Optimization of a Protein-Targeted Medicine into an RNA-Specific Small Molecule

Abstract

Protein-targeted small molecule medicines often bind RNAs and affect RNA-mediated pathways in cells. Historically, small molecule engagement and modulation of RNA have not been considered in medicine development; however, RNA should be considered both a potential on- and off-target. Kinase inhibitors have emecrged as common RNA binders with dovitinib, a classic receptor tyrosine kinase (RTK) inhibitor, inhibiting RTKs and the biogenesis of oncogenic microRNA-21 through direct engagement. In this study, we use knowledge of the molecular recognition of both protein and RNA targets by dovitinib to design molecules that specifically inhibit the RNA target but lack activity against canonical protein targets in cells. As it is now becoming apparent that RNA can be both an on- and off-target for small molecule medicines, this study lays a foundation to use design principles to maximize desired compound activity while minimizing off-target effects.

Graphical Abstract

Small molecule drug discovery is generally focused on affecting proteins and their associated pathways. (1,2) RNA, however, is emerging as a small molecule drug target. (3,4) Indeed, known small molecule medicines that bind protein also bind RNA and modulate its function. (5-7) Kinase inhibitors, (5,6) topoisomerase inhibitors, (5) and a serotonin reuptake inhibitor (7) bind to RNA and affect RNA-mediated processes, in addition to their canonical protein targets (Figure 1). In addition to those studies, work by our laboratory has demonstrated that known protein-targeted medicines bind RNA via unbiased transcriptome-wide studies and have also identified individual RNA targets. (6) Moreover, a more recent study used RNA 2-hydroxyl acylation probes to study the binding of three protein-binding small molecules transcriptome-wide, further supporting the notion that medicines indeed bind cellular RNA. (8) Historically, RNA has not been considered as an on- or off-target in the development of small molecule medicines, and thus could explain in part challenges associated with the drug discovery process. (9)

Figure 1.

Known medicines that bind and modulate RNA. Structures of protein-targeted medicines that also bind RNA and modulate its function in cells. (5-7) UMAP representation of chemical space covered by kinase inhibitors (red) compared to that of known RNA binders (gray). Dovitinib is indicated in cyan.

The purpose of this study is to develop a framework for the design of small molecules that bind specifically to RNAs. We assessed the potential of protein-targeted medicines to bind RNA based on chemical features of known RNA binders. We mapped the chemical space of kinase inhibitors against known RNA binders, extracted from a database of all publicly available RNA-binding small molecules (Inforna), (10) using UMAP (uniform manifold approximation and projection) analysis. (11) This showed an overlap between the chemical space of these two groups of molecules (Figure 1), suggesting that there are common properties that influence binding to both RNA and protein. The data also suggested that kinase-targeting small molecules might be modified to augment RNA or protein binding selectively.

The receptor tyrosine kinase (RTK) inhibitor, dovitinib, also inhibits processing of the oncogenic microRNA-21 precursor (pre-miR-21); both activities are the result of direct target engagement (Figures 1 and 2A,B). (6) RTKs are transmembrane receptors that initiate signal transduction pathways and regulate cell proliferation, differentiation, and migration. (12) They are dysregulated in cancers and are a target for anticancer medicines. (13) Like all micro (mi)RNAs, miR-21 is produced by RNA polymerase II as a primary (pri-) miRNA. Pri-miRNAs are processed stepwise first in the nucleus by Drosha, here affording pre-miR-21, (14) and then in the cytosol by Dicer. The resultant mature miRNA is loaded into the RNA-induced silencing complex (RISC) to regulate gene expression. (14) Mature miR-21 translationally represses various messenger (m)RNAs that suppress apoptosis and activate the invasive nature of cancer cells, including phosphatase and tensin homolog (PTEN). (15)

Figure 2.

Design of structural derivatives of 1 to obtain compounds specific for pre-miR-21. (A) Schematic of 1 (red sphere) inhibition of receptor tyrosine kinase (RTK) activation and signaling in cancer. (B) Schematic of 1 (red sphere) inhibition of miR-21 biogenesis and its relation to cancer. (C) Top: Interaction network of 1 bound to fibroblast growth factor receptor 1 (FGFR1; PDB 5A46). Hydrogen bonds (solid lines) and hydrophobic interactions (dashed lines) stabilize 1 within the binding pocket of FGFR1. Bottom: Molecular docking of 1 against a model of the Dicer processing site harbored within pre-miR-21 (ΔG₃₇^o = −9.03 kcal/mol). The N-methyl-1,4 piperazine group is exposed to solvent, while the aromatic benzimidazole and the 4-amino-5-fluoroquinolin-2(1H)-one form stacking interactions (solid lines) and hydrogen bonds (dashed lines) with the RNA. FGFR1 is shown in ribbon representation; pre-miR-21 is shown in stick and the surface representations; and 1 is shown in stick representation. (d) Schematic of derivatization of 1. The 4-amino-5-fluoroquinolin-2(1H)-one was modified to maximize interactions with the RNA and reduce binding to RTKs.

It was previously discovered that 1 (dovitinib) bound RNA (6) using a selection-based library-vs-library method dubbed 2-dimensional combinatorial screening (2DCS). (4) In brief, the RNA-binding landscape of all small molecules in Calibr’s Repurposing, Focused Rescue, and Accelerated Medchem (ReFRAME) library, of which 1 is a member, were defined by selecting their preferred RNA structures from an RNA library. (6) Notably, ReFRAME is a drug repurposing library that comprises small molecules that have reached clinical trials or undergone significant preclinical profiling. (16) The selectivity of 1 was reprogrammed for pre-miR-21 over RTKs by its conversion to a heterobifunctional ribonuclease targeting chimera (RiboTAC). (6) The RiboTAC binds and recruits ribonuclease L (RNase L) to pre-miR-21, effecting its degradation and inhibiting its downstream pathways in cells and in mouse models of cancer and kidney fibrosis, both driven by aberrant expression of the miRNA. (6) Indeed, this functionalization changed the selectivity of the molecule to favor the RNA by 2500-fold. (6) Herein, we designed derivatives of 1 that improved RNA specificity by implementing structural biology and medicinal chemistry approaches.

The three-dimensional (3D) structure of 1 (Figure 2C, top) bound to the RTK fibroblast growth factor receptor 1 (FGFR1) (PDB 5A46), (17) molecular docking analyses, and previously reported structure–activity relationships (SAR) (18) were used to identify molecular features that drive binding of 1 to RTKs. Combined with knowledge of features that favor RNA binding, the scaffold of 1 was molecularly edited to reduce kinase inhibition while maximining activity against pre-miR-21 (Figure 2C, bottom).

The structure of the 1–RTK complex showed that 1 binds the ATP-binding pocket with the N-methyl-1,4 piperazine group projected into the solvent while the aromatic benzimidazole and the 4-amino-5-fluoroquinolin-2(1H)-one interact with the protein; 1 occupies ~273 Ų of the binding pocket’s available surface area (Supporting Information, Figure S1A). The bound pose was stabilized by multiple hydrogen bonds (Leu484, Glu562, and Ala564) as well as hydrophobic interactions (Val561, Ile545, and Leu484) (Figure 2C, top). The interaction of 1 with Val561 is a rare interaction observed in only 15 of 490 human protein kinases. (19)

Within pre-miR-21, the binding site for 1 is harbored in the Dicer processing site, which forms an A bulge (5′GAC/3′C_G). Molecular docking of 1 into a model of the Dicer site of pre-miR-21 (5′UGUUGACUGUCG/3′ACAAC_GGUAGC) showed the N-methyl-1,4 piperazine group is exposed to solvent, while the aromatic benzimidazole and the 4-amino-5-fluoroquinolin-2(1H)-one form stacking interactions and hydrogen bonds with the closing base pairs of the A-bulge (ΔG₃₇^° = −9.03 kcal/mol) (Figure 2C, bottom; Supporting Information, Table S2). Additionally, 1 occupied ~317 Ų of the RNA binding pocket (Supporting Information, Figure S1B). (20)

Using these two complexes as a guide, we designed a panel of derivatives of 1 (n = 16; Figures 2D and 3A, and Supporting Information, Figure S2) that (i) have less favorable interactions with the RTK, (ii) introduce steric bulk in positions where the molecule is closely buried within the ATP-binding pocket of the RTK, and (iii) introduce functionalities that could confer enhanced binding affinity for pre-miR-21 (Figures 2D and 3A, and Supporting Information, Figure S2).

Figure 3.

Derivative 6 has reduced antikinase activity and increased RNA affinity compared to 1. (A) Chemical structures of derivatives of 1 and their ability to derepress miR-21’s downstream target, phosphatase and tensin homolog (PTEN), using a luciferase reporter transfected into MDA-MB-231 cells (n = 10 for vehicle; n = 8 for 1; n = 5 for derivatives of 1). (B) Left: The effect of 4, 5, and 6 on the activity of the receptor tyrosine kinase (RTK) fibroblast growth factor receptor 1 (FGFR1) was assessed in vitro (n = 3 independent experiments with two technical replicates each). Right: Effect of 1 and 6 on the activity of the RTK fms-like tyrosine kinase 3 (FLT3) in vitro (n = 3 independent experiments with two technical replicates each). (C) Blind docking of 5 (left, blue) and 6 (right, blue) against FGFR1 shows binding to a different pocket compared to 1 (green), which binds in the ATP-binding site. FGFR1 is shown in ribbon, and 1, 5, and 6 are shown in sphere representation. (D) Secondary structures of Dicer site RNA mimics used in binding studies and corresponding binding curves (n = 2 independent experiments with three technical replicates each). **, p < 0.01; ***, p < 0.001; ****, p < 0.0001, as determined by a one-way ANOVA with multiple comparisons. Data are reported as the mean ± SD.

The collection of compounds was screened for binding to pre-miR-21 by AbsorbArray, a microarray-based method where compounds are adhered to the surface and incubated with radioactively labeled RNA (Supporting Information, Figure S3A, Table S1). (5) Of the 16 derivatives, 12 molecules dose-dependently bound to pre-miR-21; seven gave greater signal on the microarray surface than 1 (Supporting Information, Figure S3B). All 17 compounds (1 and its derivatives) were docked against pre-miR-21, revealing that binders have on average a more favorable free energy of binding compared to nonbinders, by −2.5 kcal/mol (Supporting Information, Table S2). Notably, the two molecules with the highest radioactive signals, 10 and 11, also have the most favorable docking scores (ΔG₃₇^° = −15.05 kcal/mol and −14.28 kcal/mol, respectively) (Supporting Information, Table S2).

The RNA-binding derivatives (n = 12) were next evaluated in a luciferase reporter assay that measures derepression of PTEN, one of miR-21’s downstream targets, (15) in MDA-MB-231 triple negative breast cancer (TNBC) cells. (Note, this reporter contains the firefly luciferase coding region fused to the 3′ untranslated region (UTR) of PTEN, which is responsive to both miR-21 and miR-17 abundance, as previously demonstrated. (21)) MDA-MB-231 cells overexpress miR-21 and evade apoptosis by suppression of PTEN. (15) Of the 12 candidates, nine did not affect the viability of MDA-MB-231 cells at concentrations ≤5 μM (Supporting Information, Figure S4). Of these nine derivatives, 4, 5, and 6 derepressed PTEN to a similar extent as parent compound 1 (by ~60–75% at 5 μM, as normalized to Renilla luciferase; Figure 3A). Derivatives 2 and 3 increased luciferase activity to a lesser extent, and this increase was not statistically significant (Figure 3A). Together, these data suggest that 4, 5, and 6 inhibit biogenesis of pre-miR-21 and derepress PTEN.

Interestingly, while dovitinib (1) inhibited FGFR1 in vitro with an IC₅₀ of 0.04 ± 0.01 μM, consistent with previous reports, (22) 4, 5, and 6 inhibited FGFR1 much less potently with IC₅₀s of 5 ± 1 μM, >10 μM, and >10 μM, respectively (Figure 3B, left). While some data support the exocyclic amine of 1 is important for inhibition of kinase activity, (23) our analysis of the complex between FGFR1 and 1 (PDB 5A46) did not identify the exocyclic amine as having a key interaction with the protein (Figure 2C, top). While 4 lacks the exocyclic amine, it maintained a degree of RTK inhibitory activity, albeit less potently than 1. Docking studies suggest that 4 binds the RTK’s ATP-binding pocket, however, the interaction pattern is different than that of 1, affording a different structural rearrangement in the ATP binding pocket. In particular, a salt-bridge forms between Glu571 and the piperazine moiety (Supporting Information, Figure S5A) in the presence of 4 (ΔG₃₇^° = −8.84 kcal/mol) but not 1 (ΔG₃₇^° = −8.46 kcal/mol). (17) Compounds 5 and 6 have additional steric bulk that cannot be accommodated in the binding pocket. Blind docking studies show that 5 (ΔG₃₇^° = −10.56 kcal/mol) and 6 (ΔG₃₇^° = −10.14 kcal/mol) bind to FGFR1, but at sites other than the ATP-binding pocket (Figure 3C, and Supporting Information, Figure S5B,C). These results support the compound design hypothesis and correspond to a 125-fold reduction in potency between 4 and 1 and a >250-fold reduction for 5 and 6 compared to 1 (Figure 3B, left).

As 5 and 6 had >250-fold reduced potency for RTK inhibition and similar activities for derepressing PTEN at a single dose (5 μM), they were reassessed for miR-21-dependent derepression of PTEN in dose response, where the luciferase reporter was mutated to only contain the miR-21 recognition site. (21) While both 5 and 6 dose-dependently derepressed PTEN translation, 6 showed the most potent effect, derepressing PTEN to a greater extent than 1 (Figure 4A) and was carried forward for additional in vitro and cellular characterization.

Figure 4.

Derivative 6 inhibits pre-miR-21 biogenesis in MDA-MB-231 cells and reduces an invasive cellular phenotype while having no effect on RTKs. (A) Effect of 1, 5, and 6 on PTEN expression, as assessed using a luciferase reporter in MDA-MB-231 cells (n = 13 for vehicle; n = 10 for 1; n = 5 for 5- and 6-treated samples. (B) Representative Western blot and quantification of extracellular signal-regulated kinase (ERK) and phosphorylated (p)ERK abundance in MDA-MB-231 cells treated with 1 or 6 (n = 3). (C) Effect of 6 on the biogenesis of miR-21, as assessed by measuring mature and pre-miR-21 abundance in MDA-MB-231 cells by RT-qPCR (for studies of mature miR-21, n = 3 for LNA-treated samples and n = 5 for vehicle- and 6-treated samples; n = 4 for pre-miR-21). (D) Representative images and quantification of invading MDA-MB-231 cells in Boyden chamber assay with or without treatment with 6, an LNA antagomir targeted to miR-21 (LNA-21), or a scrambled antagomir (Scr. LNA) (n = 6 biological replicates for vehicle-treated samples with four images analyzed per replicate; n = 3 biological replicates for 6- or LNA-treated samples with four images analyzed per biological replicate). **, p < 0.01; ***, p < 0.001; ****, p < 0.0001, as determined by a one-way ANOVA with multiple comparisons. Data are reported as the mean ± SD.

The affinities of 1 and 6 for binding to a model of pre-miR-21’s Dicer site (Figure 3D) were measured by microscale thermophoresis (MST) or fluorescence quenching. (24) The Dicer site model comprises nt 18–42 of pre-miR-21, which includes all nucleotides endogenous to the precursor miRNA’s apical loop (Supporting Information, Table S1). In these experiments 1 bound to the Dicer site with a K_d of 4.0 ± 0.2 μM (in agreement with a previous report) (6) while 6 bound with a K_d of 1.5 ± 0.1 μM (Figure 3D). [Note that the affinities of 5 and 6 are similar (Figure 3D).] Further, 6 bound specifically to the Dicer site, as studied by measuring its affinity for RNAs with various point mutations. Here, the A bulge in the Dicer site and an adjacent U bulge were individually or simultaneously mutated to base pairs (Figure 3D, Supporting Information, Table S1). Both bulges were required for saturable binding, as all three mutants had binding curves that did not saturate (Figure 3D). Thus, 6 was specific for pre-miR-21’s Dicer site in vitro.

Binding of 1 and 6 was also assessed by nuclear magnetic resonance (NMR) spectroscopy, both by water–ligand observed via gradient spectroscopy (WaterLOGSY) (25) and saturation transfer difference (STD) NMR spectroscopy (26) (Supporting Information, Figure S6 and Table S1). The appearance of negative signals in the WaterLOGSY spectra (Supporting Information, Figures S6A-C) as well as changes in peak signal intensity in STD spectra confirmed binding of both small molecules to the model of pre-miR-21’s Dicer site (Supporting Information, Figure S6A,D,E).

Compound 6 was then docked into a model of pre-miR-21’s Dicer site and compared to the docked model of 1. Compound 6 bound to pre-miR-21 with a ΔG₃₇^° of −12.15 kcal/mol, > −3.0 kcal/mol more stable than 1. This change in free energy is driven by additional hydrogen bonds that 6 formed with the RNA backbone and strong stacking interactions formed with neighboring base pairs (Supporting Information, Figure S7A and Table S2). Additionally, 6 occupied a larger surface within the pre-miR-21 binding pocket than did 1 (~368 Ų and ~318 Ų, respectively; Supporting Information, Figure S7B). Molecular dynamics (MD) simulations (1 μs long) verified the stability of 1 and 6 in the pre-miR-21 binding pocket (Supporting Information, Movie S1, Movie S2, and Tables S3 and S4). Together, these data indicate 6 displays better shape complementarity and specificity toward pre-miR-21 than 1 and provides further support for the observed changes in selectivity.

Of note, binding studies showed that the presence of the U-bulge was required for saturable binding, which was not present in the RNA model used for docking studies. As the predicted U-bulge in pre-miR-21 is separated from the apical loop by a single base-pair, there is evidence that the structure is dynamic, interconverting from the U-bulge to a larger hairpin loop. (27) Additional evidence suggests the U-bulge influences the structure of the A-bulge binding site through a structural rearrangement. (27) Converting this U-bulge to an AU pair therefore has the potential to alter the observed dynamics, as well as the structure of the A-bulge binding site, thereby influencing molecular recognition.

We assessed the cellular specificity of 6 for activity against pre-miR-21 over the canonical RTK targets in MDA-MB-231 TNBC cells, where miR-21 is aberrantly expressed, by measuring: (i) derepression of PTEN (above), (ii) inhibition of miR-21 biogenesis, and (iii) inhibition of RTK downstream signaling, particularly the phosphorylation of extracellular signal-regulated kinase ((p)ERK). As expected, 1 dose-dependently inhibited phosphorylation of ERK (pERK) while not influencing ERK levels, as previously reported, (6) with an IC₅₀ of ~1 μM. In contrast, 6 had no effect on pERK or ERK abundance up to a concentration of 5 μM (Figure 4B), in contrast to the dose-dependent derepression of PTEN (Figure 4A). As further confirmation of these cellular RTK inhibition studies and in vitro inhibition of FGFR1 (Figure 3B, left), 6 also had reduced potency for inhibition of the RTK fms-like tyrosine kinase 3 (FLT3). While 1 inhibited FLT3 activity with an IC₅₀ of 0.004 ± 0.002 μM, in agreement with previous reports, (28) 6’s IC₅₀ was 1.7 ± 0.4 μM, a 425-fold reduction in potency (Figure 3B, right).

The derepression of PTEN by 6 in the luciferase reporter assay (Figure 4A) suggested that 6 inhibits miR-21 biogenesis. Indeed, 6 reduced miR-21 abundance while increasing pre-miR-21 abundance, supporting a mechanism of action in which 6 inhibits miR-21 biogenesis (Figure 4C, and Supporting Information, Table S5). Inhibition of miR-21 biogenesis and derepression of PTEN were sufficient to reduce the invasive phenotype of MDA-MB-231 cells, a miR-21-driven breast cancer phenotype, at a 5 μM dose (Figure 4D). (29)

To study if the observed reduction in invasion was due to inhibition of pre-miR-21 biogenesis and not RTK activity, we induced an invasive phenotype in MCF-10A cells, a model of healthy breast epithelial cells, by forced expression of WT pre-miR-21 or a mutant in which the 6-binding site was ablated (A bulge is mutated to a base pair; Supporting Information, Figure S8). This phenotype was rescued by 6 in cells expressing WT pre-miR-21 but not in cells overexpressing the mutant (Supporting Information, Figure S8). In addition to supporting that the rescue of the cellular phenotype is driven by inhibition of pre-miR-21 biogenesis and not anti-RTK activity, these studies demonstrate 6’s specificity for the pre-miR-21 Dicer site.

This study demonstrates that molecular editing, guided by modeling and analysis of high dimensional chemical space, can be utilized to fine-tune the selectivity of a dual-targeting compound (i.e., both protein and RNA) toward a single target. While RiboTAC functionality was previously reported as a strategy for modifying the selectivity of dovitinib (1) toward RNA, (6) this study improved specificity by medicinal chemistry optimization, thereby preserving the drug-like properties of the parent compound.

Herein, we showed that a medicine designed to target proteins also binds RNA and can be optimized to selectively bind a single target by iterative design and synthesis studies. This study, as well as previous studies, show that small molecules can directly bind to and affect RNA function (5-7) and also provides a guide for how to systematically evaluate activity against both protein and RNA targets, properties that should be equally considered in the drug-development process to alleviate off-target effects.