α-Amanitin Restrains Cancer Relapse from Drug-Tolerant Cell Subpopulations via TAF15

Cancer relapse occurs with substantial frequency even after treatment with curative intent. Here we studied drug-tolerant colonies (DTCs), which are subpopulations of cancer cells that survive in the presence of drugs. Proteomic characterization of DTCs identified stemness- and epithelial-dominant subpopulations, but functional screening suggested that DTC formation was regulated at the transcriptional level independent from protein expression patterns. We consistently found that α-amanitin, an RNA polymerase II (RNAPII) inhibitor, effectively inhibited DTCs by suppressing TAF15 expression, which binds to RNA to modulate transcription and RNA processing. Sequential administration of α-amanitin and cisplatin extended overall survival in a cancer-relapse mouse model, namely peritonitis carcinomatosa. Therefore, post-treatment cancer relapse may occur through non-distinct subpopulations and may be effectively prevented by α-amanitin to disrupt transcriptional machinery, including TAF15.

Functional heterogeneity of cancer cell colonies. DTCs are defined as colonies that can survive in the presence of anticancer drugs at respective CoI 50 concentrations ( Fig. 2a; Supplementary Fig. S4). Even in the face of drug treatment, resting or slowly proliferating cells can endure and reinitiate tumor growth 8 . To evaluate whether DTCs possess higher tumorigenicity than both untreated colonies and sheets, we subcutaneously inoculated single DTCs, single untreated colonies, or an equal number (1.0 × 10 4 ) of sheet-derived cells into nude mice. MKN45 cells were used for inoculation because of the highest colony-forming ability among all cell lines tested ( Supplementary Fig. S5). Cells from sheets showed continuous growth in all replicates, whereas only 50% (3/6) of recognizable tumors formed after the inoculation of DTCs and untreated colonies (Fig. 2b,c). The size variations of tumors derived from single DTCs and single untreated colonies suggest that chemotherapy does not necessarily induce or enrich "drug-tolerant" clones 8,12 . Regardless of whether the cells used for xenografting were derived from sheets, colonies, or DTCs, there all have morphological similarities based on histology (Fig. 2d). The variable degree of tumor growth from both DTCs and untreated colonies in vivo may reflect non-deterministic mechanisms relevant to tumor growth rates.
We next investigated whether DTCs have a reversible phenotype based on the finding that a chromatinmediated mechanism regulates gene expression in drug-tolerant cancer cells. 6 If a reversible mechanism plays a role in forming DTCs, then re-formation of DTCs from single-cell suspensions may "reboot" regulatory mechanisms responsible for DTC formation (Fig. 2e). The numbers of secondary DTCs were similar from re-disseminated untreated colonies (Fig. 2f, left). The CoI50 values after re-dissemination of DTCs were not higher than those of untreated colonies (Fig. 2f, right; Supplementary Fig. S6). These results suggest that the drug-tolerant properties of DTCs were lost during the processes involved in single cell suspension, re-dissemination, and re-formation of colonies due to the same reversible mechanisms.
Two major subgroups of DTCs in terms of proteomic profile. To assess changes in the levels of protein markers in DTCs, we performed CoLA assays, which involve a modified reverse-phase protein microarray (RPPA) that uses specific antibodies to detect various kinds of proteins in a quantitative manner 13,14 . We first collected a total of 2400 colonies from five cell lines (HCT116, HT29, HeLa, MCF7, and MKN45) that were exposed to four different kinds of drugs (CIS, DTX, GEF, and SOR) at five different concentrations, as well as from those without drugs ( Fig. 3a; Supplementary Table S1). Cells from these 2400 colonies were then lysed and printed onto nitrocellulose attached to a glass slide (Fig. 3a,b). Quantitative protein measurement was performed on the CoLAs using 44 specific antibodies to assess stemness, pluripotency, and epithelial and mesenchymal markers (Supplementary Table S2). Two-way hierarchical clustering identified two major colony clusters (CC1 and CC2) and two major protein clusters (PC1 and PC2) ( Fig. 3c; Supplementary Fig. S7). CC1 showed high and low levels of epithelial and stemness marker proteins, respectively, whereas CC2 exhibited a reciprocal phenotype. MCF7 and HeLa were significantly enriched in CC1 or CC2 (P < 0.001), respectively, whereas HCT116, HT29, and MKN45 were spread in both clusters (Fig. 3d). Interestingly, the epithelial/non-epithelial relationship indicated that E-cadherin was positively and negatively correlated with CK-8 and vimentin, respectively, in DTCs independent from all drug concentrations ( Fig. 3e; Supplementary Fig. S8). These observations suggest that the two distinct major subgroups of DTCs from any drug condition either existed prior to the drug treatment or were stochastically induced.
Initiation and establishment of DTCs. The hierarchical clustering from the CoLA assay revealed that MKN45 colonies had the most plastic phenotype. Since treatment of MKN45 with CIS was associated with an Scientific RepoRts | 6:25895 | DOI: 10.1038/srep25895 epithelial-low/stemness-high phenotype, the epithelial-low/stemness-high phenotype seemed to be a marker for the DTCs (Fig. 3f). We next investigated the difference in individual protein levels of putative markers for the cancer stem cells (CSC) and the epithelial-mesenchymal transition (EMT) in both CIS-treated and untreated colonies of MKN45. Although levels of several markers, including CD133 and CD24 15,16 , were significantly higher in MKN45 DTCs, none of these markers were exclusively expressed in the DTCs (Fig. 4a), which is consistent with the wide variation of tumorigenicity in individual DTCs (Fig. 2b,c). Moreover, fluorescent immunocytochemistry of CD44 in DTCs and untreated colonies was heterogeneous in both groups (Fig. 4b,c). Although high levels of CD44 have been shown to antagonize reactive oxygen species (ROS) 17 , ROS was indeed observed in both DTCs and untreated colonies. Moreover, CD44 intensity showed no significant correlation with the ROS levels in individual colonies (Fig. 4b,c).
Since changes in levels of the putative CSC and EMT markers did not seem to be directly related to a drugtolerant phenotype, we next focused on pluripotency-associated proteins. The levels of the pluripotency-associated  50 and CoI 50 concentrations for each cell line were determined by a growth suppression assay and colony formation assay, respectively. Asterisks indicate no inhibition. All experiments were performed in triplicate. Error bars represent s.e.m. (c) GI 50 /CoI 50 ratios of each drug were calculated using data in Fig. 1b. Asterisks indicate the ratio is more than 40-fold. (d) Cell cycle state of HeLa.S-Fucci2 cells in sheets and colonies. Top panels show fluorescent images of mCherry-hCdt1 (red) and mVenus-hGeminin (green) expression. Hoechst 33342 (blue) was used to stain nuclear DNA. Scale bars = 100 μ m. Corresponding density scatter plots with the threshold of red, green, yellow and no color cells are shown (bottom panels). Numbers indicate percentages of each fraction.
proteins, including SSEA4, TRA-1-60, NANOG, OCT4A, and SOX2 18,19 , were consistently higher in MKN45 DTCs than those of untreated colonies (Fig. 4d). In contrast to the levels of pluripotency-associated proteins, the mRNA level of pluripotency-inducing transcription factors was downregulated in MKN45 DTCs relative to untreated colonies (Fig. 4e). The methylation status of upstream regions of transcription start sites for  pluripotency-inducing genes was in good agreement with their gene expression levels (Fig. 4f). Interestingly, hierarchical clustering in both proteins and transcripts identified two major clusters that discriminated DTCs from untreated colonies ( Supplementary Fig. S9). The gene expression of pluripotent-inducing factors may be induced and reduced immediately once cells are exposed to drugs, and the proteins remain present longer to establish the drug-tolerant phenotype. This observation may reflect the time lag between transcriptional control and protein turnover in the context of controlling stemness properties in DTCs 20 . Taken together, these observations led us to hypothesize that the stochastic emergence of DTCs may be initiated by transcriptional regulation and subsequently established by protein expression.
High efficacy of RNAPII inhibition on DTC formation. To confirm the transcriptional steps that predominantly affect the emergence of DTCs, we screened a set of four compounds that affect steps, including chromatin formation, transcription, or protein synthesis (Fig. 5a). Trichostatin A (TSA) is an antifungal antibiotic that inhibits class I/II histone deacetylases (HDACs) 21 . Actinomycin D (AMD) is a cyclic polypeptide-containing antibiotic that has clinical applications for treating multiple types of sarcoma by acting as an inhibitor of RNAPI (at 0.05 μ g/ml), RNAPII (at 0.5 μ g/ml), and RNAPIII (at 5 μ g/ml) 22 . α -AMA is a mushroom toxin that inhibits RNAPII 22,23 . In addition, cycloheximide (CHX) is an antifungal antibiotic that inhibits translational peptide chain elongation 24 .
We first examined colony-specific inhibition of these compounds by comparing GI50 and CoI50 values. TSA and α -AMA exhibited more than 10-fold lower CoI50 than GI50 (Fig. 5b). We next determined the CoI50 concentration for each inhibitor and 50-fold higher concentration of the respective CoI50 values were used for 4 and 24-hour exposure of inhibitors in a colony formation assay. TSA clearly suppressed colony formation after 24-hour exposure (Fig. 5c). A complete suppression of colony formation by α -AMA was observed after 4-and 24-hour exposure (Fig. 5c). On the other hand, the less-specific RNAP inhibitor AMD showed no significant colony suppression after 4-and 24-hour exposure, suggesting that the selective inhibition of RNAPII plays a significant role in the inhibitory effect of colony formation (Fig. 5c). CHX produced a slight suppression of colony formation, but the effect of CHX-induced inhibition of protein synthesis reductions in colony formation seemed to be milder than that of epigenetic regulation and mRNA synthesis (Fig. 5c). We also confirmed that both TSA and α -AMA demonstrated remarkable colony suppression following 24-hour exposure, while exposure to α -AMA for 4 hours was sufficient for a nearly complete colony suppression effect for all cell lines tested (Fig. 5d). Although chromatin modification may be one of the major causes behind alterations in gene expression during colony formation 25 , these findings suggest that RNAPII-dependent inhibition of early-phase mRNA synthesis is sufficient to produce a nearly complete suppression of colony formation. The significance of mRNA synthesis during colony formation was also demonstrated in a colony formation assay-based screening with a compound library showing that RNA polymerase inhibition was ranked among the top 2.5% categories of 160 mechanism of actions ( Supplementary Fig. S10).
To examine whether colony suppression can be reproduced in vivo, we next observed nude mice after peritoneal injection of 1.0 × 10 6 MKN45 cells that had been treated with α -AMA for 4 hours. Mice receiving a peritoneal injection of MKN45 cells showed a constant decrease in body weight, while those mice injected with α -AMA-treated cells maintained their body weight, likely due to reduced alimentary tract functional failure (Fig. 5e). At 28 days after peritoneal injection of cancer cells, the number of peritoneal nodules was significantly lower in mice that received α -AMA-treated cancer cells relative to those injected with untreated MKN45 cells (Fig. 5f). A subsequent colony formation assay showed that α -AMA has the most potent activity in DTC suppression among the four compounds (Fig. 5g).
RNAPII-dependent TAF15-mediated drug tolerance. We next screened the potential molecular target of α -AMA based on transcriptional DNA microarray data using MCF7 DTCs that exhibited a high CIS tolerance (Fig. 1b). Among the top 2.5% of specifically-induced genes in DTCs, we focused on TAF15 because its gene product binds RNAPII 26,27 (Fig. 6a). The promoter methylation status and protein expression level of TAF15 in MKN45 DTCs were in good agreement with the gene expression data (Fig. 6b,c). TAF15 encodes TATA-binding protein (TBP)-associated factor 15, which is a member of the TET (TLS/FUS, EWS, TAF15) family of RNA-binding proteins 27 . A recent study demonstrated that the TAF15 protein binds the C-terminal domain of RNAPII more avidly than EWS and TLS/FUS proteins 28 . Our transcriptional screening with DTCs demonstrated that the level of TAF15 was higher than EWS and TLS/FUS levels ( Fig. 6a; Supplementary Fig. S11). The protein level of TAF15 was higher in DTCs whereas EWS and TLS/FUS proteins did not show increased levels in DTCs ( Fig. 6c; Supplementary Fig. S11). Thus, we postulate that TAF15 is specifically involved in DTC formation among the TET family. The most abundant TAF15-associated RNA component has been reported to be U1 small-nuclear (sn)RNA, which is part of the splicing machinery 29,30 . The level of TAF15-associated U1 snRNA in DTCs and untreated colonies was not significantly different ( Supplementary Fig. S12). These findings suggest that the U1 snRNA-binding capacity of TAF15 is not predominantly involved in DTC formation.
We next focused on a causal link between TAF15 upregulation and α -AMA sensitivity in DTC-forming cells. In response to α -AMA treatment, both TAF15 mRNA and TAF15 protein levels were decreased in MKN45 cells, suggesting that the RNAPII activity towards TAF15 mRNA was inhibited by α -AMA (Fig. 6d,e). Notably, a similar morphological change was observed in MKN45 cells following both TAF15 knockdown and α -AMA treatment (Fig. 6f). A subsequent colony formation assay revealed that TAF15 knockdown suppressed the emergence of both DTCs and untreated colonies, suggesting that TAF15 is a crucial target of α -AMA in the context of DTC suppression ( Fig. 6g; Supplementary Fig. S13). In contrast, the overexpression of TAF15 did not result in a greater number of DTCs compared to the mock control (Supplementary Figs S13 and S14). These results suggest that α -AMA treatment may present a therapeutic opportunity for diseases caused by DTCs, such as PC by chemotherapy, via inhibition of TAF15 mRNA synthesis.
Scientific RepoRts | 6:25895 | DOI: 10.1038/srep25895 α-AMA/CIS-combined treatment restraining PC in nude mice. Selective suppression of DTC emergence by α -AMA in vitro offers a therapeutic opportunity to prevent PC. However, α -AMA ingestion was reported to be directly linked to death due to toxicities in the gastrointestinal tract, liver, and kidney 31 . Therefore, we evaluated both the therapeutic and toxicological properties of α -AMA using a PC model established by intraperitoneal inoculation of MKN45 cells followed by drug treatment. The body weights of mice after tumor cell inoculation decreased sharply starting on day 10 in α -AMA alone, CIS alone, and no treatment groups, whereas groups receiving serial administration of α -AMA and CIS maintained their body weight (Fig. 7a). The nodules spread mainly in the mesentery and ranged from 1 to 5 mm in diameter in mice from the α -AMA, CIS, and no  treatment groups, while the majority of the mice in the α -AMA/CIS group exhibited very few nodules (Fig. 7b). In fact, the average number of nodules in the α -AMA/CIS group was significantly decreased, although CIS and α -AMA single administration groups exhibited a similar suppression effect, but to a lesser degree (Fig. 7c). The overall survival of mice in the α -AMA/CIS group was significantly longer than for those in the α -AMA group, suggesting that the α -AMA toxicity was reduced by CIS combination (Fig. 7d). A comparison of the average time to death of the four treatment groups revealed that only the α -AMA/CIS group exhibited a significant prolongation in overall survival (Fig. 7e). The extension of the average time to death with α -AMA/CIS treatment was approximately 10 days within the 70-day period compared to other treatment groups. Subsequently, we performed pathological examination and a liquid chromatography-mass spectrometry (LC-MS) analysis to assess the effect of CIS administration on α -AMA toxicity. Pathological examination of the liver and kidney at 48 hours after α -AMA administration revealed no remarkable findings at a dose of 0.4 mg/kg, which was used in the PC mouse model (Supplementary Fig. S15). Substantial pathological changes were observed only at higher doses (1.2 and 3.6 mg/kg). An LC-MS analysis for α -AMA concentration revealed that a higher dose than 3.6 mg/kg was needed to detect α -AMA from the liver samples ( Supplementary Fig. S16). Interestingly, the α -AMA level in the liver was 42% lower in the α -AMA/CIS group compared to the α -AMA single administration group, suggesting that CIS may inhibit α -AMA accumulation in the liver (Fig. 7f). In principle, these results indicate that: (i) the number of nodules in mesentery may be a surrogate marker of overall survival in PC; and (ii) α -AMA possesses the capacity to prevent cancer relapse due to PC when it is administered with CIS. Taken together, the combination of α -AMA and CIS appears to elicit the potential suppression effect of α -AMA for PC while also reducing its toxicity.

Discussion
We used DTCs as an in vitro model for relapsed cancer cells, which are subpopulations that survive in the presence of anticancer drugs. Each DTC can be initiated from either a single cell or an extremely small number of cells, which mimics the process of cancer relapse in humans after curative surgery followed by adjuvant chemotherapy. Proteomic characterization of DTCs using CoLA technology showed that there were two distinct groups of cells based on stemness-and epithelium-associated protein expression, while the expression of a set of genes known to be associated with cellular pluripotency appeared to affect colony formation 19,32 . These results led us to hypothesize that DTC formation may be suppressed by inhibiting fundamental transcriptional machinery at a relatively early stage of DTC formation and prior to protein synthesis. Consistent with the finding that transcription is involved in DTC formation, the RNAPII inhibitor α -AMA demonstrated a significant reduction in DTC formation. The mechanism of action for α -AMA appeared to involve down-regulation of TAF15, which participates in RNAPII-dependent transcriptional machinery 27 . Importantly, α -AMA/CIS exhibited an inhibitory effect on PC in vivo, suggesting that α -AMA may be used for the prevention of PC with conventional chemotherapy.
A variety of mechanisms for drug tolerance have been proposed, including target gene mutation 33 , adaptive resistance 34,35 , selection of intrinsically resistant populations 36 , and epigenetic alterations 6 . Empirical clinical observations suggest that cancer relapse does not always result from a specific population that acquired drug-tolerant properties in response to administered anticancer drugs, but may instead be due to undefined populations among total heterogeneous cancer cell populations that happen to have drug tolerance 37 or acute adaptive responses to a particular drug 35 . In fact, complete suppression of cancer relapse is often extremely challenging because of difficulties in: (i) selectively targeting cancer initiating cells; (ii) defining effective drug combinations; and (iii) selecting inhibitors that affect the fundamental components of cell division. Although there are no definitive clinical guidelines for all relapsed situations, conventional DNA damaging agents or taxanes are often used for PC patients 38,39 . In the present study, we used CIS in a combinational approach with α -AMA to elucidate its therapeutic potential based on molecular fraction screening for colony formation inhibition. α -AMA, which is a bicyclic octapeptide produced by the basidiomycetes mushroom, Amanita phalloides, is a potent inhibitor of RNAPII 23 . Very few therapeutic opportunities that exploit α -AMA have been pursued because of its high gastrointestinal tract, liver, and renal toxicities 31 . However, recent development of a cancer immunotherapy using an anti-EPCAM monoclonal antibody conjugated with α -AMA demonstrated clear growth suppression effects in a human pancreatic cancer mouse skin xenograft model without systemic side effects even at a dose of 0.1 mg/kg 40 . Our model mimics human PC after apparently "curative removal" of solid tumors in the abdomen. The combinational therapy of α -AMA and CIS indicates that inhibition of RNAPII is critical for cells to develop as a mass in the peritoneum. Importantly, our treatment with α -AMA did not show significant lethal side effects despite peritoneal injections at a dose of 0.4 mg/kg. The mouse median lethal dose for α -AMA was estimated to be higher than 1.2 μ g/kg in our validation. Importantly, α -AMA accumulation in the mouse liver was inhibited by CIS administration. These results suggest that the α -AMA dose used in the present study is tolerable and sufficient for an antitumor effect.
While searching for molecular targets of α -AMA, the TAF15 gene, which encodes a transcription factor that binds RNAPII directly, was identified. TAF15 was first isolated as a TATA-binding protein-associated factor (TAF) in the general transcription factor IID complex that initiates RNAPII-dependent transcriptional machinery and pre-mRNA splicing 27,41 . A protein family consisting of translocated in liposarcoma or fusion in sarcoma (TLS/ FUS), Ewing's sarcoma (EWS), and TAF15 is called the TET (TLS/FUS, EWS, TAF15) family and has various roles in gene expression 42 . TAF15 is involved in chromosomal translocations in acute leukemia 43 and extraskeletal myxoid sarcomas 44 , resulting in fusion proteins that may have oncologic potential through enhanced cell spreading and adhesion 45 . For TAF15-targeted immunotherapy, a human IgG PAT-BA4 was developed against the tumor-specific TAF15 variant 46 . Although the therapeutic effect of PAT-BA4 has been limited in vitro, TAF15 targeting therapy has demonstrated anti-adhesion and anti-migration effects in cancer cells, which indicates the importance of TAF15 for increased malignant potential. In the present study we demonstrated that TAF15 mRNA was one of the crucial components for initiating DTC formation. However, from the transcriptional profiling of DTCs, the formation and establishment of DTCs does not appear to be a process that requires merely a small number of specific genes or proteins. In fact, at least several hundred genes lie downstream of TAF15 47 . Hence, we suggest that an agent that can inhibit mRNA synthesis over a wide spectrum, such as α -AMA, may be more appropriate for preventing cancer relapse than those compounds that target a single molecule. From a therapeutic point-of-view, whether α -AMA doses having sufficient antitumor effects would be tolerable in humans either by ingestion or peritoneal injection remains to be determined 48 .
Scientific RepoRts | 6:25895 | DOI: 10.1038/srep25895 In summary, our study demonstrated that RNAPII-dependent mRNA synthesis is a fundamental process for DTC formation and establishment of PC. These findings provide the foundation for a promising strategy to prevent cancer relapse after apparently successful surgical and chemotherapeutic treatments. Drugs. CIS and DTX were purchased through Nippon Kayaku (Tokyo, Japan) and Sanofi KK (Tokyo, Japan), respectively; TSA was purchased through Wako (Osaka, Japan). AMD, α -AMA and CHX were purchased through Sigma-Aldrich (St. Louis, MO, USA); GEF and SOR were purchased through Tocris (Bristol, UK) and Cell Signaling Technology (Danvers, MA, USA), respectively.
Growth Suppression and Colony formation Assay. For growth suppression assay, cells were plated in a 96-well plate at 3.2 − 12.8 × 10 4 cells/cm 2 . The cell viability was determined using CCK-8 (Dojindo, Kumamoto, Japan) and a TriStar LB 941 microplate reader (Berthold Technologies, Bad Wildbad, Germany). For colony formation assay, drugs or compounds were dispensed into 48-well plates before cell dissemination. Cells were then plated at 1.3 − 2.6 × 10 2 cells/cm 2 . Colonies were stained with 0.1% (w/v) crystal violet.
CoLA. For colony lysate preparation, a total of 2400 colonies were isolated individually using suction produced by a micropipette, and individual cell pellets were lysed with 10 μ l Pink Buffer as previously described 14 . The lysates were dotted onto nitrocellulose membranes attached to a glass slide using a 2470 Arrayer (Aushon Biosystems, Billerica, MA, USA). CoLAs were incubated with specific primary antibodies (Supplementary Table S2) followed by analysis with a Catalyzed Signal Amplification System (Dako Japan, Tokyo, Japan). For detection of total protein levels, CoLAs were stained with colloidal gold (Bio-Rad, Hercules, CA, USA). An optical flatbed GT-X970 scanner (EPSON, Suwa, Japan) with the resolution set to 16-bit, 2400 dpi was used for image acquisition. To prevent skewing of image intensities, a Wedge Density Strip (Danes-Picta, Praha, Czech Republic) was used for readout range calibration 49 . Spot images were converted to raw pixel values using P-SCAN software 50 or WinPscan (http://abs.cit.nih.gov/pscan; http://www.nishizukalab.org).
Xenografts. For tumorigenicity tests, six colonies (untreated) and DTCs derived from MKN45 cells were individually injected subcutaneously into the left and right side of the backs of six 6-week-old female nude mice (BALB/cAjcl-nu/nu, CLEA Japan, Tokyo, Japan). These mice were monitored for 49 days after the inoculation or until tumors reached 10 mm in the largest diameter, and were then euthanized. For the PC model, 1.0 × 106 MKN45 cells were injected intraperitoneally into six 6-week-old female nude mice (BALB/cAjcl-nu/nu, CLEA Japan). Mice were then treated with CIS (4.0 mg/kg, intraperitoneal administration) or a combination of CIS and α -AMA (0.4 mg/kg, intraperitoneal administration). For the combination treatment, α -AMA was given 24 hours before CIS. Body weight was monitored for 28 days after the treatment. All experiments were performed in accordance with the approval of the Iwate Medical University Ethical Committee for Animal Experiment Regulation (23-067, 25-023, and 26-004).
Visualized Cell Cycle Analysis. The cell cycle in sheets and colonies of HeLa cells was monitored using the Fucci system (HeLa.S-Fucci2) 11 . In this system, cells in the S/G 2 /M or G 1 phase express Venus-hGem or mCherry-Cdt1 protein, respectively. HeLa.S-Fucci2 cells were seeded as cell sheets (3.0 × 10 4 cells/cm 2 ) or as colonies (1.0 × 10 2 cells/cm 2 ). Cells were treated for 48 hours with CIS GI 50 for sheets, and for 13 days with CIS CoI 50 for colonies. Conditions for untreated cells were used as a control. For quantification of the fluorescence intensity of individual cells, cell sheets and colonies were diluted to single cell suspensions, and the fluorescence intensity was quantified using a Tali Image-Based Cytometer (Life Technologies, Carlsbad, CA, USA). The 75 th percentile of signal in untreated or CIS-treated sheets was defined as Venus-or mCherry-positive cells, respectively ( Supplementary Fig. S3). DNA Microarray. Total RNA was extracted from two biological replicates of bulk untreated colonies and DTCs of MCF7 emerged in the presence of 0.8 μ M CIS using the RNeasy Kit (QIAGEN, Tokyo, Japan). Gene expression profiling was performed according to the manufacturer's instructions (SurePrint G3 Human Gene Expression 8 × 60 K microarrays, Agilent Japan, Tokyo, Japan). Raw data were normalized by dividing each probe signal by the 75 th percentile of the entire signal. For identification of the TAF15 gene, gene ontology analysis was performed using Database for Annotation, Visualization and Integrated Discovery (DAVID) software tools (http://david.abcc.ncifcrf.gov/). The DNA microarray data (GSE65419) is available via Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo/). Methylated DNA Immunoprecipitation Combined with Microarray. Methylated DNA immunoprecipitation combined with microarray (MeDIP-chip) analysis was performed as described previously 51,52 . Genomic DNA was extracted from bulk untreated colonies and DTCs of MKN45 emerged in the presence of 0.2 μ M CIS using the DNeasy Kit (QIAGEN). Fluorescent labeling, microarray hybridization, washing and scanning were performed according to the manufacturer's instructions (Human Meth 385K Prom Plus CpG Arrays, NimbleGen, Madison, WI, USA). Immunoprecipitated DNA samples from untreated colonies and DTCs were labeled with Cy3 and Cy5, respectively. Each raw data point was first normalized by Z score transformation, and then calculated as a Cy5/Cy3 log 2 ratio. Promoter regions with differentially methylated CpG sites were identified using SignalMap Software version 2.0 (NimbleGen). The MeDIP-chip data (GSE69203) is available via Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo/). Immunocytochemistry. Cells were seeded onto 96-or 48-well plates and incubated for 13 days in the presence or absence of 0.2 μ M CIS. Colonies were then fixed in 4% paraformaldehyde (PFA) for 15 min at room temperature and stained with 4′ ,6-diamidino-2-phenylindole (DAPI) for 20 min. After washing, cells were blocked in 5% BSA and 0.1% Tween 20 before staining with anti-CD44 antibody (156-3C11, Cell Signaling Technology, Danvers, MA, USA) at 4 °C overnight. After washing with phosphate buffered saline (PBS), CF488A-conjugated secondary antibodies (Biotium, Hayward, CA, USA) were applied for 1 hour at room temperature. After washing, wells were filled with PBS. Images were collected on an IN Cell Analyzer 2000 automated microscope and deconvoluted using IN Cell Analyzer 2000 software (GE Healthcare). The deconvoluted image projections were analyzed using Developer (GE Healthcare) to identify colonies.

ROS Detection.
For ROS detection, colonies were incubated with the cell-permeable fluorescent indicator CellROX Deep Red (Life Technologies) and Hoechst 33342 at 37 °C for 30 min. After washing, wells were filled with PBS and images were collected as described in the Immunocytochemistry section. Immunocytochemistry for CD44 was performed after ROS detection.
qRT-PCR. For single-colony analysis in Fig. 4e, total RNA was extracted from 10 individually isolated untreated colonies and DTCs derived from MKN45 cells using the RNeasy kit (QIAGEN). A 1-mm diameter single MKN45 colony (approximately 1.0 × 10 4 cells) yielded approximately 100 ng of total RNA. cDNA was synthesized from 50 ng total RNA in a 10 μ l reaction volume using PrimeScript RT Master Mix (TaKaRa Bio, Otsu, Japan).
For the time course analysis in Fig. 6d, total RNA was extracted from 1.0 × 10 6 MKN45 cells treated with DMSO or α -AMA for 0, 4, 24, and 48 hours. cDNA was synthesized from 500 ng total RNA in a 10 μ l reaction volume using PrimeScript RT Master Mix (TaKaRa Bio). Quantitative RT-PCR (qRT-PCR) was performed using the LightCycler Nano System (Roche, Mannheim, Germany). Primer sequences are listed in Supplementary Table S3.
Western Blotting. Cells were harvested by trypsinization, and lysed with Pink Buffer as previously described 14 .
Ten micrograms of lysate was resolved on NuPAGE 4-12% Bis-Tris gels (Life Technologies) and electrophoretically transferred to a nitrocellulose membrane using iBlot Dry Blotting System (Life Technologies). Membranes were blocked for 1 hr in 5% BSA in Tris-buffered saline with Tween 20 (TBST, Dako), and then incubated overnight with specific primary antibodies in 5% BSA in TBST. Horseradish peroxidase (HRP)-conjugated anti-rabbit IgG secondary antibody (1:2000, Cell Signaling Technologies) was diluted in 5% nonfat dry milk in TBST. Chemiluminescence detection reagents were incubated with the membrane, and an image was acquired using an Image Quant LAS500 (GE Healthcare). Band intensities were quantified using ImageJ version 1.49q (http:// rsbweb.nih.gov/ij/). Primary antibodies used in this analysis are shown in Supplementary Table S2.
Knockdown of TAF15. For TAF15 knockdown, 1.0 × 10 5 MKN45 cells were transfected with 5 nM siRNA using Lipofectamine RNAiMAX (Life technologies) for 48 hours at 37 °C in a humidified incubator supplied with 5% CO 2 . Morphological changes were observed after 48 hours. To assess the effect of TAF15 knockdown on DTC generation, cells were disseminated in the presence or absence of 0.2 μ M CIS at low density (2.0 × 10 2 cells/cm 2 ). Approximately 100 colonies per well from the siRNA-transfected cells appeared in the control (i.e., without CIS) 13 days after the dissemination. Colony counting was performed as described in the Colony Formation Assay section. siRNAs sequences were described elsewhere 47 .