Key Laboratory of Systems Biology, SIBS-Novo Nordisk Translational Research Centre for PreDiabetes, Shanghai Institutes for Biological Sciences, Chinese Academy of Sciences, Shanghai, China
L. Chen, Key Laboratory of Systems Biology, SIBS-Novo Nordisk Translational Research Centre for PreDiabetes, Shanghai Institutes for Biological Sciences, Chinese Academy of Sciences, Shanghai 200031, China
Extensive studies have been conducted on gene biomarkers by exploring the increasingly accumulated gene expression and sequence data generated from high-throughput technology. Here, we briefly report on the state-of-the-art research and application of biomarkers from single genes (i.e. gene biomarkers) to gene sets (i.e. group or set biomarkers), gene networks (i.e. network biomarkers) and dynamical gene networks (i.e. dynamical network biomarkers). In particular, differential and dynamical network biomarkers are used as representative examples to demonstrate their effectiveness in both detecting early signals for complex diseases and revealing essential mechanisms on disease initiation and progression at a network level.
The identification and further utilization of biomarkers are important in the development of modern medicine. Generally, a biomarker can be any actual element type, such as a clinical index , mutation pattern , gene expression [1, 3], antibody presence [4, 5], metabolic profile and image [6-9] or biomolecular structure [10, 11]. In biomarker studies, there are three principal topics of current research. The first is the biological element itself; frequently studied elements include clinical tissues, gene transcripts, translated proteins and outcome metabolites. The second is the measurable signal of a biological element. For example, expression, phosphorylation, structure–affinity and image may all reflect the particular characteristics of some molecule markers (i.e. proteins) in a disease occurrence. The third is the combination of different indicative biological elements (e.g. a group or network of elements), and is usually composed of several of the above-mentioned biological elements and their available measurements.
Focusing on high-throughput gene expression data, we summarize recent research on gene-related biomarkers, including gene and gene network biomarkers. In fact, biomarkers identified at the gene level have been widely studied, e.g. gene signatures recognized from comparisons between cancer and normal/adjacent samples [12-15] (or treated samples ), and many of them have also been successfully applied in pathogen mechanism indication, disease diagnosis and treatment prognosis.
In contrast to traditional biomarker research in an isolated and static manner, two problems have emerged from recent gene biomarker studies. The first is how to find a biomarker that can reflect weak but effective (or causal) gene associations with diseases [17-19]. The second is how to use biomarkers for early disease diagnosis, i.e. to detect the critical transition before the appearance of clinical symptoms during disease initiation or gradual disease progression [20, 21]. Solving these two problems is challenging and requires the characterization of diseases with biomarkers in a systematical and dynamical manner, particularly by exploring the rich information obtained from high-throughput analyses. To address the first problem, the identification of reliable gene biomarkers requires the extraction of a group of correlated pathogen genes and inference of their disease-specific relationships. Whereas, for the second problem, detection of the early signals of a disease requires extraction of dynamical information on genes and determination of early genes' pathogen change, which is equivalent to revealing when and which genes cause critical functional reorganization  from the normal to disease status/state during disease occurrence and deterioration.
To tackle these problems in practice, a number of novel theories and technologies have been proposed and implemented [20, 23-25] related to network biomarkers and further DNBs, which differ greatly from traditional biomarkers (e.g. individual genes or molecules). These biomarkers are able to distinguish the disease state (e.g. patient status after the appearance of clinical symptoms) or even the pre-disease state (e.g. patient status before the appearance of clinical symptoms) from the normal state in a systematical and dynamical manner. On the one hand, identifying a network biomarker (e.g. sub-network marker) [24, 26] or a module biomarker (e.g. co-expression marker)  requires the statistical inference of a set of strongly correlated genes from a group of normal or cancer samples, forming a gene network. The topological structure of such a network would be useful for detecting genes with significant ability in disease discrimination and interpretation. These network-based biomarkers are identified mainly by considering the (static) conservation or consensus aspect of a gene module (i.e. a group of interconnected genes) in the normal or disease states, thereby distinguishing normal and disease samples in a reliable way. On the other hand, detecting a dynamical network biomarker (DNB) [20, 21] or a dynamical network of biomarkers exploits dynamical or fluctuation information among samples at different time points, e.g. clinical stages. These novel dynamical biomarkers may signal emergence of the critical transition of a biological system before the appearance of clinical symptoms, i.e. distinguish between normal and pre-disease samples rather than disease samples, thereby achieving early diagnosis at the pre-disease stage .
To briefly describe the state-of-the-art research on gene biomarkers (Fig. 1), we summarize the various methods used on biomarkers below, namely: single and differential gene expression (Fig. 1A); gene set and differential enrichment (Fig. 1B); gene and differential networks (Fig. 1C); dynamical gene network and differential activity (Fig. 1D); and finally, the conclusions drawn from this review.
Single and differential gene expression
A single gene can display dysfunction in multiple ways, i.e. mutation of the coding sequence, differential regulation of mRNA expression, differential degradation of tRNA translation or differential modification of the encoded protein. All these potential changes in a single gene and its products can provide a simple way to judge the risk of human disease (Fig. 1A). Because the analysis of gene expression is becoming a high-throughput and inexpensive approach, differential gene expression analysis has been widely applied not only to identify gene biomarkers, but also to confirm the clinical utility of single genes. For example, single gene biomarkers have been already identified in cancers, e.g. the BRCA1 and BRCA2 genes for breast cancer [28, 29] or the IL28B gene for liver cancer [30, 31].
With rapid advances in next-generation sequencing technology , analysis of a single gene mutation might be more attractive in future disease research because of its clear association (e.g. causal relation) with disease initiation and progression. Moreover, driver mutations identified in a genome-wide association study , as well as the identification of recurrent somatic mutation networks , are revealing the pathogenic physical changes of many genes to be disease drivers rather than consequent results.
Gene-set and differential enrichment
Although the use of individual gene biomarkers can be readily applied for clinical applications, it has been primarily adopted in the analysis of hereditary diseases. This is because many well-known complex diseases (e.g. cancer and metabolic diseases) are usually induced not by individual genes, but by many associated genes and their dysfunctions involved in divergent biological mechanisms [18, 35, 36]. Using one gene set instead of a single gene is a main approach for biomarker identification in the current study of complex diseases . The differential expression of those genes in the identified gene-set biomarker (i.e. a set of marker genes) can be directly used to classify a person's health and disease states or potential pre-disease state [1, 38, 39]. In order to further enhance the robustness or consensus of gene-set biomarkers in dissimilar disease cohorts, a scheme has been adopted to mine a common discriminative gene group from multiple datasets .
Another alternative method to improve the accuracy and robustness of gene-set biomarkers is to use differential enrichment rather than differential expression to quantify and distinguish between the combination status of a group of genes corresponding to normal or disease samples [40-44]. This enrichment-based approach is also known as the activity-based method [43, 45], which deliberately makes full use of the significant extent and number of differentially regulated genes in one gene set (Fig. 1B). In other words, this type of method assumes that there will be many differentially regulated genes in the same gene-set biomarker that indicates the pathogenic genotypes of patients, although the differentially regulated genes of each patient may be different. As a representative method, condition responsive genes are defined as a subset of pathway genes whose expression combination has the greatest ability to discriminate the disease phenotypes . For each pathway, the activity level is simply and directly summarized from the gene expression levels of its condition-responsive genes. Meanwhile, a method has been developed to promote the inference accuracy of such pathway activity by filtering out potential noise before estimating the pathway activity ; this is known as a denoising algorithm using relevance network topology. The key steps in this approach include: (a) selecting a subset of genes correlated to the response to differential pathway activation across tumours; (b) building a relevance correlation pruned network on selected genes whose edges reflect correlations consistent with the prior information, e.g. synthetic perturbation signature; (c) inferring pathway activation over the pruned network using a topology-based metric, which emphasizes the importance of hub genes in the network by giving them greater weight. In addition, another method known as signaling pathway impact analysis further considers the (directed) complex gene interactions built in pathways, together with differential gene expressions when inferring pathway activity, although this approach estimates the pathway activity for a group of samples with the same phenotype rather than for individuals .
Actually, a set of genes is more reliable and has greater power for disease discrimination than a single gene for disease indications . In fact, it has been shown that differentially regulated pathways are able to serve as better biomarkers than single genes and the heuristically designed activity scores of module genes are also useful in the study of cancer causality [48, 49].
Gene and differential networks
As stated above, gene-set-based methods usually focus on each gene's expression individually, but disregard correlations among the expressions of a set of genes (or a gene-set-related network). Along with the development of computational systems biology studies, exploring the information from a network of genes [e.g. gene regulatory network or protein–protein interaction (PPI) network] and how it is altered is becoming an attractive approach to decipher the occurrence and progression of complex diseases [20, 50-54]. In general, there are two main categories of network-based gene expression analysis: weighted gene co-expression network analysis and the PPI-based co-expression network [23, 24, 55].
The first approach measures correlations among all the gene pairs and extracts modules based on the density of gene connections , although it is difficult to distinguish detailed topological connections among genes in a co-expression module because of the high connection density as gene co-expression. This method has been used to find conserved gene groups in many comparative studies [23, 56, 57]. In particular, derived from weighted gene co-expression network analysis, the co-expression modules for HBV/HCV-induced HCC (i.e. Hepatitis B virus or Hepatitis C virus induced Hepatocellular carcinoma) revealed distinct progression patterns from hepatitis infection to hepatocellular carcinoma . After extracting modules of samples with different virus infections (HBV or HCV) at various disease stages (I–IV), the functional enrichment of those modules displayed significant preference for differentially regulated functions during HBV–HCC and HCV–HCC progressions.
By contrast, the second method usually constructs a supervised network based on known physical interactions (e.g. a PPI network) mapped with gene co-expressions, although it may result in some false-negative interactions [24, 55]. In many representative methods, Ideker et al. first introduced an effective algorithm to identify connected regions of a global network that show significant changes in expression under a particular subset of conditions, although it disregards the pairwise gene similarity . Later, Ulitsky et al. proposed module analysis via the topology of interactions and similarity sets, which is based on a probabilistic model to divide functional modules as jointly active connected sub-networks on PPI combined with gene expressions . This method uses a likelihood score to evaluate the modularity of a group of modules or sub-networks, but several heuristic algorithms are adopted to search the best group of functional modules because of the intractable computation complexity for an exact optimization . Recently, Ulitsky et al. presented another method as dysregulated gene-set analysis via sub-networks, which is based on the set cover problem for identifying connected gene sub-networks significantly enriched for genes differentially regulated in disease specimens . It uses two numbers to evaluate the target module: both the number of differentially regulated genes in a module and the number of samples with differentially regulated genes are as large as possible. Although this approach still uses heuristic algorithms, not guaranteeing optimality to search the best gene sub-network, it has a provable performance, which is different from pervious method [59, 60]. Furthermore, Dittrich et al. have designed a new additive score (function) defined by network genes considering statistical interpretation and data integration, which searches a sub-network with the best score. This approach implemented in bionet software uses an integer–linear programming model based on the prize-collecting Steiner tree problem to calculate provably optimal sub-networks [61, 62]. As an application example of a PPI-based co-expression network, epigenomic data, gene expression data and PPI networks have been integrated to identify module biomarkers for a colorectal cancer study , which include differentially methylated genes possibly serving as causal or driver genes for colorectal cancer.
Dissimilar to the above methods with correlation networks, differential networks give another view of control-case (or normal-disease) studies [44, 63]. This type of approach considers a virtual and global network consisting of differential interactions, in which a differential interaction means that two genes have a strong correlation in the control condition, but a weak correlation in the case condition, or vice versa. In a study of gastric cancer development and progression , several differential networks have been constructed, and the common differential sub-network is actually a dynamical core module during different clinical stages of gastric cancer. The genes in such a core module have shown their superior performance in distinguishing normal and disease samples , so that they can serve as biomarkers for gastric cancer and have also been validated on independent datasets.
In fact, the local features of a differential network have also been widely studied to characterize cancer. Taylor et al. observed two types of important proteins in the biochemical interaction structure, e.g. the intermodular and intramodular hubs, which decide the modularity of human interactome . Taylor et al. further illustrated that the altered modularity might be used for cancer prognosis: (a) according to differences in the correlation of co-expression of the hubs with their interactors/partners, a network signature was identified as network hubs that are significantly different between patient groups; (b) an affinity propagation algorithm applied such a network signature to train and predict patient outcome using cross-validation . Meanwhile, Teschendorff et al. analysed the effect of the randomness of local gene expression patterns . Based on integrated PPI–mRNA expression networks in the context of cancer genomics, they detected new indicators of a metastatic cancer phenotype: (a) according to the expression correlations between interacting proteins, a constrained weighted network is defined as a stochastic information flux matrix; (b) based on this stochastic matrix, a new entropy measurement for a local pattern of information flux is proposed to quantify the degree of randomness corresponding to single genes; (c) it is then observed that metastatic breast cancers can be characterized by a small, yet significant, increase in the degree of randomness in local expression patterns . Furthermore, West et al. proposed differential network entropy based on a heat kernel stochastic matrix instead of general stochastic information flux matrix, and detected more characteristics of local expression pattern indicating cancer hallmarks: (a) cancer cells tend to be characterized by an increase in network entropy; (b) differences in the gene expression between normal and cancer tissues are anti-correlated with local network entropy changes; (c) in particular, many genes driving cell proliferation in cancer cells or encoding oncogenes are usually associated with the reductions in network entropy .
Recently, another new approach has proposed the construction of a differential expression network by complementarily considering both differential expression of genes and dysfunctional correlation of gene pairs (i.e. both differential and nondifferential interactions), a detailed computational scheme for which is illustrated in Fig. 2 . Compared with the overall procedure to construct a differential expression network, traditional schemes for differential gene sets and differential networks are also shown. The construction of a differential expression network includes the three steps detailed below, in which a background network (e.g. PPI network) and expression data for case and control (e.g. disease and normal) are assumed to be available.
Extracting differential interactions
Genes are selected in a pairwise manner instead of individually. In other words, an edge from a background network, e.g. PPI network, is selected only if its corresponding two genes are strongly correlated (e.g. the correlation coefficient between expression of these two genes is greater than or equal to some threshold, such as 0.7) in one condition, although not (e.g. < 0.5) in another. Such an edge implies that the interaction between a gene pair would be perturbed under the disease state and hence is denoted as a ‘differential interaction’. All of these interactions comprise a dysfunctional network. Interestingly, it has been observed that for many cases, there are usually no or few overlaps between the two sets of genes chosen by ‘differential genes’ and ‘differential interactions’, respectively. This indicates that the two schemes of extracting differential expression information are complementary and demonstrates the need for the integration of differential genes and differential interactions.
Extracting nondifferential interactions
Each nondifferential interaction is chosen based on the differential expression of the two genes on each edge of the background network. At first, there are many commonly employed extraction techniques for ‘differential genes’, any of which can be adopted to identify significantly differentially expressed genes. These genes are then mapped onto the background network, e.g. PPI network. Finally, only the edges in the background network for which two nodes (genes) are both differentially regulated are taken as nondifferential interactions. The network comprising such edges is defined as an active network.
Integrating differential interactions and nondifferential interactions
Clearly, the aforementioned two methods (differential and nondifferential interactions) to identify disease-related genes (nodes) or interactions (edges) are complementary. The final scheme is to use the union of these two types of interactions to construct a differential expression network, which is able to characterize alterations in the expression and correlation of the genes (i.e. nodes and edges) between disease and normal samples simultaneously.
The main contribution of the above new scheme is that a molecular network (e.g. PPI network) is integrated with gene expression profiles by bringing together ‘differential interactions’ and ‘nondifferential interactions’. Clearly, compared with traditional schemes, the differential expression network exploits more information about both gene expression and gene interaction related to the disease. It can be expected that this scheme may extract novel disease-associated interactions (and genes) and further discover gene regulatory mechanisms involved in disease progression.
Dynamical gene network and differential activity
Network-based biomarkers are able to distinguish between disease and normal samples by considering their differential features in a static, but networked manner. However, they cannot generally characterize the dynamical properties of a biological system and its temporal changes during disease initiation and progression. By contrast, early diagnosis of a complex disease is becoming increasingly important in clinical research, which necessitates dynamical biomarkers to detect the early signals of diseases by exploring both dynamical and network information [27, 68]. In general, disease progression can be considered to have three stages: the normal stage, the pre-disease stage and the disease stage, where the pre-disease stage, in particular, is critical (e.g. the pre-diabetic stage in the progression of diabetes) when individuals are in a state (i.e. pre-disease state) just before the appearance of disease symptoms [20, 38]. Thus, detection of the pre-disease state is crucial to achieve the early diagnosis of diseases. Different from both the normal state and the disease state, the pre-disease state has been shown to have significant dynamical features, e.g. the appearance of a group of strongly correlated but fluctuating genes when a biological system approaches the pre-disease state. Such theoretical results have been applied to various diseases to identify their critical stages by exploring information from gene and protein expression data [20, 21, 69].
The pre-disease stage can be considered to be the limit of the normal stage just before the critical transition from the normal stage to the disease stage during disease initiation and progression [20, 70, 71]. A new concept, i.e. the DNB as a general criterion, has recently been developed to detect early-warning signs of the critical transition to or the sudden deterioration caused by a disease. Intuitively, a DNB is a group of genes that have strong correlations, but their expressions vary greatly at the pre-disease stage rather than other stages. These genes also have clear physical and biological implications in relation to the disease, i.e. DNB genes mark the first changes from the normal state to the disease state, and thus consist of the leading network. Therefore, DNB genes not only can be used to identify the pre-disease state, but also are considered to be causally related to disease initiation and development. In the following, we first summarize the theory of DNB and then describe a recent approach using a progressive module network model, which gives a network-constrained DNB. Applying this new model to type 1 diabetes mellitus (T1DM) shows its potential to reveal significant functional reorganization during the early development of T1DM.
Theory and criteria for a DNB
The detailed mathematical derivation of the DNB as an early-warning sign during disease initiation and progression can be found in Chen et al. , and the following three criteria of the DNB are briefly summarized in both biological observation and mathematical description.
When a biological system approaches a critical transition, there exists a group of molecules (i.e. DNB or dominant group), e.g. genes (or proteins), satisfying the following:
correlations among the expression of DNB genes become stronger, e.g. significantly co-expressed;
by contrast, correlations among the expression of DNB and non-DNB genes become weaker; and
variances in the expression of these DNB genes become larger rather than steady-state expression.
When all the above three conditions are satisfied simultaneously, this group of molecules (e.g. genes) is a dominant group in the biological system and is called as the leading network (or leading module). In fact, each of the three conditions represents a criterion, and their combination is naturally expected to be a strong signal of or indicator for the pre-disease state. The leading network is also coincident with DNB, and the existence/appearance of DNB suggests that the biological system is in the pre-disease state. Therefore, the above three statistical criteria have been simply combined to construct a composite index CI (or signal index) such that:
where PCCd is the numerical measurement of criterion (a), which is the absolute value of the average correlation (e.g. Pearson's correlation coefficient, which is a well-known mathematical measurement on linear correlation , including positive or negative correlations ) of the expressions of DNB genes; PCCo is the numerical measurement of criterion (b), which is the absolute value of the average correlation between the expression of DNB genes and other genes; and SDd is the numerical measurement of criterion (c), which is the average standard deviation in the expressions of DNB genes. Thus, we can identify the DNB to detect the early signs of a disease by identifying when the signal index achieves its largest value. From a theoretical viewpoint, the DNB or the leading network is a group of genes that drives the whole system from one state to another, and therefore the DNB genes are causally related to disease initiation and progression.
Progressive module network model
In recent studies, a new computational framework known as the progressive module network model has been implemented to assess network modules relevant to the pre-, early and advanced-disease stages by reconstructing the progressive module network based on DNB theory. This framework primarily consists of four steps (Fig. 3). (a) Tissue- and time-specific networks (context-based networks) are constructed from the available gene-expression data and biological interactions, and each context-based network is divided into sub-networks (i.e. by a Markov cluster algorithm ). All sub-networks/modules consist of a module pool (shown in Fig. 3A,B). (b) For a tissue, each module in such a pool is used to calculate its signal index (Eqn (1)) over time, and the module with the largest index value is considered to be the pre-disease module in the tissue (the time when this pre-disease module has the largest score is the critical time; Fig. 3C). (c) The observation index score (Eqn (2)) for each module is also calculated across different tissues and times, and the significant high-ranked modules are considered to be disease-responsive modules (Fig. 3D) because these modules have been detected as many times as possible. (d) For each tissue, the progressive module network is recalculated using a weighted shortest-path algorithm based on the pre-disease modules and disease-responsive modules at the critical time (Fig. 3E).
Here, we give the biological representations and mathematical notations of signal index I (Eqn (1)) and observation index F (Eqn (2)). In fact, the signal index is the composite index in the above DNB theory, which is a measurement of the three criteria of DNB detection. The module with the largest signal value is selected as the pre-disease module. Meanwhile, the observation index reflects the size of a module and its observation frequency in spatial–temporal data. The module with the largest observation value is selected as the disease-responsive module because of its nontrivial high occurrence during disease development and progression. In addition, note that MG,S is a module, where G is a sub-network of the biological network (e.g. PPI) and S is a subset of times when the module is observed; for a sub-network G, V(G) is its node set, representing the genes in this sub-network, E(G) is its edge set, representing interactions between the genes in this sub-network, and is its cut edge set, representing interactions between the genes inside G and outside G; sdv is the standard deviation for expressions of genes represented by node v, and pccv1,v2 is the absolute correlation between expressions of two genes represented by nodes v1 and v2.
Analysis on the early development of T1DM
A progressive module network model has been used to analyse T1DM based on integrated gene expression data from three datasets  deposited in the Gene Expression Omnibus from the National Center for Biotechnology Information; these are pancreatic lymph nodes within GSE15150, spleen within GSE21884 and peripheral blood cells within GSE21897. The dynamical module biomarkers of different tissues for T1DM are identified, and the module organizations of tissue-specific progressive module networks are strongly associated with the early development of T1DM.
For the pre-disease modules (or DNBs) in three tissues (genes within red, orange and blue circular nodes in Fig. 4), there is no gene overlap. However, they are functionally related because they can be directly connected by a group of genes, such as the genes for G-protein-coupled receptors shown in Fig. 4 (e.g. gene Gng12 in brown). This finding demonstrates that, during the initiation of T1DM, the driving cause in different tissues may be different, but dysfunction of the transmission signals from hormones (functions of G-protein-coupled receptors) would be a general trigger for the initiation and early development of T1DM . However, the essential progressive genes observed in all three tissues (genes within pink circular nodes in Fig. 4) tend to be closer to the disease-responsive modules (genes within purple circular nodes in Fig. 4) rather than pre-disease modules. Thus, they would be potential targets for therapy to control disease progression. In fact, recent studies have shown that CD4 (gene Cd4 in Fig. 4) is important in inhibiting the progression of T1DM, and IL17ra, which is associated with Cd4 (this gene association does not exist in the background network but can be found in the STRING database) has the ability to reverse established T1DM symptoms [77-79]. In addition, many essential progressive genes identified using a progressive module network model are also important candidates as therapy targets.
Both computational analysis and independent public experimental data demonstrate the potential of the progressive module network to better understand the early development of T1DM. Compared with the original DNB method, the progressive module network model considers the network constraint on the construction of a DNB and therefore has a clear biological meaning. In addition, it is also a general method to identify dynamical module biomarkers for other complex diseases.
With the rapid accumulation of high-throughput data on complex diseases, the identification of reliable gene biomarkers is in great demand both for early diagnosis of diseases and to help decipher the essential mechanisms of disease progression in a systematical and dynamical manner. In fact, a single disease generally results from many correlated factors rather than single or a few individual factors. In other words, many complex diseases can be considered as a problem in a system, which is caused by the gradual accumulation of multiple genetic or epigenetic changes. From the viewpoint of nonlinear dynamics, early signals of disease progression, and in particular the critical changes before transitioning between clinical stages, may be reflected by the dynamics of the biological system state, e.g. time- or stage-course gene-expression data. Gene biomarkers from different levels will help us extract pathogen information from differential gene expression and correlation data. By considering the information of both networks and dynamics, the network biomarker and DNB can provide novel insights on elucidating disease initiation and progression at the molecular level. In particular, the DNB has been shown to have the ability to detect the early-warning signs of a disease, i.e. distinguish pre-disease and disease samples, which is a key for early diagnosis and accurate prognosis. In future, to further improve the accuracy, robustness and interpretability of these biological indicators on diseases, sophisticated biomarker methods are needed to integrate not only different types of information (e.g. static and dynamical information) but also different types of indicators (e.g. clinical markers , mutation markers  or epigenetic markers ).
We sincerely thank Prof. Ettore Appella and Dr Lisa Jenkins from National Cancer Institute, NIH, Bethesda, Maryland for their valuable suggestions and editing which significantly improved the manuscript. The authors also would like to thank anonymous reviewers who gave valuable suggestion to improve the quality of the manuscript. This work was supported by the NSFC (Nos. 61134013, 91029301, 61072149, 31200987, 61171007, 91130033, 11326035 and 11131009), the Chief Scientist Program of Shanghai Institutes for Biological Sciences (SIBS), Chinese Academy of Sciences (CAS) (No. 2009CSP002), the Knowledge Innovation Program of CAS (No. KSCX2-EW-R-01), the Knowledge Innovation Program of SIBS of CAS with Grant No.2013KIP218, the China Postdoctoral Science Foundation Funded Project under Grant No. 20110490757, and 863 project (No. 2012AA020406). This research was also partially supported by the National Center for Mathematics and Interdisciplinary Sciences of CAS, Shanghai Pujiang Program, and the FIRST program from JSPS initiated by CSTP.