Single-cell RNA-sequencing of migratory breast cancer cells: discovering genes associated with cancer metastasis

Abstract

Considerable evidence suggests breast cancer metastasis arises from cells undergoing epithelial-to-mesenchymal-transition (EMT) and cancer stem-like cells (CSCs). Using a microfluidic device that enriches migratory breast cancer cells with enhanced capacity for tumor formation and metastasis, we identified genes differentially expressed in migratory cells by high-throughput single-cell RNA-sequencing. Migratory cells exhibited overall signatures of EMT and CSCs with variable expression of marker genes, and they retained expression profiles of EMT over time. With single-cell resolution, we discovered intermediate EMT states and distinct epithelial and mesenchymal sub-populations of migratory cells, indicating breast cancer cells can migrate rapidly while retaining an epithelial state. Migratory cells showed differential profiles for regulators of oxidative stress, mitochondrial morphology, and the proteasome, revealing potential vulnerabilities and unexpected consequences of drugs. We also identified novel genes correlated with cell migration and outcomes in breast cancer as potential prognostic biomarkers and therapeutic targets to block migratory cells in metastasis.

Graphical Abstract

Single-cell RNA-sequencing of migratory breast cancer cells: discovering genes associated with cancer metastasis

Introduction

Metastasis remains the major cause of death for patients with breast cancer. Despite advances in early detection and treatment, breast cancer remains incurable once metastases develop^1,2. Metastasis requires dissemination of cancer cells from a primary tumor; survival in the circulation; and proliferation of malignant cells in the secondary sites. In this complicated multi-step process, local migration and intravasation of malignant cells into vasculature are key initial steps in dissemination and metastasis. To investigate migration of metastatic cancer cells, considerable efforts have focused on Epithelial-Mesenchymal Transition (EMT), a developmental program in which epithelial cells acquire migratory and invasive phenotypes. While controversy exists about the requirement for transitions between EMT to Mesenchymal-Epithelial Transition (MET) states in metastasis, studies strongly suggest that EMT and cell migration constitute key components of metastasis^3,4. In recent decades, various markers of EMT, including membrane proteins (e.g. E-CAD, N-CAD), cytoskeletal markers (e.g. Vimentin, Cytokeratins), and transcription factors (e.g. Snail, Slug, ZEB1, ZEB2, Twist), have been identified^5–7. The wide range of markers associated with EMT underscores problems of marker-based approaches across multiple cancers: 1) different types of EMT, such as partial EMT, may express different subsets of markers; 2) multiple sets of markers have been used to define EMT even within a single type of cancer; and 3) markers are inconsistent across different malignancies⁵. Inconsistency of existing EMT markers highlights the need for innovative approaches to better identify cancer cells with high potential for metastasis.

To overcome limitations of marker-based systems, we hypothesized that isolating highly migratory cancer cells, a hallmark feature of EMT, would provide a label-free approach to enrich cells with enhanced metastatic potential. We previously developed a microfluidic device placing single cancer cells at the entrance of microchannels to isolate cancer cells that migrated through the channels in response to a chemotactic gradient of serum⁸. Migratory cancer cells enriched by in vitro cell migration assays formed more tumors and metastasized to a significantly greater extent than control non-migratory cells in human tumor xenografts^{8, 9}. By comparing gene expression profiles for populations of migratory versus non-migratory cells, we discovered novel cancer-regulating genes, such as phosphatidylserine decarboxylase (PISD), associated with tumor initiation and metastasis⁸. These results validate our experimental platform to identify candidate regulators of cancer cell migration, tumor initiation, and metastasis.

Since single or small subsets of cells may drive metastasis, we leveraged our microfluidic migration platform and single-cell RNA sequencing to investigate heterogeneity among individual migratory breast cancer cell lines and patient-derived cells^10,11. For all cell populations, we observed gene expression signatures clearly distinguishing between migratory and wild-type cancer cells. While all migratory breast cancer cells consistently showed elevated signatures of EMT and CSCs, we discovered considerable heterogeneity in expression of specific markers of these states. We observed extensive heterogeneity among migratory cells. Some migratory cells expressed both epithelial and mesenchymal markers, consistent with partial EMT states observed recently in circulating tumor cells from patients^12–14. Migratory breast cancer cells showed distinct gene expression profiles for genes related to oxidative stress, mitochondrial morphology, and the proteasome. These unique features of migratory cells point to potential mechanisms regulating EMT and CSCs. Our research also revealed novel genes correlated with breast cancer cell migration and significant differences in outcomes for patients. These genes represent potential new prognostic biomarkers and/or drug targets specifically for migratory cancer cells and breast cancer metastasis.

Results

Microfluidic motility-based migratory cell enrichment for single-cell RNA-sequencing

To isolate migratory breast cancer cells for single-cell transcriptome analysis, we designed a cell migration platform consisting of three main channels (left/center/right) for cell loading/retrieval and thousands of channels for migration-based selection (Fig. 1)⁸. Since the typical diameter of a breast cancer cell exceeds the height of a migration channel, cells are initially positioned at the entrance of migration channels (Supplementary Fig. 1). Loaded cells migrated toward a gradient of serum chemoattractant in the center channel (Supplementary Fig. 1). We defined migratory cells as those reaching the center channel. We adjusted the time allowed for migration from 12-48 hours based on movement of various cancer cells so we could sort the top 0.5-1% (15-30 cells) of migratory cells. After retrieving migratory cells from the center channel by trypsinization, the retrieved cells proliferated for two weeks to expand the migratory cell population to the thousands of cells needed for functional analysis and single-cell sequencing. We repeated motility-based sorting for three breast cancer cell lines (SUM149, SUM159, and MDA-MB-231) and cells adapted to two-dimensional culture from a breast cancer patient derived xenograft (GUM36). Migratory and wild-type populations of all four cell lines were then processed by Hydro-Seq for single-cell transcriptome analysis using an RNA barcoding technique, which allows all mRNAs from one cell to be uniquely labeled and identified after sequencing^10,11. Hydro-Seq microfluidic bead-cell pairing scheme can precisely isolate single cells without contamination from multiple cell capture. In addition to transcriptome analysis, we examined morphology and migration of recovered cells. When plated in low-density, clones from single wild-type cells remained clustered, while migratory cells showed an EMT phenotype of wide dissemination (Fig. 1b and Supplementary Fig. 2). We further performed migration assays to compare expanded migratory and wild-type cancer cells⁸. Migratory cells retained enhanced chemotaxis for at least two weeks after sorting (Fig. 1c–e and Supplementary Fig. 2). The functional studies validated that our migration-based selection enriches for cancer cells with an EMT phenotype.

Figure 1. Selection of migratory breast cancer cells for single-cell RNA-Seq.

(a) Workflow of the presented work. Microfluidic migration was performed to isolate migratory cells. The isolated cells were validated functionally and then processed for single-cell RNA-Seq. (b) The different clonal morphologies of migratory and wild-type cells, including MDA-MB-231, SUM159, and SUM149. The clones (day 3) derived from wild-type cells are connected, while the ones from migratory cells are disseminated. (scale bar: 50 μm) (c-e) Microfluidically sorted migratory cell populations have distinct migratory behaviors as compared to wild-type ones. Graphs show positions of individual cells (after migration of 24 hours) and box plot and whiskers summaries for migration of (c) MDA-MB-231, (d) SUM159 and (e) SUM149 cells toward 10% fetal bovine serum (FBS) media. (n = 600 channels). *** refers to P < 0.001.

Distinct transcriptome profile of migratory cancer cells and pathway analysis to pinpoint critical mechanisms

After validating distinct morphology and migration phenotypes, we performed single-cell RNA-sequencing to identify main contributors to elevated migration of breast cancer cells. We first visualized single-cell data using principal component analysis (PCA) and t-distributed stochastic neighbor embedding (tSNE) clustering methods. We found obvious separation between migratory and wild-type cells for each population, indicating distinct transcriptome profiles (Fig. 2 and Supplementary Fig. 3). The results match well with their different morphological and functional features. We repeated motility-based cell selection using MDA-MB-231 cells, and we found a very similar gene expression profile between migratory cells selected independently (Supplementary Fig. 4). To define key pathways associated with migratory breast cancer cells, we selected significantly altered genes defined by thresholds of fold change and portion of expressing cells for each cell line and applied pathway analysis using the NCI-Nature 2016 pathway database. We observed striking similarities in the top-ranked pathways among all cell lines (Table 1, Supplementary Table 1–4). Among the overlapping pathways, we found known pathways related to EMT, including “HIF-1-alpha transcription factor network” and “TGF-β receptor signaling.” There are pathways related to membrane proteins critical to cell migration, such as syndecan and integrins. Interestingly, we also found cell cycle related pathways including “Validated targets of C-MYC transcriptional activation,” “Aurora signaling,” and “CDC42 signaling events” and other previously reported pathways related to AP-1 transcription factor and PDGFR-beta signaling as top-ranked ones. In addition to pathway analysis of individual cell lines, we also performed pathway analysis using overlapping up-regulated and down-regulated genes from all four cell lines. The analysis generated a similar set of pathways related to EMT and cell cycle (Supplementary Table 5).

Figure 2. tSNE clustering demonstrates the distinct gene expression profiles of migratory and wild-type breast cancer cells.

tSNE plot of single-cell transcriptome analysis for migratory (red color) and wild-type (blue color) cells of four breast cancer cell lines: (a) SUM149, (b) SUM159, (c) MDA-MB-231, and (d) GUM36. X-axis represents tSNE1, and Y-axis represents tSNE2. Each dot represents a cell. Two populations of the same cell line are clearly separated, indicating distinct gene expression profiles.

Table 1.

Top-ranked altered pathways (NCI-Nature pathway database) of migratory population of three breast cancer cell lines (SUM149, SUM159, and MDA-MB-231) and patient-derived cells GUM36.

SUM149	P-value	SUM159	P-value
Validated targets of C-MYC transcriptional activation	3E-08	HIF-1-alpha transcription factor network	2E-05
Beta1 integrin cell surface interactions	1E-06	Validated nuclear estrogen receptor alpha network	6E-05
HIF-1-alpha transcription factor network	1E-06	IL6-mediated signaling events	8E-05
Integrins in angiogenesis	5E-06	Validated targets of C-MYC transcriptional activation	8E-05
Beta3 integrin cell surface interactions	4E-05	Syndecan-4-mediated signaling events	2E-04
Alpha9 beta1 integrin signaling events	8E-05	Aurora B signaling	3E-04
Syndecan-4-mediated signaling events	4E-04	PDGFR-beta signaling pathway	4E-04
Hypoxic and oxygen homeostasis regulation of HIF-1-alpha	5E-04	ErbB1 downstream signaling	5E-04
Arf1 pathway	7E-04	Aurora A signaling	8E-04
VEGFR1 specific signals	9E-04	Syndecan-2-mediated signaling events	1E-03
MDA-MB-231	P-value	GUM36	P-value
Validated targets of C-MYC transcriptional activation	1E-07	Integrins in angiogenesis	2E-06
a6b1 and a6b4 Integrin signaling	2E-05	Validated targets of C-MYC transcriptional activation	2E-06
ErbB1 downstream signaling	4E-04	AP-1 transcription factor network	4E-06
TGF-beta receptor signaling	4E-04	Beta1 integrin cell surface interactions	8E-06
Signaling events mediated by c-Met	6E-04	Arf1 pathway	9E-05
Regulation of nuclear beta catenin signaling	8E-04	HIF-1-alpha transcription factor network	1E-04
CDC42 signaling events	1E-03	C-MYB transcription factor network	2E-04
Class I PI3K signaling events mediated by Akt	2E-03	Beta3 integrin cell surface interactions	2E-04
Integrin-linked kinase signaling	2E-03	Alpha4 beta1 integrin signaling events	2E-04
HIF-1-alpha transcription factor network	2E-03	VEGFR1 specific signals	3E-04

Single-cell RNA-sequencing identifies elevated EMT transcriptome profiles of migratory breast cancer cells

Due to the known association between cell migration and EMT, we first examined established epithelial, mesenchymal, and EMT regulators in migratory cancer cells. A dot plot (Fig. 3a) shows expression of 13 genes as representative epithelial and mesenchymal/EMT markers among 8 cell populations (migratory and wild-type of four cell lines). The single-cell data provide not only the overall average expression by color but also the percentage of single cells expressing that gene by size of dot. Overall, we observed a clear trend that migratory cells expressed more mesenchymal and EMT genes and less epithelial genes as compared to wild-type ones. However, obvious differences exist among cell lines. Expression data from SUM149 cells match well with the canonical EMT model among all presented genes (Fig. 3b, c)^{6, 7}. As a highly mesenchymal-like cancer cell line, MDA-MB-231 cells express little CDH1 and EpCAM. Surprisingly, expression of EMT-related transcription factors (HIF1A, SNAI1, SNAI2, TWIST1, and ZEB2) are also low for both migratory and wild-type populations of MDA-MB-231 cells (Supplementary Fig. 5). As another mesenchymal-like cell line, SUM159 cells also have low expression of all epithelial genes but clear up-regulation of mesenchymal genes in the migratory population (Supplementary Fig. 6). Patient-derived cells GUM36 demonstrate more mixed expression of both epithelial and mesenchymal genes yet clearly up-regulate EMT-related transcription factors in the migratory population (Supplementary Fig. 7). The results suggest that conventional EMT markers may not correlate well with cell migration across all cell types. The inconsistency between cell markers and behavior also highlights the value of microfluidic isolation of migratory cells to identify potential cell-type specific regulators of EMT. Violin plots with statistical test of relevant EMT genes are included in Supplementary Fig. 8–11.

Figure 3. Migratory breast cancer cells express more mesenchymal/EMT related genes and less epithelial related genes than wild-type ones.

(a) Dot plot shows the expression of 13 genes among 8 cell populations (migratory and wild-type of 4 cell lines). Larger dot means higher percentage of single cells expressing that gene, and smaller dot means lower percentage of single cells expressing that gene. Grey color represents the lowest expression, and red color represents the highest expression. The expression is logarithmically normalized. (b, c) Gene expression and clustering of migratory and wild-type SUM149 cells. X-axis represents tSNE1, and Y-axis represents tSNE2. Each dot represents one cell. Red color represents high (90th percentile) expression of a gene, and grey color represents low (10th percentile) expression of a gene. The expression is logarithmically normalized. (b) The expression of epithelial genes: keratin 18 (KRT18), keratin 8 (KRT8), keratin 7 (KRT7), epithelial cell adhesion molecule (EpCAM), and cadherin 1 (CDH1). (c) The expression of mesenchymal and EMT genes: cadherin 2 (CDH2), vimentin (VIM), fibronectin 1 (FN1), integrin beta 1/CD29 (ITGB1), hypoxia inducible factor 1 alpha (HIF1A), transcription factor SOX-4 (SOX4), snail transcriptional repressor 1 (SNAI1), snail transcriptional repressor 2 (SNAI2), Zinc finger E-box-binding homeobox 1 (ZEB1), and Zinc finger E-box-binding homeobox 2 (ZEB2). Migratory SUM149 cells express more mesenchymal and EMT genes as compared to wild-type ones yet exhibit significant cellular heterogeneity.

CSCs and heterogeneous sub-populations in migratory cells based on transcriptome profiles

As we previously observed that sorted migratory cells were more tumorigenic⁸, we further investigated expression of known markers of CSCs (ALDH isoforms, CD44, CD24, CD29, CD90, and CD133) in migratory and wild-type cancer cells (Fig. 4, Supplementary Fig. 8–14)¹⁵. Overall, transcriptome analysis suggests enrichment of CSCs in migratory cells, represented by up-regulated expression of ALDH isoforms, CD44, and ITGB1/CD29. Since heterogeneity may exist within CSCs, single-cell analysis allows precise identification of distinct subsets of these cells rather than merely average expression profiles. In this study, we found enrichment of both EMT-CSCs (CD44^high/CD24^low) and MET-CSCs (ALDH^high) in migratory cell populations (Supplementary Table 6)¹⁶. More interestingly, we found that migratory cells can be further split into two distinct populations by unbiased clustering: epithelial and mesenchymal sub-groups for SUM149 cells (Fig. 5a). These analyses show that conventionally defined epithelial-like (CDH1^high, EPCAM^high, KRT8^high, and KRT18^high) cells can migrate to a comparable extent as mesenchymal-like (CDH2^high, CD44^high, VIM^high) breast cancer cells (Fig. 5c,d). By comparison, pan-ALDH and HIF1A expression levels do not distinguish between the two sub-populations of migratory cells (Fig. 5c,d). When looking at CSCs, EMT-CSCs and MET-CSCs concentrated as expected in the mesenchymal and epithelial sub-groups, respectively (Fig. 5b). Remarkably, we also identified rare cells simultaneously expressing EMT-CSC and MET-CSC markers^{12, 16}. These cells may be endowed with the highest degree of cellular plasticity and metastatic potential (Fig. 5b). The results again demonstrate advantages of single-cell analysis to define heterogeneity within populations of CSCs. Violin plots with statistical test of relevant genes are included in Supplementary Fig. 15.

Figure 4. Migratory breast cancer cells express more CSC related genes than wild-type ones.

(a) Dot plot shows the expression of 13 genes among 8 cell populations (migratory and wild-type of 4 cell lines). Larger dot means higher percentage of single cells expressing that gene, and smaller dot means lower percentage of single cells expressing that gene. Grey color represents the lowest expression, and red color represents the highest expression. The expression is logarithmically normalized. (b-d) Gene expression and clustering of migratory and wild-type SUM149 cells. X-axis represents tSNE1, and Y-axis represents tSNE2. Each dot represents one cell. Red color represents high (90th percentile) expression of a gene, and grey color represents low (10th percentile) expression of a gene. The expression is logarithmically normalized. (b) The expression of MET-CSC genes: pan-aldehyde dehydrogenase isoforms (Pan-ALDH), aldehyde dehydrogenase 1 family member A3 (ALDH1A3), and aldehyde dehydrogenase 1 family member A1 (ALDH1A1). (c) The expression of MET-CSC genes: CD44 and CD24. (d) The expression of stem-cell related genes: CD29/ITGB1, polycomb complex protein BMI-1 (BMI1), thymus cell antigen 1 (THY1), ATP-binding cassette super-family G member 2 (ABCG2/CD338), Wnt family member 5A (WNT5A), pan-NOTCH isoforms (Pan-NOTCH), NOTCH homolog 2 (NOTCH2), transcription factor SOX-9 (SOX9), GATA binding protein 3 (GATA3), and epidermal growth factor receptor (EGFR). Overall, migratory SUM149 cells express more stemness related genes than wild-type ones.

Figure 5. tSNE clustering demonstrates the cellular heterogeneity of migratory SUM149 cells.

(a) tSNE plot of single-cell transcriptome analysis suggests two populations among migratory SUM149 cells: mesenchymal-like (red color) and epithelial-like (blue color) cells. X-axis represents tSNE1, and Y-axis represents tSNE2. Each dot represents a cell. Two populations are clearly separated, indicating distinct gene expression profiles. (b) EMT-like CSCs (CD44 high), MET-like CSCs (ALDH1A3 high), and dual positive CSCs (both CD44 and ALDH1A3 high) on the tSNE plot. epithelial-like CSCs are generally distributed in the epithelial cluster, and mesenchymal-like CSCs are distributed in the mesenchymal cluster. (c, d) Gene expression and clustering of migratory SUM149 cells. Each dot represents one cell. Red color represents high (90th percentile) expression of a gene, and grey color represents low (10th percentile) expression of a gene. The expression is logarithmically normalized. (c) The expression of epithelial/MET-CSC genes: Pan-ALDH, ALDH1A1, ALDH1A3, CDH1, EPCAM, KRT7, KRT8, KRT18. (d) The expression of mesenchymal/EMT-CSC genes: CD44, CD24, VIM, CDH2, FN1, ITGB1, HIF1A, Enhancer of Zeste Homolog 2 (EZH2), SNAI1, SNAI2, ZEB1, and ZEB2.

Migratory cells exhibit distinct profiles for anti-oxidant defense, mitochondrial morphology, and protein degradation

Beyond markers of EMT and MET states, CSCs have been characterized based on functions related to oxidative stress, mitochondria, and the proteasome. First, transcriptome data indicate migratory cells possess greater capacity to counter reactive oxygen species (ROS), as evidenced by increased expression of nuclear factor (erythroid-derived 2)-like 2, (NFE2L2), microsomal glutathione S-Transferase 3 (MGST3), and glutathione-disulfide reductase (GSR) (Supplementary Fig. 16)¹⁷. Enhanced anti-oxidant defense previously has been associated with CSCs, and cells isolated based on low ROS have enhanced properties of CSCs. Second, expression data suggest greater fragmentation of mitochondria in wild-type versus migratory cells. Migratory cells consistently show reduced expression of mitochondrial fission 1 protein (FIS1) and increased levels of the mitochondrial metalloprotease YME1L1 (Supplementary Fig. 17)¹⁸. Low expression of YME1L1 generates cells with greater fragmentation of mitochondria. While potentially context-dependent, recent studies suggest that mitochondrial fragmentation impairs EMT and stem cell properties^{19, 20}. Finally, migratory cells consistently show down-regulation of proteasome genes as compared to wild-type ones (Supplementary Fig. 18). While the direct relationship between proteasome activity and migration is not clear, we and others have validated low proteasome activity as a marker for CSCs²¹. These results collectively strengthen links between isolation of migratory cancer cells and other functional hallmarks of CSCs.

Identification of novel genes associated with cell migration

In addition to known EMT and CSC transcriptome profiles, our data provide the opportunity to discover new regulators of cell migration as potential prognostic markers and/or drug targets. For each cell line, we compared migratory and wild-type populations to identify significantly altered genes. When analyzing four different breast cancer cell lines altogether, we identified 24 genes up-regulated in migratory cells from all four cell lines, 139 genes up-regulated for three cell lines, and 562 genes up-regulated for two cell lines. Also, we identified 34 genes down-regulated for migratory cells from all four cell lines, 174 genes down-regulated for three cell lines, and 609 genes down-regulated for two cell lines (Supplementary Table 7–9). We used a publicly available database (kmplot.com) to determine correlations with prognosis in breast cancer²². Combining transcriptome analysis of migratory cells and data mining for patient outcomes, we successfully identified top-ranked candidate genes over-expressed in migratory cells that were concordant with significant differences in survival of patients with breast cancer (Fig. 6a and Supplementary Fig. 19). Among the gene targets we identified, some have known association with unfavorable prognosis of breast cancer patients, such as CD46 and Actin Regulator ENAH. Actin Related Protein 2/3 Complex Subunit 5 (ARPC5) and Nuclear Factor, Erythroid 2 Like 2 (NFE2L2) were identified as oncogenes for lung squamous cell carcinoma but not investigated in breast cancer^23–26. More interestingly, we identified less studied genes including Anillin Actin Binding Protein (ANLN) (Fig. 6b), Small Nuclear RNA Activating Complex Polypeptide 1 (SNAPC1) (Fig. 6c), Poly(A) Binding Protein Cytoplasmic 3 (PABPC3) (Fig. 6d), and Centromere Protein F (CENPF) (Fig. 6e), microsomal glutathione S-transferase 3 (MGST3), and UBX domain protein 4 (UBXN4) as genes with high expression in migratory cells and poor prognosis in breast cancer patients.

Figure 6. Sequencing of migratory breast cancer cells identifies new markers for prognosis prediction, treatment guidance, and mechanistic studies.

(a) Dot plot shows the expression of 10 genes among 8 cell populations (migratory and wild-type of 4 cell lines). Larger dot means higher percentage of single cells expressing that gene, and smaller dot means lower percentage of single cells expressing that gene. Grey color represents the lowest expression, and red color represents the highest expression. The expression is logarithmically normalized. Migratory breast cancer cells express more listed genes as compared to wild-type ones. (b-e) The Kaplan-Meier plot shows that high levels of (b) Anillin Actin Binding Protein (ANLN), (c) Small Nuclear RNA Activating Complex Polypeptide 1 (SNAPC1), (d) Poly(A) Binding Protein Cytoplasmic 3 (PABPC3), and (e) Centromere Protein F (CENPF) expression correlate with reduced relapse-free survival (RFS) in breast cancer.

Migratory cells maintain distinct transcriptome profile with time in cell culture

After demonstrating distinct transcriptome profiles between migratory and wild-type cancer cells, we further examined to what extent those differences can be maintained and inherited. Two weeks after migration-based selection of MDA-MB-231 cells, we performed the first single-cell RNA-sequencing experiment. Then, we continued growing the cell population and sequenced it again four and six weeks after initial isolation. The PCA plot shows that the migratory cells at different time points are all distinct from wild-type cancer cells (Supplementary Fig. 20). More interestingly, three migratory cell populations mixed well on PCA plot, suggesting the unique migratory transcriptome profile can be maintained over time.

Discussion

Considerable evidence suggests that many malignancies are driven by subsets of cells with properties of CSCs. Mechanisms underlying generation and maintenance of CSCs remain incompletely defined, hindering ongoing efforts to successfully target these cells therapeutically. While CSCs can be identified by expression of cell surface markers or enzymatic activity, these methods are limited by phenotypic heterogeneity and plasticity. As an alternative to marker-based approaches, we developed a microfluidic cell migration platform to sort cells based on enhanced migration^{8, 27}. This approach capitalizes on the established link between EMT, a driver of cell migration, and formation of CSCs. We previously demonstrated isolation of migratory cells with our microfluidic platform. The direct functional sorting enriched for metastatic and tumorigenic cells without relying on markers for CSCs. In the current study, we harvested migratory cells from this device to compare them with wild-type ones and then further dissected heterogeneity among migratory cells by single-cell RNA sequencing.

Pathway analysis revealed many pathways related to EMT, including HIF-1α, hypoxia, and TGFβ, among the top-ranked differences between migratory and wild-type cells, validating our experimental approach. Several pathways related to integrins showed preferential activation in migratory cells. These pathways include “Integrins in angiogenesis,” “Beta1 integrin cell surface interactions,” “Beta3 integrin cell surface interactions,” “Alpha4 beta1 integrin signaling events”, “a6b1 and a6b4 Integrin signaling,” and others. These pathways also represent established regulators of EMT. Focusing on specific subunits of integrins, we found consistent overexpression of integrin subunit alpha V (ITGAV) and integrin subunit beta 1 (ITGB1) in migratory populations (Supplementary Fig. 21, 22)^{28, 29}. Differences in expression of these two integrin subunits reinforce associations with CSCs in breast cancer and other malignancies, indicating these proteins may represent therapeutic targets to eliminate the subset of CSCs^{30, 31}. We also observed activation of cell cycle related pathways, especially “C-MYC transcriptional activation” as a top-ranked pathway distinguishing wild-type from migratory cells isolated from all four cell lines. (Supplementary Fig. 23) Though c-Myc is a known oncogenic transcription factor in tumor progression, our data show greater expression of the c-Myc pathway in wild-type cancer cells as compared to migratory ones³². We also examined other cell cycle related pathways (“Aurora signaling,” and “CDC42 signaling events”) and proliferation related genes (Supplementary Fig. 24) in wild-type cells³³. While recognizing prior studies have described different effects of cMyc in cell migration, our data support a dichotomy with wild-type cells in a more MET-like state that favors cell proliferation^{34, 35}. By comparison, EMT-like migratory cells generally reduce expression of proliferative genes to promote movement.

While we could establish a strong correlation between migration and EMT by pathway analysis, we found remarkable differences in established regulators of EMT among all four cell lines. For highly invasive MDA-MB-231 and SUM159 cells, expression of CDH1 and EPCAM is very low in both wild-type and migratory cells, so levels of these epithelial markers cannot distinguish the subpopulation of migratory cells. However, migratory SUM159 cells clearly up-regulate mesenchymal markers including CDH2, VIM, and HIF-1α. Migratory SUM149 demonstrates a canonical profile of EMT with low expression of EPCAM and cytokeratins and high expression of CDH2, VIM, and HIF-1α. The expression profile of migratory GUM36 cells is characterized by high expression of VIM, HIF-1α, and SNAI2. Pronounced differences in gene expression among four breast cancer cell lines highlight existence of multiple mechanisms driving cells to an EMT state. For purposes of isolating EMT cells based on markers, our results highlight limitations of relying on established markers, which likely will identify only a biased, incomplete sub-population of cells.

While migratory cells overall show over-expression of genes related to EMT, we identified sub-populations of migratory SUM149 cells with profiles of epithelial cells and epithelial-like CSCs (MET-CSCs, ALDH1A3^high). We also detected migratory cells with a partial, or hybrid, EMT phenotype. All migratory SUM149 cells have elevated expression of HIF-1α and Pan-ALDH, markers of EMT and MET-CSCs, respectively, reflecting a hybrid state. Further clustering of migratory SUM149 cells revealed discrete sub-populations expressing epithelial (CDH1, EPCAM, and cytokeratins) or mesenchymal (CDH2, VIM, SNAI2, and ZEB1) markers, respectively. Recent studies focused predominantly on circulating tumor cells have documented cells with hybrid EMT states in patients with breast cancer. Hybrid states exist along a continuum as cells transition between full EMT and MET states^12–14, generating extensive heterogeneity among circulating tumor cells. Cancer cells in a hybrid EMT state may define sub-populations of cells with markedly elevated potential for metastasis. A hybrid EMT state favors clustering of circulating tumor cells, reducing anoikis and promoting metastasis. Hybrid EMT states also are associated with drug resistance. These data emphasize that migration-based selection enriches for EMT-CSCs, MET-CSCs, and hybrid states. Co-existence of multiple cell states within the migratory population highlights plasticity of migratory cells and the need for single-cell analysis to detect the full range of cellular heterogeneity.

In addition to cells with components of EMT, our migration-based selection strategy isolated cells expressing known markers of CSCs. This result matches prior studies showing that EMT generates CSCs and our recent publication demonstrating that isolating migratory cells with our device enriches for cancer cells with increased tumor-initiating potential⁸. Using single-cell RNA-Seq, we measured expression of established markers of CSCs. Overall, we observed up-regulation of pan-Aldehyde dehydrogenase isoforms (pan-ALDH), CD44, and IGTB1/CD29 in migratory cells (Fig. 4a)^36–38, indicating that migration-based selection enriches both EMT-CSCs (CD44^high/CD24^low) and MET-CSCs (ALDH^high)¹⁶. We note greater enrichment of EMT-CSCs from SUM149 and GUM36 cells, both of which contain more cells with epithelial properties in the starting populations. We observed low expression for THY1/CD90, prominin-1 (PROM1/CD133), and ATP-binding cassette sub-family G member 2 (ABCG2/CD338), which are other genes commonly associated with CSCs^39–41. Polycomb complex protein BMI-1 (BMI1), a gene more commonly associated with proliferation rather than migration, did not differ between migratory and wild-type cells⁴². Upregulation of NOTCH isoforms (pan-NOTCH) and transcription factor SOX-2 occurred only in migratory GUM36 cells, suggesting that while these signaling pathways are associated with CSCs, they are not required to generate CSCs in all breast cancer cells. For SUM149 cells, we observed co-expression of CD44, IGTB1/CD29, and Wnt family member 5A (WNT5A), while transcription factor SOX-9 (SOX9) and epidermal growth factor receptor (EGFR) segregated with ALDH1A3. The co-expression data again highlight the value of single-cell sequencing in the investigation of complicated CSC systems.

Migratory cells from all tested breast cancer cell lines consistently exhibited reduced expression of genes associated with the proteasome, a large protein complex that degrades intracellular proteins. Low expression of proteasome genes and/or proteolytic function of this organelle has been used to define CSCs in a variety of malignancies, including lung, breast, glioma, and colon ^{43, 44}. In head and neck cancer, low expression of proteasome genes correlates with reduced survival⁴⁵. Reduced proteasome activity has been proposed to promote CSCs through Notch signaling⁴⁴. However, we identified upregulation of the NOTCH pathway only in migratory GUM36 cells, indicating existence of alternative mechanisms linking reduced proteasome activity to a CSC phenotype. A recent study reported that inhibitors of the proteasome drove breast cancer cells to an EMT, CSC state by activating the TGF-β signaling pathway⁴⁶, consistent with our data demonstrating preferential expression of TGF-β pathway genes in migratory cells. The association of reduced proteasome gene expression and function with EMT and CSCs may explain failure of proteasome inhibitors in breast cancer and other malignancies except multiple myeloma.

Our results also demonstrated that migratory breast cancer cells retain distinct expression profiles over at least six weeks in cell culture. We found similar gene expression profiles of migratory MDA-MB-231 cells over time as demonstrated by PCA plot. In addition, migratory populations sequenced at three different time points enriched for 23 common pathways (Supplementary Table 10), including critical pathways associated with EMT and the cell cycle. These data suggest surprising stability in the EMT phenotype in these cells. In the future, we will investigate to what extent heritable changes in gene expression arise from genetic and/or epigenetic regulators in migratory cells^{47, 48}. Overall, this work identifies multiple levels and types of heterogeneity in migratory breast cancer cells; highlights NFE2L2 and other mediators of anti-oxidant defense as therapeutic targets for CSCs; and points to possible new prognostic biomarkers and/or drug targets to improve treatment of breast cancer.

Materials and Methods:

Microfluidic chip design and fabrication

The migration devices were fabricated from a single layer of PDMS (Polydimethlysiloxane, Sylgard 184, Dow Corning), which was fabricated on a silicon substrate by standard soft lithography, and a glass slide. Two masks were used to fabricate the multiple heights for main channels (40 μm height) and the migration channel (5 μm height). One device contains 900 migration channels (450 channels on one side), and the migration channel 30 μm in width, 5 μm height, and 1 mm in length. The PDMS layer was bonded to a glass slide after activation by oxygen plasma treatment (80 Watts, 60 seconds) to form a complete fluidic channel. The microfluidic chips were sterilized by UV radiation prior to use. Before cell loading, collagen (Collagen Type 1, 354236, BD Biosciences) solution (1.45mL Collagen, 0.1mL acetic acid in 50mL DI Water) was flowed through the device for one hour to enhance cell adhesion. Devices were then rinsed with PBS (Gibco 10082) for one hour to remove the residual collagen solution.

Cell culture

We purchased MDA-MB-231 cells from the ATCC and cultured cells in Dulbecco’s Modified Eagle Medium (DMEM, Gibco 11965) supplemented with 10% fetal bovine serum (FBS, Gibco 10082) and 1% Penicillin/Streptomycin (Pen/Strep, Gibco 15070). We cultured SUM149 and SUM159 cells (gift of Dr. Stephen Ethier, Medical University of South Carolina) in F-12 (Gibco 11765) media supplemented with 5% FBS (Gibco 10082), 1% Pen/Strep (Gibco 15070), 1% GlutaMax (Gibco 35050), 1 μg/mL hydrocortisone (Sigma H4001), and 5 μg/mL insulin (Sigma I6634). We maintained GUM36 cells (obtained from Dr. Max Wicha lab, University of Michigan), patient derived cell line (The original PDX was derived from an ER-/PR-/Her2+ breast cancer patient) adapted to standard two-dimensional culture environments, in Dulbecco’s Modified Eagle Medium (DMEM, Gibco 11965) supplemented with 10% FBS (Gibco 10082) and 1% Pen/Strep (Gibco 15070). We maintained all cells at 37°C in a humidified incubator with 5% CO₂.

Cell migration assay and cell retrieval

Cells were harvested from culture plates with 0.05% Trypsin/EDTA (Gibco, 25200) and centrifuged at 1,000 rpm for 5 minutes. Then, the cells were re-suspended in culture media to a concentration of 3 x 10⁵ cells/ml. 100 μL of this cell suspension was pipetted into left and right inlets, and 20 μL of serum free media was pipetted into the left and right outlets. After 5 minutes, the solution in all inlets/outlets was replaced by 50 μL medium containing 10% serum. The device maintained static flow while remaining in an incubator for 30 minutes to enhance cell adhesion, and we confirmed cell adhesion by microscopy. All cells used in this study adhered within 30 minutes. Then, the media in the left/right inlet was replaced with 200 μL serum-free culture media, and 200 μL of medium with 10% serum was applied to the central inlet to induce chemotactic migration. Due to the nature of diffusion, the concentration of the chemoattractant in the migration channel increases linearly along the channel from left/right channels to central channel. Then, the entire chip was put into a cell culture incubator. Migration distance was measured based on the final cell position in each migration channel after 24 hours of incubation without media replenishment. For data collection, cells were stained by LIVE/DEAD® Viability/Cytotoxicity Kit (Invitrogen, L3224) to distinguish live and dead cells. To have consistent results, we only used data from central 300 (out of 450) migration channels. The difference in cell motility between upstream and downstream is less than 15%, and the results are consistent among devices⁸. For retrieving highly-migratory cells from the device, we pipetted 100 μL of PBS to wash the microfluidic channels for 5 minutes, followed by 100 μL of trypsin for 5 minutes in incubator. The devices were then examined under microscope, and we found more than 95% of cells were retrieved from migration devices. Then, migratory cells with the solution (~ 5 μL) in the outlets were collected and placed in 96-well plates for cell culture.

Single-cell RNA-Seq

We isolated migratory cells from 3 breast cancer cell lines (MDA-MB-231, SUM159, SUM149) and a patient derived breast cancer cell line adapted to standard cell culture (GUM36). Two weeks after isolation, the harvested migratory cells were validated by confirming their elevated migrating capability using microfluidic migration chips. Then, we performed high-throughput single-cell barcoding transcriptome sequencing for each cell population^10,11. After barcode beads captured mRNA from cells, we performed RT (Thermofisher Maxima RT kit), PCR (Kapa HiFi Hotstart PCR Readymix), and library preparation (Illumina Nextera XT Library Prep Kit). The cDNA samples were then quantified and pooled by the University of Michigan Sequencing Core for sequencing. We obtained approximately 30 million reads (paired-end: one side 26 base pairs for barcode and the other side 115 base pairs for mRNA quantification) for each population. Reads were aligned using STAR and processed by the standard flow suggested by Dropseq¹⁰. Then, we used open-source SEURAT kit (http://satijalab.org/seurat/) to analyze single-cell sequencing data. Cells with more than 2,000 genes detected were considered to be successfully sequenced, and the cells having more than 5% mitochondrial gene expression were discarded for their poor viability¹⁰. After quality check, we got 131 wild-type SUM149, 448 migratory SUM149, 202 wild-type SUM159, 216 migratory SUM159, 272 wild-type MDA-MB-231, 1452 migratory MDA-MB-231 (Replicate 1 two weeks after migration sorting: 394, Replicate 1 four weeks after migration sorting: 439, Replicate 1 six weeks after migration sorting: 361, Replicate 2 two weeks after migration sorting: 258), 342 wild-type GUM36, 247 migratory GUM36 cells. For each cell type, we identified significantly altered genes defined by logarithmic fold change as 0.25 and minimal portion of expressing cells as 10% for pathway analysis following data analysis workflow in SEURAT and Dropseq¹⁰. Pathway analysis was performed using Enrichr (http://amp.pharm.mssm.edu/Enrichr/) tool with NCI-Nature pathway database⁴⁹. For cellular heterogeneity analysis, we segregated a cell population into distinct sub-populations based on shared nearest neighbor (SNN). We visualized all data either by t-distributed stochastic neighbor embedding (tSNE) or principal component analysis (PCA). Statistical significance was calculated using Wilcoxon Rank Sum test.

Identification of genes associated with patient prognosis

We first categorized genes significantly over- and under-expressed between migratory and wild-type cells in SUM149, SUM159, MDA-MB-231, and GUM36 breast cancer cells. Among overlapping genes (both over- and under-expressed in migratory cells from all cell lines), we used a publicly available database, kmplot.com²², to determine correlations with prognosis in breast cancer. To analyze the prognostic value of a particular gene, patient samples are split into two groups based on expression of that gene and compared by a Kaplan-Meier curve to calculate the hazard ratio with 95% confidence intervals and log rank P value.

Image acquisition

The microfluidic chips were imaged using an inverted microscope (Nikon). The bright-field and fluorescent images were taken with a 10x objectives and a charge-coupled device (CCD) camera (Coolsnap HQ2, Photometrics). A FITC/TRITC filter set was used for the fluorescent imaging of LIVE/DEAD® Viability/Cytotoxicity Kit (Invitrogen, L3224). Bright field imaging was performed using an exposure time shorter than 10 ms, and the fluorescent imaging was performed using an exposure time shorter than 100 ms, minimizing the phototoxic effect on cells. The microfluidic cell chamber array was scanned with a motorized stage (ProScan II, Prior Scientific). Before each scanning, the stage was leveled to ensure the image remained in focus throughout the whole imaging area.

Automated cell counting program

The locations of live cells were obtained using a custom MATLAB program developed by our lab⁸. For each LIVE/DEAD staining fluorescence microscopy image, there were three overlaid channels: bright field, FITC, and TRITC. This program obtained the fluorescent intensity value of each pixel on FITC channel and TRITC channel. Each pixel had a value ranged from 0 to 255 indicating its brightness, and the pixels with values greater than the pre-defined threshold were treated as “bright pixel” by the program. Any block containing over a pre-defined number of “bright pixels” was counted as one valid cell, so that any noise, cell debris, and device defects were not included because of their small size or low fluorescent intensity. Dead cells were also excluded based on the signal in TRITC channel. In each migration channel, the program marked the live cells moving farthest as the migration frontier in that channel. The whole data analysis process takes less than 1 minute. We have validated the software by comparing manual distance measurement and computer-aided measurement. The difference between two trials is within 3%.

Cell migration data analysis and processing

As cell migration results do not follow normal distribution, a non-parametric Mann-Whitney U test was used for comparisons of cell motility with a significance level of 0.05 considered statistically significant. * refers to P < 0.05, ** refers to P < 0.01, and *** refers to P < 0.001. The box graphs were plotted using Origin 9.0. The bottom and top of the box are the first and third quartiles, and the band inside the box is always the second quartile (the median). The ends of the whiskers represent the 5th percentile and the 95th percentile. The square inside the box indicates the mean, and the x outside the box indicates the minimum and maximum of all the data.