## Introduction

Patient stratification for cancer therapy is an excellent illustration of precision medicine, and biomarker-based treatment selection has tremendously aided in the success of current cancer therapies1. In this sense, the current ability to molecularly define and differentiate breast cancer (BCa) into molecular subtypes2,3 has allowed the identification of patients at risk of relapse4 and has led to biomarker signatures used to spare low-risk patients from aggressive chemotherapy5.

Tumours are heterogeneous entities and most cancers retain a differential fraction of cells with increased self-renewal capability (cancer stem or initiating cells)6. Cancer-initiating cells (CICs) exhibit a unique spectrum of biological, biochemical and molecular features that have granted them an important role in disease recurrence and metastatic dissemination in BCa7,8. Despite the accepted relevance of CICs in cancer progression, the molecular cues governing their activity and function remain largely unknown. The sex determining region Y Box 9 (SOX9) is a recently described regulator of cell differentiation and self-renewal9,10,11 and is found upregulated in BCa12,13,14.

The promyelocytic leukaemia (PML) protein negatively regulates survival and proliferation pathways in cancer, functions that have established it as a classical pro-apoptotic and growth inhibitory tumour suppressor15,16. PML is the essential component of multi-protein sub-nuclear structures commonly referred to as the PML nuclear bodies. PML multimerizes to function as a scaffold critical for the composition and assembly of the entire complex, a process that is regulated by Small Ubiquitin-like Modifier (SUMO)-mediated modifications and interactions15,16. Despite the general perception of being PML a bona fide tumour suppressor in cancer, a series of recent studies have demonstrated that PML exhibits activities in cancer that go far and beyond tumour suppression17. The work in chronic myeloid leukaemia has evidenced that PML expression can be promoted in certain cancers, providing a selective advantage to tumour cells18,19. Moreover, PML is found upregulated in a subset of BCa20. However, to which extent PML targeting could be a valuable therapeutic approach in solid cancers remains obscure.

In this study, we reveal the therapeutic and stratification potential of PML in BCa and the molecular cues, underlying the therapeutic response unleashed by PML inhibition.

## Results

### PML silencing hampers BCa-initiating cell capacity

The elevated expression of PML in a subset of BCa17,20 strongly suggests that it could represent an attractive target for therapy. To ascertain the molecular and biological processes controlled by PML in BCa, we carried out short hairpin RNA (shRNA) lentiviral delivery-mediated PML silencing in different cellular systems. Four constitutively expressed shRNAs exhibited activity against PML (Fig. 1a; Supplementary Fig. 1a–d). PML knockdown elicited a potent reduction in the number of ALDH1-positive cells and in oncosphere formation (OS, readout of self-renewal potential7,21), in up to three PML-high-expressing basal-like BCa (BT549 and MDA-MB-231) or immortalized (HBL100) cell lines tested (Fig. 1b–d; Supplementary Fig. 1e–g). This phenotype was recapitulated with a doxycycline-inducible lentiviral shRNA system targeting PML (sh4; Fig. 1e,f; Supplementary Fig. 1h).

Self-renewal capacity is a core feature of CICs7. On the basis of this notion, we hypothesized that PML could regulate tumour initiation in BCa. We performed tumour formation assays in immunocompromised mice, using MDA-MB-231 cells (PML-high-expressing triple-negative breast cancer (TNBC)) transduced with non-targeting (shRNA Scramble: shC) or PML-targeting shRNAs. PML silencing exhibited a profound defect in tumour formation capacity, resulting in a decrease in the frequency of tumour-initiating cells from 1/218 (shC) to 1/825 (sh5) and completely abolished (1/infinite) in sh4 (Fig. 1g; Supplementary Fig. 1i).

To extrapolate these observations to the complexity of human BCa, we characterized a series of patient-derived xenografts (PDXs; Table 1; Supplementary Fig. 1j). The distribution of PML expression in the different subtypes of engrafted tumours was reminiscent of patient data, with a higher proportion of PML-high-expressing tumours in basal-like/triple-negative subtype20. Taking advantage of the establishment of a PML-high-expressing PDX-derived cell line (PDX44), we sought to corroborate the results obtained in the PML-high-expressing cell lines. As with MDA-MB-231 cells, PML silencing was effective in the PDX44-derived cell line (Fig. 1h) and resulted in a significant decrease in OS formation (Fig. 1i). In vivo, PML silencing decreased tumour-forming capacity of PDX44 cells (tumour-initiating cell frequency was estimated of 1/39.6 in shC, 1/100 in sh5 and 1/185 in sh4; Fig. 1j; Supplementary Fig. 1k).

These data demonstrate that PML expression is required for BCa-initiating cell function in TNBC cells.

### PML sustains metastatic potential in BCa

CIC activity is associated with tumour initiation and recurrence7,22. We have previously shown that PML expression is associated to early recurrence20, which we validated in an independent data set23 (Supplementary Fig. 2a). The development of metastatic lesions is based on the acquisition of novel features by cancer cells24. On the basis of our data, we surmised that the activity of PML on CICs could impact on the survival and growth in distant organs. To test this hypothesis, we measured metastasis-free survival (MFS) in two well-annotated large messenger RNA (mRNA) data sets3,25,26. First, we evaluated the impact of high PML expression in MFS in the MSK/EMC (Memorial Sloan Kettering Cancer Center-Erasmus Medical Center) data set25,26. As predicted, PML expression above the mean was associated with reduced MFS (Fig. 2a). Second, we validated this observation in the METABRIC data set, focusing on early metastasis (up to 5 years)3. On the one hand, we confirmed the MSK/EMC data (Fig. 2b; hazard ratio (HR)=1.31, log-rank test P=0.006). On the other hand, a Cox continuous model demonstrated an association of PML expression with the increased risk of metastasis (HR=2.305, P=0.002). Of note, we tested the expression of PML in patients with complete pathological response or residual disease after therapy27, but could not find a significant association of these parameters in two data sets (Supplementary Fig. 2b).

The molecular alterations associated to metastatic capacity can be studied using BCa cell lines, in which metastatic cell sub-clones have been selected through the sequential enrichment in immunocompromised mice28. If PML is a causal event in the acquisition of metastatic capacity, then changes in its expression should be observed in this cellular system. As predicted, PML mRNA and protein expression were elevated in three distinct metastatic sub-clones compared with their parental counterparts25,26 (Fig. 2c; Supplementary Fig. 2c).

Metastasis surrogate assays provide valuable information about the capacity of cancer cells to home and colonize secondary organs29. TNBC cells exhibit metastatic tropism to the lung30, and the molecular requirements of this process have begun to be clarified through the generation of highly metastatic sub-clones31. Our patient analysis suggests that PML expression is favoured in primary tumours, with higher capacity to disseminate. Moreover, cell sub-clone analysis further reveals that PML expression is selected for in the process of metastatic selection. With this data in mind, we asked to which extent PML would be responsible for the enhanced metastatic capacity. To address this question, we silenced PML in a highly metastatic sub-clone derived from MDA-MB-231 and injected these cells in the tail vein of nude mice. We chose tail vein injection due to the fact that other metastasis models based on the orthotopic implantation of cells in the mammary fat pad25 are influenced by primary tumour formation, which we reported to be altered by PML (Fig. 1). The reduction of PML was confirmed in the injected cells (Fig. 2d). Strikingly, PML silencing led to a significant reduction in lung metastatic foci formation (Fig. 2e). When evaluating the immunoreactivity of PML in the metastatic lesions (Fig. 2e–g), we observed a direct association between PML silencing at the time of injection (Fig. 2d) and the immunoreactivity of PML in metastatic foci (Fig. 2g). We evaluated whether the lack of PML could be limiting metastatic growth capacity by eliciting an apoptotic response, rather than CIC capacity. However, no differential apoptosis was detected by the means of cleaved caspase-3 staining (Supplementary Fig. 2d–e).

These data demonstrate that the genetic targeting of PML results in a tumour-suppressive response, characterized by decreased BCa-initiating cell function and consequently, reduced tumour initiation and metastasis.

Our data demonstrate that PML is transcriptionally regulated in BCa. PML gene expression is regulated upon various external stimuli, including type I and II interferons and interleukin 6, which are mediated by interferon regulatory factors and signal transducers and activators of transcription (STATs), respectively32,33,34,35. Specifically, it has been reported that activated STAT3 but not STAT1 correlates with PML mRNA and protein levels in fibroblasts, HeLa and U2OS cell lines34. Since, STAT3 is activated in oestrogen receptor (ER)-negative BCa36, we hypothesized that this transcription factor may be responsible for the transcriptional activation of PML in this tumour type. We silenced STAT3 with two different short hairpins (sh41 and sh43), and showed that this approach led to the decrease in PML protein and gene expression in the different cell lines tested (Fig. 3a; Supplementary Fig. 3a–b). Moreover, pharmacological inhibition of the Janus kinase/signal transducers and activators of transcription (JAK/STAT) pathway at two different levels (SI3-201, an inhibitor of STAT3 phosphorylation and activation; TG1013148, a potent and highly selective ATP-competitive inhibitor of JAK2) decreased PML levels (Fig. 3b,c). In coherence with the activity of PML, genetic and pharmacological inhibition of STAT3 in MDA-MB-231 cells reduced the primary OS formation capacity (Fig. 3d–f). Importantly, PML gene expression levels in a cohort of 448 patients (MSK/EMC) correlated with the activity of STAT3, as confirmed with two different STAT3 signatures (Fig. 3g; ref. 37; http://software.broadinstitute.org/gsea/msigdb/cards/V$STAT3_01). In addition, immunohistochemical analysis confirmed an association between the PML immunoreactivity and phosphorylated STAT3 levels in the Marseille cohort (Fig. 3h). Our results provide strong support for the role of STAT3 as an upstream regulator of PML in BCa. ### Elevated PML expression predicts response to arsenic trioxide PML can be pharmacologically inhibited with arsenic trioxide (Trisenox, ATO), which induces SUMO-dependent ubiquitylation and proteasome-mediated degradation of the protein38,39. Similar to our results obtained by knocking down PML via shRNA, low doses of ATO decreased PML levels and exerted a negative effect on the OS formation capacity both in MDA-MB-231 and PDX44 cells (Fig. 4a,b). Moreover, ATO reduced the tumour formation capacity in a xenograft model derived from MDA-MB-231 cells in full coherence with the genetic approach (tumour-initiating cell frequency was estimated of 1/279 in vehicle and 1/703 in ATO; Fig. 4c; Supplementary Fig. 4a). We hypothesized that cells with elevated PML would be ‘addicted’40 to the expression of the protein and hence be more sensitive to the action of PML inhibitors. To prove this notion, we studied additional cell lines with high (BT549, HBL100) or low (MCF7, T47D) PML expression. With this approach, we could demonstrate that the effect of PML silencing on the OS formation was exquisitely restricted to PML-high-expressing cells (Fig. 4d). This effect was recapitulated with ATO (Fig. 4e), where PML-low cells remained refractory to the drug in terms of the OS formation capacity. Our results open a new avenue for the treatment of tumours that exhibit elevation in PML expression. PML elevation is predominant in ER-negative tumours (Supplementary Fig. 4b), which also present worse prognosis than ER-positive BCa2,41. Whereas luminal subtypes present better overall prognosis, there is a subset of patients within this subtype that exhibits aggressive disease42. We hypothesized that within this PML-low-expressing subtypes, the worse prognosis subgroup would exhibit increased PML levels. Indeed, MFS analysis within each intrinsic subtype confirmed that ER-positive BCa (luminal A and luminal B) contained a subset of patients with higher PML and worse prognosis (Supplementary Fig. 4c–g). Our results in ER-positive tumours indicate that the PML expression is enriched in patients harbouring tumours of poor prognosis2,3. These results are coherent with our data in metastatic clone selection (Fig. 2c), suggesting that the acquisition of aggressive features is accompanied by the elevation of PML expression and ‘addiction’ to the protein. We therefore sought to study whether metastatic ER-positive cell sub-clones, which present elevated PML expression, would exhibit sensitivity to PML inhibition, in contrast to the parental cells. Indeed, ATO reduced the OS formation selectively in PML-high-expressing metastatic cells derived from MCF7, whereas the parental cells remained refractory to the drug (Fig. 4f,g). Our results strongly suggest that PML elevation in BCa is associated to a dependence on its expression and hence enforces the need for patient stratification based on PML levels before the establishment of PML-directed therapies. ### PML regulates BCa-initiating cell function through SOX9 To ascertain the molecular mechanism by which PML regulates BCa-initiating cell function, we first evaluated the expression levels of this gene in a sorted population of ALDH1-positive versus -negative MDA-MB-231 cells (Fig. 5a,b), and in adherent cultures versus OS (CIC-enriched cultures) (Fig. 5c). Strikingly, PML expression increased in both experimental approaches (Fig. 5b,c), together with the levels of well-established stem cell regulators (Fig. 5c). On this basis, we hypothesized that PML might control the expression of stem cell factors, as a mean to regulate BCa-initiating cell function. SOX9 is a recently described regulator of cell differentiation and self-renewal9,10,11 and is upregulated in BCa12,13,14. Constitutive (Fig. 5d; Supplementary Fig. 5a–b) and inducible (Fig. 5e) PML silencing exerted an inhibitory effect on SOX9 expression that correlated with the OS formation capacity (Fig. 5f). PML pharmacological inhibition also induced a decrease on SOX9 expression (Fig. 5g; Supplementary Fig. 5c–d). This regulatory activity was corroborated in the PDX44 cell line (Fig. 5h,i; Supplementary Fig. 5e), and in a correlative manner in the PDX data set (Fig. 5j), as well as in the aforementioned Marseille data set (Fig. 5k). We next ascertained the molecular cues regulating SOX9 expression downstream PML. Since the regulation was observed at the mRNA level, we interrogated SOX9 promoter in silico and in public datasets. The ENCODE project has provided a vast amount of information about regulators and binding sites43. SOX9 promoter exhibited a 2 kb region of acetylated H3K27 (H3K27Ac), which would indicate the proximal regulatory region. To our surprise, we found PML among the 10 proteins with highest confidence DNA-binding score in SOX9 promoter region (Fig. 5l; cluster score=527 (refs 44, 45, 46)). There is limited evidence of the capacity of PML to regulate gene expression in concordance with transcription factors through association with DNA47,48,49. We performed chromatin immunoprecipitation (ChIP) analysis of ectopically expressed and endogenous PML, as well as SOX9 expression analysis in these conditions. We confirmed that PML is in close proximity to SOX9 promoter region and that its ectopic expression upregulates SOX9 transcript and protein levels (Fig. 5m; Supplementary Fig. 5f–j). It is worth noting that PML does not present a canonical DNA-binding domain, and it is therefore plausible that it lies in close proximity to SOX9 promoter through the interaction with intermediary DNA-binding proteins. The regulation of SOX9 led us to hypothesize that this transcription factor mediated the effects of PML on the regulation of CIC function. On the one hand, we ascertained whether SOX9 silencing would recapitulate the effects of PML inhibition. We set up two shRNAs targeting SOX9 (Fig. 6a) that exhibited a potent effect on primary (Fig. 6b) and secondary (Supplementary Fig. 6a) OS formation. Moreover, SOX9 silencing in MDA-MB-231 cells reduced the tumour formation capacity in vivo (tumour-initiating cell frequency was estimated of 1/71.7 in shC, completely abolished (1/infinite) in sh9.1 and 1/4145.5 in sh9.2; Fig. 6c; Supplementary Fig. 6b–c), in agreement with other reports12. On the other hand, we evaluated the capacity of ectopically expressed SOX9 to bypass the effects of PML silencing on CIC function. Ectopic SOX9-expressing BCa cells were refractory to PML genetic inhibition in terms of the OS formation (Fig. 6d,e) and tumour formation (tumour-initiating cell frequency was estimated of 1/139.8 in shC/Mock, 1/57.5 in shC/SOX9, 1/1506 in sh4/Mock and 1/270.8 in sh4/SOX9; Fig. 6f; Supplementary Fig. 6d). Importantly, the in vitro observation was recapitulated in ATO-treated cells (Fig. 6g). These data reveal a novel molecular mechanism by which PML controls the expression of the stem cell factor SOX9 to regulate BCa-initiating cell function (Fig. 6h). It is worth noting that we found PML at the promoter region of other stem cell genes, such as LGR5 (Supplementary Fig. 6e–g), indicating that the capacity of this protein to regulate CIC function could involve a larger and more complex transcriptional program. ## Discussion Finding successful targeted treatment strategies for women at risk of metastatic BCa is of outstanding clinical interest. Our data unveil the therapeutic potential of targeting PML in combination with a stratification companion that identifies patients harbouring PML-high-expressing BCa. We demonstrate that PML targeting impacts on BCa-initiating cell function, and hence on cancer initiation and dissemination. In addition, we observed that PML expression is increased in BCa-initiating cells, highly metastatic sub-clones and in BCa patients at risk of metastasis. These data suggest that, in a subset of BCas, PML sustains the function of BCa-initiating cells and in turn supports the metastatic dissemination capacity6. We show that PML-directed therapies are efficient in BCa cells with elevated expression of the protein. Such phenomenon has been defined as ‘addiction’40,50,51, and it represents an exciting avenue in the establishment of novel therapeutic initiatives. Importantly, targeted therapies have been particularly successful when combined with a predictive biomarker. The availability of a clinically validated protocol to detect PML immunoreactivity52 offers a unique opportunity to define the patients that would benefit from therapies based on PML inhibition. In addition, our proof-of-concept demonstration of the therapeutic efficacy of PML pharmacological inhibition with ATO indicates that (1) repositioning of ATO (that is currently used in the treatment of acute PML) for BCa therapy is a viable approach, (2) there is strong support for the development of novel and more effective PML inhibitors and (3) the identification of combined therapies with PML inhibitors in BCa is a novel and exciting area of investigation. Mechanistically, our data demonstrate that PML is in close proximity to the promoter region of SOX9, and positively regulates the expression of the gene. SOX9 has been recently established as a central regulator of normal and cancer stem cells9,10,11,12,13,53,54,55,56,57,58,59,60,61. This activity is executed in part through the functional interplay with epithelial-to-mesenchymal transition regulators such as SLUG12,53,56. In BCa, SOX9 is found overexpressed in the TNBC subtype, and regulates the WNT/beta-catenin pathway14. In addition, this transcription factor is a main driver of the transcriptional signature of this subtype of BCa62. All these features make SOX9 an ideal target for BCa therapy. However, development of small molecules targeting transcription factors has been an outstanding challenge with limited success63. Our data demonstrating that PML sustains SOX9 expression in aggressive BCa opens the possibility to bypass this limitation and inhibit the function of the transcription factor through upstream PML targeting. In summary, our data provide proof-of-concept demonstration of the fact that PML-inhibiting compounds could exhibit strong potential for BCa therapy upon PML-based stratification. ## Methods ### Cell culture MDA-MB-231, BT594, HBL100, MCF7 and T47D cell lines were obtained from the American Type Culture Collection (Manassas, VA, USA) or from Leibniz-Institut—Deutsche Sammlung von Mikroorganismen und Zellkulturen GmbH (DMSZ, Germany), who provided an authentication certificate. None of the cell lines used in this study was found in the database of commonly misidentified cell lines maintained by ICLAC and NCBI biosample. PDX44-derived cell line was generated by Dr Ibrahim and Dr Serra starting from xenograft tumours. Cell lines were routinely monitored for mycoplasma contamination and quarantined, while treated if positive. All cell lines were maintained in DMEM media supplemented with 10% (v/v) foetal bovine serum and 1% (v/v) penicillin–streptomycin. OS formation assays were carried out as previously described64. In brief, single-cell suspensions were plated in six-well tissue culture plates covered with poly-2-hydroxyethyl-methacrylate (Sigma, St Louis, MO) to prevent cell attachment, at a density of 3,000 cells per ml in serum-free DMEM supplemented with 1% penicillin/streptomycin, 1% B27 (Invitrogen, Carlsbad, CA, USA), 10 ng ml−1 epidermal growth factor (EGF) (Sigma, St. Louis, MO) and 2 ng ml−1 fibroblast Growth Factor, basic (FGFb) (Invitrogen, Carlsbad, CA, USA). After 6 days in culture, OS were counted using a light microscope. For secondary OS formation, following the same protocol, 100,000 cells were plated in 100 mm dishes and collected by gentle centrifugation (200g) and dissociated enzymatically (5 min in 1:1 TrypLE solution at 37 °C, Life Technologies, cat: 12604013) and single cells were re-plated at a density of 3,000 cells per ml in six-well tissue culture plates for 6 days. ### Generation of stable cell lines 293FT cells were used for lentiviral production. Lentiviral vectors expressing shRNAs against human PML, STAT3 and SOX9 from the Mission shRNA Library were purchased from Sigma‐Aldrich or Addgene. Cells were transfected with lentiviral vectors following standard procedures, and viral supernatant was used to infect cells. Selection was done using puromycin (2 μg ml−1) for 48 h. As a control, a lentivirus with scrambled shRNA (shC) was used. Short hairpins sequence: sh1PML (TRCN0000003865): CCGGCAATACAACGACAGCCCAGAACTCGAGTTCTGGGCTGTCGTTGTATTGTTTTT, sh2PML (TRCN0000003865): CCGGCAATACAACGACAGCCCAGAACTCGAGTTCTGGGCTGTCGTTGTATTGTTTTT; sh4PML (TRCN 0000003867): CCGGGCCAGTGTACGCCTTCTCCATCTCGAGATGGAGAAGGCGTACACTGGCTTTTT; sh5PML (TRCN 0000003867): CCGGGTGTACCGGCAGATTGTGGATCTCGAGATCCACAATCTGCCGGTACACTTTTT; shC: CCGGCAACAAGATGAAGAGCACCAACTCGAGTTGGTGCTCTTCATCTTGTTG. sh41STAT3 (TRCN0000020841): CCGGGCTGAAATCATCATGGGCTATCTCGAGATAGCCCATGATGATTTCAGCTTTTT; sh43STAT3 (TRCN0000020843): CCGGGCAAAGAATCACATGCCACTTCTCGAGAAGTGGCATGTGATTCTTTGCTTTTT. sh1SOX9 (Addgene, GenBank ID: RHS3979-9587792; GCATCCTTCAATTTCTGTATA); sh2SOX9 (TRCN0000342824): CCGGCTCCACCTTCACCTACATGAACTCGAGTTCATGTAGGTGAAGGTGGAGTTTTTG. Sub-cloning of shC and sh4PML into pLKO-Tet-On vector was done introducing AgeI and EcoRI in the 5′-end of top and bottom shRNA oligos, respectively (following the strategy provided by Dr Dmitri Wiederschain65, Addgene plasmid: 21915). HA-PMLIV was sub-cloned into a TRIPZ vector using Age1–Mlu1 sites. ### Immunoassays Western blot analysis was carried out as previously described20. Uncropped scans are provided as part of the Supplementary Information (Supplementary Fig. 7). In brief, cells were seeded on six-well plates and 4 days (unless otherwise specified) after seeding cell lysates were prepared with RIPA buffer (50 mM TrisHCl pH 7.5, 150 mM NaCl, 1 mM EDTA, 0.1% SDS, 1% Nonidet P40, 1% sodium deoxycholate, 1 mM sodium fluoride, 1 mM sodium orthovanadate, 1 mM beta-glycerophosphate and protease inhibitor cocktail; Roche). The following antibodies were used for western blotting: rabbit polyclonal anti-PML, 1:1,000 dilution (cat: A301-167A; Bethyl laboratories), rabbit polyclonal anti-phospho-STAT3 (Tyr705) 1:1,000 dilution and total STAT3 1:1,000 dilution (cat: 9145, 9132 respectively; Cell Signaling), rabbit polyclonal anti-SOX9 1:2,000 dilution (cat: AB5535; CHEMICON International), HA-Tag polyclonal antibody 1:2,000 dilution (cat: C29F4, Cell Signaling Technology, Inc) and mouse monoclonal anti-beta-ACTIN 1:2,000 dilution (clone: AC-74, catalogue: A5316, Sigma-Aldrich). After standard SDS–polyacrylamide gel electrophoresis and western blotting techniques, proteins were visualized using the enhanced chemiluminescence (ECL) system. For immunofluorescence, cells were seeded on glass cover slips in 24-well plates and 4 days after seeding, cells were fixed with 4% paraformaldehyde (15 min), PBS (three times wash), 1% Triton X-100 (5 min), PBS (three times wash), 10% goat serum (1 h) and anti-PML antibody 1:100 ditution (catalogue A301-167A; Bethyl laboratories) was added overnight (4°C) in goat serum. Cover slips were washed with PBS three times and incubated with secondary antibody (anti-rabbit Alexa488; Invitrogen-Molecular Probes) for 1 h (room temperature). Cover slips were washed with PBS three times, and 4,6-diamidino-2-phenylindole added to stain nuclei (10 min), followed by mounting with Mowiol. Immunofluorescence images were obtained with an AxioImager D1 microscope. For immunohistochemistry, tissues were fixed in 10% neutral-buffered formalin and embedded in paraffin according to standard procedures. Three to four μm-thick sections were stained for PML (clone PG-M3 Santa Cruz Biotechnology Inc, sc-966, 1:200 dilution), and vimentin (1:1,000, NCL-L-VIM-V9, Novocastra). Antigen retrieval was performed with citrate buffer (pH 6). Detection was performed with the ABC Kit from Vector Laboratories and 3,3′-diaminobenzidine (DAB)-based development. Sections were counterstained with haematoxylin. The PML general immunoreactivity scoring system (used in Fig. 1h) is described in ref. 20. For the Marseille data set, PML (1:200), SOX9 (1:400, Millipore) and phospho-STAT3 (Tyr705) (1:100, Cell Signalling: M9C6) immunostaining was performed as reported11,20. The percentage of PML-high (Ph) and -low (Pl) immunoreactive tumour cells in the Marseille data set was quantified separately and the h-score was calculated, attributing a relative value of 1 × to Pl and 2 × to Ph intensity nuclear signal (h=(1 × Pl)+(2 × Ph)). For SOX9 automated quantification and construction of tissue microarrays (TMAs) in the Marseille data set was carried out as reported23. In brief, cores were punched from the selected paraffin blocks, and distributed in new blocks including two cores of 0.6 mm diameter for each tumour. All the TMA blocks were stored at 4 °C. TMA serial tissue sections were prepared 24 h before immunohistochemistry processing and stored at 4 °C. The immunoperoxidase procedures were performed using an automated Ventana Benchmark XT auto-stainer. This device allowed identical well-controlled procedures for antigen retrieval and Ventana kits. ### Quantitative real-time PCR Cells were seeded as for western blot. Total RNA was extracted from cells using NucleoSpin RNA isolation kit from Macherey-Nagel (ref: 740955.240C). Complementary DNA was produced from 1 μg of RNA using qScript cDNA SuperMixt (Quanta Bioscience, ref: 95048). Taqman probes were obtained from Applied Biosystems. Amplifications were run in a Viia7 Real-Time PCR System (Applied Biosystems) using the following probes: PML (Hs00971694_m1, cat: 4331182) and SOX9 (Hs01001343_g1, cat: 4331182). For STAT3, SOX2 and LGR5 amplification, Universal Probe Library (Roche) primers and probes were employed (STAT3, For: cccttggattgagagtcaaga, Rev: aagcggctatactgctggtc; probe: 14; SOX2, For: gggggaatggaccttgtatag, Rev: gcaaagctcctaccgtacca; probe: 65; LGR5, For: accagactatgcctttggaaac, Rev: ttcccagggagtggattctat; probe: 78). β-actin (Hs99999903_m1, cat: 4331182) and GAPDH (Hs02758991_g1, cat: 4331182) housekeeping assays from Applied Biosystems showed similar results (all quantitative PCR with reverse transcription data presented were normalized using GAPDH). ### ALDH1 activity by FACS To measure the ALDH1 activity present in the cells, the ALDEFLUOR assay was carried out according to manufacturer’s (Stemcell Technologies) guidelines. In brief, dissociated MDA-MB-231 cells were resuspended in ALDEFLUOR assay buffer at a final concentration of 1.106 ml−1. ALDH substrate, bodipyaminoacetaldehyde was added to the cells at a final concentration of 1.5 mM. Immediately, half of the cells were transferred to an Eppendorf tube containing a two fold molar excess of the ALDH inhibitor, diethylaminobenzaldehyde. Both tubes were incubated for 45 min at 37 °C, and after this incubation cells were centrifuged at 250g for 5 min at 4 °C and resuspended in ice-cold ALDEFLUOR assay buffer. Cells were analysed using a FACSAria1 (Becton Dickinson) flow cytometer. DRAQ7 (BiostatuS) was added prior analysis to each tube for dead cell exclusion. FACSAria1 was also used for sorting cells. Data were analysed using the FACSDiva software. ### Reagents For in vitro experiments, SI3-201 (Sigma-Aldrich, SML0330) was prepared at 10 mg ml−1 in dimethylsulfoxide and used at the indicated concentrations. TG101348 (Santa Cruz, sc-364740) was prepared 100 mg ml−1 in dimethylsulfoxide and used at the indicated concentrations. ATO (Sigma-Aldrich) was prepared at a concentration of 100 mM in NaOH 1 N and subsequently diluted to 0.1 mM in PBS for a 1,000 × working solution. ATO was used at 150 nM either 3 or 6 days as indicated in figure legends. For in vivo experiments a dose of 5 mg kg−1 per day was intraperitoneally administered. ### Mice Xenograft experiments were carried out following the ethical guidelines established by the Biosafety and Welfare Committee at CIC bioGUNE and Biodonostia Institute. The procedures employed were carried out following the recommendations from AAALAC. Xenograft experiments were performed as previously described66, injecting either 5.105 or 5.104 cells per condition (unless otherwise specified), four injections per mouse. Metastasis experiment was approved by the institutional animal care and use committee of IRB-Barcelona. For tail vein injections, cells were resuspended in PBS and injected into tail vein of mice using a 26 G needle (1.2 × 105 cells per mouse), as previously described25. Cell lung colonization capacity was scored 21 days post inoculation by human vimentin. PML expression was scored as undetectable (PML 0) and detectable (PML 1+, 2+ and 3+). All mice (female Hsd:Athymic Nude-Foxn1 nu/nu) were inoculated at 8–12 weeks of age. ### ChIP ChIP was performed using the SimpleChIP Enzymatic Chromatin IP Kit (cat: 9003, Cell Signaling Technology, Inc). MDA-MB-231 cells were grown in 150 mm dishes either with or without 50 ng ml−1 doxycycline during 3 days. Cells from three 150 mm dishes (2.5 × 107 cells) were cross-linked with 35% formaldehyde for 10 min at room temperature. Glycine was added to dishes, and cells incubated for 5 min at room temperature. Cells were then washed twice with ice-cold PBS, and scraped into PBS+PMSF. Pelleted cells were lysed and nuclei were collected following manufacturer’s instructions. Nuclear lysates were digested with micrococcal nuclease for 20 min at 37 °C and then sonicated in 500 μl aliquots on ice for three pulses of 15 s using a Branson sonicator. Cells were held on ice for at least 1 min between sonications. Lysates were clarified at 11,000g for 10 min at 4 °C, and chromatin was stored at −80 °C. HA-Tag polyclonal antibody (cat: C29F4, Cell Signaling Technology), rabbit polyclonal anti-PML (cat: A301-167A; Bethyl laboratories) and IgG antibody (cat: 2729, Cell Signaling Technology, Inc), were incubated overnight (4 °C) with rotation and protein G magnetic beads were incubated 2 h (4 °C). Washes and elution of chromatin were performed following manufacturer’s instructions. DNA quantification was carried out using a Viia7 Real-Time PCR System (Applied Biosystems) with SybrGreen reagents and primers that amplify the predicted PML binding region to SOX9 promoter (chr17:70117013-70117409) as follows: left primer: ccggaaacttttctttgcag and right primer: cggcgagcacttaggaag. ### Patient data sets, bioinformatics and statistical analysis All studies involving human subjects were approved by the corresponding committees with informed consent as stated in the original publications3,23,25. The use of MSK/EMC and Marseille cohorts were previously described23,25. For MSK/EMC, MFS curves were plotted using Kaplan–Meier estimates and compared using the Gehan–Breslow–Wilcoxon test. Two groups were compared using mean PML expression values as the cutoff between PML high and low. Kaplan–Meier survival and correlation analysis in patient samples: publicly available and clinically annotated BCa cohorts with gene expression profiles (GSE2603, GSE2034, GSE5327 and GSE12276) were pooled as described above. To remove systematic biases, before merging the expression measurements were converted to z scores for all genes. For intrinsic subtype classification, we carried out the following analysis: for luminal genes, ESR1 and PGR1 presented a bimodal distribution. We used package mclust to fit a mixture of normal distributions with two components and obtain the posterior probability that each patient belongs to the luminal low and luminal high components. A patient was considered luminal low if the posterior probability of belonging to this group was >80%. The same criterion was used for luminal high. When a patient was neither luminal high nor luminal low, it was considered luminal intermediate. Proliferation status (Prol) and ERBB2 expression did not present a bimodal distribution. Therefore, half of the patients with lowest mean values were considered proliferation low. The rest were considered proliferation high. After defining high and low populations for each parameter, the subtypes were constructed as follows: luminal A: Prol low, ESR1 intermediate or high, luminal intermediate or high; luminal B: Prol high, ESR1 intermediate or high, luminal intermediate or high; HER2 enriched: Prol high, ESR1 intermediate or low, luminal intermediate or low; ERBB2 high; basal like: Prol high, ESR1 low, luminal low; ERBB2 low, PGR1 low. Sixty-four patients could not be assigned to any subtype according to PAM50’s classification. A Cox proportional hazards model was fitted to compute HR. Likelihood ratio tests were performed to compute P values. The HR was checked for constancy over time, fulfilling Cox model assumptions. For Curtis data set patients, RNA was extracted from 1,980 tumours as described3. RNA hybridizations were performed using Illumina HT-12 v3 platform and analysed using the bioconductor bead array package67. The BASH algorithm68 was applied to correct for spatial artefacts in the arrays. Bead-level data were summarized and re-annotated as described in ref. 3. Log-intensity values for PML expression were scaled to z scores. Probe selection was performed on the basis of probe quality, 3′-position, no other genomic matches and no single-nucleotide polymorphisms in the region. On the basis of these criteria, PML probe ILMN_1731299 was selected for analysis. Survival analysis was done using as endpoints MFS at 5 years (distant metastasis as event). Two groups were compared using mean PML expression values as the cutoff between PML high and low. We used the log-rank test as implemented in the survival R package69. For therapy response analysis, publicly available data sets (GSE22093 and GSE23988) were downloaded from Gene Expression Omnibus (GEO), and subjected to background correction, log2 transformation and quartile normalization. For correlation analysis with STAT3 signatures, gene sets were extracted (ref. 37, and http://software.broadinstitute.org/gsea/msigdb/cards/V$STAT3_01) and average signal value in the MSK/EMC data set was calculated. These values were used to perform the correlation analysis with PML signal values (Pearson correlation).

No statistics were applied to determine sample size. The experiments were not randomized. The investigators were not blinded to allocation during experiments and outcome assessment. Data analysed by parametric tests are represented by the mean±s.e.m. of pooled experiments unless otherwise stated. n values represent the number of independent experiments performed or the number of individual mice or patient specimens. For each independent in vitro experiment, at least three technical replicates were used and a minimum number of three experiments were performed to ensure adequate statistical power. Analysis of variance test was used for multi-component comparisons and Student's t-test for two-component comparisons. In the in vitro experiments, normal distribution was confirmed or assumed (for n<5) and Student's t-test was applied for two-component comparisons. Two-tailed statistical analysis was applied for experimental design without predicted result, and one tail for validation or hypothesis-driven experiments. The confidence level used for all the statistical analyses was of 0.95 (alpha value=0.05). Tumour-initiating cell frequency was estimated using ELDA software as previously described70.

### Data availability

Data from public repositories analysed throughout this manuscript (see the ‘Patient data sets, bioinformatics and statistical analysis’ section) is available as indicated in the referenced publications.