) methods (Figure 3(b)), small molecules that bind to the target are separated from the target prior to identification. One of the emerging indirect AS-MS assays is the automated ligand identification system (ALIS), which allows for the HTS of small molecules that bind to an unlabeled target. Originally used against protein targets, ALIS was adapted by Merck for use in the screening, ranking, and characterization of RNA binders [21,98,121,122]. In an ALIS assay, the target is first incubated with small-molecule compound mixtures. Then, the target–small-molecule complexes are purified by size exclusion chromatography, and the complexes are dissociated using reverse-phase chromatography. Finally, the bound small molecules are detected by MS. In seminal work by Rizvi et al. [21], structurally diverse small molecules that bind to the FMN riboswitch were identified using the ALIS approach, and the binders induced a change in the riboswitch conformation different from that induced by the cognate FMN ligand. Very recently, Aguilar et al. [98] applied ALIS to the study of the non-coding Xist RNA. They identified the drug-like compound, X1, with an indazole–benzimidazole scaffold to have submicromolar affinity to Xist RNA. By binding to Xist, X1 displaces interacting proteins, suppresses histone H3K27 trimethylation, and blocks initiation of X-chromosome inactivation.

3.3. Small molecule microarray (SMM)

SMM (Figure 3(c)) has proven to be a robust and scalable method for discovering RNA-binding small molecules. In this method, a collection of small molecules is printed onto a glass surface using a robotic microarrayer [123]. Then, the microarray is incubated with a fluorescently labeled target biomolecule and washed to remove unbound target. Next, the microarrays are imaged using a fluorescence scanner, and the fluorescence intensities of each spot on the microarray are quantified. Last, the spots with increased fluorescence upon incubation, representing interactions between the target and the small molecules, are revealed by statistical analysis.

The Schneekloth group used SMM as a platform for identifying RNA binders. In one example, they pooled a library of 20,000 drug-like compounds and incubated the library with Cy5-labeled HIV-1 TAR RNA (target) and a Cy5-labeled precursor of miR-21 hairpin (competitor) [99]. After analysis of the microarray data, they identified a compound that selectively binds to TAR and does not resemble previously known TAR-binding small molecules. Further evaluation demonstrated that one of the analogues shows specificity for only the TAR region, lacks toxicity, and has good antiviral activity.

An ingenious variation on the standard SMM is the two-dimensional combinatorial screen (2DCS) developed by the Disney lab [124]. 2DCS is a library-versus-library approach, wherein a small-molecule library conjugated onto an agarose-coated microarray is hybridized with a γ-³²P-labeled RNA library, along with competitor oligonucleotides. 2DCS has enabled the assessment of millions of RNA–small-molecule interactions and the identification of novel RNA binders. The data generated by 2DCS has been incorporated into Inforna [125,126], which was used to identify binders of precursor microRNA-96. The binders inhibit the biogenesis of the microRNA, upregulate FOXO1, and induce apoptosis in cancer cells [125].

3.4. Fragment-based screening

Fragment-based screening is gaining attention in recent years since it shows promise in developing RNA binders with drug-like properties [127]. Briefly, fragment-based screening employs biophysical techniques to identify low molecular weight compounds. The fragments would have weak but specific interactions, but they may be optimized using a variety of strategies to obtain more potent and selective compounds.

Fragment-based screening was implemented to identify novel ligands for the E. coli thiamine pyrophosphate (TPP) thiM riboswitch [128,129], which is a regulatory element in the non-coding region of mRNA that controls gene expression by sensing TPP. A library of 1300 fragments was screened against the thiM riboswitch using a combination of equilibrium dialysis, NMR spectroscopy, and isothermal titration calorimetry. Structurally diverse fragments with equilibrium dissociation constant (K_d) values ranging from 22 to 670 µM were identified. The hit fragments were validated using X-ray crystallography, small-angle X-ray scattering, and SHAPE, which showed that the fragments compete with TPP but rearranges the riboswitch into a structure different from that of the cognate complex [130].

Garavís et al. [103] used ¹⁹F-NMR-based fragment screening (Figure 3(d)) to identify small molecules that bind to telomeric repeat-containing RNAs (TERRAs), which fold into G-quadruplexes. TERRA G-quadruplexes act as scaffolds for the formation of telomeric heterochromatin in cancer cells and are thus ideal therapeutic targets [131]. The screen identified 20 hit fragments from a 355-member fluorinated fragment library. Examination of the interactions of several of the hits suggested that the fragments favor the parallel propeller-like conformation of telomeric sequences.

3.5. DNA-encoded library (DEL) screening

DEL screening offers several advantages including the ability to screen extensive libraries, low cost, simplicity, and speed [132]. A DEL is a collection of small-molecule compounds in which each compound is bound to a segment of synthetic DNA (Figure 3(e)) [133]. Each DNA segment serves as a barcode that contains information about the structure of the small molecule attached to it. DEL screening is an affinity-based process, which involves (1) incubation of the DEL compounds with the target, (2) removal of the weak or nonspecific binders, (3) sequencing of the DNA codes of the isolated binders, and (4) data analysis of the resulting DNA sequences to reveal the identities of the binders.

The Disney group recently reported the use of a strategy that integrates DEL with DCS to probe binding between RNAs and small molecules [104]. They screened a 73,728-member DEL against 4,096 RNAs, thereby evaluating a total of ~300 million interactions. Using fluorescence-activated cell sorting, they identified both the DEL compounds that specifically bind to the RNAs and the selective RNAs. One of the notable findings of the study is a small molecule that binds with nanomolar affinity for primary microRNA-27a, inhibits miRNA biogenesis, and rescues a migratory phenotype in triple-negative breast cancer cells.

The biggest issue with DEL screening against RNA targets is that it could result in false positives due to the nonspecific binding of the DNA tags to the RNAs. In a recent report, HitGen circumvented this issue by utilizing RNA patches, competitive elution, and an algorithm to differentiate DNA–RNA binding from RNA–small-molecule binding [105]. Using these strategies, they identified two compounds, from a 46.3 billion-member DEL, with low nanomolar affinity to FMN riboswitch and that compete with the cognate FMN ligand.

3.6. Virtual screening

Virtual screening can significantly reduce costs and expand the chemical space accessible, as well as remove inactive compounds prior to experimental testing. One of the virtual screening approaches involves the docking of virtual libraries to the 3D structure of the target, followed by evaluation of the binding energies using a scoring function (Figure 3(f)). There are a few programs that were developed specifically for docking small molecules to RNA, such as MORDOR [134] and rDock (formerly, RiboDock) [135]. In addition, there are programs, originally designed for protein targets, that were adapted or reparameterized for small-molecule docking to RNA, including Dock6 [136], ICM [137], and AutoDock [138].

Daldrop et al. [107] used the program DOCK 3.5.54 to identify novel binders to the guanine riboswitch (GRA). First, they docked 23 compounds (test set)–8 known ligands and 15 decoys–to the crystal structure of the GRA carrying a C74U mutation [139]. This resulted in a clear separation of the ligands and decoys. Then, they assembled a larger database containing 2,592 compounds plus the 23 compounds in the test set, docked the database to GRA, and ordered the compounds by score. A near perfect discrimination between known ligands and database compounds/decoys was obtained. Encouraged by these results, they performed more studies on the top scoring hits and found four new ligands of GRA with micromolar level binding affinities, two of which have scaffolds not resembling those of previously known ligands.

The Al-Hashimi group considered the large conformational changes in RNA targets by docking small molecules on RNA dynamic ensembles [108]. The group first carried out 80 ns molecular dynamics (MD) simulations on NMR structures of HIV-1 TAR RNA to select 20 conformers that would make up the dynamic ensemble. Then, they subjected the TAR ensemble to virtual screening against ~51,000 small molecules. Fluorescence-based assays were performed using the top 57 hits, resulting in the identification of six compounds that bind to TAR that were not previously known to bind to TAR. One compound, netilmicin, exhibits exceptional selectivity for TAR and inhibits interaction with Tat and HIV replication.

3.7. Phenotypic screening

Using a mouse model for spinal muscular atrophy (SMA), PTC Therapeutics and Hoffmann-LaRoche performed phenotypic screening (Figure 3(g)) to identify splicing modifiers of survival of motor neuron 2 (SMN2) with high selectivity and oral bioavailability [113]. They employed a cell-based system in which the luciferase reporter protein was only expressed when exon 7 of the SMN2 minigene was included after splicing. Later studies revealed that the small molecules function by binding to the SMN2 pre-mRNA and the U1 snRNP [140]. The SMN2 splicing modifiers were optimized, resulting in the discovery of the now FDA-approved risdiplam [114].

Around the same time the PTC-Roche SMN2 splicing modifiers were identified, a team at Novartis also performed phenotypic screening to identify small molecules with the ability to reduce exon 7 exclusion [115]. An NSC34 motor neuron cell line expressing an SMN2 minigene reporter was employed to screen their compound library (~1,400,000 compounds). Two active compounds – branaplam (originally NVS-SM1) and NVS-SM2 – were found to have good efficacy, bioavailability, and distribution to the brain. Branaplam is currently in a Phase 1/2 clinical trial in infants with SMA (NCT02268552) [141]. Interestingly, branaplam also promotes the inclusion of a pseudoexon in huntingtin (HTT) transcripts, thereby downregulating HTT mRNAs and protein [142].

A third example of the successful application of phenotypic screening in identifying RNA binders was reported by researchers at Merck, who screened a chemical library of ~57,000 compounds for antibacterial activity against an E. coli in the presence of riboflavin [20]. They found that ribocil – a small molecule that does not resemble riboflavin – completely suppresses E. coli bioactivity. Later, they identified the target of ribocil to be the FMN riboswitch by selecting E. coli ribocil-resistant mutants and mapping ribocil-resistant mutations to their target using whole-genome sequencing. Further evaluation revealed that ribocil is a racemic mixture of isomers and that one of its enantiomers, ribocil-B, predominantly inhibits riboflavin biosynthesis and bacterial growth.

4. Biophysical techniques to confirm in vitro binding of the hit compounds

After screening for compounds that potentially bind the RNA targets, the next critical steps are: (1) to confirm that such compounds bind to the target, (2) assess the functional effects of the compounds, and (3) establish target engagement. In this section, several of the biophysical techniques used for confirming the binding of small molecules to an RNA target (Figure 4) will be described. To select the most appropriate technique, one should review the range, strengths, and limitations of each biophysical technique. Ideally, one should perform multiple biophysical assays in parallel or in a funnel format from the highest to the lowest throughput. Furthermore, appropriate experimental design and conditions, use of suitable controls, skill in the application of the technique, and correct and adequate data analysis are crucial for ensuring accurate and repeatable results. Given the multitude of biophysical techniques, this review will only cover several of the most powerful techniques for probing RNA–small-molecule interactions.

Figure 4. Biophysical techniques for confirming and validating hit compounds. (a) In an ITC experiment, the reference and sample cells should contain the same buffer because mismatched buffers lead to large background heats during the titration. Each peak in the raw data counts for one injection of the ligand solution, and the respective enthalpy is measured. The data is analyzed by fitting a sigmoidal curve to the released or absorbed enthalpy per mol plotted against the molar ratio. The binding affinity and stoichiometry can be determined from the fitted curve. (b) For SPR measurements, it is generally recommended to use low surface densities of the immobilized RNA and high ligand flow rate for accurate determination of kinetic constants. SPR causes a dip in the intensity of the reflected light at the sensor surface. A shift in the curve represents binding of the ligand to the RNA. An SPR sensogram can be divided into the association, steady state or equilibrium, and dissociation phases. (c) Since BLI experiments are conducted in uncovered plates and using small volumes, the entire experiment should be conducted swiftly (< 4 hours) to limit excessive evaporation, which can impact experimental results. In BLI, changes in the number of molecules bound to the biosensor tip shift the interference pattern that can be monitored in real-time. Association and dissociation are measured by dipping the sensor in solutions with or without the ligand, respectively. (d) Prior to MST assays, different assay conditions should be tested because the type of buffer, pH, and ionic strength can influence the thermophoretic movement of molecules and dramatically change the outcome. The RNA and ligand solutions are incubated and filled into a glass capillary. At a specific temperature, an infrared laser generates a temperature gradient, and RNA-small molecule complexes move slower in that gradient than unbound RNA. A binding curve can be generated from the difference between the fluorescence signals of the bound and unbound states, from which a binding constant can be derived. (e) The H1 and H3 protons of guanosines and uridines, respectively, as well as the H5–H6 crosspeaks of pyrimidines, can be perturbed after small molecule binding. By obtaining the ¹H and 2D ¹H −1H NMR spectra of the RNA target prior to and after the addition of small molecules, one can monitor these perturbations and determine if the small molecule binds to the RNA and which RNA residues bind the small molecule.

4.1. Isothermal titration calorimetry (ITC)

The gold standard for obtaining binding constants and thermodynamic parameters for target–ligand interactions is ITC [143]. During an ITC experiment, a ligand (small molecule) solution is typically injected into a sample cell containing the label-free target biomolecule (RNA) (Figure 4(a)). A sensitive calorimeter measures the heat released or absorbed during titration until the binding reaction has reached equilibrium. The measured heat values are then used to obtain the binding affinity (K_D), stoichiometry (N), and change in enthalpy (∆H), from which the changes in free energy (∆G) and entropy (∆S) are calculated. Obtaining these thermodynamic parameters for binding of the hit compounds to the target RNA can help in compound classification and prioritization at a later stage. Although a trusted technique, ITC requires large amounts of sample and high solubility of the small-molecule compound.

4.2. Surface plasmon resonance (SPR)

Characterizing the kinetics of binding can lead to an understanding of the efficacy and safety of the small molecule and therefore enhance the success of the drug discovery program. SPR is the most extensively used technique for measuring the kinetics of small-molecule binding to an RNA [144]. In an SPR experiment using a microfluidic system, the target biomolecule (RNA) is immobilized onto a sensor chip, and a pump is used to flow the liquid containing the ligand (small molecule) over the sensor chip (Figure 4(b)). An optical system measures the changes in refractive index at the interface of the liquid and the sensor, which are plotted over time in the form of a sensogram. From the SPR experiment, one can obtain the kinetic parameters, such as the association and dissociation rate constants (k_on and k_off, respectively) and K_D. SPR is highly sensitive, relatively high throughput, and has a small sample requirement. However, the target immobilized onto the sensor must be stable over time, and a high level of expertise and experience are necessary for carrying out a successful SPR experiment.

4.3. Biolayer interferometry (BLI)

Another optical-based technique, BLI, can be used to measure RNA–small-molecule interaction [96,145,146]. BLI analyzes the interference patterns of white light reflected from the surface of a glass fiber biosensor (Figure 4(c)). The biotinylated target (RNA) is first immobilized onto streptavidin biosensor tips, then the tips are dipped into a solution containing the ligand (small molecule). The interaction of the ligand with the target can be monitored by recording the changes in light interference, which directly correlate with changes in the thickness of the biolayer resulting from the association of the ligand. Like SPR, one can obtain the k_on, k_off, and K_D values from a BLI experiment. Although SPR is widely used in the pharmaceutical industry, BLI is gaining popularity because it is cheaper and easier to use, enables significantly higher throughput, does not rely on fluidics, and is less technically demanding compared to SPR.

4.4. Microscale thermophoresis (MST)

One of the emerging biophysical techniques that is being applied for probing RNA–small-molecule interactions is MST [147]. In an MST experiment, a capillary containing a fluorescently labeled molecule (the target RNA) and a nonfluorescent ligand (small molecule) is gently heated with an infrared laser to generate a temperature gradient, along which the molecule migrates in a phenomenon known as thermophoresis (Figure 4(d)). The fluorescent molecule is mixed with different concentrations of the nonfluorescent ligand, and the effect of different concentrations of the ligand on the migration of the molecule is quantified to measure binding. While MST is versatile in its applications and does not require immobilization of the sample, preparing the samples may be time consuming and labeling may alter the binding properties of the small molecule. Moreover, MST does not provide the kinetic constants like SPR and BLI do and has a lower sensitivity than ITC.

4.5. Nuclear magnetic resonance (NMR)

RNA-observed NMR techniques provide a wealth of information regarding how a ligand interacts with its target RNA, and such information is extremely helpful for understanding the mechanism of recognition and rational design. The most straightforward method is chemical shift mapping with ¹H NMR spectroscopy using the imino proton region, in which the perturbations in the solvent exchangeable H1 of guanosines and H3 of uridines can be monitored upon gradual addition of the ligand (Figure 4(e)) [148]. The H1 and H3 imino protons serve as hydrogen donors in base pairing interactions and are, thus, perturbed when a small molecule interacts with the RNA. Because of the distinct chemical shift ranges and minimal spectroscopic overlap, perturbations in these imino protons can be monitored in a simple 1D manner. One disadvantage of imino ¹H NMR spectroscopy is that, because the unstructured imino protons cannot be observed, it cannot detect perturbations when they occur at the loops and bulges on the RNA. To monitor more changes in the RNA brought about by small-molecule binding, it is advisable to pair imino-based NMR characterizations with 2D ¹H-¹H total correlated spectroscopy (TOCSY, otherwise known as Homonuclear Hartmann-Hahn or HOHAHA). This method measures the H5–H6 cross-peak of pyrimidines (Figure 4(e)), and since the H5 and H6 protons are non-solvent exchangeable and carbon-bound, TOCSY can robustly monitor perturbations even those of pyrimidines in unstructured loops and bulges.

5. Assessing the functional effects and target engagement of RNA-binding small molecules

As mentioned above, aside from determining the binding of small molecules to the RNA target with in vitro experiments, it is necessary to assess the functional and biochemical effects of the small molecules in a physiological context using cell-based assays. These assays can provide information about the efficacy, mechanism of action, cytotoxicity, cell permeability, stability, and off-target effects of the RNA binders. As is the case for protein-targeted small-molecule drug discovery, one must carefully select from a broad range of cell-based assay types and cell sources to simulate a physiological or disease-specific environment.

In addition, it is critical to assess whether the functional effect of the small molecule is a consequence of direct engagement of the RNA target. The techniques that have been used to assess engagement of small molecules to RNA targets are summarized in Table 3. It can be expected that drug regulatory authorities will expect a thorough investigation on the potential off-targets – both RNA and protein – of the small molecules especially since there are fewer examples of RNA-targeted small-molecule drugs.

Table 3. Techniques to assess engagement of small molecules to RNA targets.

Technique	Description	References
Chem-CLIP	Steps: Preparation of the Chem-CLIP probe: Two functional units are appended with the small molecule–a chemical cross-linking unit that allows the small molecule to attach to RNAs and a purification unit (biotin) for easy isolation of the bound RNA Incubation of cells with the Chem-CLIP probe Isolation and purification of bound RNAs RT-qPCR or RNA-seq to identify the bound RNAs	[146,149–157]
C-Chem-CLIP	By using C-Chem-CLIP, one can determine the RNA bound not by the compound but by the cross-linking and purification units. Cells are incubated with the Chem-CLIP probe and the competing non-cross-linking parent compound. RNAs bound by the parent compound are depleted in the pull-down.	[146,150–153,155]
Chem-CLIP-Map-Seq	Chem-CLIP-Map is an extension of the Chem-CLIP technique. It was developed for repeat sequences since it is difficult for RT to read through the repeats and to PCR amplify. Chem-CLIP-Map is performed in the following manner: Traditional Chem-CLIP experiment Digestion of bound RNAs with antisense oligonucleotide and RNase H Analysis of digested fragments using RT-qPCR	[150,152]
Global proteomics	Global proteomics can be used to study the effects of RNA-targeted small molecules on proteins. This can be performed by incubating the cells with the small molecules, then harvesting and analyzing the proteins using LC-MS/MS.	[156]
PEARL-seq	PEARL-seq employs photoaffinity labeled probes to assess engagement of a small molecule to its intended RNA target and identify other RNAs it binds to. The technique is carried out by: Preparing small-molecule compounds with photoaffinity labels (photoprobes) Addition of the photoprobes to RNA extracted from cells Cross-linking, biotinylation, and fragmentation of the RNA Avidin capture, reverse transcription, and cDNA elution Second adapter ligation PCR amplification and sequencing	[158,159]
RiboSNAP	RiboSNAP is a technique that uses small molecules capable of cleaving an RNA target to profile the binding of small molecules directly from total RNA obtained from cells. There are three variations of RiboSNAP. In the first variation, the small molecule is appended with a light-activatable cleavage module (e.g. N-HPT). After irradiation with light, the cleavage module produces radicals that cleave the target RNA. The second variation uses RIBOTACs – cleavage-inducing probes comprised of RNA-binding and nuclease-recruiting units – to target RNAs and activate RNase L, leading to the degradation of the bound RNAs. The third variation was developed to attenuate the issues associated with light-activatable cleavage modules (limited tissues to which light can be applied) and RIBOTACs (complex cleavage patterns). For a cleavage module, this variation uses bleomycin A5, which, when appended with the RNA-binding small molecule, allows for RNA-specific cleavage.	[152,160]

*Abbreviations used: Chem-CLIP, chemical cross-linking and isolation by pull down; C-Chem-CLIP, competitive Chem-CLIP; N-HPT, 1-hydroxy-6-thioxo-1,6-dihydropyridine-2-carboxylic acid; LC-MS/MS, liquid chromatography with tandem mass spectrometry; RiboSNAP, small molecule nucleic acid profiling by cleavage applied to RNA; RIBOTAC, ribonuclease-targeting chimeras; RNA-seq, RNA sequencing; RT, reverse transcriptase; RT-qPCR, RT quantitative polymerase chain reaction

6. Techniques for determining 3D structure

Although not a strict requirement in drug discovery, the 3D structure of the target–small-molecule complex helps us better understand how a small molecule binds to its target and gives us clues on how to optimize the small molecule for increased potency and selectivity. The three techniques for determining the 3D structures of RNA–small-molecule complexes are X-ray crystallography, NMR, and cryo-electron microscopy (cryo-EM) (Figure 5).

Figure 5. Comparison of the techniques for the determination of the 3D structures of RNA–small-molecule complexes. (a) X-ray crystallography can provide atomic resolution structures for RNAs of any size but requires a large amount of sample that must be crystallized. The crystal structure of the FMN riboswitch in complex with the compound, BRX1151 (red), is shown [22]. (b) Unlike X-ray crystallography, solution-state NMR structures provide information about the dynamics of the RNA target, but it also requires large amounts of sample and is not suited for solving the 3D structures of large RNAs. The solution structure of the UGGAA-UGGAA RNA pentad in complex with the naphthyridine carbamate dimer (NCD, red) is shown [8]. (c) In recent years, the resolutions of cryo-EM structures of RNAs are improving, but it is still only suitable for large RNAs or RNA–protein complexes. Cryo-EM requires only a small amount of sample, and the obtained structures are closer to the RNA’s native state than the X-ray structures. The cryo-EM structure of the SAM IV riboswitch–SAM (red) complex is shown [161].

6.1. X-ray crystallography

The first step in X-ray crystallography is obtaining single crystals of sufficient size and quality. After obtaining the crystal(s), a diffraction experiment is performed, where the crystal is illuminated with X-ray beams, producing a diffraction pattern. The crystal is gradually rotated to collect a set of diffraction patterns, which are converted into a 3D model that shows the electron density within the crystal. Based on the electron density map, the atomic model of the RNA–small-molecule complex is built and subsequently refined using computational tools for structure determination.

Over the past few decades, various crystallization strategies – the use of protein chaperones [162], tetraloops [163], kissing loops [164], the tRNA scaffold [165], and model RNAs [166] – have improved the likelihood of obtaining crystals for RNA and its complexes. Moreover, crystallization robots have significantly reduced the amount of sample required for crystallization experiments and the experimental burden. The introduction of covalently or non-covalently bound heavy- or anomalously scattering atoms [167–169] also led to successful crystallographic phase determination of RNAs. In addition, microfocus beamlines, detector technologies, and X-ray-free electron lasers enabled structure determination from substantially smaller crystals. Finally, the availability of highly automated and comprehensive computational systems, such as the Phenix software suite [170], that can arrive at an initial model of a structure with minimal human intervention, has largely reduced the time consumed for structure determination.

X-ray crystallography can be used for RNAs of any size and provides atomic resolution images of the RNA–small-molecule structures (Figure 5(a)). Although a powerful tool that delivers highly reliable structural information, X-ray crystallography requires that a sample be crystallizable, that the compound be soluble, and that a substantial percentage of binding sites be occupied by the ligand. In addition, an X-ray crystal structure represents only one (or at most, two) static form of the RNA–small-molecule complex.

6.2. Solution-state NMR

An NMR experiment is often performed on an RNA–small-molecule complex dissolved in sodium phosphate buffer with salt, usually sodium chloride. The water suppression technique allows measurement with a buffer containing 5–10% D₂O without having to replace the buffer with 100% D₂O. Prior to NMR experiments, it is advisable to test different annealing and salt conditions to determine the proper folding conditions for the RNA target. In some cases, the use of ¹³C/¹⁵N labeled RNA in multinuclear experiments can provide more chemical shift information and reduce signal overlap.

Typically, the 2D-NOESY spectra, focusing on the imino proton region, are first measured and the signals are assigned to determine the secondary structure [171]. Then, NMR measurements (2D-NOESY or 2D ¹H-¹H TOCSY) for the non-exchangeable protons are performed, and the NOE cross-peaks between purine H2 (only adenine), H8, or pyrimidine H6 and pyrimidine H5 or ribose H1’ are assigned. Following signal assignment, structural information (proton–proton distance, dihedral angle, and base pairing) is derived, and a restraint file is created for structure calculations. Currently, most calculations are performed using the simulated annealing protocol, and the calculations are performed until the restraints are satisfied and the structure converges. Iterative rounds of signal assignment and structure calculations are made until an acceptable ensemble of structures is obtained.

RNA synthesis techniques have allowed the preparation of samples with different isotopic labels, which have proven to be useful in structure determination and dynamic studies [172]. Furthermore, vast improvement in electronics and cryogenic probes increased the sensitivity of NMR measurements by tenfold over the past decade. Recently, Novakovic et al. proposed an approach for collecting NMR spectra for RNAs by injecting hyperpolarized water and reported a 300-fold enhancement of imino resonances [173]. Other notable advances include computational approaches that improve the ability to probe the dynamic RNA ensembles, such as those for predicting chemical shifts from structure and generating conformational libraries from sequence [174].

The distinguishing feature of solution-state NMR is that, as its name clearly implies, the 3D structure of RNA–small-molecule complex is measured in solution (Figure 5(b)). Thus, NMR’s biggest advantage over X-ray crystallography is that it provides information about the dynamics of small-molecule binding. However, the NMR spectra of large RNAs may be very complicated and impossible to interpret, thereby limiting the application of NMR to smaller RNAs.

6.3. Cryo-EM

The continuing advances in cryo-EM offer unprecedented opportunities to visualize large RNA–small-molecule complexes. In a typical cryo-EM experiment, the samples are applied to a grid mesh and plunge frozen in liquid ethane in a process called vitrification. During this process, the sample solution is kept in an amorphous, noncrystalline state. The frozen grids are then screened for appropriate ice thickness, sample density, and particle orientation. Next, a series of two-dimensional images (micrographs) of the sample are taken, and computational tools are used to extract and average representative views of the molecules from the micrographs. In the end, atomic models are built into the EM maps and refined to generate the macromolecular structure.

Compared to X-ray crystallography and NMR, cryo-EM requires only a small amount of sample and can tolerate lower degrees of purity (Figure 5(c)). Moreover, since there is no need to crystallize the sample, cryo-EM is more straightforward and provides structures that are closer to its native state than X-ray crystallography. However, cryo-EM structures of RNAs are rare, and the rare structures are only at a moderate resolution (3.0–10 Å). This is due to the intrinsic heterogeneity and small size of RNAs and the lack of effective approaches to obtain properly folded, stable RNAs. Efforts to overcome these challenges are underway and include: (1) strategies to create larger and more stable RNA nanoarchitectures [175], (2) development of the Ribosolve workflow [176] that integrates native gel analysis, mutate-and-map by next-generation sequencing, cryo-EM, and auto-DRRAFTER RNA modeling [177], and (3) hybrid approaches combining RNA modeling, X-ray crystallography, and NMR.

7. Optimization of RNA-binding small molecules

Understanding the key interactions between an RNA target and a small molecule is highly beneficial for optimizing small molecules to improve their potency and selectivity. However, even when the 3D structure of an RNA–small-molecule complex is available, visual inspection is not enough to fully understand the molecular interactions. One of the promising methods for exploring molecular interactions is the fragment molecular orbital (FMO) method [178], which provides a comprehensive list of the interactions and their strengths in terms of energy stabilization of the complex. FMO is much faster than traditional quantum mechanics (QM) methods because FMO calculations are done by partitioning a large system into fragments.

Recently, the Frank group applied the FMO method to probe the interactions between the FMN riboswitch and nine small molecules [179]. Their analyses resulted in the identification of the hot-spot residues within the binding pocket on the FMN riboswitch, as well as in the confirmation of the dominance of electrostatic interaction in driving ligand binding. By examining the structure–energy relationships within the nine complexes, the group was also able to uncover an atypical T-shaped contact between a riboswitch residue and the ligands, which may further stabilize complex formation.

Even without comprehensive FMO analysis, one can still design potent and selective RNA-binding small molecules using the available 3D structure information, albeit with increased complexity. Vicens et al. [22] used the structure-based drug design strategy to optimize roseoflavin mononucleotide, which binds in a nonspecific manner to the FMN riboswitch. In their study, they examined the 3D structures of complexes formed by FMN riboswitches to determine the key interactions that would need to be maintained or replaced in the analogues. Through careful dissection of the effects of the modifications on binding affinity and activity using a combination of chemical probing and transcription termination assays, they discovered 5FDQD, which binds to the FMN riboswitch and shows antibacterial activity.

8. Conclusion

Here, the different technologies for identifying therapeutically compelling RNA motifs, screening for small-molecule compounds that bind to them, confirming in vitro binding, assessing the bioactivity and establishing target engagement, determining the 3D structures of RNA–small-molecule complexes, and optimizing the hit compounds to identify lead molecules with increased potency and selectivity have been discussed. The RNA community has successfully been able to repurpose the experimental techniques and adapt the computational tools that were protein-centric for use in RNA targets, as well as create technologies tailored for RNA. As more advances in both experimental and computational technologies are made and leveraged to their fullest potential, we are likely to witness an emergence of more RNA-targeted small molecules drugs, particularly for diseases with unmet medical needs.

9. Expert opinion

Small molecules have been the preferred modality for drug development because of their good oral bioavailability, broad tissue distribution, lower risk of immunogenicity, and lower manufacturing cost. However, the pharmaceutical industry’s productivity has been rapidly declining due to the diminishing pool of disease-related proteins that are the conventional targets of small-molecule drug discovery. Thus, finding novel therapeutic targets is critical. Targeting RNAs with small molecules is one way to address this problem. There are already several encouraging evidence that RNA-targeted small-molecule drugs could be developed.

Nevertheless, there remain areas for further improvement. The first is in target identification. For example, the nearest neighbor parameters that are the core of RNA secondary structure prediction algorithms have changed only modestly since they were introduced by Xia et al. [58] in 1998. More thermodynamic data – in a wide variety of RNA lengths, base pair arrangements, and physical conditions – plus a more rational way to consolidate the data are needed to achieve more accurate RNA secondary structure predictions [180]. In addition, although some RNA secondary structure prediction software can predict non-Watson–Crick pairs and higher-order structures, such as pseudoknots and G-quadruplexes, none of them can predict all or with great accuracy. Also, while the chemical probing techniques have greatly improved since their infancy, it is still difficult to obtain accurate reactivity profiles. Therefore, secondary structure predictions should be coupled with rigorous experimental verification, much like what Vicens and Kieft proposed in their paper [181]. Various groups, including ours, are developing technologies to overcome the above-mentioned limitations in target identification technologies. We can anticipate that RNA targets will be better characterized in the next 5 years.

A second area that requires improvement is the way we approach screening. With our predilection for drug-like small molecules, we tend to screen against RNA molecules with complex structures. There is a danger to this – we may preclude opportunities to discover drugs for diseases linked to RNAs with simpler structures. Repeat sequences that form stem loop structures are a good case in point. While we should continue our pursuit of drug-like modulators of RNA, we should in parallel venture outside the Lipinski space by screening with more diverse chemical libraries. Along with trying different chemical space, should resources allow, it would be beneficial to screen against a variety of RNA motifs using one chemical library, a strategy previously employed with 2DCS [124] and DEL screening [104]. This not only helps in weeding out promiscuous RNA binders but also in the comprehensive characterization of compound libraries.

The third area for improvement is in determining the 3D structures of RNA and RNA–small-molecule complexes. While many efforts have been made, there is still a dearth of RNA 3D structures; less than 1% of the total number of structures in the Protein Data Bank is for RNA. A new machine learning approach called Atomic Rotationally Equivariant Scorer (ARES) has been shown to effectively learn even from a small amount of structural data and outperform other 3D structure prediction software [182]. This is an encouraging development, but ARES’ accuracy in predictions for RNAs is lower than that of AlphaFold for proteins. To improve its accuracy, we should obtain more 3D structure information using experimental methods to procure a larger data set on which to train ARES and other machine learning tools.

The fourth and last area that would benefit from more research is hit-to-lead optimization. A structure-based computer-assisted manner of optimizing compounds is ideal but has yet to be proven effective. Since even ‘structured RNAs’ populate an ensemble of conformations, it is possible that one compound may capture one conformation from the ensemble, and another compound (even an analogue) may capture another conformation. One way to address this situation is molecular dynamics (MD). Provided they are carefully applied, MD simulations on available 3D structures can be used to explore RNA dynamics and conformational changes on the submicrosecond time scale [183]. Moreover, combining MD simulations with QM methods, such as the FMO method, would enable the evaluation of interaction energies between the RNA in its various forms and the hit compound, and that information is expected to guide drug design.

There remains an abundance of functional RNAs that are yet to be discovered. Comparative structure analyses have revealed conserved RNA structures that perform a certain function. Analyses of interactions with RNA-binding proteins hint at the functions of the previously uncharacterized RNA motifs they bind to. However, these analyses must be supported by experimental evidence. Moreover, the number of databases for RNA is continuing to grow, and this could cause confusion even to people in the RNA community. Thankfully, there are efforts being put forth to unify the data; RNAcentral is one of them [184,185]. Launched in 2014, RNAcentral is a comprehensive database containing information on non-coding RNAs spanning a broad range of functions and species. It is our hope that this database be expanded to also cover mRNA and pre-mRNA sequences that may have functions outside coding for proteins.

Article highlights

• A variety of experimental and computational technologies have been developed for the discovery of RNA-targeted small molecules

• Therapeutically valuable RNA motifs can be identified using in silico secondary structure prediction and chemical probing methods, but these methods still require further improvement for greater accuracy

• The utility of target-based and phenotypic screening approaches in the discovery of RNA-targeted small molecules has been demonstrated against a wide variety of diseases

• The binding of small-molecule compounds to the RNA target can be confirmed using in vitro biophysical techniques that are also commonly used for protein targets

• After confirming binding, it is crucial to determine the functional and biochemical effects of the RNA-targeted small molecules, establish target engagement, and obtain three-dimensional structure information

This box summarizes key points contained in the article.

Declaration of interest

E Morishita is a senior investigator at Veritas In Silico. The author has no other relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed.

Reviewer disclosures

Peer reviewers on this manuscript have no relevant financial or other relationships to disclose.

Acknowledgements

The author is grateful to Gota Kawai, Shingo Nakamura, Takashi Kamimura, Amiu Shino, and Maina Otsu for their helpful feedback and comments.

Additional information

Funding

This paper was not funded.