RESULTS AND DISCUSSION

To simulate the production of SARS-CoV-2 virus particles in a human cell we utilised an existing, community-developed human genome scale model known as RECON 2.2 (Thiele 2013; Swainston 2016) (see Methods ). This model represents the most comprehensive catalogue of metabolic reactions found in human cells, with many of its reactions associated with known genes (Swainston 2016). Within this model, we implement a pseudo reaction representing the production of SARS-CoV-2 viral particles from biosynthetic precursors (see Figure 1A andMethods ). The construction of this pseudo reaction is based on available structural information on the virus including its use of proteins for viral packaging (Bar-On 2020; Mahmoudabadi 2017; Bárcena 2009; Neuman 2006; 2008; 2011) and its genome sequence. As such, this pseudo biomass reaction accounts for the stoichiometry of nucleic and amino acids required to make a complete virus and associated energetic costs. This analysis highlights that leucine and alanine are the most utilised amino acids in SARS-CoV-2 proteins and adenosine- and uridine-triphosphate are the more common nucleotides in its RNA (Figure 1B).
Metabolic fluxes supporting SARS-CoV-2 production in a human cell are primarily in central carbon metabolism. By incorporating the SARS-CoV-2 viral biomass function into the human metabolic model and assuming a minimal media composition (see Methods ) we predict a metabolic flux distribution for optimal virus production in a human cell (Figure 1C). We then evaluated the flux variability allowed in each reaction of the model, while maintaining an optimal viral production level (Supplementary File 1). These analyses have shown that reactions which must carry flux for optimal viral biomass production include glycolysis, oxidative phosphorylation, fatty acid oxidation, and specific amino and nucleic acid biosynthesis pathways (Figure 1C andSupplementary File 1). As the optimal flux distributions are related also to the flux limits imposed on uptake reaction fluxes, we also repeated the flux variability analysis with minimal media but using an increased uptake limit and with a rich media that allows all uptake reactions of the model to be active. Increasing uptake limits did not alter the general conclusions about key active pathways sustaining optimal virus production, but resulted in higher uptake fluxes causing additional pathways relating to overflow metabolism to be active (Supplementary File 1). Simulating a rich media resulted in a much lower number of flux-carrying reactions as the cell can obtain several key compounds such as uridine-triphosphate from the media under this scenario (Supplementary File 1). Since, such a rich media is physiologically less realistic, we focus the remaining analysis on the results from simulations using the minimal media.
Inhibiting specific metabolic enzymes and enzyme combinations inhibit SARS-CoV-2 production in human cells. To identify reaction perturbations which, when inhibited, can halt or reduce virus production, we systematically simulated a knock-out of each flux-carrying reaction. Excluding reactions involved in uptake from media, this analysis highlighted 35 reaction knock-outs that can stop virus production and an additional 8 reactions that can reduce it below 80% of the original (Figure 1C). The former group of reactions tended to be involved in nucleotide and amino acid biosynthesis pathways, while the latter group included reactions primarily in glycolysis and oxidative phosphorylation (Table 1 and Supplementary File 2). Key ones among these reactions are further discussed below.
Considering that it is possible for the effects of single perturbations to be circumvented by re-directing of fluxes, we also explored combined perturbations. We created all possible pairs of flux-carrying reactions according to the flux variability analysis (over 5000 pairs) and simulated the effect of setting their reaction fluxes to zero. This has identified over 400 reaction pairs, co-inhibition of which results in the reduction of virus optima to 80% or less of the original (Table 1 and Supplementary File 2). Most of these reaction pairs involved one of the 10 single perturbations that were found to reduce virus production to less than 80% on their own, but pairing them with additional reaction increased their impact. For example, inhibition of glyceraldehyde-3-phosphate dehydrogenase (GAPD) and cytochrome c oxidase (CYOR) individually caused reduction to 62% and 60% of original virus production respectively, but combined inhibition of these reactions results in 25% of original production (Supplementary File 2). Some of these cases of increased effect arises due to co-inhibition of reactions more effectively blocking fluxes into virus biomass precursors. For example, combined blocking of GAPD and CYOR, reactions involved in respiration and glycolysis respectively, results in reduced fluxes into pyruvate and alpha-ketoglutarate (akg), a key intermediary of the tricarboxylic acid (TCA) cycle. Akg is further linked into valine production through the valine:3-methyl-2-oxobutanoate shuttle across the mitochondrial membrane (Figure 2). In the optimal flux distribution for SARS-CoV-2 production, this “valine shuttle” has a high flux and contributes to the production of both valine and multiple other amino acids via mitochondrial glutamate (Figure 2b). Perturbations to both CYOR and GAPD lead to a new flux distribution where the glutamate production in the mitochondrial matrix is sustained through a different metabolic route. The “valine shuttle” that was active in the optimal solution is now non-functional and is instead replaced by a leucine:4-methyl-2-oxopentanoate shuttle carrying a lesser flux. This in turn decreases the production of the amino acids from glutamate, and thus causing a significant decrease in virus biomass production flux (Figure 2).
In the case of simulating the minimal media with higher uptake fluxes, we have also identified pairs of completely new enzyme inhibitions, that were not causing any effect on their own (Supplementary File 2). Some of these pairs exert their effects by blocking multiple pathways from a given compound and thereby causing disruption in steady state balances in the model. For example, co-inhibition of citrate synthase (CSm) and several other enzymes such as histidase (HISD) totally prevents flux in SARS-CoV-2 biosynthesis reaction by making impossible the mass balance of protons in the cytosol (see Figure S1).
Metabolic requirements of viral production are different to those arising from host cellular maintenance. In the above discussed analyses, we considered host metabolism as optimised for viral production and evaluated impact of perturbations only on this process. Such metabolic perturbations should also be evaluated for their impact on the normal metabolic functions of uninfected host cells. In previous studies, normal state of metabolism in human cell lines are either represented through a pseudo reaction for cellular maintenance (Bordbar 2010; Thiele 2013) or through consideration of specific metabolic functions such as ATP or lipid production (Gille 2010; Mardinoglu 2014). In the former case, cell maintenance is captured by a generic account of cellular constituents such as lipids, carbohydrates and DNA and a more specific accounting of amino acid usage in protein expression (Thiele 2013; Bordbar 2010;). In RECON2.2, the protein-based component of the maintenance function is calculated from a large collection of human genes’ open reading frames (Thiele 2013).
Here, we expanded from this approach and calculated the biomass protein component using available protein expression data for lung cells from the Human Protein Atlas project (Uhlén 2015) (see Methods ). Comparing the resulting human lung cell maintenance function to the SARS-CoV-2 biomass, in terms of the building block stoichiometries, revealed differences in relative amino and nucleic acid usage (Figure 3 and Supplementary File 3). Compared to the host, there was particularly higher relative usage of phenylalanine, isoleucine, asparagine, threonine, tryptophan, and tyrosine in the virus and particularly lower relative usage of glutamate, histidine, methionine, and proline (Figure 3). Accordingly, the optimisation of the model using the host metabolic maintenance results in a different metabolic flux distribution compared to viral production (Supplementary File 4). The differences, however, were rather limited from the perspective of fluxes supporting SARS-CoV-2 production; out of all reactions that must carry a flux to sustain a virus optimal state, almost all were also required to carry flux to sustain a host optimal state (Supplementary File 4).
Flux control can ensure selective reduction in viral production. Given the above finding that the same reactions carrying flux for SARS-CoV-2 production also carry flux for host metabolic maintenance, we re-analysed the effects of enzyme perturbations on both virus and the host. We found that many of the previously identified single perturbations limiting virus production also limited significantly the host metabolic maintenance, with only one single perturbation - that involving CYOR - showing more than 5% difference in its impact on virus vs. the host (Supplementary File 5). The same finding prevailed for double perturbations. The only pairs that displayed 5% or more difference in their effects on virus vs. host are those involving CYOR paired with other enzymes (Supplementary File 5).
The limited differential impact of full inhibition of enzymes made us postulate that more refined perturbations could provide a better strategy to just impact SARS-CoV-2 production without affecting the host. In particular, given the differences in optimal metabolic fluxes between virus production and host maintenance states, we argued that there might be flux values for some reaction that are compatible with only one of these states. To explore this possibility, we systematically analysed the flux variability of each reaction given either the optimisation of host maintenance or virus production. This has allowed us to see if any of the reactions would have flux regimes that are only compatible with the optimal host maintenance but not with optimal virus production and then ‘enforce’ such flux regimes on the model. This approach allowed us to identify few double reaction perturbations that are fully selective on their effect and solely reduce virus production without causing any impact on the host (see Table 1). All of these involved threonine deaminase (THRD) and caused up to 17% reduction in SARS-CoV-2 production (Table 1).
The flux enforcement approach creates further constraints on how the metabolic fluxes in the system can be balanced at steady state. In the case of flux enforcements involving THRD this reduced set seems to leave only certain flux distributions to be possible that are specifically less optimal for the virus’ biomass biosynthesis, but still optimal for host maintenance function to be fulfilled. Expected from this, we found that the effective flux enforcements relate to compositional differences between the viral biomass and host maintenance functions, as illustrated in Figure 4 for the case of the enforcement of THRD and succinate dehydrogenase (SUCD1m). In the optimal flux distribution for SARS-CoV-2 production in the unmodified model, threonine is obtained from the medium through a threonine:leucine shuttle, with both of these amino acids being in relatively similar demand between host and virus requirements. In the case of THRD and SUCD1m fluxes enforced to specific ranges, threonine is instead obtained through a threonine:isoleucine shuttle and is now further intertwined with the arginine:isoleucine shuttle (Figure 4). This shift in fluxes happens because the flux enforcement of SUCD1m, a reaction of the TCA cycle, decreases the flux of the previously observed valine shuttle (see Figure 2) and increases the flux of alternative mitochondrial shuttle reactions including one involving leucine (see Figure 4). The reduced availability of leucine causes the utilisation of isoleucine and arginine for threonine, thereby creating a trade-off among these amino acids. Crucially, arginine is the highest and isoleucine the third highest differentially demanded amino acid when comparing virus vs. host requirements (Figure 3). This is why the described trade-off situation among these amino acids and threonine, caused by the flux enforcement, differentially impacts the SARS-CoV-2 biomass production more than it effects host maintenance.
Current metabolic drugs exist that could target predicted reactions to inhibit production of SARS-CoV-2. The metabolic approach employed here allowed prediction of several reactions and reaction combinations that could limit SARS-CoV-2 production in human cells in general and differentially in human lung cells. The most significant of these are re-summarised in Table 1 as those reactions, the perturbation of which, can reduce virus production below 80% of the original. For these reactions, we evaluated their associated enzymes in the light of existing, approved drugs using the available small molecule inhibitor and drug database DrugBank (Wishart 2018). We found several existing drugs that could inhibit some of the predicted reactions including those targeting enolase (ENO), phosphoglycerate mutase (PGM), and SUCD1m (Table 2). These drugs could be used as a starting point to experimentally test the predictions made here, using in vitroassays. In addition to these identified small molecule inhibitors, we note that it might also be possible to achieve development of de novo metabolic gene knock-out approaches using recent CRISPR and RNA silencing approaches.