NOTCH-mediated non-cell autonomous regulation of chromatin structure during senescence

Senescent cells interact with the surrounding microenvironment achieving diverse functional outcomes. In addition to autocrine and paracrine signalling mediated by factors of the senescence-associated secretory phenotype, we have recently identified that NOTCH1 can drive ‘lateral induction’ of a unique form of senescence in adjacent cells through specific induction of the NOTCH ligand JAG1. Here we show that NOTCH signalling can modulate chromatin structure both autonomously and non-autonomously. In addition to senescence-associated heterochromatic foci (SAHF), oncogenic RAS-induced senescent (RIS) cells in culture exhibit a massive increase in nucleosome-free regions (NRFs). NOTCH signalling suppresses both SAHF and NFR formation in this context. Strikingly, NOTCH-induced senescent cells, or cancer cells with high JAG1 expression, also drive similar chromatin architectural changes in adjacent cells through cell-cell contact. Mechanistically, we show that NOTCH signalling represses the chromatin architectural protein HMGA1, an association found in a range of human cancers. Thus, HMGA1 is involved not only in SAHFs, but also RIS-specific NFR formation. In conclusion, this study identifies that the JAG1-NOTCH-HMGA1 axis mediates the juxtacrine regulation of chromatin architecture.

ABSTRACT Senescent cells interact with the surrounding microenvironment achieving diverse functional outcomes. In addition to autocrine and paracrine signalling mediated by 5 factors of the senescence-associated secretory phenotype, we have recently identified that NOTCH1 can drive 'lateral induction' of a unique form of senescence in adjacent cells through specific induction of the NOTCH ligand JAG1. Here we show that NOTCH signalling can modulate chromatin structure both autonomously and non-autonomously. In addition to senescence-associated heterochromatic foci show that NOTCH signalling represses the chromatin architectural protein HMGA1, an association found in a range of human cancers. Thus, HMGA1 is involved not only in SAHFs, but also RIS-specific NFR formation. In conclusion, this study identifies that the JAG1-NOTCH-HMGA1 axis mediates the juxtacrine regulation of chromatin architecture. 20

INTRODUCTION
Cellular senescence is an autonomous tumour suppressor mechanism that can be triggered by multiple pathophysiological stimuli including replicative exhaustion 1,2 , exposure to chemotherapeutic drugs 3 and hyper-activation of oncogenes such as RAS 4,5 . Persistent cell cycle arrest, an essential feature that defines senescence, is 5 accompanied by diverse transcriptional, biochemical, and morphological alterations.
These senescence hallmarks include increased expression and secretion of soluble factors (senescence-associated secretory phenotype; SASP) 6,7 , senescenceassociated β-galactosidase activity, enlarged nuclei, and dramatic alterations to chromatin structure [8][9][10] . Importantly, the combination, quantity and quality of these 10 features can vary depending on the type of senescence. Senescent cells have profound non-cell autonomous functionality mediated primarily by the SASP. The SASP can have either pro-or anti-tumorigenic effects and act in an autocrine or paracrine fashion [11][12][13][14] . In addition, we have recently identified that NOTCH signalling can drive a cell-contact dependent juxtacrine senescence 15 . 15 The NOTCH signalling pathway is involved in a wide array of developmental and physiological processes. During development and in tissue homeostasis NOTCH has well defined roles in differentiation and stem cell fate 16  translocates to the nucleus where, together with transcriptional co-activators such as mastermind-like 1 (MAML1), it drives transcription of canonical target genes including the HES / HEY family of transcription factors 16 . NOTCH signaling has also been shown to induce a type of senescence, NOTCH-induced senescence (NIS), where cells are characterised by distinct SASP components 15, 18 . Recently, we 5 showed that during NIS there is a dramatic and specific upregulation of JAG1 that can activate NOTCH1 signaling and drive NIS in adjacent cells (called 'lateral induction') 15 .
SAHF formation is dependent on the transcriptional upregulation and accumulation of chromatin-bound High-Mobility Group A (HMGA) proteins 27 . These are a family of flexible architectural proteins, consisting of HMGA1 and HMGA2, which bind to the 20 minor groove of AT-rich DNA via three AT-hook domains to alter chromatin structure 28,29 . Despite a critical role in the formation of SAHFs during senescence, HMGA proteins are also important during development where they promote tissue growth 30,31 and regulate differentiation [32][33][34][35] . Furthermore, a large body of literature has accumulated that in most cases demonstrate an association between high HMGA expression and tumour malignancy as indicated by metastasis, chemo-resistance and poor prognosis 36,37 . For example, HMGAs are upregulated in approximately 90% of lung carcinomas 38 . Here, we characterise the chromatin phenotype in RIS and NIS cells. We demonstrate that these two types of senescent cells exhibit distinct chromatin 20 structures at a microscopic and nucleosome scale. Both gain multiple accessible regions or 'nucleosome free regions' (NFRs), which are often exclusive between RIS and NIS and reflect the transcriptional landscape of each phenotype. Strikingly, we find that autonomous and non-cell autonomous activation of the NOTCH signalling pathway in RIS cells can repress SAHFs and the formation of RIS-specific NFRs, partially by transcriptional repression of HMGAs. Our study demonstrates that chromatin structure and the nucleosome landscape can be regulated through juxtacrine signalling. The relationship between these two prominent tumour associated genes, HMGA and NOTCH1, may also have prognostic value in vivo. 5

NOTCH1 reprograms chromatin structure and abrogates SAHFs
We have previously demonstrated that ectopic NOTCH1 intracellular domain (N1ICD), an active form of NOTCH1 (Fig. 1a), can drive NIS that is distinct from RIS 10 in terms of SASP composition 15 . We noticed that NIS cells also have a unique chromatin structure and sought to investigate how this compares to RIS.
To ask whether NIS cells simply lack SAHFs or whether N1ICD actively modulates chromatin structure, we expressed N1ICD in the presence of HRAS G12V .
Interestingly, N1ICD in the context of RIS also resulted in a dramatic enlargement of nuclei but a complete ablation of SAHF formation (Fig 1b-d). This was emphasised by a 'smoothening' of chromatin as indicated by a marked reduction in the standard 5 deviation of DAPI signal measured within individual nuclei (Fig. 1b,  inhibitor DAPT to repress downstream signalling by N1ICD (Fig. 1a). We found that a greater number of SAHF-positive cells were formed and that these accumulated at earlier time points when NOTCH1 signalling was repressed (Fig. 1e). Furthermore, a dose dependent effect was evident where higher concentrations of DAPT resulted in a greater proportion of cells developing SAHF during RIS (Supplemental Fig. 1e).

Non-cell autonomous regulation of SAHFs
N1ICD-expressing cells can induce NIS in adjacent normal cells, at least in the case of IMR90 cells 15 . To determine whether N1ICD-expressing cells can also alter chromatin structure in adjacent cells, we performed co-cultures between mRFP1-10 expressing IMR90 ER:HRAS G12V cells and IMR90 cells retrovirally infected with doxycycline (DOX)-inducible FLAG-N1ICD (iN1ICD) in the presence and absence of 4OHT and DOX (Fig. 2a). Strikingly, co-culture with N1ICD-expressing IMR90 cells was sufficient to repress SAHF formation in adjacent RIS (red) cells (Fig 2b, c).

15
Of the canonical NOTCH1 ligands, we have previously observed a strong and unique upregulation of JAGGED1 (JAG1) following ectopic N1ICD expression, which we found to be responsible for the juxtacrine transmission of senescence 15 . We reasoned that N1ICD-mediated upregulation of JAG1 and subsequent 'lateral induction' of NOTCH1 signalling is a likely mechanism by which SAHFs are 20 regulated non-autonomously. To test this hypothesis we expressed ectopic JAG1 fused to mVenus (JAG1-mVenus) in hTERT-immortalised human Retinal Pigment Epithelial (RPE1) cells. We confirmed cell surface expression of ectopic JAG1 by flow-cytometry (Supplemental Fig. 2a) before co-culturing with IMR90 ER:HRAS G12V mRFP1 cells. Whilst control RPE1 cells had no effect on SAHF formation in red IMR90 RIS cells, RPE1 JAG1-mVenus cells significantly repressed the formation of SAHFs (Fig 2e, f). Note this repression did not occur when these two types of cells were co-cultured without any physical contact in a transwell format (Supplemental Fig. 2b). Our data suggest a mechanism by which lateral induction of NOTCH 5 signalling by JAG1 can block SAHFs in the context of RIS and demonstrate a novel mechanism by which higher-order chromatin structure can be regulated through cellcell contact.

NOTCH signalling inhibits HMGA expression 10
To unravel the mechanisms underpinning NOTCH1 dependent repression of SAHFs, we re-analysed previously published RNA-seq data generated from IMR90 cells expressing HRAS G12V and N1ICD 15 . We observed that N1ICD dramatically represses the expression of High-Mobility Group AT-hook 1 (HMGA1) and HMGA2 (Supplemental Fig. 3a), critical components of SAHF structure 27 . 15 To further validate that NOTCH1 signalling represses HMGA expression, we introduced constitutive N1ICD into ER:HRAS G12V expressing IMR90 cells. Ectopic N1ICD significantly repressed HMGA1 and HMGA2 at an mRNA and protein level in both the presence and absence of 4OHT-induced HRAS G12V (Fig. 3a, b). N1ICD has 20 a similar effect on HMGA expression when expressed in other cell lines in the absence of HRAS G12V , suggesting a conserved mechanism (Supplemental Fig. 3b).
In the DOX-inducible FLAG-N1ICD system, inhibition of NOTCH1 signalling by coexpression of dnMAML1-mVenus was sufficient to rescue N1ICD-mediated repression of HMGA1/ 2 (Fig. 3c, d), suggesting the effect is dependent on the canonical pathway of NOTCH signalling.
Finally, we used IMR90 ER:HRAS G12V cells expressing DOX-inducible FLAG-N1ICD to investigate whether ectopic re-expression of EGFP-tagged HMGA1 is sufficient to 5 rescue SAHFs. The introduction of EGFP-HMGA1 resulted in a partial but significant rescue of SAHF-positive cells when cells were treated with DOX and 4OHT (Fig. 3e).
Collectively, our data suggest that NOTCH1 signalling represses the formation of SAHFs at least partially by inhibiting HMGAs, establishing a novel connection 10 between these two prominent cancer-associated genes.

Non-cell autonomous inhibition of HMGA
To determine whether HMGAs are repressed non-autonomously by JAG1 expressing cells, we performed further co-cultures between RPE1 cells retrovirally 15 infected with JAG1-mVenus and IMR90 cells ectopically expressing a cell surface marker, rat Thy1, allowing for subsequent isolation using magnetic-activated cell sorting (MACS) (Fig. 3f). As expected, IMR90 cells co-cultured with JAG1 expressing cells expressed canonical NOTCH1 target genes such as HEY1 and HEY-L.
Strikingly, both HMGA1 and HMGA2 were significantly repressed in IMR90 cells 20 cultured with JAG1 expressing RPE1 cells when compared to controls (Fig. 3f), demonstrating that HMGA proteins can be repressed non-autonomously.

Gene-distal nucleosome free regions accumulate in RIS and NIS
To investigate whether NOTCH1 influences chromatin structure at a higher resolution, we employed ATAC-seq (Assay for Transposase-Accessible Chromatin using Sequencing) 46 . This method takes advantage of a hyperactive Tn5 transposase enzyme that inserts sequencing adapters into accessible chromatin, or 'nucleosome-free regions' (NFRs). Following PCR amplification these regions can be 5 sequenced as a method of identifying NFRs genome wide (Fig. 4a). Indeed, chromatin accessibility is known to be a direct reflection of cell phenotype and the transcriptional landscape 40 .
We generated at least 3 replicates from IMR90 ER:HRAS G12 cells expressing FLAG-10 N1ICD or a control vector and induced with 4OHT or not. For simplicity, these conditions were labelled as 'Growing', 'RIS', 'NIS' and 'N+RIS (expressing both N1ICD and RAS)'. Using a previously published trimmed-means normalisation approach 47 , we first generated normalised coverage files. When interrogated using a genome browser, these were comparable to each other and to ENCODE DNase-15 sequencing data from normal human lung fibroblasts (NHLF), especially around housekeeping genes (Fig. 4b). Most of the samples, excluding a single replicate from the NIS and N+RIS conditions, were of high quality with a reads in peaks percentage (RiP%) of more than 10% (Supplemental Fig. 4a). Replicates clustered well by unbiased clustering and PCA analyses (Fig. 4c, Supplemental Fig. 4b). 20 To generate a consensus set of high confidence NFRs for downstream analyses, we identified peaks present in at least two replicates of each condition (Fig. 4d Table 2). We isolated 50,361 peaks unique to RIS cells ('RIS unique NFRs'), 55,936 peaks unique to NIS cells ('NIS unique NFRs') and 54,670 peaks present in both RIS and NIS cells that were not detected in growing cells ('Novel shared NFRs'). All three of these subsets had a tendency to be gene distal when compared to peaks found in growing IMR90 cells, with RIS-unique NFRs 5 showing the greatest tendency (Supplemental Fig. 5a).
Next, we mapped NFRs within 300bp of a gene TSS to genes (in order to confidently map NFRs to genes) and performed Gene Ontology (GO) analyses with these lists.
Consistent with our previous RNA-seq data 15 , genes with RIS-unique NFRs were 10 significantly enriched in the term 'inflammatory response' whilst genes with NISunique NFRs were enriched in the terms 'extracellular structure organization' and 'extracellular matrix organization' and included genes that transcribe fibrotic proteins such as collagens (Fig. 4f).

15
Genes with 'Novel shared NFRs' were enriched in the term 'apoptotic signalling pathway', which was particularly striking as both pro-and anti-apoptotic proteins have recently been described to play important roles in senescence (Fig. 4f) [48][49][50] . For example, we detected a novel shared NFR at the promoter of the gene BCL2L1, which encodes the protein BCL-XL, an anti-apoptotic protein and potential target for 20 'senolytic drugs' 49,50 . The increased accessibility at the promoters of these genes in both RIS and NIS suggests a unique mode of apoptotic sensitivity in senescent cells, which are generally apoptosis-resistant but can be selectively killed using senolytics.
Finally, we asked whether the genes in our gene lists are differentially expressed in RIS and NIS cells by RNA-seq. To do so, we reanalysed previously generated data 15 and found that, on average, genes with RIS-specific NFRs were transcriptionally upregulated in RIS cells, whilst genes with NIS-specific NFRs were transcriptionally upregulated in NIS cells (Supplemental Fig. 5b, Supplemental Table 3). Normalised 5 ATAC-seq coverage files, when viewed using a genome browser (Fig. 4g, h, Supplemental Fig. 5c), supported these observations. We also noted that, whilst few inflammatory genes gained RIS-unique NFRs at their promoters, some had upstream enhancer elements that become accessible in RIS cells (

NOTCH signalling antagonises the formation of RIS specific NFRs
By unbiased clustering, we observed a greater correlation between NIS and N+RIS cells than between RIS and N+RIS cells (Fig 4c, Supplemental Fig. 4b). This 15 suggests a dominant effect of N1ICD over RAS on the nucleosome scale, which is consistent with our previous observations for SASP components 15 and SAHFs (Fig.   1d).
To validate whether NFRs in RIS are repressed by N1ICD, we took the 50,361 'RIS-20 unique NFRs' (Fig. 4e) and intersected them with the consensus peaks identified in the N+RIS condition. Of the RIS unique NFRs, 61% (30,775) were repressed by N1ICD (Fig. 5a). HMGA proteins have previously been shown to affect chromatin compaction by competing with Histone-H1 for linker DNA. To determine whether repression of HMGA1 is a mechanism by which N1ICD can repress RIS-unique NFRs, we generated additional ATAC-seq samples from IMR90 ER:HRAS G12 cells expressing a short-hairpin against HMGA1 27 and treated with 4OHT, hereafter referred to as 'RIS+shHMGA1'. By intersecting the 50,361 RIS-unique NFRs with consensus peaks identified in the RIS+shHMGA1 condition, we found that 30% 5 (15,292) were dependent on HMGA1 (Fig. 5a). Of these, 77.6% (11,879) were also repressed by N1ICD. These analyses illustrate that a subset of RIS-unique NFRs can be repressed by N1ICD, possibly through HMGA downregulation. RIS-unique NFRs were significantly more AT-rich than other subsets of NFRs (Fig. 5b), further supporting the involvement of HMGA in formation of RIS-unique NFRs. These data 10 suggest that, in the RIS context, HMGA1 is a key regulator of chromatin structure at the nucleosome and microscopic level.
To validate the above 'binary' approach, where we assume that peaks are either present or absent, we used our normalised coverage files to perform unbiased k-15 means clustering around novel consensus NFRs identified in RIS, NIS or N+RIS conditions (that were detected in growing cells) (Fig. 5c). NFRs separated into 3 clusters: clusters 1 and 2 were dominated by RIS whilst cluster 3 was dominated by NIS. Clusters 1 and 2 appeared to separate largely due to the level of enrichment, where NFRs in cluster 1 were stronger. Strikingly, the signal in clusters 1 and 2 was 20 reduced in the N+RIS and RIS+shHMGA1 conditions when compared to the RIS condition ( Fig. 5c). Notably, whilst peaks in cluster 3 were increased in the N+RIS condition, they did not increase in the RIS+shHMGA1 condition, suggesting an HMGA-independent mechanism of chromatin opening in NIS (Fig. 5c). Therefore, in line with microscopic SAHF structures, N1ICD alters chromatin structure in RIS at the nucleosome scale in part through repressing HMGA expression.

Non-cell autonomous regulation of chromatin by tumour cells
HMGA1 and NOTCH1 are two prominent cancer-associated genes. Both can act as 5 oncogenes or tumour suppressors in a context-dependent manner. We reasoned that the relationship between these two genes might also be important in the tumour microenvironment and asked whether tumour cells expressing JAG1 can affect HMGA expression and chromatin structure in adjacent fibroblasts.

10
To answer this question, we used the Cancer Cell Line Encyclopaedia (CCLE) 51 to identify tumour cell lines that express low (MCF7), medium (A549) and high (Hep3B) levels of JAG1, which we confirmed by immunoblotting (Fig. 6a)  To determine whether tumour cell lines can induce NOTCH1 signalling and repress HMGAs non-autonomously, we repeated the co-cultures and isolated the IMR90 ER:HRAS G12V mRFP1 cells using flow cytometry (Fig. 6c). Quantitative RT-PCR showed a dramatic upregulation of the canonical NOTCH1 target gene HEYL and a concurrent downregulation of HMGAs in fibroblasts co-cultured with JAG1expressing tumour cells, particularly A549 and Hep3B cells (Fig 6d). Two other canonical target genes, HES1 and HEY1, were not as dramatically upregulated by JAG1 expressing cell lines (Supplemental Fig. 6c), which highlights the complexity of 5 the NOTCH pathway but may also suggest that HEYL plays an important role in this interaction.
Finally, we asked whether tumour cell lines could repress 'RIS-unique NFR' formation in fibroblast cells, as was the case for ectopic N1ICD (Fig. 5a). We isolated 10 4OHT-induced IMR90 ER:HRAS G12V mRFP1 cells from co-culture with tumour cell lines by flow cytometry and performed ATAC-seq using these cells (Fig. 6c). By performing intersections between RIS-unique NFRs (50,361) (Fig. 4e) and consensus peak sets from red RIS cells previously co-cultured with tumour cell lines, we found that co-culture with MCF7, A549, or Hep3B cells repressed 57% (28,635), 15 67% (33,532) or 91% (45,840) of RIS-unique NFRs respectively (Fig. 6e, f). These data correlated well with the ability of the tumour cell lines to repress SAHFs in adjacent IMR90 (Fig. 6b). NFRs repressed by co-culture with the tumour cell lines overlapped well with each other and with the 15,292 'HMGA1-dependent' NFRs (Supplemental Fig. 6d, e). These data suggest that tumour cells expressing JAG1 20 can dramatically alter the chromatin landscape of adjacent stromal cells in a juxtacrine manner, particularly in the senescence context where HMGA proteins are highly expressed.

HEY-L and HMGA1 anti-correlate in multiple tumour types
If NOTCH1 signalling inhibits HMGA1 in vivo, we would expect an anti-correlation between NOTCH1 activity and HMGA1 expression in human tumour samples. To test whether this is the case, we first performed a pan-tissue type analysis using microarray data from the R2 database (http://r2.amc.nl) by comparing the expression 5 of HMGA1 and canonical NOTCH1 target genes (Supplemental Fig. 7a, b, c). When Z-score expression values were compared across all the datasets in R2, we observed a strong negative correlation between HMGA1 and HEY-L (R=-0.356, p < 0.0001) and between HMGA1 and HEY1 (R=-0.281, p <0.0001), but no correlation between HMGA1 and HES1 (supplemental Fig. 7a, b, c). This result was particularly 10 interesting, because HEY-L was also the most dramatically upregulated gene in IMR90 fibroblasts co-cultured with JAG1-expressing RPE1 cells (~3 fold) (Fig. 3g) and tumour cell lines (~20 fold) (Fig. 6d). Using the web-based tool KM plotter 52 , which incorporates microarray data from 5,143 breast 53 and 2,437 lung cancer 54 patients, we found that patients with low HMGA1 or high HEY-L have a better 15 prognosis than patients with the inverse expression pattern in lung and breast cancer (Supplemental Fig. 8a, b), suggesting that the relationship between these proteins may have prognostic value.
As microarray data can be dependent on the quality of the probe used, we analysed 20 the expression of HMGA1 and HEY-L using RNA-seq data generated by the TCGA Research Network 55 (http://cancergenome.nih.gov). Consistently, there was a significant negative correlation between these two genes in the majority of tumour types analysed (Fig. 7a). There was a particularly strong anti-correlation in Lung Squamous Cell (SCC) (Fig. 7b) (r=0.4842; p= < 0.0001). Moreover, when further categorised based on expression into 'HEY-L high -HMGA1 low' and 'HEY-L low -HMGA1 high' tumours we found that patients in the former category had a significantly better prognosis for Lung SCC (Fig. 7c) (p=0.00316). These data demonstrate that an anti-correlation between HMGA1 and NOTCH1 activity is 5 evident in cancer and that this correlation can be prognostic of patient outcome.

DISCUSSION
In the current study, we provide evidence for NOTCH-mediated, contact-dependent 'lateral modulation' of chromatin structure at both the microscopic and nucleosome which can provide a growth advantage 47 . Thus, our data raise an interesting possibility that HMGA1 can drive pluripotency and cancer 36,37 in part by modulating chromatin accessibility. It will be important to understand how HMGA1 facilitates both chromatin 'opening' at the nucleosome scale and the formation of SAHFs, and to determine whether the two are related. We wonder whether RIS-unique NFRs, which 5 are mostly HMGA1-dependent and gene distal, could have structural rather than regulatory functionality.
Chromatin accessibility was also increased in NIS cells although these are often at

Flow cytometry
Analysis of ectopic JAG1-mVenus expression was conducted by flow cytometry as previously described 15 . Cells were stained with anti-JAG1-APC (FAB1726A, R&D systems) or isotype control antibody before analysis on a FACSCalibur flow cytometer (Becton Dickenson). Flow data were further analysed using FlowJo v10.

Fluorescence microscopy
Analysis was performed as previously described 27

ATAC-seq
ATAC-seq samples were generated as previously described 46 using 100,000 IMR90 cells and 13-cycles of PCR amplification. Samples were size selected between 170 and 400bp (in order to isolate 'nucleosome free' and 'mono-nucleosome' fragments) 15 using SPRIselect beads (B23319, Beckman Coulter) before single-end sequencing to generate 75-bp reads on the NextSeq-500 platform (Illumina).

ChIP-seq
Chromatin immunoprecipitation (ChIP) was performed as previously described using 20 20 µg of sonicated chromatin 76 from Growing and RIS IMR90 ER:HRAS G12V cells and 5µg of anti-H3K27Ac antibody (Clone CMA309 77 ) and H3K4me1 antibody (Clone CMA302 77 ). Libraries were prepared using the NEBNext Ultra II DNA Library Prep Kit for Illumina (37645, New England Biolabs) according to manufacturers instructions except that size selection was performed after PCR amplification using SPRIselect beads (B23319, Beckman Coulter). Samples were sequenced single-end using 50-bp reads on the HiSeq-2500 platform (Illumina).

RNA-seq 5
RNA-seq data was generated from IMR90 ER:HRAS G12V cells expressing a shorthairpin targeting the 3' UTR of human HMGA1 (RIS+shHMGA1). RNA was purified as above and quality checked using the Bioanalyser eukaryotic total RNA nano series II chip (Agilent). mRNA-seq libraries were prepared from 6 biological replicates of each condition using the TruSeq Stranded mRNA Library Prep Kit 10 (Illumina) according to manufacturers instructions and sequenced using the HiSeq-2500 platform (Illumina).

RNA-seq analysis 15
Reads were mapped to the Human reference genome hg19 with the STAR (version 2.5.0b) aligner 78 . Low quality reads (mapping quality less than 20) as well as known adapter contamination were filtered out using Cutadapt (version 1.

ChIP-seq and ATAC-seq analysis
ChIP-seq and ATAC-seq reads were mapped to the human reference (hg19) with BWA (version 0.7.12) 81 . A similar filtering was carried out using Cutadapt as described for RNA-seq data, and reads falling into the 'blacklisted' regions identified 5 by ENCODE 82 were further removed. Average fragment size was determined using the ChIPQC Bioconductor package 83 , and peak calling was performed with MACS2 (version 2.1.0) 84 , using fragment size as an extension size (--extsize parameter).
High confidence peak sets for each condition were identified separately, using only those peak regions that were present in at least two replicates.  Table Browser. The enhancers were identified based on our own H3K4me1 and H3K27ac histone mark ChIP-seq data sets; all regions that had peaks in both of these marks in either growing or RIS cells were considered as enhancers.

Intersecting consensus peaks and generating Venn diagrams
The Homer (v3.12) 87 command 'mergePeaks' with default settings and the output options '-venn' and '-prefix' were used to generate values for plotting Venn diagrams and associated bed files for further analysis. Only literal overlaps (overlapping by 1bp) were considered. Venn diagrams were plotted using the R package 'Venneuler' (https://cran.rproject.org/web/packages/venneuler/index.html).

Calculating distances to Transcriptional Start Sites and GC percentages of peaks.
To calculate the distance of consensus peaks from Transcriptional Start Sites (TSSs) 5 and GC percentage of consensus peaks, the Homer (v3.12) 87 command 'annotatePeaks.pl' was used with default settings and the output option '-CpG'.

Gene ontology analysis
Consensus peaks within 300bp of a gene TSS were identified as described above. 10 Gene Ontology (GO) analysis was performed using the GO Biological Process 2015 annotation provided on the web-tool 'Enrichr' 88,89 (http://amp.pharm.mssm.edu/Enrichr/).

Generation of normalised coverage files 15
A previously described trimmed-mean approach was used to generate scaling factors for each ATAC-seq condition relative to others 47 . Briefly, we reasoned that the enrichment of reads within ATAC-seq peaks containing TSSs of genes that are both expressed (logCPM >mean logCPM) and have low variance between conditions (-0.14 < logFC < 0.14) by RNA-seq should not vary, unless there are differences in 20 ATAC-seq sample quality, preparation or sequencing. By reanalysing our previously published IMR90 RNA-seq data 15 together with newly generated RNA-seq samples for RIS+shHMGA1 cells we identified 589 genes that fit these criteria. We counted the reads from the ATAC-seq samples that map to these specific genes using Rsubread, and computed scaling factors based on the mean counts for each condition separately. Normalised coverage files (bigWig) were generated by pooling reads from all of the replicates and applying the calculated scaling factors using the 'genomecov' function in bedtools, sorting the resulting normalised bedGraph files and then converting them to bigWigs using the 'bedGraphToBigWig' function from 5

Generation of clustered heatmaps
Heatmaps were generated using normalised coverage of peaks (+/-1kb) representing novel NFRs (consensus peaks detected in any sample excluding peaks 10 detected in the growing IMR90 sample) with k-means clustering using the deepTools package 90 .

TCGA analysis
We analysed expression levels of NOTCH-associated genes in the publically 15 available RNA sequencing data generated by the TCGA Research Network: http://cancergenome.nih.gov/ 55 . Computational analysis and statistical testing of the NGS data was conducted using the R statistical programming language 91 . Filtered and log 2 normalised RNA expression data along with all available clinical data were downloaded from the GDAC firehose database (run: stddata__2015_06_01) for each 20 gene of interest from the relevant cancer-specific collections.
Correlation testing for associations between expressed genes was performed using the cor.test function in R to calculate the Pearson's product moment correlation coefficient and test for significant deviation from no correlation. Plotting of TCGA data was performed using the ggplot2 R package 92

COMPETING FINANCIAL INTERESTS 20
The authors declare no competing financial interests.               ER:HRAS G12V mRFP1 cells cultured with tumour cell lines +4OHT for 6 days relative to cells cultured alone (as described in Fig. 6c). (b, c) n = 3 biologically independent replicates. Statistical significance calculated using one-way ANOVA with Tukey's correction for multiple comparisons; *P ≤ 0.05, **P ≤ 0.01, ***P ≤ 0.001. (d, e) Venn-diagrams showing literal overlap between NFRs identified in the subsets indicated. Overlaps are between 'RIS specific NFRs' (defined in Fig 3e) that are not detected in the RIS+MCF7, RIS+A549 and RIS+Hep3B ATAC-seq samples (d) or between these and 'HMGA1 dependent NFRs' defined in Fig. 5a (e).