MLL3 Loss Drives Metastasis by Promoting a Hybrid Epithelial-Mesenchymal Transition State

MLL3 loss enhances breast tumor metastasis

MLL3 is frequently mutated in a wide range of human cancer (Extended Data Fig. 1a) ^{20–22, 30}. In breast cancer, MLL3 is altered by both point mutations and gene deletions in major subtypes (Extended Data Fig. 1b). Breast cancers with MLL3 alterations have a significantly poorer prognosis than MLL3 wild-type (WT) breast cancers (Fig. 1a). Furthermore, MLL3 expression was significantly downregulated in breast tumors compared to normal mammary tissues (Extended Data Fig. 1c) ^{22, 31}. Interestingly, MLL3 expression was further reduced in distant metastases compared to primary tumors, suggesting a potential role of MLL3 loss in metastasis (Fig. 1b–c, Extended Data Fig. 1d) ³².

To test this, we deleted MLL3 with lentiviral CRISPR in luciferase2-tdTomato (Luc2-Tomato)-labeled MDA-MB-231 cells, which expresses WT MLL3 (Extended Data Fig. 2a). These MLL3-mutant MDA-MB-231 cells developed significantly larger primary tumors than MLL3-WT cells (transduced with a non-targeting CRISPR vector) after orthotopic injection (Fig. 1D–E and Extended Data Fig. 2b). Furthermore, the MLL3-mutant tumors generated much higher metastatic burdens in the lung (11-fold) and the liver (10-fold) than MLL3-WT tumors, even when normalized by primary tumor burden (Fig. 1f–g and Extended Data Fig. 2c–d). The MLL3-WT lung metastases expressed significantly lower levels of MLL3 than the primary tumors, consistent with the result that loss of MLL3 enhanced metastasis (Extended Data Fig. 2e–f). MLL3 deletion in MDA-MB-435s cells also increased spontaneous metastasis to the lung (6-fold) and liver (3-fold), even though it did not affect primary tumor growth (Fig. 1h–k and Extended Data Fig. 2g–j). A marked increase of spontaneous metastasis by MLL3 deletion was observed in another TNBC cell line SUM159 (Fig. 1l–o and Extended Data Fig. 2k–l). Together, these results showed that MLL3 loss promotes spontaneous metastasis in multiple cancer cell models.

Loss of MLL3 promotes metastatic colonization

Circulating tumor cell (CTC) frequency reflects the capacity of tumor cells to invade, intravasate, and survive in circulation. The numbers of CTCs were comparable between mice bearing MLL3-WT and -mutant MDA-MB-231 tumors, suggesting that MLL3 deletion did not affect these early steps of metastasis by MDA-MB-231 cells (Fig. 2a). By examining the frequency of Tomato⁺ cancer cells extravasating out of FITC lectin-labeled blood vessels after tail vein injection, we found that MLL3 deletion also did not affect the rate of extravasation (Fig. 2b–c).

We then measured the ability of cancer cells to colonize the lung. One week after tail vein injection, the number of cancer cells and clone sizes were similar between the MLL3-WT and -mutant groups, suggesting MLL3-deletion did not affect the initial seeding/survival of extravasated cells (Fig. 2d and Extended Data Fig. 3a). However, MLL3-mutant metastatic nodules grew significantly larger and generated a greater metastatic burden than the WT control at later time points (Fig. 2d–e and Extended Data Fig. 3b). Consistently, we found that MLL3-mutant metastases have higher frequencies of Ki67⁺ cells than WT metastases (Fig. 2f and Extended Data Fig. 3c), while no difference was observed for cleaved caspase 3⁺ cells (Extended Data Fig. 3d). MLL3 loss also greatly increased metastatic colonization of SUM159 cells (Fig. 2g and Extended Data Fig. 3e).

To assess lung colonization in immunocompetent hosts, we generated mouse mammary stem cell (MaSC) organoids carrying Trp53 CRISPR deletion and Pik3ca^H1047R mutation and then further deleted Mll3 with lentiviral CRISPR to create Mll3-mutant or -WT Trp53^null/Pik3ca^H1047R MaSCs. After tail vein injection in syngeneic FVB/n mice, two independent lines of Mll3-mutant MaSCs generated a dramatically higher metastatic burden than Mll3-WT cells (Fig. 2h–j and Extended Data Fig. 3f). Together, these results suggested that MLL3 loss increased metastatic colonization.

MLL3 loss facilitates hybrid EMT in mesenchymal cells

While MDA-MB-231 primary tumor cells had a vimentin (VIM)⁺/E-cadherin (Ecad)⁻ mesenchymal phenotype, a fraction of cancer cells in spontaneous lung metastases showed a hybrid EMT phenotype co-expressing VIM and Ecad (Fig. 3a). Similar VIM⁺/Ecad⁺ cells were observed in MDA-MB-231 experimental metastases, indicating that MDA-MB-231 cells undergo partial MET to acquire hybrid EMT features at distant metastatic sites (Fig. 3a). We then used the CD44 and CD104 cell-surface markers to measure the acquisition of hybrid EMT, which can be identified as CD44⁺CD104^high, whereas fully mesenchymal cells are CD44⁺CD104^{low 15, 16}. The MLL3-mutant MDA-MB-231 lung metastases contained ~2-fold more CD44⁺CD104^high cells than the control metastases two weeks after tail vein injection (Fig. 3b–c). Mll3 deletion also increased the frequency of VIM⁺/Ecad⁺ hybrid EMT cells in Trp53^null/Pik3ca^H1047R MaSC organoids and lung metastases (Fig. 3d–e).

The MET occurring in distant metastases is likely driven by stromal signals in the metastatic niche, although such signals are yet to be clearly defined in the lung ³³. One signaling pathway that promotes MET is protein kinase A (PKA) ³⁴. Therefore, we tested whether MLL3 loss can facilitate PKA-induced MET. PKA activator forskolin treatment generated significantly more (3-fold) CD44⁺CD104^high hybrid EMT cells in MLL3-mutant cells than the WT control (Fig. 3f). Interestingly, forskolin upregulated the epithelial marker E-cadherin only in MLL3-mutant cells (Fig. 3g). Together, these results showed that loss of MLL3 facilitated the acquisition of the hybrid EMT state.

We next assessed whether the hybrid EMT cells increased by MLL3 deletion have higher metastatic capacity. CD44⁺/CD104^high hybrid EMT cells and CD44⁺/CD104^low mesenchymal cells were sorted from forskolin-treated WT or MLL3-mutant MDA-MB-231 cells and assessed by experimental metastasis. The MLL3-mutant hybrid EMT cells showed markedly higher lung metastasis-initiating ability than the other cell types (Fig. 4a–c). The MLL3-mutant hybrid cells also have significantly increased metastasis-initiating ability to multiple other organs, including the bone, liver, and kidney (Fig. 4d–e and Extended Data Fig. 4a–d). Furthermore, MLL3-mutant hybrid cells showed significantly higher tumor-initiating ability and faster tumor growth than the other cell types (Fig. 4f, Extended Data Fig. 4e and Supplementary Table 1). These results suggest that the induction of hybrid EMT by MLL3 loss increases multi-organ metastatic and tumor-initiating abilities.

MLL3 loss facilitates EMT and metastasis in luminal cancer

To determine whether MLL3 loss promotes epithelial-mesenchymal plasticity in epithelial cells, we deleted MLL3 in the luminal breast cancer cell line MCF7 by lentiviral CRISPR. While MLL3 deletion did not lead to a noticeable EMT phenotype in adherent culture, it significantly increased the sphere-forming ability, a property associated with cells undergoing an EMT (Fig. 5a) ³⁵. MLL3-mutant MCF7 spheres contained substantially higher frequencies of cells lacking the epithelial marker CD24 (Fig. 5b). These CD24⁻ cells had a higher sphere-forming ability than CD24⁺ cells (Extended Data Fig. 5a). Furthermore, the MLL3-mutant spheres markedly upregulated vimentin and EMT transcription factors SLUG and SNAIL while maintaining the expression of the epithelial markers E-cadherin and CD104, suggesting these cells underwent a partial EMT (Fig. 5c and Extended Data Fig. 5b). MLL3 deletion in another luminal breast cancer cell line T47D similarly increased sphere formation and the frequency of CD24⁻ and VIM⁺ cells (Fig. 5d–e and Extended Data Fig. 5c).

We further tested whether MLL3 loss potentiates the EMT in response to TGF-β in HMLE immortalized human mammary epithelial cells, a well-characterized model for TGF-β-induced EMT ³⁵. MLL3-mutant HMLE cells underwent EMT more efficiently than MLL3-WT cells responding to low concentrations of TGF-β1 and generated mesenchymal cells at faster kinetics, as measured by the induction of CD44^highCD24^low mesenchymal cells (Fig. 5f–g and Extended data Fig. 5d).

In vivo, MLL3 deletion significantly accelerated MCF7 tumor onset and growth (Fig. 5h–i). Furthermore, the MLL3-mutant tumors exhibited decreased E-cadherin expression and disorganized adherent junctions, consistent with the partial EMT phenotype in MLL3-mutant MCF7 spheres (Fig. 5j). Furthermore, MLL3-mutant MCF7 cells generated significantly more metastases than the WT cells after tail vein injection (Fig. 5k–l). These results showed that MLL3 loss enabled luminal cancer cells to acquire the hybrid EMT property and become more metastatic.

MLL3 loss activates the interferon-γ pathway

To understand the mechanisms by which MLL3 deletion promotes hybrid EMT, we compared transcriptomes of MLL3-WT and -mutant MCF7 cells. Several inflammatory signaling pathways, including interferon-γ (IFNγ), interferon-α (IFNα), and tumor necrosis factor-α (TNFα) signaling, were among the top gene signatures enriched by MLL3 deletion (Fig. 6a and Extended Data Fig. 6a–b). MLL3 deletion also increased the IFNγ gene signature in MDA-MB-231 cells (Fig. 6b). Furthermore, the hybrid EMT cells sorted from forskolin-treated MDA-MB-231 cells expressed a significantly higher level of IFNγ gene signature than the mesenchymal cells (Fig. 6c and Extended Data Fig. 6c). Ingenuity Pathway Analysis (IPA) identified IFNγ as the upstream cytokine regulator in the hybrid EMT cells (Fig. 6d). Several interferon regulatory factors (IRF1, IRF2, IRF3, IRF5, and IRF7) and STAT1 were also identified as active transcription factor regulators in the hybrid EMT cells (Fig. 6e). These results suggest that the hybrid EMT state is associated with higher IFNγ pathway activity, which is upregulated by MLL3 loss.

IFNγ signaling is required for hybrid EMT and metastasis

We tested the functional role of inflammatory cytokines and found that IFNγ treatment, but not IFNα or TNFα, robustly increased the frequency of CD24⁻ cells in MCF7 adherent culture, suggesting a unique role of IFNγ in hybrid EMT induction (Extended Data Fig. 6d). Furthermore, MLL3-mutant MCF7 cells downregulated the epithelial marker CD24 more efficiently than WT cells upon IFNγ treatment, supporting that MLL3 loss sensitizes the response to IFNγ (Extended Data Fig. 6e). Notably, IFNγ treatment in control MCF7 spheres increased the frequency of CD24⁻ cells to that of MLL3-mutant spheres (Fig. 6f). Conversely, inhibition of IFNγ signaling using an IFNGR1 blocking antibody decreased the frequency of CD24⁻ cells in MLL3-mutant spheres (Fig. 6g). Lentiviral CRISPR deletion of IFNGR1 similarly reduced the frequency of CD24⁻ cells and sphere-forming efficiency in MLL3-mutant cells (Fig. 6h–i and Extended Data Fig. 6f). IFNGR1 deletion also abolished the effect of MLL3 loss on accelerating MCF7 tumor onset (Fig. 6j). Additionally, IFNGR1 deletion diminished EMT phenotypes in MLL3-mutant MCF7 tumors, as measured by EMT markers and tumor histology (Fig. 6k and Extended Data Fig. 6g–i).

We then assessed the function of IFNγ signaling in hybrid EMT induction and metastasis in mesenchymal cells. Treating MDA-MB-231 cells with IFNγ increased the frequency of CD44⁺CD104^high hybrid EMT cells (Fig. 6l). Conversely, IFNGR1 CRISPR deletion in MLL3-mutant MDA-MB-231 cells reduced forskolin-induced hybrid EMT cells (Extended Data Fig. 7a). When MLL3-WT MDA-MB-231 cells were pretreated with IFNγ (1 or 10 ng/ml) for four days in vitro, they acquired markedly increased metastatic ability (Fig. 6m). Conversely, IFNGR1 deletion in MLL3-mutant MDA-MB-231 cells reduced lung colonization to the level of MLL3-WT cells (Fig. 6n). Furthermore, the percentages of hybrid EMT cells among disseminated tumor cells in the lung were significantly reduced by IFNGR1 deletion (Fig. 6o–q and Extended Data Fig. 7b–d). IFNGR1 deletion also dramatically decreased primary tumor growth and spontaneous metastasis by MLL3-mutant MDA-MB-231 cells (Fig. 6r–s and Extended Data Fig 7e–f). Together, these results indicate that IFNγ signaling is required for the induction of hybrid EMT and efficient metastatic colonization, acting downstream of MLL3 loss.

IFγ treatment significantly increased the expression of EMT transcription factors (TFs) SLUG, SNAIL, and ZEB1 in MCF7 and MDA-MB-231 cells, recapitulating the effect of MLL3 deletion (Extended Data Fig. 7g–h). Conversely, IFNGR1 deletion in MLL3-mutant cells reduced the levels of these EMT TFs in both cell lines. Concordant with the role of IFNγ in EMT TF upregulation, we observed a significant correlation between the IFNγ response gene signature and the EMT TF signature in the TCGA BRCA cancer samples (Extended Data Fig. 7i). High IFNGR1 levels are associated with poorer survival in TNBC, although with moderately improved survival in ER⁺ cancer³⁶. While the underlying causes of these distinct associations remain unknown, this data supports the role of IFNγ in promoting aggressive tumor phenotypes.

MLL3 loss increases H3K27ac in IFNγ response gene enhancers

To understand how MLL3 loss may enhance IFNγ signaling at the chromatin level, we performed ATAC-seq and ChIP-seq for enhancer and promoter histone marks comparing MLL3-WT and MLL3-mutant MDA-MB-231 cells. HOMER analysis³⁷ of ATAC-seq peaks showed that the motifs of IFNγ downstream TFs were enriched in the more accessible chromatin regions in MLL3-mutant cells, while the global chromatin landscape did not change significantly (Fig. 7a and Extended Data Fig. 8a). ChIP-seq for H3K4me1, H3K27ac and H3K4me3 showed that MLL3 loss did not alter the abundance and global distribution of these histone marks at the genome-wide scale (Extended Data Fig 8a–b), agreeing with previous studies in MCF7 cells ^{38, 39}.

Next, we focused on the effect of MLL3 loss on enhancers of IFNγ pathway genes. We analyzed the H3K27ac peaks by ROSE ^{40, 41} and found that MLL3 loss increased H3K27ac intensity in the enhancers of a subset of IFNγ response genes in both MCF7 and MDA-MB-231 cells (Fig. 7b–d). In contrast, H3K4me1 levels at these H3K27ac-up regions were not changed in MDA-MB-231 cells while slightly increased in MCF7 cells upon MLL3 loss (Extended Data Fig. 9a–b). Of note, MLL3 knockdown in MCF7 cells increased H3K27ac levels and established a super-enhancer at the IRF1 gene, a key mediator of IFNγ signaling (Fig. 7e–f). In MDA-MB-231 cells, MLL3 deletion markedly increased H3K27ac levels at the IRF1 enhancer upon IFNγ stimulation, although it did not affect the H3K27ac levels in the unstimulated state (Extended Data Fig. 9c–e). Interestingly, MLL3 deletion significantly increased the overall enrichment of H3K27ac in enhancers in both MDA-MB-231 and MCF7 cells (Extended Data Fig. 9f).

We next examined whether the MLL3 complex binds to the IFNγ response gene loci. We performed ChIP-seq using an antibody recognizing both MLL3 and MLL4 and another antibody against UTX, a subunit shared by MLL3 and MLL4 complexes. Both MLL3/4 and UTX bind to enhancers of IFNγ response genes, including IRF1 and STAT1. Furthermore, the intensity of MLL3/4 and UTX peaks was significantly reduced after MLL3 deletion, supporting direct occupancy of the MLL3 complex in these regions (Extended Data Fig 9a–b, g).

STAT1 phosphorylation and the subsequent auto-upregulation are critical steps in the IFNγ signaling pathway ^{42, 43}. We found that H3K27ac levels at the STAT1 enhancer/promoter regions increased significantly in MLL3-mutant MCF7 cells (Fig. 7e). Furthermore, phospho-STAT1 and total STAT1 expression were higher in these cells upon IFNγ treatment in both adherent and sphere cultures (Fig. 7g and Extended Data Fig. 10a–b). STAT1 occupancy at its own promoter was significantly higher in MLL3-mutant MCF7 cells than WT cells upon IFNγ treatment (Fig. 7h). MLL3 deletion in MDA-MB-231 cells also led to a more robust upregulation of phospho-STAT1 and total STAT1 by IFNγ (Extended Data Fig. 10c–d). These results point to a consistent effect of MLL3 loss on activating enhancers of a sub-cluster of IFNγ response genes, which leads to the upregulation of key mediators of IFNγ signaling, such as STAT1.

MLL3-deficient tumor cells are sensitive to BET inhibition

To identify molecular vulnerabilities for MLL3-mutant cells, we performed a chemical screen comparing the response of syngeneic WT and Mll3-mutant MaSCs to a collection of epigenetic compounds. We found that the BET (bromodomain and extra-terminal) protein inhibitor (+)-JQ1 inhibited Mll3-mutant cells much more efficiently than WT cells (Fig. 8a). Furthermore, Mll3-mutant organoid cells treated with JQ1 had a greatly diminished ability to re-seed secondary organoids in the absence of further JQ1 treatment, suggesting inhibition of the self-renewal ability (Fig. 8b). Using primary cultures of Mll3-WT and Mll3-mutant tumors induced by constitutively active Pik3ca (Pik3ca*) ²⁹, we showed that Mll3-mutant tumor cells are more sensitive (10-fold) to two distinct BET inhibitors, JQ1 and GSK525762 (Fig. 8c). Furthermore, while parental MCF7 cells are relatively resistant to BET inhibition ⁴⁴, MLL3 deletion increased their sensitivity to JQ1 by 15-fold. Similar effects were found for GSK525762 (Fig. 8d).

In vivo, JQ1 significantly inhibited the growth of Mll3-mutant Pik3ca* tumors but did not affect Mll3-WT tumors (Fig. 8e). Similar differences in JQ1 responses were observed between MLL3-WT and -mutant MDA-MB-231 tumors (Fig. 8f). Furthermore, JQ1 significantly inhibited the growth of MLL3-mutant MDA-MB-231 metastases in both the lung and bone but did not affect MLL3-WT metastases (Fig. 8g–h). The proliferation in MLL3-mutant metastatic lesions was significantly inhibited by JQ1, as shown by Ki67 immunostaining (Fig. 8i). TNBC cells, including MDA-MB-231, show increased sensitivity to BET inhibitors, albeit resistance frequently occurs ^45–49. Our results show that MLL3 loss sensitizes tumor cells to BET inhibition, suggesting a potential strategy for targeting MLL3-mutant cancers, including metastatic diseases.

Cell lines and adherent cell culture

MCF7, MDA-MB-231, SUM159, T47D and MDA-MB-435s cells were cultured in Dulbecco’s Modified Eagle Medium (DMEM, Corning) with 10% fetal bovine serum (FBS, HyClone) and 1% (v/v) antibiotic-antimycotic solution (HyClone). HMLE cells were maintained in 1:1 mix of MEGM (Lonza) and DMEM/F-12 (Corning) supplemented with 5 ng/ml epidermal growth factor (EGF, Sigma-Aldrich E9644), 2.5 μg/ml insulin (Sigma-Aldrich, 91077C), 0.25 μg/ml hydrocortisone (Sigma-Aldrich, H0888). For TGFβ1 treatment, HMLE cells were cultured in MEGM and DMEM/F12 media (1:1) supplemented with insulin, EGF, hydrocortisone, 5% FBS and different concentrations of TGFβ1. The medium was refreshed every 3 days.

All cell lines, except HMLE, were authenticated by short tandem repeat (STR) analysis, matching to the ATCC database. HMLE cells, which lack preexisting STR data, were authenticated by morphology and flow cytometric profile. Of cell lines used in this work, only MDA-MB-435s is in the list of commonly misidentified cell lines maintained by ICLAC. MDA-MB-435s was obtained from ATCC and authenticated by STR on November 21, 2017 with 100% match to ATCC database (Extended Data Fig. 2m). We used MDA-MB-435s at the beginning of this study as a cell line that can form spontaneous metastasis with reasonable efficiency. One report suggests that MDA-MB-435s is a breast cancer cell line that expresses melanoma genes due to lineage infidelity ⁵⁹. However, its exact origin is still controversial. Therefore, after showing the effect of MLL3 deletion in promoting metastasis in MDA-MB-435s, we used other bona fide breast cancer cell lines to confirm our findings and dissect the mechanisms.

Tumor sphere culture

Tumor sphere culture was performed as described in previous studies ^{60, 61} with slight modification. MCF7 were grown in DMEM:F12 supplemented with 2% (v/v) B27 (50x, Gibco, 17504044), 20 ng/ml EGF (Sigma-Aldrich E9644), 20 ng/ml basic fibroblast growth factor (bFGF, EMD Millipore, GF003.), 4 μg/ml heparin (Sigma, H4784), 10 μg/ml insulin (Sigma-Aldrich, 91077C), 0.5 μg/ml hydrocortisone (Sigma-Aldrich, H0888). 1,000 cells/well were plated in ultra-low attachment 96-well plates and cultured for 7 days. The spheres greater than 100 μm in diameter were counted.

Mammary Organoid Culture

Wild-type and Mll3-null MaSC organoids were generated and cultured as described in ²⁹. 500 cells/well were seeded in ultra-low attachment 96-well plates (Corning Life Sciences) with organoid medium, containing Advance DMEM/F12 (Life Technologies) supplemented with 1% L-glutamine, 5% heat-inactivated FBS (Sigma-Aldrich, F2442), 5% Matrigel (Corning Life Sciences, 354234), 10 ng/ml EGF (Sigma-Aldrich, E9644), 20 ng/ml bFGF (EMD Millipore, GF003), 4 μg/ml heparin (Sigma-Aldrich, H4784) and 5 μM Y-27632 (Cayman Chemical Company, 10005583). After 7-day culture, organoids greater than 100 μm in diameter were counted.

Lentiviral CRISPR-mediated gene knockout

The sgRNAs targeting MLL3 were designed using the CRISPR Design Tool from the Zhang Laboratory (http://crispr.mit.edu) and cloned into the lentiCRISPRv2 vector (Addgene #52961, a gift from Feng Zhang). The sgRNA targeting sequences are as follows: sgMLL3 #3-GGGGAAGCGCCATCTTTGCG, sgMLL3 #9-GTCGGAGGAGGACAAGAGCG, sgNT-GCGAGGTATTCGGCTCCGCG. Cells were transduced with the lentiviral vectors and selected by puromycin. For IFNGR1 targeting, the sgRNA sequence targeting IFNGR1 or sgNT was cloned into the lentiCRISPRv2-GFP vector (a gift from David Feldser, Addgene #82416). Cells were transduced and sorted based on GFP. The sgIFNGR1 sequence is GTATAAAATACTCACGCAGA.

Animals

NOD/SCID, NSG, C57BL/6 and FVB/n mice were imported from the Jackson Laboratory (Bar Harbor, ME) and bred in-house. Mice were housed in microisolator cages and a pathogen-free condition. Mice were on a 12-12 hour light-dark cycle, with water and food ad libitum. Temperature in the facility was maintained at 21 °C. 4-8 weeks female mice were used in the study. All animal experiments were performed according to protocols approved by the IACUC of Albert Einstein College of Medicine (Protocol 20170607 and 00001266). For primary tumors, the maximal tumor burden approved by our IACUC is 2 cm in diameter and all tumors in this study did not exceed the limit.

Metastasis assays

For spontaneous metastasis, female NOD/SID mice at 6-8 weeks of age were injected with 1 × 10⁶ cells suspended in 50 μl 75% PBS/25% Matrigel into the left inguinal mammary fat pad. Tumor growth was monitored weekly by caliper measurements (V = (4/3) × π × (L/2) × (L/2) × (D/2)) and metastasis burden was monitored longitudinally by bioluminescence imaging with an IVIS Spectrum system (PerkinElmer, Living image version: 4.3.1.0.15880). At the endpoint, organs were dissected post-mortem and imaged for bioluminescence ex vivo. For experimental metastasis, 3-5 x 10⁵ cells were injected via tail vein into 4-6 weeks old female NSG or FVB/n mice. Metastatic burdens were monitored by bioluminescence imaging, flow cytometry or tissue histology. For experimental metastasis of FACS sorted CD44⁺CD104^high and CD44⁺CD104^−/low MDA-MB-231 cells, 2 × 10⁴ sorted cells were injected per mouse.

Orthotopic tumor cell implantation

Female NSG or NOD/SID mice at 6-8 weeks of age were injected with 1 × 10⁶ tumor cells suspended in 50 μl of 75% PBS/25% Matrigel (Corning Life Sciences, 354234) into the inguinal mammary fat pad. For MCF7 experiments, mice received 200 nM 17β-estradiol (Sigma-Aldrich, E8875) in the drinking water as previously published ⁶². Tumor onset was monitored by weekly palpation. Tumor volume was measured by caliper and calculated as V = (4/3) × π × (L/2) × (L/2) × (D/2). For in vivo limiting dilution assay, 20,000, 4,000, and 400 of FACS sorted CD44⁺CD104^high and CD44⁺CD104^−/low sgNT or sgMLL3 MDA-MB-231 cells were injected into the third and fourth mammary fat pad of NOD/SCID mice. Tumor onset and progression were monitored by weekly palpation for 80 days.

Circulating tumor cell and extravasation analyses

100 μl blood samples from tumor-bearing mice were collected via retro-orbital venous plexus. After removal of red blood cells using lysis buffer, tdTomato+ tumor cells were quantified by flow cytometer. For extravasation, blood vessels were visualized via tail vein injection of FITC-labeled Ricinus communis agglutinin I (RCA I, Vector Laboratories, FL-1081) 15 min before animal sacrifice. Frozen tissue sections were processed for imaging tumor cells and blood vessels. Percentages of tumor cells outside of blood vessels were scored in at least ten random fields in each mouse.

Western blot

Cells were lysed with the RIPA buffer containing 25 mM Tris-HCL, pH 7.5, 150 mM NaCl, 1% sodium deoxycholate, 0.1% SDS, 1 mM EDTA and 1.0% NP-40 on ice. Samples were resolved by SDS-PAGE and transferred to PVDF membranes. The western blots were performed with antibodies against MLL3 (Polyclonal, Millipore Sigma, ABE1851, 1:1000), E-cadherin (clone 36, BD Biosciences, cat# 610182, 1:1000), Vimentin (clone D21H3, Cell Signaling Technology, cat# 5741S, 1:1000), Slug (clone C19G7, Cell Signaling Technology, cat#9585S, 1:1000), Snail (clone C15d3, Cell Signaling Technology, cat#3879S, 1:1000), Zeb1 (clone D80D3, Cell Signaling Technology, cat#3396S, 1:1000), Stat1 (clone D1K9Y, Cell Signaling Technology, cat#14994S, 1:1000), phosphor-Stat1(Tyr701) (clone 58D6, Cell Signaling Technology, cat#9167S,1:1000), β-actin (clone C4/actin, BD Biosciences, cat#612656, 1:10,000), H3K27ac (clone D5E4, Cell Signaling Technology, cat#8173S, 1:1000), H3K4me1 (clone D1A9, Cell Signaling Technology, cat#5326S, 1:1000), H3K4me3 (clone C42D8, Cell Signaling Technology, cat#9751S, 1:1000), or Histone H3 (clone 1B1B2, Cell Signaling Technology, cat#14269S, 1:10,000). Original scans of the western blot results are provided in the Source Data for individual figures.

Lung metastatic tumor cell isolation

Lungs were perfused with PBS through right ventricle, dissected and minced. The tissues were then digested with 300 units/ml collagenase type 3 (Worthington Biochemical, LS004182), 100 units/ml hyaluronidase (Worthington Biochemical, LS002592), 5 μM Y-27632 (Cayman Chemical Company, 10005583-10) and 10 μg/ml DNase I (Roche, 104159) in the DMEM/F-12 medium at 37°C for 1 hours on an orbital shaker. After primary digestion, the cell pellet was washed with 1x PBS. RBCs were lysed by RBC lysis buffer (eBioscience, 00-4300-54). Cells was further digested with 0.05% trypsin-EDTA for 5 minutes and 1 unit/ml neutral protease (dispase) (Worthington Biochemical, LS02109) plus 100 μg/ml DNase I (Roche, 104159) for 5 minutes. The digested cells were then filtered through 40 μm cell strainers to obtain single cells.

Flow cytometry and sorting

Multiparametric flow cytometric analyses were performed on an LSRII equipped with FACS Diva 6.1 software (BD Biosciences) and analyzed with FlowJo software (FlowJo, LLC). Dead cells were excluded by FSC, SSC and 49,6-diamino-2-phenylindole (DAPI, Sigma) staining. Cell sorting was performed using a MoFlow Astrios Cell Sorter (Beckman Coulter). Fluorophore-conjugated anti-CD44 (clone G44-26, cat#559942) and anti-CD24 (clone ML5, cat#555428) antibodies were obtained from BD Bioscience. Biotinylated-CD104 antibody (ITGB4, clone 439-9B, cat#13-1049-82) was purchased from eBioscience. Anti-human IFNGR1 antibody (Clone: GIT208, cat#I443) was purchased from Leinco Technologies. Unless otherwise specified, all antibodies were used at a 1:100 dilution. FACS gating strategies are provided in Supplementary Figure.

Immunofluorescence and immunohistochemistry

Formalin-fixed and paraffin-embedded tissue sections or OCT-embedded frozen sections were stained with standard protocols. The following primary antibodies were used: E-cadherin (clone 36, BD Biosciences, cat#610182, 1:200), Vimentin (clone D21H3, Cell Signaling Technology, cat#5741S, 1:200), CD44 (clone G44-26, BD Bioscience, cat#559942, 1:50), CD104 antibody (clone 439-9B, eBioscience, cat#14-1049-82, 1:200), Ki67 (Polyclonal, EMD Millipore, cat#AB9260, 1:200), cleaved caspase 3 (clone 5A1E, Cell Signaling Technology, cat#9664S, 1:200), anti-human mitochondria antibody (clone 113-1, EMD Millipore, cat# MAB1273, 1:200), Snail (clone C15d3, Cell Signaling Technology, cat#3879S, 1:200), Slug (clone C19G7, Cell Signaling Technology, cat#9585S, 1:200), Keratin 8 (Developmental Studies Hybridoma Bank, cat#AB_531826, 1:60), MLL3 (Polyclonal, Millipore Sigma, ABE1851, 1:100), α-smooth muscle actin (clone 1A4, Biolegend, cat#904601, 1:300). The ImmPRESS HRP goat anti-mouse IgG Polymer Reagent (Vector Laboratories, MP-7402) and the DAB/HRP detection IHC kit (Vector Laboratories, SK-4100) were used for immunohistochemistry. Fluorescent slides were imaged with an AXIO Examiner D1 microscope (Zeiss) with a confocal scanner unit (Yakagawa). SlideBook software 6.0 was used for the acquisition of images. Immunohistochemistry slides were scanned with a P250 High Capacity Slide Scanner (3DHISTECH).

Chemical screening and Cell viability assay

MaSCs were screened with an epigenetic compound library (Cayman #11076). 1,000 cells were seeded in 96-well plate. After 24 h, cells were treated with 0.8 μM compounds or 0.1% DMSO for 3 days. For treatment of organoid cultures, 1,000 cells were seeded in ultra-low attachment 96-well plate and the test compounds were applied at the same time. Cell viability was determined by the cell counting kit-8 (CCK-8, Dojindo, CK04-13) assay or alamar blue assay (Invitrogen, DAL1025) after 72 h treatment according to the manufacturer’s instructions. The optical density (OD) was measured by synergy H4 hybrid microplate reader (Biotec Inc). The cell viability (%) was calculated as the following: (OD_compound – OD_background)/(OD_vehicle – OD_background) × 100%.

In vivo JQ1 treatment

Small tumor fragments (~1 mm³) of Mll3-WT/Pik3ca* and Mll3-mutant/Pik3ca* tumors were orthotopically transplanted into 6-8 weeks female C57BL/6 mice. For MDA-MB-231 primary tumor, female NOD/SID mice at 6-8 weeks of age were injected with 1 × 10⁶ cells suspended in 50 μl 75% PBS/25% Matrigel into the inguinal mammary fat pads. When tumors reached ~250 mm³, mice received daily i.p. injections with 50 mg/kg JQ1 or vehicle control for 14 days. JQ1 was kindly provided by Dr. James E Bradner from Dana-Farber Cancer Institute (Boston, MA, USA). Tumor volume was measured by caliper and calculated as V = (4/3) × π × (L/2) × (L/2) × (D/2) every two days.

To assess the efficacy of JQ1 in the metastatic setting, Luc2-tomato-labeled MLL3-WT or -mutant MDA-MB-231 cells were injected via tail vein into 4-6 weeks old female NSG mice. When metastatic burden reached 6-8×10⁹ photon flux as measured by bioluminescence imaging, mice received daily i.p. injections with 50 mg/kg JQ1 or vehicle control for 14 days. Metastatic burdens were monitored by bioluminescence imaging weekly.

Quantitative RT-PCR

Total RNA was isolated from cells using the Quick-RNA microprep RNA kit (Zymo Research, R1050) or Direct-zol RNA Miniprep Plus kit (Zymo Research, R2072) and cDNA synthesis was performed by the High-Capacity cDNA Reverse Transcription Kit (Applied Biosystems, 4368814). Real-time PCR was performed with SYBR Green PCR Master Mix (Applied Biosystems, 4368708) with GAPDH as internal controls in an Applied Biosystems 7900HT real-time PCR machine with QuantStudio 6 Real-Time PCR system. Primer sequences are shown in Supplementary Table 2.

Microarray and RNA-seq gene expression analyses

Microarray gene expression profiling was done using total RNA on Affymetrix human Gene 2.0ST chips by the Einstein Genomic Core Facility according to the manufacturer’s instruction. For RNA-seq, total RNA was isolated using Direct-zol RNA Miniprep Plus kit. RNA-seq libraries were prepared and sequenced by Genewiz. RNA-seq reads were aligned using STAR (version 2.4.0) to the reference human genome (GRCh38.p13) with Ensembl annotation (gencode.v33). DESeq2 (v1.30.1) R package was applied to analyze the differential expression. Upstream regulators were analyzed by IPA (QIAGEN).

Histone mark ChIP-seq

ChIP-seq was performed as previously described ^{63, 64}. The following antibodies were used for immunoprecipitation: anti-H3K4me1 (clone D1A9, Cell Signaling Technology, cat#5326S, 2 μg per sample), H3K4me3 (clone C42D8, Cell Signaling Technology, cat#9751S, 2 μg per sample), and H3K27ac (Polyclonal, Active Motif, cat#39133, 2 μg per sample). Around 2 x 10⁷ cells and two biological repeats per genotype were used for ChIP-seq experiment. Crosslinking was performed with 1% formaldehyde (Sigma-Aldrich, F1635) at 37°C for 10 min. A final concentration of 125 mM glycine was added for 5 minutes at room temperature to quench formaldehyde crosslinking. Cells were washed with cold PBS and harvested on ice. Cells were lysed with 1ml of SDS lysis buffer (1% SDS, 10mM EDTA, 50mM Tris-HCl pH8) in the presence of protease inhibitor (Roche, cat#11836170001) and phosphatase inhibitor (Thermo, cat#78427). Cells were sonicated using the Branson Sonifier 250 (Branson Ultrasonics 101063588) with a 20% amplitude setting for 5 min 30 s with 10 s of on and off intervals. After sonication, 1 ml of ChIP dilution buffer (0.01% SDS, 1.1% Triton X-100, 1.2 mM EDTA, 16.7 mM Tris-HCl pH8, 167 mM NaCl) with proteinase and phosphatase inhibitor was added to prevent SDS precipitation. After centrifugation at 3,000 x g for 10 min at 4°C, 2 ml of supernatant were transferred to a 15 ml falcon tube and diluted with 3 ml of ChIP dilution buffer containing proteinase and phosphatase inhibitor.

50 μl of Dynabeads (Thermo Fisher Scientific 10009D), which were previously blocked with 1% BSA, were added to samples and incubated at 4°C with rotation for 1 hour. After pre-clearing, Dynabeads beads were removed. The supernatant was transferred to new 15 ml falcon tubes and 100 μl out of 5 ml volume was collected as 2% input. Antibodies were added to the pre-cleared samples for overnight incubation at 4°C with rotation. 200 μl of Dynabeads that were previously blocked with 1% BSA, were added to the sample and incubated for 6 hours at 4°C with rotation. After incubation, beads were collected and washed at 4°C with rotation following the procedure: 5 min with low salt buffer (0.1% SDS, 1% Triton X-100, 2 mM EDTA, 20 mM Tris-HCl pH8, 150 mM NaCl), twice 5 min with high salt buffer (0.1% SDS, 1% Triton X-100, 2 mM EDTA, 20 mM Tris-HCl pH8, 500 mM NaCl) and twice 5 min with TE buffer (10 mM Tris-HCl pH8, 1 mM EDTA). After washing, beads were resuspended in 250 μl of elution buffer (1% SDS, 0.1 M NaHCO3) and incubated at 60°C for 30 min using a thermomixer. The supernatant was transferred to new tubes and 5 M NaCl was added for overnight de-crosslinking at 65°C. 10 μl of 0.5 M EDTA, 20 μl of 1 M Tris-HCl (pH 6.5) and 1 μl of proteinase K (20 mg/ml, NEB, P8107S) were added and incubated at 45°C for 1 hour for de-crosslinking. ChIP-DNA was isolated by using the QIAquick PCR purification kit (QIAGEN, 28104). Libraries were prepared using the NEBNext^® Ultra II DNA Library Prep Kit (NEB, E7103S) and NEBNext^® Multiplex Oligos for Illumina^® (Index Primers Set 1; NEB, E7335S). Samples were pooled and submitted to MSKCC Integrated Genomics Operation core for sequencing. Libraries were run over one lane of a HiSeq 4000 in a 50bp/50bp paired-end run, using the HiSeq 3000/4000 SBS Kit (Illumina).

UTX and MLL4 ChIP-seq

ChIP was performed as described previously with some modifications⁶⁵. Briefly, around 2 x 10⁷ MLL3-WT or -mutant MCF7 and MDA-MB-231 cells were crosslinked with 1% formaldehyde for 10min, quenched by 125 mM glycine for 10 min and rinsed twice with 1X PBS. Fixed cells were incubated in the lysis buffer containing 5 mM PIPES pH 8.0, 85 mM KCl, 0.5% NP-40 and protease inhibitors on ice for 10 min and centrifuged at 850 x g for 5 min at 4°C. Nuclear pellets were resuspended with cold TE buffer (10 mM Tris-HCl pH 8.0, 1 mM EDTA), and sonicated into 200-500 bp fragments at 4°C. The experimental chromatins mixed with 20 ng drosophila spike-in chromatin (Active Motif, cat#53083) were incubated overnight at 4°C with Dynabeads Protein A (Invitrogen, 10002D), which had been pre-coated with 1 μg Drosophila-specific histone variant H2Av spike-in antibody (Active Motif, cat#61686) and 8 μg primary target antibody. Anti-UTX (UTX #2 antibody) and anti-MLL4 (MLL4 #3 antibody) were both home-made, which were described previously ^{66, 67}. The beads were washed twice with RIPA buffer, twice with RIPA buffer containing 300 mM NaCl, twice with LiCI buffer (50 mM Tris-Cl pH 7.5/250 mM LiCl/0.5% NP-40/0.5% Na-DOC) and twice with PBS. DNA was then eluted with 100 μl elution buffer (0.1 M NaHCO₃, 1%SDS and 20 μg proteinase K) and reverse crosslinked at 65°C overnight. DNA was purified using QIAquick PCR purification kit (Qiagen) and quantified. The ChIP DNA and input DNA were used to construct libraries using NEBNext Ultra^™ II DNA Library Prep kit for Illumina (NEB, E7645) following the manufacturer’s instructions. The final libraries were sequenced on Illumina NovaSeq 6000.

ChIP-seq data analysis

The raw reads were assessed for quality using FastQC v0.11.9 (http://www.bioinformatics.babraham.ac.uk/projects/fastqc/), then trimmed using Trimmomatic (v0.40) ⁶⁸ with Illumina adaptor file and default values. The trimmed paired-end reads were then aligned to the human reference genome (GRCh38) with bwa-mem (arXiv:1303.3997v2), then sorted and indexed with samtools (v1.15.1) ⁶⁹. The resulted alignment BAM files were then processed through PICARD (https://github.com/broadinstitute/picard) tools for ‘MarkDuplicates’. The marked alignments were then used for DeepTools2 for clustering and heatmaps ⁷⁰. Firstly, BigWig files were generated from the Markduplicated bam files using bamCoverage with RPKM normalization. The BigWig files were then passed to computeMatrix, which generates a matrix of scores of ChIP-seq read density across selected genomic regions (e.g., peaks), which was used for generating heatmaps. The regions of interest, including IFNγ signature genes and enhancers, were curated from UCSC table browser and BEDOPS v2.4.40 ⁷¹. Peaks were called using MACS2⁷² with a p value of 1e⁻³. Super-enhancers were identified using ROSE based on H3K27ac peaks and ChIP-seq signals. The peak annotation was also taken from the ROSE output ^{40, 41}. IFNγ response genes are grouped into two clusters according to their H3K27ac intensity at enhancers or stitched enhancer regions between MLL3-mutant and -WT cells. The H3K27ac intensity was calculated from ChIP samples minus the input control, and then averaged for replicates. If the log (fold change) of calculated intensity (sgMLL3/sgNT) is larger than 0, the gene was classified into Cluster 1, otherwise into Cluster 2. The gene list and relevant H3K27ac intensity can be found in Source Data.

STAT1 and H3K27ac ChIP-qPCR

Around 3-4 x 10⁶ cells were crosslinked with 1% PFA at room temperature for 10 minutes, quenched with 125 mM Glycine, and rinsed twice with PBS. Cells were lysed and sonicated at 4°C using a Bioruptor Sonicator (Diagenode, UCD-200TM) into 100-600 bp fragments. For STAT1 ChIP, Dynabeads Protein A (Invitrogen, 10002D) was pre-coated with 10 μg anti-STAT1 (clone D1K9Y, Cell Signaling Technology, cat#14994S) or Rabbit IgG (clone DA1E, Cell Signaling Technology, cat#3900S) antibody. For H3K27ac ChIP, Dynabeads Protein A (Invitrogen, 10002D) was pre-coated with 5 μg anti-H3K27ac (clone D5E4, Cell Signaling Technology, cat#8173S) or Rabbit IgG (clone DA1E, Cell Signaling Technology, cat#3900S) antibody. The samples were incubated with antibody pre-coated beads at 4°C overnight. ChIP samples were then washed, eluted, and reverse crosslinked by overnight incubation at 65 °C. ChIP-DNA was purified by Phenol:Chloroform:Isoamyl Alcohol (Sigma-Aldrich, 77677-100ML) and used for qPCR analysis. ChIP enrichment was calculated relative to input samples. Primers used for ChIP-qPCR are listed in Supplementary Table 3.

ATAC-seq and data analysis

100,000 FACS sorted CD44⁺CD104^high and CD44⁺CD104^−/low cells were used for ATAC-seq assay. All subsequent processing using standard DNA extraction and library preparation protocols were performed by Novogene. Briefly, nuclei are extracted from samples, and the nuclei pellet is resuspended in the Tn5 transposase reaction mix. The transposition reaction is incubated at 37°C for 30 min. Equimolar Adapter1 and Adatper 2 are added after transposition, PCR is then performed to amplify the library. After the PCR reaction, libraries are purified with the AMPure beads. After library quality control, sequencing is performed for different libraries according to the concentration and the demand of data amount on Illumina NovaSeq platform. For data analyses, after investigating the quality of raw data by FASTQC, raw reads were trimmed with BBduk.sh, with the parameters minlen=50 qtrim=r trimq=6 ktrim=r k=25 mink=11. Alignment to the human genome was done with bowtie2 (v. 2.4.1) with the parameters _mm _no-mixed _no-discordant _very-sensitive -X 1000, duplicates were removed with samblaster, and reads from chrM were removed with sambamba. Differential ATAC-seq peaks were called using Genrich (v. 0.6; https://github.com/yejg2017/Genrich), and blacklisted regions were filtered out in the process. Motif analysis on peaks was performed with Homer (v4.11).

Statistics and reproducibility

Data are represented as mean ± SEM. Statistical tests were performed with GraphPad Prism V8.0 (San Diego, CA). Data distribution was assumed to be normal, but this was not formally tested. A two-tailed Student’s t-test was used for comparisons of continuous variables between two groups. One-way ANOVA with Tukey or Dunnett test was used when three or more groups were compared. Two-way ANOVA with multiple comparisons test was used to check the effect of two factors on a dependent variable. Chi-square likelihood ratio test was used to check the limiting dilution assay. Fisher’s right tailed exact test was used in IPA. Kaplan-Meier survival curves were analyzed by using GraphPad Prism V8.0 or 9.0 (San Diego, CA) and p values were calculated using the log-rank test. Statistical details (sample size and specific test performed) for each experiment are denoted in the corresponding figure or figure legend. No statistical method was used to predetermine the sample size. Sample sizes were determined based on previous experimental experience. No data were excluded from the analyses. The in vitro cell experiments were not randomized. For in vivo tumor onset, progression, and metastasis experiments, mice were randomly assigned to experimental groups. For experiments with chemical inhibitors, samples were randomly allocated to control and experimental groups. For tumor onset and progression experiments, investigators were blind to group information. For metastasis experiments, investigators were not blind to group information, however, data reported for metastasis experiments are not subjective but rather based on quantitative total photo flux by IVIS. Investigators were not blinded to cell genotypes during in vitro experiments, because experiments required frequent intervention by investigators to maintain cell lines.

Data Availability

RNA–seq, ChIP-seq, ATAC-seq and microarray data that support the findings of this study have been deposited in the Gene Expression Omnibus (GEO) under accession codes GSE171447. The human reference genome (GRCh38) was used as a reference genome for the alignment. The human breast cancer data were derived from the TCGA Research Network: http://cancergenome.nih.gov/. The data-set derived from this resource that supports the findings of this study is available in the details of Source data. Other source data for all figures are provided with the paper. All other data supporting the findings of this study are available from the corresponding author on reasonable request.

Code Availability

The code for in-house scripts used in this study was deposited into the following link: https://github.com/zcmit/EinsteinMed/tree/main/Cui_etal.2022.