Compartmentalization of metabolism between cell types in multicellular organisms: a computational perspective

Abstract

In multicellular organisms, metabolism is compartmentalized at many levels, including tissues and organs, different cell types, and subcellular compartments. Compartmentalization creates a coordinated homeostatic system where each compartment contributes to the production of energy and biomolecules the organism needs to carrying out specific metabolic tasks. Experimentally studying metabolic compartmentalization and metabolic interactions between cells and tissues in multicellular organisms is challenging at a systems level. However, recent progress in computational modeling provides an alternative approach to this problem. Here we discuss how integrating metabolic network modeling with omics data offers an opportunity to reveal metabolic states at the level of organs, tissues and, ultimately, individual cells. We review the current status of genome-scale metabolic network models in multicellular organisms, methods to study metabolic compartmentalization in silico, and insights gained from computational analyses. We also discuss outstanding challenges and provide perspectives for the future directions of the field.

Introduction

Individual metabolic reactions are carried out in distinct locations in the body and in different compartments of the cell, which generates compartmentalization of metabolism at both cellular and organismal levels. At the organismal level, compartmentalized reactions cooperate to sustain energy and biomass production during growth, homeostasis and wound healing. A well-known example of compartmentalized metabolism is the Cori cycle in which lactate produced by anaerobic glycolysis in skeletal muscles is transported to the liver and converted to glucose, which then returns to muscles to provide energy for movement [1]. More recent studies have shown that lactate is a major metabolite circulating in the blood that fuels energy production in different tissues [2]. In complex organisms such as humans, one organ or tissue usually contains multiple cell types that each have their own metabolic activities (Figure 1). For instance, in the endocrine pancreas, only beta cells produce insulin [3]. It is not yet possible to prepare bulk samples for every cell type to trace flux or measure metabolite abundance. Therefore, experimental studies of metabolic compartmentalization at the level of entire metabolic network and in multiple tissues/organs have been challenging [4,5]. Metabolic network modeling has provided a facile alternative to reveal the metabolic state of different tissues, organs and cell types.

Figure 1: Compartmentalization of metabolism in multicellular organisms.

Complex multicellular organisms (such as humans) contain multiple organs and tissues that further consist of several cell types. Metabolism is compartmentalized at every layer, resulting in tissue- and cell- specific metabolic networks and distinct metabolic states. Organism-level GEM contains all the enzymes, metabolites, and metabolic reactions that are present in an organism according to its genome annotation (metabolites, enzymes and reactions are shown in circle, rectangle and arrow, respectively). A particular tissue or cell only expresses a subset of enzymes, which results in the compartmentalized metabolism. The magenta arrow indicates the flux distributions in tissues and cells.

Genome-scale metabolic network models (GEMs) detail the enzymatic conversions and transport reactions that can take place in an organism using the annotation of the genes that encode the corresponding enzymes and transporters, and can be used to study metabolism in silico [6]. In GEMs, nodes are metabolites and edges include the conversion reactions between metabolites, as well as metabolite transport reactions between different cellular compartments (Figure 1). Genes encoding metabolic enzymes or transporters are associated with different reactions based on homology with known enzymes in other organisms (often bacteria), unless the function of the gene has experimentally been studied in the organism of interest. A well-constructed GEM can be used with constraint-based flux balance analysis (FBA)[7], a method that calculates conversion rates of metabolites in all reactions of the GEM at steady state, and the model can be integrated with omics data such as gene expression profiling or proteomics data to derive hypotheses about metabolite buildup and flux alterations [8,9].

Several computational methods have been developed to study metabolic compartmentalization at a systems level. For instance, in a pioneering study, metabolism was predicted for ten human tissues in the form of flux distributions that describe the flux of every reaction in the metabolic network [10]. With recent advances in single-cell omics technologies, it may now be feasible to gain insights into metabolic compartmentalization at the level of distinct cell types and, ultimately, individual cells. Here, we review the current status of GEMs in multicellular organisms, methods to study metabolic compartmentalization in silico, and insights gained by applying the methods to humans and model organisms. We finish by summarizing outstanding challenges and provide perspectives for future directions.

GEMs of multicellular organisms

GEMs have been reconstructed for various multicellular organisms, including several commonly used model organisms such as Arabidopsis thaliana, Caenorhabditis elegans, and Mus musculus [11]. Human GEMs are being developed continuously, in a community-driven way. Starting from Recon 1 in 2007 [12], a series of human GEMs have been published, including Recon 2 [13], Recon 2.2 [14], Recon3D [15], and human 1 [16]. The human model has expanded over time from Recon 1, which only contains 1,496 genes, 2,766 metabolites and 3,311 reactions to the most recent model named human 1, which contains 3625 genes, 10,138 metabolites and 13,417 reactions. The consistency and quality of human metabolic models have also improved over time [16,17], allowing FBA and data integration. GEMs for model organisms have also been expanded, resulting in improved mouse [18], zebrafish [19], and nematode [20] models. Whole genome sequencing and metabolic gene annotations have also enabled the reconstruction of models for non-canonical model organisms such as the green foxtail and Pacific white shrimp [21,22].

Methods to study the metabolic compartmentalization in silico

Modeling algorithms

Algorithms have been developed to predict and analyze tissue-level metabolism via integration of GEMs with omics data (e.g., transcriptomics and proteomics). We classify these algorithms into two classes based on their purpose: ‘network builders’ and ‘phenotype predictors’ (Figure 2). Network builders aim to reconstruct context-specific metabolic network models, for instance to describe the metabolic network for a particular tissue. Different tissues express different metabolic enzymes and therefore only use a subset of the reactions included in a whole-organism GEM (Figure 1). If an enzyme is not expressed in a particular tissue, reactions associated with that enzyme should be removed from the network. Similarly, if a metabolite is not present, reactions dependent on this metabolite may be removed. With such ideas, methods have been developed to extract context- (tissue-)specific networks by integrating transcriptomic, proteomic, and/or metabolomic data with GEMs (comprehensively reviewed in: [23] and recently in [24]). The resulting tissue network models can be used to directly inform the compartmentalization of metabolic capacities (i.e., which reaction is able to carry flux in a tissue) or to perform downstream analysis such as predicting tissue-specific metabolic phenotypes (Figure 2).

Figure 2: modeling algorithms to study metabolic compartmentalization.

Omics data can be integrated with GEMs to build tissue- and cell- specific GEMs, and to directly predict the metabolic phenotypes in tissues and cells. Accordingly, modeling algorithms are classified into network builders and phenotype predictors. The tissue-specific GEMs can be further used to predict metabolic phenotypes by FBA or some phenotype predictors such as FPA. Representative network builders include iMAT [10], iMAT++ [20], INIT [41], tINIT [55], MBA [67] and FASTCORE [68]. Representative phenotype predictors include EFLUX [28], iMAT [10], iMAT++ [20], RELATCH [29], REMI [30], FPA [20], COMPASS [31] and TIMBR [33].

Phenotype predictors aim at predicting metabolic phenotypes such as flux distribution and metabolite abundance (Figure 2). Flux distributions can be predicted by performing FBA on the tissue-specific networks constructed by network builders [24]. FBA has been widely used to predict flux distributions, mostly in proliferating cells such as mircobes and cancer cells, by using biomass production (growth) as an objective function [7]. However, this approach is less effective for modeling tissues in a multi-cellular organism, as it is difficult to capture the diverse and complex metabolic functions of tissues by a single objective function. One strategy to overcome this problem is to narrow down the solution space by constraining the fluxes of some of the reactions that exchange metabolites with the environment based on measured uptake and secretion rates of metabolites in the exometabolome [25–27]. Alternatively, flux distributions can be predicted directly from the integration of omics data without the need for an a priori objective function [10,20,28–32]. Algorithms that achieve this task usually aim to find a flux distribution that best agrees with gene expression data [10,20,28–32], and sometimes metabolomic data [30,32], such that the reactions associated with highly expressed genes have active flux while those associated with lowly expressed genes have low or no flux. Network-scale integration is not perfectly quantitative since alternate flux distributions exist. Therefore, some algorithms intend to predict the relative flux levels for each reaction across tissues individually instead of making a network-scale flux distribution, such as Flux Potential Analysis (FPA)[20] and Compass [31]. Finally, metabolite abundance can also be predicted qualitatively [10] or quantitatively [20,33,34]. In addition to mechanistic modeling with phenotype predictors, a few recent studies highlight the use of machine learning to predict metabolite concentrations, with or without an underlying metabolic network model [35–38]). Incorporating machine learning has great potential to predict metabolic phenotypes that cannot be currently achieved by mechanistic modeling, provided that sufficient data is obtained to training the models [39,40].

Modeling frameworks

The simplest way is to model each tissue and cell type of an organism individually, i.e., by individually reconstructing tissue-specific networks and predicting metabolic phenotypes [41–43]. However, this simple approach neglects interactions between tissues and cells. To model inter-tissue interactions, the networks of two or more tissues can be connected by the exchange of metabolites. The pioneering work by Lewis and colleagues constructed dual-tissue networks that characterized the crosstalk between astrocytes and neurons by connecting the tissue-specific networks of the two cell types [44]. Transportable metabolites were allowed to exchange between the two networks and both networks could obtain nutrients from the environment (endothelium/blood). The same strategy has been applied to reconstruct multi-tissue networks that model the crosstalk between liver, skeletal muscle and fat tissues [45–47]. However, multi-tissue models cannot capture the metabolic compartmentalization and interaction at the whole-organism level, which includes the digestion and absorption of the diet, distribution of the nutrients throughout organs and tissues and finally the energy and biomass production of the entire body. A few recent studies pioneered the development of whole-body models to address this challenge. For example, two whole-plant models were developed for the common model plant A. thaliana [48,49]. Both models are spatially compartmentalized to contain either leaf, stem, and root or only leaf and root as different but coordinated metabolic networks, and are also temporally compartmentalized to have light and dark phases that represent the cycling of nutrients between day and night. More recently, we constructed the first whole-animal model that simulated the conversion of diet (bacterial biomass) to energy and biomass in seven major tissues of the nematode C. elegans [20] based on published second larval stage mRNA expression levels in these tissues [50]. To decrease the computational demand for the integration of the model with expression profiling data, a dual-tissue framework was used, where the intestine network was connected to each of the other six tissues (pharynx, neurons, muscle, gonad, hypodermis and glia), one tissue at a time, resulting in the nutrient influx to intestine that in turn feeds the other tissue via the exchange of transportable metabolites. Around the same time, the first whole-body human GEM was developed [51]. This model contains 26 organs and six blood cell types in two sex-specific whole-body metabolic (WBM) reconstructions, where the metabolic network of each compartment was curated to represent its known physiological properties, and the overall model was built according to human anatomy. However, the rigorous reconstruction of dozens of tissues resulted in a very large model (more than 80,000 reactions), which may become a bottleneck for applying many phenotype predictor algorithms due to the high demand for computational power.

Metabolic compartmentalization in humans and model organisms

To date, the tissue-specific metabolic networks have been built for more than one hundred normal and diseased human tissues [52]. However, only limited studies have applied phenotype predictors to gain deeper insights into human tissue metabolism [10,45]. Studies have started to utilize the tissue- and cell-type-specific networks to reveal the mechanisms and drug targets of different diseases [44,53–58]. For example, a hepatocyte-specific network has been integrated with gene expression data obtained from patients with non-alcoholic fatty liver disease to the discover biomarkers and potential therapeutic targets [53]. The multi-tissue and whole-body human models have also been shown to recapitulate known metabolic communications within the body [44–47]. For instance, the human sex-specific organ-resolved WBM [51] recapitulated known inter-organ metabolic cycles and energy use, as well as predicted known biomarkers of inherited metabolic diseases in different biofluids.

Although only few studies are available to date, modeling metabolic compartmentalization in model organisms has also proven highly useful because it allows data acquisition from well-controlled laboratory conditions and facilitates experimental validations [18,20,48,59]. In particular, C. elegans offers high throughput sampling of cells from thousands of individuals grown under precisely controlled conditions [50]. The whole-animal model of C. elegans recapitulated known tissue functions, captured metabolic properties that are shared with similar tissues in human, and provided predictions for novel metabolic functions [20]. For instance, the gonad was predicted to be the main site of selenocompound production, and possibly selenoprotein biosynthesis, which may direct future investigation of the roles of selenocompound pathway. Similarly, multi-tissue models of A. thaliana provided insights about the energetic cost of translocating carbon and nitrogen sources between different organs during day and night metabolism [48] and the partitioning of carbon and nitrogen sources between organs during growth [49].

Challenges and perspectives

The last decade has witnessed the emergence of a number of computational methods for modeling metabolic compartmentalization in multicellular organisms, however, challenges still exist. The network builder algorithms were benchmarked in a recent study [8], which found that no single algorithm universally gave the most physiologically accurate models, although popular algorithms were useful for a number of applications. The phenotype predictor algorithms need future investigations on systematic benchmarking and data-driven improvement. Future studies will also be required to develop the modeling frameworks for computationally efficient whole-body models in which multiple tissues can interact with each other while the model size is still compatible with regular computational power.

More biological insights about tissue metabolism are still to be gained despite the fact that hundreds of tissue-specific networks have been reconstructed. For instance, comparative analysis of those tissue-specific networks may reveal the fundamental paradigm of metabolic compartmentalization in multicellular organisms. In addition, phenotype predictors (e.g., FPA) could also be applied to gain insights about metabolic phenotypes in tissues and cells. Furthermore, we expect the technological developments and increasing data availability to futher improve the outcome of mechanistic modeling applications. For example, deeper and more accurate insights might be obtained by using proteomics as the integration input compared with transcriptomics, as proteomic data is more directly related to enzyme abundance. Notably, high-quality quantitative proteomics has become increasingly available, including a proteome map of 32 normal human tissues that was recently published [60]. Similarly, single-cell omics has great potential to improve the resolution of predictions on the metabolic functions of cell types. Thanks to the availability of single-cell transcriptomes, recent studies have been able to model tissue metabolism at unprecedented resolution [20,31]. However, most studies still predict the metabolic state of a group of cells belonging to the same cell-type rather than that of individual cells, as the technical noisiness of single-cell omics data and the intrisict stochasity of metabolism in a single cell make it difficult to model individual cells. Various strategies have been proposed to address this challenge, such as information borrowing strategies [20,31], stochastic simulation [61] and population modeling [62]. We foresee the active methodology development of integrating single-cell data, and meanwhile, applications of the current methods to gain knowledge about metabolism at the level of cell types inside of a tissue. Ultimately, metabolic specialization and communication may be modeled at the single cell level. To this end, established methods for modeling the metabolic heterogeneity in microbial communities [6,63,64] may inspire future method development.

Finally, we encourage the development of tools that offer user-friendly programs for online, context-specific network reconstruction, result interpretation and metabolic phenotype predictions. Although it may be computationally challenging to host such a tool on a typical server, online integration would be eventually feasible by conjoining algorithms with the high-performance computing clusters or commercial cloud services such as Amazon AWS, which has been successfully implemented for sequencing data processing [65,66]. Such tools may ultimately make metabolic network integration a routine technique for omics data analysis to the broad scientific community.

Highlights.

• Metabolic network modeling reveals metabolic states of tissues and cell types

• Algorithms include network builders and phenotype predictors based on their purpose

• Multi-tissue and whole-body models capture the interactions between tissues

• Integrating single-cell omics data provides cell-type level metabolic insights