Small-molecule targeting of MUSASHI RNA-binding activity in acute myeloid leukemia

The MUSASHI (MSI) family of RNA binding proteins (MSI1 and MSI2) contribute to a wide spectrum of cancers including acute myeloid leukemia. We find that the small molecule Ro 08–2750 (Ro) binds directly and selectively to MSI2 and competes for its RNA binding in biochemical assays. Ro treatment in mouse and human myeloid leukemia cells results in an increase in differentiation and apoptosis, inhibition of known MSI-targets, and a shared global gene expression signature similar to shRNA depletion of MSI2. Ro demonstrates in vivo inhibition of c-MYC and reduces disease burden in a murine AML leukemia model. Thus, we identify a small molecule that targets MSI’s oncogenic activity. Our study provides a framework for targeting RNA binding proteins in cancer.


Results
Ro binds to MSI2 and inhibits its RNA-binding activity. In order to identify a putative MSI RNA binding antagonist, we previously performed a fluorescence polarization (FP)-based screen using recombinant MSI1 and MSI2 and a consensus target RNA with a library of 6208 compounds 36 . We selected Ro 08-2750 (Ro) based on its RNA-binding inhibition of both MSI1 and MSI2 36 . MSI2 RNA-binding inhibition was confirmed by FP (IC 50 of 2.7 ± 0.4 μM) (Fig. 1a). We then used a chemiluminescent Electrophoresis Mobility Shift Assay (EMSA) to quantify MSI2-RNA complexes in vitro. MSI2 binding could be competed with unlabeled RNA or by increasing concentrations of Ro (Fig. 1b, c). We then performed MicroScale Thermophoresis assay (MST) and found that Ro directly interacted with MSI2 with a K D of 12.3 ± 0.5 μM (Fig. 1d). This interaction was then narrowed down to just the RNA-recognition motif 1 (RRM1), (Supplementary Fig. 1a). Also, the interaction with MSI2 could be competed with the addition of target RNA (K D of 27.5 ± 2.6 μM). These data suggest that Ro directly interacts with the MSI2 RRM1 and competes with RNA binding.
Ro interacts with the RNA recognition site of MSI2 RRM1. To study how Ro interacts with MSI2, we obtained a crystal structure of the apo human MSI2 RRM1 at 1.7 Å resolution ( Table 1, RCSB PDB accession code 6DBP) after unsuccessful co-crystallization attempts. We performed docking analysis to identify a putative binding region (Fig. 2a, b and Supplementary Fig. 2a, b). Based on Ro's ability to compete for MSI-RNA complexes, we hypothesized that the binding site is likely to be shared with the RNA binding site. Thus, we predicted a stacking interaction with F66 and R100 with Ro (Fig. 2b). Also, the K22 side chain and the NH backbone group from F97 formed stabilizing H-bonds with the oxygens from the aldehyde in the C 2 position in the pyrimidine ring and the aldehyde from the opposite ring (Fig. 2b, c). A 2D representation indicates that the R100 forms a non-covalent π-cation and two H-bonding interactions with K22 and F97 (Fig. 2c).
To directly test our docking model, we mutated the putative interacting residues (K22A, F66A, F97A and R100A) and determined their ability to bind to Ro. This resulted in significantly reduced binding (F97A K D 69.5 ± 14.7 μM, R100A K D 148 ± 65 μM, F66A and K22A K D > 200 μM compared with 10.5 ± 0.3 μM for WT) (Fig. 2d). The triple mutant (F66A/F97A/ R100A) was incapable of binding to Ro. Importantly, the single mutations did not disrupt RNA binding to MSI2 whereas the triple mutant completely inhibited its activity (K D > 50 μM) ( Supplementary Fig. 2c). These data support our docking since the single mutants demonstrated reduced Ro binding activity without altering RNA binding.
To test structure-activity relationships, we evaluated two Ro related molecules (Ro-NGF and Ro-OH, Fig. 2f). The first analog, Ro-NGF, previously synthesized and described by Eibl et al. 39 , was selected to determine if Ro's activity was related to its anti-Neural Growth Factor (NGF) activity, as this compound showed the highest affinity for neural growth factor (NGF) (K D [NGF] = 1.7 μM) in its compound series (Supplementary Data 1). The second analog, Ro-OH, was synthesized by reduction of the Ro aldehyde to the corresponding alcohol, providing a single alteration to the structure ( Supplementary Fig. 3a-c). Initial structural analysis of our docked model suggested that the Ro-OH-MSI complex lacks the R100 π-cation interaction and that Ro-NGF binds MSI in a position displaced from the RNAbinding core ( Supplementary Fig. 4a-d) as compared with Ro.  We computed similar binding free energies (ΔG bind ) for the three ligands (Ro, Ro-OH and Ro-NGF, −5.5 and −6.1 versus −5.1 kcal mol −1 for Ro-NGF) ( Supplementary Fig. 4e). Both MSI2 and ligands adopted a heterogeneous ensemble of conformations and binding poses, with the protein-ligand complex predicted to undergo a slight conformational change upon binding of Ro and Ro-OH ( Supplementary Fig. 4f). Free energy calculations for all three small-molecules suggest that the Ro-NGF-MSI complex adopts a much more diverse set of conformations (as measured by conformational clustering of the fully interacting alchemical state) than the complexes with Ro-OH or Ro (Fig. 2g). The Ro-MSI complex demonstrated the fewest clusters, with the top three clusters accounting for 92.7% of the sampled configurations ( Supplementary Fig. 4g). The Ro-OH-MSI complex showed a larger number of clusters, with the four clusters accounting for 49.1% of sampled configurations, indicating a greater degree of heterogeneity than Ro ( Supplementary Fig. 4h).
To experimentally validate these predictions, we performed EMSA of GST-MSI2 competing RNA with Ro-OH or Ro-NGF, comparing their potency with Ro and unlabeled RNA as positive controls. Accordingly, Ro-OH had reduced activity compared with Ro (∼30-40% versus 65-75%, p < 0.05), whereas Ro-NGF was completely unable to disrupt MSI2-RNA complexes (Fig. 2h, i). These results were confirmed by a FP assay with Ro-OH inhibiting with 12.5-fold less potency than Ro, and Ro-NGF failing to inhibit RNA-binding activity ( Supplementary Fig. 4i). Furthermore, in MST assays, Ro-OH showed a 27-fold lower affinity than Ro (K D of 302.0 ± 119 μM for Ro-OH versus 11.2 ± 0.6 μM for Ro) for GST-MSI2, whereas Ro-NGF failed to demonstrate any interaction up to 500 μM (Fig. 2j). Thus, our structural and biochemical experimental data support the conclusion that Ro and MSI2 interact via the RRM/ RNA binding site and that the drug can displace RNA from its binding site, thus likely inhibiting MSIrelated RNA regulation.
Ro demonstrates activity in murine MLL-AF9 leukemic cells. To test the cellular effect of Ro in a murine AML leukemia model, we used previously established MLL-AF9 expressing leukemic bone marrow (BM) cells from secondary transplants 7 . Consistent with an on-target effect on MSI inhibition and in agreement with the RNA-binding activity inhibition assays, Ro effectively inhibited leukemia cell proliferation (half-effective concentration, EC 50 = 2.6 ± 0.1 μM). By comparison, the analogs that failed to interact with MSI2 had a diminished effect (Ro-OH EC 50 = 21.5 ± 0.8 μM; Ro-NGF > 50 μM), suggesting that the antiproliferative effect is due to the ability of Ro to inhibit MSI2 RNA binding-activity (Fig. 3a). Treatment of cells with Ro resulted in an increase in differentiation at 5 μM dose and 48 h treatment as seen both quantitatively by flow cytometry (Fig. 3b) and by morphological analysis (Fig. 3c). We found a significant increase in apoptosis (Annexin V+ population as early as 8 h both at 5 and 10 μM) with the highest increase at 48 h and 10 μM Ro ( Fig. 3d and Supplementary Fig. 5a). We then assessed how MSI2 overexpression affected the plating capacity of MLL-AF9+ BM cells in the absence or presence of Ro. MSI2 overexpressing cells formed 50% more colonies than control cells transduced with an empty vector. Treatment of cells with Ro resulted in reduced colony formation in control cells by >50% and ∼75% at 1 μM and 5 μM concentrations, respectively. Furthermore, MSI2overexpressing leukemia cells demonstrated reduced activity at these doses (Fig. 3e). Consistently, MSI2's translational direct targets 7,30 SMAD3, c-MYC, and HOXA9 were reduced, whereas their abundance remained unaffected in cells that overexpressed MSI2 (Fig. 3f). Moreover, colony-forming ability was further rescued by overexpressing MSI2-Ro-binding mutants (K22A, F66A, F97A, and R100) ( Supplementary Fig. 5b, c). Overall, these data further support that MSI2's cell toxicity was related to MSI2 and its RNA binding activity. We also found that Ro blocked MLL-AF9+ BM colony formation at concentrations that did not affect the plating efficiency of normal Lin-Sca+ cKit+ (LSK) cells, indicating a potential therapeutic window between normal and malignant cells (Fig. 3g).

Ro inhibits survival of human AML lines and patient cells.
To determine if Ro has activity against human myeloid leukemia, we first tested cytotoxicity effects in MOLM13 (AML, MLL-AF9+) and K562 (CML-BC, BCR-ABL+) cell lines 4,27 . Similar to the mouse leukemia cells, Ro demonstrated an anti-proliferative effect (EC 50 ∼8 μM), whereas the two analogs (Ro-OH and Ro-NGF) revealed >4-fold weaker potency. Ro affected viability of CD34+ cord blood cells at an EC 50 of ∼22 μM, 2.6-fold higher concentration than the human leukemia cell lines ( Fig. 4a and Supplementary Fig. 6a). Ro induced myeloid differentiation and promoted apoptosis in both K562 and MOLM13 cells based on flow cytometry and morphology (Fig. 4b-d and Supplementary  Fig. 6b-d) without any effect on differentiation on normal CD34 + cord blood cells ( Supplementary Fig. 6e, f). Plating activity was >80% inhibited at the 20 μM Ro dose in the human AML cell lines (Fig. 4e). In addition, Ro demonstrated differential sensitivity in three AML patient samples in colony plating assays compared with normal human CD34+ cord blood cells (>50% inhibition in colony numbers at 5 μM compared with only a modest reduction at 20 μM Ro ( Fig. 4f and Supplementary Data 2). These results suggest that Ro can induce toxicity in human myeloid leukemia cells with a (2-fold) level of selectivity compared with normal cells.
Ro inhibits MSI2 RNA-binding and alters MSI2 gene signature. To further investigate the effect and mechanism of action of Ro, we initially performed RNA immunoprecipitation (RNA-IP with FLAG) experiments on K562-MIG (empty vector) and K562-FLAG-MSI2 (MSI2 overexpressing) cells (Fig. 5a). By incubating the drug at 10 μM (∼EC 50 ) for 1 h with the cells, we could detect a significant decrease in MSI2 mRNA binding targets (TGFBR1, c-MYC, SMAD3, CDKN1A) (Fig. 5b). These data suggest that Ro can block MSI2 binding to target mRNAs in a cellular context at a short time-point. To globally assess the proximal effect of Ro treatment on the transcriptional program, we then performed RNA-sequencing on MOLM13 and K562 cells after 4 h of treatment. Ro incubation   30,40 . Interestingly, we observed an overlap of MSI-associated signatures from our previous dataset and enrichment with MSI1 direct mRNA targets from the intestine (Supplementary Fig. 7a and Supplementary Data 6-11) 4 . Moreover, we observed a~70% overlap of the functional pathways between each individual cell line and the pathways altered after shRNA depletion of MSI2 (Fig. 5d). Among these shared pathways, 76% (543 out of 717) overlapped in MOLM13 compared with K562 cells treated with Ro, which included c-MYC, mRNA-related, and leukemia-associated gene sets ( Fig. 5d and Supplementary Data 12). Thus, Ro treatment after a short administration recapitulated a large portion of the MSI2associated gene expression program.
To determine how Ro affects previously determined MSI targets, we treated both K562 and MOLM13 cells with increasing concentrations of Ro (up to 20 μM at 4 h). In previous studies, MSI was demonstrated to maintain the protein levels of TGFβR1, c-MYC, SMAD3, and HOXA9 7,30 while suppressing P21 abundance 41 . Consistent with this, we observed a significant dose dependent reduction of TGFβR1, c-MYC, SMAD3, HOXA9 and an increase P21, while the non-target control β-ACTIN remained unchanged (Fig. 5d, e). In addition, Ro could inhibit MSI2 targets in a time-dependent manner with c-MYC, a short half-life protein, being reduced in 1 h of treatment ( Fig. 5f, g). In support of Ro altering translation of specific MSI2 targets but not generally inhibiting global translation, we found equivalent global protein synthesis after drug treatment as assessed by Opropargyl-puromycin incorporation ( Supplementary Fig. 7b). As previously noted by RNA-sequencing, there were modest effects on the mRNA expression of MSI2 targets by qPCR ( Supplementary Fig. 7c) suggesting that Ro mainly influences its direct targets through a post-transcriptional mechanism. Thus, these results support our hypothesis that Ro acts in the MSIrelated translational program.
Ro inhibits leukemogenesis in a myeloid leukemia model in vivo. Finally, we sought to determine if Ro has in vivo efficacy using an aggressive murine MLL-AF9 murine leukemia model. Mice were treated with Ro at 13.75 mg kg −1 ip, the highest dose  achievable due to limited compound solubility and the use of DMSO as an excipient. Acute treatment of Ro (4 and 12 h) reduced c-KIT protein abundance and intracellular c-MYC (Fig. 6a-c). To determine if Ro treatment could effect disease burden, we next treated a second cohort of animals and monitored them for disease progression for 19 days after transplantation (Fig. 6d). Ro administration every 3 days was well tolerated (Supplementary Fig. 8a) with mice exhibiting little to no weight loss and equivalent red blood cell, platelet, mean corpuscular volume, hematocrit, and hemoglobin counts compared with the non-treated group ( Supplementary Fig. 8b-f). Using healthy mice, we also reported no changes in liver enzymes 24 h     Fig. 8g). Although there was no change in leukemia latency in this very aggressive model, disease progression was assessed in both treated and control groups when control mice and treated mice succumbed to disease (day 19 post-transplantation). The treated group exhibited a significant reduction in spleen weights (Fig. 6e), white blood cell counts (Fig. 6f) and c-MYC levels compared with the control group (Fig. 6g). These data support the concept that targeting MSI in vivo could have therapeutic efficacy in AML.

Discussion
Inhibiting MSI RNA-binding activity could represent a novel therapeutic avenue in both hematological malignancies and solid cancers. Our previous FP-based screen identified compounds that inhibit MSI binding to RNA 36 . Here, we characterize Ro 08-2750 (Ro) as a selective MSI inhibitor with biochemical, structural, and cellular validation linking the compound to the inhibition of the MSI program. Ro falls in the low micromolar range of activity, in line with other RBP associated inhibitors [42][43][44] . We validated Ro as a MSI2 RNA-binding inhibitor with biophysical and biochemical assays. We obtained a high-resolution crystal structure of the MSI2 RRM1 which allowed us to utilize a newly developed computational molecular modeling algorithm and perform docking analysis. This docking analysis was supported through the identification of key interacting residues within the known RNA binding region finding that was confirmed by NMR chemical shift analysis. Both our novel crystal structure and the computational tools will be useful for the discovery and development of small-molecule RBPs inhibitors. We found that a single chemical reduction of Ro drastically decreased its activity in both in biochemical and in vitro cell based assays. Utilizing a related compound with high affinity binding to NGF, we found that it no longer bound MSI2 and poorly inhibited leukemia cell growth. Further studies involving medicinal chemistry with heterocycle isoalloxazines or pteridine-derived compounds could help identify more selective and potent MSI-inhibitors. Despite the potential of this class of molecules to interact with structured RNA motifs 37 , no direct RNA binding to MSI RNA probes or poly(A) was found for Ro. Other groups have identified agents that have putative MSI1 inhibitory activity. The natural phenol extracted from cottonseed ((−)-gossypol) was shown to reduce MSI1 to bind RNA 43 but this interaction was not tested for selectivity. (−)-Gossypol has been considered to be a pan-active compound that has hit in multiple HTS screens 45,46 and assigned to have activity against Bcl-2 47 . MSI1 activity was also inhibited by ω−9 monounsaturated fatty acids (e.g. oleic acid), allosterically binding and inducing a conformational change that prevents RNA to bind 48 . It remains unclear if (−)-gossypol or oleic acid have a broader RNA binding protein inhibitor activity as they were not directly tested against other RBPs 43,48 . We found that Ro could demonstrate differential binding activity to MSI2 compared to five different RRM-based RBPs, Ro's effect on colony formation and direct targets could be rescued by MSI2 overexpression and by mutants of MSI2 that bind poorly to the inhibitor. Moreover, we observed a strong enrichment for the MSI2 shRNAs gene expression signature, associated functional pathways, inhibition of MSI2 binding of target mRNAs and reduced abundance of MSI2 direct targets after Ro treatment. In contrast to other general translational inhibitors 49 , Ro did not alter global translation. These data suggest that Ro could be used to probe the acute effects of MSI inhibition in a variety of cellular contexts and cancer models.
It is also important to note that Ro inhibits both MSI1 and MSI2 and although MSI1 is expressed at low levels in myeloid leukemia it could still be blocking residual MSI1 activity.
Moreover, in other models such as the intestine where both factors act redundantly 12 , dual inhibition could provide a powerful therapeutic strategy. Based on the close conservation of the RRMs between the two proteins it might be challenging to design MSI1 or MSI2 selective inhibitors.
We demonstrated a therapeutic index for Ro in human AML patient samples versus cord-blood derived CD34+ human stem and progenitor cells. Despite the challenges for in vivo administration, we reduced the disease burden in an aggressive MLL-AF9 leukemia model and decreased c-MYC levels without overt toxicity. Interestingly, it has previously been shown that MSI2 can contribute to chemotherapeutic resistance in different cancer models 50,51 . Future studies could examine if combination therapies could provide additional efficacy.
This study identifies and characterizes Ro 08-2750 as a compound selectively inhibiting the oncogenic RNA-binding activity of MSI in myeloid leukemia. It will be important to use this compound (or other chemical derivatives) to test their efficacy in other cancer models and on MSI function related to normal physiology. We suggest that Ro provides the rationale for developing more potent compounds with improved clinical utility for the treatment of cancers that are dependent on the MSI family. In addition, as there are hundreds of RRM containing RNA binding proteins, targeting an RRM motif to block RNA activity with Ro represents a valuable proof of concept for the general inhibition of this class of RNA regulators. Thus, we provide a framework to identify and test novel RNA binding protein inhibitors in cancer.
Viral transduction of murine MLL-AF9 leukemia and normal cells. Tibia, femurs, pelvis, and arm bones from leukemia or C57BL/6 wild type mice (10-12weeks-old) were harvested, crushed, filtered, and subjected to red blood cell lysis (Qiagen). To isolate c-Kit + cells, bone marrow cells were incubated with anti-CD117 microbeads (Miltenyi Biotec), according to manufacturer's instructions, and then subjected to positive selection using autoMACS Pro Separator. For MLL-AF9+ BM cells, thawed vials from previously established secondary transplants 7 were used. All murine cells were cultured and transduced in RPMI with 10% FBS and cytokines SCF (10 ng ml −1 ), IL-3 (10 ng ml −1 ), and IL-6 (10 ng ml −1 ) and GM-CSF (10 ng ml  In vivo MLL-AF9 leukemia model and Ro 08-2750 administration. A total of 10,000 of MLL-AF9 BM secondary mouse leukemia cells previously obtained 7 were injected retro-orbitally into female C57BL/6 (10-12-weeks-old) recipient mice that had been sublethally irradiated at 475 cGy. Drug administration (Ro 08-2750, 13.75 mg kg −1 , DMSO) was performed by intraperitoneal injections (50 μL, top tolerated DMSO volume) 3 weeks after BM transplants (when showing signs of disease) for pharmacodynamic experiments (Fig. 6a), and 3 days after BM transplant for in vivo long-term studies (Fig. 6d). Mice weight was monitored every day to check for toxicity. All animal studies were performed on animal protocols approved by the Institutional Animal Care and Use Committee (IACUC) at Memorial Sloan Kettering Cancer Center.
Fluorescence polarization. To validate RNA-binding activity inhibition by Ro 08-2750 and derivatives (Ro-OH, Ro-NGF) we used Fluorescence Polarization (FP) based assay in 384-well format for dose-response curve studies 36 . The RNA oligo used (Cy3-C 9 [spacer]-rGUAGUAGU, Integrated IDT Technologies) contained 2 MSI motifs (GUAGU) and had 8-nucleotides of length, optimal to minimize background and unspecific interactions. Here, manual pipetting was used to plate the reagents and the FP reading was performed in a BioTek Synergy Neon Plate Reader (High-Throughput Screening Resource Center, HTSRC, Rockefeller University).
Microscale thermophoresis. For binding affinity studies of RNA and smallmolecules to proteins of interest, purified recombinant GST-MSI2 WT, K22, F66, F97 and R100 to single alanine (A), triple (F66A/F97A/R100A) mutants, GST-RBP controls (SYNCRIP, SRSF2, HUR, RBMX, TIA-1) and bovine serum ALBUMIN were NT647-labeled using an amine-coupling kit (NanoTemper Technologies). Runs were performed at a concentration range of 50-120 nM (MSI2 and mutants) and 60 nM (SYNCRIP) to get optimal fluorescence signal using an LED power of 40-50% in a red laser equipped Monolith NT.115 (NanoTemper Technologies) (HTSRC, Rockefeller University). Prior to each run, protein preparations were diluted in MST buffer (50 mM HEPES, 100 mM NaCl, 0.05% Tween-20, pH 7.4) and protein aggregation was minimized by centrifuging the solutions at 20,820×g for 10 min. GST-proteins or GST-protein/RNA complexes (15 min pre-incubation) were mixed with increasing concentrations of small-molecules (0.015 to 500 μM) or RNA (0.0015 to 50 μM) and loaded onto 16 Premium Coated capillaries. The RNA oligo used (rGUAGUAGUAGUAGUA, Integrated IDT Technologies) contained 4 MSI motifs (GUAGU) and was 15-nucleotides long. The MicroScale Thermophoresis (MST) measurements were taken at RT and a fixed IR-laser power of 40% and 20 s per capillary. GraphPad Prism was used to fit the normalized data and determine apparent K D values, represented as percent of fraction bound.
Isothermal titration calorimetry. Due to incompatible fluorescence interference of labeled-RNA in MST assay, we used a non-labeled RNA probe of 15-nucleotides (15-nt, rGUAGUAGUAGUAGUA, Integrated IDT Technologies), poly(A) RNA (Sigma-Aldrich) and recombinant GST-MSI2 to assess direct Ro interaction with these agents. 15-nt or poly(A) RNA (1 mM stock in RNAse free H 2 O) were diluted to a 10 μM final concentration with Isothermal Titration Calorimetry (ITC) buffer (10 mM HEPES + 10% 10 mM Citrate Phosphate and 0.05% Tween-20, pH 7.0) and was titrated against 100 μM Ro 08-2750 or palmitate [positive poly(A) RNA control] in the same buffer by using MicroCal PEAQ-ITC range (Malvern Panalytical, HTRSC, Rockefeller University). As a protein binding control and to confirm the binding affinity (K D ) obtained by MST, we titrated GST-MSI2 (fulllength) at 30 μM in ITC buffer against 100 μM and 300 μM Ro 08-2750 obtaining similar K D values (see Supplementary Fig. 1e for 100 μM). Kinetic and thermodynamic parameters were analyzed and fitted by AFFINmeter web-based software (www.affinimeter.com) and final graphs were represented by GraphPad Prism v7.0.
Chemiluminescent electrophoresis mobility shift assays. An Electrophoresis Mobility Shift Assay (EMSA) approach to assess MSI2-RNA complexes and the inhibitory effect of small-molecules was set up by using LightShift Chemiluminescent RNA EMSA kit (Thermo Scientific). In brief, GST-MSI2 (125-250 ng) was preincubated with DMSO or the small-molecule (typically 20 μM final concentration) during 1 h at RT in 1X RNA EMSA binding buffer (10 mM HEPES, 20 mM KCl, 1 mM MgCl 2 , 1 mM DTT, Thermo Scientific) supplemented with 5% glycerol, 100 μg mL −1 tRNA and additional 10 mM KCl. After this period, 40 nM of biotinylated-RNA (biotin-rGUAGUAGUAGUAGUA, Integrated IDT Technologies) was added to the mixture (20 μL final volume) and incubated 1 h at RT. During this second incubation period, a 4-20% TBE polyacrylamide gel (BioRad) was pre-run at 100 V for 30-45 min in cold 0.5X TBE (RNAse free). Five microliter of 5× loading buffer was added to the 20 μL reaction and loaded into the pre-run TBE gel and voltage set at 100 V. Samples were electrophoresed until 3/4 of the length of the gel. Samples were then transferred in 0.5X TBE at 350-400 mA for 40 min. Membranes were then crosslinked with UV-light crosslinking instrument (UV Stratagene 1800) using Auto-Cross Link function. Membranes were either stored dry for development next day or developed using the detection biotinlabeled RNA chemiluminescence kit (as indicated by the manufacturer) (Thermo Scientific) and Hyperfilm ECL (GE Healthcare).
Crystallization and structure determination. A final concentrated MSI2 RRM1 pure protein preparation (>98% by coomassie) at 2 mg mL −1 in 50 mM Tris-HCl, pH 7.5 was crystallized by sitting drop vapor diffusion. A 1 μL of protein solution was mixed with an equal volume of precipitant solution containing 100 mM Tris, 200 mM Li 2 SO 4 , 25% PEG 3350 (pH 8.5). Crystals appeared after two weeks. They were cryoprotected by mother liquor containing 25% glycerol and flash frozen in liquid nitrogen. X-ray diffraction data were collected from single crystals at the Advanced Photon Source beamline 24ID-C. The temperature was 100 K and the wavelength was 0.9792 Å. Indexing and merging of the diffraction data were performed in HKL2000 52 . The phases were obtained by molecular replacement by PHENIX 53 using PDB entry 1UAW as the search model. Interactive model building was performed using O 54 . Refinement was accomplished with PHENIX. Data collection and refinement statistics are summarized in Table 1. Ramachandran statistics were used, 100% of the residues are in the favored regions of Ramachandran plot. The crystal structure has been deposited in RCSB PDB under the accession code 6DBP. The NMR data was acquired on Bruker AVANCE series of spectrometers equipped with Z-axis gradient TCI/TXI CryoProbes TM at a sample temperature of 5°C and B o field strengths of 500.13 and 800.23 MHz, respectively. To map the binding site from chemical shift perturbations the stock solution of the compound was titrated into 100 μM 1 H/ 15 N-labeled RRM1 at five different protein to compound ratios (1:2, 1:4, 1:6, 1:8, 1:10). The backbone resonances of 1 H/ 13 C/ 15 N labeled RRM1 in the presence (1:14 ratio) and absence of the compound Ro 08-2750 were assigned at 5°C using a standard suite of triple resonance experiments 55 , HNCO, HNCA, HNCACB and CBCA(CO)NH acquired at 800 MHz. The multidimensional NMR datasets were processed in Topspin 2.1 from Bruker Biospin and the chemical shifts analyzed in CARA1.5 56 .
Immunoblot analysis. For immunoblot analysis, Ro treated and DMSO control MOLM13 or K562 cells (routinely at 0.5 × 10 6 cells mL −1 ) were counted and washed twice with cold PBS before collection. 1-5 × 10 6 cells were resuspended and lysed in 250 μl of 1× RIPA Buffer supplemented with Protease Inhibitor Tablets (Sigma-Aldrich) buffer for 30 min on ice. After centrifugation at 20,820 × g on a top-bench centrifuge, lysate (supernatant) was collected and total protein quantified by BCA (Thermo Scientific). Luminescence-based cytotoxicity assays. A total of 10,000 cells (MLL-AF9+ BM from secondary transplants or human leukemic cell lines -K562 or MOLM13-) were platted into U-bottom 96-well plates in the presence of increasing concentration of small-molecules (Ro, Ro-OH or Ro-NGF) up to 100 μM (in 1:2 serial dilutions). Cells were cultured for 72 h at 37 C in a 5% CO 2 incubator. To read cell viability, Cell-Titer Glo kit (Promega) was used. After cooling down cells to RT for 20-30 min, 100 μL of the cultured cells were transferred to opaque-white bottom 96-well plates and mixed with 100 μL of Cell-Titer Glo reagent. The mixture was incubated for 15 min at RT and read using a Synergy H1 Hybrid reader (BioTek) for luminescence. Data was normalized as percentage viability and graphed by nonlinear regression curves in Graph Pad PRISM 7.0.
RNA immunoprecipitation. To assess mRNA enrichment and blocking of proteinbinding to mRNA by the small-molecules we performed RNA immunoprecipitation (RNA IP) experiments using Magna RIP RNA-binding protein immunoprecipitation kit (#03-115, Millipore). 25 × 10 6 K562-MIG or MSI2 overexpressing cells 1 h treated with DMSO (control) or Ro μM were used. First, cells were washed with cold PBS and lysed. Five micrograms of mouse anti-Flag (clone M2, #F1804, Sigma-Aldrich) antibody incubated with magnetic beads were used to immunoprecipitate Flag-MSI2 K562 cells. After washing the immunoprecipitated, they were treated with proteinase K. RNA extraction was performed by the phenol-chloroform method, and 200-500 ng of purified RNA was converted to cDNA using the Verso cDNA kit (Thermo Scientific). qPCR was used to validate target mRNAs bound by MSI2 and control cells.
O-Propargyl-Puromycin incorporation by flow cytometry. Cells were plated at a density of 200,000 cells mL −1 and pre-treated with DMSO or Ro up to 4 h. Then, 50 μM O-propargyl-puromycin (OP-Puro; NU-931-05, Jena Bioscience) was added. Control cells were co-incubated with DMSO or Ro and treated with 150 μg mL -1 cycloheximide for 15 min. Non-OP-Puro treated cells were also used as negative controls for flow cytometry. Cells were washed twice before collection and subjected to processing using the Click-iT Flow Cytometry Assay kit (#C10418, Invitrogen) following the manufacturer's instructions. Labeled cells were analyzed using a BD LSR Fortessa instrument and graphed as Alexa Fluor 647 (AF647) Mean Fluorescence Intensity (normalized to DMSO control treated with OP-Puro).
RNA sequencing. Total RNA was isolated from 1 × 10 6 dry pellets of K562 and MOLM13 4 h treated with DMSO (control) or Ro 20 μM (n = 4 for each group) using Qiagen RNeasy Plus Mini kit and the quality assessed on a TapeStation 2200 (Agilent technologies). QuantSeq 3′ mRNA-Seq Library Prep Kit FWD (Lexogen, Vienna Austria), supplemented with a common set of external RNA controls, according to manufacturer's recommendations (ERCC RNA Spike-In mix, Ther-moFisher Scientific, #4456740). An in-house pipeline was used for read mapping and alignment, transcript construction and quantification of data generated by sequencing (HiSeq 2000, NYGC, NY, USA). This procedure was done in the Epigenetics Core from MSKCC. RNA-seq data have been deposited in NCBI GEO database with the accession code GSE114320.
Synthesis of Ro-OH by reduction of Ro 08-2750 aldehyde. To a cooled (0°C) slurry of Ro 08-2750 (19 mg, 0.070 mmol) in anhydrous MeOH (1.9 mL) was added LiBH 4 (32 mg, 1.5 mmol) in portions over 5 min. The slurry turned from bright orange to dark brown, then dark green within 10 min. The reaction mixture was removed from the ice bath and allowed to warm to room temp (22°C) over 2 h. Reaction progress was monitored by LC-MS (5-95% MeCN in H 2 O). Four portions of LiBH 4 (10 mg, 0.04 mmol) were added every 12 h until the reaction was complete. The reaction was quenched with AcOH (10 mL) and filtered. The solids were washed with water (5 mL), MeOH (5 mL), and Et 2 O (5 mL). The solid was collected and dried under vacuum to provide a pale orange solid (7 mg, 26%). Purification by HPLC (5-95% MeCN in H 2 O) afforded the product as an orange solid (3 mg, 16%). The synthesis was adapted from Salach et al. 57 . Statistical analysis. Student's t test was used for significance testing in the bar graphs, except where stated otherwise. A two-sample equal-variance model assuming normal distribution was used. The investigators were not blinded to the sample groups for all experiments. P values less than 0.05 were considered to be significant. Graphs and error bars reflect means +/− standard error of the mean except stated otherwise. All statistical analyses were carried out using GraphPad Prism 7.0 and the R statistical environment.
Modeling and system preparation for computational modeling. System preparation, modeling, and initial docking calculations were performed using the Schrödinger Suite molecular modeling package (version 2015-4), using default parameters unless otherwise noted. The MSI2 RRM1 protein structure (PDB ID: 6DBP) was prepared using the Protein Preparation Wizard 58 . In this step, force field atom types and bond orders were assigned, missing atoms were added, tautomer/ionization states were assigned, water orientations were sampled, and ionizable residues (Asn, Gln, and His residues) have their tautomers adjusted to optimize the hydrogen bond network. A constrained energy minimization was then performed. All crystallographically resolved water molecules were retained.
Potential binding sites were explored and characterized using the SiteMap 59 tool. Ligands with experimental activity and known inactives were docked into putative binding sites using Glide SP 60,61 to evaluate enrichment of known actives. Best docking scores were for the 'Ro' series for the '(−)-gossypol' binding site described by Lan et al. 43 compared with other putative pockets.
Since the receptor may not be in an optimal conformation to bind small molecule inhibitors, induced fit docking 62 of ligand Ro 08-2750 was performed to this binding pocket. Induced fit docking results were validated with the metadynamics protocol described by Clark et al. 63 . In these metadynamics simulations a biasing potential is applied to the ligand RMSD as collective variable. The resulting potential energy surface is evaluated towards how easy a ligand can move away from the initial binding mode. The underlying assumption is that a ligand pose which is closer to the real one has a higher energetic barrier to leave the pose than an incorrect pose. The pose ranked second using the induced fit docking score retrieved the best score from the metadynamics ranking protocol compared with the other induced fit docking poses. This receptor configuration was furthermore tested towards its suitability for a virtual screening by a Glide SP docking of known actives into this pocket. The docking scores using this receptor conformations were better (down to −6.2) compared with the initial protein conformation in the crystal structure. Furthermore, a WaterMap 64 calculation was done for this receptor.
Induced fit docking of Ro-NGF and Ro-OH compounds. Induced Fit Docking (IFD) was performed against the receptor pose from the selected Ro 08-2750 pose, using Schödinger molecular modeling suite (version 2017-4). Poses for Ro-NGF and Ro-OH, the top and second scored poses, respectively, were selected to most closely match the Ro 08-2750 pose.
Alchemical free energy calculations. Absolute alchemical free energy calculations were carried out to validate the putative binding poses in a fully flexible explicitly solvated system. The YANK GPU-accelerated free energy calculation code with the Amber family of forcefields was used for this purpose. Details follow: System preparation and modeling: the top poses generated by induced fit docking, as described above, were selected as input protein and ligand poses. Because proteins and ligands were already prepared, they were simply run through the pdbfixer 1.4 command line tool with add-atoms and add-residues set to None to convert residue and atom names to be compatible with Amber tleap.
Parameterization: tleap (from the minimal conda-installable AmberTools 16 suite ambermini 16.16.0) was used to solvate the complex in a cubic box with a 12 Å buffer of TIP3P water molecules around the protein. The system was parameterized using AMBER's forcefield ff14sb 65 and GAFF 1.8 66 . Missing ligand parameters were determined using antechamber 67 . The ligand was assigned charges using the AM1-BCC 68 implementation in OpenEye (OEtoolkit 2017.6.1 through openmoltools 0.8.1).
Minimization: minimization was performed using the implementation of the L-BFGS algorithm in OpenMM 7.1.1 69 with a tolerance of 1 kJ mol −1 nm −1 .
Production Simulation: production simulation was run using YANK 0.19.4 70 using OpenMMTools 0.13.4. In order to keep the ligand from diffusing away from the protein while in a weakly coupled state, it was confined to the binding site using a Harmonic restraint with an automatically determined force constant (K = 0.33 kcal mol −1 Å −2 ). The restraint was centered on the following receptor residues using all-atom selection: 2, 4, 46, 76, 78, and 80. The ligand atoms were automatically determined. The calculation was performed using particle mesh Ewald (PME) 71 electrostatics with default YANK settings with a real-space cutoff of 9 Å. A long-range isotropic dispersion correction was applied to correct for truncation of the Lennard-Jones potential at 9 Å. The system was automatically solvated with TIP3P 72 solvent and four neutralizing Cl − ions, paramterized using the Joung and Cheaham parameters 73 . Production alchemical Hamiltonian exchange free energy calculations were carried out at 300 K and 1 atm using a Langevin integrator (VRORV splitting) 74 with a 2 fs timestep, 5.0 ps -1 collision rate, and a molecular-scaling Monte Carlo barostat. Ro 08-2750 and Ro-NGF were run for 10,000 iterations (50 ns/replica) with 2500 timesteps (5 ps) per iteration, while Ro-OH was run for 15,000 iterations (75 ns/ replica) with 2500 timesteps (5 ps) per iteration. Complex configurations were stored for each replica once per iteration. Replica exchange steps were performed each iteration to mix replicas using the Gibbs sampling scheme 75,76 . The alchemical pathway was automatically determined for each compound using the YANK autoprotocol protocol trailblazing feature.
Absolute binding free energy estimates: absolute free energies (ΔG) of binding for each compound was estimated using MBAR. Samples were reweighted to a cutoff of 16 Å to correct the isotropic dispersion correction to a non − isotropic long-range dispersion. This correction is important to account for the heterogeneous density of protein. To remove the harmonic restraint bias, samples were reweighted to substitute a squared well restraint of radius 10 Å.
Clustering analysis: the fully interacting trajectory from YANK was extracted to a PDB file, discarding the following number of initial iterations, which came prior to equilibration: 77 1500 for Ro 08-2750, 1600 for Ro-OH, and 1600 for Ro-NGF. These trajectories were aligned in MDTraj 78 using only protein backbone atoms. The small molecules were then sliced out and clustered on Cartesian coordinates using the MSMBuilder 79 implementation of RegularSpatial clustering using a 1 Å RMSD cutoff. For the most populated clusters for Ro 08-2750 and Ro-OH, cluster centers were selected and shown with 10 randomly sampled cluster members. Ro-NGF produced a large number of lowly populated clusters with highly heterogeneous binding poses, and were therefore not shown.
Conformational heterogeneity analysis: to investigate the conformational heterogeneity in the presence or absence of the ligand, the fully interacting thermodynamic state (corresponding to the holo protein bound to the ligand) and fully non-interacting state (corresponding to the apo protein free of ligand interactions) for all three ligands were extracted using a 4-frame skip, discarding the initial frames as above.
Reporting summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability
A reporting summary for this Article is available as a Supplementary Information file. Coordinates and structure factors have been deposited in the RCSB Protein Data Bank (PDB), under the accession code 6DBP. RNA-seq data have been deposited in NCBI Gene Expression Omnibus (GEO) database with the accession code GSE114320. The source data underlying Figs. 1b, 1d, 2d-j, 3f, 5e-h and Supplementary Figs. 1c-e, 2b, 2d, 2 f, 5c and raw data are provided as a Source Data file. All data is available from the corresponding author upon reasonable request.