MATERIALS AND METHODS

Human genome scale metabolic model and its adaptation to human lung cells. To identify specific host metabolic reactions that can alter viral production, we make use of a generic human cell genome-scale metabolic model that has been previously developed (Thiele 2013) and that has been subsequently curated and improved by the systems biology community (Swainston 2016). This model, referred as RECON2.2, reflects the state-of-the-art in genome scale metabolic model construction applied to human cells and contains over 8000 thousand reactions, many of which have associated gene and protein information (Swainston 2016). This model also contains a pseudo reaction representing generic maintenance costs of a human cell, including ATP and precursor stoichiometries for proteins, DNA, RNA, lipids, and carbohydrates. This pseudo biomass maintenance reaction is primarily derived using information from human leukaemia cell lines (Thiele 2013). This generic human cell model represents a consensus human metabolic capacity, and as such, its use in this study allows identification of widest range of possible metabolic reactions which can then be further scrutinized and sifted in a cell specific context.
As discussed in the main text, when identifying metabolic perturbations that can selectively affect viral production without much affecting the host, the specifics of the used host maintenance representation must be cell-specific. To align this representation to a lung cell we have made use of the gene expression data available from the Human Protein Atlas project (Robinson 2020). In particular, we have used the gene expression profile for lung cells available from this project and available protein sequences from ENSEMBL database (Yates 2020) to create a lung-cell specific stoichiometry for amino acids required for protein synthesis. The proportion of each amino acid in the composition of the new maintenance function was determined by converting the protein coding RNA’s codons to amino acids, counting their frequency and weighing these with the normalised expression coefficient provided by the Human Protein Atlas project. The energy cost associated with amino acid polymerisation is also accounted for based on the length of the protein sequences, and assuming 4.3 molecules of ATP hydrolysed to ADP per amino acid polymerisation (Quek 2014). The stoichiometric coefficient for each amino acid in the new maintenance function is then scaled so as to represent the same weight as the original protein synthesis reaction in the RECON2.2 model (not accounting for ATP, ADP, Pi nor H2O in the scaling) (see Code Availability ). The remaining elements of the maintenance function was retained as in the RECON2.2 model. The final, lung-cell specific maintenance function is provided in Supplementary File 3.
Creation of SARS-CoV-2 virus biomass function. The biomass function for the SARS-CoV-2 is done as in a previous study (Aller 2018) and by accounting for the composition and stoichiometry of proteins and genomic material in the virus. The protein composition and stoichiometry of the virus particle is obtained from electron microscopy and mass spectrometry studies on other coronaviruses, including SARS-CoV (Mahmoudabadi 2017; Barcena 2009; Neuman 2006; Neuman 2008; Neuman 2011). The resulting protein stoichiometry is further checked against recent estimates specific to SARS-CoV-2 (Bar-On 2020). Composition of protein and virus genome sequences are obtained from the National Centre for Biotechnology Information (NCBI) nucleotide database (accession number NC_045512). Energetic costs in form of ATP stoichiometry is calculated as above. The final SARS-CoV-2 biomass function is provided in Supplementary File 3 , while the computer code used to calculate is made available (see Code Availability ).
Simulation of the metabolic model. The integrated genome scale metabolic model was simulated using the flux balance analysis (FBA) approach (Bordbar 2014). FBA assumes steady state of metabolic fluxes and implements linear optimisation to find one particular flux distribution across all reactions that can satisfy this assumption and that is optimal under given flux constraints and a certain optimality criterion. Here, we used the standard mathematical implementation of flux balance analysis as described before (Bordbar 2014) and used maximisation of flux through the host maintenance or viral biomass pseudo reactions. All reaction flux constraints are kept as in the original RECON2.2 model except for extracellular transport reactions. The extracellular transport reactions are normally set to carry negative flux to represent uptake of metabolites from the media. In the RECON2.2 model all extracellular transport reactions’ minimum flux values are set to -1000 mmol∙gDW-1∙h-1 (where DW stands for grams of dry weight) to represent a rich media (all exchange reactions allowed to carry flux). We have used here both this approach and additionally implemented a minimal media containing only essential metabolites, carbon and nitrogen source, and oxygen. The identification of the minimal media was achieved using a linear optimisation based algorithmic approach (Senior 2017), where a pseudo currency metabolite is added to all exchange reactions of the model and the flux for the extracellular transport reaction of this pseudo metabolite is systematically altered to identify a minimal set of exchange reactions that can still result in model optimisation. To implement the minimal media the identified extracellular transport reactions’ minimum flux values were set to -1000 or to -10 mmol∙gDW-1∙h-1, with all other extracellular transport reactions’ minimum flux set to zero. The identified media composition is provided as Supplementary File 6 and a computational implementation of the described minimal media identification approach is provided in Python (see Code Availability )