Efficient NMR Screening Approach to Discover Small Molecule Fragments Binding Structured RNA

Abstract

We describe a scalable nuclear magnetic resonance (NMR) screening approach to identify and prioritize small molecule fragments that bind to structured RNAs. This approach is target agnostic and, therefore, amenable to many RNA structures and libraries, and it provides initial hits for further synthetic elaboration and structure-based drug discovery efforts. We demonstrate the approach on the pre-miR-21 stem-loop, which is of significant interest in oncology and metabolic diseases. We screened the pre-miR-21 hairpin using a small (420 compounds) commercially available fragment library and identified 18 hits in the first round of triage screening. This was further refined to four fragments which passed all screening cascade filters. Among these four hits, a thiadiazole fragment was demonstrated to bind the Dicer cleavage site of pre-miR-21 by target-detected NMR experiments and through the observation of clear intermolecular NOEs.

The potential to dramatically increase the number of “druggable” therapeutic targets has renewed significant investments to target ribonucleic acid (RNA) with small molecules.¹⁻³ While drugging RNA is believed to be difficult given its uniform negative charge and limited chemical diversity, a few sustained projects in the biotech industry and academia have discovered druglike molecules with nanomolar binding activity against viral RNA regulatory elements.^4,5 Furthermore, FDA approval of compounds which selectively regulate RNA splicing in spinal muscular atrophy (SMA) patients has demonstrated the clinical benefit of targeting post-transcriptional gene expression mechanisms, despite relatively weak affinity.^1,6

Finding small druglike molecules which bind RNA remains challenging, limited by the very low hit rate of protein-directed small molecule libraries and the general lack of knowledge of RNA targeting chemistry. In fact, it has been stated, for example, that targeting RNA stem-loops would be very challenging and only more elaborate structures would be amenable to drug discovery efforts.⁷ However, the RNA hairpin is the most common and persistent RNA structure, likely to be locally folded even as more complex structures found in vitro are unfolded in the cell.^8,9 Therefore, developing methods and identifying compounds that bind to RNA stem-loops is of significant value. Herein we report an efficient and scalable method to identify and semiquantitatively rank compounds that bind RNA. We demonstrate the screening approach on a difficult target, pre-miR21, a therapeutically relevant noncoding RNA.

The overexpression of microRNA-21 (miR-21) leads to the downregulation of key tumor suppressors and proapoptotic factors along with polarizing T-cells and macrophages away from the tumor suppressing phenotype.¹⁰⁻¹² The suppression of mature miR-21 overexpression could reverse malignancy by exploiting tumor addiction to miR-21 in vivo, as suggested by several animal models.^13,14 Consistent with this hypothesis, targeting the mature single stranded RNA directly with antisense oligonucleotides can reverse the phenotypic consequences of miR-21 overexpression in cellular and animal models.¹⁵⁻¹⁸ Since the biogenesis of mature miR-21 expression is tightly regulated, interfering with critical regulatory steps could provide appealing opportunities for small molecule inhibitors that would achieve the same effect as antisense approaches but with much more attractive pharmaceutical profile. The canonical biogenesis pathway of mature miR-21 expression is shown in Figure 1. One potential route to regulate miR-21 levels is through inhibition of the enzymatic Dicer processing step. However, targeting the Dicer enzyme directly is likely to have global effects on microRNA processing, since Dicer knockdown is universally deleterious in animals.¹⁹⁻²¹

Figure 1.

Canonical miRNA biogenesis pathway (black arrows) initiates with transcription of the miRNA gene by PolII (1) to generate the primary miRNA transcript (2), which is cleaved by the microprocessor complex (Drosha/DGCR8) (3) to generate the precursor miR-21 hairpin (pre-miR21) (4). The pre-miR21 hairpin is exported to the cytoplasm by Exportin5 (5) for further processing by the miRISC complex (Dicer/TRBP/AGO2) (6) which cleaves off the apical loop structure to generate the mature miRNA–miRNA* duplex (7). The mature single stranded miRNA-21 (8), which coincides with the 5′ strand of the miRNA–miRNA* duplex, is incorporated in the miRISC complex with its component Ago2 and can then target mRNAs directly (9) to regulate translation or RNA stability. This biogenesis pathway offers several opportunities to target miRNA precursor species with small molecules. Here we discuss the discovery of small molecule fragments that target the conserved apical loop structure (red arrows) common to both the primary and precursor transcripts (10). Image created with BioRender.com.

Targeting the pre-miR-21 hairpin directly, on the other hand, could allow the selective inhibition of miR-21 maturation without affecting other microRNAs, provided specificity is achieved in selectively targeting the pre-miR-21 stem-loop over other cellular pre-miRNAs Recent data using linear peptides, macrocycles, natural products, and other large compounds suggest direct targeting of the pre-miR-21 structure would indeed lead to inhibition of the mature miR-21, with observed reduction of expression in vitro and in cellular assays.²²⁻²⁷ However, the relatively large molecular weight, toxicity, and general physiochemical properties of molecules identified so far make their development as drugs very unlikely. Unfortunately, screening of druglike compound libraries originating from protein-targeting chemical space and closer to pharmaceutically active compounds, typically yield very low hit rates and generally weak, mid-micromolar binding activity.²⁸⁻³¹ Fragment-based drug discovery (FBDD) approaches provide an attractive alternative, which has often been used successfully with hard-to-drug proteins.³²⁻³⁴ The FBDD approach to screening has often found initial fragments, which can be developed into genuine small molecule hits using well-established fragment growing or linking strategies.³⁵

Here we describe an approach which builds on RNA fragment screening studies reported earlier.³⁶⁻⁴⁰ The main differences are the lack of probe molecule used to preform the RNA structure and create a binding pocket, and the use of relaxation editing experiments. We demonstrate this approach on the apical loop of pre-miR21 and report the discovery of four robustly confirmed RNA binding fragments, one of which binds proximal to the Dicer cleavage site and displays intermolecular NOEs that allow modeling of the intermolecular interface.

The 72 nts full-length pre-miR-21 was truncated to focus screening efforts on the apical loop structure which regulates both Drosha and Dicer processing. We were particularly interested in the region near the Dicer cleavage site, between nucleotides A29 and G45 (Figure 2A), which is also the binding site for the L50 peptide we reported earlier.²⁷ This shortened RNA construct was experimentally screened against a commercially available fragment library following the screening cascade reported in Figure 2B, which uses both ligand and target detect NMR experiments to confirm binding and identify putative binding sites in the RNA.

Figure 2.

(A) NMR derived secondary structure of the pre-miR-21 apical loop screening construct. The Dicer cleavage site is highlighted in red and coincides with A29 at the 5′ end and G45 at the 3′ end. Weak base pairs in the apical loop are indicated with blue dashes; these were identified by the observation of weak and broad imino peaks in the 1D ¹H NMR spectrum of free pre-miR21.²⁷ (B) First, we run a series of ligand detect NMR experiments (orange) to identify binding compounds from mixtures. At this initial stage, any change in ligand line width is considered a hit. Once this set of putative hits are identified from mixture screening, the compounds are rescreened one at a time for conclusive verification. In this second step, we compare the intensity of the bound (I_B) relative to the free ligand peak (I_F) to generate a binding score (1 – I_B/I_F); this binding score is proportional to K_D. If compounds bind both mixture and single compound screens, we run a ligand detect NOESY experiment to monitor changes in the sign of the small molecule NOE peaks. Compounds that show positive hits in all three experiments are moved forward to conduct a set of target-detected experiments (gray). The target-detect cascade can also establish whether the ligand induces any change in the secondary structure or RNA base paring by monitoring for changes in chemical shift of base paired imino resonances. This step is followed by the measurement of chemical shift changes upon binding within pyrimidine residues, which is done by recording ¹H TOCSY spectra of free and ligand bound samples. Lastly, a NOESY is collected for interesting RNA-Ligand complexes to identify potential intermolecular NOEs.

The fragment library was assembled from the Maybridge “Rule of 3” collection initially designed to target enzymes with nucleoside substrates, such as thymidylate synthase.⁴¹ This 420-member fragment library consisted of heteroaromatic biaryl fragments along with mono- and poly-heterocyclic compounds, with an average molecular weight of 189 g/mol, ranging between 94 and 291 g/mol.⁴¹ The fragment library was divided into 54 individual mixtures with 5–8 ligands per mixture to maximize 1D proton resolution for NMR screening. Obviously, mixture screening increases throughput and provides intrinsic competition. However, to minimize competition between fragments with similar structures, mixtures were prepared with no more than two members sharing the same scaffold.

Primary screening against this mixture library was conducted using 1D ¹H line-broadening experiments, which we developed to examine binding of individual compounds to both RNA⁴² and proteins.⁴³ Line broadening experiments unambiguously identify a ligand bound to a macromolecular target based on changes in ligand relaxation rates between free and bound small molecule (Figure 3). When a ligand binds to RNA, the T₂ relaxation rate of the small molecule fragment increases in proportion to the molecular weight and K_D of the complex,⁴³ as expressed in eq 1:

1

In this expression, B is the fractional bound occupancy, the sum of ligand peak intensities represented by bound (I_B) and free (I_F) ligand. Plotting the fractional occupancy (B) versus the concentration of biomolecular target ([R]_T) at a single ligand concentration ([L]_T) produces a binding isotherm from which the apparent affinity (K_D) can be calculated by fitting the curve, as long as the NMR line width ratio (c) is constant. A formal derivation of the constant c was previously conducted for proteins.⁴³ While this protocol is straightforward, collecting a full titration curve and measuring K_D for each fragment greatly increases screening time. Therefore, we introduce a triage tool which uses the fractional occupancy B in a binary fashion, as measured by the ratio of bound and free ligand line widths. We show herein that this simple ratio is sufficient to rank and prioritize RNA binding ligands. This step greatly increases the throughput needed to identifying RNA binding fragments.

Figure 3.

Ligand-detect workflow used to identify a 1,2,4-thiadiazole fragment binding to the pre-miR-21 stem-loop. (A) Representative compound mixture from our library; five to seven compounds are mixed together in each mixture to increase throughput and provide intrinsic competition between fragments. The 1,2,4-thiadiazole fragment (red) was shown using this approach to be the best ligand out of our 420 compounds library, originally designed for protein enzymes. (B) 1D ¹H Line broadening experiments provide a fast and robust approach to identify binding compounds from a mixture (BI, starred) and rank compounds by using the fractional bound occupancy (1 – I_B/I_F) (BII, starred). The 1D ¹H line broadening experiments only require 1.25 min of NMR run time, and a total of ∼5 min for data to be collected on each NMR sample, including all manipulations (∼10 min total to compare free vs bound spectra). This approach dramatically improves screening efficiency compared to other ligand detect methods such as STD and WaterLOGSY, opening the door for screening libraries of the order of 10 000 compounds in the matter of days on a single NMR spectrometer. (C) Compounds that show fractional occupancy greater than 0.4 are shown to also bind to RNA via 2D NOESY screening. In this experiment, a NOESY spectrum is collected with a 600 ms mixing time on a 500 μM fragment; upon the addition of 10 μM pre-miR-21, the NOESY cross peaks change sign as the result of the increased T2, confirming binding to pre-miR-21.

The workflow to identify binding hits from the library starts with the collection of reference spectra for each fragment and ligand mixture is shown in Figure 2. All NMR samples are prepared at 100 μM concentration for each ligand in a “standard” screening buffer: 99.99% D₂O with 25 mM deuterated tris (pH 7.5), 2% d₆-DMSO and 11.1 μM DSA (4,4-dimethyl-4-silapentane-1-ammonium trifluoroacetate) as an internal reference, similar to DSS (4,4-dimethyl-4-silapentane-1-sulfonate).⁴⁴ After reference 1D ¹H spectra of each individual ligand and ligand mixture were recorded, a second set of samples were prepared in the same manner, but for the addition of 10 μM RNA. Spectra for the “bound” samples were then collected identically to the free reference spectra.

All data were collected using a standard water suppression 1D ¹H pulse sequence with 8K data points (to increase spectral resolution) and 64 scans at 298 K. These NMR screening conditions balance signal-to-noise, resolution and throughput. Total data collection time is approximately 5 min/sample, including sample changing and experimental setup (temperature equilibration, lock, tune/match/and shimming). All NMR spectra are processed in a uniform manner, using an ACD/Laboratories 1D NMR processor, to minimize any operator bias. Since the line-broadening experiments are collected in two different NMR tubes (free and bound), the total turnaround time per mixture is ∼10 min with our current instrument configuration (500 MHz Bruker DRX and 600 MHz Avance II, both instruments equipped with TCI cryoprobes).

The top image in Figure 3B shows an example of mixture screening results. Here, the black NMR signals are from the free ligands while the red NMR signals are recorded after pre-miR-21 was added. The decrease in ligand signal is specific only to compounds that bind, as identified by the asterisks (**). Compounds that do not bind the RNA show no change in ligand line width and very closely overlay the original free ligand spectra. At this triage stage of screening, every compound for which a change in ligand line width is observed is considered a hit and binned for follow up studies with individual compounds.

The results of this mixture screen step against pre-miR-21 yielded 17 unambiguous hits for further study (hit rate 4%). The chemical identity of each of the 17 hits are reported in Table 1. Given the binary (free/bound) nature of this 1D ¹H line broadening experiment, each mixture can be screened in ∼10 min. This screening approach increases throughput nearly 5-fold compared to traditional STD and WaterLOGSY experiments which typically run ∼60 min/sample using standard interleaved STD experiments on Bruker instrumentation. Using this binary line-broadening approach coupled with screening compounds in mixtures, our library of 420 compounds was screened and primary hits identified within only ∼9 h; without any sample or data handling automation. A library of 10 000 compounds can be straightforwardly screened within a week with modest implementation of automation (sample changers and liquid handling).

Table 1. Seventeen Fragments Identified in the First “Triage” Round of Screening, in a Binary Yes/No Fashion, from the Compound Mixture Library^a.

^aFollow up screening steps provided a qualitative ranking of the fragments, based on the fractional bound occupancy (1 – I_B/I_F). Compounds that show fractional bound occupancy greater than 0.4 are shown to bind to the RNA in NOESY experiments; therefore, a general cutoff of 0.4 is applied to separate and positively confirm RNA binding fragments.

After triage screening against the mixture library, we confirmed and prioritized hits for follow up by using ligand-detect NMR assays of each individual ligand. Experimentally, this is the same 1D ¹H line broadening method described above. In the bottom image of Figure 3B, the free reference spectrum for the ligand is represented in black while the bound ligand spectrum is shown in red. The chemical shifts for the individual compounds match the chemical shifts of compounds that change in the mixture, thereby confirming ligand identification (asterisks); proton assignments are labeled for reference on the thiadiazole fragment structure.

After free and bound spectra for each of the 17 individual ligands were collected, the total sum of each individual ligand peak intensity was measured relative to the internal reference standard DSA, to calculate I_F (free) and I_B (bound). When comparing hits to the same target, ligands with stronger binding will experience the largest changes in line width; this effect can be quantified as a fractional bound occupancy (B, eq 1), which represents a semiquantitative measure of binding affinity, to provide a rapid ranking and prioritization of ligands. The change in ligand line width is proportional to the molecular weight and shape of the target, along with the affinity of the ligand–RNA complex; therefore, binding compounds can be rank ordered for the same target with a simple binary approach without a full titration analysis to determine relative K_D’s,⁴³ as long as changes in the target hydrodynamic shape are not large. Once the fractional occupancy for each of the 17 ligands was calculated, the ligands were sorted by fractional occupancy (Table 1).

The binary (free/bound) approximation increases throughput while providing a qualitative rank order of affinity through (1 – I_B/I_F). However, it is important to establish what degree of line broadening is associated with a real hit, to avoid wasting NMR resources pursuing unproductive false positive or very weak hits. A generous approach claims any compound with a measurable change in binding occupancy at the individual screening step as a legitimate hit. Using this criterion, 11 compounds from our library bind the pre-miR-21 hairpin, while 7 identified in mixture experiments are revealed to be false positive. Each of these 11 compounds gave a measurable response in the individual ligand screening step (Table 1). The seven remaining compound showed poor solubility or did not show a measurable decrease in peak intensity upon RNA binding, when investigated individually.

To further triage compounds and eliminate very weak hits, we used an orthogonal ligand detected NMR assay. In this step, we collect NOESY spectra at long mixing time (600 ms) for each of the 11 fragments that gave a response in the individual ligand screening step. Here, a NOESY spectrum is collected of the free ligand at 500 μM concentration (Figure 3C left), followed by a second NOESY after the addition of 10 μM of pre-miR-21 (Figure 3C right). The NOE cross-peaks of unbound small molecules have opposite signs with respect to the diagonal; when a compound binds to a larger molecular weight molecule, the signal of the small molecule NOEs flips because the sign of the NOE effect flips at molecular weight above approximately 1000 in aqueous buffers.^36,41 Adding this step prior to target detected NMR experiments is optional, but its introduction further triages compounds, which is important because the target-detected experiments are most demanding. By adding this NOESY filter step, we further reduced out hit list to four compounds that gave positive responses in the mixture screening, the individual ligand screen and the NOESY conformation screening steps (Table 1). In practice, all four compounds had a binding occupancy greater than 0.4. We notice, however, that the binding occupancy score depends on the molecular weight of the target.⁴³ Care must therefore be used to determine this cutoff experimentally for each new target to be screened.

Based on the results of the ligand detected NMR screening filters, we nominated the compound with the largest fractional occupancy, 3-phenyl-5-(piperazin-1-yl)-1,2,4-thiadiazole, for further target-based NMR studies to identify the binding site for the fragment. First, we titrated the compound into a pre-miR-21 sample in aqueous buffer to monitor changes in base paring and secondary structure (Figure 4A). A 500 μM pre-miR-21 sample was titrated from 0:1, 1:1, to 2:1 ratios of ligand to RNA at 10C. We did not observe any significant changes in imino chemical shifts or the formation of new base pairs for pre-miR-21 upon ligand addition. However, we noticed peak broadening in the aromatic region of the spectrum. The apical loop of pre-miR-21 appears to be dynamic and in chemical exchange with solvent given the lack of any imino signals beyond G45. Therefore, the ligand could bind in the apical loop region without inducing significant chemical shift changes in the lower stem region, from which the NH signals originate.

Figure 4.

Target-detect experiments used to identify the fragment binding site on pre-miR-21. (A) 1D ¹H imino titration to monitor changes in base pairing: (I) Free pre-miR-21 RNA, (II) 1:1 complex, and (III) 2:1 complex. We do not observe any significant changes in chemical shifts for the imino signals, but the spectrum appeared broadened, particularly in the aromatic region around 8 ppm. (B) Binding site identification from 2D TOCSY spectra. The largest chemical shift changes were observed for U31, while C30 and C46 are broadened. (C) D₂O NOESY spectra at 37C identify two sets of intermolecular NOEs between the thiadiazole fragment and pre-miR21. An intense set of intermolecular NOEs was observed between the piperazine functional group (DRG HA/HB) and the G22 H8 proton. The second set of intermolecular NOEs links the U38, C41 and U43 H6 protons and the piperazine functional group (DRG HA/HB). (D). The second binding site is identified based on the observation of intermolecular NOEs between U31 H6 and the benzyl protons from the thiadiazole fragment (DRG HD).

We next collected TOCSY spectra of the free and bound RNA (Figure 4B). A 500 μM sample of pre-miR21 was titrated with 1 mM of the thiadiazole fragment. The chemical shift changes were mapped based on our previous work which provided assignments for the pre-miR-21 apical loop. The peak corresponding to U31 gave the largest chemical shift change upon adding the fragment. Other key changes were peak broadening for C46 and C30 and smaller shifts for U38, C41, and U43. All these changes are located within the upper stem and apical loop region of pre-miR-21, which would potentially explain the lack of changes in the imino signals. Lastly, we collected a NOESY of the 2:1 complex and identified seven unambiguous intermolecular NOEs between the thiadiazole fragment and pre-miR-21. One intermolecular NOE was not identified clearly since it overlaps with multiple H6/H8 protons (Figure 4C/D). The majority of intermolecular NOEs originate in the piperazine group protons.

The analysis of the NOESY spectra led to the identification of two binding sites for the thiadiazole fragment, one in the lower helix with intermolecular NOEs to G22 H8 and a second in the apical loop, with NOEs to U38, C41, and U43 H6 protons. The presence of multiple binding sites is not surprising. Small molecule binding to the terminal ends of RNA is commonly observed. Interestingly, the benzyl protons also have an intermolecular NOE to the U31 H6 proton, which, together with the intermolecular NOEs in the apical loop and piperazine protons, likely identifies a new ligand binding site for small molecules within pre-miR-21. The presence of intermolecular NOEs to the bases and the absence of intermolecular NOEs in the H1′ region of the spectra, places the fragment in the major groove, with the benzyl group located near U31 and the piperazine group pointing toward the apical loop (U38, C41, U43). Unfortunately, the low proton density for the 1,2,4-thiadiazole makes it difficult to fully establish the structure of the complex, but we can anticipate it has to be proximal to the Dicer cleavage site nucleotides A29 and C46, and therefore to have the potential to interfere with enzymatic activity.

In summary, we have demonstrated the application of a scalable approach for NMR-based fragment screening with an RNA that has been refractory to the identification of high-quality small molecule hits. Identification of high quality fragments requires in general good screening methods and screening library. In this Letter, we directly address the screening approach. We show the use of a 1D ¹H line broadening approach to triage binding compounds from mixtures increases throughput. This approach can detect fragments with affinity K_D of ∼10 μM to 1 mM and is target agnostic and does not need to be changed or optimized to any new target. Using this approach, we were able to quickly (9 h of NMR time) screen a small library of 420 fragments to identify 4 binding compounds, including 1 compound with intermolecular NOEs to residues near the Dicer cleavage site. With improvements in sample automation and nonuniform sampling approaches for 2D target detect experiments, this approach is easily scalable to much larger focused libraries of up to 10 000 compounds in approximately a week of spectrometer time. Thus, while this NMR approach is not suitable for high-throughput screening, it is an excellent method to evaluate focused fragment libraries.

The current library was designed to target thymidylate synthase, an enzyme that converts dUMP to dTMP. Given the enzyme preference for nucleoside ligands, this library may shed some light on how to design focused libraries for RNA targeting. While hit rate is low, pre-miR-21 is historically a very difficult target because of the flexibility of the apical loop. Our approach identified fragments that bind the apical loop and distinguish closely related compounds. For example, a related compound (3-phenyl-1,2,4-thiadiazol-5-amine), missing the piperazine ring, does not interact at all with the pre-miR-21 hairpin (Figure S1).

Glossary

Abbreviations

NMR

Nuclear Magnetic Resonance

RNA

Ribonucleic Acid, NOE, Nuclear Overhauser Effect

TOCSY

Total-Correlation Spectroscopy;

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsmedchemlett.1c00109.

• detailed description of the experimental materials, methods and procedures (PDF)

• Table of compound structures (XLSX)

• Mixture library, searchable table of verified hits (PDF)

Special Issue

Published as part of the RNA: Opening New Doors in Medicinal Chemistry special issue.

Author Contributions

MDS designed and executed experimental workflows, data collection, and analysis. The manuscript was written through the contributions of all authors. All authors have given approval to the final version of the manuscript.

The authors declare the following competing financial interest(s): M.D.S. and G.V. are co-founders of both Ithax Pharmaceuticals and Ranar Therapeutics.