Modeling, an experimental approach to investigate complex biological systems, has significantly contributed to our understanding of cancer. Although extensive cancer research has been conducted utilizing animal models for elucidating mechanisms and developing therapeutics, the concepts in a good model design and its application have not been well elaborated. In this review, we discuss the theory underlying biological modeling and the process of producing a valuable and relevant animal model. Several renowned examples in the history of cancer research will be used to illustrate how modeling can be translatable to clinical applications. Finally, we will also discuss how the advances in cancer genomics and cancer modeling will influence each other going forward. Cancer Res; 76(20); 5921–5. ©2016 AACR.

Complex biological systems are composed of multiple interacting elements that are usually nonlinear and involve different levels of regulation, including positive and negative feedback. In addition, they exchange information with the surrounding environment, evolving in response to resources, and stresses. Such interacting networks are widespread in biology, ranging from genetics, metabolism, and stress signals to the population level. The resulting complexity associated with huge information flow poses great challenges to our understanding of biological systems from the reductionist viewpoint and produces the old paradox, the more we know, the less we understand.

There are two general approaches to study these complex biological systems: circuit analysis and modeling. Circuit analysis addresses the deficiency of quantitative analysis of biological studies (1). Biological systems are difficult to delineate because how the input determines the output is not often well elucidated. If the function of each biological effector could be defined quantitatively, then describing the signal processing of a biological “circuit” module and further analyzing the whole system would be possible. In this process, the accumulation of knowledge can fine-tune the analysis and continuously enhance understanding of the system.

Although the quantitative approach is required for the advancement of biological studies, life as an evolving open system can mount an overwhelming number of variables, making it very difficult to identify the boundary of the system and the critical determinants. For example, research in cancer therapeutics has been evolving over decades to involve studies in tumor growth kinetics, cancer genetics, energy metabolism, microenvironment, immune response, etc. At each stage, the scope of cancer research is changing, and the quantitative approach is necessary but insufficient to address the exponentially increasing complexity. An alternative approach is modeling, which aims to identify the essential elements and their interactions for facilitated problem solving. Its conceptual process is very similar to the production of a map (2), whereby in mapping an unknown region, the boundary must be defined so it can be aligned to the known neighboring area. The information must be obtained and represented, and through further research, the represented data will be fine-tuned with new details, making it more precise. This analogy exemplifies the five steps in modeling: (i) system conceptualization, whereby the thesis question to be answered needs to be identified so that the system can be defined; (ii) model building, where the hypotheses are developed to identify the essential elements, which will be used to construct the model system; (iii) model testing, during which the model is subjected to the condition necessitated by the thesis question to generate responses; (iv) outcome evaluation, where the gap between the generated and expected outcomes is assessed; and (v) model improvement, the process by which the model is modified on the basis of new information from testing and subsequently subjected to the next test run. After each experimental run, the model can be improved further, and eventually, the critical elements for controlling the system will be identified (Fig. 1). Importantly, as a dynamic process, modeling in general follows, but is not limited to, the five sequential steps aforementioned.

Figure 1

Cancer modeling as a continuous cycle. 1, identification of the question. A good model design defines the question to be answered by the model. Determining the desired resolution (microscopic to macroscopic) given the complexity of the biological system (molecular level to population level) of study will aid in outlining the scope of the model (range of question) and its boundary (field of study), which will determine the parameters of inputs and outputs of the model. 2, model building. To build a model, a physiologically or pathologically relevant system should be tailored to make the resolution of observation match the scope of the study. The driving factors of the system, which is consistent with the parameters in the identified question, should be able to be manipulated under experimentation. The system should produce a relevant and useful readout that properly addresses the endpoints of study. 3, model testing. Testing of the model involves adjusting the parameters of the driving factors as input and generation of outputs that allow evaluation for comparison of real system. The study design should emphasize the importance of a well-defined and controlled operation, ensuring statistical power for meaningful conclusions. 4, outcome evaluation. The endpoint of the study should be compatible with that of the real system, allowing comparison between the two systems at translatable bases. The translation of modeling results to clinical outcomes depends on the resolution of the model. The high-resolution modeling (e.g., signaling pathway) allows more straightforward application of the results to clinics, while low-resolution (e.g., pathophysiologic phenotypes) modeling results require consideration of genetic background and scaling. 5, model improvement. Continual assessment should appraise the system for improvement and modifications, and the endpoints should be critiqued and evaluated within the context of the real biological system. This can inform the next test run and guide design improvements. A well-designed model can bridge the gap to translational studies and inform their design, with similar feedback from clinical studies informing the next model improvement. This interplay can advance fundamental knowledge and clinical therapies.

Figure 1

Cancer modeling as a continuous cycle. 1, identification of the question. A good model design defines the question to be answered by the model. Determining the desired resolution (microscopic to macroscopic) given the complexity of the biological system (molecular level to population level) of study will aid in outlining the scope of the model (range of question) and its boundary (field of study), which will determine the parameters of inputs and outputs of the model. 2, model building. To build a model, a physiologically or pathologically relevant system should be tailored to make the resolution of observation match the scope of the study. The driving factors of the system, which is consistent with the parameters in the identified question, should be able to be manipulated under experimentation. The system should produce a relevant and useful readout that properly addresses the endpoints of study. 3, model testing. Testing of the model involves adjusting the parameters of the driving factors as input and generation of outputs that allow evaluation for comparison of real system. The study design should emphasize the importance of a well-defined and controlled operation, ensuring statistical power for meaningful conclusions. 4, outcome evaluation. The endpoint of the study should be compatible with that of the real system, allowing comparison between the two systems at translatable bases. The translation of modeling results to clinical outcomes depends on the resolution of the model. The high-resolution modeling (e.g., signaling pathway) allows more straightforward application of the results to clinics, while low-resolution (e.g., pathophysiologic phenotypes) modeling results require consideration of genetic background and scaling. 5, model improvement. Continual assessment should appraise the system for improvement and modifications, and the endpoints should be critiqued and evaluated within the context of the real biological system. This can inform the next test run and guide design improvements. A well-designed model can bridge the gap to translational studies and inform their design, with similar feedback from clinical studies informing the next model improvement. This interplay can advance fundamental knowledge and clinical therapies.

Close modal

The first step of cancer modeling is identification of the question or problem. For example, in the 18th century, Pott observed scrotal cancer to uniquely afflict English chimney sweepers (3, 4). Questioning the etiology, his scope of investigation involved identifying factors associated with chimney sweeping that could lead to scrotal carcinogenesis. His observation of soot accumulation on the lower scrotum, where the cancerous ulcers were observed, led him to conclude that a compound in the soot induced skin carcinogenesis.

Next, hypotheses are developed to build the model. This includes at least three parts: (i) selection of an experimental system physiologically and pathologically relevant to the scope of the question; (ii) identification of the driving or limiting factors that can be manipulated in the system; and (iii) determination of parameters that can be observed and/or measured in this system. Following Pott's classical investigation, Yamagiwa and Ichikawa in the early 20th century proposed that repetitive exposure to coal tar (a component of soot) induces carcinoma on mammalian skin. They proceeded to test his hypothesis by painting coal tar on rabbit ears, resulting in the first published animal model in the history of cancer research (5).

In model testing, the inputs have to include both control and treated groups, and reproducibility is crucial for the adequacy of the experimental system. The output should allow the evaluation of any differences from the modeled phenomenon with clinical data or outcomes. Yamagiwa and Ichikawa manipulated their model by changing the location of coal tar application on the rabbit ears as well as exposure duration, identifying the conditions allowing coal tar to induce carcinoma in rabbit models. They concluded that coal tar is a carcinogen and provided early support for the concept of chemical carcinogenesis.

For outcome evaluation different aspects can be analyzed as the output, such as cancer incidence, histopathology, growth kinetics, or therapeutic response, according to the hypothesis. However, the modeled subject always has to be clinically relevant; otherwise, it has no practical value. In the case of Yamagiwa and Ichikawa's study, they showed histopathologic evidence that the tumor in their model progressed in a similar fashion to human skin cancer. In the decades that followed, polycyclic aromatic hydrocarbons (PAH) were isolated from coal tar and tested in mouse models, confirming their carcinogenic properties. Further human studies confirmed the role of PAHs in human carcinomas as well (6), giving new direction of model improvement in chemical carcinogenesis. Although mice represent the model that best combines logistics and relevance, others available with differing advantages and disadvantages include in vivo yeast, worm, and fly models, as well as in vitro models, reviewed in refs. 7–10.

In the next section, we will discuss four representative examples of cancer modeling, each illustrating the process of implementing a good model system to research a specific question.

Our first example focuses on modeling chemotherapy. In the late 1950s, methotrexate was the standard treatment of choriocarcinoma but only induced temporary remission. Using human chorionic gonadotropin (hCG) as a tumor marker, Min Chiu Li reasoned that residual disease could persist in patients after remission and eventually resulted in recurrence. He continued treating patients until hCG was reduced to a normal level, and recurrence was not observed. These results suggested that cyclic treatment was required to prevent recurrence, while reducing toxicity (11, 12). On the basis of the above observations, in the 1960s, Howard Skipper hypothesized that patient survival was determined by the effect of chemotherapy on tumor growth kinetics, especially residual disease. He modeled drug-induced cell killing by subjecting syngeneic mice, with intraperitoneal implanted murine leukemia L1210, to chemotherapeutic agents. Therapeutic activity was assessed on the basis of percent median increase in lifespan, net log10 cell-kill, and long-term survivors. By manipulating the parameters in this model, he discovered: (i) leukemia cells grow exponentially and kill the host upon reaching the limit of tumor burden; (ii) efficacious therapy has to eradicate all the residual disease; and (iii) chemotherapeutic agents kill leukemia cells as a constant fraction, not a constant number, termed the “log-kill” response (13–15). These principles, consistent with clinical observations, were subsequently confirmed in human leukemias as well (16). This provided one of the cornerstones of modern chemotherapy for cancer treatment.

Our second example focuses on the modeling of metastasis. In 1889, Stephen Paget suggested that tumor cells (the “seed”) needed to be in a favorable microenvironment (the “soil”) to generate metastatic growth (17, 18). This “seed-and-soil” hypothesis fell out of favor for another hypothesis proposing that the preferential metastatic sites were driven by mechanical forces, where hematogenous spread ended with vascular cell trapping; however, neither hypothesis was proven at that time. In the 1980s, Hart and Fidler set out to elucidate the mechanism by modeling tumor growth at metastatic sites (19). To determine the effect of “soil,” they intravenously injected mice with B16 melanoma cells, a cell line previously shown to metastasize to the lung. They controlled for both vascular lodging as a mechanism and trauma by using additional organ tissue implants, such as renal tissue, and mice with a surgical incision but no implantation. The results showed that tumor cells disseminated into multiple sites but preferentially grew in the lungs at orthotopic or ectopic sites. To model the role of the individual “seed,” subclones of B16 melanoma cells isolated from subculturing were then injected into the mice. The results demonstrated that subpopulations preferring a specific metastatic growth site, such as brain versus lung versus liver, preexisted in the primary tumor. Taken together, these results strongly supported the “seed-and-soil” hypothesis (20, 21), and similar modeling was applied to the identification of site-specific targeting signals and genes in animal and clinical studies.

In addition to providing a means of describing the fundamental biology of cancer, modeling has also played a key role in the study and development of therapeutics. Again, models are built with the needed adjustable parameters to produce output that can be evaluated in the context of the hypothesis set to answer a predetermined question. In this case, the question may be the efficacy, safety, mechanism of, or resistance to, a therapy. For example, recently, a mouse BrafV600-mutant melanoma line was developed to answer the question of whether BRAF inhibitors, such as vemurafenib, can increase the T-cell response against melanoma. The scope of this model focused on immunogenic tumors that have the capability to interact with T cells (22). Therefore, melanoma cell lines derived from Braf-mutant mice were engineered and transplanted into syngeneic immunocompetent hosts, followed by adoptive T-cell transfer. In this case, the combined treatment tested superior to either treatment alone, and further evaluation showed that this enhanced response was due to increased cytotoxicity of tumor infiltrating lymphocytes (TIL) by vemurafenib. Several pathology studies of human melanoma tumor samples have shown increased TILs in BRAF inhibitor–treated tumors (23), which is consistent with the results from this model for the assessment of its clinical relevance. Clinical trials have been initiated to test the combination of BRAF inhibitor with immunotherapy that activates T cells (24), and the results can be used to revise the mouse model. This is an iterative process, as the clinical trials are informed by the preceding preclinical model studies.

Our final example of modeling comes from studies on immunotherapy mechanisms and treatment. It has been observed that only a subset of patients responds to anti-CTLA-4 and anti-PD-1/PD-L1 antibodies, and this heterogeneous response was associated with the degree of tumor T-cell infiltration. On the basis of their studies of clinical data, genetic analysis of human tumors, and IHC, Spranger and colleagues found that cutaneous melanomas lacking significant T-cell infiltration were associated with an active β-catenin pathway (25). To explore this association, the authors built mouse models to compare BrafV600E/Pten−/− mice with mice that additionally had expression of activated β-catenin. Similar to the clinical findings, murine melanomas expressing active β-catenin barely had any TILs, whereas tumors without the active β-catenin had substantially more TILs and delayed tumor growth. Further investigation determined that the mechanism of impaired T-cell infiltration was due to defective recruitment and activation of dendritic cells, resulting in insufficient T-cell priming. Moreover, intratumoral injection of dendritic cells into the mice bearing melanomas with active β-catenin led to significantly improved response to anti-CTLA-4 and anti-PD-L1 antibodies. This model system was designed to investigate a mechanism of therapy resistance and concurrently uncover a potential route to circumvent that resistance. The replication of human tumor biology in mice enabled controlled dissection of the pathways and mechanisms involved, leading to a means toward future therapeutic improvements.

Considering the question from the perspective of the tumor-bearing host, the same group asked whether heterogeneity in therapeutic responses may be caused by differential host physiology. On the basis of previous research suggesting that the intestinal microbiome affects antitumor immune response in patients undergoing chemotherapy, they hypothesized that the microbiome may also play a role in the efficacy of immunotherapy. To build the model, two strains of C57BL/6 mice from different mouse facilities with unique microbiota were implanted with subcutaneous B16 melanoma and treated with anti-PD-L1 therapy. Indeed, the model system showed distinctively different response rates in the two separately housed mouse groups. The outcome was further evaluated by performing fecal transplants between the mouse groups, and the results supported their hypothesis (26). Clinical studies are needed to validate these initial results in patients, which in turn will generate new hypotheses for further testing by modeling.

Although all models have limitations and are imperfect representations of real systems, they can be chosen such that their limitations are minimized in relation to the biology under consideration; a good model should provide potentially translatable results as well as insights into molecular mechanisms. How to extrapolate the results from modeling can be guided by the concept of resolution. When studying resistance to a targeted therapy, it may be useful to build the model aiming at high resolution (e.g., signaling pathways). In such cases, the focused phenomenon can be modeled even in the unmatched tissue of nonmammalian animals, as long as it is well conserved among species (such as fruit fly and zebrafish cancer models). However, in other instances, studying the disease at a lower level of resolution (e.g., phenotype and emergent system properties) may be necessary for identifying the driving factors of the system. Overemphasis on molecular mechanism regardless of the required resolution, especially aiming to attribute the phenotypes to one pathway or one gene, may reduce the scope of the modeling, produce bias, and impede progression to the next cycle.

An important consideration in the evaluation of model outcomes is the compatibility of the system outputs. For example, when modeling human diseases in mice, survival is not the same endpoint as decreased disease burden. Careful analysis and critical appraisal of results are necessary for meaningful conclusions. Importantly, modeling is a continuous improvement process, rather than a goal of creating a one-time tool. Modeling results can be applied successfully to human biological mechanisms and to therapy, and shortcomings can be reevaluated in the improved model to better reflect the real system of study. Modeling has greatly informed the field of biology and, with the incorporation of new tools, can continue to provide essential, valuable knowledge to the understanding of complex systems.

The ability to improve the modeling process with modern tools will continue to enrich the means of delineating complex biological systems. This can be realized by incorporating essential new humanizing approaches, such as the use of human cancer genomic data and advances in human tissue engineering. Cancer genomics has provided extensive mechanistic and therapeutic targets to investigate, and modeling provides a controllable system for evaluating the complex genetic interplay with cancer biology. Model systems are a powerful approach for studying drug resistance, cancer recurrence, and treatment optimization and, with continued model reassessment and improvisation, can lead to an excellent means of providing guidance for clinical trials. Modeling also enables a viable means of exploring precision medicine, by mapping out unique characteristics that can significantly guide and improve therapy outcomes in a small subset of individuals.

No potential conflicts of interest were disclosed.

We thank Drs. Lalage Wakefield and Ji Luo (Laboratory of Cancer Biology and Genetics, National Cancer Institute, NIH, Bethesda, MD), and Christina Stuelten (Laboratory of Cellular and Molecular Biology, National Cancer Institute, NIH, Bethesda, MD), M. Raza Zaidi (Fels Institute for Cancer Research and Molecular Biology, Temple University, Philadelphia, PA), and Heather Farthing (University of Miami Miller School of Medicine) for constructive comments and discussion.

This research was made possible through the NIH Medical Research Scholars Program, a public–private partnership supported jointly by the NIH and generous contributions to the Foundation for the NIH from the Doris Duke Charitable Foundation, The American Association for Dental Research, The Howard Hughes Medical Institute, and the Colgate-Palmolive Company, as well as other private donors. For a complete list, please visit the Foundation website at: http://fnih.org/work/education-training-0/medical-research-scholars-program.

1.
Lazebnik
Y
. 
Can a biologist fix a radio?–or, what I learned while studying apoptosis
.
Cancer Cell
2002
;
2
:
179
82
.
2.
Costanza
R
,
Wainger
L
,
Folke
C
,
Maler
KG
. 
Modeling complex ecological economic-systems - toward an evolutionary, dynamic understanding of people and nature
.
Bioscience
1993
;
43
:
545
55
.
3.
Brown
JR
,
Thornton
JL
. 
Percivall Pott (1714-1788) and chimney sweepers' cancer of the scrotum
.
Br J Ind Med
1957
;
14
:
68
70
.
4.
Pott
P
. 
Cancer Scroti
.
Chirurgical observations relative to the cataract, the polypus of the nose, cancer of the scrotum, different kinds of ruptures, and the mortification of the toes and feet
.
London, United Kingdom
:
Hawes, Clarke, Collins
; 
1775
. p.
63
8
.
5.
Yamagiwa
K
,
Ichikawa
K
Experimental study of the pathogenesis of carcinoma
.
J Cancer Res
1918
;
3
:
1
29
.
6.
Lippman
SM
,
Hawk
ET
. 
Cancer prevention: from 1727 to milestones of the past 100 years
.
Cancer Res
2009
;
69
:
5269
84
.
7.
Lasserre
JP
,
Dautant
A
,
Aiyar
RS
,
Kucharczyk
R
,
Glatigny
A
,
Tribouillard-Tanvier
D
, et al
Yeast as a system for modeling mitochondrial disease mechanisms and discovering therapies
.
Dis Model Mech
2015
;
8
:
509
26
.
8.
Xu
X
,
Kim
SK
. 
The early bird catches the worm: new technologies for the Caenorhabditiselegans toolkit
.
Nat Rev Genet
2011
;
12
:
793
801
.
9.
Brumby
AM
,
Richardson
HE
. 
Using Drosophila melanogaster to map human cancer pathways
.
Nat Rev Cancer
2005
;
5
:
626
39
.
10.
Shoemaker
RH
. 
The NCI60 human tumour cell line anticancer drug screen
.
Nat Rev Cancer
2006
;
6
:
813
23
.
11.
DeVita
VT
 Jr
,
Chu
E
. 
A history of cancer chemotherapy
.
Cancer Res
2008
;
68
:
8643
53
.
12.
Li
MC
,
Hertz
R
,
Bergenstal
DM
. 
Therapy of choriocarcinoma and related trophoblastic tumors with folic acid and purine antagonists
.
N Engl J Med
1958
;
259
:
66
74
.
13.
Skipper
HE
. 
Effects of chemotherapy on kinetics of leukemic cell behavior
.
Cancer Res
1965
;
25
:
1544
50
.
14.
Skipper
HE
,
Schabel
FM
 Jr
,
Wilcox
WS
. 
Experimental evaluation of potential anticancer agents. XIII. On the criteria and kinetics associated with curability of experimental leukemia
.
Cancer Chemother Rep
1964
;
35
:
3
111
.
15.
Skipper
HE
,
Perry
S
. 
Kinetics of normal and leukemic leukocyte populations and relevance to chemotherapy
.
Cancer Res
1970
;
30
:
1883
97
.
16.
Berenbaum
MC
. 
In vivo determination of the fractional kill of human tumor cells by chemotherapeutic agents
.
Cancer Chemother Rep
1972
;
56
:
563
71
.
17.
Paget
S
. 
The distribution of secondary growths in cancer of the breast
.
Lancet
1889
;
1
:
571
73
.
18.
Fidler
IJ
. 
The pathogenesis of cancer metastasis: the ‘seed and soil’ hypothesis revisited
.
Nat Rev Cancer
2003
;
3
:
453
8
.
19.
Hart
IR
,
Fidler
IJ
. 
Role of organ selectivity in the determination of metastatic patterns of B16 melanoma
.
Cancer Res
1980
;
40
:
2281
7
.
20.
Fidler
IJ
,
Kripke
ML
. 
Metastasis results from preexisting variant cells within a malignant tumor
.
Science
1977
;
197
:
893
5
.
21.
Fidler
IJ
,
Gruys
E
,
Cifone
MA
,
Barnes
Z
,
Bucana
C
. 
Demonstration of multiple phenotypic diversity in a murine melanoma of recent origin
.
J Natl Cancer Inst
1981
;
67
:
947
56
.
22.
Koya
RC
,
Mok
S
,
Otte
N
,
Blacketor
KJ
,
Comin-Anduix
B
,
Tumeh
PC
, et al
BRAF inhibitor vemurafenib improves the antitumor activity of adoptive cell immunotherapy
.
Cancer Res
2012
;
72
:
3928
37
.
23.
Wilmott
JS
,
Long
GV
,
Howle
JR
,
Haydu
LE
,
Sharma
RN
,
Thompson
JF
, et al
Selective BRAF inhibitors induce marked T-cell infiltration into human metastatic melanoma
.
Clin Cancer Res
2012
;
18
:
1386
94
.
24.
Hu-Lieskovan
S
,
Robert
L
,
Moreno
BH
,
Ribas
A
. 
Combining targeted therapy with immunotherapy in BRAF-mutant melanoma: promise and challenges
.
J Clin Oncol
2014
;
32
:
2248
54
.
25.
Spranger
S
,
Bao
R
,
Gajewski
TF
. 
Melanoma-intrinsic β-catenin signalling prevents anti-tumour immunity
.
Nature
2015
;
523
:
231
5
.
26.
Sivan
A
,
Corrales
L
,
Hubert
N
,
Williams
JB
,
Aquino-Michaels
K
,
Earley
ZM
, et al
Commensal Bifidobacterium promotes antitumor immunity and facilitates anti-PD-L1 efficacy
.
Science
2015
;
350
:
1084
9
.