Abstract

   Mycobacteria adapt to infection stresses by entering a reversible
   non-replicating persistence (NRP) with slow or no cell growth and broad
   antimicrobial tolerance. Hypoxia and nutrient deprivation are two
   well-studied stresses commonly used to model the NRP, yet little is
   known about the molecular differences in mycobacterial adaptation to
   these distinct stresses that lead to a comparable NRP phenotype. Here
   we performed a multisystem interrogation of the Mycobacterium bovis BCG
   (BCG) starvation response, which revealed a coordinated metabolic shift
   away from the glycolysis of nutrient-replete growth to depletion of
   lipid stores, lipolysis, and fatty acid ß-oxidation in NRP. This
   contrasts with BCG’s NRP hypoxia response involving a shift to
   cholesterol metabolism and triglyceride storage. Our analysis reveals
   cryptic metabolic vulnerabilities of the starvation-induced NRP state,
   such as their newfound hypersensitivity to H[2]O[2]. These observations
   pave the way for developing precision therapeutics against these
   otherwise drug refractory pathogens.

   Subject terms: Pathogens, Biochemical networks, Bacterial pathogenesis
     __________________________________________________________________

   A multi-omics comparison of the mycobacterial response to hypoxia and
   starvation reveals that distinct differences in fatty acid and
   cholesterol metabolism lead to the same non-replicating, drug-tolerant
   persistence phenotype.

Introduction

   Mycobacterium tuberculosis (Mtb) remains the single most lethal human
   bacterial pathogen in the world^[46]1. With humans as its only host,
   Mtb adapts to the stress of infection by entering a state of
   non-replicating persistence (NRP) marked by limited cell division,
   immune quiescence, and increased antibiotic tolerance^[47]2,[48]3. The
   molecular mechanisms underlying the transition to and survival in NRP
   remain poorly defined, however. One clue may lie in the fact that Mtb,
   in contrast to other bacteria, devote a significant portion of genomic
   coding capacity to enzymes involved in lipogenesis and lipolysis, with
   growing evidence that lipid biosynthesis and fatty acid metabolism play
   key roles in Mtb’s pathogenicity^[49]4,[50]5. During acute infection,
   mycobacteria survive by subverting phagolysosome biogenesis and
   acidification, progressively gaining access to the cytosol and
   modulating host immune responses^[51]6. During chronic infection, it is
   believed that mycobacteria sequester themselves within macrophages,
   where oxygen and nutrient deprivation trigger the NRP phenotype. While
   the hypoxic response has been well characterized in members of the Mtb
   complex, our understanding of how mycobacteria adapt to nutrient
   starvation remains incomplete^[52]7.

   Starvation of mycobacteria has been shown to cause altered central
   carbon metabolism^[53]8, global transcriptional reprogramming^[54]9,
   the induction of virulence-enhancing toxin–antitoxin modules^[55]10,
   altered permeability to drugs and metabolites^[56]11,[57]12, and
   phenotypic resistance to mechanistically diverse
   antibiotics^[58]13,[59]14. Despite extensive study, previous
   investigations of mycobacterial starvation have predominantly relied on
   monitoring changes in gene expression triggered by growth on single
   carbon sources^[60]15, collating phenotypic and essentiality
   differences among single-gene mutants^[61]16, and mapping transcription
   factor binding sites across the genome^[62]17. The diversity of
   experimental conditions makes it difficult to compare findings between
   these studies, and few attempts have been made to compare responses to
   starvation and hypoxia in the same study, which makes it difficult to
   define how metabolic adaptations to two very different stresses lead to
   the same NRP phenotype.

   In this work, we describe an exhaustive set of systems-level analyses
   that all point to unique shifts in lipid metabolism as a feature that
   distinguishes the mycobacterial response to hypoxia and starvation
   despite the common NRP endpoint. Here we use unsupervised multivariate
   analyses to integrate a variety of biochemical parameters, phenotypic
   metabolic profiling, time-course metabolomics, longitudinal RNA-seq,
   and quantitative proteomics. These analyses revealed that, in contrast
   to a hypoxia-induced shift to cholesterol metabolism and triglyceride
   accumulation, starvation induces a metabolic shift to β-oxidation of
   fatty acids and a resulting ketotic state. This transition induces a
   unique vulnerability to H[2]O[2] that is not observed in
   hypoxia-induced NRP^[63]18. Together, these findings underscore the
   power of integrated systems approaches to understanding how pathogens
   respond to stress and suggest that the metabolomic heterogeneity of the
   NRP state could present an exploitable vulnerability for the
   development of novel TB chemotherapies.

Results

Starved mycobacteria exhibit hallmarks of NRP

   To initiate our mechanistic analysis, we first established a suitable
   experimental model for starvation-induced NRP, which was consistent
   with our previous studies of hypoxia-induced NRP in Mycobacterium bovis
   BCG str. Pasteur 1173P2 (BCG), a member of the Mtb Complex. Consistent
   with other experimental models of bacterial persistence^[64]19,
   nutrient deprivation by culturing M. tuberculosis CDC1551 (Mtb), BCG,
   and Mycobacterium smegmatis MC^2155 in phosphate-buffered saline (PBS)
   resulted in biphasic death kinetics characterized by a rapid decrease
   in both OD[600] and CFU within the first 4 days of starvation (S4),
   followed by a slower decline to 10–30% of initial CFU (S10-S30 for BCG,
   Fig. [65]1a). Using flow cytometry, we observed coalescence of
   population forward- and side-scatter values in starved BCG cultures
   (Supplementary Fig. [66]1a and Supplementary Table [67]1), suggesting
   that mycobacteria assume smaller overall cell size and reduced cellular
   complexity during prolonged nutrient deprivation. These
   starvation-induced morphological transitions are consistent with
   previous studies of MTB isolated from lung lesions^[68]6. Intriguingly,
   we found divalent cations to be major determinants of mycobacterial
   survival during starvation, with growth in divalent cation-free PBS
   (compared to Dulbecco’s PBS containing Mg^2+, Ca^2+) required to induce
   NRP in MTB, BCG, and SMG in vitro (Supplementary Fig. [69]1b, c). When
   starved in Mg^2+- and Ca^2+-containing DPBS, significantly more BCG and
   SMG survived compared to starvation in PBS (Supplementary Fig. [70]1b,
   c). This detail may explain why previous starvation studies report no
   loss of MTB viability during growth in Mg^2+- and Ca^2+-containing PBS
   (i.e., Dulbecco’s PBS)^[71]9,[72]14. We further assessed starved
   cultures for the possibility that nonviable cells (~95% of initial CFU)
   served as a carbon source for viable starved BCG persisters in culture.
   Washing and re-culturing the cells in fresh PBS did not result in
   increased cell death (Supplementary Fig. [73]1d), which rules out
   cannibalism as the basis for NRP, and cells retained their ability to
   regrow (“resuscitate”) in nutrient-replete 7H9 media (resuscitation day
   6 or R6, Fig. [74]1b). Thus, it appears that neither a transition to a
   viable but nonculturable state, nor cannibalism alone, accounted for
   mycobacterial survival over prolonged periods of
   starvation^[75]20,[76]21.

Fig. 1. Characteristics of starved mycobacterial persisters.

   [77]Fig. 1
   [78]Open in a new tab

   a Survival profiles (OD[600], CFU) of MTB, BCG, and SMG during nutrient
   deprivation in PBS. b Recovery of BCG after 4, 10, 20, or 30 days of
   starvation in PBS (S4, S10, S20, and S30, respectively) followed by 2
   (white), 6 (gray), and 10 days (black) of resuscitation in
   nutrient-replete media; MTB after 20 days of starvation followed by 2
   days (white) and 6 days (gray) of resuscitation; and SMG after 6 days
   of starvation followed by 9 h (white) and 24 h (black) of
   resuscitation. Doubling times are not significantly different by
   one-way ANOVA with post hoc Tukey's HSD. Data are shown as mean ± SE
   for n ≥ 3. c Reduction in total cellular RNA levels in BCG at S4, S10
   and S20, with restoration at R6. d Transcriptional induction of the
   stringent response associated genes relBCG, relBE2, mazEF6, vapBC3, and
   vapBC31 in BCG, and repression of cell division marker ftsZ measured by
   RNA-seq. e Expression of relMTB in BCG measured by qPCR. f Antibiotic
   susceptibility of Log (◆), S4 ( Inline graphic ), S10 ( Inline graphic
   ), S20 ( Inline graphic ), and R6 ( Inline graphic ) BCG after 48 h of
   exposure. Data represent mean ± SD for n ≥ 6. c–e Data represent
   mean ± SD for n ≥ 3; *P < 0.05 determined by one-way ANOVA with
   Dunnett's test vs. Log. See Supplementary Data [79]4 for the source
   data used in these graphs.

   After establishing that starved mycobacteria remain viable, we
   characterized starved BCG for the transcriptional, phenotypic,
   metabolomic, and proteomic hallmarks of bacterial persistence^[80]3.
   Consistent with the observed transcriptional induction of RelA-SpoT
   homologs (RSH) in the stringent responses of other bacteria^[81]22,
   mRNA levels for rel[BCG] and rel[MTB]—genes that control intracellular
   levels of pp(p)Gpp in BCG and MTB—peaked in early starvation (S4)
   (Fig. [82]1d and Supplementary Data [83]1) and were associated with a
   concomitant 40% reduction in total RNA content per cell relative to
   log-growing bacilli (Fig. [84]1c and Supplementary Figs. [85]1e and
   [86]2a). Total RNA levels decreased at S4, remained low over the course
   of starvation, and returned to pre-starvation levels upon 6 days of
   resuscitation in nutrient-replete 7H9 media (R6, Fig. [87]1c).
   Similarly, we observed the induction of several important
   toxin–antitoxin modules (relBE2, mazEF6, vapBC3, vapBC31) and reduced
   expression of the cell division gene ftsZ at S4 (Fig. [88]1d). Starved
   BCG cultures also developed tolerance against bactericidal doses of the
   conventional antitubercular antibiotics rifampicin, streptomycin,
   ethambutol and isonaizid (Fig. [89]1f). Importantly, the drug-tolerant
   phenotype was reversed by resuscitation in 7H9 medium, which is
   consistent with the phenotypic drug tolerance that is a hallmark of
   NRP^[90]13,[91]14.

   To characterize the molecular changes induced by starvation in BCG, we
   performed quantitative proteomics to longitudinally profile the time
   course of NRP in starved BCG, with the detection of 1102 proteins
   common to three biological replicates in all time points (Supplementary
   Data [92]2 and Supplementary Fig. [93]3a). A principal component
   analysis (PCA) of the proteomic changes during starvation showed
   subsets of proteins strongly distinguishing the four time points
   (Supplementary Fig. [94]3b), with KEGG pathway enrichment analysis for
   significantly upregulated (>1.3-fold) proteins underscoring phenotypic
   differences between early (S4) and late (S20) non-replicating
   persistence and resuscitation (R6). For example, starvation caused a
   significant shift away from cell wall anabolism (e.g., CmaA2, EmbB,
   MmaA2, and GlmU as negative predictors of the starvation response) and
   toward lipid catabolism (e.g., FabD2, FadE16, GlgE and BCG_0220 as
   positive predictors of starvation), with the exception of upregulated
   virulence lipid biogenesis factors phthiocerol dimycocerosates (PDIM;
   e.g., BCG 2974, PpsB, and PpsD) and mannosyl-β-1-phosphomycoketide
   (MPM; e.g., Pks12 and Pks13). BCG 2974 is a trans-acting enoyl
   reductase in the biosynthesis of PDIMs and glycosylated
   phenolphthiocerol dimycocerosates, which are major cell-surface
   virulence factors of Mtb during the early macrophage invasion step of
   infection^[95]23. KEGG analysis of the proteomics data also identified
   protein predictors that modulate host immunity (Eis), participate in
   energy homeostasis (AtpH, CoaE, CtaD, GlgP, PckA, and PgmA), maintain
   proteome integrity (HtpX, PepC, PepN, PepQ, PpiA2, and PpiB), sequester
   metal cations (CtpI, HemB and HemC), and, as discussed in detail
   shortly, serve putative roles in fatty acid β-oxidation (Acs, FadB2,
   FadB4, FadB5, FadE2, FadE16, FadE23, FadE24, FadE36, EchA8, EchA9,
   EchA16, FadA, FadA2, and FadA4). The coordinated upregulation of lipid
   dehydrogenases and oxidoreductases suggests that persisters upregulate
   fatty acid metabolism during nutrient deprivation. On the other hand,
   our analysis identified fatty acid biosynthetic enzymes AcpP and FadD26
   as strong contravariants of the starvation response, suggesting
   downregulation of de novo fatty acid synthesis during nutrient
   deprivation. While oxaloacetate would typically facilitate the
   integration of Ac-CoA from fatty acid β-oxidation into the
   tricarboxylic acid (TCA) cycle, our proteomic results indicate the
   induction of alternative metabolic pathways during starvation. For
   instance, the general upregulation of thiolase II enzymes (e.g., FadA)
   suggests increased metabolic flux towards acetoacetate (AcAc)^[96]24.

   These systematic analyses demonstrate that well-controlled nutrient
   deprivation predictably induces a stable, phenotypically drug-tolerant
   sub-population in MTB, BCG, and SMG cultures, with BCG serving as a
   slow-growing experimental surrogate of MTB. While both hypoxia and
   nutrient deprivation lead to the same drug-tolerant NRP phenotype, it
   is unclear how mycobacteria adapt to these significantly different
   stresses. We addressed this issue with a comparative multi-omic
   analysis.

Proteomics and transcriptional profiling predict distinct metabolic
adaptations in starvation- and hypoxia-induced NRP

   Here, we sought to discern how the functional expression of
   mycobacterial genes differed between the BCG model for
   starvation-induced NRP and an established BCG in vitro model for
   hypoxia-induced NRP^[97]25. We first compared the proteomics data for
   the BCG starvation time course (1102 proteins) with our previous
   analysis of the hypoxic BCG proteome^[98]26 (966 proteins;
   Supplementary Data [99]2 and Fig. [100]2a). The correspondence between
   the two stress time courses is as follows: exponentially growing
   bacilli (Log), early adaptive NRP response (NRP1; day 4 of starvation,
   S4; day 4 of hypoxia, W4), enduring NRP response (NRP2; days 10 and 20
   of starvation, S10 and S20; days 9 and 18 of hypoxia, W9 and W18), and
   resuscitation (day 6 of nutrient restoration, S-R6; day 6 of normoxia,
   W-R6). We identified 379 proteins that had quantitative coverage across
   all 5 related time points in both hypoxia and starvation proteomic
   datasets (Supplementary Data [101]2 and Fig. [102]2a). The two datasets
   differ strikingly in several respects.

Fig. 2. Functional gene responses are different between starvation- and
hypoxia-induced NRP.

   [103]Fig. 2
   [104]Open in a new tab

   a Proteins detected and quantified across time-course proteomic
   analyses of starvation- (green circle: 1102 proteins) and
   hypoxia-induced (blue circle: 966 proteins) NRP. A total of 379
   proteins were quantified across all time points in both conditions. b
   Principal component analysis reveals distinct protein dynamics during
   starvation (green ellipse: S4, S10, and S20) and hypoxia (blue ellipse:
   W4, W9, and W18) responses. Protein observations are color-coded
   according to the functional category, as follows: red = lipid
   metabolism (e.g., acetyl-CoA metabolism, β-oxidation); orange =
   cholesterol metabolism (e.g., propionyl-CoA metabolism); green = cell
   wall, virulence, detoxification and adaptation; blue = redox
   homeostasis; magenta = regulatory proteins and information pathways;
   yellow = replication and translation. c Major metabolic differences
   between starvation- and hypoxia-induced NRP informed by proteomic
   datasets. While co-catabolism of carbon substrates appears to be a
   shared feature of both stresses, starvation significantly induces the
   upregulation of proteins with known functions in fatty acid metabolism.
   On the other hand, hypoxia coordinately induces the upregulation of
   proteins with documented roles in the metabolism of cholesterol and
   odd-chain fatty acids (e.g., enzymes involved in the methylmalonyl-CoA
   pathway (MCP) and methylcitrate cycle (MCC)).^[105]30 Importantly,
   starvation significantly decreased and hypoxia increased levels of
   isocitrate lyase (Icl), the key enzyme in the glyoxylate shunt. OXO
   oxaloacetate, CIT citrate, αKG α-ketoglutamate, SUC succinate, FUM
   fumarate, MAL malate. d–g Comparison of two-component system protein
   levels during hypoxia (red) and starvation (blue time courses. Plots
   contain individual log[2](fold-change) data for 1–3 replicates, thick
   bars as mean values and thin bars as standard deviation on 0, 4, 9, and
   19 days of hypoxia or starvation and after 6 days of resuscitation (R)
   with normoxia or nutrient restoration. Note: peptide signals for MtrB
   in hypoxia did not meet the cutoff criteria for proteomic quantitation
   but we have added the data in (e) for completeness. See Supplementary
   Data [106]4 for the source data used in these graphs.

   First, a survey of transcriptional regulators reveals divergent
   pathways of gene expression expected for the response to the two
   distinct stressors of hypoxia and starvation. For instance, one of the
   most upregulated genes for hypoxia-induced persisters, at both the
   transcript and protein levels, is the chaperone HspX^[107]26,[108]27,
   which, surprisingly, is one of the most downregulated genes terms of
   mRNA and protein abundances for starvation-induced NRP (Supplementary
   Data [109]1 and [110]2). hspX is under the control of DevSR, which
   suggests a divergence of transcriptional responses resulting in a
   phenotypically similar persistent state. Indeed, as shown in
   Fig. [111]2d–g, six of seven transcription factors detected in both
   stress conditions were differentially regulated. The two-component (TC)
   system that regulates the early hypoxic response, DevRS, confirms the
   quantitative rigor of the proteomics analyses, with an expected large
   increase in both DevR and DevS in hypoxia and no changes during
   starvation (Fig. [112]2d). While there is a claim for an essential
   interaction between the housekeeping sigma factor SigA and
   DevR^[113]24, SigA levels decreased over the hypoxia time course and
   were unchanged in starvation (Fig. [114]2g). The parallel reduction of
   members of another TC system, PrrAB, in hypoxia but not in starvation
   (Fig. [115]2f) is consistent with Prr function in regulating
   respiratory and oxidative phosphorylation pathways^[116]28, as
   discussed shortly. The small increases in MtrAB TC proteins for both
   hypoxia and starvation (Fig. [117]2e) could reflect the role of this
   system in cell division and cell wall metabolism given the common NRP
   endpoint of no or slow growth^[118]29. Another example of the
   divergence of gene expression for hypoxia and starvation involves
   biosynthesis and transport of the phthiocerol dimycocerosate (PDIM)
   cell wall lipid. While starvation upregulated PDIM biosynthesis
   pathways and the PDIM membrane transporter MmpL7 (Supplementary
   Fig. [119]4a and Supplementary Data [120]2), hypoxia strongly reduced
   MmpL7 and PDIM synthesis enzymes in hypoxia (Supplementary Fig. [121]4a
   and Supplementary Data [122]2). These observations demonstrate
   similarities and differences in protein levels for hypoxia and
   starvation for a variety of BCG metabolic pathways, with parallel
   observations with metabolites discussed shortly.

   Despite some strong similarities in the broad KEGG pathways associated
   with these 379 proteins during the hypoxic and starvation stresses
   (Supplementary Fig. [123]3c–e), PCA of covarying proteins for the two
   stresses showed significant differences between hypoxia and starvation
   (Fig. [124]2b). Among the most striking differences was the
   starvation-induced upregulation of enzymes involved in fatty acid
   β-oxidation: acetyl coenzyme A synthetase (Acs), long-chain fatty
   acyl-CoA ligase (BCG 2974), NADPH oxidoreductase (FadB4, FadB5),
   acyl-CoA dehydrogenase (FadE2, FadE16, FadE23, FadE24, FadE36),
   enoyl-CoA hydratase (EchA8, EchA9, EchA16), and acyl-CoA thiolase
   (FadA, FadA2, and FadA4) (Supplementary Data [125]2 and Fig. [126]2c).
   The coordinated increase in lipid dehydrogenases and oxidoreductases
   suggests that persisters upregulate fatty acid metabolism during
   nutrient deprivation. In contrast, PCA showed that one of the strongest
   hallmarks of the hypoxia response was the upregulation of cholesterol
   and odd-chain fatty acid metabolism and a strong theme of cell wall
   biogenesis: upregulated propionyl-CoA decarboxylase (AccD4),
   acyl-[acyl-carrier protein] desaturase (DesA1 and DesA2), acyl-CoA
   dehydrogenase (FadE4 and FadE5), l-lactate dehydrogenase (LldD1),
   methylmalonyl-CoA (MutB), 3-ketoacyl-ACP reductase (FabG and FabG2),
   monooxygenase (EthA), isocitrate lyase (Icl), citrate synthase (GltA),
   diacylglycerol O-acyltransferase (BCG_3794c), and essential cell wall
   formation proteins (EmbA, Lgt and PonA1) (Supplementary Data [127]2 and
   Fig. [128]2c). Cholesterol catabolism provides a carbon source for
   energy production and restructuring of cell wall lipids, and has been
   extensively characterized in mycobacteria^[129]8. Cholesterol is broken
   down to acetyl-CoA for the TCA cycle, propionyl-CoA for the
   methylcitrate cycle, the vitamin B12-dependent methylmalonyl pathway or
   lipid synthesis (e.g., triacylglyceride and PDIM formation), and
   pyruvate for the generation of acetyl-CoA or to fuel gluconeogenesis.
   Upregulated LldD1 drives metabolic flux towards pyruvate, upregulated
   MutB facilitates the synthesis of propionate from TCA cycle
   intermediates, and upregulated GltA accelerates carbon assimilation
   into the TCA cycle by facilitating the condensation of acetyl-CoA and
   oxaloacetate to form citrate (the first reaction of the TCA cycle).
   These proteomic signals align with cholesterol and odd-chain fatty acid
   metabolism (Fig. [130]2b, c): MutB facilitates the generation of
   3-carbon intermediates in the form of methylmalonyl-CoA that can be
   used as building blocks for mycobacterial cell wall lipids, and the
   upregulation of FabG, FabG2, and BCG_3794c, which participate in TAG
   biosynthesis, agrees with the observation that mycobacteria increase
   lipid deposition during the early stages of NRP in
   macrophages^[131]8,[132]15,[133]30.

   The metabolic divergence in hypoxia and starvation is reinforced by
   analysis of fatty acid metabolism. While hypoxia causes accumulation of
   lipid droplets^[134]31, we observed starvation-induced depletion of
   intracellular triacylglycerol stores within just one day of starvation
   (Fig. [135]3a and Supplementary Fig. [136]5a) associated with increased
   α,β-esterase activity (Fig. [137]3b and Supplementary Fig. [138]5b–h),
   an established proxy for lipase activity in mycobacteria^[139]32, with
   LipY being the main lipase for the hydrolysis of
   triacylglycerols^[140]33,[141]34. Further, both transcriptomics and
   proteomics data show that hypoxia induces a general reduction in the
   levels of enzymes involved in fatty acid β-oxidation and synthesis of
   long-chain mycolic acids (Supplementary Data [142]1 and [143]2).
   Importantly, the hallmarks of β-oxidation observed during starvation
   uniformly reversed upon resuscitation (Fig. [144]3a, b and
   Supplementary Fig. [145]5a–h). Collectively, these results suggest that
   starved persisters actively reprogram their metabolism toward fatty
   acid β-oxidation, a finding that concurs with previous observations of
   in vitro stress and in vivo survival^[146]6,[147]9,[148]30,[149]35,
   while hypoxic mycobacteria upregulate pathways associated with both
   cholesterol and odd-chain fatty acid metabolism. Despite the same
   common endpoint of NRP, these distinct metabolic shifts have different
   consequences for the cell, including starvation-induced ketosis
   resulting from substantial increases in β-oxidation of fatty acids
   during starvation, which we explored next.

Fig. 3. Starvation induces shifts in lipid and ketone body metabolism.

   [150]Fig. 3
   [151]Open in a new tab

   a TAG content of BCG analyzed by densitometry of thin-layer
   chromatograms. N = 3; *P < 0.05; one-way ANOVA with Dunnett's test vs
   day 0. b Percentage of cells with elevated esterase activity (CFDA^hi)
   during starvation and resuscitation. N = 6; two-way ANOVA with
   Bonferroni post tests, ^#P < 0.05 vs log; *P < 0.05 vs resuscitation
   day 0. c Hierarchical clustering analysis of the metabolic phenotype of
   Log, S4, S10, S20, and R6 BCG on carbon sources that induced growth.
   Heatmap denotes signals from tetrazolium dye reduction, reflecting
   carbon utilization relative to the positive control. d PCA bi-plot of
   PLS-DA scores and loadings of metabolic phenotype datasets (n = 3 per
   condition) based on carbon sources with significant dye reduction.
   Statistical analysis: one-way ANOVA with Bonferroni post test. The
   PLS-DA model is based on correlation coefficients between PCA scores
   (condition) and loadings (metabolite utilization) whereby proximity of
   the carbon source to the condition characterizes the condition. These
   predictors were used to successfully differentiate independent Log and
   S30 samples. e Intracellular β-hydroxybutyrate in Log, S20, and R6
   cultures. N = 6; *P < 0.05; one-way ANOVA with Dunnett's test vs Log.
   (f) pH of BCG with nutrients or starved in PBS. Each symbol denotes the
   median pH of >50,000 cells; *P < 0.05; unpaired two-tailed t test with
   Welch’s correction. See Supplementary Data [152]4 for the source data
   used in these graphs.

Starvation induces a ketotic metabolic state in mycobacterial persisters

   The data to this point predict a starvation-induced metabolic shift to
   increased β-oxidation of fatty acids, which should lead to
   downregulation of glucose metabolism and a ketotic state caused by the
   buildup of ketone bodies such as AcAc and β-hydroxybutyrate (BHB). To
   test these predictions, we used phenotypic profiling and metabolomics
   to define the metabolic shifts that mediate mycobacterial persistence
   during the BCG starvation time course. Phenotypic profiling involved
   incubating starved BCG on an array of 71 different metabolites, of
   which 41 induced metabolic activity (tetrazolium dye reduction) and 32
   induced cell growth (recoverable CFU) (Supplementary Tables [153]2 and
   [154]3). Unsupervised hierarchical clustering of signals from
   tetrazolium reduction during culture on the 32 growth-enabling
   compounds clearly distinguished exponential growth in nutrient-rich
   media (Log and R6) from both early (S4) and late starvation (S10 and
   S20) (Fig. [155]3c). Given the observed metabolic distinctions between
   starved and exponentially growing BCG, we sought to discriminate carbon
   source preferences at each timepoint of the starvation experiment. Here