RNAi screen identifies essential regulators of human brain metastasis-initiating cells
Brain metastases (BM) are the most common brain tumor in adults and are a leading cause of cancer mortality. Metastatic lesions contain subclones derived from their primary lesion, yet their functional characterization is limited by a paucity of preclinical models accurately recapitulating the metastatic cascade, emphasizing the need for a novel approach to BM and their treatment. We identified a unique subset of stem-like cells from primary human patient brain metastases, termed brain metastasis-initiating cells (BMICs). We now establish a BMIC patient-derived xenotransplantation (PDXT) model as an investigative tool to comprehensively interrogate human BM. Using both in vitro and in vivo RNA interference screens of these BMIC models, we identified SPOCK1 and TWIST2 as essential BMIC regulators. SPOCK1 in particular is a novel regulator of BMIC self-renewal, modulating tumor initiation and metastasis from the lung to the brain. A prospective cohort of primary lung cancer specimens showed that SPOCK1 was overexpressed only in patients who ultimately developed BM. Protein–protein interaction network mapping between SPOCK1 and TWIST2 identified novel pathway interactors with significant prognostic value in lung cancer patients. Of these genes, INHBA, a TGF-β ligand found mutated in lung adenocarcinoma, showed reduced expression in BMICs with knockdown of SPOCK1. In conclusion, we have developed a useful preclinical model of BM, which has served to identify novel putative BMIC regulators, presenting potential therapeutic targets that block the metastatic process, and transform a uniformly fatal systemic disease into a locally controlled and eminently more treatable one.
KeywordsBrain metastasis Brain metastasis-initiating cell BMIC BMIC regulators TWIST2 SPOCK1 Patient-derived xenotransplant Non-small cell lung cancer
Brain metastases (BM) are the most common tumor to affect the adult central nervous system, occurring at a rate ten times greater than that of primary brain tumors . With a median survival calculated in weeks in untreated patients, extended to only 3-18 months following multimodal therapy, BM represent a leading cause of cancer mortality . Although BM pose a significant clinical burden, the metastatic process is highly inefficient as individual tumor cells maintain a variable capacity for invasion, dissemination, therapeutic resistance, and other “hallmarks of metastasis” . While it is apparent that metastatic lesions are genetically divergent from their primary tumors [4, 55], determining the subclonal architecture of primary and metastatic lesions may lead to novel therapeutic opportunities, the identification of predictive biomarkers and subsequent prophylactic treatment of those patients most likely to develop metastases.
The use of primary patient samples to study metastasis is limited by several factors , from initial procurement of specimens to gathering a sufficient number of cells for experiments. Additionally, models utilizing primary human cells have yet to be successful in recapitulating each stage of the metastatic cascade . Instead, current frameworks investigate each metastatic stage in isolation, limiting the translational efficacy of putative therapeutic targets . In the current study, we have established in-house early-passage brain metastasis cell lines termed as brain metastasis-initiating cells (BMICs) from primary patient samples of lung-to-brain metastases. We demonstrate that BMIC lines are enriched for stem cell properties as well as the ability to recapitulate the metastatic process both in vitro and in vivo. We also established a BMIC patient-derived xenotransplantation (PDXT) model to interrogate human lung-derived BM. We performed in vitro and in vivo RNA interference screens to identify SPOCK1, a TGF β target gene  and TWIST2 as important BMIC regulators. SPOCK1 in particular is a potent regulator of BMIC self-renewal, tumor initiation, and metastasis. Moreover, our results were translatable to clinical samples, where in a prospective cohort of primary non-small cell lung cancer patient slides, both targets were exclusively expressed in only those patients who developed BM. We then examined whether an indirect relationship exists between SPOCK1 and TWIST2 using protein–protein interaction networks, through which we identified a unique set of BMIC interactors that have significant prognostic value in lung cancer patient survival. One such marker, INHBA, a TGF-β ligand found mutated in lung adenocarcinoma , showed reduced expression in BMICs with knockdown of SPOCK1, further implicating a role for the TGF β pathway in BM development as well as identifying novel therapeutic targets in the pathway. With this work, we present a unique method of modeling human BM using patient-derived BMICs. Using this model, we have successfully identified and functionally validated novel regulators of primary and metastatic tumor growth, and clinically validated our targets as predictive biomarkers for BM. In blocking the metastatic process, we aim to transform a uniformly fatal systemic disease into one that is locally controlled and eminently more treatable.
Materials and methods
Patient sample processing and cell culture
Brain metastases (BM) originating from non-small cell lung carcinoma (NSCLC) primary samples were obtained from consenting patients, as approved by the Research Ethics Board at Hamilton Health Sciences. BMs were processed and maintained in Tumor Sphere Media (TSM) as previously described . BMICs were grown as tumorspheres that were maintained at 37 °C with a humidified atmosphere of 5% CO2. Each BMIC line was characterized by performing limiting dilution assays (LDA) as previously described . The most aggressive BMIC line with an enhanced engraftment capacity, BT478, was the primary BMIC line used for in vivo and RNAi experiments.
Cell proliferation assay
Single cells were plated in a 96-well plate at a density of 1000 cells/200 µL per well in quadruplicate and incubated for 4 days. 20 µL of Presto Blue (Invitrogen) was added to each well approximately 2 h prior to the readout time point. Fluorescence was measured using a FLUOstar Omega Fluorescence 556 Microplate reader (BMG LABTECH) at excitation and emission wavelengths of 535 and 600 nm, respectively. Readings were analyzed using Omega analysis software.
Zone exclusion migration assay
Flow cytometric characterization
BMIC tumorspheres were dissociated and single cells resuspended in PBS + 2 mM EDTA. Cell suspensions were stained with anti-CD133, anti-CD44, anti-CD15, anti-EpCAM or matched isotype controls (Miltenyi) and incubated for 30 min on ice. For TWIST2 and SPOCK1 co-expression analysis, single cell suspensions of BMIC lines BT478 and BT530 were stained with primary SPOCK1 antibody (Mouse monoclonal IgG, Abcam) followed by goat anti-mouse alexafluor 647 secondary antibody (1:1000, Invitrogen), and TWIST2 (Sheep Polyclonal IgG, R&D Systems) followed by donkey anti-sheep FITC secondary antibody (1:1000, Life Technologies), with 30-min incubations on ice for each antibody.
Samples were run on a MoFlo XDP Cell Sorter (Beckman Coulter). Dead cells were excluded using the viability dye 7AAD (1:10; Beckman Coulter). Compensation was performed using mouse IgG CompBeads (BD). Surface marker expression was defined as positive or negative based on the analysis regions established using the isotype control.
Immunofluorescence staining of BMIC lines for SPOCK1 and TWIST2
50,000 cells from BMIC lines BT478, BT530, and BT751 were cultured on Matrigel coated-coverslips in 24-well plates for 72 h. Wells were washed with PBS and cells fixed with 4% paraformaldehyde for 20 min at room temperature (RT), permeabilized with 0.1% Triton X-100 for 5 min, then blocked with 1% BSA at RT for 20 min. Cells were stained with primary Anti-human SPOCK1 antibody (10 µg, Mouse monoclonal IgG, Abcam) followed by donkey anti-mouse alexafluor 647 secondary antibody (1:1000, Invitrogen), or Anti-human TWIST2 (10 µg, Sheep polyclonal IgG, R&D Systems) followed by donkey anti-sheep FITC secondary antibody (1:1000, Life Technologies), for 2 h at RT then followed by secondary antibodies for 1 h at RT. Slides were then mounted with mounting medium and analyzed using Volocity imaging software.
In vivo modeling of metastasis
Intracranial injections (ICr)
Intracranial injections were performed as previously described . Briefly, 10 μL of cell suspension (Online Resource 1 Table 2) was injected into the right frontal lobe of 8- to 10-week-old mice.
Intrathoracic injections (IT)
30 uL cell suspension (5 × 105 cells) mixed with 30 μL of growth factor reduced matrigel (BD Bioscience) was injected into the right upper chest of 8- to 10-week-old mice.
Intracardiac injections (ICa)
100 μL cell suspension (2.5 × 105 cells) was injected into the left ventricle over 15 s in 8- to 12-week-old mice.
Mice were monitored weekly, and upon reaching endpoint brains and lungs (for IT and ICa injections) were harvested, sectioned, and paraffin-embedded for hematoxylin and eosin (H&E) and Anti-human COXIV immunohistochemistry. Images were scanned using an Aperio Slide Scanner and analyzed by ImageScope v126.96.36.1990 software (Aperio).
Reverse transcription and quantitative PCR of mRNA
Total RNA was isolated using Norgen RNA extraction kit (Biotek) and reverse transcribed using qScript cDNA Super Mix (Quanta Biosciences) and a C1000 Thermo Cycler (Bio-Rad). qRT-PCR was performed using the Cfx96 (Bio-Rad) with SsoAdvanced SYBR Green (Bio-Rad) using gene specific primers (Online Resource 1 Table 7) and 28SrRNA as the internal control.
shRNA dropout screen
shRNA constructs (5–8 individual shRNA constructs per gene) were obtained through the RNAi Consortium and the pool was constructed as previously described . Replication-incompetent lentivirus was produced as described in Venugopal et al. .
In vitro tumorsphere formation shRNA screen transduced cells were cultured in T150 flasks, propagated as secondary spheres, and divided into three replicates. Tumorspheres were passaged for four passages. At each passage, genomic DNA was isolated for sequencing.
In vivo screen 200,000 cells were injected intracranially into NOD-SCID mice (n = 10). At endpoint, brains were harvested, perfused, and genomic DNA isolated for sequencing.
Lentivirus was prepared for individual shRNA constructs similar to the pooled shRNA library. BMIC lines were transduced with lentivirus along with an shGFP control and treated with puromycin for 48 h prior to in vitro assessment of sphere formation, proliferation, and migration or in vivo injections by all three routes (ICr, IT, and ICa) to assess tumor formation and metastasis.
shRNA dropout screen analysis
A list of 150 genes involved in regulation of BM was expertly curated from the literature (Online Resource 1 Table 4). We used Wordle (http://wordle.net) to highlight biological processes associated with the 144 selected shRNAs (select genes in Online Resource 1 Table 5, full list in Online Resource 2). Using Pathway Data Integration Portal ver. 1 (pathDIP, http://ophid.utoronto.ca/pathDIP), we first performed pathway enrichment analysis across 19 major pathways databases and identified 532 pathways that are significantly enriched by the selected shRNAs (p < 0.05). Pathway names were corrected by removing uninformative words (e.g., Homo sapiens) prior to generating the word frequency figure (Fig. 3b).
In vitro analysis
Libraries were sequenced on an Illumina HiSeq 2500. Library preparation and sequencing were carried out as described previously . An average of 3.1 M reads were obtained per sample. Reads consisting of 21–22 nucleotides, followed by a XhoI restriction site, allowing a single mismatch, were retained for mapping. Reads were mapped against a library-specific, non-redundant FASTA file using bowtie (v 0.12.7) with the following command-line options: -l 21 -m 1 -n 2 --best –strata. On average, 99.5% of properly formatted reads aligned, allowing two mismatches and suppressing multi-mapping reads.
Mapped hairpins were enumerated from the resulting SAM file using a bespoke Java program, which also merges together all samples and hairpin annotations. Each sample was normalized by summing the total number of reads assigned to hairpins, dividing the sum by 1000,000, and then dividing the individual hairpin counts by the normalized sum. The resulting value is “reads per million mapped reads,” which is then transformed to log-base2.
Linear regression was performed for each hairpin to model the dropout trend over time. Normalized abundance counts were regressed against the passage number using the lm function in R (v 3.1.2). The majority of hairpins, as expected, had zero slope. The 5% of hairpins with the most negative slopes were selected for validation.
In vivo analysis
Sequencing and read mapping were performed as per the in vitro screen. Between 1.1 and 9.5 M reads (mean = 4.8 M) were processed per sample. Log fold-change values were computed for each hairpin in each in vivo sample by taking the ratio against the matching P0 hairpin abundance and transforming to log2. Only hairpins observed in both the P0 sample and at least one of the in vivo samples were retained for further analysis. Gene depletion was then ranked by comparing the aggregate fold-change value across all samples to the population of negative control hairpins using the Wilcoxon rank-sum test. P values were adjusted for multiple testing correction using the method of Benjamini and Hochberg. Finally, dropouts were ranked by the adjusted P value.
Immunohistochemical staining of patient samples for SPOCK1 and TWIST2
4 µm formalin-fixed paraffin-embedded sections were dewaxed in 5 changes of xylene and brought down to water through graded alcohols. Antigen retrieval was performed by pretreatment with Tris–EDTA (pH 9.0), then primary antibodies for SPOCK1 (Abnova #H00117581-M01; 1/2000) or TWIST2 (Atlas #HPA007450; 1/800) applied. Endogenous peroxidase was blocked with 3% hydrogen peroxide. The detection systems used were MACH 4 universal HRP polymer system (TWIST2, Intermedico Cat# BC-M4U534) and species-specific ImmPRESS polymer system (SPOCK1, Vector Labs). After following kit instructions, color development was performed with freshly prepared DAB (DAKO Cat# K3468). Finally, sections were counterstained lightly with Mayer’s Hematoxylin, dehydrated in alcohols, cleared in xylene and mounted with Permount mounting medium (Fisher cat# SP15-500). The optical densities (OD) of each sample were determined by densitometric analysis using ImageJ software.
Protein–Protein Interaction Mapping of SPOCK1 and TWIST2
Physical protein interactions (PPI) for SPOCK1 and TWIST2 and their interacting partners were identified using Integrated Interaction Database (IID) v.2016-03 (http://ophid.utoronto.ca/iid) . Protein interaction networks were visualized and analyzed using NAViGaTOR version 2.3 (http://ophid.utoronto.ca/navigator) . Corresponding XML file in NAViGaTOR can be downloaded from http://www.cs.utoronto.ca/~juris/data/NCB/BMnetwork.xml. We calculated centrality/betweenness by counting the number of shortest paths through a given node and edge, connecting all possible pairs of nodes in the network, resulting in all pairs shortest path count (APSP). All proteins in the network were selected for pathways enrichment analysis as described above, using Pathway Data Integration Portal (pathDIP) ver. 1 http://ophid.utoronto.ca/pathDIP).
Edge centrality was evaluated by betweenness, and nodes participating in the most central interactions were selected for further analysis of prognostic properties (all nodes, edges, and betweenness scores are listed in Online Resource 1 Table 8 and connection derivations in Online Resource 3).
Cox modeling of prognostic potential of top 20 genes identified through PPI network
To evaluate multivariate prognostic potential of the top 20 genes, we developed Cox proportional hazards model (SPOCK1, PLAT, APBB1, CELSR2, PLG, PLAU, PRSS3, PRKCQ, TCF4, INSR, TP53, ATXN1, EP300, JAG2, INHBA, LAMA3, ELAVL1, STMN2, VAV2, ARHGAP26), where the gene expressions served as only covariates. Model was derived using R’s package glmnet , applying ridge regression (also known as L2, or Tikhonov regularization) to prevent over-fitting. TCGA’s lung adenocarcinoma (ADC) RNA-seq data were standardized by converting to z scores and along with the corresponding clinical data were used as “training data.” Model was validated on the independent, publicly available lung ADC data (GSE31210) , and its prognostic performance was evaluated by concordance index (function survConcordance from R’s package survival , v 2.38.3) and area under receiver operating characteristics (ROC) curve—AUC, measured at the third year after initial time point (function AUC.cd from the R’s survAUC package, v 1.0.5). Resulting model then predicted the risk score for each patient in the validation dataset. Patients were grouped into two groups based on the predicted risk score, using the threshold that was obtained prior to validation to maximize the hazard ratio obtained on the training data (for more details see Royston et al. ). Validated hazard ratio (HR) between these two groups, as well as associated statistical significance (Log-rank test) were finally calculated (function survdiff from the survival package) and Kaplan–Meier (KM) plot showing time-dependent survival probabilities of these two groups was generated.
Generation of univariate gene expression Kaplan–Meier plots
Kaplan–Meier (KM) curves depicting patient survival by risk group based on SPOCK1 and TWIST2 expression in Fig. 6c were generated using SurvExpress, a suite for validation of biomarkers and survival analysis  using TCGA Lung Adenocarcinoma (TCGA-LUAD) dataset (version Sept. 2012) . KM curves were generated from quantile-normalized TCGA lung ADC expression data and matching information about patient survival.
Prognostic properties of the most central nodes were evaluated using http://kmplot.com (version 2015; data downloaded on March 6, 2016) . Only lung ADC samples were used, and biased samples were removed (n = 2437). Auto select best cutoff and censor at threshold was used, with JetSet probe selection. Obtained hazard ratios and corresponding P values are plotted in Fig. 7b. Resulting KM plots for overall survival are included in Fig. 7c and Online Resource 1 Fig. 11d; specifying probe sets used.
Replicates from at least three samples were used for each experiment. Respective data represent mean ± SD with n values listed in figure legends. Student’s t test and 2-way ANOVA analyses using GraphPad Prism 5. P < 0.05 were considered significant.
Patient-derived BMICs exhibit stem cell characteristics and enhanced migratory potential in vitro
Migratory capacity was variable between individual BMIC lines (Fig. 1c), but overall BMIC lines had enhanced migration compared to invasive GBM lines. Through analysis of each patient’s BMICs for known cancer stem cell (CSC) markers, we found differential expression of CSC surface markers CD133, CD15 , and CD44 [2, 10, 25], as well as epithelial cell adhesion marker (EpCAM)  (Fig. 1d, Online Resource 1 Fig. 2). Collectively, these data confirm the presence of a migratory CSC-like population in vitro and underscore the existence of patient-to-patient heterogeneity between BMIC lines. The variation of CSC expression levels between patients further confirms our previous work, where typical primary tumor CSCs markers are inefficient at identifying BMIC population  and emphasizes the necessity of novel BMIC markers.
Patient-derived xenotransplantation models effectively recapitulate human brain metastases in vivo
We have utilized our unique BMICs to establish appropriate patient-derived xenotransplantation (PDXT) models of BM that allow for serial in vivo enrichment and propagation of the functional tumor-initiating cell (TIC) population that initiates BMs. Firstly, BMIC tumor initiation capacity (TIC) was assessed in vivo through intracranial injections of two BMIC lines (BT478 and BT530) in NOD-SCID mice. This model was adapted from a PDXT model established by our lab for primary BTIC (Brain tumor initiating cell) population . In vivo LDA (Online Resource 1 Table 2) showed that BMICs were able to form tumors from as few as 100 cells (Fig. 1e). To evaluate if our BMICs retained original patient tumor marker profile, xenotransplants were stained with routine marker profiles commonly utilized to diagnose brain metastases from a primary lung cancer, and we found that xenografts recapitulated the original patient tumor histology and cyto-architecture (Online Resource 1 Fig. 3).
Secondly, to assess BMICs ability to complete different stages of the metastatic cycle in vivo, we injected two BMIC lines (BT478 and BT530) into NOD-SCID mice through two different injection routes: a) intracardiac injections (ICa) and b) intrathoracic injections (IT) (Online Resource 1 Table 3).
shRNA drop out screen identifies TWIST2 and SPOCK1 as novel genetic regulators of brain metastasis
RNA interference (RNAi) using short-hairpin RNAs (shRNAs) provides a versatile tool allowing for rapid interrogation of gene function in mammalian cells, and has been used to identify genetic regulators of tumorigenesis and metastasis of various solid cancers [32, 48, 56].
Knockdown of SPOCK1 or TWIST2 leads to inhibition of brain metastasis in vivo
BMICs were transduced with shGFP, shTWIST2, or shSPOCK1 and injected through the ICr, IT, and ICa routes. In ICr injections, which served as controls to validate tumor engraftment, shTWIST2 in BMICs reduced tumor volume when compared to the shGFP transduced control cells, while shSPOCK1 prevented tumor formation altogether (Fig. 5b, n = 2, p values not calculated). For IT injections, although BMICs with shTWIST2 did not reduce lung tumor formation, cells failed to metastasize to the brain. shSPOCK1 had both greatly reduced lung tumor formation and no BM (Fig. 5b). Mice receiving shSPOCK1 or shTWIST2 cells had a survival advantage over mice that received control cells through ICa injections. Moreover, knockdown of these genes completely ablated BMIC metastasis to the brain (Fig. 5b, c). Collectively, these data implicate SPOCK1 as a vital regulator of BMIC self-renewal and the development of brain metastasis, whereas TWIST2 was confirmed to be important to initiation of metastasis of BMICs.
Expression levels of SPOCK1 and TWIST2 in lung cancer patient samples are predictive of brain metastasis
SPOCK1 and TWIST2 expression were only observed in primary lung cancer specimens in patients who ultimately developed BM (Fig. 6b, SPOCK1 p < 0.01; TWIST2 p < 0.0003), and were also present in the corresponding patient-matched BM (n = 12). Interrogation of The Cancer Genome Database (TCGA) for SPOCK1 and TWIST2 expression in lung cancer patients determined that in a validation cohort of 255 NSCLC patients, high expression of these genes predicted poor patient survival (Fig. 6c, SPOCK1 p = 0.0036; TWIST2 p = 0.0001). These data suggest that SPOCK1 and TWIST2 may have clinical utility as predictive biomarkers of BM formation in NSCLC patients.
Protein–Protein interaction mapping between SPOCK1 and TWIST2 identifies novel pathway interactors
The invasive nature of BM and their ability to escape aggressive treatments predict uniformly poor patient outcome . Advances in screening, detection, and therapeutics for systemic cancers have led to an increase in cancer patient survival, yet leaves cancer survivors vulnerable to an increased prevalence of BM [23, 38]. Here, we have successfully developed BMIC lines from primary patient BM originating from lung cancer, where these samples represent an enriched source of human cells capable of completing the lung-to-brain metastatic cascade.
Syngeneic mouse models have significantly improved our understanding of molecular factors that govern subclone survival throughout stages of metastasis [3, 16, 18], though the clinical utility of these models is limited as transgenic murine models may not recapitulate the inter- and intratumoral heterogeneity of lesions in humans. Current models for studying metastasis investigate each stage in isolation, failing to capture the dynamic evolution of tumor subclones during primary tumor growth. Nonetheless, these models have contributed to the identification of metastatic drivers, such as ST6GALNAC5 , L1CAM  from breast cancer, and LEF1 and HOXB9  for lung cancer. When paired with RNAi screens, these models have aided the identification novel metastasis suppressor genes [15, 32].
Through our BMICs derived from patient samples of BM of lung origin, we have successfully generated an appropriate and feasible model of human BM. We build upon our previous work, where we had confirmed our BMICs to possess typical stem characteristics such as self-renewal and tumor initiation . In our current work, we show that our BMICs are capable of completing the complexities of metastasizing to the brain. Our unique PDXT model provides a unique opportunity to isolate and characterize BMICs as well as provides a novel platform to screen and validate BM treatments.
Our shRNA dropout screens performed with our in vitro BMIC and in vivo BMIC PDXT model systems identified SPOCK1 and TWIST2 as regulators of BMICs, with SPOCK1 in particular proving to be critical to BMIC self-renewal, tumor initiation, and migration. BMICs with knockdown of SPOCK1 and TWIST2 also displayed reduction in stemness as depicted by downregulation of CSC marker CD133 expression, suggesting a potential relationship between SPOCK1, TWIST2, and stemness. Functional validation of both genes by knockdown in our PDXT model resulted in complete abrogation of brain metastasis through ICa and IT injections. Interestingly, knockdown of SPOCK1 completely inhibited tumor formation in both the primary lung and secondary brain environments, implicating SPOCK1 as an essential regulator of tumor initiation irrespective of the microenvironment. The dramatic reduction in secondary sphere formation seen with shSPOCK1 further corroborated its role in governing BMIC self-renewal. Conversely, knockdown of TWIST2 had no significant effect on primary lung tumor formation but reduced BM, providing further validation for the role of TWIST2 in BMIC initiation of metastasis. Our data implicate SPOCK1 and TWIST2 in metastatic progression, and may prove to be useful as candidate genes in the detection of migrating BMICs or circulating tumor cells (CTCs). Development of anti-metastatic therapies has been limited due to a lack of understanding of the intricacies of the metastatic cycle, with most drugs targeting the end stages .
Systemic cancer progression follows two basic models: linear progression, where the metastatic cells develop within the primary tumor prior to dissemination to a secondary site, or parallel progression that is independent of primary tumor formation. Current research favors the parallel progression model, indicating a need to re-evaluate current treatment modalities . Additionally, recent studies have shown that the presence of metastatic CSCs correlates with the overall incidence of metastasis development in patients, and so offers as a novel pre-metastatic therapeutic option in metastasis prevention. However, targeting these cells faces several difficulties. For instance, the limited knowledge of the biology of metastatic CTCs challenges the ability to precisely identify this particular population , and conventional therapeutics have a short time in circulation that may result in inefficient exposure and consequently ineffective killing of CTCs . The identification of novel biomarkers of this metastatic population opens up a unique therapeutic avenue for prevention of metastatic development. Early detection and therapeutic targeting of CTCs based on SPOCK1 or TWIST2 expression could terminate the metastatic process at its initiation, ideally reducing the risk of BM development.
Given that not all lung cancer patients develop BM, the immediate clinical utility of our findings is in the ability to identify those patients who are more likely to develop a BM such that they may receive targeted therapy or an escalation in current treatment protocols of NSCLC. Interestingly, SPOCK1 and TWIST2 immunohistochemistry was only positive in those lung cancer patients who developed BM, whereas lung cancers that did not develop BMs were devoid of SPOCK1 and TWIST2 expression. These data suggest that SPOCK1 and TWIST2 may have clinical utility as predictive biomarkers of BM formation in newly diagnosed NSCLC patients.
Despite SPOCK1 and TWIST2 being implicated in other primary cancers and metastases [43, 58], to date no work has shown any relationship between these two genes. We have shown that although SPOCK1 and TWIST2 are co-expressed in a minority population of BMICs, they have no direct interaction. Through PPI networks analysis, we have identified several genes that connect SPOCK1 and TWIST2 indirectly and are specific to lung tissue. Several of these connecting genes have been implicated in neurological development  or the progression of neural diseases such as Alzheimer’s [8, 60] suggesting that these genes may predispose BMICs to home from the lung to the brain. None of these genes have been implicated in co-operating with SPOCK1 or in BM development. Of the top twenty connector genes with the highest centrality in the SPOCK1-TWIST2 PPI network (Fig. 7a), nineteen genes each had significant predictive potential for BMs in a primary lung adenocarcinoma patient cohort. When applied as a signature, these genes had an even greater impact on predicting poor patient survival. Expression of the most significant connector, INHBA, was correlated with SPOCK1 or TWIST2 expression, validating an interaction between these three genes. INHBA is a known TGFbeta ligand, where its binding to TGFB receptors promotes the downstream activation of the SMADs pathway . Activation of the TGFbeta pathway upregulates expression of several transcription factors known to regulate the epithelial-mesenchymal transition (EMT), such as TWIST2 . SPOCK1 has also been shown to be a TGFbeta downstream target gene, where its activity also promotes EMT . Inhibition of TGFbeta pathway through INHBA has been shown to reduce experimental metastatic development . From this circumstantial evidence, we suggest a possible interactive pathway of INHBA-activated TGFbeta regulation of SPOCK1 and TWIST2, promoting BMIC proliferation and metastasis through EMT.
The known involvement of TGF-β with SPOCK1 and INHBA implicates a role for the TGF-β signaling pathway in BM development [29, 41]. Our data present a novel group of genes interact with TWIST2 and SPOCK1 to influence BMIC metastasis and BM development in lung cancer patients. The multifaceted functional roles of BMIC genes in driving metastasis and the potent ability of BMICs to evade most therapies render BMIC regulatory genes as ideal therapeutic targets.
In conclusion, we have successfully developed a novel BMIC model system of human lung-to-brain metastasis through the establishment of unique patient-derived BMIC lines. Through our clinically relevant BMIC models, we identified SPOCK1 as a novel predictive biomarker of BM and critical regulators of the metastatic process for lung-derived BM. We have also identified a predictive prognostic signature of lung cancer patients and present novel therapeutic targets. Blocking the metastatic process would transform a uniformly fatal systemic disease into a locally controlled and eminently more treatable one.
This work was supported by funds from the Department of Surgery at McMaster University, Ontario Institute for Cancer Research Cancer Stem Cell Program, a Canadian Cancer Society Research Institute Innovation Grant and The Boris Family Fund for Brain Metastasis Research for Dr. Sheila Singh; Brain Canada PhD Studentship for Mohini Singh; Ontario Research Fund (GL2-01-030), McLaughlin Centre Accelerator grant (#MC-2013-05), Canada Research Chair Program (CRC #225404), and Canada Foundation for Innovation (CFI #12301, #203373, #29272, #225404, #30865) supported Dr. Igor Jurisica.
Conceptualization and design: MS, CV, SS. Methodology: MS, CV, NM, SM, DB, PV, MQ, TV, MD, KD, AT. Data analysis and interpretation: MS, CV, TT, KB, RH, JM, IJ, DK, SS. Writing, review, and/or revision of the manuscript: MS, CV, NM, SM, DB, PV, MQ, BM, JM, IJ, DSS. Supervision, JAH, JM, IJ, SS
Compliance with ethical standards
Conflict of interest
The authors declare no potential conflict of interest.
- 1.Aguirre-Gamboa R, Gomez-Rueda H, Martinez-Ledesma E, Martinez-Torteya A, Chacolla-Huaringa R, Rodriguez-Barrientos A et al (2013) SurvExpress: an online biomarker validation tool and database for cancer gene expression data using survival analysis. PLoS One 8:e74250. doi:10.1371/journal.pone.0074250 CrossRefPubMedPubMedCentralGoogle Scholar
- 45.Singh SK, Clarke ID, Terasaki M, Bonn VE, Hawkins C, Squire J et al (2003) Identification of a cancer stem cell in human brain tumors. Can Res 63:5821–5828Google Scholar
- 55.Waghorne C, Thomas M, Lagarde A, Kerbel RS, Breitman ML (1988) Genetic evidence for progressive selection and overgrowth of primary tumors by metastatic cell subpopulations. Can Res 48:6109–6114Google Scholar
- 59.Zakaria N, Yusoff NM, Zakaria Z, Lim MN, Baharuddin PJ, Fakiruddin KS et al (2015) Human non-small cell lung cancer expresses putative cancer stem cell markers and exhibits the transcriptomic profile of multipotent cells. BMC Cancer 15:84. doi:10.1186/s12885-015-1086-3 CrossRefPubMedPubMedCentralGoogle Scholar
- 60.Zhang C, Browne A, Child D, Divito JR, Stevenson JA, Tanzi RE (2010) Loss of function of ATXN1 increases amyloid beta-protein levels by potentiating beta-secretase processing of beta-amyloid precursor protein. J Biol Chem 285:8515–8526. doi:10.1074/jbc.M109.079079 CrossRefPubMedPubMedCentralGoogle Scholar