Clarke, Endocrine-Related Cancer) Systems Biology - Perspectives On Multiscale Modeling in Research On Endocrine-Related Cancers, 2010

Endocrine-Related R Clarke et al.
Multiscale modeling in cancer 26:6 R345–R368

Cancer systems biology
REVIEW
Systems biology: perspectives on multiscale

modeling in research on endocrine-related
cancers
Robert Clarke1, John J Tyson2, Ming Tan3, William T Baumann4, Lu Jin1, Jianhua Xuan5 and Yue Wang5
1Department of Oncology, Georgetown University Medical Center, Washington, District of Columbia, USA
2Department of Biological Sciences, Virginia Polytechnic Institute and State University, Blacksburg, Virginia, USA
3Department of Biostatistics, Bioinformatics & Biomathematics, Georgetown University Medical Center, Washington, District of Columbia, USA
4Department of Electrical and Computer Engineering, Virginia Polytechnic Institute and State University, Blacksburg, Virginia, USA
5Department of Electrical and Computer Engineering, Virginia Polytechnic Institute and State University, Arlington, Virginia, USA
Correspondence should be addressed to R Clarke: [email protected]
Abstract
Key Words
Drawing on concepts from experimental biology, computer science, informatics,
ff systems biology
mathematics and statistics, systems biologists integrate data across diverse platforms
ff mathematical biology
and scales of time and space to create computational and mathematical models of the
ff computational biology
integrative, holistic functions of living systems. Endocrine-related cancers are well suited
ff predictive modeling
to study from a systems perspective because of the signaling complexities arising from
the roles of growth factors, hormones and their receptors as critical regulators of cancer
cell biology and from the interactions among cancer cells, normal cells and signaling
molecules in the tumor microenvironment. Moreover, growth factors, hormones and
their receptors are often effective targets for therapeutic intervention, such as estrogen
biosynthesis, estrogen receptors or HER2 in breast cancer and androgen receptors in
prostate cancer. Given the complexity underlying the molecular control networks in these
cancers, a simple, intuitive understanding of how endocrine-related cancers respond
to therapeutic protocols has proved incomplete and unsatisfactory. Systems biology
offers an alternative paradigm for understanding these cancers and their treatment.
To correctly interpret the results of systems-based studies requires some knowledge
of how in silico models are built, and how they are used to describe a system and to
predict the effects of perturbations on system function. In this review, we provide a
general perspective on the field of cancer systems biology, and we explore some of the
advantages, limitations and pitfalls associated with using predictive multiscale modeling Endocrine-Related Cancer
to study endocrine-related cancers. (2019) 26, R345–R368
Introduction
Over the past few decades, many advances in endocrine- approach, focusing on mechanistic studies of specific
related cancers have come from the experimental fields of genes and proteins, linear signaling pathways, and
cellular and molecular biology and from their translation particular anticancer drugs and other interventions. A
into clinical applications. Generally speaking, cellular systems-based approach builds on this important work by
and molecular studies have taken a mostly reductionist providing a more holistic account of the complex networks
https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://doi.org/10.1530/ERC-18-0309 Published by Bioscientifica Ltd.
Printed in Great Britain
Downloaded from Bioscientifica.com at 11/27/2019 12:44:33AM
via free access
Endocrine-Related R Clarke et al. Multiscale modeling in cancer 26:6 R346
of interacting genes, proteins and metabolites that that a systems approach, including computational and
determine how a cancer cell survives and thrives within mathematical modeling of new data streams, is essential
the tumor microenvironment and how the host responds to transform data into actionable knowledge that leads
to the tumor. From this viewpoint, molecular networks to fundamental improvements in human health. An
and the subcellular processes they regulate are seen to overview of the organization of this review is provided in
interact with activities occurring within the tumor cell, Fig. 1. We begin with a section on why models are needed,
its microenvironment and the cancer-bearing organism. how modelers generally approach building their models,
A holistic view, where interactions can have both local and some considerations regarding the specific goals of
and distant effects, is nothing new for endocrinologists modeling. Next, we describe how models may be based
and experts in some other fields. However, in what is now on a modular structure, and how modularity can lead to
often referred to as the ‘post-genomic era’, the tools and emergent behaviors, as consequences of the dynamical
technologies available to effectively study any cancer as properties of signaling networks. We discuss deterministic,
a systems-disease have changed dramatically. In concert stochastic and Bayesian models, and how their parameters
with these advances has come greater insight into the are estimated from data and provided with error bounds.
remarkable complexity of signaling, its integration and We then discuss model performance, potential sources
the coordination evident in controlling and executing of error, the importance of independently validating
cellular functions. model predictions and modeling drug interactions.
In this review article, we hope to introduce a broad Subsequent sections discuss examples of a knowledge-
readership to the potentials and limitations of a systems guided computational tool for building networks, a
approach to improve our understanding and treatment of mathematical model of the estrogen receptor landscape
endocrine-related cancers. The scope of endocrine-related and some insights into interpreting models.
cancer systems biology is large and complex, and we For our purposes in this review, a system is a
acknowledge that some issues in this field are addressed collection of interacting components that produces a
here at a relatively simplistic level. Nonetheless, we believe defined biological output in response to specific inputs.
Figure 1
Representation of data streams and how these relate to computational and mathematical modeling in the context of systems biology. The four primary
sections of this review contain specific insights into different aspects of modeling that reflect how modeling uses data streams to build multiscale
models. We first describe why models are needed in ‘Why build models’. The second section ‘Multiscale modeling’ introduces several critical aspects of
modeling, from some basic goals of modeling, then describing how models can use a modular structure that can explain the emergent properties of
biological systems. Deterministic, stochastic, and Bayesian models are then presented, as is the critical feature for cancer therapies of strategies to model
drug interactions. These subsections are followed by a discussion of types of error in modes, assessing model performance, and validating model
predictions. The final two subsections within the section on multiscale modeling provide specific examples of tools or approaches to modeling: a
knowledge-guided computational tool for building networks, and a mathematical model of the estrogen receptor landscape. The penultimate section
‘Interpreting models’ provides some insights into the challenges and pitfalls of interpreting model solutions. The final section ‘Future directions’ offers
some brief insights into where the authors see the field going in the next few years.

via free access
To be useful, such input–output models must adequately we consider a ‘mathematical model’ as using differential
capture the complexity of the system. Complexity does equations and stochastic algorithms to create dynamic,
not necessarily mean ‘big’ (many nodes and edges). semi-mechanistic models of control networks of limited
Relatively small networks can exhibit non-intuitive scope (dozens of genes and their products). Of course,
signal-processing capabilities due to inherent feedforward such dynamical models must ultimately be simulated
and feedback loops and non-linear kinetic rate laws, for on a digital computer, but we consider a ‘computational
which small changes in input produce disproportionately model’ as something different: as using machine-learning
large changes in output. tools to explore high-dimensional data (hundreds or
Most biological systems are open, complex, dynamic thousands of genes and/or proteins).
and adaptive. While these fundamental properties may Mathematical models may be deterministic or
be missed in work that adopts a solely reductionist stochastic in nature, depending on the role of random
perspective, there would be little for systems biologists events in the system being modeled. In either case, all
to model without the data and insights obtained from models ultimately entail a statistical evaluation of how
reductionist studies. Systems biologists acknowledge both well the model’s output fits the available experimental
the complexity of biological systems and the fact that data. Both stochastic and deterministic models can be
much of what must be modeled and interpreted is still useful when used appropriately (Twycross et al. 2010).
poorly understood. Computational and mathematical At present, deterministic models are usually the initial
models are often used to analyze and integrate data from approach taken to provide a description of molecular
multiple technological platforms into new representations events in cellular control systems. However, considering
of system function. These new representations can the paucity of informative data within the flood of omics
expand our understanding of complex regulatory systems results, the unavoidable noise in biological measurements,
(Lavrik & Zhivotovsky 2014, Wang & Deisboeck 2014, and our ignorance of latent variables in regulatory
Altrock et al. 2015, Peng et al. 2016, Janes et al. 2017, Ji networks, stochastic (Wilkinson 2009) or hybrid models
et al. 2017). Ultimately, systems-based insights into the (Twycross et al. 2010) are being applied more widely.
biology of endocrine-related cancers may lead to better Some of the general limitations in modeling have been
treatments and outcomes for patients (Werner et al. 2014, discussed elsewhere (Di et al. 2006, Wilkinson 2009,
Jinawath et al. 2016, Ji et al. 2017). Twycross et al. 2010) and will not be reiterated here.
While the idea of generating mathematical models of From a clinical perspective, useful in silico models
signal flow in a biological system is not new (Le 2007, Ji et al. will have to be multiscale. For example, drug action at
2017), the sources and magnitude of data for multiscale the molecular scale must be linked to clinical outcomes at
modeling, and many of the computational/mathematical the tissue or organism scale. Multiscale models use many
tools available, have changed dramatically in recent years. different data types from multiple sources, spanning scales
Many of the newer technologies fall into the rapidly from DNA to RNA to protein, from metabolites to cells to
developing fields of omics (genomics, transcriptomics, tissues, from tissues to organisms and even to interacting
proteomics, metabolomics), an increasing number of populations. Modeling based only on genome and/or
sub-omic technologies and quantitative microscopy transcriptome data can be limited because approximately
including gene expression in single cells (Sandberg 2014, 50% of changes found in the transcriptome may not be
Buettner et al. 2015, Kanter & Kalisky 2015). Central to present in the proteome (Vogel & Marcotte 2012); an
our ability to analyze and integrate these new data streams even smaller percentage of changes in the genome may
and to build new mathematical models and computational filter through to the proteome. Hence, spanning scales
representations of the data, are the analytical approaches (provided necessary data are available) may improve the
and software tools that continue to be developed by models and provide new insights into cancer physiology
computer scientists, mathematicians and statisticians. (Deisboeck et al. 2011).
Rather than being identified with any of these particular In this review, we explore some of the basic concepts and
specializations, systems biology sits uniquely at their challenges in applying computational and mathematical
nexus. modeling to endocrine-related cancer research. Rather
We will focus our discussion on the use of than providing detailed descriptions of tools-of-the-trade,
computational and mathematical approaches to we discuss a variety of computational and mathematical
model system function in the context of endocrine- approaches that are often applied, the advantages and
related cancer biology. For the purposes of this review, limitations of each, and the specific challenges for using

via free access
them correctly and usefully. Since we will not discuss endocrine-related cancers is their cellular heterogeneity,
specific experimental designs here, readers interested in which creates a dynamic microenvironment of many cell
exploring the many tools, workflows and frameworks and types in addition to the cancer cell component and can
emerging standards for systems-based research may find also affect a tumor’s response to treatment (Junttila & de
the following sources useful (Brazma et al. 2006, Swertz & Sauvage 2013, Meacham & Morrison 2013, Martelotto
Jansen 2007, Gehlenborg et al. 2010, Ghosh et al. 2011, et al. 2014). Often, models are built with transcriptome
Wu & Stein 2012, Hofree et al. 2013, Sedgewick et al. 2013, data that reflect averaged expression values, since tissue
Wen et al. 2013, Cheng et al. 2014a,b, Hoadley et al. 2014, microdissection prior to collecting omic data remains
Creixell et al. 2015, Leiserson et al. 2015, Dimitrova et al. relatively uncommon. When applying computational
2017, Nam 2017, Keenan et al. 2018, Miryala et al. 2018). and mathematical modeling to study cell type/tissue type
Similarly, there are many sources of cancer omics data in in data from complex tissue samples, data deconvolution
the public domain that are too numerous to capture here. using either supervised or unsupervised approaches is
However, we provide examples of some widely used large a prerequisite. Supervised data deconvolution can be
omics datasets that include data from breast and other performed by integrating tissue-specific gene or protein
endocrine-related cancers in Table 1. expression profiles (Newman et al. 2015) from the Gene-
Given clear evidence of a significant lack of Tissue Expression program (GTex Consortium 2015)
reproducibility in biomedical research (Begley 2013, and The Human Protein Atlas (Ponten et al. 2011).
Mobley et al. 2013, Hatzis et al. 2014) and the potential Alternatively, in the more challenging case of intra-tumor
for systems approaches to both reduce and exacerbate this heterogeneity where subclone-specific markers are often
problem, an appreciation of some of the key challenges unknown, an unsupervised data deconvolution approach
– for which there may or may not be adequate current such as Convex Analysis of Mixtures can be exploited
solutions – is timely. While we cannot address all the to uncover the hidden subclone specificity (Wang et al.
major issues in such an interdisciplinary subject, we hope 2015, 2016, Herrington et al. 2018). While tools for
that our perspective will be pertinent to using systems supervised (Zuckerman et al. 2013, Hart et al. 2015)
biology to attain a better understanding of endocrine- and unsupervised deconvolution of averaged data from
related cancers. heterogeneous tissues (Chen et al. 2011, Wang et al. 2016)
can be used as a data processing step prior to modeling –
this preprocessing step remains uncommon.
Why build quantitative models of The properties of high-dimensional data, particularly
biological systems? data from omics technologies, present unique challenges
(Clarke et al. 2008) that are often inadequately addressed
‘The statistician knows, for example, that in nature there or fully appreciated. Nonetheless, the purpose of in
never was a normal distribution, there never was a straight silico analysis is to apply tools to extract meaningful
line, yet with normal and linear assumptions, known to be results from high-dimensional data for the purposes of
false, he can often derive results which match, to a useful generating and testing biological hypotheses (Tyson et al.
approximation, those found in the real world.’ 2011). For instance, we may wish to understand and
predict under what conditions a cancer cell will begin to
George E P Box (1919–2013)
proliferate in situ or migrate to a new location. Extracting
To extract new insights and build integrated, predictive such knowledge from large datasets by intuitive reasoning
models, particularly from experiments that generate ‘big alone can be difficult or impossible and is often associated
data’, requires some form of in silico analysis to deal with the with a high risk of operator bias and/or error. Thus, new
complexity of the data. For biological systems, complexity tools and approaches continue to emerge to deal with the
can arise from dimensionality (many genes and their challenges of working in high-dimensional data spaces
interactions) and from general properties of the system and to enable integrating the spatial, temporal and cell
that reflect its topology (feedforward and feedback loops), context-specific nature of regulatory networks (Hoadley
adaptability (redundancy, degeneracy), multimodality et al. 2014, Leiserson et al. 2015, Masoudi-Nejad et al. 2015,
(concurrent performance of multiple integrated and Tape 2016, Barberis & Verbruggen 2017, Dimitrova et al.
coordinated tasks) and dynamism (changes in time and 2017). New concepts, such as ‘master regulator proteins’
space) (Clarke et al. 2008, Tyson et al. 2011). Complexity can that may determine the transcriptional state of a cancer
also arise at the cellular level. A notable feature of several cell, also continue to arise (Califano & Alvarez 2017).

via free access
Table 1 Examples of the most commonly used endocrine-related breast cancer public omic datasets.
Database URL Data spaces
CPTAC https://proteomics.cancer.gov/data-portal Proteome

EGA https://ega-archive.org/datasets Genome, Transcriptome
EMBL-EBI https://www.ebi.ac.uk/services/all Genome, Transcriptome,
Proteome, Metabolome
GNPS/Massive https://gnps.ucsd.edu/ProteoSAFe/static/gnps-splash.jsp Metabolome
ICGC https://dcc.icgc.org/ Genome, Transcriptome
MassIVE https://massive.ucsd.edu/ProteoSAFe/static/massive.jsp Proteome
Metabolomics https://www.metabolomicsworkbench.org/ Metabolome
Workbench
NCBI-GEO https://www.ncbi.nlm.nih.gov/gds Genome, Transcriptome
ONCOMINE https://www.oncomine.org/resource/login.html Genome, Transcriptome
ProteomeXchange (PX) Consortium http://www.proteomexchange.org/ Proteome
ProteomicsDB https://www.proteomicsdb.org/ Proteome
TCGA https://portal.gdc.cancer.gov/ Genome, Transcriptome
Primary data and metadata quality vary across and within these sites. For example, clinical metadata for human subjects are often limited. The platform
used for data collection in each omics space also can vary across and within these sites. While most provide access to the raw (unprocessed) data,
ONCOMINE primarily exposes only processed data; the method of data processing can vary across individual studies.
Computational modeling can provide unbiased gene set enrichment analysis can rapidly probe a large
results from large data sets, allowing us to visualize database of genes and their hierarchically annotated
complex signaling relationships within the data functions to suggest signaling pathways closely affiliated
(Gehlenborg et al. 2010). Some of the more useful with a list of differentially expressed genes (Subramanian
approaches in this area come from applications of graph et al. 2005); for example, see http://software.broadinstitute.
theory. Graphs are mathematical structures that represent org/gsea/index.jsp. A pathways database and a search tool
pairwise relationships between nodes. Each gene/protein is also provided by the Gene Ontology Consortium (see
is a node (or vertex) and each connection with another http://geneontology.org/page/go-enrichment-analysis).
gene/protein is an edge. Graphical representations of Given adequate data, both computational and
molecular signaling are readily available on the web. For mathematical models can make quantitative predictions
example, signal transduction pathways may be found at of the biological state under investigation. One of the
the community-based Kyoto Encyclopedia of Genes and primary uses of quantitative models is to perform in silico
Genomes (KEGG; http://www.genome.jp/kegg) or the experiments where the values of specific nodes or edges
commercially supported Biocarta Pathways Project (http:// are changed and the model is used to predict how the
www.biocarta.com/genes/index.asp). These graphical change affects other nodes in the network. It is possible to
representations are mostly assembled intuitively from the run hundreds or thousands of such simulations to explore
literature to provide a static reflection of the topological both model performance and how specified changes in
features of mostly canonical signaling networks. node/edge values affect the distribution of predicted
Static maps are widely used to represent complex outcomes. For example, in silico modeling can be used to
signaling networks and to guide largely intuitive compare multiple drug combinations including the effects
interpretations of signaling, but they are of limited use of scheduling and dosing that would be very difficult
for predicting signal flows through edges of the network in animal models or even in some cell culture models
in a living cell responding to signals received from its (Tang & Aittokallio 2014, Ryall & Tan 2015, Ledzewicz &
environment. Limited dynamic information may be Schaettler 2016). Appropriate quantitative models, when
evident in the directionality of signal flow (such as, effectively applied to sufficient, high-quality data, can
protein A upregulates the production of protein B), but enable investigators to explore questions in ways that
the consequences of many such interactions in a complex, would otherwise not be possible. Visualization of the
interconnected network are challenging to predict by outputs from computational analysis of high-dimensional
intuitive reasoning alone. Appropriate computational data can be an indispensable aid in interpreting the
models can help to uncover complex associations hidden biological significance of the data (Gehlenborg et al. 2010,
in the data and often may provide a statistical assessment Cirillo et al. 2017, Pavlopoulos et al. 2017, Robinson et al.
of the strength of any predicted association. For example, 2017). Thus, multiscale modeling enables investigators to

via free access
explore complex datasets and signaling in new ways that and new predictions will be generated that can be used to
are both tractable and productive. test the modified assumptions.
A suitable framework to guide the modeling effort
is a key starting point. The framework describes, at a
Multiscale modeling high level, what is generally known about the system in
the context of integrated modules that perform specific
‘Numquam ponenda est pluralitas sine necessitate’ cellular functions. Thus, a modular function, such as cell
(Plurality should not be proposed unnecessarily) death, may be explained by a model of the signaling that
controls and executes one or more forms of cell death,
William of Occam (c. 1287–1347)
such as apoptosis. Where there is sufficient knowledge
‘Since all models are wrong the scientist cannot of an individual module, a reasonably detailed influence
obtain a ‘correct’ one by excessive elaboration. On the diagram of known or predicted signaling relationships
contrary following William of Occam he should seek an can be created to guide construction of the mathematical
economical description of natural phenomena.’ equations. This knowledge can be gained from specific
experimental data available in the laboratory, from the
George E P Box (1919–2013)
literature, or perhaps based on a static canonical model
All models are abstract representations of the system they as might be obtained from KEGG or Biocarta. Where a
are built to portray. The types of models we consider here canonical model does not exist (or there is good reason
are not intended to explain all of cancer biology. Rather, to believe that canonical signaling is inadequate),
we use models to learn something new about how a computational modeling can be used to formulate new
specific function may operate, be controlled and interact hypotheses about the topology of a control module from
with other cellular functions to affect a specific biological high-dimensional data (Clarke et al. 2011). Where there is
outcome. For example, we may wish to understand how sufficient knowledge of the components and interactions
estrogens affect the decision of some breast cancer cells to of a control system, the interaction diagram can be
enter and complete a turn of the cell cycle. Understanding translated into a set of mathematical equations that
this function could then lead to addressing larger goals, quantitatively represent dynamical fluxes through the
such as developing new therapeutic interventions to network (readers interested in exploring specific in silico
block cell cycling or predicting which patients would models can find examples in several databases including
receive the greatest benefit from blocking this action of JWS Online, available at http://jjj.biochem.sun.ac.za/
estrogens. Thus, the primary goals of modeling are to give index.html, and the Biomodels Database http://www.ebi.
insights into how a control system works at the molecular ac.uk/biomodels-main/). An example of such a framework
level and to make robust, reliable predictions about how can be seen in our roadmap for systems modeling
the system responds to a variety of natural situations and of endocrine responsiveness in breast cancer
medical interventions. (Tyson et al. 2011).
For molecular signaling studies, the latter goal can At some level, useful models need to address
be achieved by changing the values of parameters in the open, complex, dynamic and adaptive nature of
the model and experimentally validating the predicted biological systems. While we do not intend to provide
outcomes. Given a perturbation or rewiring of a control a detailed description of the concepts and methods of
network, the output of a model is a prediction of the model building, we can mention some general, widely
changed state of the cell (for example, alive or dead; applicable principles. First of all, we must keep our end-
proliferating or growth arrested). When simulations of a goal in mind (what aspect of cancer cell physiology are
model under a variety of realistic conditions inadequately we trying to understand) as well as our starting point
reflect what is already known to occur in cells, the model (what is our working hypothesis about the underlying
must be modified or extended. For example, a model may control system). Then, ideally, we would like to get
predict that reducing the expression of one gene should from the working hypothesis to accurate predictions of
increase the expression of another, but the observed result cell behavior with a model that is as simple as possible,
of this experiment (perhaps using an RNAi approach) but not so simple as to leave out crucial features of the
is the opposite. By considering how to resolve this molecular biology or cell physiology. Of course, these
discrepancy between model and experiment, new insights are vague and often antithetical requirements (what is
may be gained into how the control system works, simple? what is crucial?), but it is the job of the modeler

via free access
to make informed decisions about how much detail can An example of the classification task would be the
and should be included in the mathematical model. use of gene expression data from a patient’s tumor to
Often these uncertainties can be addressed by an iterative predict the patient’s prognosis and/or to determine the
approach, involving knowledge-guided trial-and-error or best choice of treatment. Among the simplest examples
the use of multiple feature selection tools (as an example is the heuristic guide for the treatment of breast cancer
see the Feature Selection functions by MathWorks, http:// patients based on a three-gene classification scheme:
www.mathworks.com/help/stats/feature-selection.html). estrogen receptor alpha (ER), progesterone receptor
To highlight issues that may be useful for the non- (PGR) and HER2. Knowledge of the expression of these
expert wishing to evaluate published models and/or to three genes defines three molecular subgroups: ER and/or
collaborate with modelers, we next address the utility PGR-positive (can be treated with an endocrine therapy),
and methodology of mathematical and computational HER2-positive (can be treated with an anti-HER2 therapy,
modeling. In our studies, we use computational tools to approximately half of these also express ER and/or PR and
extract small, robust and information-rich topological may also receive an endocrine therapy) and absence of
features from high-dimensional data sets. These features expression of all three – often referred to as triple-negative
can then be tested and validated experimentally, and breast cancer (TNBC) – which is usually treated with
at this stage, a simple mathematical model may be cytotoxic chemotherapy. A similar goal is exemplified by
useful in capturing this knowledge, working out its using a panel of clinical/pathological measures to predict
implications, and making predictions to guide further prognosis in breast cancer; an example being the semi-
laboratory experiments (Clarke et al. 2011). This iterative quantitative assessment that produces the Nottingham
approach requires a modeling framework (a network Prognostic Index (Galea et al. 1992). Classifiers based on
diagram), some relevant experimental data, and a basic omics data are also available and in common clinical
understanding of how components of the network may use, including the 70-gene signature that comprises the
interact to produce observed physiological responses of MammaPrint prognostic predictor (Bedard et al. 2009)
cells. The network diagram guides the construction of and the prognostic PAM50 gene signature (Parker et al.
the mathematical model, which can be used to compute 2009). Signatures that have not yet become adopted
the expected behavior of the simulated cells. To carry widely in the clinic continue to emerge (Wu & Stein 2012,
out simulations, we must first estimate the values of the Cheng et al. 2013). The output from these types of models
parameters (such as rate constants and binding constants) is a prediction of the future behavior of the cancer – a
in the mathematical model. Parameter estimation is a clinical outcome such as an estimate of patient survival
difficult problem, but it can (and must) be carried out in (prognosis) – often within a defined time period.
light of existing experimental data (Tyson et al. 2011). Omics-based classifiers (most frequently
There would be no rationale to include a parameter transcriptomic) are usually built using a supervised
without some data or direct evidence of its involvement approach, where a training set of data from samples
in reactions, and these data can provide bounds on with known outcomes is used and the predictive model
parameter values in the mathematical model. Once is subsequently validated in independent datasets.
an initial model adequately accounts for the existing Classification models often rely primarily on the statistical
data, it can be used to predict specific outcomes of new properties of each measurement/input variable and do
experiments that can be run to confirm, extend or adjust not require that these properties derive specifically from
the model. Thus, iterative modeling with the addition of any biological function of the system (Clarke et al. 2008).
new data allows both testing and refining of the model, The literature contains many different attempts to build
which leads to new biological insights (Clarke et al. 2011). classification schemes in breast cancer but often with
varying results and robustness, even for some of the most
widely used tools (Mackay et al. 2011, Venet et al. 2011).
Examples of modeling goals
While some schemes produce comparable outcomes on
Cancer systems biology studies tend to focus either on a common dataset, the features selected for classification
classification, where the goal is to predict a phenotype by each scheme often have little overlap (Imamov et al.
or outcome based on data, or on mechanistic modeling, 2005). Given the complexities in molecular signaling and
where the goal is to learn something new about how the selection of genes based on their statistical properties
the system (a tumor, a cancer cell or a signaling network to support classifier performance, it is not clear whether
within the cell) functions (Clarke et al. 2011). this observation reflects different genes representing

via free access
similar underlying processes (Imamov et al. 2005) or a lack mRNA regulation in microarray data are often used to
of robustness in feature selection unrelated to biology. validate these signatures. Studies with RNAi or cDNA
Network-based classification can also be performed overexpression, mostly done using cell lines growing
on individual patient data (Creixell et al. 2012). The key is in vitro, may also be used to further establish the influence
to develop a quantitative metric based on the topology of of gene expression on target gene regulation.
a learned network that can be applied to new observations These approaches may not account fully for the
to determine if the new observation is likely to share the complexity of a given target gene’s transcriptional
same topology. For example, once phenotype-specific regulation, such as whether factors other than the protein
networks are learned, a model-based likelihood measure complex that is detected as being bound to a specific
can be calculated to determine which topological promoter element are driving the measured differential
hypothesis is more likely generating the new observation, expression of the target gene. For example, TF1→Target
where the learned variance of network topology is used to Gene could still be driven through a latent variable(s),
support such likelihood-based hypothesis testing. since the same experimental outcomes could be seen if
The second goal of a systems analysis of data is to TF1 was knocked down and the true relationship was
generate new insights into mechanistic aspects of the TF1→TF2→Target Gene or even TF1⊣TF2⊣Target Gene.
cancer phenotype. For example, the model may be used Hence, both false-positive and false-negative regulatory
to understand why patients respond differently to a events may be obtained in addition to true events.
specific therapy or how molecular signaling regulates or For in silico modeling, including data on TF2 may or
executes a specific phenotype. Hence, the analysis may be may not affect model function. Where it does not,
structured to test if a series of proposed features might be the measurements of TF2 are superfluous and, in the
true (hypothesis testing) or to discover new features that interests of parsimony, can be eliminated from the model.
might explain mechanism (hypothesis generation). While Alternatively, there may be technical reasons that make
these models also frequently use the statistical properties the measurements of TF2 more reproducible than those
of the measurements to find signaling features of interest, of TF1. In this case, when TF1 and TF2 capture the same
there is an explicit assumption that the measurements, information, the model may perform better with TF2
and any changes in their values across phenotypes, are measurements than using those for TF1.
derived from relevant biological properties of the system.
Among the more common approaches for mechanistic
Modules and emergent behavior
studies is the use of transcriptome data to build gene
regulatory networks, as exemplified by a network of System models can be constructed as a network of
transcription factors (TFs) and the target genes that they integrated and interacting modules that perform the
are known, or predicted, to regulate. Insights from models system’s component operations in a coordinated manner
built primarily from transcriptome data can be limited (Tyson et al. 2011). The topology of signaling for a
by the often low frequency with which transcriptome module can be extracted de novo from the data, with
changes translate into similar expression changes in the functions being implied from any known activities of
proteome (Vogel & Marcotte 2012). The target genes for their member nodes (Wu & Stein 2012). However, for
TFs are identified either in silico (predicted using DNA modeling known functions where there is significant
sequence data; see MotifDb at (http://www.bioconductor. data and prior knowledge, modules can be viewed
org/packages/release/bioc/html/MotifDb.html) as an more discretely as integrated network components that
example of a tool for performing this function) or regulate and/or execute a specific function (Tyson et al.
experimentally (chromosome immunoprecipitation- 2011). For example, apoptosis could be considered as a
based methods; ChIP). These studies often produce small module that performs a cell death function; apoptosis
and mostly unidirectional maps (TF→target) and they can then be modeled as a discrete process, perhaps as a
can be noisy. For example, in silico predictions of targets closed, input–output device. Cells have other modules
based only on promoter sequences do not account for that perform similar functions, including autophagy
DNA structure/accessibility and are often incomplete. (which can produce prodeath or prosurvival outcomes).
Experimentally measured promoter occupancy (such as by These modules represent biological redundancy because
ChIP) does not always reflect functional regulation of the if an irreversible cell fate decision is made in favor of
adjacent gene. Correlations of measured (ChIP/ChIPseq) death, one of several differently constituted modules can
or predicted promoter sequence binding with differential execute that decision. Some genes may play key, but not

via free access
necessarily similar, functions in more than one of these in response to specific stresses is likely the emergent
modules. For example, BCL2 can regulate the activation property that drives both the phenotypic plasticity
of the autophagy module through its ability to sequester often attributed to cancer cells and the development of
BECN1, while also affecting execution of the apoptosis resistance to anticancer drugs. From an intuitive point-
module through its effects on mitochondrial membrane of-view, emergent properties are challenging because they
permeability (Clarke et al. 2012). Cell fate may depend on are difficult to deduce from a knowledge of the individual
the amount of BCL2 present and its subcellular location. components of the system, and the relationships between
For example, BCL2 bound to BECN1 may be unable to the emergent property and its component parts may be
protect the mitochondria, with BCL2:BECN1 complexes non-linear and dynamic (changing over time). To deal
effectively preventing the initiation of prosurvival reliably with these complexities requires comprehensive
autophagy (BECN1) and concurrently not preventing and accurate mathematical models to guide our thinking
apoptosis (BCL2). Since other prosurvival BCL2 family and predictions.
members can also bind to BECN1, the balance of Emergence may underlie many novel behaviors of
prosurvival-to-prodeath BCL2 family members (there cancer cells that cannot easily be foreseen from knowledge
is potentially significant signaling degeneracy within of the system’s individual components. In evolutionary
apoptosis), the concentration of free BECN1 remaining biology, emergence can reflect the development of
available to activate autophagy, and their respective larger or more complex functions or behaviors derived
subcellular localization(s) may all contribute to the final from the interactions among, but not shared with,
cell fate decision. The potential for cell context-specific individual smaller or less complex features (Okasha
wiring (and rewiring in response to stress) is evident. 2012, Gho & Lee 2017). New behaviors in tumors likely
A clear understanding of these interactions in ER+ arise through changes that affect interactions within
breast cancer cells requires both significant insight and and among modules. For example, changes in signaling
quantitative data from wet laboratory studies. Predicting from within the tumor microenvironment (adaptive)
cell fate outcomes robustly in the presence of various or the acquisition of a genetic/epigenetic change (such
endocrine stressors (estrogen withdrawal, exposure as activating or inactivating mutations) could alter the
to SERMs/SERDs) is unlikely to be successful without level of expression, function or subcellular location of
adequate in silico modeling. An effective dynamic model a molecule or the activity of a pathway in a network.
of these relationships could also be used to predict optimal Consequently, this pathway may now connect different
drug dosing and scheduling to drive maximal cell death modules that perform a new cellular function or
and potentially limit the emergence of drug resistance continue to perform an existing function in a different
(Tang & Aittokallio 2014, Ryall & Tan 2015). manner. Where these new emergent properties confer
Integration of modular functions allows a cancer a biological advantage, they are expected to experience
cell to coordinate and execute the activities it needs to positive selection (in a Darwinian sense) (Enriquez-
proliferate, survive, move and invade locally, respond Navas et al. 2015). Acquired drug resistance may be an
to stress and manage its metabolism to support these example of a new emergent property that is not evident
actions. Modules can be combined differentially in time in the initial cell population. Such resistance could
and space, creating some of the phenotypic diversity that be mutational (ER mutations that confer resistance
is characteristic of breast cancer cells. When modules to aromatase inhibitors in breast cancer) or adaptive
interact in complex feedback and feedforward loops, they (activation and integration of the unfolded protein
can exhibit redundancy (different modules performing response module with a prosurvival autophagy module
similar functions), degeneracy (different signaling routes that act together to confer resistance to antiestrogens)
allowing a module to perform the same function in (Clarke et al. 2011, 2012).
different ways) and novelty (the ability to perform new The emergent properties of cells in a system like an
functions or old functions in new ways). This plasticity ER+ breast tumor likely explain, in part, the phenotypic
of the response characteristics of modular networks is the heterogeneity of some breast tumors and also the diversity
origin of their ‘emergent’ properties (Bhalla & Iyengar of responses that confer drug resistance (Clarke et al.
1999). For example, an apoptosis module may be blocked 2012). The property of emergence with respect to acquired
in a cell but the cell death decision may now be executed multiple drug resistance (a function that is likely subject to
by an autophagy module. The ability to recombine positive selection), and the potential that some complex
signaling features in complex regulatory networks functions may never stabilize (the rate of appearance of

via free access
new metastatic foci may continue to increase throughout Nonetheless, like weather prediction, mathematical
the disease process), may underlie the high prevalence of models of cellular regulatory systems can be very useful
distant recurrences that are poorly responsive to available for short-term forecasting of local activity without being
systemic therapies, and so are generally fatal. reliable predictors of long-term ‘weather’ patterns on a
‘global’ scale.
Dynamics Parameters
One of the major strengths of quantitative mathematical To simulate a mathematical model, we must first estimate
modeling is the ability to capture the dynamic the values of all kinetic parameters from experimental
nature of a system (Aldridge et al. 2006, Anderson & observations. Examples of parameters include reaction
Quaranta 2008, Toettcher et al. 2009, Spencer & Sorger rate constants (such as protein synthesis and degradation,
2011, Molinelli et al. 2013). In particular, models of or phosphorylation and dephosphorylation) and binding
endocrine-related cancers have provided new insights or dissociation constants (for example, Michaelis
into the temporal development of invasive, metastatic constants for enzyme-catalyzed reactions). Estimation of
cells (Quaranta et al. 2008, Gallaher et al. 2014), drug- these parameter values is often the most difficult aspect of
treatment responses and drug-resistant states (Chen building a useful mathematical model (Liepe et al. 2014,
et al. 2013, 2014, Parmar et al. 2013, McKenna et al. Kimura et al. 2015). The goal of parameter estimation is
2017) and the origins of network plasticity (Tavassoly often not to find the ‘optimal’ set of parameter values
et al. 2015, Picco et al. 2017). Examples of some of the for fitting a selection of experimental results but rather
methods used in mathematical modeling are provided to find a representative collection of parameter sets that
in Table 2 (Tyson et al. 2019). all provide an ‘acceptable’ fit to the data (Tavassoly et al.
Despite their evident utility, dynamic models in 2015).
molecular cell biology must be interpreted cautiously. When faced with the dimensionality of data from an
Model predictions can be very accurate when restricted to omics platform, a mathematical model with thousands
conditions close to the experimental conditions on which of variables would be difficult to formulate and almost
the model was built, but less reliable when extrapolated impossible to parametrize. Currently, high-dimensional
far beyond the range for which they have been verified. data are more effectively explored using computational
Table 2 Methods of mathematical modeling.
Method Dynamic variables Time Example
Boolean networks X(t) = 0 or 1 t = integer X inhibits synthesis of Y and

Y(t) = 0 or 1 (0, 1, 2, …) Y inhibits synthesis of X
X (t + 1) = ¬Y (t )
Y (t + 1) = ¬X (t )
Ordinary differential equations X(t) = positive real number t = real number X inhibits synthesis of Y and
Y(t) = positive real number (t ≥ 0) Y inhibits synthesis of X
dX k
= sx − kdx X
dt 1 + Y p
dY ksy
= − kdyY
dt 1 + X q
Stochastic models M(t) = positive integer t = real number Propensity of mRNA synthesis = ksm
(t ≥ 0) Propensity of mRNA degradation = kdmM
Probability density function for number of
mRNA molecules in the cell is
λM
P (M ) = e − λ ksm
M ! , where λ = k
dm
Hybrid deterministic-stochastic M(t) = positive integer t = real number Genetic regulatory network:
models P(t) = positive real number (t ≥ 0) Simulate mRNA fluctuations, M(t), with a
stochastic model and protein dynamics,
P(t), with ordinary differential equations
Additional information can be found in Tyson et al. (2019). Reprinted from Journal of Theoretical Biology, Vol 462, Tyson JJ, Laomettachit T & Kraikivski P,
Modeling the dynamic behavior of biochemical regulatory networks, Pages 514–527, Copyright (2019), with permission from Elsevier.

via free access
modeling where the assumptions of the model are higher Whichever approach is selected, statistical models
level and less demanding of detailed kinetic information. (Bayesian or frequentist) have assumptions that can
For example, machine-learning techniques can learn the be violated and parameters (even non-parametric
features of molecular networks and their relationships probabilistic tools have parameters; these are not fixed in
from the data. Bayesian approaches are common in this advance but obtained from the data) that can be affected
regard and are discussed below. by the data structure and that can influence performance.
While it is not always evident which statistical model
is most appropriate for the data being analyzed,
Deterministic and stochastic models
understanding what the model outputs represent is
Deterministic models, defined usually by differential important for correctly inferring biological meanings or
equations, produce specific outcomes for a given set of appreciating the uses and limitations of the output.
parameter values and initial conditions, without any An increasingly common approach for computational
evidence of randomness. In contrast, stochastic models modeling is to build models that incorporate prior
evolve in time with significant random fluctuations knowledge of the system (Tian et al. 2014b, 2015). Prior
(Singhania et al. 2011, Barik et al. 2016). For example, a gene knowledge can be as simple as looking at the expression
regulatory network, where TFs regulate specific targets, levels of genes already known to contribute to the
could be modeled deterministically or stochastically. In a phenotype, at known interactions among molecules
deterministic model, the rate of gene transcription would such as protein–protein or protein–DNA interactions
have a definite value determined by the activity of the (PPIs or PDIs) or at relationships reported in canonical
transcription factor. In a stochastic model, the activity of signaling pathway representations. Incorporation of prior
the TF would determine only the propensity (probability knowledge, depending on the quality of the knowledge,
per unit time) of transcribing the gene into an mRNA can greatly improve the performance of algorithms to
molecule. In this case, a stochastic model represents build Bayesian networks. Indeed, a major challenge
more accurately the noisy process of gene transcription in constructing Bayesian networks is the selection of
in individual cells, but a deterministic model may capture appropriate prior probability distributions (priors) for the
adequately the average rate of expression of the gene over variables in the model. How these parameters are estimated
a population of cells responding to an external stimulus for a Bayesian approach affect its outcomes (Lampinen &
that is activating the TF. If we have data on the noise Vehtari 2001). Poorly estimated priors (relative to ground
associated with gene transcription in individual cells, truth – which is often unknown) may provide fits to
then a stochastic model may be warranted and needed. the data that are statistically acceptable and intuitively
Stochastic models have been useful for exploring the logical, but solutions that are, nonetheless, noisy and lead
dynamic responses of endocrine-related cancers (Jain to incorrect biological interpretations. Influence of the
et al. 2011, Chen et al. 2014, Morken et al. 2014). A prior can be reduced using Bayesian hierarchical models
deterministic model is simpler and more appropriate if and robust priors (Berger 2010).
we have only gross transcriptome data on populations of In Bayesian networks, the edges are directed but the
cells under constant conditions. sign is not specified. Consequently, whether the edge is
positive (such as driving) or negative (such as inhibiting)
must be inferred from sources external to the model and/or
Bayesian models
established experimentally. A further limitation is that
A general objective of computational tools is to find edges cannot be interpreted as necessarily reflecting direct
patterns (correlation structures) within data. For interactions. While some interactions may well be direct,
example, with transcriptomic data an algorithm may latent variables can also create direct edges in the model
look for patterns of changes in gene expression that are solution where none exist in the biological system. For
correlated with each other and with the phenotype(s) or example, the predicted edge of A→B in the model may
function(s) of interest (Dutta et al. 2016, Anafi et al. 2017, really be A→C→B (see also the discussion of modules and
Califano & Alvarez 2017). Some measure of the statistical emergent behavior, above). Inferring feedback loops can
strength of these correlations, using either a Bayesian also be difficult, such as A→C→B→A.
(conditional probabilistic) or frequentist (parametric or For gene network modeling, the quality of the
non-parametric probabilistic) approach, is usually applied knowledge and its incorporation into the selection of
to help identify the associations most likely to be correct. priors will improve the predictions. Two implications

via free access
follow from this observation. Firstly, a team with better Here, error propagation represents the effects of the
biological understanding of a system may build a Bayesian- variability in the input variables on their respective model
based algorithm that outperforms others on the analysis functions and on model output (Mangado et al. 2016).
of this specific system (because the model’s priors are Estimating (and reporting) uncertainty propagation and
more correctly defined by the team’s existing knowledge) its implications is an important consideration in assessing
but produces less robust/accurate predictions than other model calibration and interpretation (Vanlier et al. 2012).
algorithms when it is applied to related systems. Secondly, Methods to estimate uncertainty propagation continue to
detailed prior knowledge of a system limits what new be developed and applied (Ades & Lu 2003, Welton & Ades
knowledge can be discovered. The more that is understood 2005, Dubois 2010, Moseley 2013, Mangado et al. 2016).
about the system ahead of time, the better the model will In his discussion of error propagation in metabolomics
perform. However, the model will be making predictions studies, Moseley notes that both derived and propagated
in a shrinking space where there is less new knowledge uncertainty should be reported along with the results
to be discovered. In reasonably well understood systems, (Moseley 2013).
these latter models may have most utility in building our Measurement errors, as they apply to the relationship
confidence that what we believe to be true may indeed between a measured variable and its covariate, are additive
be true. In systems that are inadequately known, the new (Eckert et al. 1997). Integrative analyses across workflows in
knowledge space can be large and the predictions noisy; multiscale modeling, as may occur when combining data
the extent to which something is now believed to be from DNA sequence, RNA sequence/abundance and/or
true may require careful evaluation. Overall, the primary PPI studies, include many relationships between the
advantages of modeling include the ability to integrate measured variables (such as mRNA and protein expression
significant amounts of knowledge, to help researchers to levels) and covariates (such as a clinical outcome or changes
understand confounding events seen in the data and to in phenotype). Such analyses may be prone to error
answer questions of combinatorial complexity for which propagation and to error additivity or even amplification.
experimentation within the wet laboratory is prohibitive. For example, agglomerative techniques (such as some
hierarchical clustering), growing decision trees (such as
some random forest methods) or the network propagation
Error, performance and validation
algorithms that have begun to attract increased attention
Some workflows may include the output of one algorithm (Cowen et al. 2017) may be sensitive to error propagation.
as a means to guide parameter estimation for another. Once an error (node-edge connection) is made during
For example, in building a gene regulatory network from the graph build, it may remain and affect the accuracy
expression data, an investigator could take the output of subsequent local connections and of the overall model
predictions from a tool that predicts a TF and its targets solution. A build error that remains can lead to a model
as a means to define the priors for a Bayesian network solution that reaches convergence and appears ‘globally
modeling analysis of how these molecules are related correct’ but contains features that are ‘locally wrong’. The
in the data from a gene expression study. Intuitively, challenge here is that it is the local connections that are
even if the TF output is statistically noisy, it might be used to guide individual wet laboratory experiments.
expected to outperform a model with uninformed priors Studies that apply bioinformatic/biostatistic tools
where equal probabilities are assigned to each outcome. to solve problems in large data spaces are likely to be at
Nonetheless, some of the predictions will be wrong and greatest risk of experiencing the various types of errors
represent errors in the prior that may be worse than described above. The ‘hairball’ models often produced
uninformative; these types of errors will be propagated are rarely robustly tested for local error, especially when
from the output of one tool to the output of the next. the global model fit provides an apparently miniscule
Since the variables and their relationships (as captured in P value. For example, independent datasets showing
their priors) were thought to be intuitively correct, if these the same topologies are often not shown, frequently
incorrect variables persist as key features of the Bayesian because the data are not available to do independent
model solution, they could create the trap of self-fulfilling validation. The internal topology of individual cliques is
prophecy (Clarke et al. 2008). Predictions from one tool rarely tested, even using a simple n-fold cross-validation.
will also be associated with a level of error (variability), Global solutions are also rarely tested by an analogous
and this type of error will also propagate when the outputs n-fold cross-validation, such as removing entire cliques
are used as input variables for another tool in a workflow. at random. Since the overall topology of the solution

via free access
is likely to be influenced by the relationships among approach is to knockdown a target gene in cells where
discrete discovered features, without testing the effects of it is overexpressed, overexpress the gene in cells where
removing features on the remaining structures, there are its expression is low and then determine if the biological
few ways to determine topological robustness. While these function(s) is altered as the model predicts. Knockdown
‘hairballs’ will likely have met the statistical requirements is commonly achieved by an RNAi method such as siRNA
for global algorithmic convergence, how many of the or shRNA transfection. A gene may also be eliminated
local structures are correct, either internally within using CRISPR (Yin et al. 2019). How often a cell totally
each feature or externally within the global solution, is loses a gene or its expression likely requires careful
often left to human intuition and the risks therein (Clarke consideration. Total loss of a protein’s expression, as
et al. 2008). would usually occur with CRISPR, could alter a signaling
Appropriate assessments of model robustness and feature in a manner that does not occur when expression
validation are critical to the successful use of a systems is lowered but not eliminated in the phenotype(s) of
biology approach (Steyerberg et al. 2001). There are interest. While CRISPR is often preferred over RNAi, for
many tools to assess model performance and validation genes where downregulation rather than total loss is the
and a detailed technical discussion is beyond our scope. primary biological observation, RNAi may offer a more
Here, we use performance to denote assessments of the physiologically relevant validation approach. A similar
robustness or reproducibility of model predictions. For caveat applies to the use of cDNA transfection to produce
performance, biostatistical assessments of model fit overexpression of a gene. The level of overexpression
are usually incorporated into the workflow. Examples may be outside the range seen in the phenotype(s) under
of approaches to assess performance include use of a study, and so also produce changes in network features
receiver operating characteristic analysis and estimates that are not physiologically relevant. These types of
of the positive predictive value and negative predictive events could lead to misinterpretations of the validation
value. An internal n-fold cross-validation is commonly experiments. For example, the phenotype predicted by
used, particularly when data are limited (Waljee et al. the in silico model is not observed or further studies to
2014). A random portion of the data is withheld at each determine the effects of the manipulation of a gene on
interaction as a ‘validation set’, and the remaining data signaling identifies new relationships that are signaling
are used as a ‘training set’ for running the model. Multiple artifacts from a physiological relevance perspective.
iterations are run and the performance for each iteration As an example of a biological validation strategy,
is compared to assess the overall model performance. consider a prediction by an in silico model that an
A model can be tuned by adjusting its parameters until antiestrogen should induce autophagy through altering
the predictions from the training and internal validation expression of BECN1 in ER+ breast cancer cells. One
sets become sufficiently comparable. Since this approach approach to mechanistic validation of this prediction
can lead to model overfitting, the most informative could be to apply the drug and its vehicle control to ER+
assessment of model performance is obtained from the and ER− cells (negative control), measure changes in
use of independent datasets not used in model building BECN1 and autophagy and then use a molecular approach
and any internal performance analyses. A robust model to study if BECN1 knockdown or overexpression altered
is expected to produce broadly similar predictions in the regulatory effects of the antiestrogen on autophagy.
all comparable data sets. For classification studies using An underappreciated challenge with these types of studies
human tumors, the use of independent datasets may is that the experimental validation may be frustrated by
also be the only tractable option for validating model a high proportion of intuitively rational, statistically
predictions. significant, but biologically incorrect in silico model
For models that are used to predict system function predictions (the wet lab validation experiments show the
in a biological context, mechanistic or functional predictions to be invalid).
validation of a prediction is almost always required.
Here, validation refers to experimental validation in
Modeling drug interactions
the form of appropriate wet laboratory studies. These
validation studies are often done in cell lines and/or Another area of significant potential for a systems approach
animal models and can include applying perturbations is the search for drug combinations for treating a specific
to the experimental system and then measuring whether cancer in the context of a multicomponent signaling
the changes predicted by the model occur. A common network within the cancer cells (Tang & Aittokallio 2014,

via free access
Ryall & Tan 2015). Effective combination therapy, which the signaling network model parameters and the functional
is a hallmark of current cancer treatment, requires an structure of the dose–response relationship (Fang et al. 2016).
adequate understanding of signal complexity. Developing This model comprises a Hill equation for signals arriving at
and evaluating drug combinations is difficult because each receptor, a generic enzymatic rate equation to describe
the complexity of the problem increases combinatorially the transmission of signals among connecting genes, and
with the number of constituent drugs proposed to address a logistic equation to represent the cumulative effect of
an integrated driver pathway of the cancer. When the genes implicated in the onset of the cell death machinery.
possibility of sequencing drugs at different times relative These statistical models generate a global drug sensitivity
to one another is added to the mix, complexity again index based on the joint dose–response characteristics. Only
increases dramatically. Progress has been made using a the few terms with large global-sensitivity indices, much
systems biology approach. For example, the joint effects like principal components, are kept and subject to further
of multidrug combinations can be evaluated based on the experimental validation. Recently, the experimental design
mechanisms of action of the drugs (Fitzgerald et al. 2006). required for such subsequent experimentation has also been
If the constituent drugs in a combination therapy exert worked out (Fang et al. 2016, Huang et al. 2018).
their effects through known mechanisms that feed into
common pathways, the joint effect of the combination
An example of computational
may be assessed by the ‘Loewe additivity’. If the drugs
modeling: KDDN
act non-exclusively on multiple targets, the effect may
be assessed by the ‘Bliss additivity’ (Baeder et al. 2016). Cancers are often characterized by dysregulation of
Knowledge of the biological system can be used for molecular signaling (Barabasi et al. 2011, Tyson et al. 2011,
experimental design and data analysis. Thus, drugs with Creixell et al. 2012). Significant rewiring of molecular
different mechanisms of action, as revealed by systems networks can drive key phenotypic transitions that can
biology modeling, may exhibit different shapes of their occur in both a tumor and its microenvironment (Califano
dose–response relationships. Such information can be 2011, Roy et al. 2011, Ideker & Krogan 2012). The impact of
augmented by experimental data on a single drug to a treatment can spread through the network and alter the
optimally design the experiments on the joint effect of activity of functionally relevant gene products (Roy et al.
the drug combinations. 2011, Creixell et al. 2012). Most molecular components
Because the complexity of the problem increases exert their functions through interactions with other
rapidly with the number of constituent drugs, even the molecular components (Li et al. 2008, Gong & Miller
development of systems-based methods for the design 2013). How cancer cells differ from each other in their
and analysis of three-drug combinations has been responses to environments or treatments is intrinsically
only recent (Fang et al. 2017). The case of three-drug context specific (Mitra et al. 2013) and identifying such
combinations is fundamentally more difficult than two- differences may represent a ‘wicked’ problem for the
drug combinations. Finding doses of the combination, research community (Rittle & Webber 1973, Courtney
number of combinations and replicates needed to detect 2001, Clarke et al. 2011). Changes in molecular
departures from additivity depend on the dose–response interdependencies across cancer phenotypes may reveal
shapes of each of the constituent drugs. Thus, different novel hub genes and pathways, which may be suitable
classes of drugs with different dose–response shapes must targets for drug development. Instead of asking ‘which
be treated as separate cases. We designed and analyzed genes are differentially expressed?’ the question here is
a combination study of three anticancer drugs (PD184, ‘which genes are differentially connected?’ (Hudson et al.
HA14-1 and CEP3891) that inhibit the H929 myeloma cell 2009). Studies on network-attacking events will shed new
line. The three-drug combinations study used the original light on whether network rewiring is a general principle
4D dose–response surface formed by the dose ranges of of cancer cell responses, as most molecular therapies
the three drugs (Fang et al. 2017). target proteins and their networks but not genes (Califano
Methods for screening large numbers of drug 2011). Novel hypotheses inferred from the rewired TFs
combinations are being developed to reduce the problem to and their distal enhancers or partners can be proposed
one that is more experimentally manageable by using the and examined (Creixell et al. 2012, Mitra et al. 2013).
experimental data from dose–response studies of single drugs While multiscale omics data and the prior knowledge
and from a few combinations along with a systems analysis that provide insight into complex interactions are
of pathway/network information to obtain an estimate of increasingly available, models and analysis methods to

via free access
functionally integrate this information are still sorely cells adapt to the stresses of endocrine-based therapies.
needed. In particular, systematic efforts to characterize Our central hypothesis invokes a gene network that
selectively activated regulatory components and coordinately regulates those functions of a cell module
mechanisms must effectively distinguish significant that determine and execute the cell’s fate decision. Using
network rewiring from random background fluctuations. the KDDN tool, we identified three small topological
Most published biological network inferences were features and then overlaid these onto the canonical
obtained from molecular datasets acquired under a apoptosis pathway from KEGG (Fig. 2). The largest of
single condition, for which the statistically significant the three features reflected much of our prior knowledge,
network rewiring across different conditions is unknown despite not explicitly incorporating this knowledge into
or unreported (Mitra et al. 2013). The inability to the models (Zhang et al. 2009). Following the predictions
identify significant rewiring in biological networks of this topology, we uncovered some fundamentally new
represents a major limitation on the use of these results insights into molecular signaling; for example, the direct
for molecular signaling studies. The Knowledge-fused regulation of BCL2 by XBP1 and the requirement of NFκB
Differential Dependency Network (KDDN) method has for XBP1 signaling to regulate the prosurvival cell fate
been developed to infer significant rewiring of complex outcome in the context of antiestrogen treatment and
biological dependency networks, via sparse modeling and resistance (Clarke et al. 2011, Tyson et al. 2011, Hu et al.
data-knowledge integration (Zhang et al. 2009, 2011, Tian 2015). In applying KDDN to data from a rodent model, we
et al. 2013, 2014a,b, 2015). Specifically, KDDN formulates found that exposure to estrogens in utero induces a rewired
the inference of differential dependency networks (Zhang network in the mammary glands of the offspring that
et al. 2009, 2011, Tian et al. 2014a) that incorporate predicts for resistance to endocrine therapies in tumors
both conditional data and prior knowledge as a convex that arise in these glands during adulthood. Subsequent
optimization problem (Zhang & Wang 2010, Tian et al. studies showing that tumors in these mammary glands
2011) and uses an efficient learning algorithm to jointly are less responsive to tamoxifen (TAM) provided the first
infer the conserved biological network and significant direct demonstration of why many ER+ breast cancers may
rewiring across different conditions (Tian et al. 2014b, be pre-programmed to fail to respond to TAM treatment
2015). KDDN uses a minimax strategy to maximize the or respond and later recur (Hilakivi-Clarke et al. 2017).
benefit of prior knowledge while confining its negative We further pursued the functional evidence of the
impact under the worst-case scenario. Furthermore, hidden dependencies/crosstalk inferred by KDDN. For
KDDN matches the values of model parameters to the example, KDDN analysis of global protein expression data
expected false-positive rates on network edges at a from 122 TCGA ovarian cancer samples (selected based on
specified significance level and assesses edge-specific P homologous recombination deficiency, HRD, a phenotype
values on each of the differential connections. with distinct prognosis and response to therapies)
Tests on synthetic data have shown that KDDN resulted in a number of phenotype-dependent modules
produces biologically plausible results (Zhang et al. 2009, of co-expressed proteins. Several of the member proteins
2016, Herrington et al. 2018) and can reveal statistically in the modules were known to be involved in histone
significant rewiring in biological networks. The utility of modification. With the additional evidence of HRD
KDDN is evident following its application to a variety of status-dependent acetylation or deacetylation of histone
real gene and protein expression datasets including yeast proteins in the same samples, we were able, using patient
cell lines (Tian et al. 2014b), breast cancer (Tian et al. 2014b), population data, to support what has been shown in cells
ovarian cancer (Zhang et al. 2016) and medulloblastoma (Gong & Miller 2013, Tang et al. 2013) that histone protein
(Tian et al. 2014a). The method efficiently leverages data- acetylation affects the choice of DNA double-strand break
driven evidence and existing biological knowledge while repair pathways (between homologous recombination
remaining robust to false-positive edges in the prior and non-homologous end-joining) (Zhang et al. 2016).
knowledge. The network rewiring events identified by
KDDN reflect previous studies in the literature and provide
new mechanistic insight into the biological system(s) that An example of mathematical modeling:
extends beyond this earlier work. ER landscape
To study how gene networks may rewire during the
transition from normal to neoplastic breast cells, we Dynamic mathematical models track a system as it
have focused on understanding how ER+ breast cancer evolves in time. A key use of such models is to optimize

via free access
Figure 2
Differential dependency network focused on the KEGG apoptosis pathway (Kanehisa & Goto 2000). Recurrent breast cancers (uniquely featured by red
edges) showed the imbalance between apoptosis and survival with only one route into the cell through IL1B-induced inhibition of proapoptotic CASP3.
Non-recurrent breast cancer (uniquely featured by green edges) had a cascade of signaling pathways inside the cell that provides the balance between
apoptosis and survival. Copyright Kanehisa Laboratories. Reproduced with permission from KEGG.
therapeutic protocols. For example, instead of applying idea considered estrogen deprivation therapy (Chen et al.
a given drug or combination of drugs continuously for a 2014). ER+ cells were presumed to exist in three different
specified overall duration, the drug(s) can be applied for states: an estrogen-sensitive state (growth driven by
fixed durations with rest intervals in between. Alternatively, the estrogen receptor bound to estrogen), an estrogen-
several drugs can be applied in a repeating sequence for hypersensitive state (growth driven by membrane-
fixed durations. Optimizing the durations and dosing of associated estrogen receptor (ERM) bound to estrogen)
drugs is a combinatorial problem that is difficult to solve and an estrogen-independent state (growth driven by
experimentally, but relatively simple to solve via computer growth factor receptors (GFRs)). Transitions between the
simulation, assuming an accurate dynamical model states were governed by the estrogen level (high, low,
is available. Impressive results have been obtained in trace) in which the cells were grown. If cells were growing
prostate cancer and glioblastoma using two-compartment in a high (physiological) concentration of estrogen, most
models that simulate the temporal development of the cells would transition to the estrogen-sensitive state. If the
sensitive and drug-resistant populations of cancer cells estrogen concentration dropped to a low level, sensitive
(Jain et al. 2011, Leder et al. 2014, Morken et al. 2014). cells would begin to die, but some would transition to a
In the case of ER+ breast cancer and antiestrogens, hypersensitive state and continue growing.
the resistance character of the cells changes with time To model the transitions among these states, we
in response to the drugs. Hence, it is necessary first to developed a stochastic differential equation model of an
model the dynamics of development of drug resistance individual cell. States were characterized in the model
in individual cells, then to model the dynamics of a by ERM activity (high or low) and GFR activity (high
population of treated cells by linking the cellular scale or low). The model qualitatively matched observations
to the population scale, and finally to consider strategies in the literature concerning sensitivity transitions in
for optimizing drug therapy. A proof of concept of this breast cancer cells as the estrogen level was varied.

via free access
The fact that resistance to estrogen deprivation was the disease can be kept in check (similar to increasing
reversible if resistant cells were transferred back to duration of the recurrence-free survival period).
estrogen-rich medium for a sufficiently long time was This example provides a possible roadmap for how
also captured. Using techniques from statistical physics, modeling a molecular understanding of the response of
it is possible to visualize this model as a landscape upon a cancer cell to a drug can be transitioned to a tissue-
which the system makes spontaneous transitions among level model and used for therapy optimization. While
three low-lying basins (Fig. 3A), which represent the three the situation in patients is certainly more complicated
states of estrogen sensitivity. Random fluctuations in the than the model systems described here, the success
cells can occasionally cause transitions from one basin to of simple compartment models to guide therapy in
another, representing the natural heterogeneity seen in a simulated tumors provides hope that more complicated,
cell population. However, the system typically resides in molecularly-based, multiscale models will ultimately be
the lowest basin, as determined by the estrogen level. useful in guiding therapy in the clinic.
It is not efficient to simulate large numbers of these
‘model cells’ for long periods of time in order to compute
how a population would evolve in response to changes Interpreting models: caveat lector
in estrogen dose. To circumvent this problem, a cell-level
model was used to compute the transition probabilities ‘A little learning is a dang’rous thing; drink deep, or taste
among states as a function of estrogen concentration. not the Pierian spring: there shallow draughts intoxicate the
These probabilities were then used to create a population brain, and drinking largely sobers us again.’
model that efficiently tracked the number of cells in each
Alexander Pope (1688–1744)
state. A treatment regimen consisting of cycles of estrogen
deprivation followed by a drug holiday was considered, The qualitative and quantitative models that we have
and the deprivation and break durations were optimized described above produce results that can be difficult to
to drive the cancer cell population as low as possible. interpret correctly and usefully. Correct interpretation
Results are shown in Fig. 3B and C for the situation where is important, of course, because no one wants to spend
the cancer population is initially 1000 cells. For the time and precious experimental resources failing to
parameters in the model, the cancer cannot be eradicated. validate an incorrect understanding of the results of a
However, over a suitable range of therapeutic parameters, computational and/or mathematical analysis of a cellular
Figure 3
(A) The estrogen-response landscape for a particular level of estrogen stimulation. There are four basins of attraction for the cell state corresponding to
sensitive (ERM−/GFR−), hypersensitive (ERM+/GFR−) and independent (GFR+). (B) A sample intermittent treatment regimen (top panel) produces varying
proportion of cells in different states (second panel; cyan = sensitive, green = hypersensitive, blue = independent), a varying proliferation index of the
overall cell population (third panel; yellow indicates death and red indicates growth). The overall population level, starting from 1000 cells, is shown in the
bottom panel. (C) Plot of the average value of cell number <log10 N> over the interval t ϵ (2 × 104, 3 × 104) as a function of Ttreat and Tbreak. The white dot
indicates the case in (B). Any combination of Ttreat and Tbreak that puts the system within the log10N = 3 contour will suppress cancer growth. This figure is
adapted, with permission, from Fig. 3 and 6 of Chen et al. (2014).

via free access
control network. For example, gene set enrichment assessment, usually a P value, accompanies each model
is a powerful tool to explore high-dimensional data prediction, and understanding what these assessments
sets (Subramanian et al. 2005), but how its results are represent is important in evaluating the results. It is not
interpreted and used requires a thorough understanding unusual for a model to provide many predictions for which
of what the results do and do not imply. In this case, gene the P values are small (highly statistically significant), but
set enrichment analysis provides a static representation of how easily or appropriately these statistical estimates can
canonical signaling pathways, which are highly idealized be used to guide biological interpretation is not always
views of the most frequently observed events in a signal clear. Primarily, P values reflect how well each model
transduction pathway. However, once an event is identified output fits the data input, subject to the parameters and
and reported, it is more likely to be studied further and assumptions in the statistical model used. Thus, the use of
eventually to be considered as being canonical. Moreover, different statistical tools with the same input gene list and
these canonical signaling maps (examples include KEGG the same database may give different outputs, or the same
and Biocarta) are often assembled from a variety of sources outputs with different P values, because the parameters
(cell types, tissues, and species). Consequently, these and assumptions in each statistical model are different.
graphical representations may be relevant only in part to Also, if some pathways in the database are larger, better
the signaling processes under consideration in the specific annotated, or more fully (and correctly) understood than
cell context that a researcher is studying experimentally others, the P values associated with these pathways could
and trying to model computationally. be smaller (implying a more statistically significant fit)
Researchers are often limited to applying reductionist than less well represented pathways that may be a better
wet laboratory technologies to validate the predictions reflection of the underlying biological truth.
of models that attempt to explain some or all of the When the decision as to which is likely to be the
complexity in the biological system under investigation. correct solution is left entirely to intuition, it is not
Often, the cost in time and resources needed to validate surprising that the solution that best supports the
experimentally the predictions of multiscale models current hypothesis, or that is most easily explained by
can be prohibitive, making the ability to select among the operator’s existing knowledge, is often selected over
multiple solutions a necessity. Most algorithms, given other statistically significant outputs that are not easily
input variables in the correct format, will produce understood or may even refute the hypothesis. In such
outputs/predictions, but these often include false cases, the investigator is likely to fall into the trap of
positives and false negatives that are not easily identified. self-fulfilling prophesy (Clarke et al. 2008). To be able to
While model outputs are usually associated with interpret model outputs appropriately, it is often critical
probability estimations, the results of any significance to understand both what the data represent and some of
tests generally provide an evaluation only of how well the the basic principles of how the model works. This prior
model fits the available data. This statistical evaluation is knowledge is particularly important when the correct
not necessarily an estimate of how well the predictions interpretation is counterintuitive or inconsistent with the
reflect biological truth. Moreover, when sorting through hopes or expectations of the study designers.
multiple apparently statistically significant predictions, Cell context, by which we mean the unique patterns
an investigator can be left relying on subjective intuition, of genes, proteins, and metabolites that are expressed in
perhaps guided by an incomplete, inadequate, or incorrect a cell and that interact to influence the physiology of
understanding of the system. Since model predictions that cell (Clarke & Brünner 1996), is one of the central
should generally be consistent with the experimental data determinants of how signaling and function are related in
and/or the (sometimes) limited knowledge of the system biological systems. Context is clearly related to the cell/
currently available, the trap of self-fulfilling prophesy tissue type, local microenvironment, status of the host,
becomes almost unavoidable (Clarke et al. 2008). and other external and internal influences. Some aspects
A gene set enrichment algorithm may produce several of cellular context may be highly conserved, as can be
predicted pathways and functions associated with a single seen from the DNA sequence of some genes through to
set of differentially expressed genes. While some genes can the basic signaling topology of some highly conserved
certainly participate in more than one pathway or regulate functions. Nonetheless, there can also be substantial
more than one cellular function, the investigator must diversity, even within closely related species, tissues, or
determine which output(s) (which pathway, module, or cells. For cancer research, the differences between the
function) is most likely to represent the truth. A statistical normal and neoplastic state in the same tissue or cell type

via free access
is where we look most often for molecular targets that potentially powerful approach to extracting new features
can be diagnostic, prognostic, and/or therapeutic. Here, from high-dimensional data spaces (Hosny et al. 2018, Zou
even small changes in cell context can have substantial et al. 2019). It is likely that deep-learning approaches will
implications for the ability to address a specific hypothesis. be more commonly applied in the near future to guide
Despite the often fundamental importance of cellular knowledge discovery within the framework of cancer
context, it is frequently ignored. systems biology.
Some modeling approaches are of limited utility Another area that has attracted renewed interest is
either scientifically or clinically and need to be the heterogeneity arising from the presence of multiple
re-addressed. The complex hairball models often cell types and the consequent complexity of interactions
generated by some computational tools may (or may within tumor microenvironments. For molecular
not) contain valid insights into regulatory biology. signaling, a key issue in this context is whether the events
However, their complexity can be so high and the noise identified as being associated with a biological outcome
sufficiently extensive and the errors undefined that these or phenotype are intrinsic or extrinsic to the cancer cells
models cannot be tested meaningfully or interpreted and/or other cells within the microenvironment. While
reliably. Employment of a razor to shave away those most therapeutic interventions attempt to induce cell
components that do not add to the utility, robustness, death programs that are executed within the cancer cell
or accuracy of a model may be desirable, but only if its (intrinsic), many of the signals that initiate this intrinsic
application is tractable and it is evident what ‘whiskers’ activity are generated by activities originating in stromal
can be removed without a significant loss of predictive or immune cells (extrinsic). Single-cell RNAseq can address
power. Proactively incorporating feature elimination some of these issues, but this is not always feasible and the
tools during modeling (for example, applying a support technology has its own limitations (Cheng et al. 2014a,
vector machine with recursive feature elimination for Saliba et al. 2014). Moreover, many public omics datasets
classification; Guyon et al. 2003) may help to address this are populated with data representing averaged signals from
concern by attempting to arrive at the smallest model multiple cell types, as might be expected from a study that
that meets predetermined requirements of convergence used tumor biopsies as the primary material. Some form
and statistical significance. Nonetheless, the need for of data deconvolution is then required. Tools to achieve
human intuition to interpret outcomes remains central deconvolution continue to emerge but for many of these
to many study designs, and, consequently, the risk of datasets the tools must be effective when applied in an
falling into the trap of self-fulfilling prophesy must be unsupervised manner because data that could supervise
carefully avoided (Clarke et al. 2008). the analysis is often absent. Tools that can accurately and
robustly perform unsupervised data deconvolution are
likely to become more widely used in the near future.
The application of systems biology approaches to
Future directions
critical questions in endocrine-related and other cancers
The properties of high-dimensional data spaces, and may provide new insights into cancer biology and lead
the challenges and opportunities these provide (Clarke to new treatments. Many signaling networks and the
et al. 2008), remain central to the performance of many biological processes that they regulate often prove to be too
computational modeling approaches and bioinformatic complex for biostatistics, bioinformatics or mathematical
tools and workflows. Tools designed to manage these biology alone to unravel. However, the integrated use of
properties explicitly, such as support vector machines, these approaches can support the building of predictive
and workflows to address high dimensionality, such as multiscale models from a systems perspective. The
including dimensionality reduction as a preprocessing virtuous cycle of in silico model prediction, validation in
step in data analysis, are likely to remain in use. New and appropriate wet laboratory experiments, with validated
more powerful tools and workflows are likely to continue results feeding back to improve model predictions, can
to emerge, increasing the power and accuracy of predictive then drive new discovery of complex systems in a manner
models, the quality and accuracy of data interpretation, that often outstrips intuitive reasoning. In those cancers
and the utility of the new knowledge gained. Deep where hormone and growth factor receptors and their
learning, a subset of approaches within the broader signaling play a major role, systems approaches may offer
field of machine learning that generally applies neural the best means to address the complexity and dynamic
network-based modeling, has gained recent attention as a nature of signaling and how it responds to therapeutic

via free access
interventions that affect the cancer cells and their Brazma A, Krestyaninova M & Sarkans U 2006 Standards for systems
biology. Nature Reviews: Genetics 7 593–605. (https://doi.org/10.1038/
interactions within their microenvironments.
nrg1922)
Buettner F, Natarajan KN, Casale FP, Proserpio V, Scialdone A, Theis FJ,
Teichmann SA, Marioni JC & Stegle O 2015 Computational analysis
of cell-to-cell heterogeneity in single-cell RNA-sequencing data
Declaration of interest
reveals hidden subpopulations of cells. Nature Biotechnology 33
Robert Clarke is an Associate Editor of Endocrine-Related Cancer. Robert
155–160. (https://doi.org/10.1038/nbt.3102)
Clarke was not involved in the review or editorial process for this paper, on
Califano A 2011 Rewiring makes the difference. Molecular Systems Biology
which he is listed as an author. The other authors have nothing to disclose.
7 463. (https://doi.org/10.1038/msb.2010.117)
Califano A & Alvarez MJ 2017 The recurrent architecture of tumour
initiation, progression and drug sensitivity. Nature Reviews: Cancer 17
116–130. (https://doi.org/10.1038/nrc.2016.124)
Funding
Chen L, Chan TH, Choyke PL, Hillman EM, Chi CY, Bhujwalla ZM,
This work was supported in part by Public Health Service Awards U01-
Wang G, Wang SS, Szabo Z & Wang Y 2011 CAM-CM: a signal
CA184902, U54-CA149147, and DoD-BCRP-CA171885 to R Clarke, the
deconvolution tool for in vivo dynamic contrast-enhanced imaging
Georgetown-Lombardi Comprehensive Cancer Center grant (P30-
of complex tissues. Bioinformatics 27 2607–2609. (https://doi.
CA51008-19), R01-CA164717 to M Tan, and R01-CA201092 to W T Baumann.
org/10.1093/bioinformatics/btr436)
Chen C, Baumann WT, Clarke R & Tyson JJ 2013 Modeling the estrogen
receptor to growth factor receptor signaling switch in human breast
References cancer cells. FEBS Letters 587 3327–3334. (https://doi.org/10.1016/j.
febslet.2013.08.022)
Ades AE & Lu G 2003 Correlations between parameters in risk models: Chen C, Baumann WT, Xing J, Xu L, Clarke R & Tyson JJ 2014
estimation and propagation of uncertainty by Markov chain Monte Mathematical models of the transitions between endocrine therapy
Carlo. Risk Analysis 23 1165–1172. (https://doi. responsive and resistant states in breast cancer. Journal of the Royal
org/10.1111/j.0272-4332.2003.00386.x) Society, Interface 11 20140206. (https://doi.org/10.1098/
Aldridge BB, Burke JM, Lauffenburger DA & Sorger PK 2006 rsif.2014.0206)
Physicochemical modelling of cell signalling pathways. Nature Cell Cheng WY, Ou Yang TH & Anastassiou D 2013 Development of a
Biology 8 1195–1203. (https://doi.org/10.1038/ncb1497) prognostic model for breast cancer survival in an open challenge
Altrock PM, Liu LL & Michor F 2015 The mathematics of cancer: environment. Science Translational Medicine 5 181ra50. (https://doi.
integrating quantitative models. Nature Reviews: Cancer 15 730–745. org/10.1126/scitranslmed.3005974)
(https://doi.org/10.1038/nrc4029) Cheng F, Jia P, Wang Q, Lin CC, Li WH & Zhao Z 2014a Studying
Anafi RC, Francey LJ, Hogenesch JB & Kim J 2017 CYCLOPS reveals tumorigenesis through network evolution and somatic mutational
human transcriptional rhythms in health and disease. PNAS 114 perturbations in the cancer interactome. Molecular Biology and
5312–5317. (https://doi.org/10.1073/pnas.1619320114) Evolution 31 2156–2169. (https://doi.org/10.1093/molbev/msu167)
Anderson AR & Quaranta V 2008 Integrative mathematical oncology. Cheng F, Jia P, Wang Q & Zhao Z 2014b Quantitative network
Nature Reviews: Cancer 8 227–234. (https://doi.org/10.1038/nrc2329) mapping of the human kinome interactome reveals new clues for
Baeder DY, Yu G, Hoze N, Rolff J & Regoes RR 2016 Antimicrobial rational kinase inhibitor discovery and individualized cancer
combinations: Bliss independence and Loewe additivity derived from therapy. Oncotarget 5 3697–3710. (https://doi.org/10.18632/
mechanistic multi-hit models. Philosophical Transactions of the Royal oncotarget.1984)
Society of London: Series B, Biological Sciences 371 20150294. (https:// Cirillo E, Parnell LD & Evelo CT 2017 A review of pathway-based
doi.org/10.1098/rstb.2015.0294) analysis tools that visualize genetic variants. Frontiers in Genetics 8
Barabasi AL, Gulbahce N & Loscalzo J 2011 Network medicine: a 174. (https://doi.org/10.3389/fgene.2017.00174)
network-based approach to human disease. Nature Reviews: Genetics Clarke R & Brünner N 1996 Acquired estrogen independence and
12 56–68. (https://doi.org/10.1038/nrg2918) antiestrogen resistance in breast cancer: estrogen receptor-driven
Barberis M & Verbruggen P 2017 Quantitative systems biology to phenotypes? Trends in Endocrinology and Metabolism 7 291–301.
decipher design principles of a dynamic cell cycle network: the (https://doi.org/10.1016/S1043-2760(96)00127-0)
‘Maximum Allowable mammalian Trade-Off-Weight’ (MAmTOW). Clarke R, Ressom HW, Wang A, Xuan J, Liu MC, Gehan EA & Wang Y
NPJ Systems Biology and Applications 3 26. (https://doi.org/10.1038/ 2008 The properties of very high dimensional data spaces:
s41540-017-0028-x) implications for exploring gene and protein expression data. Nature
Barik D, Ball DA, Peccoud J & Tyson JJ 2016 A stochastic model of the Reviews: Cancer 8 37–49. (https://doi.org/10.1038/nrc2294)
yeast cell cycle reveals roles for feedback regulation in limiting Clarke R, Shajahan AN, Wang Y, Tyson JJ, Riggins RB, Weiner LM,
cellular variability. PLoS Computational Biology 12 e1005230. (https:// Bauman WT, Xuan J, Zhang B, Facey C, et al. 2011 Endoplasmic
doi.org/10.1371/journal.pcbi.1005230) reticulum stress, the unfolded protein response, and gene network
Bedard PL, Mook S, Piccart-Gebhart MJ, Rutgers ET, Van’t Veer LJ & modeling in antiestrogen resistant breast cancer. Hormone Molecular
Cardoso F 2009 MammaPrint 70-gene profile quantifies the Biology and Clinical Investigation 5 35–44. (https://doi.org/10.1515/
likelihood of recurrence for early breast cancer. Expert Opinion on hmbci.2010.073)
Medical Diagnostics 3 193–205. (https://doi. Clarke R, Cook KL, Hu R, Facey CO, Tavassoly I, Schwartz JL,
org/10.1517/17530050902751618) Baumann WT, Tyson JJ, Xuan J, Wang Y, et al. 2012 Endoplasmic
Begley CG 2013 Six red flags for suspect work. Nature 497 433–434. reticulum stress, the unfolded protein response, autophagy, and the
(https://doi.org/10.1038/497433a) integrated regulation of breast cancer cell fate. Cancer Research 72
Berger JO 2010 Statistical Decision Theory and Bayesian Analysis. New 1321–1331. (https://doi.org/10.1158/0008-5472.CAN-11-3213)
York, NY, USA: Springer-Verlag. Courtney JF 2001 Decision making and knowledge management in
Bhalla US & Iyengar R 1999 Emergent properties of networks of inquiring organizations: towards a new decision-making paradigm
biological signaling pathways. Science 283 381–387. (https://doi. for DSS. Decision Support Systems 31 17–38. (https://doi.org/10.1016/
org/10.1126/science.283.5400.381) S0167-9236(00)00117-2)

via free access
Cowen L, Ideker T, Raphael BJ & Sharan R 2017 Network propagation: a GTex Consortium 2015 Human genomics. The Genotype-Tissue
universal amplifier of genetic associations. Nature Reviews: Genetics Expression (GTEx) pilot analysis: multitissue gene regulation in
18 551–562. (https://doi.org/10.1038/nrg.2017.38) humans. Science 348 648–660. (https://doi.org/10.1126/
Creixell P, Schoof EM, Erler JT & Linding R 2012 Navigating cancer science.1262110)
network attractors for tumor-specific therapy. Nature Biotechnology 30 Guyon J, Weston J, Barnhill MD & Vapnik V 2003 Gene selection for
842–848. (https://doi.org/10.1038/nbt.2345) cancer classification using support vector machines. Machine Learning
Creixell P, Reimand J, Haider S, Wu G, Shibata T, Vazquez M, 46 389–422.
Mustonen V, Gonzalez-Perez A, Pearson J, Sander C, et al. 2015 Hart Y, Sheftel H, Hausser J, Szekely P, Ben-Moshe NB, Korem Y,
Pathway and network analysis of cancer genomes. Nature Methods 12 Tendler A, Mayo AE & Alon U 2015 Inferring biological tasks using
615–621. (https://doi.org/10.1038/nmeth.3440) Pareto analysis of high-dimensional data. Nature Methods 12
Deisboeck TS, Wang Z, Macklin P & Cristini V 2011 Multiscale cancer 233–235. (https://doi.org/10.1038/nmeth.3254)
modeling. Annual Review of Biomedical Engineering 13 127–155. Hatzis C, Bedard PL, Birkbak NJ, Beck AH, Aerts HJ, Stem DF, Shi L,
(https://doi.org/10.1146/annurev-bioeng-071910-124729) Clarke R, Quackenbush J & Haibe-Kains B 2014 Enhancing
Di Ventura B, Lemerle C, Michalodimitrakis K & Serrano L 2006 From in reproducibility in cancer drug screening: how do we move forward?
vivo to in silico biology and back. Nature 443 527–533. (https://doi. Cancer Research 74 4016–4023. (https://doi.org/10.1158/0008-5472.
org/10.1038/nature05127) CAN-14-0725)
Dimitrova N, Nagaraj AB, Razi A, Singh S, Kamalakaran S, Banerjee N, Herrington DM, Mao C, Parker SJ, Fu Z, Yu G, Chen L, Venkatraman V,
Joseph P, Mankovich A, Mittal P, DiFeo A, et al. 2017 InFlo: a novel Fu Y, Wang Y, Howard TD, et al. 2018 Proteomic architecture of
systems biology framework identifies cAMP-CREB1 axis as a key human coronary and aortic atherosclerosis. Circulation 137
modulator of platinum resistance in ovarian cancer. Oncogene 36 2741–2756. (https://doi.org/10.1161/CIRCULATIONAHA.118.034365)
2472–2482. (https://doi.org/10.1038/onc.2016.398) Hilakivi-Clarke LA, Wärri A, Bouker KB, Zhang X, Cook KL, Jin L,
Dubois D 2010 Representation, propagation, and decision issues in risk Zwart A, Nguyen N, Hu R, Cruz MI, et al. 2017 Effects of in utero
analysis under incomplete probabilistic information. Risk Analysis 30 exposure to ethinyl estradiol on tamoxifen resistance and breast
361–368. (https://doi.org/10.1111/j.1539-6924.2010.01359.x) cancer recurerence in a preclinical model. Journal of the National
Dutta A, Le Magnen C, Mitrofanova A, Ouyang X, Califano A & Abate- Cancer Institute 109 djw188. (https://doi.org/10.1093/jnci/djw188)
Shen C 2016 Identification of an NKX3.1-G9a-UTY transcriptional Hoadley KA, Yau C, Wolf DM, Cherniack AD, Tamborero D, Ng S,
regulatory network that controls prostate differentiation. Science 352 Leiserson MDM, Niu B, McLellan MD, Uzunangelov V, et al. 2014
1576–1580. (https://doi.org/10.1126/science.aad9512) Multiplatform analysis of 12 cancer types reveals molecular
Eckert RS, Carroll RJ & Wang N 1997 Transformations to additivity in classification within and across tissues of origin. Cell 158 929–944.
measurement error models. Biometrics 53 262–272. (https://doi. (https://doi.org/10.1016/j.cell.2014.06.049)
org/10.2307/2533112) Hofree M, Shen JP, Carter H, Gross A & Ideker T 2013 Network-based
Enriquez-Navas PM, Wojtkowiak JW & Gatenby RA 2015 Application of stratification of tumor mutations. Nature Methods 10 1108–1115.
evolutionary principles to cancer therapy. Cancer Research 75 (https://doi.org/10.1038/nmeth.2651)
4675–4680. (https://doi.org/10.1158/0008-5472.CAN-15-1337) Hosny A, Parmar C, Coroller TP, Grossmann P, Zeleznik R, Kumar A,
Fang HB, Huang H, Clarke R & Tan M 2016 Predicting multi-drug Bussink J, Gillies RJ, Mak RH & Aerts HJWL 2018 Deep learning for
inhibition interactions based on signaling networks and single drug lung cancer prognostication: a retrospective multi-cohort radiomics
dose-response information. Journal of Computational Systems Biology study. PLoS Medicine 15 e1002711. (https://doi.org/10.1371/journal.
2 101. pmed.1002711)
Fang HB, Chen X, Pei XY, Grant S & Tan M 2017 Experimental design Hu R, Warri A, Jin L, Zwart A, Riggins R & Clarke R 2015 NFkappaB
and statistical analysis for three-drug combination studies. Statistical signaling is required for XBP1 (U and S) mediated effects on
Methods in Medical Research 26 1261–1280. (https://doi. antiestrogen responsiveness and cell fate decisions in breast cancer.
org/10.1177/0962280215574320) Molecular and Cellular Biology 35 390. (https://doi.org/10.1128/
Fitzgerald JB, Schoeberl B, Nielsen UB & Sorger PK 2006 Systems biology MCB.00847-14)
and combination therapy in the quest for clinical efficacy. Nature Huang HZ, Fang HB & Tan MT 2018 Experimental designs for multidrug
Chemical Biology 2 458–466. (https://doi.org/10.1038/nchembio817) combination studies using signaling networks. Biometrics 74
Galea MH, Blamey RW, Elston CE & Ellis IO 1992 The Nottingham 538–547. (https://doi.org/10.1111/biom.12777)
prognostic index in primary breast cancer. Breast Cancer Research and Hudson NJ, Reverter A & Dalrymple BP 2009 A differential wiring
Treatment 22 207–219. (https://doi.org/10.1007/BF01840834) analysis of expression data correctly identifies the gene containing
Gallaher J, Babu A, Plevritis S & Anderson ARA 2014 Bridging population the causal mutation. PLoS Computational Biology 5 e1000382.
and tissue scale tumor dynamics: a new paradigm for understanding (https://doi.org/10.1371/journal.pcbi.1000382)
differences in tumor growth and metastatic disease. Cancer Research Ideker T & Krogan NJ 2012 Differential network biology. Molecular
74 426–435. (https://doi.org/10.1158/0008-5472.CAN-13-0759) Systems Biology 8 565. (https://doi.org/10.1038/msb.2011.99)
Gehlenborg N, O'Donoghue SI, Baliga NS, Goesmann A, Hibbs MA, Imamov O, Shim GJ, Warner M & Gustafsson JA 2005 Estrogen receptor
Kitano H, Kohlbacher O, Neuweger H, Schneider R, Tenenbaum D, beta in health and disease. Biology of Reproduction 73 866–871.
et al. 2010 Visualization of omics data for systems biology. Nature (https://doi.org/10.1095/biolreprod.105.043497)
Methods 7 S56–S68. (https://doi.org/10.1038/nmeth.1436) Jain HV, Clinton SK, Bhinder A & Friedman A 2011 Mathematical
Gho YS & Lee C 2017 Emergent properties of extracellular vesicles: a modeling of prostate cancer progression in response to androgen
holistic approach to decode the complexity of intercellular ablation therapy. PNAS 108 19701–19706. (https://doi.org/10.1073/
communication networks. Molecular BioSystems 13 1291–1296. pnas.1115750108)
(https://doi.org/10.1039/c7mb00146k) Janes KA, Chandran PL, Ford RM, Lazzara MJ, Papin JA, Peirce SM,
Ghosh S, Matsuoka Y, Asai Y, Hsin KY & Kitano H 2011 Software for Saucerman JJ & Lauffenburger DA 2017 An engineering design
systems biology: from tools to integrated platforms. Nature Reviews: approach to systems biology. Integrative Biology 9 574–583. (https://
Genetics 12 821–832. (https://doi.org/10.1038/nrg3096) doi.org/10.1039/C7IB00014F)
Gong F & Miller KM 2013 Mammalian DNA repair: HATs and HDACs Ji Z, Yan K, Li W, Hu H & Zhu X 2017 Mathematical and computational
make their mark through histone acetylation. Mutation Research 750 modeling in complex biological systems. BioMed Research
23–30. (https://doi.org/10.1016/j.mrfmmm.2013.07.002) International 2017 5958321. (https://doi.org/10.1155/2017/5958321)

via free access
Jinawath N, Bunbanjerdsuk S, Chayanupatkul M, Ngamphaiboon N, modeling: microscopic scale and multiscale approaches. Seminars in
Asavapanumas N, Svasti J & Charoensawan V 2016 Bridging the gap Cancer Biology 30 60–69. (https://doi.org/10.1016/j.
between clinicians and systems biologists: from network biology to semcancer.2014.03.003)
translational biomedical research. Journal of Translational Medicine 14 McKenna MT, Weis JA, Barnes SL, Tyson DR, Miga MI, Quaranta V &
324. (https://doi.org/10.1186/s12967-016-1078-3) Yankeelov TE 2017 A predictive mathematical modeling approach
Junttila MR & de Sauvage FJ 2013 Influence of tumour micro- for the study of doxorubicin treatment in triple negative breast
environment heterogeneity on therapeutic response. Nature 501 cancer. Scientific Reports 7 5725. (https://doi.org/10.1038/s41598-017-
346–354. (https://doi.org/10.1038/nature12626) 05902-z)
Kanehisa M & Goto S 2000 KEGG: Kyoto Encyclopedia of genes and Meacham CE & Morrison SJ 2013 Tumour heterogeneity and cancer cell
genomes. Nucleic Acids Research 28 27–30. (https://doi.org/10.1093/ plasticity. Nature 501 328–337. (https://doi.org/10.1038/
nar/28.1.27) nature12624)
Kanter I & Kalisky T 2015 Single cell transcriptomics: methods and Miryala SK, Anbarasu A & Ramaiah S 2018 Discerning molecular
applications. Frontiers in Oncology 5 53. (https://doi.org/10.3389/ interactions: a comprehensive review on biomolecular interaction
fonc.2015.00053) databases and network analysis tools. Gene 642 84–94. (https://doi.
Keenan AB, Jenkins SL, Jagodnik KM, Koplev S, He E, Torre D, Wang Z, org/10.1016/j.gene.2017.11.028)
Dohlman AB, Silverstein MC, Lachmann A, et al. 2018 The library of Mitra K, Carvunis AR, Ramesh SK & Ideker T 2013 Integrative
integrated network-based cellular signatures NIH program: system- approaches for finding modular structure in biological networks.
level cataloging of human cells response to perturbations. Cell Nature Reviews: Genetics 14 719–732. (https://doi.org/10.1038/
Systems 6 13–24. (https://doi.org/10.1016/j.cels.2017.11.001) nrg3552)
Kimura A, Celani A, Nagao H, Stasevich T & Nakamura K 2015 Mobley A, Linder SK, Braeuer R, Ellis LM & Zwelling L 2013 A survey on
Estimating cellular parameters through optimization procedures: data reproducibility in cancer research provides insights into our
elementary principles and applications. Frontiers in Physiology 6 60. limited ability to translate findings from the laboratory to the clinic.
(https://doi.org/10.3389/fphys.2015.00060) PLoS One 8 e63221. (https://doi.org/10.1371/journal.pone.0063221)
Lampinen J & Vehtari A 2001 Bayesian approach for neural networks – Molinelli EJ, Korkut A, Wang W, Miller ML, Gauthier NP, Jing X,
review and case studies. Neural Networks 14 257–274. (https://doi. Kaushik P, He Q, Mills G, Solit DB, et al. 2013 Perturbation biology:
org/10.1016/S0893-6080(00)00098-8) inferring signaling networks in cellular systems. PLoS Computational
Lavrik IN & Zhivotovsky B 2014 Systems biology: a way to make Biology 9 e1003290. (https://doi.org/10.1371/journal.pcbi.1003290)
complex problems more understandable. Cell Death and Disease 5 Morken JD, Packer A, Everett RA, Nagy JD & Kuang Y 2014 Mechanisms
e1256. (https://doi.org/10.1038/cddis.2014.195) of resistance to intermittent androgen deprivation in patients with
Le NN 2007 The long journey to a systems biology of neuronal function. prostate cancer identified by a novel computational method. Cancer
BMC Systems Biology 1 28. (https://doi.org/10.1186/1752-0509-1-28) Research 74 3673–3683. (https://doi.org/10.1158/0008-5472.CAN-13-
Leder K, Pitter K, Laplant Q, Hambardzumyan D, Ross BD, Chan TA, 3162)
Holland EC & Michor F 2014 Mathematical modeling of PDGF- Moseley HN 2013 Error analysis and propagation in metabolomics data
driven glioblastoma reveals optimized radiation dosing schedules. analysis. Computational and Structural Biotechnology Journal 4
Cell 156 603–616. (https://doi.org/10.1016/j.cell.2013.12.029) e201301006. (https://doi.org/10.5936/csbj.201301006)
Ledzewicz U & Schaettler H 2016 Optimizing chemotherapeutic anti- Nam S 2017 Databases and tools for constructing signal transduction
cancer treatment and the tumor microenvironment: an analysis of networks in cancer. BMB Reports 50 12–19. (https://doi.org/10.5483/
mathematical models. Advances in Experimental Medicine and Biology BMBRep.2017.50.1.135)
936 209–223. (https://doi.org/10.1007/978-3-319-42023-3_11) Newman AM, Liu CL, Green MR, Gentles AJ, Feng W, Xu Y, Hoang CD,
Leiserson MD, Vandin F, Wu HT, Dobson JR, Eldridge JV, Thomas JL, Diehn M & Alizadeh AA 2015 Robust enumeration of cell subsets
Papoutsaki A, Kim Y, Niu B, McLellan M, et al. 2015 Pan-cancer from tissue expression profiles. Nature Methods 12 453–457. (https://
network analysis identifies combinations of rare somatic mutations doi.org/10.1038/nmeth.3337)
across pathways and protein complexes. Nature Genetics 47 106–114. Okasha S 2012 Emergence, hierarchy and top-down causation in
(https://doi.org/10.1038/ng.3168) evolutionary biology. Interface Focus 2 49–54. (https://doi.
Li H, Xuan J, Wang Y & Zhan M 2008 Inferring regulatory networks. org/10.1098/rsfs.2011.0046)
Frontiers in Bioscience 13 263–275. (https://doi.org/10.2741/2677) Parker JS, Mullins M, Cheang MC, Leung S, Voduc D, Vickery T,
Liepe J, Kirk P, Filippi S, Toni T, Barnes CP & Stumpf MP 2014 A Davies S, Fauron C, He X, Hu Z, et al. 2009 Supervised risk predictor
framework for parameter estimation and model selection from of breast cancer based on intrinsic subtypes. Journal of Clinical
experimental data in systems biology using approximate Bayesian Oncology 27 1160–1167. (https://doi.org/10.1200/JCO.2008.18.1370)
computation. Nature Protocols 9 439–456. (https://doi.org/10.1038/ Parmar JH, Cook KL, Shajahan-Haq AN, Clarke PA, Tavassoly I, Clarke R,
nprot.2014.025) Tyson JJ & Baumann WT 2013 Modelling the effect of GRP78 on
Mackay A, Weigelt B, Grigoriadis A, Kreike B, Natrajan R, A’Hern R, anti-oestrogen sensitivity and resistance in breast cancer. Interface
Tan DS, Dowsett M, Ashworth A & Reis-Filho JS 2011 Microarray- Focus 3 20130012. (https://doi.org/10.1098/rsfs.2013.0012)
based class discovery for molecular classification of breast cancer: Pavlopoulos GA, Paez-Espino D, Kyrpides NC & Iliopoulos I 2017
analysis of interobserver agreement. Journal of the National Cancer Empirical comparison of visualization tools for larger-scale network
Institute 103 662–673. (https://doi.org/10.1093/jnci/djr071) analysis. Advances in Bioinformatics 2017 1278932. (https://doi.
Mangado N, Piella G, Noailly J, Pons-Prats J & Ballester MÁ 2016 org/10.1155/2017/1278932)
Analysis of uncertainty and variability in finite element Peng H, Tan H, Zhao W, Jin G, Sharma S, Xing F, Watabe K & Zhou X
computational models for biomedical engineering: characterization 2016 Computational systems biology in cancer brain metastasis.
and propagation. Frontiers in Bioengineering and Biotechnology 4 85. Frontiers in Bioscience 8 169–186. (https://doi.org/10.2741/s456)
(https://doi.org/10.3389/fbioe.2016.00085) Picco N, Gatenby RA & Anderson ARA 2017 Stem cell plasticity and
Martelotto LG, Ng CK, Piscuoglio S, Weigelt B & Reis-Filho JS 2014 niche dynamics in cancer progression. IEEE Transactions on Bio-
Breast cancer intra-tumor heterogeneity. Breast Cancer Research 16 Medical Engineering 64 528–537. (https://doi.org/10.1109/
210. (https://doi.org/10.1186/bcr3658) TBME.2016.2607183)
Masoudi-Nejad A, Bidkhori G, Hosseini Ashtiani S, Najafi A, Ponten F, Schwenk JM, Asplund A & Edqvist PH 2011 The Human
Bozorgmehr JH & Wang E 2015 Cancer systems biology and Protein Atlas as a proteomic resource for biomarker discovery. Journal

via free access
of Internal Medicine 270 428–446. (https://doi. Tian Y, Zhang B, Shih IM & Wang Y 2011 Knowledge-guided differential
org/10.1111/j.1365-2796.2011.02427.x) dependency network learning for detecting structural changes in
Quaranta V, Rejniak KA, Gerlee P & Anderson AR 2008 Invasion emerges biological networks. In BCB '11: Proceedings of the 2nd ACM
from cancer cell adaptation to competitive microenvironments: Conference on Bioinformatics, Computational Biology and Biomedicine,
quantitative predictions from multiscale mathematical models. pp 254–263. New York, NY, USA: ACM. (https://doi.
Seminars in Cancer Biology 18 338–348. (https://doi.org/10.1016/j. org/10.1145/2147805.2147833)
semcancer.2008.03.018) Tian Y, Chen L, Zhang B, Zhang Z, Yu G, Clarke R, Xuan J, Shih IM &
Rittle HWJ & Webber MM 1973 Dilemmas in a general theory of Wang Y 2013 Genomic and network analysis to study the origin of
planning. Policy Sciences 4 155–169. (https://doi.org/10.1007/ ovarian cancer. Systems Biomedicine 1 55–64. (https://doi.
BF01405730) org/10.4161/sysb.25313)
Robinson JT, Thorvaldsdottir H, Wenger AM, Zehir A & Mesirov JP 2017 Tian Y, Wang SS, Zhang Z, Rodriguez OC, Petricoin E, III, Shih IeM,
Variant review with the integrative genomics viewer. Cancer Research Chan D, Avantaggiati M, Yu G, Ye S, et al. 2014a Integration of
77 e31–e34. (https://doi.org/10.1158/0008-5472.CAN-17-0337) network biology and imaging to study cancer phenotypes and
Roy S, Werner-Washburne M & Lane T 2011 A multiple network responses. IEEE/ACM Transactions on Computational Biology and
learning approach to capture system-wide condition-specific Bioinformatics 11 1009–1019. (https://doi.org/10.1109/
responses. Bioinformatics 27 1832–1838. (https://doi.org/10.1093/ TCBB.2014.2338304)
bioinformatics/btr270) Tian Y, Zhang B, Hoffman EP, Clarke R, Zhang Z, Shih IeM, Xuan J,
Ryall KA & Tan AC 2015 Systems biology approaches for advancing the Herrington DM & Wang Y 2014b Knowledge-fused differential
discovery of effective drug combinations. Journal of Cheminformatics dependency network models for detecting significant rewiring in
7 7. (https://doi.org/10.1186/s13321-015-0055-9) biological networks. BMC Systems Biology 8 87. (https://doi.
Saliba AE, Westermann AJ, Gorski SA & Vogel J 2014 Single-cell RNA- org/10.1186/s12918-014-0087-1)
seq: advances and future challenges. Nucleic Acids Research 42 Tian Y, Zhang B, Hoffman EP, Clarke R, Zhang Z, Shih IeM, Xuan J,
8845–8860. (https://doi.org/10.1093/nar/gku555) Herrington DM & Wang Y 2015 KDDN: an open-source cytoscape
Sandberg R 2014 Entering the era of single-cell transcriptomics in app for constructing differential dependency networks with
biology and medicine. Nature Methods 11 22–24. (https://doi. significant rewiring. Bioinformatics 31 287–289. (https://doi.
org/10.1038/nmeth.2764) org/10.1093/bioinformatics/btu632)
Sedgewick AJ, Benz SC, Rabizadeh S, Soon-Shiong P & Vaske CJ 2013 Toettcher JE, Loewer A, Ostheimer GJ, Yaffe MB, Tidor B & Lahav G
Learning subgroup-specific regulatory interactions and regulator 2009 Distinct mechanisms act in concert to mediate cell cycle arrest.
independence with PARADIGM. Bioinformatics 29 i62–i70. (https:// PNAS 106 785–790. (https://doi.org/10.1073/pnas.0806196106)
doi.org/10.1093/bioinformatics/btt229) Twycross J, Band LR, Bennett MJ, King JR & Krasnogor N 2010
Singhania R, Sramkoski RM, Jacobberger JW & Tyson JJ 2011 A hybrid Stochastic and deterministic multiscale models for systems biology:
model of mammalian cell cycle regulation. PLoS Computational an auxin-transport case study. BMC Systems Biology 4 34. (https://doi.
Biology 7 e1001077. (https://doi.org/10.1371/journal.pcbi.1001077) org/10.1186/1752-0509-4-34)
Spencer SL & Sorger PK 2011 Measuring and modeling apoptosis in Tyson JJ, Baumann WT, Chen C, Verdugo A, Tavassoly I, Wang Y,
single cells. Cell 144 926–939. (https://doi.org/10.1016/j. Weiner LM & Clarke R 2011 Dynamic modeling of oestrogen
cell.2011.03.002) signalling and cell fate in breast cancer cells. Nature Reviews Cancer
Steyerberg EW, Harrell FE Jr, Borsboom GJ, Eijkemans MJ, Vergouwe Y & 11 523–532. (https://doi.org/10.1038/nrc3081)
Habbema JD 2001 Internal validation of predictive models: Tyson JJ, Laomettachit T & Kraikivski P 2019 Modeling the dynamic
efficiency of some procedures for logistic regression analysis. Journal behavior of biochemical regulatory networks. Journal of Theoretical
of Clinical Epidemiology 54 774–781. (https://doi.org/10.1016/S0895- Biology 462 514–527. (https://doi.org/10.1016/j.jtbi.2018.11.034)
4356(01)00341-9) Vanlier J, Tiemann CA, Hilbers PA & van Riel NA 2012 An integrated
Subramanian A, Tamayo P, Mootha VK, Mukherjee S, Ebert BL, strategy for prediction uncertainty analysis. Bioinformatics 28
Gillette MA, Paulovich A, Pomeroy SL, Golub TR, Lander ES, et al. 1130–1135. (https://doi.org/10.1093/bioinformatics/bts088)
2005 Gene set enrichment analysis: a knowledge-based approach for Venet D, Dumont JE & Detours V 2011 Most random gene expression
interpreting genome-wide expression profiles. PNAS 102 signatures are significantly associated with breast cancer outcome.
15545–15550. (https://doi.org/10.1073/pnas.0506580102) PLoS Computational Biology 7 e1002240. (https://doi.org/10.1371/
Swertz MA & Jansen RC 2007 Beyond standardization: dynamic software journal.pcbi.1002240)
infrastructures for systems biology. Nature Reviews: Genetics 8 Vogel C & Marcotte EM 2012 Insights into the regulation of protein
235–243. (https://doi.org/10.1038/nrg2048) abundance from proteomic and transcriptomic analyses. Nature
Tang J & Aittokallio T 2014 Network pharmacology strategies toward Reviews: Genetics 13 227–232. (https://doi.org/10.1038/nrg3185)
multi-target anticancer therapies: from computational models to Waljee AK, Higgins PD & Singal AG 2014 A primer on predictive
experimental design principles. Current Pharmaceutical Design 20 models. Clinical and Translational Gastroenterology 5 e44. (https://doi.
23–36. (https://doi.org/10.2174/13816128113199990470) org/10.1038/ctg.2013.19)
Tang J, Cho NW, Cui G, Manion EM, Shanbhag NM, Botuyan MV, Wang Z & Deisboeck TS 2014 Mathematical modeling in cancer drug
Mer G & Greenberg RA 2013 Acetylation limits 53BP1 association discovery. Drug Discovery Today 19 145–150. (https://doi.
with damaged chromatin to promote homologous recombination. org/10.1016/j.drudis.2013.06.015)
Nature Structural and Molecular Biology 20 317–325. (https://doi. Wang N, Gong T, Clarke R, Chen L, Shih IeM, Zhang Z, Levine DA,
org/10.1038/nsmb.2499) Xuan J & Wang Y 2015 UNDO: a Bioconductor R package for
Tape CJ 2016 Systems biology analysis of heterocellular signaling. Trends unsupervised deconvolution of mixed gene expressions in tumor
in Biotechnology 34 627–637. (https://doi.org/10.1016/j. samples. Bioinformatics 31 137–139. (https://doi.org/10.1093/
tibtech.2016.02.016) bioinformatics/btu607)
Tavassoly I, Parmar J, Shajahan-Haq AN, Clarke R, Baumann WT & Wang N, Hoffman EP, Chen L, Chen L, Zhang Z, Liu C, Yu G,
Tyson JJ 2015 Dynamic modeling of the interaction between Herrington DM, Clarke R & Wang Y 2016 Mathematical modelling
autophagy and apoptosis in mammalian cells. CPT: Pharmacometrics of transcriptional heterogeneity identifies novel markers and
and Systems Pharmacology 4 263–272. (https://doi.org/10.1002/ subpopulations in complex tissues. Scientific Reports 6 18909.
psp4.29) (https://doi.org/10.1038/srep18909)

via free access
Welton NJ & Ades AE 2005 Estimation of Markov chain transition 26th Conference on Uncertainty in Artificial Intelligence, UAI 2010, pp
probabilities and rates from fully and partially observed data: 701–708. Waterloo, ON, Canada: AUAI Press. (available at: https://
uncertainty propagation, evidence synthesis, and model calibration. event.cwi.nl/uai2010/papers/UAI2010_0221.pdf)
Medical Decision Making 25 633–645. (https://doi.org/10.1177/02729 Zhang B, Li H, Riggins RB, Zhan M, Xuan J, Zhang Z, Hoffman EP,
89X05282637) Clarke R & Wang Y 2009 Differential dependency network analysis
Wen Z, Liu ZP, Liu Z, Zhang Y & Chen L 2013 An integrated approach to to identify condition-specific topological changes in biological
identify causal network modules of complex diseases with application networks. Bioinformatics 25 526–532. (https://doi.org/10.1093/
to colorectal cancer. Journal of the American Medical Informatics bioinformatics/btn660)
Association 20 659–667. (https://doi.org/10.1136/amiajnl-2012-001168) Zhang B, Tian Y, Jin L, Li H, Shih IeM, Madhavan S, Clarke R, Hoffman EP,
Werner HM, Mills GB & Ram PT 2014 Cancer systems biology: a peek Xuan J, Hilakivi-Clarke L, et al. 2011 DDN: a caBIG(R) analytical tool
into the future of patient care? Nature Reviews: Clinical Oncology 11 for differential network analysis. Bioinformatics 27 1036–1038. (https://
167–176. (https://doi.org/10.1038/nrclinonc.2014.6) doi.org/10.1093/bioinformatics/btr052)
Wilkinson DJ 2009 Stochastic modelling for quantitative description of Zhang H, Liu T, Zhang Z, Payne SH, Zhang B, McDermott JE, Zhou JY,
heterogeneous biological systems. Nature Reviews: Genetics 10 Petyuk VA, Chen L, Ray D, et al. 2016 Integrated proteogenomic
122–133. (https://doi.org/10.1038/nrg2509) characterization of human high grade serous ovarian cancer. Cell
Wu G & Stein L 2012 A network module-based method for identifying 166 755–765. (https://doi.org/10.1016/j.cell.2016.05.069)
cancer prognostic signatures. Genome Biology 13 R112. (https://doi. Zou J, Huss M, Abid A, Mohammadi P, Torkamani A & Telenti A 2019 A
org/10.1186/gb-2012-13-12-r112) primer on deep learning in genomics. Nature Genetics 51 12–18.
Yin H, Xue W & Anderson DG 2019 CRISPR-Cas: a tool for cancer (https://doi.org/10.1038/s41588-018-0295-5)
research and therapeutics. Nature Reviews Clinical Oncology 16 Zuckerman NS, Noam Y, Goldsmith AJ & Lee PP 2013 A self-directed
281–295. (https://doi.org/10.1038/s41571-019-0166-8) method for cell-type identification and separation of gene expression
Zhang B & Wang Y 2010 Learning structural changes of gaussian microarrays. PLoS Computational Biology 9 e1003189. (https://doi.
graphical models in controlled experiments. In Proceedings of the org/10.1371/journal.pcbi.1003189)
Received in final form 25 March 2019

Accepted 8 April 2019
Accepted Preprint published online 8 April 2019

via free access

Clarke, Endocrine-Related Cancer) Systems Biology - Perspectives On Multiscale Modeling in Research On Endocrine-Related Cancers, 2010

Uploaded by

Copyright:

Available Formats

Clarke, Endocrine-Related Cancer) Systems Biology - Perspectives On Multiscale Modeling in Research On Endocrine-Related Cancers, 2010

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Clarke, Endocrine-Related Cancer) Systems Biology - Perspectives On Multiscale Modeling in Research On Endocrine-Related Cancers, 2010

Uploaded by

Copyright:

Available Formats

Endocrine-Related R Clarke et al.

Multiscale modeling in cancer 26:6 R345–R368

Systems biology: perspectives on multiscale

Correspondence should be addressed to R Clarke: [email protected]

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

Database URL Data spaces

CPTAC https://proteomics.cancer.gov/data-portal Proteome

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

Method Dynamic variables Time Example

Boolean networks X(t) = 0 or 1 t = integer X inhibits synthesis of Y and

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

Received in final form 25 March 2019

https://erc.bioscientifica.com © 2019 Society for Endocrinology

You might also like

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology

https://erc.bioscientifica.com © 2019 Society for Endocrinology