Reverse engineering biomolecular systems using −omic data: challenges, progress and opportunities

Abstract

Recent advances in high-throughput biotechnologies have led to the rapid growing research interest in reverse engineering of biomolecular systems (REBMS). ‘Data-driven’ approaches, i.e. data mining, can be used to extract patterns from large volumes of biochemical data at molecular-level resolution while ‘design-driven’ approaches, i.e. systems modeling, can be used to simulate emergent system properties. Consequently, both data- and design-driven approaches applied to –omic data may lead to novel insights in reverse engineering biological systems that could not be expected before using low-throughput platforms. However, there exist several challenges in this fast growing field of reverse engineering biomolecular systems: (i) to integrate heterogeneous biochemical data for data mining, (ii) to combine top–down and bottom–up approaches for systems modeling and (iii) to validate system models experimentally. In addition to reviewing progress made by the community and opportunities encountered in addressing these challenges, we explore the emerging field of synthetic biology, which is an exciting approach to validate and analyze theoretical system models directly through experimental synthesis, i.e. analysis-by-synthesis. The ultimate goal is to address the present and future challenges in reverse engineering biomolecular systems (REBMS) using integrated workflow of data mining, systems modeling and synthetic biology.

INTRODUCTION

Recent advances in high-throughput biotechnologies have impacted the progress in reverse engineering biological systems, especially for biochemical pathways and cells. The large volume of biochemical data at molecular-level resolution has made ‘data-driven’ approaches possible, where patterns may be extracted directly from the data to reveal new insight [1–6]. Such data-driven approaches are also known as ‘data mining’. At the same time, ‘design-driven’ approaches are also increasingly used to replicate features of biological systems such as scale-free organization [7], motif distribution [8] and feedback [9] so as to simulate and abstract emergent properties of biological systems [10, 11]. Such design-driven approaches may also be called ‘systems modeling’. Thus, high-throughput data mining coupled with design-driven systems modeling for biochemical pathways and cells represents a new era of reverse engineering biological systems, which we refer to as ‘reverse engineering biomolecular systems’ (REBMS) in this context (Figure 1).

Figure 1.

Integrated pipeline for reverse engineering biomolecular systems (REBMS) using high-throughput –omic data that combines data mining, modeling and simulation and testing and validation

On the whole, the key challenges in REBMS are to (i) integrate heterogeneous biochemical data for data mining [12–16], (ii) combine top–down and bottom–up approaches for systems modeling [17–22] and (iii) validate system models experimentally [23–28]. In this article, we first present an overview of common high-throughput platforms to recap existing data acquisition issues in Section 2. In Section 3, we review challenges and strategies for data mining and systems modeling, and in Section 4, we illustrate some recent progress made in the field of REBMS using two case studies. We close in Section 5 by exploring the emerging field of synthetic biology, which is an exciting approach to analyze and validate theoretical system models directly through experimental synthesis.

DATA ACQUISITION AND HANDLING

In this section, we review high-throughput data acquisition technologies for genomic, proteomic, metabolomic and cytometry data such as microarrays [29], next-generation sequencing (NGS) [30], mass spectrometry [31] and microfluidics [32]. It is important to recap existing issues encountered during data acquisition in the hope of recognizing the potential ‘garbage in—garbage out’ problem during downstream mining, modeling and validation. Then, while recognizing that ever-increasing data volumes present associated challenges in data handling such as storage security, retrieval efficiency, manageability of databases [33], we also review minimum information standards in particular because of its development as a community-driven effort to manage and integrate different data types and formats through standardized database schema and visualization [15].

High-throughput omic data acquisition

DNA microarrays [29] are a common high-throughput technology to acquire genomic data, namely gene expression, at relatively low cost. Data artifacts in microarrays regularly arise due to hybridization and spatial effects [34], slide variation [35] and dye-related signal correlation bias in two-color assays [36]. Subsequently, various engineering and statistical solutions have been developed to handle these artifacts [37, 38]. Recently, NGS technologies, such as Illumina/Solexa, Roche/454 and SOLiD, have also emerged as high-throughput genomic data acquisition platforms but not without some technical and analytical issues. NGS data volume is much larger than that of DNA microarray data, leading to issues in data storage, transfer and analysis [39]. However, it is still difficult to determine expressions from NGS data because of the low specificity of sequence alignment and the presence of splice variants [40]. Accordingly, several algorithms and tools have also been developed to address such NGS-specific issues, namely to improve the accuracy of sequence alignment, quantifying gene expression and to present sequence data visually [41–44]. Two common types of transcriptome assembly strategies are: (i) reference-based algorithms, such as SpliceMap [45], Blat [46] and TopHat [47], that work with a reference genome for assembly, (ii) the so called ‘de novo’ strategies, such as Trans-ABySS [48] and Trinity [49], that work by traversing collapsed De Bruijn graphs constructed from all substrings of reads with a specific length. In recent years, the advent of these methods in conjunction with high performance computing frameworks provided us with a complete landscape of full set of transcripts (Figure 1).

For proteomic and metabolomic data, mass spectrometry is a popular technology with applications ranging from proteomic characterization to perturbation analysis of model organisms [31, 50]. Furthermore, recent research in tissue imaging mass spectrometry (TIMS) effectively adds a second dimension to traditional mass spectrometry data by acquiring multiple mass spectra at different spatial locations of the same tissue sample. Consequently, TIMS may be used to study spatial progression by identifying and quantifying biomarkers of interest that are associated with particular tissue regions. For instance, the human cancer proteome may be analyzed in situ using MALDI TIMS [51] and TIMS image pixels may be correlated spatially with mass spectra clustering to highlight functionally similar regions in tumors [52], among other applications [53–55]. Detection sensitivity and data reproducibility [56–58] remain problematic for mass spectrometry.

On the level of whole cells, microfluidics is another high-throughput platform that enables the study of live cells, including unicellular organisms, in both population and single-cell settings by manipulating various chemical and physical conditions of the micro-environment. The functionality of microfluidics that enables a variety of experiments to be performed on a single device is also sometimes referred to as ‘lab-on-a-chip’. While microfluidics facilitates real-time observation of live cells under real and artificial conditions [32, 59], it remains challenging to design and manufacture appropriate chip features so as to capture the phenotype or behavior of interest from a given group of cells.

Minimum information standards

The development of minimum information standards is a community-driven effort to provide guidelines for sharing high-throughput –omic data in the challenge of large datasets. The goal of these standards is to guide researchers to systematically annotate experimental protocols and results. Different standards have been proposed for different types of data and different studies. For example, the MIAME (Minimum Information About a Microarray Experiment) and MIFlowCyt (Minimum Information about a Flow Cytometry Experiment) standards are based on specific data acquisition modalities [60, 61], while MIAMET (i.e. Minimum Information About a Metabolomics Experiment) and MINI (i.e. Minimum Information about a Neuroscience Investigation) are based on specific research areas [62, 63]. For REBMS, MIRIAM (Minimum Information Requested In the Annotation of biochemical Models) provides standards for describing and sharing computational models [64], while MIASE (i.e. Minimum Information About a Simulation Experiment) provides a framework to document in silico, or simulation, experiments performed using a model [65].

Some minimum information standards are now well-accepted by the community-at-large. For example, adherence to MIAME is required by a number of high-profile journals, including Nature and Science [66]. Additionally, some curated biological databases encourage data submissions adhering to these community standards; Table 1 presents a selected list of these databases. Accepted minimum information standards may also guide the development of other community standards; for example, accommodation of MIRIAM annotations in Systems Biology Markup Language (SBML) [67, 68]. However, because minimum information standards are developed by various interest groups, there is some uncertainty with regard to the appropriate standards for interdisciplinary research. The MIBBI portal (http://mibbi.org) makes an attempt to address this problem by providing a resource hub for minimum information standards. The MIBBI project aims to promote the accessibility, transparency and collaborative nature of interdisciplinary research by providing a central resource, from which standards can be tracked, compared, coordinated and adopted [61]. At the time of writing, 32 minimum information projects are registered in the MIBBI portal.

Table 1:

Resources for REBMS using –omic data

Type	Name	URL	Description
Genomic data	GenBank	www.ncbi.nlm.nih.gov/genbank	Annotated collection of all publicly available DNA sequences; international collaboration among USA, Europe and Japan
	EMBL Nucleotide Sequence Database	www.ebi.ac.uk/embl
	DNA Database of Japan (DDBJ)	www.ddbj.nig.ac.jp
Proteomic data	UniProt	www.uniprot.org	Protein sequence and function
	BRENDA	www.brenda-enzymes.org	Enzyme functional data
	SABIO-RK	sabio.villa-bosch.de	Biochemical reactions, kinetic equations with parameters and experimental conditions under which these parameters were measured
Metabolomic data	LIPIDMAPS	www.lipidmaps.org	Comprehensive resource for lipid biology, e.g. structure and lipid-associated proteins
Pathways	KEGG	www.genome.jp/kegg	Integrated resource for building blocks and functions of biological systems
	Reactome	www.reactome.org	Open-source, manually curated and peer-reviewed pathway database
System models	BioModels	www.ebi.ac.uk/biomodels-main	Peer-reviewed, published, computational models from systems biology
Portal sites	Pathguide	www.pathguide.org	Pathway- / molecular interaction-related
	ExPASy	expasy.org	Bioinformatics

DATA MINING AND SYSTEMS MODELING

Data mining and systems modeling are complementary efforts for REBMS. Data mining processes large datasets by extracting correlated data features, while systems modeling creates a unified theoretical framework, through time and space, by aggregating and associating relevant data features to predict other system behaviors.

Challenges and strategies for data mining

Data mining typically involves (i) exploratory analysis, i.e. discovering biologically interesting relationships in the data, (ii) feature selection, i.e. mining for biomarkers or interesting features and (iii) prediction, i.e. building statistical models for either diagnostic/prognostic applications or predicting system behavior. Exploratory analysis methods can reveal the natural organization of high-throughput data and identify patterns [69]. The most common unsupervised clustering method applied to high-throughput data is hierarchical clustering [70], others include k-means clustering [71], principal component analysis (PCA) [72] and bi-clustering [73]. For example, unsupervised clustering identifies groups of genes with similar expression patterns, or groups of samples with similar molecular profiles [74]. On the other hand, many high-throughput data experiments generate samples with known underlying grouping or clustering information. Thus, we can use this information to supervise the identification of features that are important for modeling. Many supervised feature selection methods exist for mining gene expression data, but these methods may be applied to high-throughput data in general [75]. Exploratory analysis and feature selection ultimately lead to prediction analysis, an important aspect of systems modeling. Regardless of the biological application (e.g. clinical diagnosis/prognosis or systems modeling), statistical modeling must be systematically evaluated to overcome challenges and avoid common pitfalls [76].

A well-known issue in –omic data mining is ‘curse-of-dimensionality’, where the number of biological samples is significantly smaller than the number of feature dimensions [77]. Especially for REBMS, the power of data mining is limited by this issue. Pooling several datasets across different studies for training to increase sample size and filling in missing data may improve data quality and quantity [78]. Extensive cross-validation during training is also proposed to overcome the limits of poor data such as small sample size and sample noise [79, 80]. In a recent issue of Nature Biotechnology, using DNA microarrays as a common high-throughput platform to measure gene expression, multiple challenges for data mining such as ‘batch effect’, ‘cross-platform difference’, ‘classifier selection’, ‘performance metric’ and ‘prevalence’ etc., and strategies to counter these challenges were studied by an international consortium consisting of 100 s of data mining researchers across 36 institutions. The outcome of this study is the publication of sensible ‘good-practice’ guidelines for each of these issues [78, 81–84].

Models must be predictive in nature in order to be scientifically useful [85]. Recently, omic-based predictive models have received increased attention in the scientific literature because of their potential for clinical applications [84, 86, 87]. That is, for high-throughput measurement such as genomic (or proteomic) expression, data mining is expected to lead to diagnostic and prognostic models that can predict disease state. Thus, another challenge in data mining is to ensure consistent performance of prediction models across both training and testing datasets (i.e. the cross-validation performance within training datasets is similar as the external-validation performance with testing datasets). Simon et al. and Quackenbush [76, 86] examine key steps and common pitfalls involved in building and evaluating predictive models. Simon stresses the importance of correctly estimating the accuracy of prediction models on future samples. This involves proper division of samples into training and testing sets before any analysis, so that none of the test samples are used in training predictors. Michiels et al. [88] reinforce this recommendation after re-analyzing several large cancer prediction studies. Their results show that many of these studies predict no better than random chance, and the selection of features greatly depends on the samples. They recommend a method of repeated random sampling to better estimate the mean and variance of prediction error. Besides, the process for building and evaluating predictive models, and identifying factors that affect predictive performance has been studied extensively [76, 84, 89, 90]. The observation is that, for genomic-based diagnostic models, prediction performance is difficult to evaluate. It involves either (i) retrospective evaluation after collecting an adequate number of future samples, or (ii) estimation of predictive performance using cross validation or other statistical sampling methods [91–93]. Despite the difficulties and pitfalls involved in building predictive models, -omic predictor models are feasible for diagnostic or prognostic in clinical use [94]. Once the features are extracted from the high-throughput data, they can be analyzed to build networks for REBMS.

Systems modeling—static network inference

Static network inference is the process of reconstructing the topology of biological systems to represent data features (e.g. genes, proteins or metabolites) as nodes and interactions as edges to predict new interactions. Static networks can be deterministic or probabilistic. However, they are only able to model a snapshot of complex dynamics of biological systems. The availability of extensive databases and sharing of high-throughput datasets has promoted genome-scale static network modeling. The modeling methods include correlation networks and Bayesian networks etc. The correlation network is one of the simplest models [95] and has low computational cost: two genes are predicted to interact if their expression profiles are highly correlated. Several other network inference algorithms such as ARCANE [96] and RELNET [97], which are based on mutual information have also been proposed. The limitation of correlation networks is that only instances of pairwise interaction are captured, and epistasis, where one gene interacts with several other genes, is not. Bayesian networks are also increasingly used for static network inference. The power of Bayesian networks comes from the provision to capture uncertainty in any contributing factor, including model parameters, hidden variables and observations, using probabilistic methods. This makes Bayesian inference a useful framework to model inherent uncertainty in biological data, which may be a result of biological or technical variability [98]. Bayesian networks may also be used to model epistasis where standard statistical tools such as multivariate logistic regression models have failed [99, 100].

In recent disease-centered applications of static network inference, Sebastiani et al. [100] used Bayesian networks to identify SNPs predictive of stroke in sickle cell patients; Tran et al. [101] used Bayesian networks in combination with copy number variation analysis to identify potential driver genes in breast cancer; and Carro et al. [102] used the mutual information-based algorithm ARACNE to identify key transcriptional regulators of the mesenchymal cellular phenotype in gliomas, which is associated with more aggressive disease and poorer prognosis. In another study, Faith et al. [103] used both their own microarray data and data from nine other publications to perform genome-scale network inference on Escherichia coli, a model organism for which substantial information about transcriptional regulatory interactions is known. Existing data was thereby used as a test-bed for evaluating the performance of network inference algorithms. The authors showed that at 60% precision, the predictions of the best-performing algorithm, based on mutual information, included 338 known interactions documented in the database RegulonDB and 741 novel interactions. Of the predictions at all precision levels, 268 interactions were experimentally tested via ChIP and 21 were confirmed. In another recent study, Zhu et al. [104] performed large-scale gene network inference by integrating genotypic, gene expression, transcription factor-binding site and protein–protein interaction datasets in yeast, which is one of the few model organisms for which comprehensive high-throughput datasets of multiple types are available. The authors showed that in terms of predicting known causal regulators in gene sets, the Bayesian networks constructed using multiple data types were superior to the Bayesian network based on gene expression data only; in addition, five previously unknown interactions made by the Bayesian network built using all four data types were experimentally confirmed.

While static network inference has produced promising results, it still faces several challenges: (i) when performing genome-scale analysis, the number of possible parameters to be found can greatly exceed the number of data points, which leads to high computational complexity; (ii) limited data availability from the available measurement may require simplifying assumptions about the complex, non-linear biological mechanisms [105], and (iii) algorithm performance can vary greatly, even for the same algorithm applied to different datasets. Marbach et al. [106] recently published the results of the DREAM3 in silico community-wide challenge, in which simulated network data was used to perform double-blind tests of gene network inference methods. Notably, a large proportion of the tested methods (11 of 29) performed poorly, with results not significantly improved over random guessing. Additionally, systemic errors in prediction were associated with ‘fan-in’, ‘fan-out’, ‘cascade’ and ‘feed-forward loop’ motifs, which describe situations including coregulation and combinatorial regulation. Similar variation in performance was linked to the properties of different biological networks by Ooi and Phan [107], who recently used known gene– and protein–scale interactions from enriched pathways of different classes in keloid fibroblasts to compare the performance of ARACNE [96] and BANJO (Bayesian Network Inference with Java Objects) [108]. Their findings indicated that the transcriptional networks were better suited to network inference analysis [107]. These results indicate that there are substantial opportunities in researching static network inference of REBMS such as to develop increasingly practicable methods that take into account fundamental biochemical mechanisms and motifs.

Systems modeling—network dynamics

REBMS is not complete without modeling system dynamics in addition to static network topology. A fully deterministic model that can be used to capture the complex dynamics of biological systems is the Boolean network. It is noteworthy to mention that models can be categorized in different types. They can be considered as static versus dynamic or deterministic versus probabilistic. For example, the Bayesian network is a probabilistic network that is static, but the Boolean network is a deterministic rule-based model that can capture dynamics. The Boolean network model [109] is a major effort in gene regulatory network inference, where gene expression is quantized to only two states: ON and OFF. The Boolean state of each gene is functionally related to other genes using Boolean functions. This assumes a fully deterministic environment that does not account for innate uncertainty in biological systems, data and model selection [110]. Beside this, the use of Boolean networks for network topology inference is limited by two major challenges: (i) gene expression cannot be described adequately by only two states and (ii) inference of Boolean functions for a large number of genes is computationally expensive. Introducing probabilistic Boolean networks was a major breakthrough to capturing not only dynamics of biological systems but also uncertainty in data, model selection and biological systems [110]. As the probabilistic Boolean network is the ‘unrolled’ version of Boolean networks in time, the dynamic Bayesian network [111] is the counterpart for the standard Bayesian network. Besides probabilistic Boolean networks and dynamic Bayesian networks, Ordinary Differential Equations (ODEs) have been used as an alternative to model gene regulatory networks. Tyson et al. [112] and Goodwin [113] have used non-linear differential equations to reverse engineer a kinetic model of gene regulation process. There are some issues in modeling biological systems dynamics using linear and non-linear ODEs. First, these models need a detailed a priori knowledge about the form of reaction rate functions usually not available in practice. Second, the limited number of samples commonly available for these studies makes inference of parameters used in reaction mechanisms very difficult. Piecewise-linear differential equations have been proposed to alleviate the aforementioned problems for modeling reaction mechanisms [18, 114]. Savageau [20, 115] has proposed power-law simplification of non-linear differential equations as an alternative approach. Furthermore, the stochastic master equation has been used to capture the delays induced by various reaction rates of biochemical processes [21].

Because reaction rate parameters are critical to these traditional models, the methods for modeling systems dynamics are sometimes reduced to parameter estimation problems. Examples of reaction rate parameters include rate constants for mass action kinetics [116], kinetic orders in biochemical systems theory [20, 115], stoichiometric coefficients and flux rates for flux balance analysis [117] and control coefficients and elasticities for metabolic control analysis [118]. Parameter estimation cannot be avoided and can easily lead to the problem of data over-fitting. In cases of over-fitting, parameters are overly sensitive to data noise, and fitted values differ clearly from measured values [119, 120]. In biological systems measured by high-throughput ‘-omics’ technologies, significantly more parameters are required than measurements available in terms of temporal points and in terms of number of samples. Thus, the parameter estimation problem is worse. In addition, dynamic biochemical systems need to be able to predict the behavior of real systems and to extrapolate components of the system at future time points.

Based on the observation that certain emergent properties in biological systems, namely robustness, resemble properties of engineering systems, this has led to a recent (or renewed) wave of efforts in REBMS in terms of cybernetics and control theory [10, 11, 121]. Cybernetics is the study of structural complexity in animal and machine that enables communication and control [122] and is closely related to control theory. In the context of biological systems, this approach was first applied to study organ systems in physiology, for instance, in circulation [123], immunology [124] and the central and peripheral nervous systems [125, 126]. More recently, cybernetics has been applied to study metabolic pathways [17, 127] at molecular levels. While cybernetics emerged as the science for effective organization within systems, control theory was developed to guide the behavior of dynamic systems. On the other hand, control, i.e. regulation of motion, is sufficient to achieve system stability and robustness. For instance, bacterial chemotaxis in E. coli is an example of relatively simple but robust cell behavior that may be modeled in terms of classical control theory. Specifically, exact adaptation in bacterial chemotaxis is observed to be robust and not affected by changes in protein levels [128, 129]. Consequently, such behavior may be modeled as the result of single-variable integral control that ensures convergence to steady state without error [22], which is a global systems property. Furthermore, details of possible mechanisms that enable the robust behavior can be probed using time-varying stimuli [130]. Modern control theory may also be useful to model regulatory behaviors in larger metabolic pathways [19]. Although recent studies do show promise in the use of cybernetics and control theory to model dynamics in REBMS, at the same time, the key challenge remains: conditions for robustness established using cybernetics and control theory are sufficient but not necessary. Thus, substantial experimental validation is required to prove the uniqueness of proposed system models for REBMS.

PROGRESS IN REBMS: TWO CASE STUDIES

In this section, we present two case studies to illustrate recent progress in the field of REBMS. In the first case study, we discuss the data mining and analysis of genomic interactions in breast cancer by integrating gene expression, SNP, CNV and DNA methylation data. In the second case study, we discuss the systems modeling of robust adaptation in bacterial chemotaxis by integrating bottom–up and top–down approaches. Ultimately, for REBMS using –omic data, integrating various data and methods is necessary and most likely fruitful.

Integrative analysis of genomic interactions in REBMS of breast cancer

To date, there are several studies that attempt to use integrative omic analysis to understand human diseases, most of which were focused on cancer due to the availability of disease tissue besides blood samples [131–137].

A recent published paper by the Cancer Genome Atlas Research Network [137] conducted integrative analysis of the single nucleotide polymorphism (SNP), copy number variation (CNV), expression profiling of coding and non-coding RNAs and global DNA methylation profiling datasets with 309 clinically annotated high grade serous ovarian cancer patients and an additional 180 patients with all but SNP data. The authors mapped the original DNA reads to the human genome, excluding duplicate reads, used CHASM [138] and MutationAssessor to identify functional mutations, used GISTIC analysis to extract quality copy number data, used consensus clustering [139] to analyze mRNA, miRNA and methylation data, used HotNet [140] to analyze protein–protein interactions and used PARADIGM [141] to estimate differential integrated pathway activity. They found that late stage ovarian cancer tumors have specific mutations of interest at both high and low prevalence; TP53 mutations were found in 96% of tumors and somatic mutations in nine genes including BRCA1, BRCA2, CDK12, NF1, RB1 and were statistically recurrent, despite low prevalence. They went on to identify 113 significant copy number aberrations in DNA and 168 genes with promoter methylation events. Furthermore, they characterized four subtypes of ovarian cancer transcription, three subtypes of miRNA and four subtypes of promoter methylation. Survival duration was found to be associated with a transcriptional signature, explaining the impact of BRCA1/2 and CCNE1 aberrations on tumor survival. Lastly, in about half of the tumors considered, they identified defects in the homologous recombination pathways, suggesting a role for Notch and FOXM1 in the pathophysiology of serous ovarian cancer. The outcomes of this case study show that in the era of high-throughput data, it is possible to integrate different data to derive a holistic view of the biomolecules for REBMS.

Modeling robust adaptation in bacterial chemotaxis

The E. coli chemotaxis network is a model system that has been successfully reverse engineered in terms of (i) establishing a framework to describe robustness of a biological system and (ii) elucidating the mechanisms that may be responsible to achieve the robustness. That is, in bacterial chemotaxis, it is observed that the steady-state tumbling frequency in a homogeneous ligand environment is insensitive to the value of ligand concentration [129]. Previously, researchers thought that the rates of specific network reactions need to be fine-tuned in order to maintain the same steady-state behavior under different conditions [142]. However, more recent studies suggest that adaptation is a robust network property that is independent of particular rates of reaction [128, 129] and the robust adaptation in chemotaxis may be the result of integral feedback [22], which is a common setup in engineering modeling that ensures the convergence to the desired steady-state.

The robust adaptation in bacterial chemotaxis by feedback has led to the integration of bottom–up and top–down approaches in systems modeling. From a bottom–up perspective, individual network components (i.e. biochemical reactions) are quantified and characterized using rates of reaction. Because these individual reactions constitute the network, the observed system properties arise directly from adding all pieces of biochemical reactions together with parameter estimation. On the other hand, from a top–down perspective, the system demonstrates robust behavior even when the system input is perturbed. Thus, the bottom–up approach may be viewed as an ‘open-loop’ system based on mass action modeling and parameter estimation, while the top–down approach may be viewed as a ‘closed-loop’ system. In this case study, both bottom–up and top–down approaches are integrated to describe robust adaptation in bacterial chemotaxis.

MODEL VALIDATION AND ANALYSIS-BY-SYNTHESIS: DUAL ROLES OF SYNTHETIC BIOLOGY

Recently, research in synthetic biology has emerged as an exciting opportunity for model validation and analysis-by-synthesis, serving dual roles in REBMS (Figure 2). Synthetic biology, i.e. the creation of novel biological components (macromolecules), devices (pathways) and systems (organisms) by modifying existing ones or using artificial materials, is promising but also challenging because of the unpredictability and poor reproducibility of experimental results.

Figure 2.

Dual role of synthetic biology in reverse engineering biomolecular systems (REBMS): (left) to test and validate models as a result of upstream mining and modeling, i.e. as the third step in a conventional workflow and (right) to create biomolecular components, devices and systems for analysis-by-synthesis, i.e. synthetic biology is the driver for REBMS

Analysis-by-synthesis is gaining more interest in REBMS because of its potential to design, manufacture and observe viable biomolecular components and devices within predictive, scalable and reproducible systems. It is an active process, common in the field of speech and communications, to analyze signals where only and all the signals to be analyzed can be generated [143]. Biomolecular components and devices refer not only to individual biochemical metabolites but also to stable self-contained modules. Then, biomolecular systems may mean any process, computational or otherwise, that is the result of clever combinations of biomolecular components and devices. Furthermore, such systems may also give rise to other components and devices. Thus, in the context of synthetic biology, the premise of analysis-by-synthesis is to be able to generate a representative variety of artificial biomolecular components, devices and systems.

Successful methods for generating biomolecular components and devices need to be scalable with fast growing high-throughput technologies. For instance, ontology mining tools such as GoMiner [144] are scalable to handle large ontology datasets so as to mine for biomolecular components and devices from linked ontology terms. On the other hand, the synthesis of artificial biomolecular components and devices using equivalent circuits in systems modeling and synthetic biology is not yet scalable. The efficacy and efficiency of assembling metabolites is directly correlated to the dexterity with which researchers may manipulate the biochemistry of the particular metabolites. Additionally, because of the diversity and heterogeneity of biomolecular systems and signals, individual (groups of) efforts in analysis-by-synthesis must be communicated effectively to make any collective progress as a community and to create impact through clinical translation. Examples include the IUPS Physiome Project (www.physiome.org.nz), where biosensors, bioreactors and multi-scale models for diagnosis and treatment. Thus, through an iterative process of data acquisition, mining, modeling and validation, emergent properties of reverse engineered biomolecular systems predicted by mathematical models may be verified and validated in synthetic biology for REBMS.

Synthetic biomolecular components (macromolecules), devices (pathways) and systems (organisms)

The primary challenge of synthetic biology for REBMS is to create and reproduce synthetic biomolecular components (macromolecules), devices (pathways) and systems (organisms) that can replicate the functions of its natural counterparts in vivo. For example, synthetic enzymes that mimic metabolism have been created [145] even though most results to date are relatively simple proof-of-concept creations; toehold reactions have been created to implement DNA circuit control mechanisms without enzymes [146], which has potential uses in automatic drug delivery; and synthetic pathways such as gene networks have also been created to mimic pathways for pathogenic virulence [147]. However, the inability to manufacture and splice DNA reliably, together with the need for extensive post-synthesis modification, has hindered the synthesis of larger, more intricate gene networks. On the other hand, a number of advancements have been made in terms of regulating gene networks, for instance, by modulating the expression of specific proteins [148] with synthetic ribosome-binding sites [149] and selectively introducing mutations in the TATA boxes of promoters to reduce unwanted network interactions (or noise) [150]. At the level of whole cells, including unicellular organisms, a number of experiments have recently demonstrated the potential of synthetic biology to modify the existing genome of bacterial cells to perform novel functions, e.g. to detect light–dark edges in images [151], implement a genetic clock [152] and inhibit biofilm formation [153]. More recently, Gibson et al. [154] reported the creation of a bacterial cell controlled by a synthetic genome, which opens the possibilities of creating minimal cells with desired properties [25].

The reliability and reproducibility of synthetic biology may be improved by in silico modeling and simulation, such as in computational structural biology for synthetic macromolecules to simulate folding and tertiary interactions. Blake et al. [23] described a method for scalable, cost-effective, sequence-independent assembly to construct large DNA components. For synthetic pathways, there are a variety of simulation software using BioBricks to construct standardized models of biochemical networks to predict kinetic outcomes [24], such as SynBioSS [155], Eugene [156] and TinkerCell [157]. While computer-aided design (CAD) tools may be useful to design and simulate synthetic macromolecules and pathways in silico, the challenge remains to: (i) validate such designs experimentally and (ii) close the simulation-experimentation information loop by incorporating CAD knowledge to improve synthetic biology in practice. Crosstalk, i.e. noise, can also occur between host and synthetic circuitry that violates the modularity of components and devices in expressing synthetic biomolecular systems through host organisms. This problem is difficult to fully address because of the complexity of intracellular signaling, but an optimization protocol, which was tested in E. coli, has been developed to allow targeted modification of existing circuits that simplifies design and testing [27]. Wang et al. [28] also developed multiplex automated genome engineering (MAGE), which enables large-scale cell programming and evolution.

The challenges and opportunities make synthetic biology an exciting topic for REBMS, where theoretical outcomes from data mining and systems modeling may be verified and validated directly through analysis-by-synthesis. While there is some progress toward addressing the overall key challenge of model validation directly through synthesis in terms of macromolecules, devices and systems, issues remain to ensure the fidelity, reproducibility and scalability of such synthetic biomolecular components, devices and systems [24–26, 28, 158].

Synthetic biology: a case study

Recent breakthroughs in synthetic biology made it possible to synthesize and implant an entire genome to create living cells. Specifically, a bacterial genome can be synthesized and transplanted to another bacterium [154]. This bottom–up approach involves constructing a very small genome and re-engineering a reduced version of the entire genome for implantation and viability testing. In this study, Gibson et al. resequenced the Mycoplasma mycoides genome and disassembled it into 1078 overlapping cassettes of 1080-bp long. Then they implanted the chemically synthesized cassettes into yeast, whose DNA-repair enzymes linked the short sequences together. These generated medium-sized stings of DNA were then transferred into E. coli and back into yeast. By repeating this procedure three times, a complete genome of M. mycoides, comprising 1.08 Mb was constructed.

The complete M. mycoides genome was then transplanted into another bacterium, namely Mycoplasma capricolum. Although few genes were deleted or disrupted in this genome insertion, they observed that properties of the transplanted M. capricolum are the same as M. mycoides. While this first construct of a living cell is just a proof of concept, the methods and technologies developed hold great promise for many other applications.

CONCLUDING REMARKS

Data-driven mining and design-driven modeling have led to significant progress in REBMS using high-throughput –omic data. In addition to theoretical analyses, the next step is to combine analysis with synthesis through synthetic biology. Research in synthetic biology is shown to serve dual roles for REBMS in model validation and analysis-by-synthesis. Thus, the overall key challenges in REBMS are to (i) integrate heterogeneous biochemical data at molecular-level resolution for data mining, (ii) combine top–down and bottom–up approaches for systems modeling and (iii) validate theoretical system models directly by experiment, in particular synthesis (Table 2). While significant progress has been made toward addressing specific problems, key issues remain to (i) increase the accuracy of feature extraction in data mining, (ii) improve the simulation and abstraction of emergent system properties in systems modeling and (iii) ensure the fidelity of assembled macromolecules and pathways in synthetic biology. By incorporating data mining, systems modeling and synthetic biology to create an integrated and iterative research pipeline, we are hopeful that these remaining issues in REBMS will be overcome.

Table 2:

Summary of challenges, progress and remaining issues and opportunities for REBMS

	Key challenge	Progress	Issues/opportunities
Data mining	Integrate and analyze high-throughput data	Minimum information standards for data storage and retrieval [60–63, 65, 66] Data combination to fill in missing values and increase sample sizes [12, 82, 83] Iterative cross-validation [77, 80]	Increase accuracy of feature extraction
Systems modeling	Use top–down with bottom–up approaches	Static network topology inference using correlation, Boolean and Bayesian networks [13, 95, 97–99, 110, 111, 159–162] Dynamics modeling using bottom–up approaches, e.g. probabilistic and dynamic Bayesian networks, mass action, flux balance and S-systems [18, 20, 21, 114, 115, 118, 163, 164] Dynamics modeling using top–down approaches, e.g. cybernetics, classical and modern control theory [19, 22, 121–127]	Improve integration of top–down and bottom–up models
Synthetic biology	Validate systems models	Macromolecule synthesis [23, 26, 145, 146] Gene pathway simulation using CAD tools [24, 148–150, 154–157] In vivo artificial organisms using synthetic genetic circuits [25, 27, 151–154, 165]	Ensure fidelity of assembled synthetic components and devices

Key Points.

• To deal with high-throughput data in REBMS, data-driven mining approaches can extract patterns and features for downstream modeling.

• Motivated by similarities between biological stability and engineering robustness, design-driven approaches can simulate and abstract emergent properties of complex biomolecular systems.

• Challenges in data analysis and modeling for REBMS may be overcome by integrating heterogeneous datasets for mining and combining both top–down and bottom–up approaches for modeling.

• Synthetic biology is an emerging field that proves to be an exciting opportunity for REBMS, serving dual roles in validation and analysis-by-synthesis.

FUNDING

This work was supported in part by grants from the National Institutes of Health Bioengineering Research Partnership R01CA108468, Center for Cancer Nanotechnology Excellence U54CA119338, and 1RC2CA148265; Georgia Cancer Coalition Distinguished Cancer Scholar Award to MD Wang, and grants 5R21DA025168-02, 1R01HG004836-01 and 4R00LM009826-03 to G. Alterovitz.

Biographies

Chang F. Quo received his BSc in Biomedical Engineering from the Georgia Institute of Technology in 2005 and is currently a doctoral student in Bioengineering at the Georgia Institute of Technology. His research is focused on the use of modern control theory to model regulation in metabolic systems.

Chanchala Kaddi received BSc in Biomedical Engineering from the Georgia Institute of Technology in 2008 and is currently a doctoral student in Bioengineering at the Georgia Institute of Technology. She is an NSF Graduate Research Fellow. Her research is focused on high-throughput data analysis for tissue imaging mass spectrometry.

John H. Phan, PhD, received BSc in Computer Engineering from the University of Oklahoma in 2003 and PhD in Biomedical Engineering from the Georgia Institute of Technology and Emory University in 2009. He is currently a Postdoctoral Fellow at Emory University. His research is focused on knowledge-driven mining of high-throughput biological data, high-performance and grid computing and translational bioinformatics.

Amin Zollanvari, PhD, received PhD from Texas A&M University in 2010. He is currently a Postdoctoral Fellow at the Harvard-MIT Division of Health Sciences and Technology. His research is focused on modeling of complex diseases and behaviors using genome-wide association data, Bayesian theory, statistical pattern recognition, asymptotic analysis of statistics in large dimensionality space and ontological mapping and inference.

Mingqing Xu, PhD, is an Full Professor of Genetics at Bio-X Institutes jointly established by the Chinese Academy of Sciences and Shanghai Jiao Tong University. He is joint professor at Shanghai Jiao Tong University School of Medicine. He earned his PhD degree in Human Genetics at Shanghai Jiao Tong University in 2005 and then did his Postdoctoral training in Statistical Genetics and Epidemiology at Harvard University. He has been serving as Director of the Translational Research Center for Integrative Omics at Bio-X Institutes of the Chinese Academy of Sciences and Shanghai Jiao Tong University since 2011. His research interests are focused on translational omics and epidemiology, including high-throughput omic data analysis and its application to personalized medicine; statistical genomics; and environmental and molecular epidemiology in human complex diseases.

May D. Wang, PhD, is a tenured Associate Professor in the Wallace H. Coulter Department of Biomedical Engineering jointly established by Georgia Institute of Technology and Emory University in Atlanta, Georgia. She is a Georgia Cancer Coalition Distinguished Cancer Scholar, and has served as Director of Biocomputing and Bioinformatics Core in Emory-Georgia Tech Cancer Nanotechnology Center focusing on Bio-Nan-Info Integration for Personalized Oncology. Dr. Wang has also served as a co-Director in the Georgia Tech Tissue Imaging Mass Spectrometry Center since 2005. Her primary research interests are translational bioinformatics, imaging informatics, and health informatics for personalized systems medicine: including high-throughput −omic data quality control and analysis for clinical biomarker identification; tissue imaging informatics; pathway modeling; and delivering personalized health informatics solution by mobile health.

Gil Alterovitz, PhD, is an Assistant Professor at the Harvard/MIT Health Sciences and Technology Division Children's Hospital Informatics Program (CHIP). He is affiliated with the Department of Electrical Engineering and Computer Science at MIT and Harvard Medical School Partners’ Center for Genetics and Genomics (HPCGG). His research interests involve development of novel, interdisciplinary approaches that bridge engineering and medicine. Specifically, he is involved in developing methods for studying biological networks and signal processing within proteomics.