Biomaterial Scaffolds Recruit an Aggressive Population of Metastatic Tumor Cells In Vivo

In most types of cancer, the formation of distant metastases is the point at which the disease is no longer considered curable. Tumor cell colonization of distal sites occurs through a sequence of events, initiated by generation of premetastatic niches by cells originating in the bone marrow (1–3). A small subset of tumor cells that have successfully invaded the tumor vasculature then home to these preconditioned locations (4). Tumor cells recruited to metastatic sites are distinct from both primary tumor (PT) and the major population of circulating tumor cells (CTC), and are often more aggressive (5), resistant to therapy (6), and cancer “stem cell” like (7). This suggests that neither the PT nor CTCs are representative of cells with metastatic potential. At present, methods to detect metastasis involve radiologic imaging, which are capable of detecting metastatic foci 7–10 mm in size (8). Imaging is typically performed for evaluation of clinical symptoms indicating compromised tissue function, and thus foci of this size are usually associated with late-stage disease.

Most early detection strategies currently developed focus on the use of blood as a “liquid biopsy”. The umbrella term liquid biopsies includes CTC, circulating tumor DNA, and exosome detection in the blood, all of which are emerging as platforms to stage patients beyond the presence, size, and molecular characteristics of the PT (9). However, while each of these is associated with metastatic risk, the interplay between the presence and relative number of each of these markers and presence of metastatic disease is tenuous. Biomaterial scaffolds that capture metastatic tumor cells (10) extend beyond the power of the liquid biopsy to capture immune cells associated with the metastatic niche and metastatic tumor cells themselves (11–15). These technologies have been successful in mouse models of breast (11–15), ovarian (16), prostate (17), melanoma (18), and hematologic cancers (19), and have captured tumor cells, reduced metastatic burden (14, 15), and improved survival (15).

A crucial piece of scaffolds as a detection platform and alternative to liquid biopsy is understanding the phenotype of tumor cells recruited to biomaterial scaffolds in vivo and testing the hypothesis that these cells are truly metastatic and similar to tumor cells that have colonized an organ. As poly(ϵ-caprolactone) (PCL) scaffold implantation in combination with surgical resection results in a survival advantage relative to mock surgery (15), we expect that the scaffold may be capturing an aggressive population of metastatic tumor cells. Identifying the phenotype of scaffold-captured tumor cells (SCAF) within the continuum of tumor cell phenotypes will inform the use of scaffolds as surrogates for metastatic sites, facilitating the development of therapeutic strategies targeting metastatic disease.

In this article, we derived cell lines from the MDA-MB-231BR breast cancer xenograft PT, as well as matched lung micrometastasis, and biomaterial SCAF. In vitro and in vivo assays were utilized to characterize phenotypic differences between these cell lines. Finally, we performed RNA-seq and Hi-C to elucidate transcriptional and chromatin configuration differences that generate these phenotypic characteristics. These studies support the utility of SCAF cells as a metastasis surrogate to reveal molecular mechanisms and identify potential therapeutic targets for metastatic cancer.

PCL microspheres were prepared as described previously (15). In brief microspheres were made by emulsification of a 6% (w/w) solution of PCL (Lactel Absorbable Polymers; inherent viscosity 0.65–0.85 dL/g) in dichloromethane in a 10% (w/v) poly(vinyl alcohol) solution followed by homogenization at 10,000 rpm for 1 minute. The solution was then stirred for 3 hours to evaporate dichloromethane solvent. Microspheres were then collected by centrifugation at 2,000 x g for 10 minutes and washed at least five times in deionized water. Finally, microspheres were lyophilized for 48 hours.

Microporous PCL scaffolds were prepared by mixing PCL microspheres and sodium chloride crystals (250–425 μm in diameter) at a 1:30 (w/w) ratio. This salt and polymer microsphere mixture was then pressed in a steel die for 45 seconds at 1,500 psi. Polymer/salt disks were then heated at 60°C for 5 minutes per side to melt polymer microparticles around salt crystals to form a continuous structure. Salt crystals were subsequently removed by immersion in water for 1.5 hours. Scaffolds were then disinfected for animal studies using 70% ethanol, rinsed with sterile water, and dried on a sterile surface.

All animal studies were performed in accordance with institutional guidelines and protocols (PRO00007801) approved by the University of Michigan Institutional Animal Care and Use Committee. Scaffolds were implanted into the dorsal subcutaneous space of 8-week-old female NOD/SCID-IL2Rγ^−/− (NSG) mice (Jackson Laboratory) immediately above the shoulder blades and as physically distant from the fourth right mammary fat pad as possible. For the implantation procedure, animals were anesthetized via isoflurane (2%, inhaled), prepared with carprofen analgesia (5 mg/kg, s.c. injection), the upper back was shaved and prepped using a betadine swab followed by an ethanol swab and this procedure was repeated three times. A fenestrated sterile field was draped over the surgical area and a 1-cm incision was made in the upper back. Following incision, subcutaneous pockets were created perpendicular to the incision, into which sanitized scaffolds were inserted (two scaffolds/mouse). The skin was then closed using sterile wound clips (Reflex 7 mm, Roboz Surgical Instrument Co).

Tumor inoculations were performed by injection of 2 × 10⁶ MDA-MB-231BR-tdTomato-luc2 (parental line, Northwestern University Developmental Therapeutics Core) or 231BR-SCAF,-PT, or -LUNG–derived cell lines in 50 μL PBS (Life Technologies) into the fourth right mammary fat pads of 10-week-old female NSG mice (Jackson Laboratory). Cell lines were confirmed to be pathogen- and Mycoplasma-free and authenticated by short tandem repeat DNA analysis and compared with the ATCC STR profile database (DDC Medical).

Mice were euthanized 4 weeks following tumor inoculation, and scaffolds and organs were retrieved and washed in Hank's Balanced Salt Solution (HBSS, Life Technologies). Samples were minced using microscissors in Liberase TL or TM (0.38 mg/mL, Roche) and incubated at 37°C for 20 minutes. Following incubation with liberase, the enzyme was neutralized via addition of 0.5 mol/L EDTA (Life Technologies) and digested tissues were strained through a 70-μm filter in FACS buffer [PBS (Life Technologies) with 0.5% BSA (Sigma Aldrich) and 2 mmol/L EDTA (Life Technologies)]. Cells were counted using a Countess Automated Cell Counter (Invitrogen) and magnetically sorted using a Mouse Cell Depletion Kit (Miltenyi) according to the manufacturer's instructions. Human cell fractions were cultured in DMEM D6429 (Sigma Aldrich) with 10% FBS and 1% penicillin/streptomycin until growth of tumor cell colonies was evident. Following the first passage of cells, culture was continued without antibiotics. Care was taken to minimize passage number of tumor cells, for all experiments the passage number was less than 10.

For quantification of in vivo metastatic ability of 231BR-SCAF, -PT, and -LUNG–derived cell lines, mice were euthanized 3 weeks after tumor inoculation, and retrieved scaffolds and lungs were washed in HBSS (Life Technologies) and processed with liberase as described for Magnetic Mouse Cell Depletion above. Following isolation of a single-cell suspension, tdTomato⁺ tumor cells were quantified using a MoFlo Astrios Flow Cytometer (Beckman Coulter). The detection sensitivity for cancer cells via flow cytometry was 0.002% (i.e., 5 cancer cells in 25 × 10⁴ total cells; ref. 14). Gating strategy is provided in Supplementary Fig. S1A. Metastasis assays were repeated twice.

Scratch assays were performed as described by Justus and colleagues (20). Briefly, cells were seeded in 24-well plates and allowed to grow to form a confluent monolayer. A sterile 200 μL pipette tip was used to create a wound in the monolayer. The surface was washed with PBS (Life Technologies) to remove any detached cells and replaced with media. Images were taken immediately after wounding, 5 hours, and 10–12 hours after wounding. Results were quantified by identification of the proportion of scratches that had no closure at endpoint (no cells touching from either side of wound), partial closure (at least two cells from either side of wound in direct contact), or complete closure (no evidence of wound remains). Assays were repeated three times.

Transwell assays were performed as described by Justus and colleagues (20). Briefly, Matrigel Invasion Chambers (BD Biosciences) were rehydrated for 30 minutes at 37°C using 500 μL serum-free media. A total of 750 μL of chemoattractant media (serum containing) was placed in the receiver well plate. Cells were harvested using TrypLE (Life Technologies), counted and suspended at 5 × 10⁴ cells/mL in serum-free media. Rehydrated invasion chambers were transferred to receiver well plate containing chemoattractant and 500 μL of cell suspension was added to each insert. Care was taken to prevent trapped air between the insert and chemoattractant media. Assays were incubated for 24 hours at 37°C and subsequently stained and fixed using crystal violet in ethanol. Invaded cells were imaged directly on the membrane using a Nikon Eclipse inverted microscope and imaged at 10×. Four images per well were captured and cell numbers were quantified in ImageJ using automatic particle counting with constant image thresholds. Assays were repeated three times.

Proportion of cancer stem cell (CSC)-like tumor cells in each cell line was determined by quantification of EpCAM⁺CD44⁺CD24⁻ cell population by flow cytometry. Tumor cells were harvested using TrypLE (Life Technologies), counted, and suspended at 1 × 10⁷ cells/mL. Cells were evenly split between control and test conditions and stained with anti-human BV421 EpCAM (clone EBA-1, BD Biosciences), PerCP-Cy5.5 CD44 (clone G44-26, BD Biosciences), and PE/Cy7 CD24 (clone ML5, BioLegend). Samples were run on MoFlo Astrios Flow Cytometer (Beckman Coulter) and data were processed using FlowJo (Tree Star Inc.). Gating strategy is provided in Supplementary Fig. S1B. CSC marker flow cytometry was repeated three times. Mammosphere assays were performed as described by Tarasewicz and colleagues (21).

Cells were harvested using TrypLE (Life Technologies), counted, and suspended at 5 × 10⁶ cells/mL. RNA was isolated using High Pure RNA Isolation Kit (Roche) according to the manufacturer's instructions. Isolated RNA concentrations were quantified using NanoDrop 2000 (Thermo Fisher Scientific) and resuspended at 80 ng/μL. RNA samples were kept at −80°C until further use. RNA quality was assessed using an Agilent 2100 Bioanalyzer. The DNA Sequencing Core at University of Michigan (Ann Arbor, MI) performed library construction and sequencing using the Hiseq-4000 platform using single-end 50-base sequence reads with a multiplex of six per sequencing lane.

Transcriptome analysis was performed as described previously (22). For RNA-seq data processing, the raw reads were checked with FastQC to identify potential quality issues in the raw data. Next, Tophat and Bowtie were used to align the reads to a human reference transcriptome (HG19). Alignment settings were kept on default except for “-b2-very-sensitive”, “-no-coverage-search”, and “-no-novel-juncs” to limit the search to known transcripts only. FastQC was then performed again on the mapped reads. Cufflinks/Cuffdiff was then utilized to quantify expression and perform differential expression analysis using parameter settings “—multi-read-correct” and “upper-quartile-norm”. CummeRbund was then used to format the cufflinks output. Benjamini–Hochberg procedure was used in CummeRbund for multiple testing method correction. Gene level analysis was performed using fragments per kilobase per million reads values outputted by Cufflinks and log₂ fold change with pseudocounts. Bin-level gene expression vectors were calculated using raw counts outputted by cufflinks and adding up the counts for all the genes in each bin then normalizing by million reads to convert them to FPKM. Gene set enrichment analysis (GSEA; refs. 23, 24) was performed with GSEA v3.0 software.

Hi-C library construction was performed as described by Chen and colleagues (25). Briefly 2 × 10⁶ cells for each condition were cross-linked with 1% formaldehyde (Thermo Fisher Scientific) in serum-free media for 10 minutes at room temperature and subsequently quenched with glycine (Sigma Aldrich) to a final concentration of 0.125 mol/L. Cross-linked cells were flash frozen in liquid nitrogen and stored at −80°C until the construction of libraries. Cross-linked cells were thawed on ice and resuspended in 1 mL lysis buffer [10 mmol/L Tris-HCl, 10 mmol/L NaCl, 0.2% Igpel (Sigma Aldrich)] and incubated for 15 minutes. Cells were next homogenized on ice and the lysate was transferred to a 1.7 mL tube. Cells were collected by centrifugation (5 minutes at 2,000 × g) and washed twice in NEB buffer. HindIII digestion was then performed in NEB buffer at 37°C overnight. Following digestion, restriction overhang ends were filled and labeled with biotin. Samples were then ligated at 16°C for 4 hours in a reaction containing 1 × ligation buffer, 1% Triton X-100 (Sigma Aldrich), 1 mg/mL BSA (Thermo Fisher Scientific), 10 mmol/L ATP (Sigma Aldrich), and 50 μmol/L T4 DNA ligase (Life Technologies). Proteinase K (0.5 mg, Life Technologies) was then added and incubated for 4 hours at 65°C. After this incubation an additional 0.5 mg proteinase K was added to each tube and the incubation was continued overnight at 65°C. DNA was then extracted using phenol:chloroform (1:1) and desalted using AMICON Ultra Centrifugal Filter Unit (Millipore) with 1 × TE buffer. Biotin was then removed from unligated ends, DNA purified with single phenol extraction and precipitated by addition of ethanol. DNA was resuspended in water and fragmentation was performed using a sonicator (Covaris, Inc.). Fragments of 200–400 bp were recovered with Agencourt AMPure XP (Beckman Coulter) utilizing the manufacturer's protocols. DNA fragment ends were repaired and purified with a MinElute column, and A-tailing was performed. Streptavidin pull-down of biotinylated Hi-C DNA was performed. Illumina adapter ligation was then performed and then DNA was amplified by 15 PCR cycles for Illumina HiSeq Sequencing. A standard quality control protocol was performed prior to sequencing and all libraries passed. Samples were pooled and sequenced in a single lane of a flow cell on HiSeq 4000 to generate paired-end sequence reads at 100 bases per end read.

A standard pipeline was used to process Hi-C sequence data using the University of Michigan Bioinformatics Core facilities (Ann Arbor, MI). Raw sequence reads were processed with FastQC for data quality control and paired-end reads were mapped to the HG19 reference human genome using Bowtie2. HOMER was used to develop the contact matrix at 100 kb and 1 Mb resolution.

Hi-C matrices were normalized, plotted, and analyzed in conjunction with RNA-seq data for cell lines using the 4D Nucleome Analysis Toolbox (4D-NAT) described by Seaman and colleagues (26).

SCAF, PT, and LUNG cell lines were cultured and passaged as described above. Roughly 2 × 10⁶ cells were isolated and genomic DNA was isolated with Qiagen DNEasy Kit according to the manufacturer's instructions. Samples were submitted to the UMich DNA sequencing core and the core performed quality control, exome capture using the Nimblegen Human Exome Capture Kit, and paired-end sequencing at 150 bases per end read using the Illumina HiSeq-4000 platform with approximately 100× coverage for each cell line.

Genestack platform was used for analysis of whole-exome sequencing data. Quality control was performed using FastQC algorithm. Adaptors and contaminants were trimmed according to methods published by Aronesty and colleagues (27) and then sequences were mapped onto the human reference genome HG19. Quality control of the mapped reads was again performed using the FastQC algorithm. Variant calling and effect prediction was then performed using SAMtools and SnpEff algorithms, respectively, and variants were selected to only include those resulting in missense or nonsense mutations with a quality >60. Gene ontology analysis of mutant genes was then performed using the Gene Ontology Consortium Software (28).

Figures 2–7 have associated raw data. Data are provided as a private link to a OneDrive folder https://1drv.ms/f/s!AmVnzTmi4yrLg8FbakFHWvhuImHOxA with password: cancerresearch until publication, at which point it will be made publicly available.

We generated cell lines from various tumor cell locations in vivo including cells derived from subcutaneously implanted scaffolds and compared them with cell lines derived from PTs or lung metastases derived from the same set of mice. On the basis of previous reports, the earliest metastatic cells should be captured in the scaffold (14, 15), with metastases arriving in the lung at subsequent times. This difference in time of arrival may underlie some variations in these metastatic cell populations. Using a mouse cell depletion kit and magnetic-activated cell sorting, we isolated human MDA-MB-231BR human breast cancer cells from each location. Approximately 1 × 10⁴ cells were plated for PT-derived cells, whereas approximately 100 and approximately 500 cells were isolated for scaffold- and lung-derived cells, respectively. Representative images of each are shown (Fig. 1). While PT cells attached and spread at day 1 (Fig. 1A), scaffold- and lung-derived cells took much longer to attach and begin growing, likely due to plating density (Fig. 1B). By day 12 of culture, scaffold-captured cells had attached and formed approximately 15 individual colonies and lung-derived cells had formed hundreds of colonies, whereas PT formed confluent cultures without visible colonies (Fig. 1C). Cells were passaged and cultured to develop stable cell lines, hereafter referred to as SCAF (scaffold-captured), LUNG (lung micrometastasis–derived), and PT (primary tumor–derived) cell lines.

We investigated the metastatic capacity of these cell lines in vivo. Cell lines were orthotopically inoculated into NSG mice and allowed to grow for 21 days. Tumor and spleen mass were similar for all lines. Representative bioluminescent images of the lungs indicate metastasis, with differential metastatic capacity for each (Fig. 2A). Metastasis was quantified by the presence of tdTomato⁺ cells in tissue homogenates by flow cytometry. Metastatic tumor burden was significantly higher in implanted scaffolds (2.4 ± 1.4 times higher; Fig. 2B) and lungs (27.3 ± 22.5 times higher; Fig. 2C) for SCAF tumor–bearing mice relative to PT tumor–bearing mice (P < 0.05 via t test). Metastatic tumor burden was not significantly different between LUNG and PT or LUNG and SCAF in implanted scaffolds (1.9 ± 1.2 times higher than PT; Fig. 2B) or lungs (18.1 ± 22.3 times higher than PT; Fig. 2C). As a control, we also developed a tumor cell line from the lung of a mouse that did not have scaffold implants, termed LUNG2. We did not observe a difference in metastatic ability from LUNG to LUNG2 (Supplementary Fig. S2). Collectively, these studies demonstrate that SCAF are more metastatic than PT and more comparable with LUNG.

As SCAF cells were highly metastatic in vivo, we next investigated the lines in vitro to identify mechanisms of enhanced metastasis. Assays were performed to analyze migration and invasion as well as CSC characteristics. Using a standard scratch assay, migration into a defect was quantified. Representative images demonstrate migrating tumor cells at 0 and 12 hours following the scratch (Fig. 3A). PT and LUNG samples demonstrate failure to generate closure (7/12) or partial closure (5/12), and show no samples with full closure (Fig. 3B). For the SCAF cell line, a statistically significant increase was observed in both partial (9/12) and total (2/12) closure and a significant decrease in no closure (1/12) was observed relative to PT and LUNG (P < 0.05 via Fisher Exact Test).

The invasion characteristics were measured using a transwell invasion assay. Representative images are provided (Fig. 3C). PT and LUNG invaded at a rate of 44 ± 17 and 51.3 ± 18.1 cells/image, respectively (Fig. 3D). Consistent with the scratch assay, a significant increase in the number of SCAF cells were observed, with 109 ± 44 cells/image, respectively. The results demonstrate that SCAF cells have a higher capacity for invasion and migration than either PT or LUNG.

We next investigated expression of CSC markers in each cell population. Using flow cytometry, we evaluated the expression of EpCAM, CD44, and CD24. Representative plots are provided, showing relative levels of EpCAM⁺ and CD44⁺CD24⁻ cells in cell lines (Fig. 4A). PT cells had 6.2% ± 0.2% EpCAM⁺CD44⁺CD24⁻ (CSC⁺) cells. A significantly increased proportion of CSC⁺ cells (7.3% ± 0.2%) was observed for SCAF (Fig. 4B). LUNG had a significantly decreased percentage of CSC⁺ cells relative to both PT and SCAF (5.3% ± 0.4%; Fig. 4B).

The stem cell characteristics were also investigated via mammosphere assay, which measures self-renewal in suspension. Representative images are provided (Fig. 4C), demonstrating mammosphere formation for SCAF, PT, and LUNG cell lines. PT generated 4.3 ± 1.5 spheres/well of a 24-well culture dish, which was comparable with the LUNG, which was 3.9 ± 1.9 spheres/well (Fig. 4D). SCAF had significantly increased mammosphere formation capability with 9.4 ± 3.1 spheres/well (P < 0.05 relative to PT and LUNG). In addition, mammospheres in the SCAF line were much larger than PT or LUNG, however we cannot rule out that this was a result of aggregation of individual spheres thus only sphere number was quantified. Collectively, these results demonstrate that the scaffold sequesters a population of cells that have increased capacity for migration and invasion, and increased proportion of CSCs, both of which could contribute to the survival advantage provided by the scaffold.

We next investigated the hypothesis that phenotypic differences between SCAF, PT, and LUNG result from the differential accumulation of mutations in these populations. To this end we performed whole-exome sequencing on SCAF, PT, and LUNG. Variant calling was performed on filtered reads mapped to the human reference genome HG19 (GRCh37) and these variants were further filtered to include only mutations resulting in a missense or nonsense mutation and compared for variants between each line (Supplementary Fig. S3; Supplementary Table S1). The vast majority of variants relative to HG19 were shared across all lines (3,191 or 95.5% of total variants) with each showing 10–13 unique variants (0.3%–0.4% of total variants). Gene ontology was then used to investigate the functional impact of these differential mutations. Mutations for all lines were implicated in the same ontological terms indicating selection based on additional accumulated mutations after leaving the PT is unlikely in this setting (Supplementary Table S2).

We next utilized RNA-seq to characterize transcriptomic differences between the cell lines that might account for their phenotypic properties. We identified 14,232 genes with measured expression, 2,901 of which were differentially expressed (Benjamini–Hochberg corrected P < 0.05 and log fold change > 0.6) between SCAF and PT cells and 14,572 genes with measured expression, 2,398 genes of which were differentially expressed between SCAF and LUNG cells. A small subset of the most differentially expressed genes (Supplementary Fig. S4A) was confirmed using qRT-PCR (Supplementary Fig. S4B). Of this set of genes, the vast majority (19/20) were found to be more similarly expressed between SCAF and LUNG compared with SCAF and PT. The PAM50 gene set (29) was interrogated, and the majority of genes (35/50) were more similarly expressed in SCAF and LUNG when compared with SCAF and PT. These genes included Myc, MDM2, ESR1, ERBB2, EGFR, CCNB1, CXCR1, IL8, IL6R, and IL6ST. Using unsupervised hierarchical clustering with the top 15 differentially expressed genes, SCAF cells clustered separately from all PT samples and all but one LUNG sample (Fig. 5A). The same result was obtained by hierarchical clustering with all genes with differential expression (Supplementary Fig. S4A). Using GSEA, TNFα via NFκB signaling and epithelial-to-mesenchymal transition (EMT) were positively enriched and glycolysis was depleted in SCAF (FDR q-value < 0.25 and P < 0.05) relative to PT (Fig. 5B). EMT was also enriched in SCAF relative to LUNG (Fig. 5C) along with E2F targets, Myc targets, and G₂–M checkpoint. Similarly, estrogen response and myogenesis were depleted in SCAF relative to both PT and LUNG, and peroxisome and IL6, JAK/STAT3 signaling were depleted in SCAF relative to LUNG (Fig. 5C). No pathways were enriched in PT relative to LUNG by GSEA; however, TNFα via NFκB signaling was enriched in LUNG relative to PT (NES 1.969 with FDR q-value < 0.001). Collectively, this analysis of gene expression indicates that the SCAF phenotype aligns more closely to LUNG than to PT.

We subsequently compared SCAF and PT cell lines by Hi-C (30) to investigate chromatin structure and genome organization differences that may give rise to differential gene expression and resultant functional behavior including metastatic phenotype. Hi-C data were collected from SCAF and PT lines and analyzed in conjunction with RNA-seq results at three scales: 1 Mb, 100 kb, and gene/pathway level binning (Supplementary Fig. S5). Overall, the Hi-C matrices appear similar with the same translocations and copy number changes visible within the Toeplitz normalized matrices at 1 Mb binning (Fig. 6A). With filtering to the 95th percentile of each matrix, translocations are visible as blocks of strong trans interchromosomal interactions and unbalanced and balanced translocations have an “L” or “X” appearance, respectively, and were found to be matched between SCAF and PT (Supplementary Fig. S6A). Copy number changes can also be visualized by looking at the sum of Hi-C contacts at 100 Mb binning for each chromosome and these were also found to be concordant between SCAF and PT (Supplementary Fig. S6B). In addition, the copy number normalization method published by Seaman and colleagues (22) was utilized to investigate differential breakpoints. No differential breakpoints were found between SCAF and PT samples. Normalized 1 Mb Hi-C contact matrices were first investigated for large-scale structural differences by summing the counts for each 1 Mb bin in one dimension. Bins with a value more than four SDs above the mean were identified as altered in structure and starred in red (Fig. 6B) with all having higher contact density in SCAF compared with PT. These altered bins were then investigated for changes in gene expression (Fig. 6C), which showed that the increased genomic interaction measured by Hi-C was accompanied by an increase in expression in SCAF relative to PT for the majority of bins. For bins with altered gene expression the average number of genes per bin that were differentially expressed with a log₂ fold change > 0.6 was significantly increased relative to randomly selected bins of the same number of genes (P < 0.05). Next, eigenvalue decomposition (Fig. 6D) and calculation of entropy difference between samples was performed on 100 kb binned matrices for each chromosome (Fig. 6E) demonstrating enhanced entropy in SCAF relative to PT for the majority of chromosomes. In addition, the Fiedler vector (the eigenvector associated with the second smallest eigenvalue of the Laplacian matrix; ref. 31) was calculated for each 100 kb binned matrix for each chromosome and plotted with the Hi-C contact matrix and RNA-seq at the same binning. Structural differences from PT to SCAF were found in the Fiedler vector for chromosomes 11 (Fig. 7A), 18, and 21.

Next, Hi-C contact matrices were generated at the gene level for the whole genome (Fig. 7B, top) and individual chromosomes 18 and 21 identified as altered in structure at 1 Mb and 100 kb (Fig. 7B). Finally, changes in centrality were calculated for each gene (Hi-C) and combined with gene expression (RNA-seq) in Principal Component Analysis to reduce the dimension of the data while maintaining meaningful variability. Projections of each bin onto the first two principal components are shown (Fig. 7C). To quantify which genes have the greatest change, the distance in the principal component space from PT to SCAF samples was calculated. Genes whose distance from PT to SCAF changed more than four SDs from the mean whole-genome distance were identified as altered in structure/function. Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways were then selected on the basis of RNA-seq (Fig. 5C) to investigate enrichment of structure/function changes within subsets of genes (Supplementary Table S3). By centrality analysis the KEGG MAPK, ErbB, HIF-1α, FoxO, mTOR, PI3-Akt, AMPK, VEGF, focal adhesion, TNFα, leukocyte transendothelial migration, cancer, and breast cancer pathways were significantly altered from PT to SCAF (P < 0.05; Fig. 7C). Similarly, distances greater than four SDs above the mean could be observed for specific genes within pathways including COL6A1, FN1, and CTNNB1 for the focal adhesion pathway (Fig. 7C). Finally, based on GSEA implicating NFκB as a key regulatory pathway, genes with NFκB-binding sites were investigated for changes in structure/function. By centrality analysis, genes with NFκB-binding sites were significantly altered in structure and function. Altered genes more than four SDs above the mean were identified including HSP90AB1, COL7A1, EIF5A, and BRD2 among others (Fig. 7C). Taken together, these results indicate we can identify genome structure organizational differences that are associated with development of a metastatic phenotype.

In this study, metastatic tumor cells derived from those captured by microporous PCL scaffolds in vivo were investigated for their phenotype relative to tumor cells from the PT, and metastatic lung lesions. This work was motivated by the idea that SCAF cells may be more easily sampled, while accessing organ-derived metastatic cells would necessitate an invasive biopsy with potential for high morbidity and false-negative results dependent upon the extent of disease progression. Our work and that of others have demonstrated that scaffolds recapitulate multiple elements of the premetastatic niche in vivo along with the capacity to capture metastatic cells within a biomaterial implant (10). Analyzing these cells along the continuum of cancer cell phenotypes provides a foundation for using a scaffold biopsy to impact patient care, such as for the development of personalized therapies (32). We developed stable cell lines from metastatic locations to have sufficient numbers of cells for in depth in vitro behavioral, functional, and molecular assays and in vivo metastasis assays. Notably, the scaffold-derived tumor cells were highly metastatic in vivo, demonstrated substantial invasiveness in vitro, and molecular analysis indicated these cells were similar to cells derived from metastatic lung lesions.

SCAF and LUNG were more highly metastatic than PT in vivo, supporting the notion that cells captured by the scaffold represent a population with the ability to successfully metastasize and colonize a tissue. Interestingly, the SCAF line showed comparable ability to metastasize to both scaffold and lung tissues, which differs from literature that has reported cell proclivity to metastasize to the location of origin (33–35). The concordant metastatic tropism to lung for both SCAF and LUNG suggests that the metastatic traits of SCAF are not site specific yet are indicative of overarching traits required for metastasis. Consistent with increased aggressiveness in vivo, SCAF cells were more migratory, invasive, and stem-like in vitro than cells from the PT. These results provide the first evidence that tumor cells derived from biomaterial scaffolds are a functionally aggressive and metastatic population of cells, with the capacity to spread to distal sites and compromise organ function.

Unsupervised hierarchical clustering of RNA-seq data identified that SCAF clustered more closely with LUNG than with PT. In addition, upon investigation of the top differentially expressed genes between SCAF and PT, 19 of 20 of these genes showed greater homology between SCAF and LUNG when compared with SCAF and PT. Similarly, for the PAM50 gene set 35 of 50 genes had more similar expression between SCAF and LUNG than for SCAF and PT. Collectively, these results indicate that scaffold and lung-derived cells, both originating from metastatic sites, show greater homology compared with scaffold-captured cells and those derived from the PT.

Further investigation of the tumor cell phenotypes by GSEA identified pathways significantly enriched/depleted in SCAF cells relative to PT and LUNG. Hallmark pathways (36) enriched in SCAF relative to PT included TNFα signaling via NFκB, EMT, angiogenesis, and complement. TNFα signaling through NFκB has been shown to enhance stemness of breast cancer cells (37), angiogenesis and complement are indicators of metastasis and survival in patients (38, 39), and EMT is closely linked to metastasis and stemness (40, 41). The SCAF line is more migratory and invasive than the LUNG line in vitro, and yet SCAF and LUNG demonstrate similar metastatic ability in vivo. This apparent contradiction may be explained by the following observations. First, EMT is enriched in SCAF relative to both PT and LUNG, indicating the SCAF line is more mesenchymal and thus more migratory and invasive in vitro. Second, in previous work it has been shown that the earliest metastatic tumor cells colonize the scaffold prior to the lung (14). Third, the SCAF cells are more homogeneous than LUNG, as demonstrated by hierarchical clustering. Fourth, IL6 JAK/Stat3 signaling was enriched in LUNG relative to SCAF, which may enhance metastasis in vivo through paracrine communication with myeloid-derived suppressor cells enhancing stemness but may not influence phenotype in vitro (42). Finally, SCAF cells are more stem-like and demonstrate higher entropy/plasticity. Hallmark pathways enriched in SCAF relative to LUNG also included E2F targets (43), Myc targets (44), and G₂–M checkpoint (44): all important axes in breast cancer. Peroxisome and IL6 JAK/Stat3 signaling were enriched in LUNG relative to SCAF and have been implicated in therapeutic resistance (45) and stemness (42). Importantly, TNFα signaling through NFκB was enriched in both SCAF and LUNG relative to PT, indicating this may be the crucial pathway for gaining enhanced metastatic ability for these cells. These results suggest that LUNG and SCAF may evolve toward similar phenotypes through the same dominant transcriptional program (TNFα via NFκB) but with different accessory programs facilitating site-specific adaptations (i.e., SCAF with EMT and LUNG with IL6 JAK/Stat3 signaling; ref. 42), which is consistent with other observations of metastatic lines (46). This supports the use of scaffold as a sampling location and also suggests that both metastatic locations are derived from a similar pool of cells yet may experience differential evolutionary pressure and adaptation at different distal sites. The scaffold has the capacity to provide a sampling site that would be clinically easy to access relative to an organ biopsy and contains a source of cells with clear phenotypic and molecular similarities to the pathologic metastatic site.

The underlying genome/chromatin organization role in the development of metastatic phenotype from PT to SCAF was identified with Hi-C. Using a multiscale approach, we found structural changes linked to functional outcomes at all points. Critically, for all except chromosome 4, entropy for SCAF was higher than PT, indicating a more dysregulated network consistent with the concept of increased aggressiveness and phenotypic flexibility that are hallmarks of stemness and metastasis (47–49). We chose to analyze SCAF as a surrogate for other metastatic sites due to the in vivo and transcriptomic similarity of these cells to the LUNG. Importantly, no copy number changes or additional translocations were observed, indicating the karyotype from PT to SCAF was conserved. Even at 1 Mb binning analysis of whole-genome contact density, we found structural differences between SCAF and PT indicating areas of chromosomes 11, 18, and 21 with a higher density of intrachromosomal contacts for SCAF. Interestingly, there were no locations identified where PT had higher contact density. Investigating these altered bins further, we found that vast majority had higher expression in SCAF, indicating increased connectivity resulted in enhanced gene expression. Similarly, we observed subtle changes in both eigenvalue decomposition and Von Neumann entropy for each chromosome. At 100 kb we found chromosomes 11, 18, and 21 identified by contact density at 1 Mb to have differences in Fiedler vector consistent with a change in structure from SCAF to PT. We also investigated the centrality of each chromosome at the gene level (structure) and gene expression (function) for the whole genome and identified significantly altered pathways in structure/function. Implicated genes in these pathways include many genes associated with metastasis and stemness including COL6A1, FN1, ITGA3, IL11, IL6ST, IL8, and TGFβR2. In particular, IL8 has been implicated in enhanced stemness, aggressiveness (50, 51), and resistance to therapy (52). Finally, based on GSEA of RNA-seq data implicating TNFα via NFκB signaling as the common transcriptional program associated with metastatic phenotype in both SCAF and LUNG cells, we investigated structure/function relationships for genes with NFκB-binding sites. We found this gene set to be significantly altered from PT to SCAF in structure/function. This further supports NFκB as one of the major drivers of phenotypic change from nonmetastatic to metastatic phenotype (53–55). Our data are among the first to connect genome/chromatin structure alterations to changes in functional gene expression driving metastatic phenotype.

Our results provide primary evidence that tumor cells recruited to biomaterial scaffolds are a functionally aggressive and metastatic population capable of distal organ colonization. Taken together, this work demonstrates scaffold-recruited tumor cells are similar to other metastatic sites, but also represent a population of highly aggressive, stem-like cells whose capture may hold promise for reducing breast cancer mortality. In addition, these results support the possibility that metastatic phenotype can be gained through changes in genome structure impacting functional gene expression rather than mutational alterations. Transcriptional alterations in tumor cell phenotype and plasticity that are maintained in culture may be equally as important as mutational changes for development of metastatic ability. The scaffold platform technology provides a unique opportunity to analyze the earliest metastatic cells, enabling the analysis of tumor cells that have left the PT and colonized a distal site. This article identifies for the first time, the many similarities and differences that are present between the metastatic cells recruited to a biomaterial scaffold and the lung. Future work should focus on the impact of heterogeneity, timing at which tumor cells arrive at the niche, and the impact of varied microenvironments on tumor cell phenotype. This platform provides an unprecedented opportunity for analysis and development of tools for personalized medicine.

L.D Shea has ownership interest (including stock, patents, etc.) at Cour Pharmaceutical. No potential conflicts of interest were disclosed by the other authors.

Conception and design: G.G. Bushnell, T.P. Hardas, J.S. Jeruss, L.D. Shea

Development of methodology: G.G. Bushnell, T.P. Hardas, R.S. Oakes, I. Rajapakse, M.S. Wicha, J.S. Jeruss

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): R.M. Hartfield, Y. Zhang, R.S. Oakes, H. Chen, I. Rajapakse, J.S. Jeruss

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): G.G. Bushnell, T.P. Hardas, S. Ronquist, I. Rajapakse, J.S. Jeruss, L.D. Shea

Writing, review, and/or revision of the manuscript: G.G. Bushnell, T.P. Hardas, R.M. Hartfield, Y. Zhang, R.S. Oakes, M.S. Wicha, J.S. Jeruss, L.D. Shea

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): L.D. Shea

Study supervision: J.S. Jeruss, L.D. Shea

Thanks to the NIH for their support through 5R01CA173745-06, R01CA214384, and R35 CA197585. G.G. Bushnell is a recipient of the NSF Graduate Research Fellowship and NRSA F31 CA224982-01. Thanks to the University of Michigan DNA Sequencing Core for assistance in sequencing and analysis. Thanks to the Developmental Therapeutics Core at Northwestern University for donation of MDA-MB-231-BR-tdTomato+Luc2+ parental line.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.