Tumour resistance: single-cell or population effect?

Tumour resistance: single-cell or population effect?

We are searching data for your request:

Forums and discussions:
Manuals and reference books:
Data from registers:
Wait the end of the search in all databases.
Upon completion, a link will appear to access the found materials.

Drug resistance can arise through a number of mechanisms. For instance, EGFR mutation when treating with EGFR inhibitors, or compensatory activation of alternative survival pathways. But does it occur at the level of a single cell, or would it be a population effect? In other words, would cells exposed to e.g. targeted kinase inhibitors slowly compensate with other pathways (and that would happen in every such cell) or would there be a sub-population of cells that is sensitive to inhibitors and another subpopulation inherently resistant and predisposed towards amplifying other pathways.

There certainly can be co-operation or competition among cancer cells within a tumor that can provide some types of population-based influences on cancer-cell survival. Present evidence, however, suggests that resistance to therapy is almost always due to resistant sub-populations of cancer cells within the tumor. In many cases, such as after treatment with EGFR inhibitors, the resistance mutations have been identified. Sometimes these mutations are in such a small fraction of the cancer cells that they can't be detected in the original tumor before treatment, but in most cases they probably were already present before therapy, caused by the mutations that are inevitable during tumor growth. This review on intra-tumor heterogeneity is a good place to start learning about this topic.

Integration of single-cell RNA-seq data into population models to characterize cancer metabolism

Affiliations Dept. of Molecular Cell Physiology, Faculty of Earth and Life Sciences, VU University, Amsterdam, The Netherlands, Manchester Centre for Integrative Systems Biology, School of Chemical Engineering and Analytical Science, University of Manchester, Manchester, United Kingdom, Swammerdam Institute for Life Sciences, Faculty of Science, University of Amsterdam, Amsterdam, The Netherlands

Roles Project administration, Writing – review & editing

Affiliations SYSBIO Centre of Systems Biology, 20126, Milan, Italy, Dept. of Biotechnology and Biosciences, University of Milan-Bicocca, 20126, Milan, Italy

Roles Project administration, Writing – review & editing

Affiliations SYSBIO Centre of Systems Biology, 20126, Milan, Italy, Dept. of Biotechnology and Biosciences, University of Milan-Bicocca, 20126, Milan, Italy

Roles Funding acquisition, Project administration, Resources

Affiliations Dept. of Informatics, Systems and Communication, University of Milan-Bicocca, 20126, Milan, Italy, SYSBIO Centre of Systems Biology, 20126, Milan, Italy

Of single cells and amplification

The key to single-cell biology, of course, is isolating a single cell. This typically is accomplished using micromanipulation, microfluidics, or fluorescence-activated cell sorting (FACS), the method Stepanauskas favors.

One alternative, developed in the laboratory of Nancy Allbritton, chair of the joint biomedical engineering department at the University of North Carolina at Chapel Hill (UNC-CH) and North Carolina State University (NCSU), is the microraft array (MRA).

An MRA is an array of tiny transparent, magnetic, polystyrene elements, each small enough to fit into a tiny well on a plate or slide. Commercialized by Cell Microsystems, MRAs typically contain about 10,000 wells, Allbritton says, though some measure in the millions. Researchers plate their cells such that there are, on average, zero or one cells per element. They can then image the array immediately, or allow the cells to grow and develop complex phenotypes. Cells of interest are gently isolated using a microneedle to pierce the bottom of the array and dislodge the MRA element. Because they are magnetic, these elements are easily captured, at which point they can be analyzed or clonally expanded.

According to Allbritton, the system enables isolation strategies that might otherwise be impossible, such as isolating cytotoxic T lymphocytes based on their ability to kill target cells. In one study, UNC-CH collaborator Scott Magness, working with Allbritton, used an MRA to study intest-inal stem cells. Among other things, the team used the platform to investigate whether intestinal stem cells must physically contact other cells in order to reach their full potential. “What he showed quite clearly is they really need to be touching a Paneth cell to grow, prosper, and divide and get the highest outgrowth rates,” Allbritton says.

Vertes takes a lower-throughput approach to metabolomics. Using a sharpened capillary and a micromanipulator—a tool typically used in patch clamping and embryonic manipulation—his team aspirates a fraction of a picoliter of material from adherent cells stuck to the bottom of a culture dish and injects it directly into a mass spectrometer.

In that way, Vertes says, his team can quantify some 22 metabolites and 54 lipids from each of about 30 cultured hepatocytes—not the whole metabolome, of course, but enough to reveal for instance, the cells’ adenylate energy charge—a measure of cellular health. It was always possible to measure such variables at the population level, Vertes notes. But by going to the single-cell level, he can observe the distribution
of energy states, and correlate those values with such parameters as morphology or metabolic activity. “For every cell, we can tell [if it] was a healthy cell, a cell half-dead, or a cell programming itself to die,” he says.

A Review of Cell-Based Computational Modeling in Cancer Biology

Cancer biology involves complex, dynamic interactions between cancer cells and their tissue microenvironments. Single-cell effects are critical drivers of clinical progression. Chemical and mechanical communication between tumor and stromal cells can co-opt normal physiologic processes to promote growth and invasion. Cancer cell heterogeneity increases cancer’s ability to test strategies to adapt to microenvironmental stresses. Hypoxia and treatment can select for cancer stem cells and drive invasion and resistance. Cell-based computational models (also known as discrete models, agent-based models, or individual-based models) simulate individual cells as they interact in virtual tissues, which allows us to explore how single-cell behaviors lead to the dynamics we observe and work to control in cancer systems. In this review, we introduce the broad range of techniques available for cell-based computational modeling. The approaches can range from highly detailed models of just a few cells and their morphologies to millions of simpler cells in three-dimensional tissues. Modeling individual cells allows us to directly translate biologic observations into simulation rules. In many cases, individual cell agents include molecular-scale models. Most models also simulate the transport of oxygen, drugs, and growth factors, which allow us to link cancer development to microenvironmental conditions. We illustrate these methods with examples drawn from cancer hypoxia, angiogenesis, invasion, stem cells, and immunosurveillance. An ecosystem of interoperable cell-based simulation tools is emerging at a time when cloud computing resources make software easier to access and supercomputing resources make large-scale simulation studies possible. As the field develops, we anticipate that high-throughput simulation studies will allow us to rapidly explore the space of biologic possibilities, prescreen new therapeutic strategies, and even re-engineer tumor and stromal cells to bring cancer systems under control.

Cancer is a complex systems problem that involves interactions between cancer cells and their tissue microenvironments. 1-3 Therapeutic approaches that narrowly focus on cancer cells frequently lead to disappointing outcomes, including resistance, tissue invasion, and treatment failure. Such failures are partly due to the unexpected behaviors that emerge from the dynamical systems of cancer tissues. Therapies act as selective pressures, even while cancer cells use increased genetic variability to broadly sample survival strategies and adapt. 3,4 Chronic hypoxia, another selective pressure, leads to metabolic changes, selection for cancer stem cells that resist treatment, invasion, and angiogenesis. 4-6 Tumor cells communicate biochemically and biomechanically with stromal cells, which allows them to co-opt normal physiologic processes. 1-3,7,8 Mathematical models can serve as "virtual laboratories" with fully controlled conditions where scientists and clinicians can investigate the emergent clinical behaviors that result from basic cell hypotheses and can evaluate new therapeutic strategies. 1,9

This review surveys cell-based methods for simulating cancer. Also known as discrete models, agent-based models, or individual-based models, cell-based models simulate individual cell behaviors within tissue environments. These models have several advantages. Each cell agent can track a fully independent state with individual parameters that reflect heterogeneity in cancer. Modelers can directly implement cell rules that reflect observations of single-cell behavior and cell-cell interactions, which allow us to translate biologic hypotheses to mathematical rules quickly run simulation experiments that explore the emergent behaviors of these hypotheses and compare against new data to confirm, reject, or iteratively improve the underlying hypotheses. 1,9,10

Cell-based models represent individual cells with two main paradigms—lattice-based models that track cells along a rigid grid and off-lattice models that have no such restriction. Figure 1 classifies most cell-based modeling approaches. Table 1 lists major open source modeling packages.

FIG 1. A schematic classification of cell-based modeling approaches.

TABLE 1. Computational Methods and Open Source Toolkits

Lattice-based models can use regular structured meshes (eg, Cartesian 11 [two- or three-dimensional [2D/3D], dodecahedral [3D]) 12 or unstructured meshes. 13 Structured meshes are simpler to implement, visualize, and combine with partial differential equation (PDE) solvers, but their structure can lead to grid biases. 13 Unstructured meshes can avoid these issues 13 but with greater complexity.

We can further categorize lattice-based methods by their spatial resolution. In cellular automaton (CA) models, each lattice site can hold a single cell. 14-17 At each time step, each cell is updated with discrete lattice-based rules: remain, move to a neighboring lattice site, die (free a lattice site), or divide to place a daughter cell in a nearby site. 14-17 These methods usually update the lattice sites in a random order to reduce grid artifacts. 14,15

In lattice gas CA (LGCA) models, a single lattice site can contain multiple cells. 14,15,17,18 LGCA models track the number of cells that move through channels between individual lattice sites rather than the motion of each individual cell. They can simulate very large numbers of cells efficiently over long periods while also connecting to statistical mechanics theory this facilitates analysis and provides a bridge to continuum methods that model cell densities or populations instead of single cells. 17,18

Some problems may require resolution of individual cell morphologies. Cellular Potts models (CPMs) use multiple lattice sites to represent each cell. 14,15,19 At each time step, CPMs visit each pixel (2D) or voxel (3D), test a random swap with a neighboring pixel/voxel, and accept or reject the swap (probabilistically) on the basis of whether it would reduce a global energy. Although CPMs can model cell morphologies and mechanics that cannot be incorporated in CA models, they are much more computationally intensive. Also, the calibration of Monte Carlo steps to physical time can be challenging. 20

We can divide off-lattice models into center-based models (CBMs) that focus on cell volumes (or masses) and models that focus on cell boundaries. We can further classify these approaches by level of morphologic detail.

CBMs track each cell’s center of mass or volume, typically by using a single software agent per cell. 13-15,21 Some CBMs represent cells as points, whereas others explicitly model cell volumes. CBMs typically update the cells’ positions by explicitly formulating the adhesive, repulsive, locomotive, and drag-like forces exchanged between cell centers. 13-15,21 Most CBMs approximate cells as spheres however, some approximate cells as deformable ellipsoids to better represent their morphologies. 22,23

CBMs can model cell morphology in greater detail by breaking cells into subcellular elements 24,25 : Each cell is represented by multiple center-based agents that interact with adhesive and repulsive forces. These models better approximate cell biomechanics but at increased computational cost. Conversely, cells can be organized into clusters or functional units (eg, breast glands or colon crypts) that are simulated as agents that interact by mechanical forces or other rule-based motions 26,27 this allows modelers to incorporate heterogeneous details into individual clusters of cells but with greater computational efficiency than traditional CBMs.

Vertex-based methods (eg, Fletcher et al 28 ) model cells as polygons (2D) or polyhedra (3D) and compute the forces that act on their vertices they are particularly useful for modeling confluent tissues. 29 For greater spatial resolution, front-tracking methods, such as the immersed boundary method (IBM), solve PDEs for fluid flow inside and between cells and then advect boundary points along the cells’ membranes in this flow. 30 Level set methods have been applied to implicitly track the movement of cell boundaries, 31 and VCell (see Connecting to Molecular Effects) recently added front-tracking capabilities. 32,33 These are among the most computationally intensive cell-based methods, but they are useful for coupling detailed cell mechanics to fluid and solid tissue mechanics.

Most cell-based models are hybrid discrete-continuum they couple a discrete cell model to continuum models of the microenvironment. 1,14,15 In general, these models use reaction-diffusion PDEs to simulate biotransport of oxygen, growth factors, and drugs. Ghaffarizadeh et al 34 developed BioFVM to solve diffusive transport of tens to hundreds of chemical substrates in 3D tissues it is the underlying PDE solver for PhysiCell (a center-based simulation framework). 21 In this framework, modelers write rules to relate individual cell phenotypes to local chemical substrate conditions. 21

Many discrete models include systems of ordinary differential equations (ODEs) to model molecular processes in individual cells. 35,36 VCell can simulate reacting flows of many proteins within a single detailed cell, 32,33 and many modeling packages (eg, Chaste, 37 CompuCell3D, 38 and EPISIM 39 ) support systems biology markup language (SBML) to include systems of ODEs that simulate molecular effects in individual cells. Others use discrete models within individual agents: Gerlee and Anderson 40 used small neural networks to simulate individual cell phenotypic "decisions" on the basis of microenvironmental inputs, whereas PhysiBoSS 41 combines the Boolean network modeling approach of MaBoSS 42,43 with PhysiCell 21 to simulate molecular processes in individual cells.

We now explore a series of modeling themes that illustrate the use of cell-based modeling in cancer biology. Although we cannot comprehensively review all cell-based modeling in cancer (or even sample all major use cases for cell-based modeling), these themes are drawn from across the field to demonstrate scientific problems with significant cell-scale effects where cell-based models can yield new insights.

Many groups have used cell-based models to investigate tumor growth in hypoxic tissues and more generally, the effect of diffusive transport limits. Gatenby et al 44 and Smallbone et al 45 used CAs to examine hypoxia-driven switching to invasive phenotypes in ductal carcinoma in situ (DCIS). They incorporated cellular metabolic adaptations to hypoxia, which allowed them to study early tumor invasion ( Fig 2A ). Anderson and colleagues 46,47 extended earlier CA results by adapting IBCell 30 (an IBM) to mouse mammary (EMT6/Ro) tumor cell proliferation in hypoxic tissues. As before, they found that hypoxic gradients could drive tissue invasion, but IBCell’s improved modeling of cell adhesion and biomechanics predicted more rounded invasive tips 47 ( Fig 2B ).

FIG 2. Cell-based models of hypoxia in breast cancer. (A) A cellular automaton model of breast cancer that explores cellular metabolic changes and early development of invasion. Reprinted with permission from Gatenby et al. 44 (B) An immersed boundary model to simulate cancer invasion under hypoxic gradients. Adapted with permission from Anderson et al. 47 (C) PhysiCell (a center-based model [CBM]) simulation of ductal carcinoma in situ as it advances in breast ducts under diffusive growth limits. Note the brown necrotic core. Adapted with permission from Ghaffarizadeh et al. 21 (D) Adapted PhysiCell simulation of hanging-drop tumor spheroids. Oxygen diffusive limits lead to hypoxic gradients, greatest proliferation on the outer edge, an interior quiescent region, and an central necrotic core (brown). Note the network of fluid-filled pores that emerges from the necrotic core mechanics. These are observed in experiments. The inset shows a fluorescent image of a hanging-drop tumor spheroid. Adapted with permission from Ghaffarizadeh et al. 21 (E) CBMs of tumor spheroids pioneered by Drasdo and Höhme produced similarly layered structures. Reprinted with permission from Drasdo and Höhme. 48 (F) A CBM of tumor cords growing around a blood vessel and showing a reversed structure with viable tissue in the interior. Adapted with permission from Szymańska et al. 49

Macklin et al 50 and Hyun and Macklin 51 applied a CBM to study oxygen-driven proliferation and necrosis in solid-type DCIS with comedonecrosis. After calibrating to individual patient pathology data (tissue specimens immunostained for the Ki67 protein to detect cycling cells, cleaved caspase 3 to detect apoptosis, and annotated with viable rim sizes and cell density 50 ), they were able to simulate comedonecrosis and microcalcifications as emergent properties of the simulations along with realistic, constant rates of tumor advancement along the breast ducts. Ghaffarizadeh et al 21 refined the DCIS model and extended it to 3D as well as simulated the hypoxic interiors of hanging drop spheroids calibrated to match MCF-10A birth and death kinetics in culture ( Figs 2C and D ). As in early 3D work by Drasdo and Höhme 48 on EMT6/Ro cells ( Fig 2E ), they predicted a layered structure—an outer proliferative rim surrounding a quiescent perinecrotic region and an interior necrotic core. They were the first to predict networks of fluid-filled pores in the necrotic cores that emerge from the competing effects of necrotic cell shrinking and adhesion these structures are observed in experimental models ( Fig 2D inset). Szymańska et al 49 used a CBM of EMT6 cells to simulate a growing tumor cord—a solid tumor that grows around a blood vessel. They predicted a similar three-layer structure but in reverse order—a proliferating core nearest the blood vessel, quiescent interior, and necrotic exterior ( Fig 2F ).

Tumor-induced angiogenesis allows lesions to grow to clinically detectable sizes. 3 McDougall and colleagues 52,53 modeled sprouting angiogenesis with a CA model of vessel tip migration: Sprout tip agents followed chemotactic and haptotactic signals to migrate toward hypoxic tumor regions and left a trail of functional vessels. They incorporated a detailed vascular network flow model, including dynamic wall shear stress rules for vessel branching and anastomosis (vessel looping), and used this framework to explore therapeutic delivery from tumor-associated vasculatures ( Fig 3A ). Bauer et al 54 used a CPM to simulate tumor-induced angiogenesis, by adding a detailed microenvironment, including extracellular matrix (ECM) and multiple vascular endothelial growth factor isoforms they concluded that variations in the spatial distributions of proangiogenic factors greatly affect capillary morphology and that inhomogeneities in nonvascular tissue naturally lead to capillary anastomosis. Boas and Merks 55 used a CPM to investigate novel hypotheses on cell overtaking: Cell-cell biomechanical and chemical communication can cause endothelial cells in the stalk to assume the role of migrating tip cells by migrating to the front of an advancing vessel ( Fig 3B ). Shirinifard et al 56 used a CPM to investigate tumor growth with angiogenesis and showed that tumor size increases with increasing angiogenesis and that the tumors grow along the vasculature ( Fig 3C ).

FIG 3. Tumor-associated angiogenesis and vascular flow. (A) A two-dimensional (2D) cellular automaton model of sprouting angiogenesis used to study drug delivery from tumor vasculatures. Reprinted with permission from McDougall et al. 53 (B) A 2D cellular Potts model of angiogenesis. Stalk cells can overtake tip cells to become new tip cells. The arrows show these role swaps. Reprinted with permission from Boas and Merks. 55 (C) A 3D cellular Potts model of sprouting angiogenesis driven by vascular endothelial growth factor released by hypoxic tumor cells. Adapted with permission from Shirinifard et al. 56 (D) A 2D cellular automaton model (left) to investigate drug delivery to simulated tumors (right). Adapted with permission from Cai et al. 57 (E) A discrete angiogenesis model of McDougall et al 53 combined with a continuum tumor growth model 58 used to investigate the effect of interstitial fluid pressure and lymphatic drainage on therapeutic delivery. Shown are tumor and the discrete vasculature (left) fluid extravasation from blood and lymphatic vessels (middle) and interstitial fluid velocity (right), which hinders drug delivery. Adapted with permission from Wu et al. 59

Cai et al 57 used a CA model of tumor cells in a continuous ECM coupled with a discrete angiogenesis model that included flow effects and substrate perfusion from the vasculature ( Fig 3D ). They showed that the final vessel configuration depends on emergent, dynamic feedback mechanisms in vascular remodeling rather than on initial conditions. Wu et al 59 extended an earlier hybrid discrete-continuum model 58 (that was based on that of McDougall and colleagues 52,53 ) to investigate the influence of interstitial fluid pressure, interstitial fluid flow, and lymphatic drainage on drug delivery in growing tumors. 59 They found that elevated interstitial hydraulic conductivity and high interstitial fluid pressure limit the transvascular delivery of nutrients and therapeutics ( Fig 3E ).

Models of cancer stem cells (CSCs) offer valuable insights into the driving forces of cancer biology. Fletcher and colleagues 37,60,61 developed a 3D CBM of colonic crypts to explore the role of stem cells (in the bottom of the crypt) in colorectal carcinogenesis. Neighboring cells were connected by linear springs, and stem-cell division and differentiation were driven by Wnt gradients along the crypt axis. The geometry of the stem-cell hierarchy (proliferation at the crypt base, expansion and differentiation along the middle and top) created an overall base-to-top proliferative cell flux. This flux has an anticancer protective effect wherein it pushes any mutated cell and its progeny out of a crypt before they can spread throughout a crypt, unless the mutation occurs in a stem-cell niche ( Fig 4A ).

FIG 4. Cancer stem cells, invasion, and the "go or grow" hypothesis. (A) Top view of a three-dimensional (3D) center-based model of colon crypts (left plots) where the stem-cell niche is in the center. A nonstem mutation (blue cells) is swept out of the crypt by the proliferative cell flux. On the right, is a 3D view of four such ducts that feed cells to a central villus, which is based on the same simulation model. Adapted with permission from Fletcher et al 61 (left) and Mirams et al 37 (right). (B) A 3D cellular automaton (CA) model (with stem-cell effects) of how chemical signaling with fibroblasts and macrophages can drive triple-negative breast cancer. Among these findings, if stromal cells can promote increased cancer cell migration, the overall tumor grows. Reprinted with permission from Norton et al. 62 (C) A 2D CA model to investigate the spread of traits in growing tumors, when cancer cells and their progeny could carry four tumor traits. Traits disseminate largely radially, with clear implications for tumor needle biopsies. Adapted from Poleszczuk and Enderling. 11 (D) A 2D cellular Potts model of stem cells in glioblastoma that shows their role in building resistance to radiotherapy. Reprinted with permission from Gao et al. 63 (E) Lattice gas CA models of the "go or grow" hypothesis in glioblastoma multiforme. As cells spend more time proliferating, they contribute to better growth up to a critical transition point beyond this point, decreased migration is insufficient to open space for cell division. Adapted with permission from Hatzikirou et al. 64 (F) A center-based model to explore the "go or grow" hypothesis in glioblastoma multiforme. Here, G0 is the models' growth rate parameter. Adapted with permission from Kim et al. 65

Norton et al 62 built a 3D CA model to examine the interaction between triple-negative breast cancer and stromal cells. Stem cells proliferated and differentiated into progenitor cells, and cancer cells exchanged chemical signals with fibroblasts and infiltrating macrophages. Among their results, they found that increasing the stromal effect on cancer cell proliferation decreased overall tumor size, whereas increasing the stromal effect on cancer cell migration increased tumor size ( Fig 4B ).

Poleszczuk et al 66 developed a 2D CA model of CSCs and nonstem cancer cells, which tracked four traits in each individual cell: migration rate, apoptosis, symmetric CSC division, and cancer cell proliferation potential. They found that increasing the cancer cell proliferation potential could reduce tumor growth because the increased cancer cell population competed with CSCs for space and inhibited CSC division. They also found that traits propagated radially from the centers of growing tumors this has implications for biopsies of tumor heterogeneity ( Fig 4C ). Gao et al 63 used a CPM to investigate the role of glioma stem cells (GSCs) in glioblastoma growth and radiation therapy response ( Fig 4D ). They found that switching from asymmetric to symmetric division or fast GSC cycling was necessary to explain clinical observations of glioma repopulation after radiotherapy and that the expanded GSC fraction could reduce radiosensitivity.

Alfonso et al 67 also explored radiotherapy treatment paradigms with respect to a heterogeneous population of CSCs and cancer cells using a 3D CA model. They found that CSCs, which are typically more radioresistant, segregated to the center of the tumor across a range of proliferation and death parameters. This emergent phenomenon is due to the faster cycling time of the cancer cells compared with CSCs. When these cell arrangements were subjected to radiotherapy, they found that radiotherapy is more effective at tumor control when it is concentrated on the tumor center where CSCs are located rather than when it is spread homogeneously across the entire tumor.

Tektonidis et al 68 examined the "go or grow" (GoG) hypothesis in gliomas, where tumor cells must make a "decision" between migration (go) or proliferation (grow). They modeled data from 3D spheroid cell cultures with a 2D LCGA model and attempted to recapitulate three experimental observations: nonidentical spreading rates of the invasive rim and central core, radially persistent and symmetric cell motion, and a highly proliferative central core compared with the remaining tumor. Tektonidis et al evaluated the emergent model behavior under a variety of cell phenotype rule sets to determine which rules were required to predict the three observations. They found that a proliferative-motility dichotomy (the GoG hypothesis), cell-cell repulsion, and density-dependent switching between the proliferative and motile states were required to match experimental observations. They concluded that disruption of the GoG mechanism to favor proliferation could limit the required tumor resection volume in surgical interventions.

Hatzikirou et al 64 investigated the GoG hypothesis in glioblastoma multiforme (GBM) using a 2D LGCA model. In their work, cells could divide, re-orient, migrate, or apoptose on the basis of local oxygenation conditions. They modeled the GoG hypothesis by switching hypoxic cells to a motile phenotype and reverting to a proliferative phenotype after escaping hypoxia. They found that increasing the cell bias toward proliferation increases overall tumor growth, but crossing a threshold could decrease overall tumor growth when motility is insufficient to open new space for cell division ( Fig 4E ). Comparable results were obtained by Gerlee and Nelander 69 using stochastic switching between the two phenotypes (migrate or proliferate). From their CA model, they derived a system of coupled PDEs to investigate further the relationships between cell-level parameters and tumor-scale dynamics.

In related work, Böttger et al 70 explored a 2D LGCA GoG model for GBM to provide a more quantitative parameter space analysis of a tumor’s invasive dynamics. Systematically varying model parameters for proliferation and motility led to counterintuitive results about invasion. Specifically, invasion speed depended on two competing processes: emptying space as a result of cell migration and filling space as a result of cell proliferation. In later work, Böttger et al 71 used similar techniques to find that if cell motility decreases with increased cell density, then small tumors self-extinguish. On the other hand, increasing cell motility with increasing cell density leads to self-sustaining growth, similar to the Allee effect frequently observed in ecology.

Kim et al 65 used a CBM to explore GBM using miR-451 as an intracellular detector of glucose, with an ODE model of miR-451 as the effector for selecting migration versus proliferation for glioma cells ( Fig 4F ). They found that cell migration depended not only on glucose, but also on mechanical spacing between cells. They also predicted that the placement of chemoattractants at the edges of a resected tumor could reduce GBM cell migration from the resection site. These GoG examples highlight the key role played in single-cell decisions in GBM and the potential for cell-based models that explore the clinical behaviors that emerge from single-cell effects. 72

Cancer invasion is essential to metastatic progression. 3,72-74 Cancer cells acquire a motile phenotype to escape primary tumors and invade nearby tissues (as conceptually modeled in epithelial-to-mesenchymal transition [EMT] 75 ), invade and travel within blood and lymphatic vessels, 76,77 and finally colonize distant metastatic niches. 78 Because single-cell effects are critical, many cell-based models have investigated cancer invasion with or without explicit modeling of EMT.

Reher et al 79 developed a 2D LGCA model to simulate the effects of a heterogeneous cell-cell adhesion in an epithelial layer, with decreased cell-cell adhesion representing one of the effects of EMT. Cell-cell adhesion was modeled by varying the initial and maximum number of adhesion receptors for each virtual cell. They found that increased adhesion heterogeneity as well as decoupling receptor number from environmental signals (cell-cell contact) lead to increased dissemination.

Kim and Othmer 80 developed a center-based model with a continuous ECM description to investigate the interactions of tumor cells, signal-secreting stromal cells, and the ECM. Stromal levels of fibroblast-secreted protein initiated EMT, which switches cells to an invasive, motile phenotype. Invasive cells also secreted a tumor-associated protease, which degrades the basement membrane and ECM. Degradation of the basement membrane results in more cell exposure to fibroblast-secreted protein, which results in more phenotypic switching more invasion more ECM degradation and ultimately, collective cell invasion ( Fig 5A ). These results suggest that inhibiting fibroblast secretions could affect invasive potential.

FIG 5. Cancer invasion and immunosurveillance. (A) In a center-based model, signals secreted by stromal cells (red) induce tumor cells (gray) to degrade the basement membrane and invade the stroma (blue mesh). Adapted with permission from Kim et al. 80 (B) A three-dimensional (3D) cellular automaton (CA) model to study selection in heterogeneous brain cancers. Cells could mutate their signaling network parameters, which leads to more invasive clones. Adapted with permission from Zhang et al. 81 (C) An immersed boundary method of contact-based signaling and polarization in breast acini. 82,83 Cells with altered signaling could fill the lumen or invade the stroma. Adapted with permission from Anderson et al. 82 (D) A 2D CA model of tumor-immune interactions. Immune cells (blue dots) become exhausted after too many successful tumor cell kills and create fibrotic tissue (yellow). Tumor encapsulation, tumor elimination, and chronic response are observed in the model. Adapted with permission from Kather et al. 84 (E) A sophisticated 3D CA model of treatments targeting programmed cell death-1 (PD-1) and programmed death ligand 1 (PD-L1) in cancer cells. (PD-L1 + cells express PD-L1 PD-L1 − cells do not.) Reprinted with permission from Gong et al. 85 (F) A 3D center-based model of immune responses to an immunostimulatory factor in a heterogeneous tumor (shaded by immunogenicity yellow cells are most immunogenic). Immune cells (red) seek and adhere to cancer cells, test for immunogenicity, and induce apoptosis. The immune response failed after immune cells aggregated near a local maximum in the signaling factor, which allows the tumor to repopulate. Adapted with permission from Ghaffarizadeh et al. 21 This work was explored further with high-performance computing. 10

Zhang et al 81 used a 3D CA model to investigate tissue invasion and tumor cell heterogeneity in glioma. Using a subcellular signaling network, cells divide and move on the basis of external concentrations of nutrients and signaling molecules. They acquire oncogenic mutations upon division to produce new clones. Each successive clone is more proliferative and nutrient seeking, which increases overall invasive potential. By starting with a sphere of the least oncogenic cells in the center of the simulation, Zhang et al found that heterogeneity increased in all regions of the simulation followed by a decrease in heterogeneity in the regions that contain the nutrient source and eventual recovery of heterogeneity. The decrease was attributed to an outgrowth of the most oncogenic subclone, which potentially explains the asymmetric invasive growth seen in experimental and clinical reports ( Fig 5B ).

Anderson et al 82 used an IBM to investigate morphologic changes in breast acini as a result of variation in cell polarization and anoikis behaviors in response to cell contact with other cells and the basement membrane. In this, and additional work by Rejniak et al, 83 Anderson et al showed that atypical behavior in a single cell can lead to eventual intraductal and stromal invasion ( Fig 5C ).

In a sophisticated investigation of metastatic colonization, Araujo et al 86 developed a hybrid CA model of prostate cancer (PCa) bone metastasis. In their work, mesenchymal stromal cells differentiated into osteoblasts that created bone material, whereas osteoclasts degraded bone. In the absence of tumor cells, these cell populations coordinated through transforming growth factor-β and receptor activator of nuclear factor kappa beta ligand (RANKL) signaling to maintain healthy bone tissue. When PCa cell agents were introduced into the bone, tumor-secreted transforming growth factor-β promoted osteoblast activity, but osteoblast-osteoclast feedbacks simultaneously degraded bone to release new growth factors that could drive additional PCa proliferation and invasion. Araujo et al studied the model to compare potential improvements of RANKL inhibitors and bisphosphonates (two standard-of-care treatments for bone metastases). They found that additional refinements to bisphosphonates would yield little clinical improvement over current treatments, whereas improvement of RANKL inhibitors from the current (model-estimated) 40% inhibition closer to a theoretical maximum 100% inhibition could dramatically improve outcome.

Individual interactions between tumor cells and immune cells are critical to immunosurveillance. 3,7,87 Cell-based models are uniquely capable of examining these while considering the roles of stochasticity and heterogeneity.

Kather et al 84 developed a CA model to study complex interactions between tumor cells and immune cells. In their model, immune cells could randomly appear (an influx model), migrate toward tumor cells, and kill them. Immune cells were assumed to be exhausted after killing a maximum number of times and would induce tissue fibrosis that impaired cell migration. Tumor cells could proliferate, die, remain stationary, or migrate on the basis of the number of open neighboring lattice sites and their distance to the nearest open lattice site (a phenomenologic model of necrosis). Depending on the relative migration and proliferation parameters for tumor and immune cells, this model could exhibit diverse tumor-immune interactions, including successful immunosurveillance (eradicated), tumor encapsulation by fibrotic tissue (immune excluded), and ongoing immune responses ( Fig 5D ). Gong et al 85 recently developed a 3D CA model to investigate the tumor-immune responses to programmed cell death-1 and programmed death-ligand 1 inhibition ( Fig 5E ).

Ghaffarizadeh et al 21 developed a 3D off-lattice model of an immune attack on a tumor with a heterogeneous oncoprotein (mutations increased both proliferation and immunogenicity). After simulating initial growth, they added simulated immune cells that chemotaxed toward tumor-released immunostimulatory factors. Each immune cell tested for mechanical collision with cells, formed an (Hookean) adhesion, tested for immunogenicity, and stochastically attempted to induce tumor cell apoptosis. Attached immune cells eventually would succeed in killing the tumor cell and detach or continue the attempt before detaching and continuing their search for new targets. Their simulations showed initial tumor regression but eventual tumor regrowth when immune cells passed some tumor cells and formed large clumps near maxima of the immunostimulatory factor ( Fig 5F ). The authors have expanded their investigation to supercomputers to explore further the effect of stochastic migration on the overall efficacy of the immune response 10 this work showcases the potential for using supercomputers to explore large therapeutic design spaces in high throughput.

Interactive resources for schools

Red blood cells

Carry oxygen in the blood. They are also known as erythrocytes.

Blood vessels

The tubes through which blood is carried around the body eg arteries, veins and capillaries


Treatment of disease using medicines that destroy cancer cells


Treatment of disease using X-rays or radioactive substances which kill cells

Bone marrow

Found in the centre of bones, it contains adult stem cells which divide and differentiate to produce red and white blood cells


Cells adapted to carry information in the form of electrical impulses


A chemical messenger produced by a particular gland or cells of the endocrine system. Hormones are transported throughout the body in the blood stream but they produce a response only in specific target cells


A group of cells in an organism that are specialised to work together to carry out a particular function.


A large organ in the upper abdomen which manufactures, stores and breaks down substances as required by the body

Cell division, mitosis and cancer

Multi cellular organisms, like humans, are made up of billions of cells. These cells need to divide and copy themselves for a variety of reasons. For example:

  • cells wear out and need to be replaced
  • new cells allow the body to repair damaged tissue
  • new cells allow the body to grow


The most common form of cell division is called mitosis . It is used for growth and repair. During mitosis, a cell makes an exact copy of itself and splits into two new cells. Each cell contains an exact copy of the original cell's chromosomes in their 23 pairs. This is the reason why all the cells in an organism are genetically identical.

To find out more about mitosis go to the ABPI genes and inheritance resource.

Cancer: mitosis out of control

Mitosis is closely controlled by the genes inside every cell. Sometimes this control can go wrong. If that happens in just a single cell, it can replicate itself to make new cells that are also out of control. These are cancer cells. They continue to replicate rapidly without the control systems that normal cells have. Cancer cells will form lumps, or tumours, that damage the surrounding tissues. Sometimes, cancer cells break off from the original tumour and spread in the blood to other parts of the body. When a tumour spreads to another part of the body it is said to have metastasized. They continue to replicate and make more tumours. These are called secondary tumours.

Medicines that are used to treat cancer are sometimes aimed at killing cells that are rapidly dividing by mitosis. They inhibit the synthesis or function of DNA - this type of treatment is called chemotherapy. More modern medicines target specific cancers in different ways. Many inhibit the growth signals for that type of cell.

Fighting cancer: stopping tumour cells from growing

There are many different types of cancer. They depend on which type of cell was the original one that started to replicate out of control. This means that there is not just one treatment for cancer. Treatments may include a combination of surgery, medicines and radiation therapy (radiotherapy).

As researchers have come to understand more about cancers, new and targeted therapies are constantly being developed. For example, a type of breast cancer that is influenced by the hormone oestrogen can be treated with hormone therapy that blocks the action or synthesis of oestrogen. Other medicines can block growth signals to the cancer cell and so slow the development of a tumour or block the growth of new blood vessels into tumours. This effectively 'starves' the cancer cells of the nutrients they need to grow.

Question 3

Mitosis is involved in the growth, repair and replacement of cells. Not all cells go through mitosis at the same speed.

Look at the types of cell below and decide how often they are replaced by mitosis. In each case, choose an answer using the radio buttons.

3 Discussion

Tumor heterogeneity in hepatobiliary tumor represents a main obstacle to personalized cancer treatment. Thus, it is highly desirable to explore such heterogeneity and its impacts on drug response using a research model that can faithfully recapitulate the in vivo phenotype of hepatobiliary tumors. Single-cell genomic provides a viable strategy to understand the genetic and phenotypic diversity at the single-cell level, which may also help to understand complex ecosystems in tumor. Here, we applied scRNA-seq to characterize patient-derived hepatobiliary tumor organoids, which was currently recognized as the powerful tumor research model. We found evidence of inherent variable of transcriptional programs related to cell cycle and epithelial expression across hepatobiliary tumor organoids. Biological and transcriptomic heterogeneity of CSCs within tumor organoids were also found, which was related to chemo-resistance. Interestingly, further analysis revealed that resistant subpopulations with unique metabolic circuitry were response to distinct molecular signatures and drug resistance. Our findings may provide a mechanistic explanation as to why some patients respond while others do not, and provide insight into the heterogeneity of hepatobiliary tumor organoids and define drug resistance associated with CSCs.

Among our key findings is the identification of biological and transcriptomic heterogeneity in patient-derived hepatobiliary tumor organoids. Hepatobiliary tumors are characterized by a high degree of tumoral heterogeneity. Unlike other malignant tumors, such as breast cancer or lung cancer that has multiple markers of genetic mutations to determine tumor behaviors, the gene mutation spectrum of hepatobiliary tumors is wide and lacks clear characteristics, resulting extensive genetic and phenotypic variation. In this study, by generating transcriptional atlas of hepatobiliary tumor organoid in single-cell level, we showed that different samples are inherently various in cell cycle and epithelial expression, which is consistent with their proliferation ability, potential drug-resistance risk, and tumoral malignancy, respectively. Notably, we further identified multiple reported malignancy-related genes (e.g., MET, PIK3R1, PRKCA, PTEN, SHC1, and STAT3) upregulated in HCC272, and demonstrated that these genes were mainly enriched for cancer-related functions, which were associated with broad drug resistance.

Another important finding is the presence of CSC heterogeneity within tumor organoids, which tempers our understanding of drug resistance. CSC plasticity [ 4 ] is a prominent cause of genetic heterogeneity in cancer, playing a vital role in tumor survival, proliferation, metastasis, and recurrence. First, cell surface markers analyses clearly showed that CSCs varied greatly among individual organoids. Second, it is of interest to note that CD44 high cells in HCC272 exhibited a distinctive pattern, which might cause distinct transcription and more drug resistance than other tumor organoids. Third, by exploring co-expressed genes, trajectory, and pseudo-time analysis, we defined the CTNNB1-enriched subpopulation as the proliferation advantage cluster and the GAPDH-enriched as the metabolism advantage one. Specifically, the metabolism advantage organoid HCC272 could remodel tumor microenvironment through accelerating the usage of glucose, enhancing hypoxia-induced HIF-1 signaling, and leading to the upregulation of NEAT1 in CD44 high cells, which induce the hyper-activation of Jak-STAT signaling eventually caused drug resistance. It is suggested that understanding the distinctive metabolic circuitry in resistant subpopulations may help us characterize the CSC heterogeneity and predict therapeutic response.

A limitation of this study is that the findings are mainly based on a small number of clinical samples, and the interpretation of tumor heterogeneity is somewhat limited. However, scRNA-seq still provides deconstructive analysis and the discovery of potential mechanisms provides credible help for precision treatment of individuals. Encouragingly, we have now established a huge hepatobiliary tumor organoid biobank based on gene variation spectrum and genetic characteristics of the Chinese population that might allow us to acquire high-quality single cells for single-cell transcriptome analysis. Further studies will be focused on using our established patient-derived hepatobiliary tumor organoid biobank to perform integrative analysis from multiple-levels, including genome, transcriptome, metabolome, and epigenome, to provide more valuable resources for clinical practice.

In summary, our study herein provides important insights into hepatobiliary tumor heterogeneity, especially the diversification of CSC distribution and the complexity of cell evolution trajectory. Meanwhile, we revealed that CD44 positive subpopulation is responsible for drug resistance by hyper-activating Jak-STAT signaling pathway, which is induced by NEAT1 upregulated in hypoxia burden. Further studies with larger sample size should be warranted to better clarify the association between tumor heterogeneity and unfavorable clinical features after resection or drug treatment.


Tumor tissue acquisition and processing

We acquired fresh tumor tissue from patients undergoing surgical resection for glioma. De-identified samples were provided by the Neurosurgery Tissue Bank at the University of California San Francisco (UCSF). Sample use was approved by the Institutional Review Board at UCSF. The experiments performed here conform to the principles set out in the WMA Declaration of Helsinki and the Department of Health and Human Services Belmont Report. All patients provided informed written consent. Tissues were minced in collection media (Leibovitz’s L-15 medium, 4 mg/mL glucose, 100 u/mL Penicillin, 100 ug/mL Streptomycin) with a scalpel. Samples dissociation was carried out in a mixture of papain (Worthington Biochem. Corp) and 2000 units/mL of DNase I freshly diluted in EBSS and incubated at 37 °C for 30 min. After centrifugation (5 min at 300 g), the suspension was resuspended in PBS. Subsequently, suspensions were triturated by pipetting up and down ten times and then passed through a 70-μm strainer cap (BD Falcon). Last, centrifugation was performed for 5 min at 300 g. After resuspension in PBS, pellets were passed through a 40-μm strainer cap (BD Falcon), followed by centrifugation for 5 min at 300 g. The dissociated, single cells were then resuspended in GNS (Neurocult NS-A (Stem Cell Tech.), 2 mM L-Glutamine, 100 U/mL Penicillin, 100 ug/mL Streptomycin, N2/B27 supplement (Invitrogen), sodium pyruvate).

CD11b + cell isolation

In total, 20 μL of CD11b microbeads (Miltenyi Biotec 130-093-634) were mixed with the 80 μL single-cell suspension (produced as above) in PBS supplemented with 2 μM EDTA and 0.5% bovine serum albumin (BSA) (MACS buffer) and incubated at 4 °C for 15 min. Cells were washed twice with MACs buffer, centrifuged for 10 min at 300 g, and resuspended in MACs buffer. The suspension was then applied to a MACS LS column in the magnetic field of a MACS Separator. Columns were washed three times with MACs buffer and magnetically labeled cells were then flushed into a collection tube. The purity of CD11b + cells was assessed via flow cytometry: CD11b + and CD11b– fractions were staining with phycoerythrin-conjugated anti-CD11b (Clone M1/70) antibody for 15 min cells were then washed twice and analyzed on a FACsCaliber flow cytometer using FACSDIVA software (Additional file 2: Figure S1a).

Single-cell RNA sequencing

Fluidigm C1-based scRNA-seq

Fluidigm C1 Single-Cell Integrated Fluidic Circuit (IFC) and SMARTer Ultra Low RNA Kit were used for single-cell capture and complementary DNA (cDNA) generation. cDNA quantification was performed using Agilent High Sensitivity DNA Kits and diluted to 0.15–0.30 ng/μL. The Nextera XT DNA Library Prep Kit (Illumina) was used for dual indexing and amplification with the Fluidigm C1 protocol. Ninety-six scRNA-seq libraries were generated from each tumor/Cd11b + sample and subsequently pooled for 96-plex sequencing. cDNA was purification and size selection was carried out twice using 0.9X volume of Agencourt AMPure XP beads (Beckman Coulter). The resulting cDNA libraries were quantified using High Sensitivity DNA Kits (Agilent).

10X genomics-based scRNA-seq

Tissue was dissociated by incubation in papain with 10% DNAse for 30 min. A single-cell suspension was obtained by manual trituration using a glass pipette. The cells were filtered via an ovomucoid gradient to remove debris, pelleted, and resuspended in Neural Basal Media with serum at a concentration of 1700 cells/uL. In total, 10.2 uL of cells were loaded into each well of a 10X Chromium Single Cell capture chip and a total of two lanes were captured. Single-cell capture, reverse transcription, cell lysis, and library preparation were performed per manufacturer’s protocol.

Sequencing for both platforms was performed on a HiSeq 2500 (Illumina, 100-bp paired-end protocol).

Exome-sequencing and genomic mutation identification

The NimbleGen SeqCap EZ Human Exome Kit v3.0 (Roche) was used for exome capture on a tumor sample and a blood control sample from each patient. Samples were sequenced with an Illumina-HiSeq 2500 machine (100-bp paired-end reads). Reads were mapped to the human grch37 genome with BWA [43] and only uniquely matched paired reads were used for analysis. PicardTools ( and the GATK toolkit [44] carried out quality score re-calibration, duplicate-removal, and re-alignment around indels. Large-scale (>100 Exons) somatic copy number variants (CNVs) were inferred with ADTex [45]. To increase CNV size, proximal (< 1 Mbp) CNVs were merged. Somatic SNVs were inferred with MuTect ( for each tumor/control pair and annotated with the Annovar software package [46].

Single-cell RNA-sequencing data processing and neoplastic-cell classification

Data processing was performed as described previously [14]. Briefly, reads were quality trimmed and TrimGalore! ( clipped Nextera adapters. HISAT2 [47] was used to perform alignments to the grch37 human genome. Gene expression was quantified using the ENSEMBL reference with featureCounts [48]. Only correctly paired, uniquely mapped reads were kept. In each cell, expression values were scaled to counts per million (CPM). Low-quality cells were filtered by thresholding number of genes detected at 800 and at least 50,000 uniquely aligned reads. tSNE plots visualizing groupings of cells were carried out using the Seurat R package [49]. CNVs that were called in matched exome-seq data were quantified in individual cells as previously described [15]. Briefly, megabase-scale CNVs were identified in tumor/normal paired exome-seq datasets, and then quantified in individual cells using a control sample from non-malignant brain.

Public data acquisition

Expression matrices from bulk RNA-seq (performed in triplicate) were downloaded from GEO for the following samples: representing BMDM, we obtained M0 (GSE68482) [50], M1, M2 macrophages (GSE36952) [51], and monocytes (GSE58310) [52]. We also obtained data for microglia purified from epilepsy-related surgical specimen (n = 3) and post-mortem brain (n = 5) (GSE80338) [53]. Lists of genes that are differentially expressed between blood-derived murine TAMs and microglial murine TAMs, in two murine glioma models, were downloaded [9]. Normalized scRNA-seq counts were obtained from GEO for astrocytoma (GSE89567) and oligodendroglioma (GSE70630). Analysis was restricted to TAMs, as classified in the BROAD single-cell data portal ( Normalized counts from TCGA RNA-seq data were obtained from the Genomics Data Commons portal ( Patients diagnosed as GBM and wild-type IDH1 expression (n = 144) as well as those with LGG classification and IDH1 mutation (n = 414), as given in [54], were normalized to log2(CPM + 1) and used for analysis. Z-score normalized counts from regional RNA-seq of 122 samples from ten patients was obtained via the web interface of the IVY GAP ( database. Furthermore, images of in situ RNA hybridizations in glioma tissue sections were downloaded for two patients: BIN1: W11-1-1-E.1.04, 57-year-old man, glioblastoma TGFBI: W8-1-1-B.1.04, 49-year-old woman, glioblastoma.

Derivation of ontogeny-specific expression signatures

Genes differentially expressed between blood-derived TAMs and microglial TAMs, recurrently in both of Bowman et al.’s two murine glioma models, were used as a starting point [9]. We identified homologues of these differentially expressed mouse genes with the biomaRt package in R [55]. The resulting set of genes was filtered for genes expressed in our human-TAM scRNA-seq data. Genes with a mean expression > 1 CPM were retained. This set of genes was used as the basis for subsequent PCA and single-cell consensus clustering (SC3). Expression values, defined as log2 (CPM/10 + 1), of genes in the human-TAM scRNA-seq data were z-score normalized, across cells from within each single-cell platform (SMARTer vs SMART-Seq2) independently. Subsequently, PCA followed by Varimax rotation was performed. Sample scores, along PC1, were partitioned using a two-component Gaussian mixture model. Genes strongly associated with PC1 in either direction were identified by applying a threshold of abs(loading) > 0.2 to the gene loadings. MFA was performed on the Smart-Seq2 and C1 data, using the FactoMineR ( R package, using the 237 mouse homologue genes. Genes strongly loading PC1 in the PCA were compared to RNA-seq data from microdissections of defined glioma anatomical structures, via the IVY atlas (, and visualized with morpheus ( SC3 clustering [56] (k = 2) was also performed in the human-TAM scRNA-seq data, restricted to the set of human counterparts of lineage-specific murine-TAM genes. Both classifiers produced highly similar classification results (Matthews correlation coefficient = 0.946). To identify genes significantly co-occurring in single cells, we calculated the odds ratios (OR) and p values as described in [57]. P values were corrected for multiple testing with Benjamini–Hochberg.

Calculation of ontogeny scores and survival analysis

For each sample in the TCGA dataset (described above) we calculated the average expression of microglial-TAM genes and blood-derived TAM genes, respectively. To compare the relative amount of infiltration between glioma subtypes, we utilized the glioVis portal [58] to classify isocitrate dehydrogenase 1/2 (IDH1/2) wild-type GBM samples into three transcriptional subtypes: Classical Mesenchymal and Proneural. IDH1/2-mutant LGGs were subdivided into astrocytomas (n = 110) and oligodendrogliomas (n = 117) based on histology and the presence/absence of a 1p/19p co-deletion.

For both the microglial and blood-derived TAM survival analysis, Progene V2 was used. High- and low-expression cohorts were defined as cases with expression scores above and below the median score, respectively. GBMs and LGGs were considered separately. We adjusted for age and gender by adding these covariates to a cox proportional hazards model [59].

Analytical flow cytometry

De-identified fresh glioma tissues were obtained as described above, in “Tumor tissue acquisition and processing.” Tissue was mechanically dissociated, resuspended in 70% Percoll (Sigma-Aldrich), overlaid with 37% and 30% Percoll, and centrifuged for 20 min at 500 × g. Enriched leukocyte populations (TIL) were recovered at the 70–37% interface, washed twice in PBS, and resuspended in flow staining buffer (PBS + 1% BSA) containing Human TruStain FcX (Biolegend). Cells were then incubated at 4° for 30 min with antibodies, washed twice in flow staining buffer, and analyzed on a BD FACSAria cell sorter.

The following antibodies were purchased from Biolegend: FITC anti-mouse/human CD282 PE anti-human P2RY12 PE/Cy7 anti-human CD204 APC/Fire™ 750 anti-mouse/human CD11b APC anti-human CD49d PerCP/Cy5.5 anti-human HLA-DR and BV421 anti-human CD206. All antibodies were used according to the manufacturers’ recommended usage.


Cancer remains one of the major malignant diseases that endangers human life and health and comprises complex biological systems that require accurate and comprehensive analysis. Since the first appearance of high-throughput sequencing in 2005 [1], it has become possible to understand life activities at the molecular level and to conduct detailed research to elucidate the genome and transcriptome. As an essential part of high-throughput sequencing, RNA sequencing (RNA-seq), especially single-cell RNA sequencing (scRNA-seq), provides biological information on a single tumor cell, analyzes the determinants of intratumor expression heterogeneity and identifies the molecular basis of formation of many oncological diseases [2, 3]. Thus, RNA sequencing offers invaluable insights for cancer research and treatment. With the advent of the era of precision medicine, RNA sequencing will be widely used for research on many different types of cancer. This review summarizes the history of the development of RNA sequencing and focuses on the latest studies of RNA sequencing technology in cancer applications, especially single-cell RNA sequencing and spatial transcriptome sequencing. In addition, we provide a general introduction to the current bioinformatics analysis tools used for RNA sequencing and discuss future challenges and opportunities for RNA sequencing technology in cancer applications.

Further Use of Single Cell Transcriptomics

Apart from using single-cell RNAseq to analyze transcriptomics to define cell types, its combination with other molecular and bioinformatics technics can increase its power to study immune compartments of the TME. Thus, single-cell transcriptomics can be combined with other omics technics. For example, the combination of single-cell genomics and transcriptomics can help to draw a connection between alterations in the cancer genome and its influences on immune cells. This is possible for example with methods like G&T-Seq, that separate poly-adenylated RNAs from genomic DNA by using biotinylated poly(A) primer prior to sequencing (193). ScTrio-Seq goes even further by combining three omics approaches, genomics, transcriptomics, and epigenomics on the same single cell (152). This has the potential to also study the influence of epigenetic changes on cells of the TME by simultaneous analyzing the genome and transcriptome.

Single-cell RNAseq data could be used for deconvolution of bulk RNA-seq data. Deconvolution infers by mathematical modeling the presence and proportion of cells of a specific compartment in a complex tissue based on the expression of reference genes in a certain compartment (178). This strategy can be applied to large databases built from bulk RNAseq data of tumor tissues (such as the TCGA) to analyze the composition of the TME. A deconvolution approach called Epigenomic Deconvolution (EDec) has been developed to in silico model the cell composition of complex tissues, in this case breast tumors, based on DNA methylation profiles (179). The algorithm uses reference profiles of a certain tissue to select loci based on different methylation profiles (feature selection). RNAseq data is then dissected into subgroups based on a reference-free deconvolution approach based on their molecular profile. These two datasets are then combined to identify cell types by comparing molecular profiles with methylation profiles. With this approach, the proportion of immune cells in breast tumors was inferred and used to predict patient outcome. Single-cell epigenomics data could improve this approach by producing information of methylation pattern of rare cell types to get an even more accurate modeling of cell type compositions.

Another application of scRNAseq data is their usage for reconstructing cell trajectories to follow cell differentiation. To model cell state transitions several bioinformatics algorithms, using unsupervised or supervised (with a little prior information about cell types and marker gene expression) modeling, have been developed to reconstruct cell trajectories (194�).

A very recent published algorithm is CellRouter, which is very robust in reconstructing trajectories between early cell states and transitory cell states to model cell differentiation (197). CellRouter works in a way that is not dependent on a priori knowledge about cell structure relationships. Single-cell data are first split into subpopulations based on community detection algorithms and then trajectories are determined across subpopulations based on calculated weights, which are weaker between unrelated cell types and stronger within related subpopulations.

ScRNAseq has a great potential to increase our knowledge about immune cancer tolerance to help us find new targets for immunotherapies. However, the main limitation of scRNAseq approaches is that cells are isolated from their environment, making difficult the analysis of connections between cells of different compartments (198, 199). The field of single-cell spatial transcriptomic is of intense research and multiple technologies have been developed in the last few years like seqFISH (200), MERFISH (201), FISSEQ (202), or TIVA (203). SeqFISH and MERFISH use single-molecule FISH (smFISH) techniques, using probes that bind to the same mRNA, but have different fluorophores attached. During several rounds of in situ hybridization and stripping off probes RNAs get a unique fluorescent barcode. This allows the simultaneous detection of many transcripts, although theoretically the usage of 4 dyes and 8 rounds of hybridization would cover the whole transcriptome (4 8 = 65,5336) these methods are based on background information about marker genes that could be provided by scRNAseq data. The FISSEQ (fluorescent in-situ) technique uses reverse transcription in situ to convert RNA into cross-linked cDNA amplicons followed by a sequencing-by-ligation technique (SOLiD). Finally, TIVA (transcriptome in vivo analysis) uses a photoactivatable biotin-labeled TIVA-tags, which upon photoactivation enable mRNA capture from single cells in live tissue. All these methods allow the detection of expressed genes in vivo in the context of a specific tissue architecture, and thus inference of the interactions between different cell types. Thus, they are a potential approach to study the connection of epithelial cancer cells and the TME in vivo to model the influence of immune cells to cancer cell progression.

G. Zalcman has received speakers bureau honoraria from BMS and AstraZeneca and is a consultant/advisory board member for BMS, AstraZeneca, MSD, and Roche. F. Mechta-Grigoriou received research support from Innate-Pharma, Roche, and BMS. No potential conflicts of interest were disclosed by the other authors.

Conception and design: G. Zalcman, A. Vincent-Salomon, F. Mechta-Grigoriou

Development of methodology: Y. Kieffer, H.R. Hocine, G. Gentric, L. Albergante

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): H.R. Hocine, G. Gentric, F. Pelon, B. Bourachot, S. Lameiras, C. Bonneau, A. Guyard, S. Baulande, G. Zalcman, A. Vincent-Salomon

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): Y. Kieffer, C. Bernard, L. Albergante, C. Bonneau, K. Tarte, A. Zinovyev, F. Mechta-Grigoriou

Writing, review, and/or revision of the manuscript: Y. Kieffer, K. Tarte, A. Zinovyev, G. Zalcman, A. Vincent-Salomon, F. Mechta-Grigoriou