Systems Biology of Cancer Metastasis

Abstract

Cancer metastasis is no longer viewed as a linear cascade of events, but rather as a series of concurrent, partially overlapping processes, as successfully metastasizing cells assume new phenotypes while jettisoning older behaviors. Lack of a systemic understanding of this complex phenomenon has limited progress in developing treatments for metastatic disease. Because metastasis has traditionally been investigated in distinct physiological compartments, the integration of these complex and interlinked aspects remains a challenge for both systems-level experimental and computational modeling of metastasis. Here, we present some of the current perspectives on the complexity of cancer metastasis, the multi-scale nature of its progression, and a systems-level view of the processes underlying the invasive spread of cancer cells. We also highlight the gaps in our current understanding of cancer metastasis as well as insights emerging from interdisciplinary systems biology approaches to understand this complex phenomenon.

eTOC blurb:

Cancer metastasis is a complex disease, arising from a growing tumor from which cells escape to other parts of the body. For long, cancer metastasis was considered as a combination of steps, which were studied separately, limiting our understanding of this complex disease. Here we present the new developments, and our perspective on how the new systems biology approach is changing our view of cancer metastasis as an integrated multiscale phenomenon comprising of interlinked parts that allow tumors to metastasize.

Introduction

Cancer metastasis, the processes involving dissemination of cancer cells from a primary lesion to distal organs, is the principal cause of cancer lethality. Dissemination of cells from a primary tumor involves a variety of cellular mechanisms. These include invading through, or colluding with, stroma, escaping immune surveillance by inhibiting or co-opting their anti-tumorigenic processes, evading and modulating the tissue microenvironment, and evolving resistance to therapeutic intervention(Fischer et al., 2015; Kalluri, 2016; Li et al., 2016b; Massague and Obenauf, 2016). Recent reports provide strong evidence that metastasis is non-linear, and involves multiple parallel overlapping routes(Harper et al., 2016; Lambert et al., 2017; Te Boekhorst and Friedl, 2016). The reductive disease models that were necessary to establish the field of metastasis research and provide foundational concepts are limited in completely characterizing metastasis owing to the integrated and complex nature of its constituent processes. The multi-parametric and multi-scale nature of cancer metastasis warrants a renewed focus on comprehensive experimental and computational approaches that provide systems-level insight, versus the investigation of isolated steps in a complex network of events. Systems biology approaches that result in predictive and testable models of complex phenotypes through integration of expertise from diverse fields including cancer biology, oncology, genetics, mathematics, bioinformatics, imaging, physics, and computer science could provide a more holistic understanding of the complete metastatic process. In this Review, we present insights into the complexity of cancer metastasis, the multi-scale nature of its progression, and a systems-level view of investigating the processes involved in invasive spread of cancer cells. We highlight the gaps in our understanding of steps involved in tumor metastasis and insights emerging from interdisciplinary systems biology approaches to this important cancer process.

Cancer Metastasis: A Dynamic Selection Process

Cancer cells exist in a continuum of phenotypic states that facilitate transition between residence in the primary lesion to local invasion, systemic circulation, and eventual seeding and colonization of distal secondary sites (Figure 1). A metastatic lesion is the result of one or several cells acquiring the capacity to circumvent a series of molecular and biophysical hurdles, which would present unsurmountable obstacles to their non-metastasized counterparts (Aceto et al., 2014; Lambert et al., 2017; Wang et al., 2017). Notably, many of these capabilities may rely on cell behaviors that are mutually incompatible or incongruent with those needed to establish the primary tumor. For example, while cells in the initial tumor mass may be proliferative, many reports indicate that disseminating cells or cells undergoing an epithelial-to-mesenchymal transition (EMT) largely suspend proliferation (Rojas-Puentes et al., 2016; Vega et al., 2004; Zheng et al., 2015). Metastatic cancer cells exhibit an anti-correlation between the so-called “grow” and “go” phenotypic states; however, once a metastatic cell is arrested in a secondary site, it could resume proliferation, even after remaining dormant for prolonged periods(Giese et al., 1996; Hatzikirou et al., 2012). A comprehensive patient cohort study of metastasis revealed that proliferative cancers were associated with increased metabolism and stress response, while the ones with more EMT like phenotype were more inflamed(Robinson et al., 2017). Current and future experimental and computational modeling of metastasis must accommodate this plasticity while integrating the many steps in the metastatic process. These steps include the initiation of metastasis and local invasion, travel to and colonization of distant metastatic sites, and evasion of the immune system, often manifested by a state of dormancy (Figure 1).

Figure 1. Overview of the observed complex and concurrent routes of metastasis.

Metastasis is a complex, multiscale process which involves multiple sub-processes occurring in parallel through partially overlapping routes. Emerging evidence suggests that pre-malignant lesions are capable of giving rise to distant, latent metastasis and are thus not only associated with late stage primary tumors. However, it is still commonly thought that metastasis occurs mainly through dissemination from malignant lesions when microenvironmental stressors induce cellular reprogramming events that facilitate cellular migration and invasion towards more nutrient-rich niches. These stressors, associated with metabolic reprogramming, can trigger phenotypic changes in cancer cells to adopt more mesenchymal-like states that are not binary, but plastic, with the cells capable of sampling these dynamic states throughout the metastatic process. While this may advance a cell’s ability to metastasize, it is now appreciated that it is likely not the only mechanism by which metastasis occurs. Indeed, there are multiple parallel mechanism co-opted by cancer cells. Lymphatics and blood vasculature are the primary route of cell seeding into the common metastatic organs across cancer types (lymph nodes, liver, lung, bone marrow, and brain), though the tropism cancer exhibit for specific organs is still poorly understood. The combination of genetic and epigenetic changes, and interactions with the diverse milieu of cells in the host microenvironment, determines cancer cell survival and outgrowth.

Initiation of Metastasis

The notion that metastatic lesions are formed from cancer cells that have disseminated from advanced primary tumors has been substantially revised following the identification of disseminated tumor cells in the bone marrow of patients with early stage disease(Schardt et al., 2005),(Klein et al., 1999). This phenomenon has also been demonstrated in spontaneously metastasizing, autochthonous animal models of breast cancer, where cells shed from premalignant lesions survived in distal organs and later gave rise to micro-metastases(Harper et al., 2016). Similarly, early dissemination outside of the preneoplastic lesion has been illustrated in pancreatic cancer models(Muzumdar et al., 2016; Rhim et al., 2012). Presentation of micro-metastases in early stages of disease adds complexity to diagnosis, detection and ultimately, to patient treatment (Figure 1).

The introduction of technologies such as intravital live cell imaging and carefully engineered in vitro systems have revealed that parallel mechanisms can co-exist for cancer cells to escape from a primary lesion. For example, it is now appreciated that invasive programs are more diverse than initially understood, with evidence suggesting that single or clusters of cells could disseminate from the primary tumor mass(Aceto et al., 2014; Massague and Obenauf, 2016; Ye et al., 2015; Yu et al., 2013), and that these programs are both cell intrinsic(Alexander et al., 2008), and triggered in response to the extracellular microenvironment(Hofschroer et al., 2017; Wolf et al., 2007). Cancer cells are capable of morphologically adapting to the physical constraints of the microenvironment, including deformation of the rigid nucleus, which, in turn, can lead to chromosomal instability, altered gene expression, and metastasis(Bakhoum et al., 2018). Such invasion-induced chromosomal instability suggests a potential mechanism for increased genomic heterogeneity at sites of metastasis. The notion that epithelial cells must first adopt an exclusively mesenchymal phenotype to promote metastasis has also been recently disputed. Work using organoid cultures of primary breast cancer epithelial cells has shown that during collective epithelial cell migration, the leading cell remains positive for the basal epithelial marker cytokeratin 14 (K14)(Cheung et al., 2013). Similarly, intravital imaging has captured E-cadherin retained by clusters of cancer cells exiting the tumor(Friedl and Gilmour, 2009). Studies in animal models of pancreatic and breast cancer further suggest that suppression of EMT through genetic depletion of transcription factors implicated in EMT has no effect on the rate of metastatic disease(Zheng et al., 2015). In contrast, recent reports indicate that the EMT transcription factor, Zeb1 is required for pancreatic tumor metastasis(Krebs et al., 2017), and that post-EMT mesenchymal-like cells cooperate with epithelial cells to eventually become metastatic by paracrine signaling(Neelakantan et al., 2017).

The initiation of metastasis is not simply a cell-autonomous event but is heavily influenced by complex tissue microenvironments. It has long been recognized that interactions between cancer cells, stromal fibroblasts, endothelial cells, immune cells as well as alterations in tissue oxygen tension and the architecture of the adjacent extracellular matrix profoundly impact tumor progression. Stabilization of the hypoxia-inducible factor (HIF), a transcription factor, can trigger a switch from collective migration to amoeboid migration as the oxygen tension fluctuates in the tumor, stimulating reciprocal signaling between mesenchymal stem cells and cancer cells promoting metastatic phenotype(Lehmann et al., 2017),(Chaturvedi et al., 2013). Tumor associated macrophages (TAMs) can be stimulated by cancer cell-secreted lactate to promote angiogenesis(Colegio et al., 2014), a requirement for distant metastasis, and more recent findings suggest that TAMs induce early dissemination of Her2+ breast cancer cells(Linde et al., 2018). Generally, the influence of the microenvironment on cancer cell invasion and migration are both cell autonomous and tissue context-dependent, adding a further layer of complexity to this process(Spill et al., 2016). Understanding the complex interactions between cancer cells and the tumor microenvironment that lead to metastasis will require integration of extensive molecular characterization data collected from in vitro and in vivo experimental models. An example is the use of data-driven modeling of protein phosphorylation in pancreatic cancer, which facilitated understanding of reciprocal molecular interactions between cancer cells and stroma, allowing system-wide delineation of the role of heterotypic cell-cell interactions in tumor growth(Tape, 2016; Tape et al., 2016).

Metastatic Colonization and Cancer Outgrowth

Colonization of distant tissues by disseminated tumor cells is an extremely inefficient process. While relatively numerous circulating tumor cells (CTCs) are detected in the blood of cancer patients, with reports indicating > 1000 CTCs/ml of blood plasma, disproportionally few metastases are clinically detectable(Nagrath et al., 2007). Following arrest in the vascular bed, a successfully metastasized cell has to extravasate and survive in a new tissue microenvironment that may or may not be conducive to survival. Overt metastatic lesions are primarily detected in select organ sites (liver, lung, bone, brain) but rarely in others (kidney, heart, stomach) (Figure 1)(Fidler, 2003). Therefore, metastatic colonization is not merely an outgrowth of rogue cancer cells from the primary organ, but arises via complex interplay between disseminated cancer cells and tissue microenvironments across the organism.

While little is known about the preference of cancer subtypes for distinct tissues, or about the receptiveness of a tissue as a metastatic site, various efforts are being pursued to further our understanding of such tissue tropism. A parabiosis mouse model approach has uncovered a strong preference for ovarian cancer hematogenous metastasis to the omentum regulated by specific ligand-receptor interaction between the two compartments(Pradeep et al., 2014). Meanwhile, downregulation of the metastasis suppressor RARRES3 facilitates breast cancer tropism to the lung by increasing cellular adhesion to lung parenchyma(Morales et al., 2014). Circulating cytokines and growth factors, as well as microRNA loaded exosomes are hypothesized to contribute to conditioning of pre-metastatic niches (Figure 1). Animal studies have uncovered pro-metastatic roles for exosomes via their effects on increasing vascular permeability and their priming of the resident cells of the metastatic site to create a pro-inflammatory and metabolically active niche(Becker et al., 2016; Costa-Silva et al., 2015; Schillaci et al., 2017). Recent reports have also suggested role of exosomes in tissue tropism, via specific integrin receptors on the exosome(Hoshino et al., 2015). However, our understanding of the systemic role of exosomes, their interactions with recipient cells, the duration of the effect, and whether they specifically target distinct cells or organs remains incomplete. Future systems biology models of metastasis, both experimental and computational, will benefit from explicit incorporation of intercellular communication which could provide a way to integrate dynamic multi-scale characteristics of the metastatic cascade. This approach may lend deep insight into the role of the primary tumor in determining tumor tropism, as well as metastatic niche formation.

Cancer Dormancy

What endures as one of the most confounding clinical phenomenon is that patients may undergo tumor resection and then remain apparently disease-free for months, years, and even decades only to relapse and be diagnosed with late stage metastatic disease(van Maaren et al., 2016). This may be a result of cell seeding from minimum residual disease after resection of the primary tumor or preexisting clinically undetectable micrometastases, but may also arise from early-disseminated cells that have remained dormant and resistant to therapy until suddenly reawakened to initiate proliferation into clinically detectable macrometastases. Emerging experimental observations note that dormant metastatic cells that later develop into overt lung metastases disseminated during the early stages of primary tumor development(Harper et al., 2016). Hypoxia within the primary tumor microenvironment may also predispose a sub-population of to-be-disseminated cells to dormancy(Fluegen et al., 2017). Where dormant cells reside and how they maintain their cell cycle arrested state is being explored through interdisciplinary approaches that combine live-cell imaging, laser capture microdissection and single cell transcriptomics. Some studies focus on the cell-intrinsic state of the dormant cancer cell, such as their epigenetic state, while others provide evidence for elements of the microenvironment, such as the resident quiescent vascular cells, for maintaining adjacent disseminated tumor cells in a dormant, growth arrested state^53,54,(Ghajar et al., 2013). Cellular latency can also emerge via the stochastic switch to self-imposed quiescence of a fraction of the cells, for example, by inhibiting WNT pathways, as the proliferative population is eliminated by immune surveillance(Malladi et al., 2016). Immune surveillance induced mass dormancy, therefore, can give rise to cellular dormancy, with macrometastatic growth occurring due to either the removal of immune pressure or acquiring other evasive traits.

Questions remain about how the mechanisms controlling dormancy are disrupted to allow for re-emergence and metastatic outgrowth years or decades later. Whether cell-intrinsic, microenvironmental and/or systemic changes related to normal physiological processes such as aging or diet impact dormancy is actively being investigated. Astrocyte-derived exosomes suppress PTEN in dormant metastatic cells, allowing for the outgrowth of lesions in the brain(Zhang et al., 2015), thus demonstrating a role for the microenvironment in guiding the reactivation of dormant cancer cells. Other studies have provided evidence that the primary origin of dormant cells may also play a role. For example, cancer cells from organs of endodermal lineage that cycle more slowly (i.e. liver, pancreas, lung) may behave differently than cancer cells originating from sites with persistent cellular turnover such as the colon or gut(Furukawa et al., 2015; Taylor et al., 2013; Wells et al., 2013). In addition, the unique characteristics of the metastatic site will likely impact reactivation of disseminated tumor cells arriving in that organ. Cells assuming latency by avoiding immune surveillance by WNT signaling induced quiescence can stochastically proliferate if immune surveillance is removed(Malladi et al., 2016). Indeed, it is a common observation that after resection of primary tumors, which typically serves as the chief source of paracrine signaling creating pre-metastatic niches, already metastasized cells remain in dormancy to be activated many years later. Mechanisms causing emergence from latency are not yet well understood, but in the absence of pre-metastatic niches, a regrowing tumor may also signal to create new niches. A recent study showed that inflammatory signals like IL-6 could make liver a more favorable pro-metastatic environment for pancreatic cancer(Lee et al., 2019). Considering that the latent cell has already gone through the process of metastasis and is now also proliferating, it is much more primed to create secondary node than a primary tumor cell would be, and therefore a reoccurring disease is probably more aggressive. Resolving the question of when cancer cells could take advantage of the pre-metastatic niche versus entering a dormant state will involve detailed mapping of the intercellular signaling of primary, quiescent, and metastatic cells with the stroma and the cell state trajectories during this communication.

It is tempting to consider that tumor cell dormancy follows an evolutionary conserved mechanism adapted from other organisms and is launched in response to stress(Carcereri de Prati et al., 2017; Seidel and Kimble, 2011; Senft and Ronai, 2016; Sosa et al., 2013),(Herrick et al., 2017; McGillivray et al., 2015; Pal et al., 2016; Schubert et al., 2015). Genetics, tissue microenvironment, and timing are almost certainly some of the factors that impact a cell’s capacity to enter and exit from dormancy. The development of therapeutic strategies targeting mechanisms that facilitate dormant cell reactivation could prevent expansion of disseminated tumor cells or minimally residual disease. However, prior to drug development efforts, a multi-parameter approach is necessary to better understand how inhibiting the exit from dormancy may impact normal physiological processes-.

Systems Approaches to Studying Metastasis

Understanding how the complex molecular-level behavior of cancer cells and their interactions with the tumor microenvironment lead to metastasis will require integration of physiological metastasis models and extensive phenotypic and molecular characterization. Sophisticated informatics and computational approaches will be necessary to make sense of these dynamic and multivariate relationships and to generate testable hypotheses that eventually lead to better patient treatment options (Table 1). Also shown are both experimental and computational systems approaches to study cancer metastasis at different stages of progression (Table 2 and Figure 2). Implementation of such approaches will involve transforming the concept of individual events in metastasis to an integrative, multi-scale process characterized by system abstraction as shown in Figure 3.

Table 1.

A consolidated list of experimental and computational techniques utilized in Systems Biology of Cancer Metastasis.

Systems Biology of Cancer Metastasis: Experimental and Computational Approaches
	Advantages	Cautionary Notes	Requirements
Experimental techniques
Single cell RNAseq (Bush et al., 2017; Gierahn et al., 2017; Leung et al., 2017; Ting et al., 2014; Tirosh et al., 2016)	Useful in studying cellular heterogeneity, state of signaling with high confidence, identification of small subpopulations driving a phenotype	Dropouts in read-counts mean that sufficient number of cells are required to gain confidence; different analysis methods from bulk RNAseq; hard to study very rare subpopulations; need for dissociation of single cells can markedly alter transcriptomics	Dissociated cells
Cellular lineage tracing during metastasis (Sikandar et al., 2017)	Direct verification of rare metastatic driving events, and cellular identity of drivers of invasion
Patient derived xenograft models (Jespersen et al., 2017; Jiang et al., 2015; Rongvaux et al., 2017; Whittle et al., 2015; Zheng et al., 2018)	Possible to focus on patient specific tumor	Mostly used for drug screens, rather than mechanistic studies, requires adequate time to observe any metastasis	Access to patient tumor cells, humanized mouse models
Systemic perturbation libraries: CRISPR/Cas9 (Chow et al., 2017; Konen et al., 2017; Shen et al., 2017)	Able to find direct causal mutations for survival, growth, or metastasis	Differential cell proliferation in cells in library can confound phenotypic enrichment assessment	Unbiased gene knockout libraries
Computational techniques
Mechanistic Modeling of biochemical/genetic pathways: ordinary/partial differential eqns (Heinrich et al., 2002; Jia et al., 2017; Jolly et al., 2016; Kirouac et al., 2013; Kochanczyk et al., 2017; Korkut et al., 2015; Nguyen et al., 2013; Olivenca et al., 2018; Park et al., 2017; Ryu et al., 2015)	Detailed mathematical modeling, the effect of every possible parameter can be investigated	Usually limited by the known quantitative parameters to simulating systems of a few constituent proteins.	Quantitative parameters such as reaction rate constants, protein concentrations, degradation rates from experiments
Higher-level pathway modeling: Petri nets (Pennisi et al., 2016), Boolean networks (Lu et al., 2015; Steinway et al., 2014), Bayesian networks (Friedman et al., 2000)	Larger systems can be simulated, by increasing the level of abstraction where detailed parameters are not available.	Only higher level details are captured	Known parameters from previous studies, or expression, proteomic, metabolic etc. data to fit parameters, or a combination of the two.
Reconstructed genetic regulatory networks (Ahmad et al., 2012; Liang et al., 2012; Walsh et al., 2017)	Ability to integrate large scale and sometimes multimodal experimental data	Limited to transciptomic processes, ignoring metabolic and proteomic regulation
Whole cell metabolic network modeling (Aguilar et al., 2016; Tsai et al., 2009; Wu et al., 2017): Flux balance and constraint-based reconstruction and analysis (COBRA) (Becker et al., 2007)	Able to integrate metabolomic and genomic data	Only allows analysis of processes that are limited by metabolic constraints	Knowledge of metabolic rate constants and identity of genes catalyzing each metabolic reaction
Multicell spatial simulations (Blinov et al., 2017; Ghaffarizadeh et al., 2018; Ibrahim-Hashim et al., 2017; Pennisi et al., 2009; Robertson-Tessi et al., 2015; Scott et al., 2013) with hybrid cellular automaton and agent based models	Can model spatial effects, cellular interactions during invasion and growth	Limited detail for processes within the cell	Parameters for intercellular interactions, spatial cell location and size distributions
Systems biology model description standards (Hucka et al., 2003; Konig et al., 2012; Le Novere et al., 2006; Lopez et al., 2013): SBML, pySB etc.	Ability to use models from many different studies, and integrate the different aspects such as tumor, normal, and stromal cells	The flexibility of SBML implies that published models may use different levels of abstraction	Databases of SBML models, experimental data
Computer vision/ deep learning for histological and fMRI images in metastatic lesions(Litjens et al., 2016; Liu et al., 2017b; Wang et al., 2016)	Not limited by human expert time, able to analyze and segment/label large number of images	Limited to the analysis/labeling that has been learned. Possibility of errors for individual cases.	Access to large datasets and computational resources, while learning the model. Inference is computationally inexpensive and does not require large datasets.
Interpretable machine learning models (Broecker et al., 2017; Hua et al., 2006; Ma et al., 2018; Michael et al., 2018; Samek et al., 2017)	Able to generate non-trivial causal, testable hypotheses from raw data	Not always easy to incorporate previous knowledge	Need to develop/customize non-standard algorithms. Large, systematic experimental data

Table 2: Systems Methods to Understand Physiological Stages of Metastasis.

Physiological stages of metastasis present unique questions, as well as experimental avenues to explore the progression of metastasis.

Systems Methods to Understand Stages of Metastasis
Stage	Techniques
Tumor Dissemination	i. Organoids: Creating spheroids mimicking primary tumor mass to study early dissemination events. Variations typically include genetic perturbations in tumor cells, microenvironmental factors (rigidity, matrix etc.), metabolic environ, as well as presence of soluble factors (Cheung and Ewald, 2016; Cheung et al., 2016; Lehmann et al., 2017; Zanoni et al., 2016). ii. SC-Seq: RNA profiling at single cell resolution to study tumor evolution, test for placement on phenotypic continuum of dissemination iii. Spatial Seq: Laser ablated, or spatial deconstruction by other means, to identify spatial correlates of altered transcriptome in the spheroid. iv. Invasion Parametrization: Organoid characterization (e.g. shape analysis, biosensor activities) driven by microscopy for high throughput screening, or mechanistic insights into dissemination (Shirure et al., 2018) v. Intravital imaging: Owing to a predetermined site of orthotopic xenotransplant to study this stage, intravital imaging is useful in studying early disseminating events (Ilina et al., 2018; Vennin et al., 2016) vi. Computational Modeling: Rich parameters are now available to mechanistically model signaling pathways to understand cellular state in the phenotypic continuum (Hatzikirou et al., 2012)
Stromal Invasion	i. Intravital imaging: Imaging live in an animal through a window (sometimes in reporter hosts) to observe interaction of disseminating cancer cells with the stromal compartment (Entenberg et al., 2017) ii. Ligand-Receptor Association: Typically through SC-Seq, wherein expression of ligands and receptors could identify heterotypic interactions in stroma (Tirosh et al., 2016) iii. Stromal invasion assays: Quantitative methods to systematically study genetic and environmental drivers of stromal invasion (Kshitiz Gupta, 2019) iv. Stromal secretome analysis: Single cell secretome analyses by immune and other cell type in stroma can reveal cancer-stroma crosstalk (Neelakantan et al., 2017; Xue et al., 2015) v. Spatiotemporal Computational Modeling: Heterotypic cellular interactions can be modeled using spatially defined modeling with partial differential equations, or cellular automatons (Tape, 2016; Tape et al., 2016)
Intravasation/Extravasation	i. 3D vasculature models: Tissue models to study cancer-vasculature interactions(Zervantonakis et al., 2012) ii. Permeability measurements: Study effect of cancer secreted factors on vascular permeability in high throughput (Harney et al., 2015; Uhl et al., 2018) iii. Phenotypic screens of intravasation: Barcoded KO libraries selected for intravasation potential (Jeon et al., 2014; Zervantonakis et al., 2012) iv. Microfluidics: Approaches to mimic vasculature and lymphatics, and incorporate defined mechanical parameters into observation (Shirure et al., 2018)
Premetastatic Niche	i. Exosomes: Microvesicles secreted by cancer cells bearing microRNA and other components, which could prepare niches in other tissues for micrometastases (Costa-Silva et al., 2015; Hoshino et al., 2015). ii. Highly multiplexed imaging cytometry: Spatially resolved histological analysis with CyTof(Giesen et al., 2014). iii. Biomimetic niches: Biomaterial based approaches to study cancer cell interactions with perturbed environment mimicking premetastatic niches (Aguado et al., 2018)
Secondary Metastasis	i. Staging analyses by Genomics: Understanding tumor subtypes and staging by genomic profiling in comparison to histological profiling (Court et al., 2015; Enokida et al., 2005; Goldsmith et al., 2019; Villani et al., 2017) ii. SC-Seq/Metabolomics/Genomics: Characterization of the metastatic nodes at multiple granularities and scale can reveal the evolution of secondary node formation (Enokida et al., 2005; Goldsmith et al., 2019) iii. Single cell secretome: Identifying the molecular communication between neighboring cells and metastatic cancer cells (Blanco et al., 2012; Li et al., 2016a) iv. MRS/PET: functional imaging can reveal environmental factors correlating with metastasis (Herzog et al.,2013; Koush et al., 2019)
Dormancy/Senescence	i. Tissue/ whole organism multiplexed histology: Identifying dormant cells unaffected by therapy (Qutaish et al., 2018; Roy et al., 2009) ii. SC-Seq: Comparative transcriptomics to reveal senescent, stem-like, or dormant cells(Marlow and Dontu, 2015; Milanovic et al., 2018) iii. In vitro, or ex-vivo models: Studying emergence of dormancy by ex-vivo tissue culture models (Clark et al., 2018; Marlow and Dontu, 2015; Wheeler et al., 2014) iv. Mechanistic computational models: Cellular automata, and other models to predict emergence of a senescent phenotype (Poleszczuk et al., 2015) v. Cell line variants: Selection of clones from metastatic sources showing latency (Marlow and Dontu, 2015)
Drug Resistance	i. Multivariate statistical analysis and regimen optimization: Systemic variation of drug regimen, combination etc to study a response (Korkut et al., 2015; Lee et al., 2012) ii. Organs on chip: Miniature tissue mimetics to systemically test drug response in varied microenvironments (Li et al., 2019; Ronaldson-Bouchard and Vunjak-Novakovic, 2018; Saengwimol et al., 2018) iii. Cancer specific drug kinetics prediction: Systems analysis of drug response by simulating signaling or metabolic state of cells (Kochanczyk et al., 2017; Persi et al., 2018) iv. Deep learning and machine learning: Artificial intelligence has found many applications in network analysis, compound property and prediction of drug activity (Camacho et al., 2018; Chen et al., 2018)

Figure 2. Systems techniques to study disease progression at different steps in the metastatic cascade.

Tumor outgrowth from the primary node to a more metastatic phenotype entails different physiological environs which pose different experimental and analytical constraints for data acquisition, visualization, sample acquisition, requiring tailored approaches to explore metastasis. Similarly, analytical techniques to mechanistically understand cancer progression at different physiological stages also differ.

Figure 3. Integrated Models to Study Cancer Metastasis.

To understand and describe cancer metastasis at the multiple scales it exists in as a disease require integration of its characteristics at multiple scales. Technological developments in sequencing, imaging, immunological assays etc. have enabled integrated collection of data at multiple scales in which cancer metastasis manifests (molecular, cellular, tissue-level, organ-level, epidemiological, and clinical), as well as along the steps involved in the metastatic cascade. Our ability to describe events on those scales will need to be integrated to develop a more holistic and systems understanding of cancer metastasis.

Computational Systems Biology Approaches to Metastasis Research

The biological mechanisms underlying cancer metastasis occur at multiple biological scales. The ability to describe events across scales is needed to understand tumor cell dissemination in an integrated manner. For a metastatic tumor cell, these scales include genetic and epigenetic alterations, protein-protein interactions, and metabolic requirements, which together control various molecular signaling networks responsible for autonomous progression to metastasis. At the tissue or organ level, homotypic and heterotypic interactions between cells, and with the microenvironment further increase the complexity of the disease phenotype.

The multi-scale perspective of cancer metastasis requires an appreciation that the phenotypes at one scale are likely informed by regulatory events at other scales. Here we provide a perspective on some of the computational approaches used to model cancer metastasis, including emerging data analysis techniques that are becoming tools in the arsenal of many systems biologists.

Mechanistic Models of Metastasis – towards linking mechanisms at different scales

Computational biology provides mathematical frameworks that represent fundamental biological processes for the purposes of generating testable, and often non-intuitive predictions about disease mechanisms. An optimal computational model of metastasis will derive observations not accessible by current experimental technologies, inspire clinically important testable hypotheses, and facilitate deeper understanding of the mechanisms of metastasis. Metastasis operates across multiple spatial and temporal scales. Therefore, individual models constructed at one scale (for example, a static picture of gene regulatory networks operating within a cell, the dynamic simulation of a signaling pathway promoting cell motility, or the spatial representation of invasion of cancer cells into the stroma) will need to be integrated into a multiscale framework representative of a systems understanding of metastasis.

Mechanistic modeling based on ordinary or partial differential equations has been used to study various aspects of the metastatic processes, such as using the genetic, transcriptional and metabolic networks to build the signaling pathways leading to cell migration and invasion. The combination of experimental observations and differential equation based modeling of signal transduction has revealed mechanisms by which extracellular ligands promote tumor growth and invasion(Kirouac et al., 2013; Kochanczyk et al., 2017; Korkut et al., 2015; Park et al., 2017; Ryu et al., 2015). A generalized mathematical model of protein kinase signaling, crucial in regulating various steps of cancer metastasis, revealed that phosphatases have more pronounced effect than kinases on the rate of duration of signaling(Heinrich et al., 2002). On these lines, individual subsystems and processes have been successfully modeled in considerable detail. For example, a recent mechanistic model was able to predict the stability of a partial epithelial-mesenchymal transition (EMT) in vivo based upon the in silico prediction that additional molecular participants in the core EMT machinery must exist to promote hybrid EMT phenotypic behaviors(Jia et al., 2017; Jolly et al., 2016). A comprehensive model of the phosphoinositide pathway could contribute to the understanding of cell polarity and the initiation of chemotaxis and invasion(Olivenca et al., 2018). In this model, all species of the pathways are modeled, along with their spatial localization within the cell, suggesting strategies to control the activities of various molecular species within the pathway. Similarly, a dynamic model of hypoxia inducible factor-1 alpha (HIF-1a) based signaling has revealed key features of HIF-1a stabilization, and its effect on transcriptional activity of the cell(Nguyen et al., 2013).

Mechanistic models have provided detailed understanding of many signaling pathways, augmented by parallel developments in biosensors, activity reporters, microscopy, microfluidics etc. However, complex phenomena like metastasis involve interactions of many more molecular species than can be observed and modeled directly, and therefore more systems approaches are necessary which do not rely on detailed observations of all the molecular species involved in the described process. Large scale cell signaling networks with quantitative reaction constants could be used for the predictive modeling of metastasis. However, the development of such genome wide networks is limited both by the ability to simultaneously observe the activity of large numbers of signaling molecules, and the computational power required to infer and simulate these networks across cells and for extended periods of time. At the gene and transcript levels, most computational approaches are informed by large-scale ‘omics studies. Modeling of larger signaling networks may involve abstraction at a higher level, using methods such as Petri nets(Pennisi et al., 2016), Boolean networks(Lu et al., 2015; Steinway et al., 2014), Bayesian networks(Friedman et al., 2000) or systemic perturbation-effect networks(Korkut et al., 2015).

Reconstructed gene regulatory networks were used to identify the regulatory program inducing metastasis in breast cancer(Ahmad et al., 2012; Walsh et al., 2017), and the role of BACH1 in bone metastasis(Liang et al., 2012). Similarly, metabolomic networks can be modeled at the whole cell level using flux balance and constraint-based reconstruction and analysis (COBRA) methods(Becker et al., 2007), predicting the genetic bottlenecks within the energy utilization program of a cell. Such analyses have been used to infer the mechanism of the oncogenic Warburg effect for the pro-metastatic microRNA-122(Tsai et al., 2009), and identify the reprogramming of adaptive metabolic and transcriptomic profiles in metastatic cancer stem cells(Aguilar et al., 2016; Wu et al., 2017).

A crucial feature of cancer metastasis, and many other biological processes, is the existence of biological hierarchies, or scales. To accommodate hierarchies inherent in cancer metastasis, more discrete models have been employed, including cellular automaton and other agent-based modeling approaches. Cellular automata models typically involve two- or multi-dimensional lattice of entities with a finite number of states, wherein latticed entities can interact with each other through an iterative process defined by a fixed set of rules. The automaton also lends itself to integrating with ODE based models within each entity, referred to as Hybrid Cellular Automaton Models. Similarly, agent based computational modeling has been used to model biological systems wherein each biological entity is referred to as an agent, with a set of attributes, which interact in a rule framework. Agent-based computational modeling of the tumor microenvironment that incorporates spatial diffusion of oxygen, nutrients and cytokines offers a platform for high-throughput hypothesis testing in silico and can help to explain intra-tumoral heterogeneity, non-uniform drug transport, and disparate therapeutic response under hypoxic and low pH conditions(Waclaw et al., 2015),(Ibrahim-Hashim et al., 2017; Robertson-Tessi et al., 2015). Agent based models allow easy compartmentalization, quintessential in biological systems at all scales, and incorporate the complexity of association and dissociation between biological entities(Blinov et al., 2017). As an example, an agent based simulator to model 3D multicellular systems has been created incorporating biophysical principles(Ghaffarizadeh et al., 2018). Another model, combined with spatial data generated from microfluidic metabolic secretion detection chambers showed that spatial organization of cells in a tumor creates spatial metabolic signatures, which drive macrophage specification and angiogenesis(Carmona-Fontaine et al., 2017). Agent based models can also be utilized to test hypotheses which are otherwise difficult to access experimentally. For example, competing hypotheses of self-seeding versus primary metastatic seeding were tested using agent based models: where circulating tumor cells repopulate the primary tumor, or secondary seeding: where metastasized cells from primary tumor form secondary nodes which return back to the primary tumor via vasculature(Scott et al., 2013). Because they replicate real biological cellular entities, hybrid automatons or agent based models can be used to model complex cell-microenvironmental interactions in a detailed mechanistic fashion in the context of metastasis(Pennisi et al., 2009; Wang et al., 2015). In addition to the biochemical signaling pathways, metastasis also involves interactions between cells in both the tissue of tumor origin, or in distal metastatic nodes. While a detailed systems understanding of the multiple steps involved in this process is lacking, there have been attempts at stochastic modeling of some of these steps at various levels of detail(In et al., 2017; Newton et al., 2012; Speer et al., 1984). While stochastic modeling may not completely describe the mechanism, it does provide an approach to generate hypotheses for metastatic site selection guided by clinical observations. Mechanistic models have been successful in describing biological phenomena where parameters are known, or can be optimally searched in a constrained manner, e.g. to model signaling motifs at subcellular levels, or using cellular automata showing interactions between a few cells. The objective is to integrate these sub-systems models to create a systems understanding of metastasis. However, such integration of sub-system models may not capture all the emergent behaviors at larger scales.

The field of systems biology could benefit from the introduction of model formalisms, the absence of which has constrained multi-scale integration of the various sub-system models. There have been attempts to formalize and describe biological interactions at different scales. Systems Biology Markup Language is a complex multi-paradigm XML-based standard for model description(Hucka et al., 2003). Multiple software packages use SBML or a subset of its capabilities for simulation, visualization, or editing systems biology models(Konig et al., 2012; Lopez et al., 2013). The Biomodels database has made more than 8000 SBML databases available for use by other researchers(Le Novere et al., 2006). While SBML is versatile and capable of representing models developed in many systems biology studies, emerging experimental knowledge is not necessarily used to continuously refine a complete and unified model covering all aspects of cancer invasion and growth. Integrative frameworks and platforms have recently been developed and adopted, including PySB(Lopez et al., 2013), and natural language processing based formalism to create automated assembly of signaling formalism from word descriptions(Gyori et al., 2017). The goal of a systems integration of the metastatic cascade would be served well if cancer biologists and systems biologists arrive towards a common formalism to describe interactions of molecules, or cells in the context of their localization and states.

The multi-scale perspective of cancer metastasis requires an appreciation that the phenotypes that manifest at one scale are likely informed by regulatory events at other scales. As an example, relevant to clinical applications, the genetic landscape of cancer determines epidemiology, while cellular and tissue level scales involving interaction of cancer cells with matrix and other cell types inform the histological outcomes. For example, technologies like direct tissue proteomics allow high resolution proteomics from paraffin-embedded microdissection biopsies. Using laser-ablated microbiopsies, this approach identified more than 400 prostate specific genes as well as distinct metabolic pathways which may contribute to progression of prostate cancer at different stages(Hwang et al., 2007). Moreover, laser capture micro-ablation, combined with single cell transcriptomics, or even flow cytometry, are able to connect a tumor’s architecture with its molecular scale(Joseph and Gnanapragasam, 2011). In another example of bridging across scales, mathematical modeling was used to present an eco-evolutionary perspective of metastasis, explaining how mutations could facilitate evolution of cancer cells to break tissue homeostasis, and in turn transform the tissue environment itself(Basanta and Anderson, 2017).

Statistical Network Models – going beyond single gene associations in cancer metastasis

Systems biology is often referred to as networks of networks, integrating the knowledge of various interacting species across different biological scales. The era of big data collection has moved beyond sequencing of the genome, or the transcriptome, and now extends to identification of large number of metabolites, proteins with post-translational modifications and high content information about their subcellular localization. Measurement of these multi-omics has prompted the use of more sophisticated bioinformatics techniques that go beyond estimating the importance of individual genes to understanding how networks of molecules cooperate to promote cancer metastasis. Network information provides a convenient format for researchers to incorporate previous biological knowledge to generate hypothesis. These networks can be genetic, signaling, or metabolic, and operate within a cell or between cells. Whole genome sequencing has facilitated attempts to identify causal genetics of metastasis, but its integration with gene regulatory networks can further elicit specific gene interactions mediating metastatic phenotypes(Leiserson et al., 2015),(Paull et al., 2013). The large number of passenger mutations in metastatic tumors, and relatively limited number of patient samples poses a challenge, but heuristics applied to interactome network can allow prioritization of genes for further investigation. Network approaches have been applied to protein interaction, regulatory, gene co-expression networks, as well as on mutation co-occurrence to provide systems molecular interaction snapshots of metastasis(Shin et al., 2017). In addition, systems network analysis can be used to generate falsifiable hypotheses linking disease to holistic systems properties, as has been shown for metastasis as a progression towards network entropy(Teschendorff and Severini, 2010; West et al., 2012).

Another level of complexity emerges from the multiple interdependent molecular snapshots of a metastatic cell: the transcriptomic state informs proteomic changes, which would then inform the metabolomics states of cells. Although techniques have emerged to get single cell snapshots in all these states, these snapshots are only partially predictive of phenotypes. As an example, highly resolved multiple drug treatment regimen and high density observations in cancer cells combined with computational modeling revealed that sequential administration of drugs re-wires signaling cascades leading to enhanced apoptosis(Lee et al., 2012; Miller et al., 2016). As the quality of interaction data improves and more sophisticated methods are devised, it should be possible to combine multiple types of interactions. For example, understanding how the signaling networks within migrating cells in metastasis integrate with the host cells using ligand receptor interactions can shed light on the functional impact of some driver mutations. Understanding the network of post-translational modifications in signaling in conjunction with the transcription factor-target regulatory interactions can help delineate the differences between the effect of mutations in protein coding regions, cis-regulatory regions, and changes in gene expression causing metastasis.

Different approaches in machine learning, e.g. learning data representations, or deep learning have also begun to derive useful inferences from experimental observations, which could serve to parametrize and scale current models of metastasis. Although deep learning has been used mostly in the fields of vision, architectures like recursive neural networks, or their variants, can be used for knowledge extraction since they are more attuned for sequential information often present in omics data(Min et al., 2017). Machine learning approaches are also well suited to glean insight from heterogeneous cell populations present in cancer microenvironment, which necessitate analysis of multiple modalities of experimental observations in high volume. For example, deep learning has been used to increase accuracy in the diagnosis of metastatic lesions from histological slides(Litjens et al., 2016; Wang et al., 2016) and fMRI images(Liu et al., 2017b). Similarly, algorithms are being developed to reclassify cancer subtypes based on more detailed genotypic data which were previously based on histological analysis, presenting opportunities for more targeted therapy(Cancer Genome Atlas Research, 2014; Han et al., 2014). Although generally viewed as “black box” approaches unable to inform about disease mechanisms, other approaches provide avenues for interpretable neural network models(Michael et al., 2018; Samek et al., 2017). Other recent developments in deep learning methods have been applied in the analysis of complex systems of interacting components seen in various physical systems(Broecker et al., 2017; Schindler et al., 2017), and similar methods may be ripe for the analysis of the dynamics of cell populations in metastasis, beyond what can be modeled using biochemical pathways within cells. Recent theoretical development of deep learning models that operate on graph structured data(Defferrard et al., 2016; Kipf and Welling, 2016) has the potential to integrate protein and gene interaction network information to relate cancer related phenotypes with genotypes and expression data(Ma and Zhang, 2018; Rhee et al., 2017).

In the future, network analysis and machine learning approaches may be used to directly inform a mechanistic understanding of metastasis. Dcell uses deep neural networks to model the growth phenotype of gene deletion, while explaining the prediction in terms of a chain of causal changes in different modules such as subcellular structures and pathways(Ma et al., 2018). Such methods can be augmented to include systems of multiple cell types to address phenotypes relevant to metastasis. Innovative applications of network inference and machine learning could bridge the gap between large amounts of biological data and incomplete mechanistic models using concepts such as equation discovery(Dzeroski and Todorovski, 2008) and hybrid mechanistic-machine learning models(Hua et al., 2006). At a minimum, they will inform, and more likely prioritize, the molecular and biochemical components that are included in more traditional systems biology-based mechanistic models.

Experimental Models and Technologies to Facilitate Systems Study of Metastasis

Systems biology is often approached through a circular iterative process, where data collection informs computational modeling to generate testable hypotheses, and in turn, the outcome of the subsequent hypothesis testing refines the computational model. Required for these iterations are quantitative multivariate experimental approaches that adequately capture the complexity (or simplicity) of the system, and, in many cases, provide longitudinal data (Table 1, Figure 2). In the study of cancer metastasis, the most often utilized experimental systems range from single-cell measurement techniques that provide a snapshot of the molecular characteristics of metastatic cells to genetically modified mouse models (GEMMs) that facilitate the study of the entire metastatic process. Further, investigating the progression of cancer metastasis may require tailored approaches to understand the mechanistic underpinnings driving those processes (Table 2). Some of the more recent developments regarding experimental systems to inform or test systems biology models are reviewed below along with a wish list for technology developers interested in contributing to this emerging area.

Single-cell approaches to studying cancer metastasis

The recent decreases in the cost and concomitant increases in the portability and flexibility of single-cell technology will facilitate a deeper understanding of tumor cell heterogeneity in metastasis(Bush et al., 2017; Gierahn et al., 2017). Particularly, single-cell DNA and RNA sequencing techniques are employed to investigate primary tumors, metastatic lesions, circulating tumor cells, and disseminated tumor cell singlets and clusters(Leung et al., 2017; Ting et al., 2014; Tirosh et al., 2016). A better understanding of the numerous molecular characteristics that give rise to metastatic cells will lead to better predictive models for patient treatment and outcome. However, when employing these approaches to parameterize or test the hypotheses derived from computational systems biology models, it is important to note that single-cell genomic and transcriptomic data pose analytical hurdles, including a low signal to noise ratio due to low read coverage and non-trivial normalization that currently makes it challenging to compare multiple datasets across patients or experimental conditions. Additionally, interrogation of small cell populations, such as dormant metastatic cells, may limit statistically meaningful insights if desired population comparisons are not carefully considered during experimental design. Considerable theoretical work has produced numerous useful data analysis techniques for single-cell genomics, transcriptomics, and proteomics, and the subsequent incorporation of these data into predictive computational models is now beginning(Liu et al., 2017a; Marcotte et al., 2016; Setty et al., 2016; Tanay and Regev, 2017; Villani et al., 2017).

One challenge in interpreting tumor heterogeneity at the single-cell level is the biological role of “rare” cancer cells captured at a single time point. Such cells could be viewed either as an important minor subpopulation, forming the tail of a distribution representing a cancer cell signature, or as traversing through a dynamic cellular state. In the absence of a longitudinal sampling strategy, coupling information at the genomic or transcriptomic level with experimental measurements of cell phenotype may provide guidance to the level of granularity required when interpreting single time point data. Techniques that can accurately quantify, at the single cell level, a physical characteristic, such as a change in cell mass or modulation of cell motility due to drug treatment or an environmental perturbation, could provide the additional data required to differentiate between a stable and transient cell state when coupled with single-cell molecular data(Rojas-Puentes et al., 2016). Even when longitudinal data is available, the integration of multiple data types, such as bulk and single-cell whole-genome sequencing and transcriptomic profiling may be required to predict an outcome with accuracy(Brady et al., 2017). Finally, single-cell sequencing of metastatic lesions coupled with computational methods and lineage tracing of metastases in animal models amenable to imaging, such as C. elegans, or zebrafish, holds the potential to address outstanding questions about the origins and clonality of metastasis, as well as the emergence of new phenotypes such as drug resistance(Brown et al., 2017; Heilmann et al., 2015; Kyriakakis et al., 2015).

Whole-organism experimental systems for metastasis research

Genetically modified animal models of metastasis have been developed and successfully utilized to study all known steps within the metastatic cascade. The benefits and drawbacks of the most frequently utilized models have been thoroughly reviewed elsewhere(Gomez-Cuadrado et al., 2017; van Marion et al., 2016). Developments in gene editing have facilitated large-scale in vivo studies of the genetic alterations that drive cancer tumor growth and metastasis(Chen et al., 2015; Kalhor et al., 2018). Computational integration of these data with information regarding modulation of epigenetic state and alteration of the transcriptome could provide insight into disease mechanism.

The development of lineage tracing tools(Sikandar et al., 2017) could allow for straightforward parameterization or testing of systems biology models through visualization of the kinetics of metastasis or by quantitatively reporting on cell populations in distant metastatic sites of interest. High temporal resolution is on the wish list of many systems biologists who study metastasis, but is difficult to obtain in mammalian models except through the incorporation of imaging windows that facilitate multiple rounds of high resolution intravital imaging(Entenberg et al., 2017; Suijkerbuijk and van Rheenen, 2017). Rapid advancement in functional imaging of metabolic or signaling state of tissues in vivo could be used to contextualize metastatic steps with the tissue microenvironment (Herzog et al., 2013; Knox et al., 2018; Miloushev et al., 2018; Shangguan et al., 2018). Zebrafish have emerged as an alternative system to study aspects of cancer metastasis at high temporal and spatial resolution (Heilmann et al., 2015), allowing for quantification of cell-cell or cell-stroma interactions and relatively easy genetic manipulation. However, in the instances when it is desirable to directly study human cells, zebrafish offer a suboptimal immune microenvironment, similar to the many xenograft-derived in vivo mouse models currently in-use. Finally, it is worth noting that a popular drug screening tool, the patient derived xenograft (PDX) mouse model, has not yet found high utility within the cancer metastasis field. There are limited PDX systems that recapitulate metastases observed in human patients (Jiang et al., 2015; Whittle et al., 2015). However, humanized mouse models combined with PDX provide avenues to study cancer metastasis in the context of the correct host immune system (Jespersen et al., 2017; Rongvaux et al., 2017; Zheng et al., 2018).

Bioengineered in vitro models for studying metastasis

In vitro and ex vivo tumor models offer simplified systems that can be perturbed with relative ease, while also being able to incorporate some of the complexity of tumor microenvironments such as variations in mechanical properties, matrix chemistry, paracrine signals, and cell-cell interactions(Katt et al., 2016; Singh et al., 2018; Zanoni et al., 2016). Andrew Ewald’s group, for example, has utilized human breast cancer organoids grown in three dimensional culture, complemented by engineered mouse models, to show that extracellular matrix characteristics inform invasive behavior and demonstrated that Twist1 regulates Protein Kinase D1 (PKD1) expression to drive metastatic dissemination (Shamir et al., 2016). Jan Lammerding has studied the role of mechanical stress during transendothelial migration of cancer cells using microfluidic cancer transmigration models incorporating tissue mechanics, biochemical properties, and cell-cell interactions (Cao et al., 2016; Singh et al., 2018). A liver microphysiological system, in which all the major cell types in the liver microenvironment are cultured together, including liver-specific immune cells, is being employed by Alan Wells and colleagues to study hepatotoxicity of cancer drugs as well as metastasis and dormancy of breast cancer cells in human liver tissue(Wheeler et al., 2014).

Systemic Perturbation Approaches

With the increasing adoption of CRISPR/Cas9 technology to target the human genome, image-guided tissue isolation, sequencing, and informatics, related approaches have been developed to systematically identify genetic basis for the various steps involved in the metastatic cascade. These screens typically involve creating a library of starting cellular cohort with single, or multiple genetic manipulation in each cell, and selectively sorting the cells that successfully pass through a phenotypic filter, identifying the genes that are enriched in the selected cohort. Enrichment hits could be revalidated using the same filter, or in a more complex system. Experimental models have also been developed that combine library screens with a phenotypic filter such as cell motility or combine a phenotypic functionality, such as dissemination combined with subsequent bioinformatics analysis(Konen et al., 2017). CRISPR/Cas9 based screens are increasingly being utilized to create autochthonous models that facilitate investigation of the combinatorial genetic mutations that drive cancer metastasis (Chow et al., 2017; Shen et al., 2017). Similarly, platforms investigating the role of extracellular matrix at different stages of metastases have been developed to explore non-cell autonomous processes. As an example, Claudia Fischbach’s group observed that the link between cancer malignancy and obesity could be partially attributed to increased collagen deposition by adipocyte stem cells in response to conditioned medium from tumor cells. These findings correlate with observations that adipose tissue in obese patients is stiffer, more collagenous, and may abet cancer metastasis by selective enrichment of cancer stem cell subpopulations(Chandler et al., 2012; Seo et al., 2015).

Precision medicine: deeper omics understanding in the clinical setting

In addition to informing both drug development and a causal understanding of metastasis, systems biology has also been instrumental in the development of patient specific causal analysis and treatment modality(Jia and Zhao, 2014). Techniques like patient specific xenografts allow the identification of targetable tumor specific antigens and the optimization of drug regimens(Karamboulas et al., 2018; Lange et al., 2018; Malaney et al., 2014). Patient and sample specific analysis is even more pertinent for cancer metastasis because the range of possible genetic, epigenetic, or transcriptomic variant combinations that may be causal for overcoming each obstacle from cancer initiation to proliferation, invasion, etc. may be more than the number of patient samples. Therefore, the classical approach of finding statistically significant differences between pathological and control samples is inadequate, with the number of parameters growing much larger than the number of samples. Single sample (N of 1) analysis methods seek to solve these problems in several ways. Classically, N of 1 studies relied on the collection of paired control-intervention data from the same individual by applying some temporal sequence of treatment and placebo/control. For cancer and metastasis, the comparison of matched samples of some combination of normal, primary tumor, and metastatic tumor has been used to find causal mutations and/or suitable interventions(Al-Ahmadie et al., 2014; Brannon and Sawyers, 2013; Kadakia et al., 2015). Califano et al. have developed a network analysis method to differentiate between causal and passenger mutations for individual patients based on the plausible set of causal genes that may be affecting the master regulators through a regulatory network(Chen et al., 2014). Identification of the pathologically important regulators allows the targeted drug screens on the patient derived xenografts to arrive at a treatment plan. We believe similar approaches that integrate the existing knowledge of cancer biology (via the construction of regulatory network), patient specific genotyping, and experimental techniques are an ideal application of systems biology in the service of patient care.

The development of less invasive tests, such as liquid biopsies for cancer biomarkers(Bettegowda et al., 2014) has opened an avenue for the wider application of precision medicine for pre-emptive surveillance for cancer initiation, metastatic growth(Li et al., 2010), chances of metastasis after treatment on the primary site, or for the selection of appropriate therapy. These biomarkers could be the expression of certain genes(Zheng et al., 2007), methylation patterns(Barault et al., 2018; Garrigou et al., 2016) or mutations(Board et al., 2010; Diehl et al., 2008) in circulating tumor DNA, the types of circulating tumor and immune cells(Ilie et al., 2012), and detected from blood or other material such as sputum, urine, or stool. The discovery and correct interpretation of these biomarkers is again dependent on the statistical issues regarding small sample size and larger number of parameters.

Integrating Models of Metastasis

Cancer metastasis is a multidimensional process, unfolding at many biological scales, including molecular signaling networks, protein-protein interactions, metabolism, cell-cell and cell-ECM interactions, organ-level control, disease manifestation and epidemiology. Our ability to describe events in each of those scales will need to be augmented to also inform events in other scales. Integrating the knowledge of all these snapshots at a single cell level would allow determination of consistently active motifs driving metastatic phenotypes. These biological scales can be practically useful as nodes of data integration, and positioning our understanding of the systemic nature of metastasis (Figure 3). Integration of results across all scales may be very challenging, and so a piecemeal approach is warranted. Multi-scale modeling across scales defined by informatics, e.g. genomic, epigenomic, transcriptomic scales is now attempted by specialized next generation sequencing (NGS) techniques, including ChIP Seq (Chromatin Immunoprecipitation Sequencing), ATAC Seq (to access chromatin accessibility genome wide), PRO Seq (Precision nuclear run-on sequencing)(Meyer and Liu, 2014), and many other technologies. The objective of these technologies is essentially to systematically explore the mechanism of gene regulation resulting in a given transcriptomic output. Several attempts are now made to correlatively analyze the transcriptomic, proteomic, and even metabolomics scales systematically in a physiological system(Bahado-Singh et al., 2017; Garcia et al., 2012; Rabinowitz et al., 2017; Zhao et al., 2018). Such integrative approaches will be crucial in mechanistically understanding cancer metastasis as a multiscale disease (Figure 3).

Integration of data across scales, both not characterized by “sequence information” is fraught with challenges that are more profound. Well characterized phenotypes with richly-defined feature sets could be mapped to transcriptomic, and genomic scales; e.g. characterizing tumor organoid shape and other characteristics for precision medicine, as well as RNA Sequencing (Clark et al., 2018; Ewald, 2017). Tissue level details can be integrated with transcriptomics by spatial scRNA sequencing, for which more and more methods are being developed (Moncada et al., 2019; Stahl et al., 2016). Similarly, signaling state of cells can be correlatively assessed with the transcriptomic state(Zhang et al., 2019), which is likely to be benefitted by single cell RNA sequencing correlated with parallel multiplexed imaging or cytometry. Intercellular communication between metastatic cancer cells and stromal cells, including the immune subpopulations, could be integratively explored by scRNA sequencing, simplistically by creating ligand-receptor interaction maps, and in a more detailed manner by combining parallel means of data collection, e.g. secretome analysis. These integrative methods aim towards validating and confirming the information flow from one physiological scales to another, however, more mechanistic integrative efforts are being made combining computational modeling of signals between cells constrained by defined boundaries in the context of a tissue(Ghaffarizadeh et al., 2018; Letort et al., 2019; Zangooei and Habibi, 2017).

Experimental in vitro or ex vivo models that faithfully recapitulate the entire process of metastasis are coming online with the advent of tissue-on-a-chip and organ-on-a-chip technologies (Table 1). These systems are likely to offer a controlled environment to build and test computational models that describe the various sub-systems of the metastatic cascade. These types of multi-scale systems biology approaches will offer a way to test large numbers of hypotheses regarding development and treatment of metastatic disease before advancing to the infinitely complex in vivo environments. It is only through such systems-level understanding of cancer metastasis that we can hope to decrease the lethality of the disease and continue to develop detection and treatment strategies. A global and mechanistic model of metastasis with a predictive flow of information across larger scales is still a work in progress, and requires methods to formalize data collection across scales, techniques to validate derived changes at a scale informed by another scale, as well as systems approaches to understand progression of metastasis combining several spatio-temporal scales (Figure 3).

Conclusion

Cancer metastasis has continued to confound researchers owing to its complex, and not completely determinable, pathological trajectory. The complexity in disease progression arises because essentially cancer is a disease of abnormal cell proliferation, and metastasis involves successfully encountering physiological hurdles, both aspects resulting in a strong evolutionary thrust defining the metastatic cascade. Cancer metastasis, is therefore an evolving disease and is a combined outcome of cells that metastasize, as well as a series of microenvironmental factors they interact with, collude, or surmount. Although, each instance of metastasis could be unique, the quest is to find commonalities which could be therapeutically targeted. The complexity of metastasis through its chronological progression, and its manifestation in various biological scales calls for a systems approach to mechanistically understand metastasis. This approach will involve integrating information gained across different scales, and mechanistically modeling disease progression as a functional outcome of the complex interaction between cancer and the ecology it interacts with. Important technical advancements in recent years to gain knowledge at a systems level at various biological scales have raised a hope that cancer metastasis may itself be modeled using a multiscale and systems approach. The need is to develop methods to integrate these approaches across scales, and stage of metastasis towards creating a systems mechanistic model of cancer metastasis.