Using systems biology approaches to understand cardiac inflammation and extracellular matrix remodeling in the setting of myocardial infarction

Abstract

Inflammation and extracellular matrix (ECM) remodeling are important components regulating the response of the left ventricle (LV) to myocardial infarction (MI). Significant cellular and molecular level contributors can be identified by analyzing data acquired through high-throughput genomic and proteomic technologies that provide expression levels for thousands of genes and proteins. Large scale data provide both temporal and spatial information that need to be analyzed and interpreted using systems biology approaches in order to integrate this information into dynamic models that predict and explain mechanisms of cardiac healing post-MI. In this review, we summarize the systems biology approaches needed to computationally simulate post-MI remodeling, including data acquisition, data analysis for biomarker classification and identification, data integration to build dynamic models, and data interpretation for biological functions. An example for applying a systems biology approach to ECM remodeling is presented as a reference illustration.

CARDIAC REPAIR POST-MYOCARDIAL INFARCTION: THE PROBLEM INTRODUCED

Following the onset of myocardial infarction (MI), the left ventricle (LV) undergoes a series of molecular, cellular, and extracellular matrix alterations that combined are known as LV remodeling. LV remodeling includes four overlapping phases: cardiomyocyte death, an acute inflammatory response, granulation tissue formation, and scar formation.¹ Irreversible cardiomyocyte death begins after approximately 20 min of ischemia, and myocytes loss has been attributed to necrosis, apoptosis, and other cell death mechanisms. Cardiomyocyte death releases complement C5a, which stimulates the inflammatory response and begins the wound healing cascade.² Figure 1 shows the time course of cardiac healing and the processes involved at each time period.

Figure 1.

The time course of cardiac healing after myocardial infarction (MI), demonstrating the overlap of processes involved at each time period. About 20 min following MI, myocytes undergo irreversible necrosis and/or apoptosis to induce complement C5a release and initiates an acute inflammation. During the acute response period, infiltrated neutrophils and macrophages release inflammatory mediators and proteinases to degrade extracellular matrix (ECM) and engulf dead tissue debris. Thereafter, granulation tissue is formed, characterized by the presence of macrophages, myofibroblasts, new blood vessels, and ECM components. Eventually, a scar enriched in collagen forms and matures.

An early increase in cytokines and chemokines stimulates neutrophil infiltration into the ischemic area, making neutrophils the first responder leukocyte. Neutrophils release inflammatory mediators and proteinases to degrade ECM and provide chemotaxis signals for subsequent macrophage infiltration.³ During the initial macrophage response, classical M1 pro-inflammatory macrophages prevail and facilitate inflammation and ECM degradation. Subsequently, macrophages convert to an alternative M2 anti-inflammatory subtype that promote resolution of inflammation, ECM synthesis, and scar formation.^{4, 5} A timely switch and dynamic balance between M1 and M2 macrophages is believed to be required to promote appropriate cardiac repair.

Granulation tissue is formed in the infarct area between 2 and 14 days after MI and is characterized by the enhanced presence of macrophages, myofibroblasts, new blood vessels, and ECM components. At this stage, myofibroblasts in both the infarct and remote non-infarct regions produce abundant ECM, with the myofibroblasts in the infarct region being more active than those in the non-infarct region.⁶ In the final stages of wound healing, granulation tissue matures into a collagen-rich, cross-linked scar that spans the entire infarct region and infiltrates into the remote region. The time period for completion of each phase varies depending on the species, with rodents showing an accelerated wound repair and healing compared to larger animal models and humans.⁷

During the different phases post-MI, inflammation and ECM turnover play significant roles in LV remodeling. Inflammation is a bi-directional response, with both necessary and deleterious roles in the repair process. Appropriate inflammation fosters removal of tissue and cell debris to form a stable scar. Excessive and prolonged inflammation or insufficient inflammation both hinder normal cardiac repair by stimulating continual tissue destruction and infarct expansion.⁸

ECM provides the scaffolding structural support for granulation tissue formation. The balance between ECM degradation and synthesis is critical for stable scar formation. Excessive ECM accumulation can increase wall stiffness and stimulate the development of diastolic dysfunction. On the other hand, insufficient or improper ECM deposition can promote cardiac rupture and maladaptive cardiac remodeling.^{5, 9} The ECM environment plays a key role in scar formation by transducing mechanical, electrical, and chemical signaling pathways.¹⁰ The ECM coordinates responses between myocytes in the viable non-infarcted remote region and neutrophils, macrophages, endothelial cells, and fibroblasts in the infarct region. Therefore, investigating the mechanisms of inflammation and ECM remodeling post-MI will help us to better understand MI responses that can stimulate adverse LV remodeling and lead to the development of congestive heart failure.

As described above, one critical feature of post-MI LV remodeling is the importance of the time course of inflammation and matrix turnover. For example, early neutrophil recruitment provides a key signal for subsequent macrophage trafficking and polarization, which coordinate myofibroblast phenotype and function. Due to different roles of these cellular and molecular players in MI response, significance of the players varies with time and temporal profiles of these players carry such information for us to discover. With this consideration, time course data are necessary to identify key markers and their underlying regulatory mechanisms. To address this need, systems biology provides useful techniques including temporal clustering for identification of key markers and differential equation models to characterize and interpret regulatory mechanisms with time course data.

Challenges of systems biology approaches using –omics technologies

During the last three decades, advancements in molecular and cellular biology have yielded dramatic progress in biomarker discovery and functional characterization that improve our understanding of mechanisms of LV remodeling.¹¹ There is a vast amount of experimental data deposited in public databases with respect to gene and protein expression, cellular responses, and tissue remodeling post-MI. These data cover temporal as well as the spatial response between infarct and remote non-infarcted myocardium.¹² The recent emergence of genomics and proteomics datasets provides a chance for researchers to examine the whole system instead of focusing on only one or several molecules.^{13, 14} The generation of these very large datasets has also generated issues on how to best analyze and extract useful information from the data.

To address this need, systems biology approaches have been proposed and developed to elucidate the mechanisms of biological response that can be applied to our example of post-MI LV remodeling.¹⁵ Systems biology is the conceptual framework that analyzes and integrates complex and highly diverse information acquired largely from isolated high-throughput -omics datasets.¹⁶ This approach is a transition from the molecular cell biology method that studies each data set individually to a more complete systemic analysis and integration to interpret the dynamics of the entire biological system. According to the different levels of the organism being studied, systems biology has adopted different processing methods. It incorporates all -omics technologies, including genomics, proteomics (and its subspecialties, such as matridomics and degradomics), matabolomics, and interactomics with signal processing algorithms. It also allows characterization of the supramolecular networks and functional modules that represent the most essential aspects of cell organization and physiology with mathematical modeling and graphical networks.¹²

One problem encountered when applying systems biology approaches is how to handle very large datasets. The advanced high-throughput –omics technology can offer expression levels for >1,000 genes or proteins in one experiment. Additionally, such experiments might be quickly repeated many times for temporal progression information, in different organs from different species for comparative information, or with different dosages of interventions, which increases the dimensionality of the data compared to previous snap-shot data.^17–19 In addition to this big data set problem, integrating datasets collected from gene, protein, and cellular components to depict whole spectra of the dynamic progression is another challenge for systems biology generally and the cardiac research field specifically.

In this review, we will summarize current methods of systems biology for data acquisition, analysis, integration, and interpretation to understand cardiac inflammation and ECM deposition in the setting of MI. A systemic flow chart of this approach is presented in Figure 2.

Figure 2.

Comparison between the gene expression levels of 84 extracellular matrix (ECM) genes after myocardial infarction (MI) in the infarct region of saline treated mice and 3 treatment groups (aliskiren- A; valsartan- V; and combination-A+V) for the 4 most significant layers obtained by SSVD biclustering. The biclustering method was used to find potential ECM biomarkers of cardiac remodeling after MI. The original matrix is shown at the top, the reconstructed matrix containing the four most significant layers is shown in the middle, and the reconstructed matrix containing layer 1 is shown at the bottom. The highlighted columns represent the genes identified as potential ECM biomarkers. In the color legend, the red color means that a gene expression is up-regulated, while dark blue represents down-regulation of a gene expression with respect to the expression levels at day 0.

ACQUIRING LARGE DATASETS TO OBTAIN SUFFICIENT INFORMATION

Recent technologies such as next generation sequencing techniques can produce huge amounts of genomic data in less than a week, while assembling the first human genome data took about ten years and the cooperation of >200 research groups worldwide.¹⁷ As another example, proteomics studies frequently utilize multidimensional techniques for data acquisition, including different approaches for separation of proteins (1D vs. 2D polyacrylamide gel electrophoresis in buffers of differing solubility), several techniques for separating peptides (1D liquid chromatography (LC) vs. 2D LC vs. isoelectric focusing), and a variety of methods for mass spectrometry (MS) such as matrix-assisted laser desorption/ionization, electrospray ionization, and ion trap or time of flight analysis to generate MS and LC-tandem MS data.¹⁸ Measurement of a large number of biomarkers (e.g., covering the whole genome) of a single cell for several samples over a period of time is also an example that shows the high dimensionality possible in data acquisition.¹⁹

Using arrays to examine the molecular level

Gene and protein arrays provide a relatively high-throughput method for examining multiple molecules simultaneously, and it is useful to segregate molecules based on their potential involvement according to patterns of change in levels. Genes or proteins that do not increase or decrease post-MI are assumed to maintain a steady state during the response, assuming that the right time window was sampled. Both inflammatory and ECM gene and protein arrays are commercially available (e.g., from Qiagen, Biorad, and R&D). While it is well accepted that mRNA levels do not always reflect protein levels, measuring multiple gene levels is technically easier than measuring multiple protein levels. Additionally, measuring mRNA or protein levels does not give information on protein quality. This is an especially important consideration for ECM, as ECM proteins generally undergo at least one post-translational modification. Typical modifications include glycosylation and proteolytic processing. Therefore, proteomics data will be much more important for studying the dynamics of ECM remodeling for protein quantification. Additionally, MS-based proteomics has the capability of evaluating proteins and their complex interactions. Techniques to obtain proteomics data for ECM specifically are reviewed and summarized below.

Using matridomics to examine the extracellular matrix level

Matridomics is defined as the comprehensive study of all ECM protein components expressed in a tissue at the time of evaluation. This proteomic array technique describes the global, integrated changes of ECM in a given condition, which has several advantages over transcriptomics or traditional one target studies.²⁰ Matridomics provides direct information on ECM protein quantity and quality. Barallobre-Barreiro and colleagues have used an innovative proteomic approach to analyze the cardiac ECM in a porcine model of ischemia/reperfusion.²¹ The detection and identification of the relatively low abundant cardiac ECM is a challenge that requires a sequential extraction procedure to enrich for ECM proteins. Using such an approach led to the identification of 139 cardiac ECM proteins by liquid chromatography/mass spectrometry, and 15 of these ECM proteins had not been previously shown to be involved in cardiac remodeling. A comparison between the ischemic region and border zone revealed much higher proteins in the infarct area but similar mRNA levels in both regions, which highlights the disparity between mRNA and protein levels. Further, the ECM proteins delineated a signature of early- and late-phase LV remodeling, with transforming growth factor β₁ (TGFβ₁) signaling at the core of the interaction network. Importantly, these novel cardiac ECM proteins identified by proteomics approach in the pig were also validated in human samples.

Our team has expanded on this approach by recently exploring a decellularization technique where the mouse LV tissue was extracted into fractions using three extraction buffers of increasing solubility potential, which we named the Texas-3 step protocol.²² The LV tissue was first incubated in a neutral low salt buffer with proteinase inhibitors to extract the most easily soluble components. The LV tissue was then decellularized with 1% sodium dodecyl sulfate with proteinase inhibitors to collect soluble cellular proteins, including the large amount of mitochondrial proteins present in LV tissue. As a final step, the decellularized tissue was subjected to acid extraction, deglycosylation, and solubilization using a 4 M GnHCl and 50 mM sodium acetate buffer with proteinase inhibitors to collect the insoluble ECM fraction. This protocol allows qualification and quantification of ECM components in tissue samples by removing highly abundant cellular and mitochondrial proteins.²²

Using degradomics to examine the matrix metalloproteinase level

Degradomics is defined as a proteomic approach to catalogue all proteinases, inhibitors, and substrates in a tissue at the given time of evaluation.²³ Multiple matrix metalloproteinases (MMPs) and the tissue inhibitors of metalloproteinases (TIMPs) have been shown to be involved in post-MI cardiac remodeling by regulating ECM turnover, inflammation, apoptosis, and angiogenesis.^{5, 24} Accumulating evidence suggests that ECM fragments cleaved by proteinases, the matricryptins, have biological effects and actively regulate angiogenesis, inflammation, and fibrosis.^{10, 25} Matricryptins are active ECM fragments that contain a cryptic domain that is only exposed after conformational changes of the parent ECM.²⁶ The presence of cryptic sites may provide key signals to regulate cell migration, proliferation, differentiation, morphogenesis, survival, ECM turnover, angiogenesis, and tissue repair by binding to corresponding receptors (e.g., integrins).^{10, 27}

Collagen fragments generated due to MMP-9 cleavage have been shown to inhibit angiogenesis and stimulate MMP-9 expression, which forms a positive feedback loop to maintain MMP-9 levels.^{28, 29} Collagen IV can be cleaved by MMPs to generate both proangiogenic and antiangiogenic fragments.³⁰ In addition, matricryptins mediate the inflammatory response by modulating chemotactic activity for inflammatory cells, amplifying phagocytic function, activate immune responses, and altering the gene expression profiles of leukocytes.³¹

Our group has identified in vivo candidate substrates for MMP-7 and MMP-9 in 7 day post-MI infarct tissue using 2-dimensional gel electrophoresis analysis and mass spectrometry based approaches.^{32, 33} In the MMP-7 null mice, the infarct area showed a lower intensity of spots that were identified to include fibronectin (Fn1) and tenascin-C (TnC). In vitro cleavage assays verified that Fn1 and TnC fragments were generated by MMP-7. Further, infusion of exogenous recombinant MMP-7 restored the production of Fn1 and TnC fragments in MMP-7 null mice, confirming that Fn1 and TnC are in vivo MMP-7 substrates.³² This was the first report to identify Fn1 and TnC as in vivo MMP-7 substrates using a proteomics approach.

Using a similar proteomics approach, we found Fn1 could also be cleaved by MMP-9 in the MI setting.³³ In these studies, using only infarct tissue provided a way to naturally focus on ECM, because ECM proteins are enriched in the scar tissue at day 7 post-MI. Under normal conditions, mitochondria accounts for >30% of the myocyte volume, and myocyte comprise >90% of the LV volume. Compared to these levels, ECM components are in very low abundance and harder to analyze by mass spectrometry because of the noise contributed by mitochondria and other intracellular components. One goal of our ongoing projects is to investigate ways to further enhance ECM representation in our analyses.

Using genomics and proteomics deposition databases

Tremendous amounts of gene expression data have been deposited in public databases. Currently, a query of NCBI Gene Expression Omnibus (GEO) leads to public data repositories including 11,752 platforms, 61,727 samples, 39,713 series, and 3341 datasets- as of July 15, 2013. A total of over 1 million microarray results were available by the year of 2008, and the number of available microarrays doubles every 2–3 years. A summary of the data deposition resources available online is provided in Table 1.

Table 1.

Public resources of available data and tools for systems biological approaches

Database	Specification	Website
Genomics	GEO	http://www.ncbi.nlm.nih.gov/geo/
Proteomics	UniProt	http://www.uniprot.org
	PRIDE	http://www.ebi.ac.uk/pride/
	GPM	http://gpmdb.thegpm.org
	DIP	http://dip.doembi.ucla.edu/dip/Main.cgi
	MIPS	http://mips.helmholtzmuenchen.de/proj/ppi/
	HPRD	http://www.hprd.org
	BioGRID	http://thebiogrid.org
	IntAct	http://www.ebi.ac.uk/intact/
	HomoMINT	http://mint.bio.uniroma2.it/HomoMINT/Welcome.do
Data integration	Mathematical models	http://bioeng.washington.edu/Models/).
Data interpretation	DAVID	http://david.abcc.ncifcrf.gov
	iProClass	http://pir.georgetown.edu/pirwww/dbinfo/iproclass.shtml

GEO- Gene Expression Omnibus; UniProt-Universal Protein Resource; PRIDE-Proteomics Identifications Database; GPM-The Global Proteome Machine; DIP-Database of Interacting Proteins; MIPS-Mammalian Protein-Protein Interaction Database; HPRD-Human Protein Reference Database; BioGRID-Biological General Repository for Interaction Datasets; IntAct-Mollecular Interaction Database; HomoMINT- Human Protein Interactions in the MINT Database); DAVID-The Database for Annotation, Visualization and Integrated Discovery; PIR-Protein Information Resource.

Similar to genomics databases, several worldwide organizations have provided online proteomics database repositories, including the PRoteomics IDEntifications (PRIDE) and Global Proteome Machine (GPM) databases. PRIDE is a public, user-populated proteomics data repository.^{34, 35} Users can upload, download, and view raw data generated by mass spectrometry proteomics experiments, including raw spectral data, peptides, protein identifications, and associated statistics through a free web interface as listed in Table 1. GPM gains the advantage of allowing researchers to use its proteomics data and tools to interrogate a number of proteomes.³⁶ Currently, there is no specific data service focusing on cardiovascular research proteomics, and such a resource would significantly benefit the field.

It is worth mentioning that the protein-protein interaction databases such as Database of Interacting Proteins (DIP), mammalian protein-protein interaction database (MIPS), Human Protein Reference Database (HPRD), Biological General Repository for Interaction Datasets (BioGRID), IntAct database, and HomoMINT database, among others, provide information on a total of >70,000 proteins and 330,000 interactions.^37–42 The reported data represent gene and protein expression profiles from different species under varying experimental conditions. Therefore, there is an urgent need to systemically integrate and analyze such data to elucidate the underlying regulatory mechanisms.

PERFORMING DATA ANALYSIS TO HARNESS THE INFORMATION

Before data analysis can be performed, the data must be cleaned to remove possible sources of noise, such as those coming from experimental design errors, measurement noise, and technical errors. These noises overlap with the inherited individual differences of biological processes, leading to a difficulty in determining true measurements. Thus, data cleaning is needed to filter noises and control the quality of data before any further investigation of the biological system can be undertaken. Several statistical methods have been used for detecting and correcting measurement outliers. These methods include finding the average or median of the data, computing the p-value, using boxplots for visualizing outliers, and using boot strapping methods to assign measures of accuracy to sample estimates.⁴³

Data analysis or mining is defined as an automatic process of retrieving and discovering important and useful information in a large data set. Many computational methods have been proposed to find new significant biomarkers with relevant biological meaning. These techniques include unsupervised machine learning techniques such as clustering and principal component analysis (PCA), as well as supervised learning approaches that utilize the prior knowledge of the system under investigation such as support vector machine (SVM).

Clustering methods are generally used to classify objects of a large population into smaller groups called clusters. The similarity between members in each cluster is higher than that among members of different clusters. Clustering is often applied to biological studies to find gene or protein co-regulatory partners. The underlying assumption is that genes or proteins sharing similar behavior may regulate a biological process together and can be classified as one cluster. This approach has also been used to find sets of biomarkers associated with a specific function.⁴⁴ Clustering methods are useful for predicting which group a biomarker belongs to and what functions that biomarker has. As more and more data are accumulated, researchers face the problem of comparing and integrating data collected under different attributes such as treatment dosages, species, and genotypes. This has led to biclustering methods that can identify not only significant markers but also the contribution of each attribution.

We have used a biclustering method to find the significant biomarkers of MI that could differentiate treatment strategies.⁴⁵ This biclustering algorithm was performed on gene expression levels in the LV of mice before MI (control) and 28 days after MI with saline, aliskiren, valsartan, or a combination of both drugs treatment that was started at 3h post-MI (Figure 3). In another study, the differentially expressed genes were classified based on their functions, and >200 genes were identified as being related to MI.⁴⁶

Figure 3.

Flow chart of data acquisition, analysis, integration, and interpretation used for systems biology approaches.

PCA is a mathematical method that converts a set of possibly related observed variables into a set of linearly uncorrelated variables by orthogonal transformation. PCA is usually used to reduce the dimensionality of expression data that will eventually identify groups of genes that have similar characteristics.⁴⁷ Barallobre and colleagues used a PCA approach to identify the 20 most significantly expressed proteins in the ECM that were related to MI.²¹ These applications provide a promising direction to integrate systems biology approaches into experimental research to focus on hypothesis generation and data analysis.

In addition to clustering approaches, SVM is another popular classification method that was originally developed by Vapnik and Cortes in 1995.⁴⁸ The SVM algorithm can predict which group the input belongs to by maximizing the margin between support points and using nonlinear mapping based on the model obtained from the training dataset. It has been widely applied to breast cancer diagnosis, protein sequencing, and ECG signal analysis for cardiovascular research on arrhythmias.^49–52 To date, no application of SVM on inflammation and ECM remodeling has been conducted.

INTEGRATING DATA TO DEVELOP INTERPRETATIONS

Data analysis can elucidate the relation of data based on correlation analysis, an approach familiar to engineers, bioinformaticians, and biologists. More advanced integration, however, is required to interpret the spatiotemporal dependence embedded in the data. Furthermore, clear biological meaning of the data analysis should be presented to bridge the knowledge gap between biologists and bioinformaticians. Different integration schemes including differential equation and graphical network modeling have been proposed to integrate experimental data.

Mathematical modeling for temporal dynamics using differential equations

Ordinary differential equations (ODE) and partial differential equations (PDE) have been applied to model the temporal and spatial changes of biological and physical variables in a continuous format.^53–55 Due to the complexity of PDE solving and modeling techniques, we use ODE to illustrate the modeling method here. ODE-based algorithms are able to infer the nonlinear dynamic nature of high dimensional regulatory mechanisms and predict the behavior of a system based on the constructed model.^{53, 54} There are two approaches in studying regulatory relationships characterized by physical and influence methodologies.^56–59 In the physical approach, exact physical interactions between molecules such as transcription factors and their promoters are investigated and the impact of other factors in the regulatory mechanism is ignored. In the influence approach, regulatory interactions among molecules are studied such that every regulation is considered as an input and output. The regulatory process is not controlled by real interactive molecular bindings, although the model sheds light on hidden indirect associations among molecules. The general format of ODEs for this approach is written as follows:

$${X_i}^{\prime }=f_i(X_1,\dots ,X_N,Y,\theta ),$$ (1)

where X_i denotes molecular concentration of the i^th gene or protein at time t, with time-derivative term ${X_i}^{\prime }=\frac{d{X}_i}{dt}$, the regulatory effect of other external stimuli such as chemical compounds or genetic perturbation are illustrated by Y, and the network dynamic model is characterized by f_i(.), which reveals dynamic interactions among all molecule X_is. θ is represented as the model-parameter vector. To simplify the ODE equation (1) and reduce the computation load, the network dynamic model is approximated such that the changing rate of molecular concentration of the i^th gene or protein is affected by weighted accumulation of all molecular expressions. The kernel of such ODE model is in linear format and represented in equation (2): ^{58, 59}

$${X_i}^{\prime }=\sum _{j=1}^N{\theta }_{ij}{X}_j.$$ (2)

The construction of the model described in equation (2) is formulated by determining associated weights, θ_ijs, that describe the influence of molecule j on molecule i. In the presence of some prior biological knowledge about their kinetic rates, only a subset of parameters, θ_ij, is required to be estimated. Least square and Bayesian approaches are wildly applied to estimate the unknown parameters.^{60, 61}

The difficulties of parameter estimation lie in the fact that many parameters in the model need to be estimated, and there is a lack of data samples to estimate the parameters accurately. In this case, we need to limit the number of unknown parameters or take advantage of sparsity of such regulatory networks to reduce the dimension of the dataset.^{58, 62, 63} Different parameter estimation approaches with time series data have been proposed for both linearly and nonlinearly parameterized systems.^{54, 61, 64–68}

While using such a linear format simplifies the modeling techniques, problems that arise with such a simplification approach include 1) how to accommodate the innate nonlinear dynamics of biological regulation due to multiple protein binding sites for protein interaction; and 2) how to approximate nonlinearly parameterized chemical reactions such as a hill equation with linear format.

ODEs and PDEs are deterministic mathematical model, i.e., the output of the system is determined given one set of parameters, initial condition, and input, while real biological processes demonstrate strong individual differences. With the same input, same biological processes may have different initial conditions and parameters. In most current research, variations of the system outputs are due to different initial conditions and noises. Sensitivity of parameters are examined to further simplify the differential equation model by focusing on key sensitive parameters and ignoring non-sensitive parameters in the model.⁶⁹ Under such direction, bifurcation analysis has been applied to illustrate the effect of parameter variation on system outputs.⁷⁰ Additionally, the effect of noises is confined within a stable system, i.e., a bounded input leads to bounded outputs. Currently, very few studies have been reported to use stochastic differential equation model to represent individual differences.^71–73 There is a lack of research linking the distribution of key parameters to distribution of the output.

Many differential equation models of cardiovascular systems can be found in the website supported by the Bassingthwaighte group (http://bioeng.washington.edu/Models/). The complexity of ECM metabolism and inflammation signaling in the post-MI setting lends itself to ODE approaches. Our team has established an ODE model of cardiac remodeling post-MI to quantify the balance between the construction and destruction of ECM.⁷⁴ To establish this ODE model, we first determined key variables representing ECM metabolism using our experimental results. We examined the expression of 84 ECM genes at 7 days post-MI and identified 17 differentially expressed genes (>2.5 fold and p<0.05). Among these genes, collagen I was selected as a key factor, because it is the major collagen in the myocardium. Osteopontin expression levels increased 206-fold in the infarct region compared to the control group, suggesting strong macrophage activation at day 7 post-MI. TIMP-1 expression increased 31-fold in the infarct compared to the control group, suggesting strong inhibition of proteolytic activity. Specifically, TIMP-1 inhibits MMP-9 and MMP-9 degrades collagen. Periostin gene expression level was increased 5.5-fold, and TGFβ1 gene expression level was increased 2.6-fold in the infarct sample at day 7, suggesting significant fibroblast functions post-MI. Based on these experimental results, the regulatory network of ECM metabolism includes the cellular contributions of macrophages and fibroblasts, and molecular contributions of TGFβ₁, MMP-9, collagen I, and TIMP-1 as main components influencing LV remodeling.

With the determination of these variables, we further examined the regulatory relationship with in vitro experiments and published literature to determine the parameters in the ODE model. For example, we determined the fibroblast secretion rate of collagen stimulated by TGFβ₁ and the macrophage infiltration rate stimulated by TGFβ₁. After we determined the variables and parameters of the ODE model, we simulated the model and validated the model by comparing our computational results against temporal profiles of these variables from other research groups. Our computational predictions agree well with the observed experimental data.

Besides our group, Vovodotz and colleagues have developed several differential equation models for inflammatory responses by combining experimental and computational approaches.⁷⁵ The Popel laboratory has developed theoretical model for collagen degradation by MMPs at the molecular level.^{76, 77} ODE models have also been widely adopted for other cardiovascular research topics in addition to ECM metabolism and inflammatory signaling. ODE models on cardiac electrophysiological from cellular level to organ level have been generated.^78–81 Noble and colleagues developed a cardiac electrical activity model in 1985.⁸² Since then, Luo-Rudy models and Jafri-Rice-Winslow model have been reported and widely used to describe mechanisms of cardiac Ca²⁺ regulation in myocytes.⁸³ All of these models illustrate the use of ODE as a promising way to interpret time course progression such as that seen in the MI response.

While differential equation models can represent biological measures, such as molecular concentration and cellular density, with respect to temporal and spatial dynamics, most models are built with influence methodologies without attempting to interpret the detailed binding information. In addition, the modeling procedure and analysis heavily depend on engineers and mathematicians due to the technical details, and this process is not very user unfriendly for typical biologists to understand or use the model. Graphical network models, including Bayesian network and protein-protein interaction (PPI) network, are new emerging representations that can improve model usage by biologists.

Using protein-protein interaction networks

A PPI network visualizes the interactions among proteins by representing each protein as a vertex and each interaction as an edge. Different measurements have been used to investigate structural properties of a network, including centrality analysis (shortest path, betweenness centrality, closeness centrality, and eccentricity) and clustering analysis (local/global clustering coefficient for community detection, average clustering coefficient with respect to degree). The shortest path is a pairwise concept that defines the minimum number of edges required to move from one vertex to another. Betweenness centrality measures the frequency of appearance of a vertex or link in the set of all shortest paths.^{84, 85} High betweenness centrality indicates that a vertex or edge is more frequently used to flow the information in the network. Closeness centrality measures how many steps are required to access every other vertex from a given vertex.⁸⁴ The vertex with the smallest closeness centrality requires the least number of edges to spread information to other reachable vertices in the network. Eccentricity of a vertex measured the shortest path distance from the farthest vertex in the graph.⁸⁶ Local clustering coefficient describes the connectivity of the neighbors of a vertex.⁸⁷ A higher clustering coefficient means the neighbors are densely connected to each other. A tremendous amount of recent research has investigated the structural properties of PPI networks using these network concepts.^88–90 However, studies are needed to identify key proteins of PPI network using only the structural properties.

Key proteins in a PPI network can be identified with enrichment analysis from DAVID with respect to Gene Ontology (GO) for functional linkage.^{91, 92} GO is a controlled vocabulary of terms that characterizes gene products in terms of their cellular components, biological processes, and molecular functions in different species. Significant effort has been contributed to such interpretation and available public resources such as Uniprot (http://www.uniprot.org/) and iProClass (http://proteininformationresource.org/).^93–97

When the number of proteins studied in a disease is small, it is possible to predict and validate functionality of each important protein. When many proteins are involved in the process, finding the most useful proteins as biomarkers is tedious and hard if not impossible. LV remodeling is a highly complicated process that includes vast amounts of protein changes with respect to signaling pathways, inflammatory response, and angiogenesis, all of which are coordinated by ECM remodeling.

Targeting the MI-specific proteins provides the opportunity to study MI responses in detail and propose preventive or treatment strategies. Our team has established a PPI network based on a set of known seed proteins and compared it with 100,000 randomly generated networks to demonstrate the specificity of the MI-specific network.⁹⁸ Azuaje and colleagues have established a PPI network focusing on the inflammatory response post-MI to find the new prognostic biomarkers for LV dysfunction.^{28, 99} They identified inflammatory proteins, including TNF receptor-associated factor 2, SH3 domain-containing-binding protein 1, and Ubiquitin C protein family members. Their associations of these proteins with clinical outcomes post-MI provided the functional characterization of the network.

Ren and colleagues provided a NetCAD platform to construct a PPI network for coronary artery disease. Their PPI network provides physicians and scientists important information about the molecular mechanisms and potential targets for therapies of coronary artery disease.¹⁰⁰ There have also been several groups that studied the effects of drugs on MI outcomes using PPI networks.^101–103 In a recent study, we introduced a multiple-delayed linear regression model to estimate the time delay between the transcription factors and their targeting genes post-MI.¹⁰⁴ A regulatory network was generated with different time delays where genes with high degree of connectivity were identified as potential biomarkers for MI, including specificity protein 1 (Sp1), histocompatibility 2-K1 (H2-K1), v-rel reticuloendotheliosis viral oncogene homolog A (Rela), nuclear factor NF-kappa-B p105 subunit (Nfkb1), and interleukin 18 (Il18).

BIOMARKER IDENTIFICATION: A CASE STUDY

We applied a sparse singular value decomposition (SSVD) biclustering algorithm to analyze gene expression data for significant biomarkers post-MI under different treatments methods (Figure 2).⁴⁵

Data acquisition (Figure 3A and B): data includes expression of 84 ECM genes in the LV infarct region of C57BL/6J wild type (WT, n=30) mice. The mice were divided into 5 groups; control group at day 0 (number = 6) and 4 treatment groups at 28 days post-MI. The treatment groups include MI with saline treatment (S, number = 6), MI with Aliskiren treatment (A, 50 mg/kg/day, number = 6), MI with Valsartan treatment (V, 40 mg/kg/day, number = 6), and MI with Aliskiren and Valsartan treatment (A+V, number = 6).

Data analysis (Figure 3 C and D): We examined the p-value of gene expression to control the data quality. We then computed fold changes of genes expression at day 28 compared to its average value at day 0 and applied the SSVD biclustering algorithm. The most significant feature of the fold change data found by this algorithm is shown in Figure 3. Out of 84 ECM genes evaluated, 10 were differentially regulated in the infarct region.

Data integration (Figure 3 F): We further examined the biological functions involved with these significantly expressed genes to confirm the coherence between analytical results and biological meanings. For each gene set, we searched DAVID for enriched gene ontology (GO) terms for their associated biological processes.⁸⁰ We observed 23 biological processes that are significantly involved with the 10 genes in the infarct region.

LIMITATIONS AND FUTURE DIRECTIONS

While the above approaches give promise for uncovering underlying mechanisms of LV remodeling post-MI using systems biology approaches, there are several areas where future work is needed (Table 2).

Table 2.

Future advances needed for better use of experimental and computational approaches to define post-MI LV remodeling.

Experimental	Computational
Better solubilizing ECM proteins to fully catalogue changes in the post-MI LV Optimizing plasma/serum evaluations to provide surrogate markers of tissue remodeling	Applying systems biology approaches to personalized medicine Using synthetic biology to reconstruct the LV remodeling system to include molecular, cellular, and tissue level functions

For cardiac ECM proteomic analysis, tissue solubility is a challenge that needs to be considered. Although the decellularization procedure greatly removes intracellular components and enriches ECM, not all of the ECM proteins are easily dissolved, especially high molecular weight glycoproteins, proteoglycans, and cross-linked ECM. This may mask potentially important ECM protein changes. Therefore, future studies are warranted to further improve proteomic protocols that focus on ECM.

Plasma or serum can be a valuable source of biomarkers for MI patients, and more research is needed to determine how well changes in the blood mirror changes in the LV tissue. Most studies using animal models analyze cardiac tissue to understand the molecular mechanisms of MI-induced remodeling. Indeed, tissue examination offers invaluable information on the pathogenesis. Because cardiac tissue samples cannot easily be collected in clinics to evaluate patient prognosis, however, having biomarkers that spill over into the blood to serve as surrogate indicators will be critical. The Human Proteome Organisation (HUPO) has an ongoing project to identify all proteins in the human plasma. Berhane and colleagues identified 3020 proteins in human plasma using mass spectrometry.¹⁰⁵ Subsequently, a subset of these proteins has been annotated as cardiovascular related proteins based on the published literature. Several groups have analyzed the plasma proteomics after depletion of albumin and identified hundreds of proteins.^{106, 107} Given the magnitude of differences among the least and most highly expressed proteins, plasma evaluations remain challenging. Identifying which of these proteins best track with LV remodeling is the next step.

Applying systems biology approaches to achieve personalized medicine has been a hot topic over the past five years. In personalized medicine, there is a paradigm change from diagnosis and treatment of disease for an overall population to personalized monitoring and preventive medicine for individuals.¹⁰⁸ This aim may not be achievable without gathering the whole dataset for each patient through –omics technologies, together with incorporating novel methods to find the most informative biomarkers of the disease and will likely include methods that are able to combine spatial and temporal data and integrating the relevant –omics data into more comprehensive models to predict the behavior of the biological system.^{109, 110} Whether we can reduce the number of variables that must be assessed for accurate model building remains to be determined.

Synthetic biology is another trend of systems biology, which builds up a biological system from its cells to the whole organ or living organism.¹¹¹ In this approach, each cell is considered to be a computer that is involved in the flow of information by signaling interactions from and to their neighborhood environment. The living organism is the integration of all these modules (cells). This approach requires extensive data acquisition, big data management, and integration that is currently challenging on several levels.

CONCLUSIONS

In this review, we introduced the significance and challenges of applying systems biology approaches, using the inflammatory response and ECM remodeling following MI as an example. We summarized possible systems biology approaches to handle vast sums of data at multiple stages from acquisition, analysis, integration, and interpretation to elucidate the spectra for systemic studies. This review should provide a useful reference guide for researchers new to this area by summarizing the possible algorithms and resources available for computational modeling.