Rational discovery of antimetastatic agents targeting the intrinsically disordered region of MBD2

The intrinsically disordered region of MBD2 is a promising target for treating cancer metastasis.


INTRODUCTION
Although at least 650,000 protein-protein interactions (PPIs) might occur in humans, only one PPI inhibitor has been approved for clinical use to treat cancers (1), suggesting that the field of PPI inhibitors remains largely unexplored. A variety of proteins and their PPIs have emerged as prospective drug targets to treat tumors because of the extreme heterogeneity and plasticity of cancer (2,3). Ligands with the potential of binding to a specific site of a target protein with known structure can be efficiently identified by virtual screening. However, the structural plasticity of target proteins usually works against yielding an effective drug candidate. For example, selected compound treatment of cells/organisms often fails to elicit the anticipated effects due to in vivo structural alterations of the target protein caused by various posttranslational modifications (PTMs) and/or unanticipated interactions of the compound and/or its target protein with other molecules (4,5). Furthermore, many critical proteins regulating various biological processes do not have unique structures as a whole or in some functionally important regions (6,7). Structures of these intrinsically disordered proteins (IDPs) or IDP regions (IDPRs) are extremely dynamic, depending on the environment, and might change during function (4,8). Many signaling IDPs/IDPRs undergo characteristic disorder-to-order transitions (DOTs) upon interactions with specific binding partners and/or through PTMs (9,10). Targeting the IDPs/IDPRs capable of DOT is generally considered an attractive but challenging task for developing anti-PPI inhibitors. In this regard, a recently identified small-molecule compound, 10058-F4, serves as a pioneering success of anti-PPI inhibitor that binds to an IDPR of c-Myc undergoing a DOT upon binding to its partner Max (11,12). 10058-F4 was discovered by a random screening using a yeast two-hybrid system (11), followed by experimental identification of its specific binding site (residues 402 to 412 of c-Myc) as an IDPR. Drug leads like 10058-F4 targeting IDPs/IDPRs cannot be found by conventional computational methods that rely on fixed conformations, such as crystallographic structures of target proteins. No computer-aided drug discovery platform is currently available for the systematic exploration of IDPRs as potential drug-target sites (3).
To fill this gap, we developed a novel platform for the discovery of drug leads based on molecular docking and molecular dynamics (MD) simulations of the DOT-associated IDPRs of target proteins. Figure 1A describes this protocol. First, intrinsic disorder predispositions of drug-target proteins are analyzed, and potential disorder-based binding regions that can undergo DOTs are evaluated. A search of the protein structure database [Protein Data Bank (PDB)] is also performed to identify known PPIs and DOTs. Once the potential drug-target sites (DOT-based PPI regions) are determined, the corresponding structures retrieved from the PDB are used for molecular docking with druggable compounds from the ZINC compound library (13). Together with the docking scores, off-target probabilities assessed by the similarity ensemble approach (SEA) (14)(15)(16) analysis are also considered for selection of lead compounds from the molecular-docked hit compounds. Last, prospected candidate compounds are suggested via MD simulations that evaluate the mode and efficiency of the compound binding.
The feasibility of the proposed approach was validated in this study by targeting an IDPR of MBD2 that undergoes a DOT upon association with its binding partner p66 for the integration of the Mi-2/NuRD chromatin remodeling complex (CRC). The integrated Mi-2/NuRD CRC includes one CHD (either CHD3 or CHD4), one HDAC (HDAC1 or HDAC2), two DOC1, three MTA (MTA1, MTA2, and MTA3), six RbAp46/48, two p66 (p66 or p66), and one MBD (MBD2 or MBD3)  molecules (17,18), where the molecular interaction of MBD2 with p66 critically mediates the proper assembly of CRC (17,19). This CRC performs an important epigenetic function in normal development and differentiation by suppressing gene expression by binding directly to the DNA methylation sites and to the DNA methyltransferases (20,21). CRC also contributes to the development of human diseases, including cancer (22,23); for example, the epigenetic regulation by Mi-2/ NuRD CRC includes multiple tumor suppressor genes (23,24), and several CRC components, including MBD2, were also observed to be oncogenic and/or closely correlated with the aggressiveness of several cancers (23,25,26). In particular, the function of Mi-2/NuRD CRC is known to be associated with the cellular process of epithelialmesenchymal transition (EMT; the conversion of adhesive epithelial cells into migratory, invasive mesenchymal cells) that drives wound healing and cell migration and invasion (27,28). In cancer, EMT necessarily mediates the metastasis of cancers and thus also enables carcinoma cells to acquire cancer stem cell (CSC) properties, malignancy-associated traits, and drug resistance (29)(30)(31). Given that the metastasis is responsible for more than 90% of contemporary cancer deaths and yet no marketed antimetastatic drug is currently available (32), developing these drugs to target the cancer spreading is an essential and urgent task for oncological therapy. In this context, functional inhibition of CRC or modulation of its individual components might serve as a novel strategy for effective anticancer therapy to prevent the progression of cancer to metastatic stage. In particular, it has been observed that down-regulation of MBD2 and/or p66, which triggered derepression of epithelial regulators via epigenetic reprogramming of the Mi-2/NuRD CRC into the MBD2-free or disentangled CRC, resulted in promoted epithelial differentiation and loss of tumor-initiating ability. Therefore, targeting MBD2 specifically at its IDPR would be a promising approach to the development of antimetastatic agents by inhibiting its DOT-based PPI with p66 that is essential for the integration of CRC and thus for its critical function in EMT. In addition, no noticeable adverse effects displayed by MBD2 inhibitors can be expected from the fact that down-regulation of MBD2 expression is essential for normal cell differentiation (33), and yet, MBD2 knockout (MBD2 −/− ) mice exhibit normal survival and reproduction (34).
Hence, in this study, the MBD2 IDPR and its DOT-based interaction with p66 for the CRC integration were selected as a highly promising target system to evaluate the efficiency of our platform for rational drug discovery. Using this novel approach, we identified two small-molecule compounds capable of inhibiting the PPI of MBD2 and thereby efficiently suppressing the cancer metastatic potentials. In vivo efficacy of both leads in inhibiting cancer metastasis was also evident in a murine xenograft tumor model. Therefore, our novel method renders IDPRs available for rational discovery of anticancer drugs targeting DOT-based PPIs. In particular, the identified compounds provide a basis for the development of previously unidentified inhibitors capable of controlling metastasis of various carcinomas.

RESULTS
In silico analysis for determination of target site and search for potential ligands As our study was inspired by the discovery of 10058-F4, which binds to the c-Myc IDPR to inhibit its DOT for interaction with Max (11,12), we compared the PPI site of MBD2 with that of c-Myc using our computational platform. Sequence analysis (see fig. S1 for sequence and structure information) revealed that disorder profiles of the PPI site of MBD2 (residues 360 to 393 for p66 interaction) (17,35) closely resembled that of c-Myc (residues 400 to 434 for Max interaction) (36,37) (Fig. 1B), characterized by a positive slope in its disorder profile. As both MBD2 and c-Myc are folded in complexes with their cognate partners (p66 and Max, respectively) (17,35,37), this analysis suggests that the PPI sites of MBD2 and c-Myc could undergo a similar type of DOT upon complex formation.
Subsequently, a nuclear magnetic resonance (NMR) ensemble structure of MBD2 360-393 in its complex with p66 138-178 (PDB ID: 2L2L) was retrieved, and the lowest-energy conformation of the ensemble was extracted for molecular docking analysis using the four residues (D 366 , I 369 , V 376 , and L 383 ) of MBD2 360-393 engaged in the coiled-coil interaction, with p66 (35) as the initial target site in the molecular docking. From the molecular docking-based virtual screening of 2 × 10 6 chemical compounds in the ZINC library, 10 promising compounds (compounds #1 to #10 in Fig. 1C) capable of interaction with MBD2 at the designated target site were selected. As a control, the Y 402 -targeted molecular docking of 10058-F4 to c-Myc 395-430 ( Fig. 1D; note that the key residue for the c-Myc interaction with 10058-F4 is Y 402 ) (35) was compared with the MBD2 360-393 docking of the 10 selected hit compounds (table S1). The MBD2 369targeted docking of two compounds {compounds #2 and #3 in Fig. 1D named herein ABA [2-amino-N-(2,3-dihydro-benzo [1,4] dioxin-2-ylmethyl)-acetamide] and APC [3-(2-amino-acetylamino)pyrrolidine-1-carboxylic acid tert-butyl ester], respectively} was found as the most favorable. In ABA:MBD2 369 and APC:MBD2 369 dockings, these compounds formed three intermolecular hydrogen bonds and had relatively low DOCK scores (−35.2 and −33.3 kcal mol −1 , respectively) of the DOCK binding. These binding features could be compared favorably with those of the 10058-F4:c-Myc 402 docking, which showed the DOCK score of −6.77 kcal mol −1 and just one intermolecular hydrogen bond (table S1).

Selection of lead compounds by in silico assessment of off-target probability and activity test in cells
Concerning the potential side effects of the selected hit compounds, their off-target probabilities were assessed by the SEA analysis (14,16), which has served as an eminent bioinformatics resource aiding in target identification for drug development by profiling multiple protein targets of chemical compounds as probes (15). For this analysis, the c-Myc inhibitor 10058-F4 and two anticancer drugs, imatinib (Gleevec) and sorafenib (Nexavar), were also compared as controls, and 2060 human proteins in the database were searched as potential targets. Given that a significant binding is feasible when both the Max T c value more than 0.5 and E value lower than 10 −10 are relevant, no suggestible off-target was predicted for 7 of the 10 hit compounds including both ABA and APC, whereas four proteins were found as the probable 10058-F4 targets ( Fig. 2A and table S2). Two of the other compounds also showed a small number of putative off-target proteins (six and two proteins for compounds #4 and #10, respectively), whereas 35 and 26 targets were suggested for imatinib and sorafenib, respectively ( fig. S2A and table S2). Therefore, we screened nine compounds with low off-target probability for cellular activity dysregulating MBD2. In particular, the cell migration assay was used for this preliminary test of the compounds on the basis of the previous observation that knockdown of MBD2 in cancer cell lines resulted in decreased migration of the cells. The result implicated most of the hit compounds in actual suppression of the migration of breast adenocarcinoma MDA-MB-231 (LM1) and colorectal carcinoma HCT116 cells ( Fig. 2B and fig. S2B). In particular, ABA (compound #2) and APC (compound #3), which accomplished the most favorable target binding in the aforementioned molecular docking experiments, also showed the least MI 50 (concentration for half-inhibition of cell migration) values. Therefore, these two molecules were selected as lead compounds for subsequent evaluation in detail.

In silico analysis of target binding for selected lead compounds
To assess target-binding feasibility and mode of binding of the two selected leads, we conducted MD simulation using the structures resulting from the ABA: In particular, the highest contact density value at the most contacted residue (D368 contact) in the ABA:MBD2 369 trajectory was higher than that (L404 contact) in the 10058-F4:c-Myc 402 trajectory, suggesting stronger binding. Next, MD simulations for the ligand:MBD2 360-393 complex were extended to include D 366 -, V 376 -, and L 383 -targeted docking (Fig. 3C). Consistent with the ABA:MBD2 369 trajectory, D 368 was the most contacted residue in the heatmaps for heavy atom contacts of the ABA:MBD2 376 trajectory, although no preferential contact was found in the other ABA:MBD2 360-393 trajectories and in the APC:MBD2 360-393 MD simulation sets. Collectively, the MD simulation indicated that the actual binding of ABA and APC to MBD2 360-393 would be as promising as the 10058-F4 binding to c-Myc 395-430 , although detailed interaction modes can be different between the two compounds. Therefore, it was subsequently examined whether the targeted binding of the compounds to MBD2 would influence specific PPI of the protein.

Inhibition of the DOT-mediated PPI of MBD2 by the lead compounds
It has been suggested that 10058-F4 evokes a local conformational change (36) or conformational equilibrium shift (38,39) of the c-Myc IDPR at its binding sites, and this small but significant alteration is critically involved in the functional inhibition of the DOTmediated PPI of c-Myc with Max. Detailed inspection of the MD simulation results suggested that the MBD2 IDPR could also undergo a local conformational perturbation upon the binding of ABA and APC. For instance, in the ABA:MBD2 369 and APC:MBD2 369 trajectories, both  and  backbone torsion angles of the most contacted residue (D 368 ) in the compound-contacting states were significantly (t test, P < 0.05) different from those in the noncontacting states ( fig. S3B). The compound-bound conformation also appeared to be different between ABA and APC, as the D 368  angles in the compound-contacting states were significantly different in between ABA:MBD2 369 and APC:MBD2 369 trajectories, although  angle differences were not significant (t test, P = 0.574). Therefore, to further analyze the possible conformational perturbation, we com-  (table S4), which is distinguished from the similarity in / of D 366 and  of I 369 , suggested that I 369 served as a turning point for the observed torsion angle differences more evident in its C-terminal region from I 369 . Collectively, comparative MD simulations of MBD2 360-393 in different states (apo-, compound-, and p66 138-178 -bound) suggested the compound-specific induction of local conformational perturbation of MBD2, especially at its p66-interacting site, which would most likely interfere with the MBD2-p66 interaction. Therefore, we next examined whether these leads can actually inhibit the PPI of MBD2, with p66 both in vitro and in cell, by fluorescence resonance energy transfer (FRET) and co-immunoprecipitation (co-IP) assay.
As the coiled-coil interaction between MBD2 and p66 occurs in an antiparallel fashion (17), MBD2 was fused with a FRET acceptor protein dTomato at its N terminus, whereas the FRET donor enhanced yellow fluorescent protein (eYFP) was C-terminally fused to p66   (33) for in vitro FRET. Unfortunately, the fulllength p66 was not available for the in vitro FRET studies due to the inclusion body formation in the Escherichia coli system for the recombinant production. The in vitro FRET result evidenced that both ABA and APC efficiently interfere with the MBD2-p66 interaction by provoking significant reduction of FRET, which, at 1 to 1.5 equimolar concentrations of the compounds, reached half of the value recorded for the MBD2-p66 1-206 complex (Fig. 4A  and fig. S4A). The FRET analysis in 293T cells by transient cotransfection of eYFP-MBD2 and mCherry-p66 expression constructs also showed the noticeable FRET reduction, which was dependent on the concentrations of the compounds used for the treatment ( Fig. 4B and fig. S4B). Furthermore, the half maximal inhibitory concentration (IC 50 Fig. 2B). Last, the results of the co-IP assay to capture the endogenous MBD2-p66 complex corroborated the fact that ABA and APC inhibit the MBD2-p66 association with the submicromolar IC 50 (Fig. 4C). Therefore, as the interruption of the MBD2-p66 interaction is anticipated to result in the prevention of the proper assembly of Mi-2/NuRD CRC, we subjected the compounds to an in-depth evaluation of biological activities targeting the function of Mi-2/NuRD CRC in cellular EMT and thereby in cancer metastasis.

Specific action of the two identified leads suppressing metastatic potentials of cancer cells
The cellular EMT process that drives cell migration and invasion is critical not only for wound healing but also for cancer metastasis, including promotion of CSC and drug-resistant properties of cancer cells (29-31). As we have previously observed that the MBD2 and/or p66 down-regulation in cancer cell lines resulted in the depressed EMT and conversely promoted epithelial differentiation, we reasoned that disrupted PPI between MBD2 and p66 by the ABA and APC compounds could result in suppression of metastatic potentials of cancer cells by regulating the Mi-2/NuRD CRC-mediated EMT. In agreement with these hypotheses, in mesenchymal type of cancer cells (triple-negative and basal-type breast cancers and aggressive colon cancers) treated with ABA or APC, the increased levels of epithelial markers (CDH1 and CTNNB1) were appreciable, whereas the mesenchymal marker (VIM, SNAIL, SLUG, and CDH2) expressions were suppressed. On the other hand, such an alteration indicative of mesenchymal-epithelial transition (MET) was not apparent in the epithelial cancer cells (luminal breast cancers and less aggressive colon cancer) (Fig. 5, A and B, and fig. S5A). Subsequent analyses confirmed that the compounds suppressed wound healing and migration/invasion abilities of the cancer cells (Fig. 5, C and D, and fig. S5B). In addition, flow cytometric measurements of the cell surface markers CD44 and CD24 indicated that the LM1 cells of the stem-like phenotype (CD44 hi /CD24 lo ) were switched over to the nonstem phenotype (CD44 lo /CD24 lo ) by the compound treatments (Fig. 5E), although the compounds did not induce significant alterations in the proliferation rates and cell cycle progression of the cells tested (Fig. 5, F and G, and fig. S5, C and D). Furthermore, the compound-treated cancer cells showed reduced capability of mammosphere formation (Fig. 5H and fig. S5E), thereby resulting in enhanced susceptibility of the cells to chemotherapeutic drugs including doxorubicin and cisplatin ( Fig. 5I and fig. S5F). Last, mRNA sequencing (mRNA-Seq) results showed that global gene expression profiles of the ABA-or APC-treated cells were highly comparable to those of MBD2-or p66-knockdown cells but markedly discriminated from the profiles of nontreated wild-type cells (Fig. 5J), supporting no significant off-target effects as initially predicted by SEA ( Fig. 2A). Together, these observations established antimetastatic activity of the lead compounds, ABA and APC, by demonstrating that the compounds actioned so specifically on the MBD2-p66 PPI system that the EMT process was efficiently modulated to induce transition of CSC-like cells from a mesenchymallike state to a bona fide epithelial state.

Verification of antimetastatic efficacy of the two selected lead compounds
Antimetastatic efficacy of the two selected lead compounds in vivo was analyzed using xenograft mice transplanted with the LM1 cells, which were chosen for its potent ability to readily metastasize to lung in mice (40). Here, ABA (10 g kg −1 ) and APC (20 g kg −1 ) compounds were administered by intravenous injection six times every 3 days from day 10 after the subcutaneous injection of the green fluorescent protein (GFP)-tagged LM1 cells, followed by sacrifice of the mice (after 4 days of the last administration) for subsequent analysis of tumors (Fig. 6, A and B). Notably, although growth inhibition of original tumor was not significant (Fig. 6, A, C, and D), both ABA and APC compounds exhibited a potent inhibition of the cancer metastasis to lung (represented by the number of nodules developed in lung; Fig. 6C), with no significant effects on body weight of the xenograft mice (Fig. 6B). It was also confirmed by immunohistochemistry that the injected LM1 cells were responsible for the origination of tumor and the metastasized tumor nodules in lung (Fig. 6D). In contrast, histological properties of major organs (Fig. 6E) and complete blood count (CBC) result (Fig. 6F) of the compound-administered mice remained normal. Thus, both ABA and APC appear to be promising antimetastatic agents that are unlikely to cause adverse effects in normal tissues.

DISCUSSION
IDPs/IDPRs are important not only for normal cellular processes but also for the development of various human diseases. In particular, proteins validated as potential drug targets have been increasingly identified to contain IDPRs crucial for PPI mediation. However, the dynamic structure of IDPs/IDPRs limits their use in rational structurebased drug discovery. There are some successful examples of finding of compounds that can bind to and regulate the IDPR-containing proteins (e.g., the c-Myc IDPR-targeting compound 10058-F4). However, most of the current approaches to discover compounds targeting functional IDPR are based on random screening. Meanwhile, because many IDPRs undergo characteristic DOTs upon specific PPIs (9, 10), related structural information can be retrieved from their complexed structures. This, together with the in-depth insights into the compound binding modes (38) and the rapidly accumulating knowledge of the IDPR structural properties (6,7), suggests the possibility for utilization of the structure-based rational approach as a feasible route for efficient discovery of drug leads targeting specific IDPRs engaged in DOT-based PPIs. The present novel approach to an antimetastatic agent development provides a prime example of a collaborative work of in silico, in vitro, in cell, and in vivo analyses to discover the drug candidates targeting a pharmacologically important IDPR. In particular, we propose here a three-step computational platform for finding these drug leads. First, IDPRs with DOT potential are selected as potential drug-target sites. We speculate that these regions can be identified based on the characteristic features of their intrinsic disorder predisposition profiles similar to those observed in the known DOTbased PPI regions of MBD2 (residues 360 to 393) and c-Myc (residues 395 to 430) (Fig. 1B). Second, for virtual screening, ordered conformation is taken from the structure of selected IDPR complexed with binding partner. Third, MD simulation is conducted for the selected drug leads targeting IDPRs. Because the structure of target IDPR is dynamic (6, 7) and because the presumably entropy-driven compound binding also occurs in a dynamic fashion (38), MD simulations of the compound-target complex structures are essential for detailed evaluation of the binding feasibility. In this study, MD simulation indicated the compound binding-specific conformational perturbations of MBD2, particularly at its critical PPI site with p66, which could provide a structural basis for the molecular inhibition of the DOT-based PPI of MBD2. In general, specific molecular interactions of IDPs/IDPRs are known to be accomplished in distinctive ways such as DOT, avidity, allovalency, and fuzzy binding; the last three involves multivalent binding sites, whereas the first represents a simple two-state binding involving a single binding site (41,42). The present MD simulation result suggests that the ABA and APC binding of the MBD2 IDPR resembled a dynamic, multivalent interaction at low entropic cost, rather than the DOT-based interaction relevant to its p66 binding. The entropy-driven compound binding and structural multiplicity of the compound-bound IDPR have been identified earlier in the case of 10058-F4 binding to c-Myc 402-412 , which also requires just a few stable atomic interactions were analyzed using Student's t test. Ab, antibody; IgG, immunoglobulin G. (38,39). In this regard, increased fuzziness of the MBD2 IDPR by the compound binding may conversely lead to decreased propensity for DOT for its p66 interaction, although the exact mode of binding of our compounds to the MBD2 IDPR, which can ultimately underlie their PPI inhibition mechanism, remains to be characterized in detail. Our computational platform also contains an additional in silico study using the SEA, which was practical to assess off-target probability of the suggested compounds that is potentially associated with adverse effects in actual usage. In subsequent studies, mRNA-Seq results in cells (Fig. 5J) were consistent with the SEA result ( Fig. 2A) that predicted no significant off-target probability, and in vivo administration of the suggested compounds raised no significant toxicity in normal tissues (Fig. 6, E and F).
It is generally appreciated that identifying and understanding molecular regulation and signaling network involved in the EMT process are essential to provide a molecular basis for antimetastatic drug development (43,44). Concerning this study, we have recently identified the MBD2-p66 molecular system in Mi-2/NuRD CRC as a promising target for EMT modulation by observing the induction of MET (conversed process of EMT) by knockdown of MBD2 and/or p66 in cancer cells. Together with this parallel effort, the present discovery of novel antimetastatic agents targeting a component of Mi-2/NuRD CRC validates that this epigenetic machinery can serve as an emerging target system for efficient antimetastatic drug developments. Both ABA and APC disrupting the specific PPI of MBD2 were able to suppress cellular EMT processes, thereby inducing epithelial differentiation of the more aggressive CSCs. Last,   our compounds potently inhibited the cancer metastasis in vivo. Furthermore, considering that they raised no noticeable adverse effects on blood and normal tissues, the present results provide a basis for a novel safe control of cancer metastasis. Hence, found in this study, low-molecular weight (<250 g mol −1 ) compounds constitute a pioneering example of antimetastatic agents acting on a specific Mi-2/NuRD CRC component. In addition, the present observation that the compound treatments rendered the cancer cells more sensitive to anticancer drugs (Fig. 5I) provides important implications in combination therapy for cancer.
In conclusion, this study successfully used a rational approach to search for the novel antimetastatic agents acting via inhibition of the DOT-based PPI in an IDPR. As IDPs/IDPRs play crucial roles in diverse cellular processes (6, 7), we believe that this platform can be applied for the discovery of innovative drug leads targeting DOTbased PPI regions in proteins associated with various cancers and other diseases.

Study design
This study was designed to develop a novel platform for the discovery of drug leads based on molecular docking and MD simulations of the DOT-associated IDPRs of target proteins and, as a proof of concept, to identify candidate drugs, suppressing metastatic potentials of cancer cells in vitro and in vivo, by targeting an IDPR of MBD2  that undergoes a DOT upon association with its binding partner p66 for the integration of the Mi-2/NuRD CRC. These objectives were addressed by (i) analyzing intrinsic disorder predispositions of drug-target proteins and evaluating potential disorder-based binding regions (45), (ii) doing molecular docking with druggable compounds from the ZINC compound library to the potential drug-target sites, (iii) selecting two lead compounds based on the docking scores and off-target probabilities and experimental validation of target binding, (iv) evaluating the mode and efficiency of the compound binding via MD simulations, (v) assessing the identified leads for biological effects suppressing metastatic potentials of cancer cells, and (vi) verifying antimetastatic efficacy in a murine xenograft tumor model.
In animal studies, mice were randomly assigned to treatment and control groups. Numbers of tested mice were specified in each figure.
Outliers were removed only if mice died at an early stage of the treatment according to the Hanyang University Institutional Animal Care and Use Committee (IACUC) dimension guideline. The primary end points were tumor size and cancer metastasis to lung. Mice were euthanized when moribund or at the end of the prespecified treatment period. All procedures were performed in accordance with institutional protocols approved by the IACUC of the Hanyang University. Pathology analysis was performed in a blinded fashion.

Statistical analysis
Data were presented as means ± SE. The sample size for each experiment, n, was included in Results and the associated figure legend. Everywhere in the text, the difference between two subsets of data was considered statistically significant if the one-tailed Student's t test gave a significance level P (P value) less than 0.05. Multiple comparisons, more than two means, were performed using a univariate analysis of variance (ANOVA), where a Scheffe posttest was performed in some cases or Kruskal-Wallis test. GraphPad Prism was used to generate MI 50 curves for cell lines treated with ABA and APC in vitro. In addition, IC 50 curves for FRET assay were also generated by GraphPad Prism. Statistical analyses were performed using IBM SPSS statistics 23.

SUPPLEMENTARY MATERIALS
Supplementary material for this article is available at http://advances.sciencemag.org/cgi/ content/full/5/11/eaav9810/DC1 Supplementary Materials and Methods Fig. S1. Structural information on MBD2 and c-Myc. Fig. S2. SEA and cell migration analysis for the nine selected hit compounds targeting MBD2. Fig. S3. MD simulations of the selected compound-docked structures of MBD2 and c-Myc. Fig. S4. FRET dynamics of ABA and APC to the MBD2-p66 interaction. Fig. S5. Effects of ABA and APC on the expression of EMT markers and CSC properties in various breast and colon cancer cells. Table S1. Molecular docking result (H-bond, hydrogen bond; N/A, not available). Table S2. Selection of compound by in silico assessment of off-target probability by SEA analysis. Table S3. Backbone torsion angle variations (95% confidence interval) of the four key residues in the four different MD simulations of MBD2. Table S4. T test and P values on the backbone torsion angle summarized in table S3. Table S5. Primer sets for vector construction. Original data file S1. Figure 1D PDB files. References  View/request a protocol for this paper from Bio-protocol.