MDA-MB-231 and HEK293 cells were obtained from ATCC. MDA-MB-231 cells were cultured in EMEM (Sigma Aldrich) containing 10% FBS, L-glutamine, vitamin mix, non-essential amino acids, sodium pyruvate and penicillin/streptomycin. HEK293 cells were grown in DMEM (Sigma Aldrich) containing 10% FBS and penicillin/streptomycin. Cells were regularly tested for Mycoplasma and confirmed to be Mycoplasma free. Whenever specified, MDA-MB-231 cells were labeled using CellVue Claret^® Far Red Fluorescent dye (Sigma Aldrich) according to manufacturer’s instructions.

For in vitro assays, MDA-MB-231 cells were seeded at 1x10⁵ cells per 6-well. High glucose DMEM/F-12 medium (Thermofisher, 11320033) was used for experiments with 2-DG (Sigma Aldrich) and medium replenished daily.

Neomycin resistance gene was inserted downstream of PGK promoter in pTREAutoR3 lentiviral vector (8). GFP gene downstream of a doxycycline-inducible minimal CMV promotor was replaced with BGN gene. Vector without an insert was used as s control. Lentivirus was generated as previously described (9). Following transduction, MDA-MB-231 cells were maintained in medium containing TET Systems approved FBS (Thermofisher). Neomycin (500µg/ml) was added to the medium until all non-transduced cells perished. Cells transduced with pTREAutoR3_Neo_BGN and control vector were named MDA-MB-231/BGN and MDA-MB-231/CON, respectively. BGN expression was induced by adding doxycycline (1µg/mL) for at least 2 days. For in vivo experiments, cells were tagged with Firefly luciferase as described (9).

Six- to eight-week-old C.B.17 SCID mice (C.B-lgh-1^b/lcrTac-Prkdc^scid) were purchased from Charles Rivers Laboratories. Brain tumor xenografts were generated by injection of 1x10⁵ CV-labelled MDA-MB-231 cancer cells into the internal carotid artery (10) or through implantation of firefly luciferase-tagged MDA-MB-231/BGN and MDA-MB-231/CON cancer cells (1x10⁵) into the striatum as previously described (9). Doxycycline (100 mg/kg BW) was administered i.p. on days 1, 3 and 5 post-cancer cell injection, and in water (2 mg/mL) from day 6 on. Tumor growth was monitored in vivo by bioluminescence imaging, using IVIS Spectrum (Perkin Elmer). Living Image software (Perkin Elmer) was used for quantification of bioluminescence signals.

All procedures were approved by the University of Leeds Animal Welfare & Ethical Review Committee and performed under the approved UK Home Office project license.

Left hemisphere of the brains isolated from mice at 4 weeks post-intracarotid injection of CV-labeled GFP+ cancer cells were enzymatically dissociated in EMEM containing 3mg/ml collagenase and 250 U/ml hyaluronidase for 20 minutes at 37°C. Tissue was washed in cold incubation buffer (0.5% BSA and 2mM EDTA in PBS) and strained. Myelin was removed using Myelin Removal Beads II (Miltenyi Biotec). GFP+CV- and GFP+CV+ cancer cells were sorted from the resulting cell suspension using the Influx v7 Sorter (BD Biosciences). Cultured MDA-MB-231 cells (GFP-tagged, untagged CV-labelled, GFP-tagged and CV-labeled, and untagged/unlabeled) were used for compensation. Analysis was performed using FlowJo.

Total RNA from sorted cancer cells was isolated using the Arcturus^® PicoPure™ RNA Isolation Kit (Thermo Fisher), followed by DNA removal using the RNase-free DNase Set (Qiagen). Smart-seq2 protocol (11) and the Nextera XT DNA Library Prep kit (Illumina) were used to generate full-length cDNA and sequencing libraries. PCR amplification was performed for 15 cycles. Libraries were pooled and paired-end mRNA sequencing was performed using the Hiseq3000 platform (Illumina).

Data processing was performed using R/Bioconductor. Reads were quality-assessed and trimmed using FastQC (12) and Cutadapt (13), respectively. Primary assemblies and comprehensive gene annotations for human, release 31 (GRCh38.p12), and mouse, release M22 (GRCm38.p6), were sourced from GENCODE (14). Human and mouse chromosomal identifiers were renamed for disambiguation, and assemblies concatenated using Biostrings (15). Reads were aligned using STAR aligner (16), with Qualimap (17) and Picard Tools (18) used for quality assessment. Read quantification, transcriptome merging and count mapping was performed using RSubread (19). Multi-mapping reads were included. Mouse-mapping reads, as identified from prior chromosomal renaming, were discarded as the desired cancer cells were of human origin. Read counts were converted to integers, and size factor normalization was performed using DESeq2 (20). Ensembl IDs were converted to gene symbols using Ensembl Gene ID Converter (21).

Genes were analyzed using an FDR =< 0.05 cutoff, unless otherwise stated. Functionally implicated transcription factors were predicted using TFactS (22). Functional protein annotation networks were visualized using STRING (23). Heat maps (unsupervised hierarchical clustering), constructed at the transcript level, and principal component analysis (PCA) plots, were generated using ClustVis (24). Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis was carried out using ClusterProfiler (25), with no output statistical cut-off, however the input gene list was restricted to a minimum fold change of 2.

The mRNAseq datasets generated for this study can be found in the Gene Expression Omnibus database with the accession code GSE220017. Publicly available data sets analyzed in this study included GSE2034, GSE5327, GSE12276, GSE14017 and GSE43837.

Raw data from GSE2034, GSE5327, GSE12276, GSE14017 and GSE43837 was extracted and independently normalized using the R package Affy (26). Gene expression data from Varešlija et al. were downloaded from the GitHub repository rpriedig (27). Pre-normalized BGN gene expression data was further normalized to that of POLR2A expression prior to analysis.

Total RNA from cells grown in vitro was isolated using the RNeasy Mini Kit (Qiagen). cDNA synthesis was performed with Superscript III Reverse Transcriptase kit (Invitrogen). Taqman PCR was performed as previously described (9). All assays were from ThermoFisher: BGN (Hs00959143_m1), GAPDH (Hs02786624_g1), ITGB1 (Hs01127536_m1), ITGB2 (Hs00164957_m1), ITGB4 (Hs00236216_m1), and POLR2A (HS00172187_m1).

Cell lysis and Western blot were performed as previously described (9). Primary antibodies were directed against biglycan (Proteintech, 16409-1-AP; 1:800), YAP1 (Santa Cruz Biotechnology, sc-101199; 1:1000), phospho-YAP1 (Fisher scientific, PA5-17481; 1:1000), and vimentin (DAKO, M0725; 1:1000). HRP-conjugated secondary anti-mouse and anti-rabbit antibodies were from Cell Signalling. Band intensity was quantified using Fiji image processing package.

Cells were incubated for 30 minutes with 10μM BrdU prior to harvest, and re-suspended in PBS, followed by dropwise addition of 9x volume of cold 70% ethanol and incubation on ice for 30 minutes. Cells were first re-suspended in denaturation buffer (PBS, 2M HCl, 0.5% Triton X-100) for 30 minutes at RT, followed by incubation in neutralization buffer (PBS, 0.1M Na₂B₄O₇.10H₂O, pH 8.5) for 30 minutes at RT, and then re-suspended in PBS, 1% w/v BSA, 0.5% v/v Tween-20 containing anti-BrdU-APC antibody (eBioscience, 17-025-152), and incubated for 1 hour at RT. Cells were washed 3x in PBS and incubated in PBS with 5μg/ml RNase A (Qiagen) and 10μg/ml propidium iodide (Sigma Aldrich) for 30 minutes at RY prior to flow cytometry analysis using the CytoFLEX Flow Cytometer (Beckman Coulter).

Cells were grown on plastic and fixed in 4% PFA for 10 minutes at RT. Mouse brain tissue was fixed and processed for floating sections, and staining performed as previously described (9). Incubation with all primary and secondary antibodies was for 1 hour at RT. Primary antibodies were directed against biglycan (Proteintech, 16409-1-AP; 1:200), GFP (Abcam, Ab13970; 1:1000), and YAP1 (Novus Biologicals, NB110-58358; 1:50). Secondary antibodies were from Jackson Immunoresearch. Nuclei were stained with DAPI. Images were acquired using AxioCam MRm (Zeiss) and AxioVision software (Zeiss), or the A1R confocal microscope equipped with Confocal NIS-Elements software (Nikon).

Statistical analyses of data not derived from mRNAseq outputs were carried out using GraphPad Prism version 8.0.0 for Windows (GraphPad). The error bars on all graphs represent the standard error of the mean (SEM). Statistical significance between the experimental groups was determined by t-test or One-way ANOVA followed by multiple comparison analysis, as specified in figure legends.

We previously reported that, following its injection into the internal carotid artery, the triple negative breast cancer cell line MDA-MB-231 displayed a low efficiency of cancer cell outgrowth in the brain following initial cancer cell seeding (Lorger and Felding-Habermann, 2010). Analysis of coronal brain sections by immunofluorescence 4 weeks after the injection of green fluorescence protein (GFP)-tagged MDA-MB-231 cells confirmed the presence of cancer cells in the brain, either as small cell clusters or single cells, with very few larger cancer lesions observed ( Figure 1A ). To distinguish between cancer cells that have undergone proliferation following their arrival in the brain versus those that remained dormant, GFP+ cancer cells were labelled with CellVue Claret (CV) vital dye. We confirmed in vitro that in proliferating cultured MDA-MB-231 cells this dye is diluted below the limit detectable by flow cytometry within 14 days, as the dye is being equally split between daughter cells at each cell division ( Supplementary Figure 1A ). As such CV loss can be used to detect proliferating cells, while non-proliferating cells are expected to retain the dye.

CV-labelled GFP+ MDA-MB-231 cells were injected into the internal carotid artery of CB17SCID mice ( Figure 1B ). We have previously demonstrated that cancer cells start extravasating into the brain at ~day 3 post-injection and all cells are extravascular by day 7 (10). To determine whether GFP+CV+ cancer cells can be detected in the brain prior to resuming proliferation, we isolated and dissociated whole brains 4 days post-cancer cell injection and analyzed them by flow cytometry. At this early time point, all cancer cells appeared within the GFP+CV+ gate and with no events observed within the GFP+CV- gate. The latter was set based on the analysis of brains isolated from mice following the injection of GFP+ cancer cells not labelled with CV and based on cultured GFP+, CV+ and GFP+CV+ MDA-MB-231 cells ( Figure 1C ; Supplementary Figure 1B ). This confirmed a reliable separation of GFP+CV- and GFP+CV+ cancer cell populations in mouse brains by flow cytometry.

We next isolated by FACS non-proliferating dormant (GFP+CV+) and proliferating (GFP+CV-) MDA-MB-231 cancer cells from the mice brains 4-weeks after the injection of GFP+ cancer cells labeled with CV into the internal carotid artery. Due to the low number of cancer cells in the brain, 9-10 mice brains were pooled. The majority of cancer cells lost CV dye and appeared within the GFP+CV- gate as expected, as proliferating cells outnumbered any dormant cells within the 4-week period due to their expansion. However, a clear population of GFP+CV+ cancer cells remained ( Figure 1C , right). We also confirmed that the cells were GFP+CV- or GFP+CV+, respectively, by immunofluorescence ( Figure 1D ). To obtain biological triplicates, 3 independent experiments were performed, and isolated cells subjected to the genome-wide gene expression analysis by mRNAseq. As summarized in Figure 1E , GFP+CV+ cancer cells represented between 0.58 to 2.02% of all isolated GFP+ cancer cells.

Principal component analysis (PCA) revealed separate clustering of dormant and proliferating cell transcriptomes ( Figure 1F ). Unsupervised hierarchical clustering of samples identified 1161 genes that were differentially expressed between dormant and proliferating cancer cells (FDR < 0.05) ( Figure 1G ; Supplementary Table IA ). The most significantly differentially expressed gene was BGN (FDR 1.7x10^-100), which encodes a small extracellular matrix protein biglycan (28), with 474-fold upregulation in dormant versus proliferating cancer cells ( Figure 1H ).

To determine whether any of the previously identified dormancy markers (29) are associated with a dormant cancer cell phenotype in the brain, we analyzed the expression of these markers in our data set. While overall dormancy-associated markers were enriched in GFP+CV+ MDA-MB-231 cells isolated from the brains, this was not the case for all markers ( Supplementary Figure 1C ; Supplementary Table IB ), suggesting a microenvironment and/or cancer type-dependent regulation of dormancy.

While most MDA-MB-231 cancer cells in the brain were biglycan-negative at 4 weeks post-cancer cell injection as analyzed by immunofluorescence, we detected rare biglycan-positive GFP+ cells ( Figure 2A ), confirming the existence of biglycan-expressing cancer cells in vivo at the protein level.

Similarly to proliferating cancer cells isolated from the murine brain, BGN was also absent from the fast proliferating MDA-MB-231 cells grown in 10% serum in vitro ( Figure 2B ). To investigate whether biglycan is functionally implicated in the regulation of cancer cell growth in vivo, we generated MDA-MB-231 cells stably expressing BGN (MDA-MB-231/BGN) under the control of a doxycycline-inducible promoter. We used an inducible promoter because BGN over-expression under a strong constitutive promoter resulted in cell growth arrest and precluded cell expansion (data not shown). Cells transduced with an empty vector were used as a control (MDA-MB-231/CON). Upon addition of doxycycline, a strong induction of biglycan could be detected in MDA-MB-231/BGN cells by Western blot in contrast to biglycan-negative MDA-MB-231/CON cells ( Figure 2C ). Firefly luciferase-tagged MDA-MB-231/CON and MDA-MB-231/BGN cells cultured in the presence of Dox for 2 days were subsequently implanted into the brain of CB17SCID mice ( Figure 2D ). Biglycan expression was maintained by doxycycline administration in vivo for the duration of the study. BGN expression in tumors was analyzed by qPCR at the endpoint and confirmed to be comparable to in vitro expression levels ( Figure 2E ). We observed a delayed tumor growth in mice with BGN-expressing tumors as compared to the control ( Figure 2F ; Supplementary Figure 2 ). This appeared to be due mainly to a delay in the cancer cell outgrowth following the initial 2-week lag-phase, as demonstrated by a significantly lower difference in the tumor burden increase between days 14 and 19 in the BGN over-expressing group ( Figure 2G ). This suggested that BGN may be primarily inhibiting the initial outgrowth of cancer cells. Notably, BGN overexpression also significantly inhibited the growth of MDA-MB-231 cells in vitro ( Supplementary Figure 1D ).

To establish a potential role of BGN in brain metastases in breast cancer patients, we analyzed publicly available gene expression data. This revealed a significant downregulation of BGN in brain metastases as compared to the patient-matched primary breast tumors ( Figure 2H ). Further analysis of non-matched primary breast tumors and breast cancer metastases from different organs revealed a significantly downregulated BGN expression in lung and brain metastases, and a tendency towards reduced BGN expression in liver metastases ( Figure 2I ). This demonstrated that cancer cells that can proliferate in the brain and develop into large metastases are associated with reduced BGN expression levels in patients, in line with our data in a preclinical model.

To identify potential differences in the activity of transcription factors between dormant and proliferating cancer cell populations, differentially expressed genes were analyzed using TFactS software (22). This revealed that HIF1α and MYC were significantly repressed in dormant as compared to the proliferating cancer cells ( Figure 3A ). Accordingly, HIF1α and MYC-dependent genes were differentially expressed between dormant and proliferating cancer cell samples ( Figure 3B ). Both transcription factors have been shown to induce a metabolic reprogramming towards aerobic glycolysis (36). Notably, nine of the HIF1α and MYC regulated genes that were downregulated in dormant cells are known to be involved in the regulation of glycolysis ( Figure 3C ). Thus, to further investigate whether inhibition of glycolysis induces reversible cell growth arrest in breast cancer cells, MDA-MB-231 cells were incubated with different concentrations of glucose analogue 2-deoxyglucose (2-DG). This reduced the growth rate of cancer cells in a dose-dependent manner ( Figure 3D ). Notably, at 10 mM 2-DG concentration, the number of live cells remained constant over a course of 9 days ( Figure 3D ) and only a low number of floating cells was observed, comparable to the control without 2-DG ( Figure 3E ). In contrast, higher 2-DG concentrations (20 and 50 mM) increased the proportion of dead floating cells, while lower 2-DG concentration (5 mM) failed to inhibit cell growth completely. Importantly, when 2-DG was removed to restore glycolysis, cancer cells resumed proliferation ( Figure 3D ), demonstrating that inhibition of glycolysis with low 2-DG concentration (10 mM) results in a reversible growth arrest in the absence of increased cell death, as seen in dormancy. In line with this, cancer cells retained CV dye when cultured in the presence of 10 mM 2-DG, while in the absence of 2-DG the CV was lost within 10 days ( Figure 3F ), mimicking our in vivo observations.

Inhibition of glycolysis with 10 mM 2-DG for 9 days significantly increased the percentage of MDA-MB-231 cells in G2/M phase as compared to control and reduced the percentage of cells in the S phase, while the proportion of cells in G0/G1 phase remained unaltered ( Figures 3G, H ). This cell cycle arrest was reversible following 2-DG removal, with percentages of cancer cells in different cell cycle phases returning to control levels at 5 days post-2-DG removal ( Figures 3G, H ). This suggested that inhibition of glycolysis is causing a reversible growth arrest of cancer cells in G2/M phase.

KEGG analysis on genes differentially expressed between dormant and proliferating cancer cells isolated from the mouse brains identified several differences ( Figure 4A ; Supplementary Table II ). Hippo signaling pathway was amongst the top enriched pathways in dormant cells ( Figure 4A ), with Scribble planar cell polarity protein (SCRIB) and Disks large homolog 3 (DLG3), two upstream activators of Hippo pathway (37, 38), being significantly upregulated in dormant as compared to the proliferating cells (5.1- and 2.1-fold change, respectively; Supplementary Table I ). Transcription factor Yes associated protein (YAP) is at the core of the Hippo signaling pathway. As glycolysis has been previously implicated in the regulation of YAP activity (39) and we have demonstrated that mild inhibition of glycolysis induces reversible cancer cell growth arrest reminiscent of dormancy, we sought to further investigate a functional link between glycolysis and Hippo pathway in this context. When Hippo pathway is activated, YAP becomes phosphorylated, which prevents its translocation into the nucleus and blocks YAP-dependent gene transcription and proliferation (37). Inhibition of glycolysis with 10 mM 2-DG resulted in a significant increase in YAP phosphorylation ( Figures 4B, C ) and its cytoplasmic retention ( Figure 4D ), suggesting that inhibition of glycolysis leads to Hippo activation. This further suggests that downregulated glycolysis may induce dormancy through Hippo pathway in vivo.

Notably, YAP is involved in the regulation of cell adhesion to the extracellular matrix by regulating the expression of various integrins. We observed a significant downregulation of ITGB4 in dormant as compared to the proliferating cancer cells in vivo, and in 2-DG-treated as compared to vehicle-treated cancer cells in vitro ( Figures 4E, F ), while the expression of ITGB1 was unaltered. This suggests that glycolysis may be involved in dormancy potentially by regulating integrin expression and consequently cell adhesion to the basement membrane.