Similar to the reconstruction of the global human metabolic networks, one could follow the protocol [17] to collate literature evidences and gene annotations, and manually identify the context-dependent reactions to reconstruct a condition-specific metabolic network. This has been applied in [28,29] to achieve comprehensive reconstructions of hepatic and neuronal cells. However, the process requires manual evaluation of thousands of papers, and curates thousands of genes and metabolites, which is lengthy, and requires tremendous effort and labor. Therefore, many studies have focused on automating the reconstruction of a cell-type or tissue-specific metabolic network, by incorporating high throughput gene expression data rather than manual curation of the cell/tissue-specific network.

Assuming changes in gene expression drives changes in the metabolic states; the basic idea in automated reconstruction is to identify active reactions by incorporating condition-specific gene expression profiles. Most reconstruction approaches start by analyzing the gene expression data to determine if a gene is “present” (highly expressed) or “absent” (low expression level) for the condition being investigated, then selects the active reactions according to their corresponding gene/enzymes’ expression level. For example, the Gene Inactivity Moderated by Metabolism and Expression (GIMME) algorithm, developed by Becker et al. [30], uses gene expression data to determine active and suppressed genes with an expression threshold, and determines active reactions based on the state (active or suppressed) of the corresponding enzymes. “Inactive reactions” are removed unless they are required for a desired functionality (according to a predefined objective function). Another approach to study human tissue-specific metabolic states was developed by Shlomi et al. [31]. The approach does not require an objective function but matches active reactions with expression data by solving a network flux to maximize the number of enzymes that are highly expressed and catalyze flux-carrying reactions. For a better quality reconstruction, Jerby et al. [32] developed a Model Building Algorithm (MBA), which determines the active “core reactions” with multiple sources of information including literature, transcriptome and proteomic data. The MBA then reconstructs a consistent network (no gaps or zero-flux reactions) with all the pre-defined core reactions (evidences obtained from both literature and data), adding as many of the likely active reactions (evidences obtained only from high throughput data), and as few of the other reactions. Nevertheless, manual curation is necessary with this approach to collate and analyze the literature and high-throughput data.

Constraint based modeling and simulation based on a condition-specific human metabolic network can be used to predict the flux distribution in the network for that specific metabolic phenotype. Such in silico analysis can be used to generate hypotheses on the cell growth, ATP production, or the states of specific metabolic functions upon perturbation [33].

Modeling and simulations of a metabolic model of human kidney reconstructed with the GIMME algorithm were used to evaluate the metabolic phenotypes associated with the side-effects of a drug treatment [34]. The side effect of a particular drug is determined by its off-targets, which are the enzymes/genes that are not the therapeutic targets but nevertheless are predicted to bind and be inhibited by the drug. FBA was performed on the perturbed network where the reactions catalyzed by the off-target enzymes were inhibited by the drug, to evaluate the systematic consequences of the drug and determine if the treatment leads to deficiency in metabolic functions [34]. This systematic approach has also been applied to metabolic disorders of the liver [32] and cancer cells [35,36], and was able to correctly identify many of the genes essential for the metabolic disorders [32], interpret metabolic state-changes (e.g., Warburg effects [36]), and predict drug targets for the metabolic system [35]. In the cancer study [35], gene expressions in the cancer cell lines are analyzed to identify the highly expressed metabolic enzyme-encoding genes in the cancer, which are used on the MBA algorithm to reconstruct a “cancer metabolic network model”. FBA is then applied to predict metabolic states (cell proliferation) across different gene knock-downs. Specifically, in each prediction one could turn off the reaction associated with the gene that is knocked-down and apply FBA on the constrained model to determine if cell proliferation (represented by an objective function) is reduced. The genes predicted to be important for cancer metabolism are confirmed to be highly essential in a shRNA gene knockdown dataset which lists experimentally identified cancer growth-supporting genes. FBA is also applied on non-cancer cells (reconstructed metabolic network with expression data of normal cells) to determine genes important for metabolism of normal cells. Genes that only affect cancer cells are predicted to be drug targets. Many known targets of FDA-approved metabolic anticancer drugs have been re-discovered with this approach. Folger et al. [35] further simulated double gene knockdowns to explore combinations of synthetic lethal drug targets.

In summary, the reconstruction and simulation of human metabolic networks enable a systematic study of the regulatory system and is applicable to a wide variety of cases. One could rely then on in silico perturbations to predict novel targets, interpret systemic effects, and identify alterations in the metabolic states to better understand the genotype-phenotype relationships.

Current reconstructions of the global human metabolic network are incomplete due to limitations in our knowledge of the complexity of the human genome. Both the human Recon 1 and Edinburgh Human Metabolic Network have gaps and missing parts. For example, there are many reactions in the networks that do not have corresponding regulatory genes/enzymes annotated. For the reactions with known genes/enzymes in either network, there are only around one thousand genes that are common to both, with some collated in only one of the networks. It would be better to combine the two reconstructions to achieve a more comprehensive network. Nevertheless, the integration would be difficult since the standards are different between the two networks—Different compound names, different reaction names, and different IDs are used for the network components.

The community needs to drive the effort for better reconstruction and integration of these network models. A community approach of “reconstruction jamborees” [37,38] was suggested for modeling the metabolic network of different species and systems. A “jamboree” is a focused work meeting that promotes collaborations between experts from diverse fields (e.g., genetics, physiology, metabolomics, engineers, and computer science), to define and evaluate protocols, standards and ontologies for the curation of models, and to resolve discrepancies. Such a collaborative effort would help refine and update the human metabolic network to further enhance the applications of these networks.

Applications of the metabolic model require reconstruction of context-dependent networks, thus computational approaches to automate the generation of condition-specific models are of great interest. Current approaches rely on present and absent (high/low expression) of the genes to determine potentially active reactions, which could be problematic since a small change in the threshold in the expression analysis could result in very different lists of present and absent genes and affect the reconstruction. Most genes’ expressions are not naturally bimodal but have many levels. A discretization process separates such expression levels into ON and OFF states with some arbitrary thresholds (e.g., based on percentile, clustering, or differential expression), thereby the analysis based on discretization will be affected by the threshold that is chosen. A change in the threshold will change many of the genes’ ON/OFF categorization, which could affect the network reconstruction, since the reconstruction process is also bimodal with reactions being absent or present according to the ON and OFF state of the genes. For example, a gene expressed a little lower than the median may be classified as absent/OFF, then the reaction it is associated with may be removed. However a gene is not necessary in an entirely OFF state when expressed at a low level, and a reaction may not necessary be completely abandoned just because there are fewer enzymes than normal. Current approaches rely on the stoichiometric constraints to mitigate the possible inconsistencies. Future reconstruction approaches could impose more realistic models of the functional relationship between genes and reactions. For example, a novel approach developed by Colijn et al. [39] associates gene expression levels with the constraints on the reactions, i.e., the lower a gene is expressed, the lower the flux that could be conducted through its corresponding reactions. Although it could be more sensitive to changes (noises) in the gene expression by directly associating the gene expression level with flux bounds, the approach is less dependent on the discretization of the gene expression and the present/absent call of the reactions, which could provide a more realistic model. It has been applied to bacteria to predict metabolic modulators and responses to different drugs [39].

With the rapid development of metabolomics, metabolic flux data becomes more readily available and novel approaches are being developed to more efficiently monitor flux distribution in mammalian cells with Metabolic Flux Analysis, by employing isotope-labeling (e.g., C13 label) techniques [40]. Researchers have been studying metabolic alterations in human cells by comparing gene expression profiles with the associated metabolic fluxes measured with isotope-labeling [6,41]. Future modeling and reconstruction of metabolic networks could also incorporate metabolic flux data to determine condition-specific flux distributions and metabolic states.

Context-dependent metabolic models of different cells could be incorporated to achieve higher level systems to study cell-cell interactions. A recent study by Bordbar et al. [42] incorporated metabolic models of human alveolar macrophage and M. tuberculosis to study host-pathogen interactions. Simulation of this metabolic system under different infection states helped identify differentially active pathways and potential gene targets. Another pioneering work by Lewis et al. [29] reconstructed a brain metabolic network for different types of neurons and their interactions with astrocytes, to study energy metabolism in the different cell types. The multi-cellular in silico model was used to analyze the gene expression data to identify changes in the metabolic states as well as important genes in Alzheimer’s disease. In both studies, the incorporation of the metabolic networks was based on manual curation, since there are no computational approaches for reconstructing a multi-cellular metabolic system to date. In the future we should expect novel approaches and applications to higher level systems modeling that involve different cells and tissues.

The gene-reaction associations in current reconstructions of the human metabolic networks account for enzymes, isozymes, transcript variants and protein complexes, but however are limited to metabolic-related genes and enzymes. There are other layers of regulation of gene expression and enzyme activity, including transcriptional regulation, post-transcriptional and post-translational processes. Many genes involved in transcriptional or post-transcriptional regulation of the metabolic genes or enzymes could also affect the metabolic states, which are excluded from current applications. Although it is difficult to have a comprehensive model due to limitations of our knowledge of biological regulation, we could incorporate more information into the metabolic network to provide a better model of the regulation of the human metabolic system. Given the increasing amount of transcriptome data, protein interaction data and other high-throughput data, they could be capitalized upon to enhance the regulatory model. Future studies should consider incorporating multiple layers of regulation to better understand the molecular mechanisms behind the regulation of human metabolism.