Systems Biology of the Cancer Cell

Abstract

Questions in cancer have engaged systems biologists for decades. During that time, the quantity of molecular data has exploded but the need for abstractions, formal models, and simplifying insights has remained the same. This review brings together classic breakthroughs and recent findings in the field of cancer systems biology, focusing on cancer-cell pathways for tumorigenesis and therapeutic response. Cancer cells mutate and transduce information from their environment to alter gene expression, metabolism, and phenotypic states. Understanding the molecular architectures that make each of these steps possible is a long-term goal of cancer systems biology pursued by iterating between quantitative models and experiments. We argue such iteration is the best path to deploying targeted therapies intelligently so that each patient receives the maximum benefit for their cancer.

1. INTRODUCTION

Today’s descriptions of complexity in cancer far outpace our understanding. Extracting insight from the glut of raw information requires holistic ways to synthesize the known and suggest something organizationally new. We need simplifying quantitative concepts that generalize across cancer types and subtypes. A framework for achieving these goals was drafted by the systems biology field and refined for more than 20 years. It is too early to declare success or failure, but no other disease has sustained as much support for (and interest from) systems biologists as cancer.

This review traces the path of cancer systems biology—its origins, status, and future trajectory. We cover fundamental cancer cell biology and the application of systems principles to cancer phenotypes. We hope this structure is useful to trainees and others seeking to appreciate seminal work in the field. We needed to make hard choices about what to include (Section 1.2) and apologize to those whose contributions were omitted. Consistent with this Annual Review series, our view skews toward engineering (1). Foundational contributions have certainly been made by systems biologists who are not engineers, but we may present their work in engineering terms. Such distinctions should continue to fade now that the first generation of native cancer systems biologists has transitioned to independence.

1.1. From systems to systems biology to cancer systems biology

The term “system” applies to any set of interconnected parts that coordinate an operation or function. New properties may arise from connected systems that are not obvious when considering the parts in isolation. Systems-level properties are assessed with a computational model—any mathematically simplified representation that is based on theory or data and updatable with new evidence (Figure 1a–c). A systems model recognizes the parts, their interconnections, and boundaries that delimit the system. Defining systems boundaries is a challenge for biology, where everything seems linked, but it is crucial for tractability and interpretability of the accompanying model.

Figure 1 |.

Model-experiment iterations in cancer systems biology. (a) Build-first paradigm of physicochemical models (9). (b) Experiment-first paradigm of data-driven models (10). (c) Iterate-first paradigm of prior-knowledge network models (11). (d) Experiment-analysis pipeline of genomics and proteomics. Steps are numbered sequentially.

The first introduction of “systems biology” as a discipline was highly algorithmic—identify all the components, perturb them, measure them, revise the model (2). This pioneering vision laid out systems biology as a deductive hypothesis test of a quantitative or logical model with inductive experimental methods that were hypothesis-generating. It was limited, however, to organisms where exhaustive profiling and perturbation were immediately feasible (e.g., yeast, bacteria). Moreover, the algorithm pre-required a working mechanistic model. At the time, cancer biology was just beginning to organize around disease hallmarks (3), and systems thinking existed only by analogy to engineered circuit boards (4). Early systems biology was remarkably ahead of cancer research at that time.

We now know cancer is a disease that calls for its own kind of systems biology. Malignancy is caused by multiple genetic alterations that permanently break or create regulatory connections in biomolecular subnetworks (5; 6), giving the disease a competitive advantage to progress. No other human malady somatically and heritably rewires our regulatory networks, thus motivating several of the “cancer systems biology” approaches used today (7; 8). Cancer systems biology is like NASA’s response to a malfunction—forensic reassembly of a broken parts list with enough depth to inform past, future, and alternate trajectories.

Cancer-associated alterations are described by omics of different kinds and bioinformatics (Figure 1d), which relate to cancer systems biology but have different goals. Omics elaborates the variance of cancer, and bioinformatics provides tools to organize this variance toward generating hypotheses. What distinguishes cancer systems biology is the central role of a computational model and the cyclic iteration of that model with experiments. The need for a model could stem from the complexity of the biological question or the data needed to address it. Cancer systems biology models may predict new results or explain old ones. Regardless, the computational model formalizes a hypothesis that is testable by experiments and revisable based on the outcome.

Cancer systems biologists often generate large datasets and use computational algorithms to visualize or analyze them, but these methods are secondary to model-experiment iteration in pursuit of new insight. Particular confusion arises in studies that create or use networks to connect biological components. A network might be purely descriptive; or, it could yield predictions and explanations that are authentically revisable by hypothesis-testing experiments. Model refinement can take many forms, but without iteration the study of cancer systems drifts into endless data collection and hypothesis generation (Figure 1).

1.2. Scope of the review

We focus on cancer systems biology of the malignant cell to complement recent systems biology-themed reviews of whole-tumor modeling (12) and cancer metastasis (13). Taking the cancer cell as a collection of dynamic signaling, transcriptional, metabolic, and epigenetic subsystems, we summarize the intracellular pathways and networks that became systems biology landmarks because of the models built upon them. Tools and formalisms used to construct systems models are mentioned but deemphasized so that key cancer insights are not obscured by methodological details (Figure 1a–c). We omit recent applications of machine learning (14) while remaining enthusiastic about future network-constrained architectures that improve interpretability (15; 16). Similarly, evolutionary cancer models are only just now becoming experimentally constrainable with molecular detail (17). Instead of surveying all mathematical approaches, we gathered cancer systems biology research that influenced the most-widespread methods today.

Strictly adhering to the cancer cell as our system requires us to exclude several topics of broader interest. Notably, we avoid the ever-expanding universe of reciprocal cell-cell interactions in the tumor microenvironment. It is not meant to imply that tumor immunology, engineered or native paracrine–juxtacrine circuits, progenitor-cell crosstalk, or collective cell behaviors are unimportant; there are simply better cancer systems biologists to review those topics than us (18). Despite the intentional omissions and surely a few inadvertent oversights, we hope our assembled Systems Biology of the Cancer Cell achieves its goals and remains timely for years to come. We comply with space limitations by allocating subsections between the main and supplementary text (Table 1). Earliest work on a pathway is covered in the supplementary text in favor of highlighting the most current understanding in the main text.

Table 1 |.

Organization of Sections 2 and 3 of this review.

2. Systems approaches to cancer biology
	2.1. Oncogenes	2.2. Microenvironment	2.3. Transcription	2.4. Metabolism	2.5. State switches
M	2.1.1. RTKs I (1)	2.2.3. Death receptors I (2)	2.3.1. bZIP TFs (3; 4)	2.4.2. Gains–losses I (5; 6)	2.5.4. EMT (7–9)
M	2.1.2. Ras–ERK I (10–13)	2.2.5. Oxidative stress (14; 15)	2.3.3. Master regulators I (16–19)		2.5.5. Persistence I (20–22)
M	2.1.5. Cell cycle I (23–25)	2.2.6. DNA damage (26–33)
S	2.1.1. RTKs II (34; 35)	2.2.1. ECM adhesion (36–39)	2.3.2. NF-κB (40–42)	2.4.1. Tumor–normal (43; 44)	2.5.1. Epigenetics (45–47)
S	2.1.2. Ras–ERK II (48; 49)	2.2.2. TGFβ (50–55)	2.3.3. Master regulators II (56)	2.4.2. Gains–losses II (57; 58)	2.5.2. Epi-adaptations (59; 60)
S	2.1.3. PI3K–Akt (61)	2.2.3. Death receptors II (62–64)		2.4.3. Hallmarks (65; 66)	2.5.3. Metastability (67–70)
S	2.1.4. Wnt–β-catenin (71–74)	2.2.4. Hypoxia (75; 76)			2.5.5. Persistence II (77)
S	2.1.5. Cell cycle II (78–80)
3. Applications of cancer systems biology
		3.1. Phenotypes	3.2. Pharmacology	3.3. Individualization
M		3.1.1. Proliferation (81–83)	3.2.1. MoA (84–87)	3.3.1. Therapy (88–93)
M			3.2.4. Dynamics (94; 95)
S		3.1.2. Cell death (96–98)	3.2.2. Adaptive–bypass (99–101)	3.3.2. Onco-metabolism (102–105)
S		3.1.3. Migration (106–109)	3.2.3. Combinations (110–112)
S			3.2.5. Drug discovery (113–115)

M, main text. S, supplementary text.

2. SYSTEMS APPROACHES TO CANCER BIOLOGY

The earliest forays of systems biology into cancer occurred through intracellular pathways that were evolutionarily conserved, biochemically defined, and directly related to hallmarks of proliferation, death, or stress. Studies of normal pathway function set the stage for later investigations of how system properties change with oncogenic insults. The architecture and discovery timeline of pathways affected how much attention they garnered from budding systems biologists. For example, the Raf–MEK–ERK pathway was elaborated in the early 1990s, and its three-tiered architecture has been usefully modeled dozens of times for different purposes (Section 2.1.2). By contrast, Hippo-YAP/TAZ arrived ten years later, and there are no established computational models of normal regulation or cancer misregulation because its mechanisms and feedbacks remain enigmatic. Excellent signaling biochemistry in cancer had been diluted out by a tidal wave of sequencing (19), despite the importance for metabolism (Section 2.4) and cytoskeletal regulation (Section 3.1.4).

Section 2 organizes pathway classes from the plasma membrane to the nucleus and orders them by characteristic time scales (t_char) from smallest to largest. Approximations based on time scales (e.g., pseudo-steady processes) are helpful for defining system kinetic boundaries of the system (Section 1.1). Biology often challenges these approximations by aggregating core pathways and processes with similar time scales. That issue notwithstanding, the most-durable systems models make justifiable compromises about their operational t_char and stay true to that scaling assumption.

2.1. Oncogenic signaling networks (t_char ~ seconds–minutes)

With its historical emphasis on biophysical and enzymological quantitation, signaling biochemistry is a natural fit for systems approaches. Signal-transduction pathways and the networks surrounding them are laden with bona fide oncogenes and tumor suppressors, which uncouple homeostatic regulation and drive cancer hallmarks autonomously. We summarize non-intuitive findings about regulation and misregulation that derived from computational models of core signaling circuitry.

2.1.1. Receptor tyrosine kinases.

Receptor tyrosine kinases (RTKs) are some of the most mechanistically well-studied initiators of cell signaling. The epidermal growth factor receptor (EGFR) is the quintessential example because the complex structure-function relationships and ordered steps leading from ligand binding to receptor tyrosine phosphorylation are known, and rate constants have been measured for many steps. This knowledge led to mechanistic models, often consisting of coupled ordinary differential equations, that connect the intermediate steps between ligand binding, receptor phosphorylation, and downstream pathway activation by laws of mass action. These models seek to resolve non-intuitive observations of signaling dynamics or to identify the key context-dependent rate processes that lead to signal attenuation, propagation, amplification, switch-like activation, or robustness toward a goal of signal engineering for therapeutic purposes. RTK models often describe normal biology but extend to cancer cell settings.

RTKs straddle extracellular and intracellular environments and undergo intermolecular and intramolecular interactions in both domains. Such interactions are perturbed by changes in RTK localization during trafficking, which generally occurs with larger time scales than do RTK catalytic and binding interactions. RTKs are thus among the most intrinsically multiscale proteins in signal transduction. Systems biology models often capture this complexity through compartmentation, with the well-mixed assumption enabling coupled ordinary differential equations. Compartmentalized models have revealed fundamental aspects of RTK biology, including the similar EGFR dephosphorylation kinetics at the plasma membrane and within endosomes (20) and a role for receptor endocytosis in resetting cells for robust fold-change responses to growth factor (21). Furcht and coworkers built a model to explain the experimentally observed maintenance of GAB1-SHP2 complexes in the cytosolic compartment by membranous EGFR through intermediary Src family kinases (22). Repeatedly, systems models pinpointed trafficking as a regulator of overall receptor phosphorylation (e.g., in response to immobilized vs. diffusible isoforms of VEGF (23)) or downstream signaling (e.g., different RTK agonists of angiogenesis (24)).

RTKs transmit information through shared oncogenic pathways (Sections 2.1.2 and 2.1.3), and models have quantified synergy or cooperativity strengths at a convergent downstream node (25) as well as the strength and directionality of pathway crosstalk mechanisms downstream of common RTK ligands (26; 27). Larger network models quantified and explained qualitative aspects of signaling biology that were long appreciated by experimentalists, including the ligand dependence of system sensitivity (28) and the robustness of RTK networks to individual processes perturbations (27).

Modeling studies of cancer-specific mutations, drugs, or signaling often point to the processes already discussed. For example, receptor trafficking was key to modeling how a peptide-mimetic interferes with MET/integrin receptor interactions (29). Elsewhere, impaired EGFR endocytosis tied to constitutive EGFR activation was required to explain the preferential signaling of the mutant receptor in non-small cell lung cancer (30). Elaborated models that resolve RTK spatial effects have revealed specific therapeutic vulnerabilities in systems that rely heavily on clustering for activation, such as the RTK AXL (31).

2.1.2. Ras–ERK.

The Ras GTPase activates a three-tiered MAPK cascade of RAF, MEK, and ERK to transmit a mitogenic signal that is co-opted by upstream mutation or amplification in many cancers. The Ras–ERK architecture is conserved across biology and most of its constituents have approved drugs, making it one of the most highly modeled pathways. Models have focused on mechanisms of signal amplification–attenuation, ultrasensitivity, and stability given that ERK dynamics control cell fates as much as or more than steady-state outputs.

Surrounding the RAF–MEK–ERK cascade, there are multiple negative feedback and feedforward loops whose roles have been examined for system properties, pathway crosstalk, and response to interventions. Asthagiri and Lauffenburger (32) incorporated feedback from ligand-receptor interactions through a fully elaborated RAF–MEK–ERK cascade to identify when the ERK pathway attenuates before the initiating receptor has fully activated. The model encodes negative feedback inspired by ERK-catalyzed phosphorylation events that disrupt GRB2-SOS interaction and RAF phosphorylation-dependent membrane localization. Both modes of negative feedback explained experimental data, but the adaptor feedback mechanism generated crosstalk with other pathways that share the adaptor, and later studies favored the RAF phosphorylation as the dominant negative feedback (33).

Drawing an analogy to circuit design, Sturm and coworkers likened the feedback-endowed ERK pathway to a negative feedback amplifier from circuit design (34). ERK activation is blunted by these feedbacks such that it exhibits a less switch-like activation than do standard amplifiers. Beyond the Ras–ERK core, additional pathway nodes participate in positive- and negative-feedback loops. For example, models including Raf kinase inhibitor protein (RKIP) predict that it simultaneously promotes ERK oscillatory responses and switch-like behavior of MEK (35).

Feedback architectures can produce unanticipated consequences for different types of perturbation. Jensen and coworkers (36) found by modeling of three-tiered A → B → C cascades that negative feedback from C –| A favored drug perturbation of B over knockdown. The prediction was verified with inhibitors and shRNA targeting MEK, and extended models with different feedback architectures predicted when such differences in potency emerge.

Regulation of mutant Ras has been of great interest for systems analysis to gain better understanding of the hydrolysis, binding, and reaction steps that engender oncogenic activity. An early contribution by Stites and coworkers used a literature-constrained mass-action model to identify parameters whose univariate perturbation most affected Ras activity (37). The four parameters—GDP dissociation, GDP and GTP catalysis, and GDP binding— are altered by mutations that commonly arise in cancer, such as the G12V mutation that reduces sensitivity to GTPase-activating proteins (GAPs). An extension of this approach was later used to predict the unique sensitivity of the G13D Ras mutant to EGFR antagonism based on its distinct biochemical properties that were encoded (38).

2.1.5. Cell cycle entry–exit.

The normal mammalian cell cycle contains multiple decisive transitions enforced by feedback (39). To achieve self-sufficient proliferation (3), cancers must evade cell-cycle checkpoints that ordinarily trigger quiescence or arrest. Systems biology made early contributions to understanding the restriction point (40), defined by the duration of growth-factor stimulation sufficient for irreversible exit from G0 and G1 entry. Central to the restriction point is the tug-of-war between Rb-family tumor suppressors and E2F-family transcription factors that regulate cyclins important for G1 (cyclin E) and S (cyclin A). In quiescent cells, Rb inhibits E2F activity and the restriction point is active. Growth factors induce a separate G1 cyclin (cyclin D), which activates CDK4/6 kinases that phosphorylate and inhibit Rb. Induction of cyclin E by disinhibited E2F leads to activation of CDK2, which phosphorylates and inhibits Rb. This double-negative feedback is the minimum requirement for all-or-none bistability in E2F activity (41) (Figure 2a).

Figure 2 |.

Key regulatory motifs in the G0/G1 restriction point. (a) Rb–E2F–CDK2-cyclin E (CycE) double-negative feedback confers bistability and hysteresis depending on prior growth factor stimulation (green). (b) MYC feed-forward regulation of cyclin D (CycD) and E2F enables the circuit to be reset when growth factors are withdrawn (magenta). (c) MYC amplification renders the restriction point independent of growth factors (brown).

Another system property is a resettable restriction point—normal cells should return to the E2F-inactive state when growth factor is withdrawn. Yao and coworkers (41) showed that resettability was conferred by the transcription factor MYC, which is stabilized by growth factors and induces cyclin D. MYC is also required for full E2F activity, creating a feed-forward architecture that enables resetting of the hysteretic Rb-E2F switch when growth factor is removed and MYC levels decline (Figure 2b). Surpassing the restriction point depends on the amplitude of E2F, which is modulated by MYC and timed by the abundance of D- and E-type cyclins (42). Cancers with MYC amplification lose resettability (Figure 2c), and the network traps cells in a chronic state of high E2F activity (43).

Steps following the restriction point and on exit from the cell cycle are covered in the supplementary text.

2.2. Microenvironmental signaling networks (t_char ~ seconds–minutes)

Cancer cells sense and missense their local context according to varied external stimuli (Section 2.1.1), and several input-output networks are archetypes for classic systems biology models relevant to cancer cells. Others are murkier because of their diffuse inputs or legacy connections to non-transformed cells of the tumor stroma. Appreciating this spectrum of inputs is crucial to developing cancer models that generalize beyond a few controlled settings. We compare strengths and weaknesses in model–experiment iteration among subsections when possible.

2.2.3. Death receptors.

Evading programmed cell death is a founding cancer hallmark (3). Cells receive extrinsic death signals from the immune system through tumor necrosis factor (TNF) superfamily ligands, including FasL and TRAIL. Ligand-receptor complexes recruit multiple adaptors, triggering proteolytic activation of caspase-8, which cleaves and activates caspase-3 to execute apoptosis. Caspase-8 also initiates a feed-forward loop involving Bcl2-family proteins, which form competitive complexes that promote or inhibit pore formation on mitochondrial outer membranes. Mitochondrial permeabilization releases factors that cooperate with caspase-3 to drive apoptosis. Active caspase-3 feeds back positively on upstream cytoplasmic and mitochondrial processes. Different cells are poised or primed to resist cell death through inhibitor-of-apoptosis proteins (IAPs), which antagonize caspase activation. TNF-family ligands also activate non-caspase pathways, such as stress MAPKs (Section 2.1.2), whose dynamic range is important for modulating the core apoptotic circuity (44). In cells where apoptosis is blocked, TNF elicits a secondary necroptotic pathway. Necroptosis involves a series of ubiquitylation events culminating in MLKL phosphorylation and plasma membrane permeabilization. Both pathways are relevant to cancer, as anti-apoptotic Bcl2-family proteins are recurrently amplified in breast and lung adenocarcinoma.

Models of death receptor-induced necroptotic signaling are actively developing with the essential regulatory mechanisms clarified in the last 15 years. Ildefonso and coworkers recently reported a physicochemical model of TNF receptor complexes and the ubiquitylation cascades leading to necroptosis (45). Bayesian parameter estimation of bulk phospho-MLKL dynamics suggested four possible modes of necroptosis that were distinguished by the predicted roles of two deubiquitylating proteins. In the model, either of the two proteins could singularly drive necroptosis, or one could antagonize necroptosis driven by the other. The distinction is important, because one protein is constitutively expressed, whereas the other is a target gene of NF-κB (Section 2.3.3) downstream of TNF, creating a feed-forward loop. Differences in basal expression and inducibility may confer cell-type specific differences in necroptosis analogous to those clarified for apoptosis (46). Like caspase models, systems models of necroptotic signaling will likely undergo considerable refinements as better live-cell reporters come available (47).

2.2.5. Oxidative stress.

Cancer progression usually requires adaptation to chronic oxidative stress. Multiple proliferation-driving oncogenes (Sections 2.1.2, 2.1.3, 2.1.5) elevate intracellular reactive oxygen species (ROS) through changes in metabolism (Section 2.4). ROS are also generated by activated fibroblasts and immune cells of the tumor stroma. Oxidative stress is also faced by normal tissues, including the heart and liver. Accordingly, cells have elaborate mechanisms to detoxify ROS with thiol-reactive scavenging pathways, which are coupled by the shared coenzyme, NADPH. Unscavenged ROS react with free thiols on cytoplasmic proteins, including ROS sensors that cause downstream transcriptional induction of compensatory pathways (Section 2.3.1). ROS sensor-inducer pathways are constitutively activated by mutation in many cancers (48).

ROS systems biology is frustratingly complex—the pathways have been known for ages but the intermediates are short-lived, difficult to quantify, and contextual. Adimora and coworkers first attempted to integrate multiple detoxification pathways in a ROS model of a T-cell lymphoma line (49). Uncertain of protein thiol susceptibility to undergo oxidation, the authors proposed two competing alternatives: a “slow” model constrained by literature estimates of bulk reactivity and a “fast” model that allowed some proteins to react 20–1000 times faster. When the models were calibrated with time-course data after bolus addition of H₂O₂ and tested against separate dose-response experiments, the fast model was preferred. Although others have questioned the validity of the well-mixed assumption for highly reactive H₂O₂ (50), the work illustrated how lumped ROS processes could be abstracted and distinguished computationally. Work is ongoing to extend these principles to the single-cell level (51).

In lieu of explicit ROS modeling, others have leveraged metabolic network reconstructions (Section 2.4) and focused on the generative fluxes for cellular NADPH (52). Investigating the response of melanoma cells to mutant-BRAF inhibitors, Paudel and coworkers (53) used transcriptomics to guide flux balance analyses of melanoma lines. The models inferred greater NADPH flux through glutathione reductase activity in the inhibitor-resistant lines, as confirmed by experiment. Follow-on small-molecule perturbations of glutathione- or NADPH-dependent pathways enhanced the bioactivity of mutant-BRAF inhibitors in otherwise resistant lines. The path to oncogene-induced oxidative-stress tolerance thus impacts the response to withdrawal when the oncogene is inhibited.

2.2.6. DNA damage.

In response to DNA damage, cancer cells stabilize the p53 tumor suppressor, a transcription factor inducing target genes that pause proliferation to allow for repair or promote apoptosis if damage is too extensive. These processes are non-uniform among cells, leading to heterogeneous outcomes in tumors and outgrowth of resistant cells often harboring p53 pathway disruptions. Normally, p53 is turned over by the MDM2 ubiquitin ligase, but MDM2 is itself a p53 target gene, creating a time-delayed negative feedback that can oscillate (Figure 3a). Time-lapse imaging has revealed pulses of p53 stabilization, target gene induction, and upstream kinase activation, which hint that cell outcomes are encoded in the strength and periodicity of the pulses.

Figure 3 |.

Models of DNA damage-induced p53 signaling. (a) The core p53–Mdm2 feedback of (66) was expanded to include p53 induction of Wip1 phosphatase in (57), which acts on p53 and ATM (white) activated by ionizing radiation but not ATR activated by ultraviolet radiation (59). Other p53 target genes include PUMA and PIG3, which accumulate differently based on mRNA and protein stability as modeled in (63). (b,c) p53, Mdm2, and Wip1 dynamics modeled by delay differential equations in (59).

The earliest work used mass-action models and fluorescent reporters to examine the feedback processes coordinated between p53 and MDM2. Cells with steady oscillatory feedback pass oscillations on to sister cells, but they diverge after ~11 hours or two oscillatory periods (54). A deterministic model perturbed with noise at different frequencies identified slow variations in protein production rates as a probable source of the observed variance in pulse amplitudes. Separately, a Fourier analysis of p53 reporter cells concluded that uncorrelated white noise sustains p53/MDM2 oscillations once activated (55). Such noise is plausibly conferred by the transcriptional bursting of p53-driven promoters that have also been profiled, modeled, and classified (56).

Initially, p53 pulses were assumed to arise as an inherent feature of p53/MDM2 feedback. Revised models with delay differential equations suggested that p53 pulses are triggered by pulsing of the upstream kinase ATM, which repeatedly surveils damaged DNA until repair is complete (57). A minimalistic model featuring nonlinear interaction terms between p53 and MDM2 constrained by observations of a simple synthetic system concluded that a single p53 feedback enabled tight timing control, but additional system components were needed to capture the dependence of p53 amplitude on input strength (58) (Figure 3a). Interestingly, the origin of DNA damage involves different feedback that controls the dynamics, with ultraviolet radiation leading to a single pulse with amplitude and duration scaling with input and ionizing radiation leading to a series of fixed pulses (59) (Figure 3a–c). p53 pulses are conserved but variable across species, with p53/MDM2 feedback particularly strong in rodents and showing more rapid oscillations than humans, monkeys, or dogs (60).

p53 dynamics ultimately control cell fates, and mathematical models have extended our understanding of how system inputs achieve desired phenotypic outcomes. For example, a model of p53 pulsing in response to double-strand breaks predicted how a series of timed drug inputs creates a sustained p53 activation that leads to senescence instead of DNA repair (61). Extending to p53 target genes (Figure 3a), short p53 pulses transmitted more faithfully to mRNAs with shorter half-lives (62). Likewise, p53 targets with slower protein decay rates integrate the pulses instead of transmitting them (63). Due to these secondary and tertiary effects, entirely different proteins are induced when p53 oscillates compared to when it rises monotonically (64). Although much of the work in this area has focused on the highly synchronized DNA damage resulting from radiation, pulses of p53 are also relevant after chemotherapy. A study combining imaging with modeling identified chemotherapy-driven p53 pulses that encode cell fates in drug-sensitive and -resistant cancer cells (65).

An important p53 target gene is the cyclin-dependent kinase inhibitor p21, and models have clarified p53-induced p21-dependent mechanisms controlling cell cycle arrest after DNA damage (67). For example, modeling revealed a feedback loop between p21 and reactive oxygen species as responsible for establishing a senescent phenotype days after cell exposure to ionizing radiation (68). For ionizing radiation responses, other models revealed that p21 contributes to heterogeneous cell death response according to variable rates of p21 degradation in S-phase (69).

2.3. Oncogenic and non-oncogenic transcriptional networks (t_char ~ minutes–hours)

Transient signals of intracellular networks often culminate in gene-expression changes through transcription factors. Sometimes, it is impossible to uncouple signaling and transcription. For example, models of DNA damage networks rely intimately on transcriptional feedback from p53 (Section 2.2.6), and cell-cycle models cannot exist without considering genes induced by E2F or MYC (Section 2.1.5). This section focuses on other transcriptional circuits that act as convergence points for multiple signaling inputs and are thus worth examining in their own right. The level of interconnectedness dictates whether mechanistic or statistical models are better for iterating with experiments (Figure 1a–c).

2.3.1. Basic leucine zipper (bZIP) transcription factors.

bZIP transcription factors are crucial effectors of fast-acting processes in cancer. In response to oxidative stress (Section 2.2.5), the bZIP family member NRF2 is stabilized, and its nuclear accumulation induces antioxidant response genes. ROS also elicit a DNA damage response (Section 2.2.6), and a subset of p53 target genes converge with those of NRF2. Building on prior work (57), Pereira and coworkers (70) synthesized a mechanistic model combining NRF2 and p53 pathways including transcriptional feedback downstream of ROS. The model captured dynamic coupling and synergy of the NRF2-p53 pathways in coordinating breast epithelial cell response to oxidative stress. The authors used the integrated model to predict NRF2–p53 coordination and sensitivity to NRF2 perturbation in p53-mutant triple-negative breast cancer lines, which were verified in 3D culture. By contrast, applying the same approach to primary triple-negative breast cancers yielded chaotic predictions, suggesting that the ROS adaptations in cell lines are different from those seen in vivo. NRF2 is a permissive non-oncogene in breast cancer, but NRF2 and its upstream antagonist are recurrently mutated in head-and-neck and lung cancers.

An important bZIP subfamily is the AP-1 group of transcription factors, many of which are induced and activated by phosphorylation downstream of Ras–ERK (Section 2.1.2). The stability, heterodimerization, and transactivation characteristics of AP-1 proteins is daunting to encode in models. Instead, Comandante-Lou and coworkers (71) adopted a data-driven approach to model AP-1 regulation in BRAF-mutant melanoma and its relationship to melanocytic differentiation. The group collected iterative immunofluorescence on 11 AP-1 family members plus six AP-1 phosphorylation states and four markers of differentiation that were interrelated through partial least squares regression modeling. The models implicated FOSL2 and JUN proteins in the AP-1 family as negative predictors of the melanocytic state and positive predictors of the undifferentiated state. Knockdown of FOSL2–JUN caused synergistic or additive increases in melanocytic markers, consistent with model predictions. Holistic approaches that capture primary and secondary consequences of targeted perturbations will be required to deconstruct bZIP transcription factor networks in the future.

2.3.3. Master regulators.

A prevailing modeling challenge is the cell context-dependent impact of transcription factors. As an organizing principle for cancer networks, Califano and Alvarez proposed “master regulators” to define factors whose abundance quantitatively tracks with specific transcriptional programs across contexts (72). Using mutual information, Alvarez and coworkers (8) built an inference approach for estimating thousands of master-regulator activities from transcriptomic data. The initial network model was refined as statistical dependencies were falsified or reinforced with mechanistic tests (Figure 1c). After drafting many network contexts, the authors showed that new samples were adequately modeled by identifying the best surrogate network among those already assembled (73). The concept of master-regulator activity now extends to their inversion with FDA-approved drugs in preclinical studies (74) and early-stage trials (75).

Master regulators may be predefined, but they can also be discovered. Chitforoushzadeh and coworkers (76) sought out master-regulator themes that emerged upon combined ligand stimulation of RTKs (Section 2.1.1) and death receptors (Section 2.2.3). Starting with time-resolved observations of signaling and gene expression in colon adenocarcinoma cells, the authors built a latent-variable model to connect the two tensors and suggest new mechanisms. Latent-variable projections uncovered a gene cluster containing an oncogenic RTK (MET) and Bcl2-family member (MCL1) induced by death receptors and inhibited by Akt signaling (Section 2.1.3). Promoter analysis and extensive follow-on experiments implicated the transcription factor GATA6 and its posttranslational regulation by GSK3 downstream of Akt. GATA6 modulates the tumorigenic potential of many endodermal tissues and is key to maintaining the stem-progenitor state of some colorectal cancers (77). The inferred master-regulator activity of GATA6 also partly stratifies breast cancer subtypes (72).

2.4. Cancer metabolic networks (t_char ~ minutes–hours)

Malignant transformation requires multiple permissive adaptations in tumor-cell metabolism that are highly interdependent. Generic models of human metabolism have been customized to cancer repeatedly, but most have been limited to one-way interpretations of transcriptomic or metabolomic data (Figure 1d). We concentrate on the handful of metabolism studies where model-experiment iterations are emblematic of a systems biology approach.

2.4.2. Metabolic gains and losses in cancer.

Tumor initiation and progression involve (epi)genetic alterations that rewire cancer metabolism in non-obvious ways. For instance, we have repeatedly mentioned the importance of NADPH in cancer cell ROS tolerance (Sections 2.2.5, 2.4.1), but how is this NADPH pool generated and maintained? Biochemistry textbooks point to the pentose phosphate pathway, but Fan and coworkers (78) could only ascribe about half the NADPH flux to this pathway. To identify alternative sources, the authors used a genome-scale metabolic model. When total enzyme concentration was minimized, the model predicted that a major contributor to NADPH production was one-carbon metabolism of serine with methylenetetrahydrofolate as a precursor. The prediction was supported by ²H- and ¹³C-tracer studies along with genetic perturbations. The results explain the widespread transcriptional upregulation of the methylenetetrahydrofolate dehydrogenase MTHFD in most cancers (79).

Another one-carbon puzzle resolved with modeling involves the de novo serine biosynthesis enzyme, PHGDH, which is amplified in breast cancer and melanoma. Reid and coworkers (80) built a localized metabolic model of carbon fluxes into nucleotide metabolism and parameterized it with ¹³C-tracer experiments. Model-guided experiments revealed that PHGDH supports nucleotide synthesis not through the production of serine as a nucleotide precursor but rather by supporting the TCA cycle and pentose phosphate pathways. Supplementing cells with α-ketoglutarate (a TCA intermediate) and ribose (an output of the pentose phosphate pathway) restores growth of cancer cells treated with a PHGDH inhibitor.

2.5. Cancer state switches (t_char ~ hours–years)

Regulatory states of cancer cells are more plastic than their normal differentiated counterparts. Our appreciation of cell-state heterogeneity has grown (81), leading to various theories and modeling formalisms that abstract transitions in settings as small as a gene locus and as large as a cell population. We highlight work that has provided simplifying understanding to an otherwise endlessly complex facet of cancer systems biology.

2.5.4. Epithelial-mesenchymal transition (EMT).

EMT is a normal process in cell development that occurs aberrantly in carcinomas. Although a strict requirement for EMT in metastasis is debated, EMT clearly occurs early in cancer dissemination and promotes resistance to therapy. The most widely studied EMT agonist is TGFβ (Section 2.2.2), but many TGFβ-independent EMT paths exist. Systems models have addressed questions surrounding the number of EMT states and mechanisms that lead to switching. EMT is more than epithelial and mesenchymal states based on the transcriptional and posttranslational circuits that control it. Investigators have identified a hybrid or partial EMT state exhibiting both epithelial and mesenchymal characteristics that plays a unique role in collective migration of cancer cells. This third state emerges from reciprocally suppressive microRNA-transcription factor circuits. Lu and coworkers considered the miR-34/Snail and miR-200/ZEB circuits, with three crosstalk mechanisms and two auto-regulatory loops (82). On its own, the miR-200/ZEB circuit exhibits tri-stability. Integration with the mono-stable miR-34/Snail circuit leads to a tri-stable system that filters noise and exhibits asymmetry in that the hybrid/partial EMT phenotype arises only during the forward EMT process but not the reverse. A contrasting model (83) considered the same circuits but configured them sequentially as a reversible pathway to a partial EMT state followed by an operationally irreversible step leading to complete EMT. Testing in breast epithelial cells confirmed three stable states as defined by the epithelial marker E-cadherin and mesenchymal marker vimentin (84) and supported the two-step model. A similar switching model identified the transcription factor NRF2 as a specific node that limits complete EMT and stabilizes partial EMT by simultaneously promoting E-cadherin and ZEB1 expression (85).

Outside the core EMT circuitry lie other inputs that promote phenotypic switching. In one model of Wnt- and EGF-mediated signaling, the role of RKIP featured prominently in multiple feedback loops leading to switch-like regulation of E-cadherin (86). Elsewhere, larger network models have nominated signaling pathways relevant during EMT. To reduce complexity, Boolean frameworks are often used to capture coarse-grained logic features. Steinway and coworkers built a 70-node Boolean model, simplified to experimental observables, and used the model to identify the simultaneous induction of Wnt and Sonic hedgehog signaling in response to TGFβ, which was validated in hepatocellular carcinoma cell lines (87). An extension of this framework explored the systematic effects of single and combinatorial pathway interventions in TGFβ-mediated EMT (88). The authors identified numerous perturbations leading to partial EMT and nominated Notch and Ras as nodes that antagonize SMADs to promote a stable epithelial state.

To discover new EMT nodes and edges, Brown and colleagues used regularized linear modeling with pathway and gene set overdispersion scores from scRNA-seq data to nominate signaling pathways that explain hypoxia-mediated EMT in human pancreas cancer tumors (89). Model testing confirmed cooperative roles for ERK and JNK in a hypoxia-initiated feedback between histone methylation and MAPK activation leading to an especially long-lasting EMT. Follow-on work suggested AXL-mediated disinhibition of YAP signaling through the cell adhesion protein FAT1 as one potential hypoxia-specific EMT mechanism (90).

2.5.5. Drug Persistence.

Treatment of cancer cells causes a reversible “drug-tolerant persister” state to emerge in the remaining population. During continuous therapy, persister cells die slowly amidst slow, error-prone proliferation that ultimately gives rise to heritably resistant subclones. Taking the persister state switch as a stochastic process, Cassidy and coworkers (91) built an age-structured model where daughter cells randomly interconvert at different prescribed probabilities. Using the model under different probabilistic transitions, the authors found that for virtual tumor populations prone to switch with increasing age since last division, there was predicted benefit for intermittent therapy withdrawal. During these drug holidays, accumulated persister cells convert back to the sensitive state and are thus responsive to subsequent therapy rechallenge. The group extended the model to consider informed dosing schedules that disfavored persister accumulation and reduced overall tumor burden. The study extends the concept of adaptive therapy, which was conceived to thwart heritably resistant subclones (92).

The discovery of persisters triggered a search for their origins—are persisters induced by the therapy or do they pre-exist? Unfortunately, the answer appears to depend on the cancer type. In colorectal cancers treated with chemotherapy and allowed to regrow, there are no detectable biases in tumor complexity, and simple probabilistic models of clone-size distribution suggest that all cells can become persisters (93). Russo and colleagues (94) extended the classic Luria-Delbrück fluctuation analysis to investigate the question more thoroughly. For colorectal cancer cells treated with targeted agents, they included an elevated mutation rate for persisters compared to the spontaneous mutational rate normally estimated by the analysis. When considering the late emergence of resistant clones, the authors found that a zero to near-zero fraction of preexisting cells was favored alongside a 7–50-fold increased mutation rate in persisters. Cotreatment of cells with inhibitors of error-prone polymerases associated with the persister state drastically inhibited or eliminated resistance.

Contrasting the above results are studies in melanoma, where resistance to mutant-BRAF inhibitors pre-exists by conventional Luria-Delbrück fluctuation analyses (95). Resistance is tied to the rare, coordinated expression of many genes, some of which confer resistance when selective pressure is applied with inhibitor. Schuh and coworkers (96) asked whether such coordinated states could arise spontaneously without any inhibitor treatment under a limited set of network assumptions. The authors incorporated transcriptional bursting into a leaky telegraph model of expression that leads to simple, additive gene-regulatory effects on other genes in networks of up to 10 nodes. They found that coordinated states arose rarely but consistently in loosely connected networks, consistent with the loose functional categories of genes that appear to be coordinately high in rare cells. Therefore, melanoma persisters result from a natural expression state that is fortuitous only when mutant-BRAF inhibitors are applied.

3. APPLICATIONS OF CANCER SYSTEMS BIOLOGY

Systems models of cancer biology are useful for both discovery and deployment. However, persistent hurdles abound: i) global cancer-cell fates are normally too poorly specified to frame as biochemical reaction networks; ii) drug (off-)targets are incompletely enumerated; iii) no two cancer cases are identical. Applied cancer systems biology requires maneuvering through a series of tradeoffs and compromises to achieve usability.

Section 3 organizes notable examples where pragmatism took precedence over perfection to good ends. The path first needed to be blazed for cancer phenotypes that were biochemically multivariate or ill-defined. Second came the systems biology of primary and secondary responses to chemotherapeutics, targeted agents, and combinations therein. There remains active interest in applying systems models to the precise biochemistry of individual cancers by extending approaches from Section 2.

3.1. Phenotypic models of cancer hallmarks

None of the latest hallmarks of cancer (97) can be distilled to a single biochemical readout. Grappling cancer cell phenotypes with systems models usually requires a data-driven statistical interface to link with upstream networks (98). Topics are ordered by the molecular granularity that is currently feasible.

3.1.1. Proliferation, cell-cycle arrest, and senescence.

Tumor initiation, recurrence, and metastasis are oligo-clonal events where default models of exponential growth do not necessarily apply. Beyond stochastic competition of cell division and death, cooperative processes may accelerate growth rates once a subpopulation has reached a critical size. Johnson and coworkers (99) posed different stochastic models of birth-death probabilities with various dependencies on population size. Each model was identifiable, and their complexities could be ranked by standard information criteria. After tracking oligo-clonal growth of a luminal breast cancer cell line seeded at densities of 2–10 cells, the authors compared the quality of model fits and found a detectable reduction in birth rate with a two-cell inoculum. This work quantitatively supports the long-standing observation that clumps of 6+ cells initiate pulmonary metastases more efficiently than singletons (100).

Sustained proliferation is interrupted during chemotherapy when cancer cells are treated with double-strand break-inducing topoisomerase inhibitors such as doxorubicin. Tentner and coworkers (101) investigated signaling and phenotypic interactions between doxorubicin and TNF (Section 2.2.3). In osteosarcoma cells with unmutated p53, there was little-to-no modulation of doxorubicin-induced arrest by TNF, although the two synergistically induced apoptosis. Cell death and cycle arrest were statistically separable in a partial least squares regression model that used biochemical measurements of DNA-damage, ERK, and NF-κB pathways. Prediction of a role for ERK in early doxorubicin-induced G1/S arrest and apoptosis was tested–confirmed pharmacologically. Netterfield and coworkers (102) built upon this work, showing that the cell-cycle exit led to a senescent cell fate at the lowest doses of doxorubicin. Tracking a similar signaling panel (101) but with subcellular resolution afforded by immunofluorescence, the authors modeled tensors of signaling and senescence–proliferation–apoptosis as latent variables of time and doxorubicin dose (102). The model identified phosphorylation of JUN (Section 2.3.1) as a key early trigger of senescence, and pharmacologic blockade of its kinase JNK prevented cell-cycle exit even more potently than inhibiting ERK. Late-phase ERK–JNK signaling reinforced the cell-state transition by coordinating AP-1 maintenance of the IL-6 and IL-8 secretory phenotype characteristic of terminally senescent cells. Signaling to cell cycle-coupled phenotypes is difficult to envision as a physicochemical model, but it is well posed for data-driven abstractions (10; 98).

3.2. Systems pharmacology of anti-cancer agents

The response of transformed cells to therapy has been a pillar of cancer systems biology for some time. Specific drugs cause remarkably widespread secondary changes that intersect with other interventions in surprising ways. For brevity, we omit multi-omic predictions of drug sensitivity (103) or repositioning lacking a front-end model that refines iteratively with experiments (Figure 1).

3.2.1. Drug mechanisms of action.

Systems models encoding cancer cell responses to drugs have discovered unanticipated mechanisms of action. Interesting examples come from the literature on kinase inhibitors, where different drugs target or stabilize the active or inactive conformations of a catalytic domain. A mechanistic model by Rukhlenko and colleagues encoded known differences in thermodynamic, kinetic, and structural characteristics of RAF inhibitors targeting the two enzyme conformations and concluded that, for dimerizing proteins, inhibitor combinations targeting both active and inactive kinase forms should work synergistically (104). To connect network-level descriptions of cancer cells to drug phenotypic responses and their mechanisms of action, statistical models are useful. Partial least-squares regression successfully predicted the cell background and stimulus response characteristics needed to capture inhibitor responses for different kinase targets (105). By comparison, standard cancer subtype classifications were poor predictors of cell response. The same model class has been redeployed for identifying network effects responsible for surprising phenotypic responses of cancer cells to investigational therapeutics. For example, Day and colleagues built a model that identified PTEN nuclear localization as an unexpected effect of SHP2 inhibitors that antagonizes glioblastoma cell death in response to temozolomide (106). Using a focused panel of regressands in such models enables known pathways to be prioritized alongside mechanisms that are new or hypothetical.

Simple models have revealed confounding artifacts in seemingly straightforward metrics of efficacy. For example, when comparing drug efficacy across cell types, differences in baseline cell proliferation obscure true differences in drug efficacy resulting from different network configurations. These effects were quantified by Hafner and colleagues to derive an alternative growth-normalized metric for drug efficacy (107). The concept extends to the dynamics of drug effects relative to cell proliferation rates (108). Recognizing that cell responses to targeted therapeutics reflects a mixture of proliferation, quiescence, and death, Tyson and colleagues developed an approach to integrate time-lapse imaging of intra-population response heterogeneity with a cell growth model and demonstrated that, even in nominally oncogene-addicted cell settings, driver gene inhibition does not cause apoptosis exclusively (109), creating alternative paths for escape.

3.2.4. Dynamic targeting.

Systems biologists have considered time as a manipulatable variable for drugs. Behar and coworkers (110) built a representative set of theoretical modules based on known signaling circuits and asked whether early- vs. late-phase readouts could be targeted selectively. Often, they successfully identified rate processes that when inhibited were predicted to disrupt some readouts but not others. The authors applied the principle to a detailed model of NF-κB signaling (111) and found translational, proteasomal, and epigenetic rates that impacted target genes differentially and were later experimentally validated (110). Another feedback-laden cancer signaling module with rich dynamics is p53 (Section 2.2.6; Figure 3), which may be amenable to a similar analysis.

Drugs affect the timing of responses, as does the timing of the drug itself. Lee and coworkers (112) screened for time-staggered inhibitors that potentiated the response to DNA-damaging agents and found that pretreatment with EGFR inhibitors was a potent sensitizer. The group built a partial least squares regression model based on sparse measurements of growth factor, DNA damage, and apoptotic networks. The model highlighted that EGFR inhibition increased caspase-8 activation following different DNA-damaging agents, which was verified by knockdown experiments. The result extended to lung cancer cell lines and formed the basis of a timed-release nanoparticle formulation that was effective preclinically (113). Others have gone on to screen for time-staggered synergies systematically (114).

3.3. Individualized cancer cell models

Each cancer rewires and imbalances normal intracellular circuitry in its own way. From the first comparative models of normal and transformed cells (115) to the latest protein-protein interaction maps of cancer (6), the field has taken various stabs at individualized systems biology. Case-specific modeling is much more than a search for cancer biomarkers, which is notoriously plagued by hidden confounders. These models seek to distill the important differences that make a cancer behave uniquely.

3.3.1. Individualized responses to anti-cancer therapy.

Different cancers of the same type have markedly different responses to targeted agents, even though the underlying molecular networks are largely conserved. Such differences arise because of i) mutations that break or, more rarely, create network edges and ii) abundance changes in network nodes that alter the overall signal transmission. Network-focused cancer systems biology models are thus highly conducive to understanding how variable druggability arises. Tejeda Zañudo and coworkers (116) revised a dynamic network model of hormone and RTK signaling to distinguish resistance mechanisms in PI3K-mutant luminal breast cancer. The update required unlumping of CDK inhibitors (p21 and p27) and Bcl2-family proteins (BCL2 and BCL-XL). By customizing a Boolean network model of generic cancer hallmarks for prostate cancer, Montagud and coworkers (117) simulated 488 cases based on mutational and transcriptomic profiles from The Cancer Genome Atlas. Similar repository data have been used to initialize physicochemical models of oxidative stress in a cell-line and patient-specific way (70). Kirouac and coworkers (118) extended this idea to simulate clinical trials for colorectal cancer treated with different combinations of ERK-pathway inhibitors. Individualized systems modeling will grow as quantitative estimates of protein across cancers become more widespread.

Pathway models also explain measured differences in drug responsiveness. For 62 breast cancer lines, Tognetti and coworkers (119) adopted an experiment-first approach, using mass cytometry to examine inhibitor profiles after growth-factor stimulation. The authors regressed these data against a prior-knowledge model to infer speed and edge-transmission parameters specific to each cellular context. Integrated regression parameters were uniquely informative for drugs targeting the PI3K–Akt–mTORC pathway (Section 2.1.3). Eduati and coworkers (120) reached similar conclusions about the contextual importance of PI3K–Akt in pancreas cancer biopsies that were perturbation-screened and modeled by logic-based ordinary differential equations. In a twist on deep learning, Kuenzi and coworkers (15) constrained a visible neural network with the hierarchical organization of protein complexes and cell biological processes. Upon learning the response to 684 drugs based on the mutant genotypes of 1235 cancer cell lines, the model revealed that sensitivity to paclitaxel (a microtubule stabilizer) was dependent on the state of glycolysis and the ubiquitin-proteasome system, both of which were confirmed by experiment.

4. EMERGING OPPORTUNITIES

We see numerous opportunities for existing systems tools to tackle cancer-related challenges that are coming to the fore. Emerging classes of precision drugs are one area given allosteric mechanisms of action that are less direct than earlier targeted therapies. Inhibitors of KRAS mutants (121) and the protein tyrosine phosphatase SHP2 (122) are good examples. Proteolysis-targeting chimeric (PROTAC) drugs are another major class of drugs that are ripe for systems analyses because of the network-level effects anticipated. Systems tools should help identify context-specific paths of resistance and suggest routes to overcome it. Data-driven approaches that consider multiple drugs in a single dataset are ready to begin incorporating quantitative descriptors of drug molecular features as independent variables akin to quantitative structure-activity relationships. We anticipate that molecular features of drugs themselves hold quantitative clues about pathways to cancer cell sensitivity that, when decoded, will aid in drug selection and design.

The above distinction between mechanistic and data-driven modeling highlights another emerging opportunity to integrate the two approaches specifically at the interface of predicting phenotypic outcomes for quantitative perturbations to network inputs. For receptor-mediated signaling, such an interface lies between the receptor-proximal steps that activate a set of effector pathways that control complex cell fate decisions through mechanisms not amenable to mechanistic abstraction (98). An early example of this approach explains how EGFR ligands with vastly divergent receptor affinities produce different cell outcomes through regulation of receptor trafficking and effector pathway dynamics (123). We envision extensions that leverage the dynamic component of integrated models to predict optimal trajectories (e.g., minimal dosing) for therapeutic combinations or to achieve specific therapeutic outcomes. Ultimately, such considerations will need to integrate with tumor evolution models (124) to navigate the adaptive resistance that invariably arises.

We also see untapped opportunities to tackle cancer cell systems and the role of functional intracellular compartments in determining cell phenotype. Compartmentalization has been considered in many systems models ranging from nucleo-cytoplasmic shuttling mechanisms to receptor endosomal dynamics. As the reach of systems models expands to integrate kinase-mediated signaling with metabolic networks (Section 5), we see new opportunities to quantify compartment-specific shuttling, partitioning, and interaction mechanisms of fate-determining proteins. Cell organelles are a general class of compartments where we anticipate cell biology and systems biology will continue to meet and advance in tandem.

5. FUTURE CHALLENGES

If every gene, protein, modification, and metabolite were measurable with cellular resolution, could we reliably distinguish cancers from pre-cancers and non-cancers? Could we predict how such deep cellular profiles would change in response to chemotherapy, radiation, or targeted agents? Answers to these questions are still far away, but we see challenges on the horizon that future cancer systems biologists might tackle. Notably, the compartmentalized organization of this review (Table 1) reminds that more thought is needed at intersections within cancer systems biology and with other fields. Models of cancer biophysics can be nicely formulated (125), for example, but links to cancer cell signaling have thus far been statistical (126) or phenomenological (127). Likewise, although signaling influences metabolism and vice versa, and models of the two are well vetted in cancer (Sections 2.1 and 2.4), there are conceptual and technical barriers for integration to overcome. Intentional cross-training and sustained thinking are needed to devise new frameworks for modeling intracellular transduction when the inputs and outputs are very different.

Cancer systems biology has always had a strong technologic bent, and we continually seek new tools to interrogate cellular processes in disease settings. Time-resolved omics remains problematic, but trading off depth for dynamics is a compromise that one day might be avoidable. Engineered signaling reporters now use space (128) and time (129) to increase plexity of readouts in living cells, and it is worth pushing the limits of such designs. Having a few dozen continuous and concurrent indicators of cancer cell state would be greatly empowering for systems biology; at the population level, far fewer than that are needed to predict cell outcomes (130).

Finally, we must deliberate on how cancer models and repositories of the past can be fully leveraged for cancer systems biology of the future (Figure 1). What does iteration look like when so much data and so many tools are publicly available? Our opinion is that the best central conduit for information exchange is a reliable and adaptable systems model. Public datasets aside from sequencing are rarely structured for alternate uses—models can help to fill the voids and guide choices about which new experiments are really worth doing. Reciprocally, models should aspire to bridge major chasms in cancer biology that would be impossible with data alone: in vitro vs. in vivo, mouse vs. human, etc. Aggressive stress testing will cause models to fail, but that is the point. It is by codifying then falsifying assumptions and prior knowledge that cancer systems biologists learn where to head next. Likewise, data collection becomes more efficient when it is done with a model in mind.

TERMS AND DEFINITIONS (other acronyms are universal gene names)

RTK

receptor tyrosine kinase, a type of transmembrane receptor with roles in proliferation, growth, and differentiation

Ras

small GTPase corresponding to KRAS, NRAS, and HRAS

MEK

mitogen-activated protein kinase kinases corresponding to MAP2K1 and MAP2K2

ERK

mitogen-activated protein kinases corresponding to MAPK1 and MAPK2

RKIP

RAF kinase inhibitor protein corresponding to PEBP1

PI3K

phosphoinositide 3-OH kinases corresponding to PIK3CA, PIK3CB, PIK3CD, and PIK3CG

mTORC

mechanistic target-of-rapamycin complex containing MTOR and regulatory subunits

p21

cyclin-dependent kinase inhibitor corresponding to CDKN1A

Cdt2

ubiquitin ligase corresponding to DTL

ECM

extracellular matrix

TRAIL

TNF-related apoptosis-inducing ligand corresponding to TNFSF10

ROS

reactive oxygen species including O₂•^–, HO•, and H₂O₂

p53

tumor suppressor gene corresponding to TP53

bZIP

large family of transcription factors defined by the basic leucine zipper structural element

NRF2

antioxidant transcription factor corresponding to NFE2L2

AP-1

activator protein-1 transcription factor comprised of FOS, JUN, and ATF dimers

IkB

inhibitor of kB proteins corresponding to NFKBIA, NFKBIB, and NFKBIE

p65

NF-kB transcription factor corresponding to RELA

TCA

tricarboxylic acid, a metabolic cycle essential for central carbon metabolism

SCLC

small-cell lung cancer, a lung cancer strongly tied to smoking

EMT

epithelial-mesenchymal transition

JNK

c-jun N-terminal kinase, a mitogen-activated protein kinase corresponding to MAPK8, MAPK9, and MAPK10

SHP2

Src-homology 2 domain-containing phosphatase 2 corresponding to PTPN11