To investigate whether breast cancer spatial genome organization showed rearrangements specific to a distant organ metastatic microenvironment, we needed to assemble genome spatial compartment data from different types of normal and cancerous cells for both breast and a secondary organ. Since most breast cancer subtypes show a preference for lung metastasis, we first focused on this metastatic site (Bennett et al., 2022). We selected a cohort of breast and lung cell lines, both normal and cancerous, for which publicly available genome organization (Hi-C) is available (Supplementary file 1). A total of 15 cell lines with available data, including primary cells and cancer cell lines from both organs, were selected. A classification of the cells used in this study based on their malignancy status and other molecular signatures is shown in Figure 1a. Briefly, for the breast cancer system, we considered both HMEC and MCF10A as normal (non-tumorigenic) breast epithelial cells and placed them at the starting node of the model. Cancer cell lines were then divided into three subcategories based on breast cancer molecular signatures: luminal (ZR751, BT474, ZR7530, MCF7, and T47D), HER2-enriched (HCC1954 and SKBR3), and triple-negative (HCC70, BT549, and MDA-MB-231). We then assigned cell lines to ‘localized’ or ‘metastatic’ status based on the site from which they were originally collected in the patient (see details in Supplementary file 1; Engel and Young, 1978; Gazdar et al., 1998; Wang et al., 2021; Stone et al., 1978). Specifically, within the luminal subtype, ZR751 and BT474 were designated as localized cancers, and the remaining three cell lines, ZR7530, MCF7, and T47D, were used as models for ascites, lung, and lung metastatic breast cancers, respectively. Similarly, for the HER2-enriched subcategory, HCC1954 and SKBR3 were considered localized and lung metastatic cancers, respectively. Finally, in the case of the triple-negative subtype, HCC70 and BT549 were used to model localized cancer whereas MDA-MB-231 represents lung metastatic cancer. For the lung portion of the system, HTBE primary cells represent normal lung epithelium, and both the A549 and H460 cell lines represent localized lung cancer. The valuable MetMap study in mouse has measured the organs that can be colonized by human-derived cancer cell lines when they are injected into mice (Jin et al., 2020). The cell lines we consider here may have different or additional ‘potential’ metastatic sites in this type of mouse model. But, it is known that species differences can lead to divergent organotropism in mouse models as compared to human patients (San Martin et al., 2017; Simons et al., 2019). Therefore, we use the site of collection as compared to the site of origin in the human patient as our reference point for primary and secondary site organs.

We compared the supervised ordering of the cell lines described above with the ordering derived from unsupervised analysis of genome-wide A/B compartment organization data. Compartment identity strengths were calculated using eigenvector decomposition on the 250 kb binned Hi-C data for each cell line (Methods). In this analysis, a positive value signifies a euchromatic A compartment region, while a negative value represents a heterochromatic B compartment association.

When we applied hierarchical clustering on the genome-wide compartment data, we observed a grouping of cells that was congruent with the cell ordering constructed based on molecular and clinical features (Figure 1b). For example, all the normal cell types, including both the breast and lung systems, were clustered together. Similarly, cells from the luminal subtype and the lung carcinoma cells were grouped in their respective clusters. However, we also observed a few interesting exceptions from the clustering result. MDA-MB-231, a breast cancer cell line isolated from lung pleural effusion, was grouped with the lung carcinoma cell lines, suggesting its compartment organization shares more features with lung cancer than do other breast cancer cell lines. We also found that HCC70 and HCC1954, which were both derived from localized cancer, clustered relatively close to the normal epithelial cells. Collectively, this unsupervised analysis shows that the different subtypes of cancer exhibit distinctive 3D genome organization features at the compartment level (Figure 1b), as previous studies suggested (Kim et al., 2022).

Performing principal component analysis (PCA) on the genome-wide compartment data revealed similar trends (Figure 1c). Interestingly, the ordering of cell types along the first principal component (PC1) captured a transition between epithelial and mesenchymal-like cells. Cells characterized as highly epithelial were found on the left extreme of the PC1 axis, and as we move toward the right along PC1, the cells acquire a mesenchymal phenotype. For example, MDA-MB-231 cells are derived from healthy cells of epithelial origin but have undergone an EMT and show a mesenchymal-like phenotype (Charafe-Jauffret et al., 2006; Yin, 2011). In our Hi-C principal components analysis, this cell line has a highly positive PC1 value. In contrast, MCF7 cells maintain a more epithelial morphology and are found at the opposite end of the PC1 spectrum from MDA-MB-231. The organization of cell lines along PC1 from spatial compartmentalization matches quite well with the previously characterized epithelial–mesenchymal classification of these cell lines based on transcriptomic data (Figure 2—figure supplement 1; Le et al., 2018; Tan et al., 2014). We noticed that the PC2 axis tends to segregate cell lines according to the presence or absence of HER2. As expected for an unsupervised analysis like PCA in a complex system, the separation is not perfect, as HER2-negative cell lines like MCF7 are plotted near HER2-positive cells. But the trend of separation suggests that in breast cancer cell lines, the presence or absence of this receptor and its downstream signaling network has important effects on the genome organization at the compartment scale. To further characterize the genomic regions that showed distinctive compartment signatures across cell lines, we extracted lists of all genes that fell into genomic bins with the 200 most extreme eigenvector loadings for PC1: 100 strongest positive and 100 strongest negative bins. Gene ontology enrichment results for these gene lists coincide with the idea that PC1 represents the EMT axis: PC1 positive genes have functions related to cell adhesion (tenascin C, JAG1, ACTA2), extracellular matrix organization (APP, LAMA2, FGF2, BMP4) and positive regulation of cell migration, consistent with a mesenchymal phenotype, while PC1 negative genes show GO terms related to epithelial maturation (Figure 1d) Overall, our analyses show that the genome organization is not only able to distinguish different subtypes of cancers but also exhibits changes along the EMT.

We next analyzed the gene expression signatures that emerged across different subtypes of breast cancer (data sources described in Supplementary file 1) and compared those with the patterns obtained from the compartmental analysis. We first performed PCA on the transcriptome profile of all the breast and lung system cells considered in this study (Figure 2a). From the low-dimensional PC projection result, we observed that similar breast cancer subtype cells distribute in the PC space based on their transcriptome profile. As in the compartment signature, MDA-MB-231 cells localize close to lung cancer cells A549 and H460, suggesting that this lung metastatic cell line has transcriptional similarities to lung cancer. We also noticed that along the largest PC axis (PC1), cells are ordered based on their EMT state. Performing GO term enrichment analysis using the genes from the regions with the top 100 most positive and 100 most negative eigenvector loadings for this PC1, we find that genes associated with PC1 positive values relate to epithelial organization, differentiation, and negative regulation of cell migration, denoting a more epithelial phenotype (Figure 2b). In contrast, PC1 negative genes relate to the mesenchymal functions with enriched terms for wound healing, hemostasis, cell movement, and ECM remodeling (Figure 2b). Along with the unsupervised PCA, we also performed gene expression analysis in a supervised fashion using two different methods – gene set variation analysis (GSVA) and non-negative principal component analysis (nnPCA) (Figure 2c, d, see Methods).

Briefly, for each cell type, an E (epithelial) score and an M (mesenchymal) score were calculated based on the expression levels of a curated set of 232 epithelial and 193 mesenchymal genes (Tan et al., 2014; Panchy et al., 2022). A higher E score represents more epithelial characteristics, whereas a higher M score corresponds to a mesenchymal phenotype. The GSVA and nnPCA transcriptome analyses recapitulate a similar ordering of cell lines according to epithelial–mesenchymal characteristics.

Numerous studies have already demonstrated that cancer cells segregate along an EMT axis by transcriptomics (Le et al., 2018; Tan et al., 2014; Panchy et al., 2022; Jung et al., 2020; Cook and Vanderhyden, 2022; Nguyen et al., 2018). Given that spatial compartmentalization is often correlated with gene expression (Li et al., 2024), a simple possible explanation for the EMT segregation of cells by compartment signatures could be that genes which change expression also change compartment. An alternative hypothesis would be that transcription and 3D genome organization change represent two independent aspects of the EMT transition. To evaluate these possibilities, we compared the genes obtained from the compartmental PCA analysis, the transcriptome PCA analysis, and the curated epithelial–mesenchymal gene set (Figure 3a, gene lists found in Supplementary file 2). We found that the genes in regions that change spatial compartmentalization have very little overlap with the sets of genes that most varied in transcription across these cell lines or were curated as differently expressed EMT genes (Figure 3a). To further investigate this result, we examined the transcription status of genes that changed compartment across the EMT spectrum and, conversely, the compartment status of genes that changed transcription (Figure 3b–d). As expected, hierarchical clustering of the detected highly varying compartments (Figure 3b, top) and highly varying transcripts (Figure 3c, d, bottom) resulted in clear segregation of the cell types into similar groupings (i.e., HMEC and MCF10A together, BT474 and SKBR3 together with an opposing pattern). But the genes which strongly changed compartments between cell lines showed no consistent patterns of correlated expression change (Figure 3b, bottom). Likewise, very few compartment differences existed across cell lines at genes whose transcription segregated cells by EMT status (Figure 3c, d, top). Notably, for both the transcriptome PCA analysis and the curated EMT gene set, the majority of the differentially expressed genes are found in the active A compartment. Overall, our results favor the hypothesis that compartmentalization and gene expression signatures are capturing two independent factors in EMT transition. The genes that differ most in expression across these cell lines are largely already accessible (A compartment) and therefore their expression is likely not controlled by spatial compartmentalization. Meanwhile, the additional EMT relevant genes that change compartment, but not transcription, may be now spatially poised for future gene regulation given additional environmental cues, as has been shown in other contexts (Das et al., 2023; Stik et al., 2020; Leemans et al., 2019; Honda et al., 2022). Overall, the set of genes that changes compartments does not have as clear epithelial and mesenchymal functional enrichment as the transcription change set of genes (as shown by the analysis of these gene sets presented in Figures 1d and 2b). This could indicate that some of the compartment changes that occur with EMT are not directly gene regulatory but rather enable an overall conformational change of the chromatin that is needed for the alterations in physical cell state or to accomplish long-distance gene regulation changes. Indeed, we observe specific cases where a compartment change region is located near but not at a gene expression change region (Figure 2—figure supplement 1c). For example, TFF3 is in the set of genes that decreases in expression in M vs. E type breast cancer cell lines. While the gene itself remains in the A compartment in most all cell types, it lies near several regions that switch from A to B in M vs. E type cancer cell lines. Therefore, some of the discordant gene expression and compartment change regions may reflect the impact of compartment shifts on gene expression in neighboring regions. Notably, this same genomic region on chr21 showed striking compartment switches with cancer progression in prostate cancer as well (San Martin et al., 2022).

To survive a metastatic environment, a cancer cell must express a repertoire of binding factors, extracellular matrix proteins, and ligands that positively interact with the secondary tissue microenvironment and foster colonization and growth. We hypothesized that for that to happen, the genome organization of the metastatic cell should change to resemble that of the cells at the secondary site, as a prelude to transcriptional activation. To evaluate this hypothesis, we compared the 3D genome organization of the metastatic and localized breast cancer cells with lung epithelial vs. lung cancer cells (Figure 4a, b). As a baseline, we first compared the breast cancer compartment identity data with that of the normal breast epithelial cells. For each comparison, this resulted in four possible compartment identity combinations: AA (A compartment in both normal breast and breast cancer), AB (A compartment in normal breast and B compartment in breast cancer), BA (B compartment in normal breast and A compartment in breast cancer), and BB (B compartment in both normal breast and breast cancer). Similarly, the compartment identity of lung cancer cell lines was also compared with that of the normal lung cell. We then performed a cross-comparison between the breast and lung cancer comparisons, leading to a total of 16 compartment identity combinations (Figure 4b). Of these, we were particularly interested in cases where the compartment state of healthy breast epithelial cells consistently differed from lung. It is notable that this set is relatively small to begin with (on the order of 200 250-kb bins), which emphasizes the overall similarities between epithelial cell types, whether they come from breast and lung. But for these regions that differ between healthy breast and lung, we next asked whether the breast cancer compartment would retain its breast identity or switch to match lung. In all the breast cancer cell lines derived from lung metastases, we found that more than half of these genomic regions switched compartments to match the lung compartment state (Figure 3—figure supplement 1). We term these switches that make the breast cancer cell compartment identity more like the lung as ‘lung-permissive genome organization changes’. They include: AB_BB (genomic regions which are in the B compartment in both normal lung and localized lung cancer, and which change from A to B in the transition from normal breast to breast cancer) and BA_AA (genomic regions which switch from B to A in breast cancer compared to normal breast epithelium, and which are in the A compartment in both normal lung and lung cancer). These combinations allow us to distinguish changes that could foster lung-specific survival changes from cancer-specific traits such as AB_AB and BA_BA (compartments that change in the same direction from normal to cancer cells in both breast and lung). We then evaluated whether lung-permissive 3D genome changes occurred more frequently in lung metastatic breast cancer than in localized breast cancer.

To quantify the frequency of lung-permissive changes overall, we calculated what fraction of the compartment alterations in breast cancer vs. normal breast tissue were also counted as lung permissive. We divided cells according to breast cancer subtype for this comparison. For the subtypes where several cell lines were available, we considered compartment identity combinations that were consistent between members of the group. We found that for luminal and triple-negative subtypes, metastatic breast cancer exhibits a higher level of lung-permissive changes than localized cancer (Figure 4c).

After selecting the genomic regions that exhibit lung-permissive changes in metastatic breast cancer cell lines, we extracted the underlying genes and used the GEPIA2 tool (Tang et al., 2017) to analyze the expression of these gene sets in TCGA breast and lung cancer patient data (Figure 4d–f; left panels). We found that in general, the gene expression levels change concordantly with the compartment. The results show that for the regions exhibiting lung-permissive changes, the switch in compartmental identity is associated with the gene expression change from normal to tumor. That is, for BA_AA regions, the breast tumor expression of these genes is higher than normal breast tissue expression, coinciding with higher expression in the lung (Figure 4d). Conversely, for AB_BB regions, the breast tumor expression of genes in these regions is lower than for normal tissue, corresponding to lower expression in the lung (Figure 4e). Notably, these lung-permissive compartment change genomic regions do not strongly overlap with the key compartment switch regions we identified by principal components analysis (Table 1). Correspondingly, the enriched functions of genes in these lung-permissive switch regions (Figure 4d–f) are different than the EMT-related functions found among the genes identified by PC1.

To verify that ‘lung-permissive changes’ truly occurred more than expected in lung metastatic cells, we used a different organ system – brain metastasis and brain cancer – as a negative control. Breast cancer can metastasize to the brain (Bachmann et al., 2015; Yoneda et al., 2001), but the specific metastatic cell lines for which we have data were not derived from brain metastases. Therefore, we would expect fewer ‘brain-like’ compartment changes in metastatic breast cancer that metastasizes to lung. Before comparing percent permissive changes between organs, we developed an adjusted version of the calculation to account for differing baseline probabilities of compartment change at the distant organ site (Figure 4—figure supplement 1 and Methods). The adjusted calculation still results in a higher level of lung-permissive changes in lung metastatic breast cancer cell lines as compared to localized (Figure 4—figure supplement 2d). For the brain comparison, we used the NHA cell line (normal human glia) and the SF9427 cell line (glioblastoma) as our reference models (Figure 4—figure supplement 2a, b). We then calculated the adjusted level of brain-permissive changes shown by localized and metastatic breast cancers (Figure 4—figure supplement 2c). We found that the metastatic breast cancer cells used in this study showed lung-permissive changes more frequently than brain-permissive compartment changes (Figure 4—figure supplement 2c, d). This is congruent with the fact that these metastatic breast cancer cell lines were all isolated from lung pleural effusions. We note that Hi-C data for brain metastatic breast cancer cell lines is not currently available for comparison.

The analysis above suggests that lung metastatic breast cancer cell lines show more lung-permissive than brain-permissive genome organization changes. Next, we asked whether a cancer type that metastasizes to the brain exhibits brain-permissive compartment changes. For this purpose, we chose prostate cancer cell lines derived from localized (lymph node) and brain metastatic sites (LNCaP and DU145) and compared their genome architecture to the non-tumorigenic prostate epithelial cell line RWPE (Figure 4—figure supplement 3a, b). We found that the brain metastatic cells showed more brain-permissive changes than the one derived from a lymph node (Figure 4—figure supplement 3c). Also, particularly in the B to A direction, there were greater brain-permissive changes in this prostate cancer system (Figure 4—figure supplement 3c) than in the breast cancer system (Figure 4—figure supplement 3c). This confirms that a cancer type that metastasized to the brain matched the brain genome compartmentalization profile more often than a cancer type that had not metastasized to the brain. We found that the genes in the switched compartment regions have functions that could relate to leaving the prostate epithelium and adapting to the brain environment (Figure 4—figure supplement 3d). For example, the DU145 (brain metastatic) brain-permissive B to A switched genes were enriched for functions in regulation of synapse organization (including calcium channel CACNA2D3). Conversely, the regions that switch from A to B to match the brain B compartment and are enriched for keratinization and cell adhesion, suggesting a downregulation of prostatic epithelial properties.

To process Hi-C sequencing reads, we used the HiC-Pro (https://github.com/nservant/HiC-Pro; Servant, 2022) pipeline with default parameters (Servant et al., 2015). The reads were aligned to the hg19 reference genome. Also, for each cell type, all the available reads from all replicates were merged together to improve the quality of contact data for further downstream processing. After obtaining the valid pair interactions, we binned the interactions at 250 kb resolution for each individual chromosome and normalized the matrices using the ICE technique (Imakaev et al., 2012). Following that, the number of normalized contacts of each chromosome was scaled to 1M using ‘scaleMatrix.pl’ script from the cworld-dekker (https://github.com/dekkerlab/cworld-dekker; Lajoie, 2019) tool suite. And finally, in order to get the compartmental organization for each individual chromosome, principal components analysis was performed on the normalized and scaled matrices using the ‘matrix2compartment.pl’ script from the cworld-dekker tool suite. The compartment analysis assigns A/B compartment identity to genomic regions based on the sign of the elements of the eigenvector (A-positive and B-negative).

The RNA-seq files were processed as per the steps described in the previous publication (Das et al., 2023). Briefly, the BBDuk (https://github.com/kbaseapps/BBTools; Riehl et al., 2025) tool was used first for adapter and quality trimming (Bushnell et al., 2017). After that, the reads were mapped to hg19 human reference genome using STAR aligner (RRID:SCR_004463; Dobin et al., 2013). Once aligned reads were generated, HTSeq-Counts (https://github.com/simon-anders/htseq; Anders and Zanini, 2020) tool was used to perform gene level feature count (Anders et al., 2015). As the RNA-seq files were obtained from various sources and generated using different experimental techniques, they suffer from the batch effect which masks the biological signal. To remove the batch effect, ComBat-seq (https://github.com/zhangyuqing/ComBat-seq; Zhang, 2020) tool was applied on the raw gene counts data (Zhang et al., 2020).

To cluster different cell types based on their compartmental organization, hierarchical clustering was performed on the genome-wide compartment identity data. First, the pairwise similarity score between all the cells was calculated using Pearson’s correlation. The similarity data was then further converted to distance data by subtracting from 1. And finally, hierarchical clustering was performed on the distance data using ‘ward’ linkage.

GSVA was performed to compute the enrichment scores for individual cell lines using RNA-seq data (Deshmukh et al., 2021; Hänzelmann et al., 2013). A list of 232 epithelial signature genes and a list of 193 mesenchymal signature genes were used to compute the E enrichment score (E score) and M enrichment score (M score), respectively (Tan et al., 2014; Panchy et al., 2022; Cursons et al., 2018).

We performed nnPCA using the gene sets mentioned above. nnPCA provided higher resolutions for large number of samples in the EMT spectrum compared to GSVA according to previous studies (Panchy et al., 2022; Panchy et al., 2021). nnPCA determines the approximately orthogonal axes with non-negative coefficients (loadings) for features (genes). Variances of projections of data points (cell lines) onto these axes are maximized via an optimization method (Sigg and Buhmann, 2008).

It is possible that while comparing the host organ-permissive changes across different secondary organs, the compartment switch ratio of different secondary localized cancers might include inherent bias to the comparison. To eliminate the bias, for the secondary organ systems, we first calculate the probability of a genomic region to be in a specific compartment in localized cancer given that the region belongs to the same compartment in normal cell (Figure 4—figure supplement 1). For the A compartment, p(TA|NA) represents the probability of a region in the A compartment in localized cancer cells given that the region also belongs to the A compartment in normal cells. Similarly, p(TB|NB) for the B compartment. If we have more than one cell line representing the secondary site localized cancer, the minimum probability across all the cells is considered. Once we have those, the probabilities are subtracted from one and the resultant values are multiplied with corresponding secondary organ-specific percentages. For example, (1 − p(TA|NA)) is multiplied to the percentage of BA_AA and (1 − p(TB|NB)) to the AB_BB.