Systems-level metabolic flux profiling identifies fatty acid synthesis as a target for antiviral therapy

Glycolytic, pentose phosphate and nucleotide flux

Flux profiling was conducted 48 h after HCMV infection, when large changes in metabolite concentrations first appear¹⁸. A battery of five assays was used: (i) direct measurement of key cellular metabolic influxes (glucose and glutamine) and effluxes (pyruvate, lactate, alanine and glutamate), (ii) kinetic flux profiling with uniformly ¹³C-labeled glucose, (iii) kinetic flux profiling with uniformly ¹³C-labeled glutamine, (iv) probing of the branch point between glycolysis and the pentose phosphate pathway (PPP) using [1,2-¹³C]glucose and (v) probing of the branch point of pyruvate oxidation versus carboxylation using [3-¹³C]glucose. The kinetic flux profiling experiments involved quick transfer of cells from an unlabeled to a labeled carbon source by aspiration of unlabeled medium and its replacement with otherwise identical medium containing the labeled carbon source. Samples were collected at various time points after the isotope switch for analysis by LC-MS/MS.

HCMV infection increased the uptake of glucose and glutamine and the excretion of lactate and glutamate (Fig. 1a). Consistent with the increased glucose uptake and lactate excretion, kinetic profiling of glycolysis indicated increased glycolytic flux: labeled glucose was more rapidly converted into labeled glycolytic intermediates in HCMV-infected than in mock-infected cells (Fig. 1b). Labeling of the five carbon species of the PPP was similar in infected and uninfected cells (Fig. 1c). Consistent with viral upregulation of glycolysis relative to the PPP, viral infection reduced the fraction of lactate with a single ¹³C-labeled carbon after feeding of [1,2-¹³C]glucose (Fig. 1d). Such singly-labeled lactate is formed when [1,2-¹³C]glucose is metabolized by the non-oxidative PPP but not by glycolysis. Although PPP flux was not upregulated, incorporation of ribose-5-phosphate into nucleotides was increased (Fig. 1e), consistent with virally induced upregulation of nucleotide biosynthesis¹⁶.

Tricarboxylic acid cycle and citrate shuttle

A particularly marked increase in labeling from [¹³C]glucose occurred for the first compound of the TCA cycle, citrate, which accumulated more than 20-fold faster in infected cells than in mock-infected cells (Fig. 2a). This reflected a larger citrate pool in the infected cells, as well as more rapid and complete labeling of that pool (compare Fig. 2b,c). Although almost all citrate in virally infected cells became labeled with ¹³C from glucose within 15 min, less than 30% of the malate pool was labeled (Fig. 2d,e). In contrast, [¹³C]glutamine (but not glucose) labeled citrate, α-ketoglutarate and malate all to a similar extent (Figs. 2e,f). This suggested that some of the carbon passing from glucose to citrate was redirected from the TCA cycle.

Citrate, in addition to being a TCA cycle intermediate, also shuttles two carbon units from the mitochondrion to the cytosol, where they are used for fatty acid and cholesterol biosynthesis. The observed TCA cycle labeling patterns (Fig. 2) suggested virus-induced activation of this shuttle. As shown schematically in Figure 2g, after [¹³C]glucose feeding, the citrate shuttle transfers labeled acetyl units (derived from glucose) from the mitochondrion to the cytosol through citrate with two ¹³C atoms. Four- and five-carbon compounds of the TCA cycle do not become labeled. Unlabeled cytosolic oxaloacetate, released after cytosolic cleavage of citrate, is reduced by NADH to form malate, which is either decarboxylated to form pyruvate and NADPH or pumped back into the mitochondrion to reenter the TCA cycle²⁴.

Absolute quantitation of metabolite concentration and flux

Although manual inspection of labeling data is a useful tool to generate hypotheses, the numerous fluxes that feed to and from the TCA cycle complicate data interpretation. These include formation of oxaloacetate through pyruvate carboxylase, which we probed directly using [3-¹³C]glucose (Supplementary Fig. 1 online) and fluxes to and from TCA-related amino acids. Of particular importance is the rapid interconversion of α-ketoglutarate and glutamate, an abundant intra-cellular metabolite. The large pool of glutamate can dilute labeled carbon atoms coming from glucose, rendering qualitative interpretation of TCA cycle labeling patterns unreliable²⁵. To achieve a more quantitative understanding of central carbon metabolic flux and its modulation by HMCV infection, we developed an ordinary differential equation model (Fig. 3 and Supplementary Fig. 2 online) that predicts isotope labeling kinetics and patterns under the assumptions of fixed metabolic fluxes and pool sizes—that is, flux balance²⁶.

We began by experimentally determining the absolute pool sizes of metabolites in the model. The approach involved extensive labeling of intracellular metabolites by simultaneously feeding uniformly ¹³C-labeled glucose and glutamine for 1 week in culture²⁷. We then extracted the labeled intracellular metabolites in the presence of known concentrations of unlabeled internal standard compounds, and determined the ratios of labeled to unlabeled compounds by LC-MS/MS. Corrections were applied to account for incomplete labeling of certain metabolites (Supplementary Table 1 online). The resulting values are shown in Supplementary Table 2 online. The use of internal standards throughout extraction, sample handling and analysis substantially enhanced the reliability of these measurements, compared to measurements based on extracted quantities per se.

We then used the pool size data, kinetic flux profiling data for glucose and glutamine, metabolite influx and efflux data, and specific branch-point data to search for combinations of fluxes consistent with the experimental results. Complete experimental results are shown in Supplementary Tables 2–4 and Supplementary Figures 3 and 4 online. Fluxes consistent with the experimental data were obtained using a global parameter identification algorithm, with a cost assigned to flux sets that did not recapitulate observed experimental data within ± 2 s.e. of their mean²⁸ (Supplementary Methods online). Treating all fluxes that fit the experimental data within the 95% confidence interval as equivalent is a crucial precaution against overfitting, a common problem in flux elucidation²⁹. To further mitigate the risk of overfitting, instead of identifying a single set of fluxes that best recapitulated the experimental results, we identified 100 different flux sets that approximated the observed data. This enabled estimation of confidence limits on the observed fluxes. Median flux values and their confidence limits for uninfected versus virally infected cells are shown in Supplementary Table 5 online.

Absolute pool sizes, absolute fluxes and their fold change after HCMV infection are shown schematically in Figure 3. Pools and fluxes scale exponentially with their diagrammed sizes, with marked differences in absolute pool sizes and fluxes across metabolites and reactions. The most active pathway in uninfected cells was glycolysis, and the majority of pyruvate produced was excreted as lactate. TCA cycle flux was ~1% of glycolytic flux and was driven largely by inflow from glutamine to α-ketoglutarate through glutamate.

Global metabolic changes induced by HCMV

Colors in Figure 3 indicate changes in pool sizes and fluxes induced by HCMV infection, with increases in red and decreases in blue. Notably, almost the entire pathway diagram (with the exception of the PPP) is red, indicating a nearly global upregulation of metabolic flux by HCMV. Glycolytic flux increased approximately two-fold (Fig. 3 and Supplementary Table 5), consistent with the enhanced glucose uptake in virus-infected cells¹⁷. Nucleotide biosynthetic flux increased approximately three-fold, consistent with the efficacy of nucleotide-based antimetabolites as HCMV therapeutics³⁰. TCA cycle fluxes increased even more markedly, with substantial upregulation of the core cycle fluxes (for example, ~20-fold from citrate to α-ketoglutarate), the pyruvate dehydrogenase pathway feeding the cycle (~80-fold from pyruvate to acetyl-CoA), the anapleuretic pathway for formation of four-carbon units (~4-fold from pyruvate to oxaloacetate) and the efflux of two-carbon units to fatty acid biosynthesis (~20-fold from citrate to malonyl-CoA). Consistent with upregulation of the fatty acid biosynthetic pathway, malonyl-CoA—the committed intermediate in fatty acid synthesis—increased in concentration from undetectable levels to amounts greater than ten-fold above the detection limit (Fig. 4a).

Fatty acid biosynthesis as an antiviral target

Many of the upregulated fluxes reflect essential processes that are not viable antiviral targets. For example, impairing canonical steps of the TCA cycle would be likely to lead to serious metabolic complications in oxidation-dependent cell types, such as cardiomyocytes and neurons. In contrast, fatty acid synthesis—although important in growth and development—is not acutely essential in mammals^31–33. Given the substantial upregulation of fatty acid synthesis, we wished to explore its role in viral infection further. To test more directly whether viral infection increases glucose-driven lipid biosynthesis, we fed uninfected or HCMV-infected fibroblasts ¹⁴C-labeled glucose. We then extracted and saponified the lipids and counted radioactivity in the resulting lipid-derived fatty acid and glycerol fractions. The resulting data confirmed that HCMV infection induced a marked increase in glucose assimilation into both the fatty acid and glycerol moieties of lipids (Fig. 4b,c).

We next used pharmacological inhibitors of the fatty acid biosynthetic enzymes acetyl-CoA carboxylase (ACC) and fatty acid synthase (FAS) to determine whether HCMV-induced upregulation of fatty acid production is necessary for viral replication (Fig. 5a). Treatment with 5-tetradecyloxy-2-furoic acid (TOFA), an ACC inhibitor³⁴, resulted in a dose-dependent inhibition of HCMV replication with a >1,000-fold effect at 10 μg ml⁻¹ TOFA. Treatment with C75 (trans-4-carboxy-5-octyl-3-methylene-butyrolactone), an inhibitor of FAS³⁵, resulted in similar dose-dependent inhibition, with a >100-fold effect at 10 μg ml⁻¹. Taken together, these results indicate that fatty acid biosynthesis is necessary for normal HCMV replication.

HCMV infection induces a temporally coordinated cascade of viral gene expression and subsequent protein synthesis. To determine whether inhibition of fatty acid biosynthesis affects this cascade, we analyzed the accumulation of the immediate early HCMV protein IE1, the early protein UL26 (which consists of two isoforms) and the late protein pp28 (Fig. 5b). Accumulation of all three proteins was unimpaired by C75 or TOFA treatment. Because the accumulation of pp28 depends on viral DNA synthesis³⁶, these results indicate that the defect in viral growth caused by inhibition of fatty acid biosynthesis occurs subsequent to DNA replication.

Inhibition of FAS as a result of malonyl-CoA accumulation can be toxic and has been shown to induce apoptosis in some transformed cell lines³⁷. The normal accumulation of viral proteins in the presence of C75 and TOFA suggested that the tested drug concentrations were not cytotoxic under the present conditions. To test directly for host cell cytotoxicity, we exposed uninfected fibroblasts to C75 or TOFA for 96 h (the length of the viral growth assays) and analyzed for viability by Trypan blue exclusion and by assessing cellular growth after drug removal. Both endpoints were identical in treated and untreated cells. To test directly for apoptosis induction, we examined the cleavage of the caspase-3 target poly(ADP-ribose) polymerase (PARP). Cleavage of PARP was undetectable in mock- or HCMV-infected cells after 24 or 72 h of TOFA or C75 treatment, whereas induction of apoptosis through osmotic shock with 0.5 M sorbitol resulted in the 85-kDa fragment characteristic of caspase-mediated PARP cleavage (Supplementary Fig. 5 online). Taken together, these results indicate that the tested doses of TOFA and C75 block HCMV replication without inducing host cell toxicity or apoptosis.

The requirement of fatty acid biosynthesis for a late event in HCMV replication may reflect an essential role for de novo synthesized fatty acids in viral envelopment or in lipid modification of viral proteins. To investigate whether this requirement extends to other enveloped viruses, we assessed the growth of influenza A in the presence of the ACC inhibitor TOFA or the FAS inhibitor C75. TOFA inhibited influenza A replication by >1,000-fold, and C75 inhibited replication by >10-fold (Fig. 5c). Thus, inhibitors of fatty acid biosynthesis impair the normal replication of two evolutionarily diverged enveloped viruses.

Biological reagents and cell culture

MRC-5 fibroblasts (ATCC) were cultured in Dulbecco modified Eagle medium (DMEM, Sigma) containing 7.5% fetal calf serum and 4.5 g l⁻¹ glucose. Before infection, fibroblasts were grown to confluence in 10-cm dishes, resulting in ~1.5 × 10⁶ cells per dish. After incubation for 3–5 d at confluence, serum-containing medium was removed and serum-free medium added. Cells were then maintained in serum-free DMEM for 24 h, which has been shown to synchronize cells in the G₀ stage of the cell cycle¹⁸. Cells were then mock-infected or infected with HCMV at a multiplicity of infection (MOI) of 3.0 plaque-forming units (PFU) per cell. We used the HCMV strain BADwt, which is derived from a bacterial artificial chromosome (BAC) clone of the AD169 strain of HCMV⁵³. The BAC was inserted into the genome of HCMV without deletion of any viral sequence and was excised by a cotransfected Cre recombinase, which mediates recombination at the loxP sites flanking the BAC, leaving only the loxP site in the viral clone. This clone has been tested in various assays and has always shown a wild-type AD169 phenotype. After a 2-h adsorption period, the viral inocula were aspirated and fresh serum-free DMEM was added.

TOFA (Biomol International) and C75 (Calbiochem) were maintained as 10 mg ml⁻¹ stocks in DMSO. To assess HCMV growth in the presence of metabolic inhibitors, densely confluent fibroblasts were infected in the presence of inhibitors at the indicated dose in DMEM containing 10% serum. After 2 h of incubation to allow viral adsorption, unbound virus was inactivated through a sodium citrate wash (40 mM sodium citrate, 10 mM KCl and 135 mM NaCl, pH 3.0) followed by a DMEM wash and then incubated at the indicated inhibitor dose in DMEM containing 10% serum. Forty-eight hours after infection, the cellular medium was changed, and fresh medium and metabolic inhibitors were added. At 96 h, cells were scraped and viral titers were determined by standard plaque assay on MRC-5 cells. To assess influenza A growth, fully confluent Madin-Darby canine kidney epithelial cells (MDCK cells; ATCC) were infected with the A/WSN/33 influenza strain (ATCC) in Flu infection buffer (DMEM containing 0.2% BSA, 0.01% CaCl₂, 0.01% MgCl₂, 1 μg ml⁻¹ trypsin (Worthington) and 0.1% fetal calf serum). Viral titers were determined by standard plaque assay on MDCK cells.

Potential toxicity of TOFA and C75 was analyzed by treating uninfected cells with various doses of pharmaceutical inhibitors in DMEM (for fibroblasts) or Flu infection buffer (for MDCK cells) for the same duration as the viral growth assay (96 h for fibroblasts and 24 h for MDCK cells). After incubation with inhibitor, treated cells were analyzed for Trypan blue exclusion and found to be >95% viable. Inhibitor-treated cells were also replated in the absence of inhibitor and found to grow with kinetics similar to those of untreated cells. For MDCK cells, 50 μg ml⁻¹ of TOFA was used for the influenza A growth assay, as this dose was found to be well tolerated by MDCK cells. Lower TOFA doses may be effective in nontransformed host cells, as cancer cells generally have strongly upregulated de novo lipid biosynthesis.

Metabolomic experiments

Labeled DMEM medium was prepared from DMEM without glucose or glutamine (Sigma) by addition of 10 mM HEPES and the appropriate forms (labeled or unlabeled) of glucose and glutamine to a final concentration of 4.5 g l⁻¹ glucose and 0.584 g l⁻¹ glutamine (labeled glucose and labeled glutamine from Cambridge Isotope Laboratories), followed by sterile filtration. For kinetic flux profiling experiments, samples were switched to fresh unlabeled medium 1 h before the switch into ¹³C-labeled medium. This minimized metabolome perturbations at the time of the isotope switch resulting from removal of accumulated metabolic waste products. Metabolome quenching and extraction were conducted as previously described¹⁸. Absolute metabolite quantitation involved extended labeling of cellular metabolites with uniformly labeled [¹³C]glucose and [¹³C]glutamine and extraction in the presence of known concentrations of unlabeled standards (for details, see refs. ^23,27 and Supplementary Methods).

Uptake of glucose and glutamine and excretion of all measured metabolites (excretion of pyruvate, lactate, alanine and glutamate were found to be significant) was determined from medium samples taken over an 8-h time period, centered at 48 h after infection. Glucose was measured by enzyme assay (E00715251, R-Biopharm). The other compounds were measured by LC-MS/MS, with inclusion of isotopic internal standards for glutamine, glutamate, pyruvate, lactate and alanine.

Estimation of the relative carbon flux between glycolysis and the PPP was carried out using [1,2-¹³C]glucose as described³⁸, with 4 h of incubation in labeled medium and detection of labeled forms of lactate by LC-MS/MS. Estimation of the relative carbon flux between pyruvate dehydrogenase and pyruvate carboxylase was carried out based on the passage of labeled carbon from [3-¹³C]glucose into malate, aspartate and citrate over 6 h (for details, see Supplementary Methods).

Liquid chromatography–tandem mass spectrometry analysis

Two different LC separations were coupled by electrospray ionization (ESI) to triple-quadrupole mass spectrometers operating in multiple reaction monitoring mode. The LC method coupled to positive-mode ESI was hydrophilic interaction chromatography on an aminopropyl column at basic pH, as described⁵⁴. The LC method coupled to negative-mode ESI was reversed-phase chromatography using an amine-based ion pairing agent (a variation of the method used in ref. 55). The stationary and mobile phases were identical, but the gradient was altered as follows: t = 0, 0% B; t = 5, 0% B; t = 10, 20% B; t = 20, 20% B; t = 35, 65% B; t = 38, 95% B; t = 42, 95% B, t = 43, 0% B; t = 50, 0% B; where B refers to the methanol-containing mobile phase. For LC, we used an LC-20 AD HPLC system (Shimadzu) with autosampler temperature of 4 °C and injection volume of 10 μl. For MS, we used a TSQ Quantum Ultra or Discovery Max triple-quadrupole mass spectrometer (Thermo Fisher Scientific). Mass spectrometry parameters were as described in ref. 54, with the addition of extra multiple reaction monitoring scans to measure partially labeled compounds. All data were corrected for the natural abundance of ¹³C (see Supplementary Methods).

Measurement of lipid synthesis

To directly examine the flux of glucose into lipids, we transferred fibroblasts (infected or mock-infected for 48 h) into medium containing radioactive glucose (8 μC ml⁻¹ [¹⁴C]glucose, 1 g l⁻¹). After incubation for 4 h, the culture medium was aspirated, cells were washed with PBS and phospholipids were extracted by addition of 500 μl of 60:40 hexane:isopropanol. The culture dishes were then washed with an additional 400 μl of the hexane:isopropanol mixture. The resulting total extract was dried under N₂ gas, resuspended in 500 μl of 1 N KOH in 90:10 methanol:water and incubated at 70 °C for 60 min to saponify lipids. Sulfuric acid (100 μl, 2.5 M) was then added, followed by hexane (700 μl) to extract the saponified fatty acids. The organic and aqueous phases were separated by centrifugation and scintillation-counted.

Computational determination of fluxes

An ordinary differential equation model of central carbon metabolism was constructed based on the diagram in Supplementary Figure 2. The model consisted of 69 differential equations, written to maintain flux balance. Equations of the model described the rates of loss of unlabeled forms of metabolites (and the creation of particular labeled forms) after feeding of [U-¹³C]glucose or [U-¹³C]glutamine media, and were written in the following manner:

where metabolite B is directly downstream in the model from metabolites A₁, A₂, …, A_N, F^tot is the total flux into (and equivalently out of) the total B pool, B^X is the pool size of a specific isotopic form of compound B; F_i is the i^th flux into B; $A_i^X$ is the sum of all isotopic forms of A_i that feed into B^X through flux F_i, $A_i^{\text{tot}}$ is the total pool size of A_i (that is, the sum of all labeled and unlabeled forms of A_i), B^tot is the total pool size of B and N is the number of fluxes into B. The model did not explicitly include nucleotide oxidation or reduction or phosphate transfer reactions (such as NAD⁺ reduction or ATP hydrolysis), as the rates of these reactions cannot be deduced using ¹³C-tracers.

Parameters (fluxes and unmeasured concentrations) were identified by a genetic algorithm implemented in C/C++ that seeks parameter values that minimize the differences between the experimental observations and the computational results. Costs were applied only when the model output fell outside of the 95% confidence limits (mean ± 2 s.e.) of the laboratory data. A series of 20 independent genetic algorithm runs were conducted on both the mock and virally infected data to obtain the distribution of best flux estimates shown in Supplementary Tables 5–7 online. For computational details, see Supplementary Tables 8–10 online, Supplementary Methods and Supplementary Computational Code online.

In creation of the flux-balanced model, amino acid efflux to protein biosynthesis was denoted by the symbol X (as shown in Supplementary Fig. 2). The efflux of each amino acid in the model was written as a fraction of X, where X is the sum of efflux from all amino acids not present in the medium (alanine, aspartate, asparagine, glutamate and proline) and glutamine to protein. These fractions were estimated based on the relative abundance of the amino acid in protein, including in the human proteome, in the HCMV proteome⁵⁶ and in collagen. The abundance ratios were approximated as 6:4:3:5:5:4 (Ala:Asp:Asn:Glu:Pro:Gln) for both infected and uninfected cells. The fraction shown in Supplementary Figure 2 for glutamate (10/27) accounts for glutamate and its product proline, and that shown for aspartate (7/27) also accounts for asparagine. As shown in Supplementary Figure 2, X is fully constrained by other model fluxes and therefore does not appear in the model as a parameter.

Protein analysis

Protein accumulation was assayed by western blot analysis. Cells were washed with PBS, scraped and solubilized in disruption buffer (50 mM Tris (pH 7.0), 2% SDS, 5% 2-mercaptoethanol and 2.75% sucrose). Resulting extracts were sonicated and centrifuged at 14,000 g for 5 min to pellet insoluble material. Equivalent fractional amounts were subjected to electrophoresis in an SDS-containing 10% polyacrylamide gel and transferred to a nitrocellulose sheet. Blots were then stained with Ponceau S to ensure equivalent protein loading and transfer, blocked by treatment with 5% milk and reacted with primary antibody. Antibodies used were specific for UL99-encoded pp28 (10B4-29; ref. 57), UL123-encoded IE1 (1B12; T.S., unpublished data), UL26 (7H19; ref. 58), PARP (Santa Cruz Biotechnology) or tubulin (Epitomics). Protein bands were visualized through enhanced chemiluminescent detection according to the manufacturer’s instructions (Pierce). Analysis of tubulin levels served as an additional protein loading control.