A systematic evaluation of Mycobacterium tuberculosis ...Dany JV Beste 3,O and Rigoberto Rios-Estepa...

A systematic evaluation of Mycobacterium tuberculosis Genome-Scale 1

Metabolic Networks 2

Víctor A López-Agudelo 1,2, Emma Laing 3, Tom A Mendum 3, Andres Baena 2,4, Luis F 3

Barrera 2,5 Dany JV Beste 3,O and Rigoberto Rios-Estepa 1,* 4

1 Grupo de Bioprocesos, Departamento de Ingeniería Química, Universidad de Antioquia UdeA, 5

Calle 70 No. 52-21, Medellín, Colombia 6

2 Grupo de Inmunología Celular e Inmunogenética (GICIG), Facultad de Medicina, Universidad 7

de Antioquia UdeA, Calle 70 No. 52-21, Medellín, Colombia 8

3 Department of Microbial Sciences, Faculty of Health and Medical Sciences, University of 9

Surrey, Guildford GU2 7XH, UK 10

4 Departamento de Microbiología y Parasitología, Facultad de Medicina, Universidad de 11

Antioquia UdeA, Calle 70 No. 52-21, Medellín, Colombia 12

5 Instituto de Investigaciones Médicas, Facultad de Medicina, Universidad de Antioquia UdeA, 13

Calle 70 No. 52-21, Medellín, Colombia 14

15

o Corresponding Author: [email protected] 16

* Corresponding Author: [email protected] 17

Abstract 18

The metabolism of the causative agent of TB, Mycobacterium tuberculosis (Mtb) has recently re-19

emerged as an attractive drug target. A powerful approach to study Mtb metabolism is to use a 20

systems biology framework, such as a Genome-Scale Metabolic Network (GSMN) that allows the 21

dynamic interactions of the many individual components of metabolism to be interrogated together. 22

Several GSMNs networks have been constructed for Mtb and used to study the complex relationship 23

between Mtb genotype and phenotype. However, their utility is hampered by the existence of 24

multiple models of varying properties and performances. Here we systematically evaluate eight 25

recently published metabolic models of Mtb-H37Rv to facilitate model choice. The best performing 26

.CC-BY 4.0 International licenseavailable under awas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

The copyright holder for this preprint (whichthis version posted November 10, 2019. ; https://doi.org/10.1101/837401doi: bioRxiv preprint

https://doi.org/10.1101/837401http://creativecommons.org/licenses/by/4.0/

2 of 39

models, sMtb2018 and iEK1011, were refined and improved for use in future studies by the TB 27

research community. 28

Introduction 29

Mycobacterium tuberculosis (Mtb) is the causative bacterium agent of the global tuberculosis (TB) 30

epidemic, which is now the biggest infectious disease killer worldwide, causing 1.6 million deaths in 31

2017 alone [1]. Mtb is an unusual bacterial pathogen, as it is able to cause both acute life threatening 32

disease and a clinically latent infections that can persist for the lifetime of the human host [2,3]. 33

Metabolic reprogramming in response to the host niche during both the acute and the chronic phase 34

of TB infections is a crucial determinant of virulence [4–7]. With the worldwide spread of multi- and 35

extensively-resistant strains of Mtb thwarting the control of this global emergency, new drugs against 36

Mtb are urgently needed and metabolism presents an attractive target [8,9]. 37

Genome-scale constraint-based modelling has proved to be a powerful method to probe the 38

metabolism of Mtb. The first Genome Scale Metabolic Networks (GSMNs) of Mtb were published 39

in 2007 by Beste (GSMN-TB) [10] and Jamshidi (iNJ661) [11] and have been used as a platform for 40

interrogating high throughput ‘omics’ data, by simulating bacterial growth, generating hypothesis and 41

drug discovery and are also the backbone for subsequent model iterations [12–20]. In 2014, Rienksma 42

and colleagues developed a genome-scale model of Mtb metabolism (sMtb) building upon three 43

previously published models [15]. More recently Rienksma and colleagues utilized RNA sequence 44

data to provide condition-specific biomass reactions and identify metabolic drug responses during 45

Mtb infection of human macrophages [21,22] thus providing a model more suited to exploring the 46

metabolism of Mtb during human infection. In 2018 the first consolidated genome-scale metabolic 47

network, iEK1011 was constructed using standardized nomenclature of metabolites and reactions 48

from the BiGG model database [23,24]. The utility of Mtb GSMNs will continue to increase, as more 49

biological data is published and existing models are updated [25]. 50

With so many well-annotated GSMN’s of Mtb available (Fig 1), a crucial first step in any genome 51

scale exploration of the metabolism of Mtb is the selection of an appropriate model. Here, we 52

systematically evaluate the performance of eight recently published GSMN of Mtb-H37Rv. In 53




3 of 39

addition to comparing the metrics of the models descriptively in terms of size, connectivity, number 54

of blocked reactions and gaps in the network, we also identify the thermodynamically infeasible, and 55

energy generating cycles that could significantly impact on the accuracy of flux simulations. Using 56

Flux Balance Analysis (FBA) we performed growth analysis and compared the models’ ability to 57

predict gene essentiality when grown on different carbon and nitrogen sources including growth on 58

cholesterol, which is thought to be the major carbon source for intracellular Mtb within its human 59

host macrophage. 60

This work provides an inventory of the available GSMN-TB and their utility in recapitulating aspects 61

of Mtb metabolism. In addition, we present updated versions of the highest performing models 62

iEK1011 and sMtb2018 (iEK1011_2.0 and sMtb2.0) which provide multi-scale simulation platforms 63

available for the TB research community and will therefore inform future studies on the metabolism 64

of this deadly pathogen. 65

66




4 of 39

Fig 1. The evolution of Genome Scale Metabolic models of Mtb. Models highlighted in 67

gray were analysed in this study. Numbers denote genes/intracellular reactions. Black 68

circles are indicative of merged Mtb models. 69

Results and Discussion 70

Descriptive evaluation of the models 71

Each of the GSMNs analysed in this study (Fig 1, S1 Appendix) combine knowledge from genome 72

annotations, literature and measured biochemical compositions of Mtb. The complex linkage between 73

genotype and phenotype is made by gene-protein-reaction (GPR) associations, implemented as 74

Boolean rules in order to connect gene functions to enzyme complexes, isozymes or promiscuous 75

enzymes, and then to biochemical reactions [26]. Using set theory, we computed the intersection 76

between all sets of the model’s genes (Fig 2, andS1 Table). In accordance with expectations, the 77

pairwise matrix (Fig 2) demonstrates that Mtb models constructed from the same ancestor (iNJ661 or 78

GSMN-TB), are more similar (Fig. 1, Fig. 2). By contrast the consolidated models iEK1011 and 79

sMtb2018 share gene similarities (>60%, 60%,

5 of 39

biosynthesis. Likewise, the iNJ661v_modified model (Fig 1) has poor annotation of genes involved 92

in lipid-metabolism (e.g. β-oxidation, cholesterol degradation, fatty acid biosynthesis, lipid 93

biosynthesis and mycolic acid biosynthesis). These models therefore have limitations for in silico 94

simulation of Mtb growing on these physiologically relevant lipid sources and for modelling growth 95

in vivo. 96

97

98

Fig 2. Pairwise matrix of shared genes among Mtb models. Values in black (genes in 99

common between the models), blue (genes specific to the model indicated), and red (genes 100

specific to the model indicated) represent the number and percentage of specific model 101

genes specified in the y- and x-axis. 102

Checking mass and charge balances of biochemical reactions 103

Currency metabolites like water, protons, ATP, and cofactors like NADH, NADPH, FADH2, 104

CoA, etc. are ubiquitous and essential for metabolism. The addition of these cofactor metabolites 105

in GSMNs and in the biomass reaction considerably improves phenotype predictions and is a 106

hallmark of good quality reconstructions [19, 30]. In order to check this we converted the 107

GSMNs into substance graphs (using a local script) where metabolites (nodes) are connected by 108




6 of 39

edges (undirected and unweighted) if they appear in the same reaction [31] and computed node 109

degrees (number of edges connected to the node) (Table 2, and S4 Table), 110

Table 2. Degree values for currency metabolites of Mtb GSMNs. PI: Phosphate, PPI: 111

diphosphate, ACP: acyl-carrier protein, MK: menaquinone. 112

Currency

Metabolite GSMN-TB1.1 iOSDD890 sMtb iCG760 iSM810 iNJ661v_mod iEK1011 sMtb2018

H 72 701 512 102 127 667 741 511

CO2 150 202 266 146 160 193 204 268

H2O 5 507 624 0 5 472 569 624

ATP 236 353 320 264 242 295 315 320

AMP 101 170 115 103 103 117 132 115

PI 189 293 266 213 196 268 293 267

PPI 175 235 180 177 180 168 196 181

COA 175 163 230 175 190 142 194 230

ACP 98 48 136 94 103 48 57 136

NADH 102 141 262 104 113 111 176 263

NADPH 133 151 192 132 146 147 150 192

FADH2 40 47 55 42 48 21 62 55

MK 37 20 11 37 37 20 18 20

O2 37 76 64 40 41 68 91 65

113

From this analysis the degree values indicated that GSMN-TB 1.1, iCG760 and iSM810 models 114

have the lowest participation of currency metabolites (Table 2). Water and protons were the most 115

underrepresented metabolites (low degree values), indicating that these models may not be 116

correctly balanced. It is important that biochemical reactions are charge and mass balanced. 117

Unbalanced reactions in GSMNs may allow proton or ATP production out of nothing [32]. In 118

order to test whether the Mtb GSMNs are mass and charge balanced, we used the COBRA 119

Toolbox function “checkMassChargeBalance” [33]. Unfortunately, we were not able to perform 120

this analysis for GSMN-TB1.1, iCG760 and iSM810 due to the lack of standard metabolite 121

formulas in these models (Table 3). iEK1011 has the lowest number of imbalanced reactions (4) 122

compared with sMtb2018 (8), sMtb, (12), iNJ661v_modified (13) and iOSDD890 (78). The 123

majority of the imbalanced reactions belonged to cell wall biosynthetic pathways, including 124




7 of 39

arabinogalactan, peptidoglycan, and mycolic acid biosynthesis (S5 Table) reflecting the 125

difficulties in rebuilding accurate metabolite formula for complex cell wall components. 126

Another potential error in GSMNs is the molecular weight of biomass, which should be defined as 127

1 g/mmol. Discrepancy in biomass weight can arise as a result of unbalanced reactions which impacts 128

on the reliability of flux predictions using Flux Balance Analysis (FBA). This affect can be amplified 129

when host-pathogen interactions are simulated by integrating host and pathogen metabolic models 130

[34]. Chang and colleagues (2017) developed a systematic algorithm for checking if the biomass 131

molecular weight from GSMNs deviates from 1 g/mmol,. Using this algorithm [34] we found 132

deviations of

8 of 39

Of those models that contained a pathway for cholesterol degradation (GSMN-TB1.1, iCG760, 152

iSM810, iEK1011, sMtb, and sMtb2018). The GSMN-TB1.1 cholesterol degrading pathway 153

contained a number of dead end metabolites making this model unsuitable for exploring the 154

metabolism of this important in vivo carbon source. 155

Table 3. Global features of the Mtb metabolic models analyzed in this study. DEM: Dead-End 156

Metabolite, Metabolites wo F: Metabolites without formula, Imbal. Reactions: Imbalanced Reactions, 157

URs: Unbounded Reactions, TICs: Thermodynamically Infeasible Cycles, diss. Flux: dissipation 158

flux, MW: Molecular Weight, UD: Undetermined. 159

GSMN-

TB 1.1 iOSDD890 sMtb iCG760 iSM810

iNJ661v_

mod iEK1011 sMtb2018

Reactions 876 1152 1311 965 938 1054 1228 1321

Intracellular Reactions 876 1055 1192 864 938 956 1118 1200

Metabolites 667 961 1047 754 724 840 998 1049

DEMs 25 162 33 84 53 90 110 34

Blocked Reactions 98 (11%) 290 (25%) 92 (7%) 89 (9%) 117 (12%) 153 (14%) 138 (11%) 92 (7%)

Metabolites wo F 667 0 2 754 724 0 4 2

Imbal. Reactions UD 78 12 UD UD 13 4 8

URs 27 (3%) 52 (5%) 70 (6%) 45 (5%) 33 (3.5%) 67 (7%) 20 (2%) 75 (6%)

TICs 8 17 16 8 8 23 4 17

ATP diss. Flux 0 0 0 0.33 0 0 0 0

GTP diss. Flux 0 0 0 0.33 0 0 0 0

CTP diss. Flux 0 0 0 0.33 0 0 0 0

UTP diss. Flux 0 0 0 0.33 0 0 0 0

Biomass MW UD 1.0070 1.0126 UD UD 1.0094 1.0937 1.0126

160

Thermodynamic and energetic properties 161

Integrating thermodynamics data into GSMNs is extremely useful in order to check the feasibility of 162

reactions and their directionality [43,44]. Although, Mtb GSMNs have been built from 163

thermodynamics information, current Mtb GSMNs have never been checked for infeasible internal 164

flux cycles (also named type III pathways), which are reactions that do not exchange metabolites with 165

the surroundings and therefore violate the second law of thermodynamics [43,45,46]. A tractable way 166

to identify reactions participating in these thermodynamically infeasible cycles (TICs) is by defining 167

the set of reactions required for an unbounded metabolic flux under finite or even zero substrate 168

uptake inputs. Using Flux Variability Analysis (FVA ) the Unbounded Reactions (URs) can be 169

identified as those reactions with fluxes at the upper and/or lower bound constraints. Thus we 170




9 of 39

identified the thermodynamically infeasible cycles (TIC) using a local script following methodologies 171

based on FVA and the analysis of the null space of the stoichiometric matrices (S7 Table) [44,47]. 172

Using this approach we show that models descended from GSMN-TB (GSMN-TB1.1, iCG760, and 173

iSM810) have a lower percentage of unbounded reactions as compared with iNJ661 ancestors 174

(iSM810, iNJ661v_modified). Interestedly, the sMtb2018 model has an increased number of 175

unbounded reactions as compared to the original sMtb (Table 3). 176

Fritzemeier and colleagues demonstrated that over 85% of genome-scale models that lack exhaustive 177

manual curation contain Energy Generating Cycles (EGCs) [48,49]. These cyclic net fluxes are 178

entirely independent of nutrient uptakes (exchange fluxes) and therefore have a substantial effect on 179

the predictions of constraint-based analyses, as they basically generate energy out of nothing. Using 180

FBA with zero nutrient uptake [49] but maximizing energy dissipation reactions for ATP, GTP, CTP 181

and UTP we show that iCG760 is (Table 3), the only Mtb genome-scale model that contains EGCs. 182

Gene essentiality metrics 183

An effective and commonly employed predictive matrix for GSMNs is the ability to reproduce high 184

throughput gene essentiality data [50]. Several high throughput transposon mutagenesis screens have 185

been performed for Mtb [51–57] in different in vitro conditions. To compare our models we used a 186

transposon insertion sequence dataset produced by Griffin et al [54] in which genes that were 187

differentially required for growth on cholesterol as compared with glycerol [54] were identified. 188

Cholesterol is an important intracellular source of carbon when Mtb is growing within its host and 189

cholesterol metabolism has been highlighted as a potential drug target [58]. Whilst Griffin and 190

colleagues identified differentially required genes, this data was not used to identify the genes 191

required for growth on cholesterol as an individual condition. We reanalyzed the Griffin’s transposon 192

sequencing data with the statistical Bayesian/Gumbel Method incorporated into the software 193

TRANSIT [59], to identify genes required for growth on glycerol, or growth on cholesterol (S8 and 194

S9 Tables). Only genes categorized as essential (ES) and non-essential (NE) were considered for this 195

analysis. 196




10 of 39

We evaluated the overall predictive power of all the Mtb GSMNs versus four high throughput gene 197

essentiality data [54,56,57] by computing the Area Under the Curve (AUC) of the Receiver Operating 198

Characteristic (ROC) (S1 Fig). The predictive power of the six GSMNs which contain the cholesterol 199

degradation pathway (GSMN-TB1.1, iCG760, iSM810,sMtb, sMtb2018, and iEK1011) showed that 200

for both cholesterol and glycerol minimal media the models derived from GSMN-TB [10] as a core 201

metabolic network have better predictive capacities than those using iNJ661 [11] (S10 and S11 202

Tables). However the recently curated model iEK1011 had the highest predictive capability overall. 203

The supremacy of iEK1011 also was confirmed by comparing the predictive power of the models 204

for Mtb grown in standard Middlebrook 7H9 OADC [56] data (S1C Fig and S12 Table). 205

We further used the essentiality dataset generated by Minato and colleagues who identified 206

conditionally essential genes on different in vitro medium including a rich medium called “MtbYM”, 207

which contains several carbon sources, nitrogen sources, amino acids, nucleotide bases, cofactors, 208

and other nutrients [57]. Although this media supported growth similarly to 7H9 medium, the overall 209

predictive power of the Mtb GSMNs in this medium (S1D Fig amd S13 Table) was 10% lower 210

compared to the other media (S1A-S1C Figs), probably because biomass objective functions of the 211

Mtb GSMNs ancestors were built and validated using growth on 7H9, 7H10, Youman’s, CAMR, and 212

Roisin’s minimal medium [10,11]. However these analyses were able to demonstrate the power of 213

ability of these models to accurately predict gene essentiality even under un-anticipated nutritional 214

conditions. 215

We identified the genes that all of the GSMN’s were unable to correctly assign essentiality (S14 216

Table, Fig 3). Using a fixed threshold value of 5% of the maximum wild-type growth rate (WTGR) 217

we identified in silico essential and non-essential genes and using experimental high throughput gene 218

essentiality data identified True Positives (TP), True Negatives (TN), False Positives (FP-a gene 219

which is essential for in silico growth but non-essential by Tn-seq) and False Negatives (FN-in silico 220

the gene is non-essential but the biological data predicts essentiality). FN genes include genes known 221

to have a major role in Mtb central carbon metabolism e.g., icl (Rv0467, isocitrate lyase), gltA 222

(Rv0896, Citrate synthase), glpD2 (Rv3302c, Glycerol-3-phosphate dehydrogenase), pyk (Rv1617, 223




11 of 39

Pyruvate Kinase), sucC and sucD (Rv0951, Rv0952, Succinyl-CoA ligase), among others (S14A 224

Table). Some of these genes e.g. icl and gltA are considered conditionally essential genes in the 225

Online GEne Essentiality (OGEE) database [60], because they are classified as NE genes in 7H10 226

medium but ES in minimal medium [53,55]. This may reflect the presence of alternative routes in 227

silico that are not feasible in vivo due to regulatory constraints. However, they may also reflect 228

inaccuracies in the predictive abilities of transposon mutagenesis studies. Some of the FP genes are 229

involved in mycolic acid biosynthesis (S14B Table), e.g., mmaA2 (Rv0644c, Cyclopropane mycolic 230

acid synthase), and mas (Rv2940c, mycocerosic acid synthase). These genes were inaccurately 231

classified as ES in silico but experimentally as NE. This is a direct reflection of our incomplete 232

knowledge of these mycolate producing pathways and the essentiality of the various mycolates 233

produced. 234




12 of 39

235

Fig 3. False Negative and False Positive predictions in the evaluated media. (a) Venn 236

diagram for predicted false negative (FN) genes; (b) Venn diagram for predicted false 237

positive (FP) genes. Genes in the gray box represents the intersection of all FN and FP genes 238

in the four media. Genes are classified as True-positives (TP) if model simulation predicts 239

no growth when essential genes are deleted, False-positives (FP) if model simulation 240

predicts no growth when not essential genes are deleted, True negatives (TN) if model 241

simulation predicts growth when not essential genes are deleted and False negatives (FN) if 242

model simulation predicts growth when essential genes are deleted. 243

244




13 of 39

Growth metrics 245

Mtb is able to metabolise several carbon and nitrogen sources both in vitro and when growing in the 246

host [61–64], and therefore we evaluated the growth metrics of Mtb GSMNs on 30 sole carbon and 247

17 sole nitrogen sources (Fig 4). The in silico results were compared with available in vitro 248

experimental data from Biolog Phenotype microarray and minimal media data [14,65]. Interestingly 249

the recent consolidated models, iEK1011 and sMtb, had the poorest performance of all the models in 250

predicting growth of Mtb in unique carbon and nitrogen sources (Fig 4A and 4B). A fundamental 251

issue with the Mtb models descended from iNJ661 is that they all require glycerol for growth as this 252

is present in the biomass formulation. Both iEK1011 and sMtb were unable to grow in silico on 253

cholesterol, acetate, oleate, palmitate and propionate on sole carbon sources. We hypothesise that this 254

is as a result of inaccuracies in the reactions associated with redox metabolism and oxidative 255

phosphorylation and specifically including menaquinone-dependent reactions such as fumarate 256

reductase and succinate dehydrogenase would correct this anomalie. To test this hypothesis we added 257

a menaquinone-dependent succinate dehydrogenase reaction into sMtb (Q[c] + SUCC[c] -> QH2[c] 258

+ FUM[c]). In support of our hypothesis this corrected the in silico growth phenotype of Mtb growing 259

on acetate, cholesterol, propionate and fatty acids (Fig 4A and S15 Table). Although iEK1011 also 260

contains a fumarate reductase reaction that is linked to menaquinone/demethylmenaquinone, it does 261

not contain a menoquinone-dependent succinate dehydrogenase (the reverse reaction). As was the 262

case for sMtb, the addition of a new menaquinone-dependent succinate dehydrogenase reaction 263

(mqn8[c] + succ[c] -> fum[c] + mql8[c]) to iEK1011 significantly improves its growth predictions 264

on sole carbon sources (S15I and S15J Tables). These simulations are supported by experimental data 265

demonstrating that fumarate reductase and succinate dehydrogenase are essential for the survival of 266

Mtb in glycolytic and non-glycolytic substrates [66,67]. Succinate dehydrogenase is a bifunctional 267

enzyme that is part of the TCA cycle and complex II of the electron transport chain, coupling the 268

oxidation of succinate to fumarate, with the corresponding reduction of membrane-localized quinone 269

electron carriers [67,68]. Mtb has multiple succinate dehydrogenases and fumarate reductases that are 270

essential for the survival of Mtb survival during hypoxia [66,67,69–71]. Succinate is central to much 271

of Mtb’s lipid metabolism: host derived cholesterol, uneven chain length fatty acids or methyl 272




14 of 39

branched amino acids all generate propionyl-CoA that can be channeled into the methylcitrate cycle 273

to produce succinate (S2 Fig), while acetyl-CoA produced by β-oxidation of host derived even-chain 274

fatty acids is metabolized through the glyoxylate shunt to also produce succinate (S2 Fig). Succinate 275

oxidation by succinate dehydrogenases is therefore a critical step, as the enzyme couples the TCA 276

cycle with electron transport chain and oxidative phosphorylation [69]. Having multiple succinate 277

dehydrogenases provides Mtb with the metabolic flexibility to survive within the human host where 278

the REDOX and nutritional environment are liable to vary. 279

Whilst carbon metabolism has been intensively studied in vitro and ex vivo, attention has only recently 280

been directed to nitrogen metabolism [72–75]. Similar to carbon consumption, iEK1011 and sMtb 281

were poor at predicting growth on sole nitrogen sources (Fig 4B, Mathews Correlation Coefficient 282

(MCC) = 0.54 and 0.48, respectively). However, like carbon the addition of the menaquinone linked 283

succinate dehydrogenase reaction into iEK1011 and sMtb significantly improves the predictive power 284

of these models on sole nitrogen sources (S16I and S16J Tables). Specifically, the correct growth 285

prediction were obtained for Mtb growing on branched chain amino acids (isoleucine and valine) and 286

proline (Fig 4B). The possible explaination for these correct prediction is that complete degradation 287

of these amino acids can converge on succinate via methyl citrate cycle (degradation of isoleucine 288

and valine) or the GABA shunt (degradation of proline) which couples TCA cycle with oxidative 289

phosphorylation by succinate dehydrogenase. 290

291




15 of 39

292

Figure 4. Predictive capacity of Mtb genome-scale models for the utilization of sole carbon 293

and nitrogen sources; (a) Growth predictions of Mtb genome-scale models using sole 294

carbon sources; (b) Growth predictions of Mtb genome-scale models by using sole nitrogen 295

sources. Model’s performance was evaluated by computation of the Mathews Correlation 296

Coefficient (MCC). Experimental growth data were obtained from [14,65]. The 1 and 0 297




16 of 39

values represent growth and not growth in a specific substrate, respectively. Carbon or 298

Nitrogen substrates are classified as TP if model simulation predicts growth while growth 299

is observed experimentally, FP if model simulation predicts growth while no growth is 300

observed experimentally, TN if model simulation predicts no growth while no growth is 301

observed experimentally and FN if model simulation predicts no growth while growth is 302

observed experimentally. 303

Refining Mtb GSMNs 304

Overall iEK1011 and sMtb2018 were the best overall GSMN’s in terms of genetic background, 305

network topology, number of blocked reactions, mass and charge balance reactions and gene 306

essentiality predictions (Fig 2, Table 2, Table 3, and S1 Fig) and therefore we selected these models 307

to refine further. iEK1011 has the advantage of containing standardized BiGG nomenclature of 308

metabolites and therefore can easily be integrated into the human GSMN Recon3D [76] to simulate 309

intracellular growth, while sMtb2018 has the utility that this model supports in silico growth in a 310

wider variety of different nutritional conditions. Our analysis also highlighted some fundemental 311

issues with these models which we addressed in order to improve the performance of these exempler 312

GSMN’s. 313

As already discussed the new sMtb2018 includes an updated respiratory chain including menaquinone 314

and menaquinol as electron carriers in all respiratory chain reactions and selected ubiquinone-315

dependent reactions in sMtb were changed to menaquinone-dependent. Six new menaquinone-316

dependent reactions were included in the sMtb model e.g., succinate dehydrogenase, and cytochrome 317

bc1 menaquinone-dependent, fumarate reductase, and malate dehydrogenase [22]. This improved the 318

predictive growth metric of sMtb and importantly allowed in silico growth on cholesterol (see 319

sMtb2018, Fig 4A). Similarly, we added to iEK1011 a menaquinone-dependent succinate 320

dehydrogenase to improve the performance of this model when growing in fatty acids and cholesterol. 321

Further improvements were also made to cholesterol metabolism by updating both models to include 322

the biochemical degradation of the C and D rings of cholesterol which was not known when these 323

models were reconstructed (S17A and S17B Tables) [77]. 324




17 of 39

Although the functionality of the vitamin B12 biosynthetic pathway remains undefined, recent 325

elevated non-synonymous nucleotide substitution rates found on cobalamin-dependent enzymes of 326

3798 Mtb strains suggest positive selection on these genes and a possible functional biosynthetic 327

pathway [41]. Therefore, we facilitated co-factor metabolism in the models by unblocking B12 328

synthesis and including the co-factors biotin and pyridoxal-5-phosphate to enhance the phenotype 329

prediction of sMtb2018 and iEK1011 as recommended by Xavier and colleagues [19]. As iEK1011 330

contains glycerol within its biomass formulation, which prevent in silico growth in glycerol-replete 331

media, we incorporated the iNJ661v_mod biomass objective function into this model (S17B Table). 332

We also added additional genes into sMtb2018 based on iOSDD890 genes that were classified as true 333

negatives and true positives by the gene essentiality analysis (S18 Table). Fifty-one reactions that 334

belong to pathways such as glycolysis, gluconeogenesis, TCA cycle, amino acid metabolism and 335

mycolic acid pathway were improved in terms of GPR annotation. All updates for the new models 336

sMtb2.0 and iEK1011_2.0 are summarized in S17 Table. 337

The sMtb2.0 and iEK1011_2.0 models were curated to remove possible reactions that participate in 338

TICs respectively (methods section). (S19 Table). The stoichiometric matrix associated with these 339

reactions was built and the null space basis vector was computed (S20 Table) [44,47] identifying 340

twelve TICs within sMtb2.0; (Fig 5A , S2 Appendix) and seven TICs in iEK1011_2.0 (Fig 5B, S3 341

Appendix) ; The major TIC of sMtb2.0 were within the reactions required for folate metabolism 342

catalysed by thymidylate synthase (thyA and thyX) enzymes and dihydrofolate reductase (DFRA) 343

enzymes, which is an essential step for de novo glycine and purine biosynthesis, and for the 344

conversion of deoxyuridine monophosphate (dUMP) to deoxytimidine monophosphate (dTMP). 345

(Fig 5A). As recommended by Pereira and colleagues [30], we modified the model to use 346

NADPH/NADP+ by retaining the DFRA2 and DFRA4 reactions and eliminating the NADH/NAD+-347

dependent reactions, DFRA1 and DFRA3. Our thermodynamic calculations suggested that THYA 348

and THYX (∆𝑟𝐺𝑚𝑖𝑛=-123 kJ/mol, ∆𝑟𝐺𝑚𝑎𝑥=- 9.3 kJ/mol and ∆𝑟𝐺𝑚𝑖𝑛=-160 kJ/mol, ∆𝑟𝐺𝑚𝑎𝑥=- 33 349

kJ/mol, respectively) are irreversible in the forward direction (S21 Table). 350




18 of 39

Our analysis showed that two-ubiquinone oxidoreductases (QRr, NADH2r) and a transhydrogenase 351

reaction (NADTRHD) were thermodynamically unstable within the iEK1011_2 as both QRr and 352

NADH2r were reversible. The Gibbs free energy computations suggest these reactions should be 353

unidirectional in the direction of ubiquinol production ( ∆𝑟𝐺𝑚𝑖𝑛 = −134.9𝑘𝐽

𝑚𝑜𝑙, ∆𝑟𝐺𝑚𝑎𝑥 =354

−20.7 𝑘𝐽/𝑚𝑜𝑙 and ∆𝑟𝐺𝑚𝑖𝑛 = −131.7𝑘𝐽

𝑚𝑜𝑙, ∆𝑟𝐺𝑚𝑎𝑥 = −20.1 𝑘𝐽/𝑚𝑜𝑙, respectively) and therefore 355

we changed the model accordingly to eliminate this TIC. 356

357

Figure 5. Thermodynamic Infeasible Cycles found in sMtb2.0 and iEK1011_2.0 with 358

proposed modifications; (a) TIC at folate metabolism in sMtb2.0; (b) TIC affecting 359

ubiquinone oxidoreductases in iEK1011_2.0. 360

The sMtb2.0 encompasses 1321 reactions, 1054 metabolites and 989 genes, while iEK1011_2.0 361

comprises 1237 reactions, 977 metabolites and 1013 genes. 362




19 of 39

Table 4. Genome-scale model features for sMtb2.0 and iEK1011_2.0. 363

Metrics sMtb2.0 iEK1011_2.0

Genes 989 1013

Reactions 1321 1237

Intracellular Reactions 1198 1119

Exchange Reactions 122 119

Metabolites 1054 977

% of Unbounded Reactions (49) 4% 20 (1.8 %)

MCC Cholesterol minimal medium 0.55 0.63

Evaluated Genes 834 866

True Positive Genes 209 210

True Negative Genes 449 503

False Positive Genes 58 21

False Negative Genes 118 132

MCC Glycerol minimal medium 0.54 0.62






MCC 7H9 medium 0.56 0.69






MCC MtbYM medium 0.43 0.52






MCC Unique Carbon Source 0.58 0.67

Evaluated Metabolites 30 30

True Positive Metabolites 20 20

True Negative Metabolites 5 6

False Positive Metabolites 4 3

False Negative Metabolites 1 1

MCC Unique Nitrogen Source 0.75 0.75

Evaluated Metabolites 17 17

True Positive Metabolites 11 11

True Negative Metabolites 4 4

False Positive Metabolites 2 2

False Negative Metabolites 0 0

364

Similar to the above sections, the predictive capability of these modified models was evaluated by 365

simulating gene essentiality predictions using experimental data. In this case, we use the 5% of the 366

specific growth rate as essentiality threshold; if a specific growth rate of no more than 5% of the wild-367

type is obtained, the given gene was considered as essential, otherwise was considered non-essential 368

(S22 Table). iEK1011_2.0 has the highest predictive performance of gene essentiality in the four 369

media conditions tested (glycerol and cholesterol minimal medium, Middlebrook 7H9, and YM 370




20 of 39

medium) compared with all the Mtb GSMNs evaluated, including sMtb2.0 (8% higher in Cholesterol 371

and Glycerol minimal medium, and 13 % higher in 7H9 medium, Table 4). 372

The ability of these models to predict growth sole carbon and nitrogen sources is also better in the 373

updated GSMN’s as compared with their predecessors. These results show that iEK1011_2.0 and 374

sMtb2.0 are currently the most suitable models for studying host-pathogen interactions, as they are 375

able to correctly predict growth on several important carbon sources that Mtb uses intracellularly [62]. 376

By systematically evaluating eight of the recent Mtb GSMNs, we have identified a number of areas 377

in which model veracity can be improved. Mtb models descended from GSMN-TB (GSMN-TB1.1, 378

iSM810 and iCG760) contain many reactions that are unbalanced by charge and mass, often because 379

protons and water have not been accounted for. We observed that all of the Mtb models have dead-380

end metabolites particularly in cofactor metabolism and related pathways. The best performing 381

models were identified as sMtb2018 and iEK1011 by virtue of higher gene coverage, fewer blocked 382

reactions, better mass and charge balances, and overall best predictive power of gene essentiality data. 383

Whilst our analysis identified that iEK1011 and sMtb2018 were the overall best performing GSMN’s 384

we also highlighted areas for improvements. For example the iEK1011 model had poor predictive 385

power for growth on sole carbon and nitrogen sources and unlike Mtb cannot grow when cholesterol 386

and fatty acids are provided as sole carbon sources. We thus updated these two models by the addition 387

of new reactions, gap filling of cofactor metabolism, and the identification and curation of TICs, to 388

give sMtb2.0 (derived for sMtb2018) and iEK1011_2.0 (derived from iEK1011). 389

The improved GSMN’s, sMtb2.0 and iEK1011_2.0 are now available (S1 File) to simulate and 390

predict the metabolic adaptation of Mtb (by the integration of OMICS data) in a plethora of in vitro 391

and in vivo intracellular conditions. We encourage researchers to continue to curate these models as 392

new data and methods become available. Improved GSMN’s including macrophage-Mtb models 393

provide a platform for more reliable simulations and ultimately a better understanding of the 394

underlying biology of Mtb. 395

Methods 396




21 of 39

All simulations were conducted on a laptop running Windows 10 (Microsoft) using MATLAB 2016a 397

(MathWorks Corporation, Natick, Massachusetts, USA), COBRA Toolbox version 3.0 [33] and 398

Gurobi Optimizer version 7.5.2 (Gurobi Optimization, Inc., Houston, Texas, USA). All code written 399

for this study is available in supplementary information (S2-S5 Files). Genome-scale models of 400

Mtb 401

Models were obtained from supplementary information of published papers and modified as 402

follows: 403

GSMN-TB1.1 – from [14] supplementary info. 404

iOSDD890 – from [16] supplementary info. 405

sMtb – from [15] supplementary info. Modification included were the addition of exchange 406

reactions to allow constraints by growth medium components. 407

iCG760 – from [17] supplementary info. 408

iSM810 – from [18] supplementary info. 409

iNJ661v_modified – from [19] supplementary info. 410

sMtb2018 – from [22] supplementary info. 411

iEK1011 – from [24] supplementary info. 412

Network connectivity evaluation 413

GSMNs of Mtb were transformed to substrate networks by local scripts after eliminate biomass 414

reaction. Node-specific topology metrics were carried out using the plugin Network Analyzer [78] in 415

Cytoscape 3.4 [79]. Two main topological parameters were evaluated for each model: 1) the node 416

degree of each metaboliteand 2) the clustering coefficient (S4 Table). 417




22 of 39

MC3 Consistency Checker algorithm [38] was used to identify Single Connected and Dead End 418

metabolites, and zero-flux reactions in each metabolic network model of Mtb. This algorithm uses a 419

stoichiometric-based identification of metabolites connected only once in each metabolic network 420

and utilize Flux-Variability-Analysis (FVA) for identifying reactions that cannot carry flux [80]. 421

Biomass Molecular Weight Check 422

Testing the biomass Molecular Weight consistency was done by running the script of Chan and 423

colleagues [34] . A biomass reaction is not standardized when the Molecular Weight of the biomass 424

formula is not equal to 1 g/mmol. However, the accuracy of the results relies on the correct chemical 425

formulae of metabolites in the tested GSMNs. 426

Identification of Unbounded Reactions (URs) 427

A straightforward way to identify all reactions that participate in one or more TICs is by performing 428

flux variability analysis (FVA). All the Infeasible loops are evidenced as a set of reactions able to 429

carry an unbounded metabolic flux under finite or even zero substrate uptake inputs. The URs are 430

those reactions that by applying FVA [80], their fluxes will hit the values defined by the upper and/or 431

lower bounds constraints [47]. Therefore, we performed FVA with the eight Mtb GSMNs with all the 432

uptake media constraints defined by 1.0 mmol/gDW/h. 433

Identification of the core set of TICs 434

Schellenberger and colleagues [81] used a methodology for identifying the core set of TICs, which 435

form the basis of all such possible cycles. This core can be obtained by the computation of the null 436

space basis of the stoichiometric matrix (all possible thermodynamically infeasible cycles form the 437

null space of the stoichiometric matrix). Consequently, the set containing all the reactions that we 438

previously identified participate in TICs was used to build a stoichiometric matrix. Therefore, the null 439

space basis of this set was computed and the different cycles composed by two and more reactions 440

were identified by a local script. 441

Checking the existence of Energy Generating Cycles (EGCs). 442




23 of 39

Energy generating cycles (EGCs) exist in metabolic networks and can charge energy metabolites like 443

ATP, GTP, CDP, and UTP without any input of nutrients; therefore, their elimination is essential for 444

correcting energy metabolism [49,82] . Fritzemeier and colleagues developed a methodology for 445

identifying if genome-scale models contain EGCs [49]. We applied it in two steps : 1) Addition of 446

Energy dissipation reactions (EDR) for ATP, GTP, CTP and UTP in the form : H2O[c] + XTP[c] -> 447

H[c] + XDP[c] + Pi[c] and 2) maximization of each EDR flux 𝒗𝒅 while no substrate uptake is 448

allowed into the model as follows: 449

𝐦𝐚𝐱 𝒗𝒅 (𝟏) 450

𝒔. 𝒕 𝑺𝒗 = 𝟎 (𝟐) 451

∀𝒊 ∉ 𝑬: 𝒗𝒊𝑳𝑩 ≤ 𝒗𝒊 ≤ 𝒗𝒊

𝑼𝑩 (𝟑) 452

∀𝒊 ∈ 𝑬: 𝒗𝒊 = 𝟎 (𝟒) 453

Here, 𝑺 is the stoichiometric matrix, 𝒗 the vector of fluxes, d the index of EDRs, 𝒗𝒊𝑳𝑩 and 454

𝒗𝒊𝑼𝑩 the vector of lower and upper bounds, respectively, and 𝑬 is the set of indices of all 455

exchange reactions of the model. If the optimal value of 𝒗𝒅 for this optimization is 𝒗𝒅 > 𝟎, 456

there exist in the genome-scale model at least one EGC that is able to generate energy 457

metabolites like ATP, GTP, CTP, or UTP. 458

Curation of TICs 459

Two types of modifications were performed on the curated Mtb genome-scale metabolic network in 460

order to eliminate TICs [47]. 461

i) TICs formed by linearly dependent reversible reactions: Usually, these arise when there 462

are two reactions (NAD+- and NADP+-dependent) with the same catalytic activity. In this 463

instance, we forced the use of NADPH/NADP+ in anabolic reactions and NADH/NAD+ 464

for catabolic reactions, as recommended by Pereira and colleague [30]. If two irreversible 465




24 of 39

reactions that catalyze the forward and backward direction exist, both reactions (and GPR 466

rules) are lumped together in just one reversible reaction. 467

ii) TICs formed by erroneous directionality assignments: we restricted the reaction 468

directionality based on Gibbs free energy change (NExT algorithm) [83,84] as long as 469

gene essentiality predictions are not compromised 470

The later modification was based on the utilization of the NExT (network-embedded thermodynamic 471

analysis) algorithm [83,84]. This algorithm allows identify new irreversible reactions by calculating 472

the thermodynamically feasible range of Gibbs energy of reactions and metabolite concentrations. 473

NExT was implemented for those reactions participating in TICs under Matlab [84] with 474

physiological conditions adapted for Mtb (Table 5). Standard Gibbs energy of formation (∆𝒇𝑮𝒊) (in 475

kJ/mol), number of hydrogen atoms, and charge of all metabolites involved in TICs were obtained 476

from the Biochemical Thermodynamic Calculator, eQuilibrator [85,86]. 477

Table 5. Biophysical properties and concentration ranges for intracellular Mtb. 478

Properties Values Reference

Redox potential, cytosol -275 mV Bhaskar et al., 2014

pH, intracellular 5.7 Zhang et al., 2003

pH, extracellular (activated

macrophage) 4.5

Rohde et al., 2007; Vandal et

al., 2009

Ionic strength 0.15 M Kümmel et al., 2006; Martínez

et al., 2014

Oxygen Concentration (mM) 0.0001 - 0.1 Martínez et al., 2014

[CO2],[Pi] (mM) 1 - 100 Kümmel et al., 2006;

Haraldsdóttir et al., 2012

Other metabolites (mM) 0.0001 – 0.1 Martínez et al., 2014

NADH/NAD 0.0001 – 0.1 Martínez et al., 2014

NADPH/NADP 0.0001 – 0.1 Martínez et al., 2014




25 of 39

If a reaction specified to be reversible in the set of TICs and it has its maximum ∆𝒓𝑮 calculated to 479

be negative, the reaction is considered to occur in the forward direction. In contrast, if the minimum 480

∆𝒓𝑮 is positive, the reaction is considered to occur in the reverse direction. No direction can be 481

inferred when the minimum ∆𝒓𝑮 is negative and the maximum is positive. Changes in directionality 482

of reactions were done strictly when gene essentiality predictions in the curated genome-scale model 483

were not compromised. 484

Gene Essentiality Analysis. 485

To identify essential genes of Mtb grown on individual conditions (cholesterol minimal medium and 486

glycerol minimal medium) from Griffin’s paper [54], we use the Bayesian/Gumbel method of 487

TRANSIT, version 2.02 [59]. The Bayesian/Gumbel method determines posterior probability of the 488

essentiality of each gene (zbar). When zbar value is 1, or close to 1, the gene is considered essential 489

(ES), if zbar is 0, or close to 0, the gene is considered non-essential (NE), uncertain (U) genes are 490

those with zbar values between 0 and 1, and for too small (S) genes zbar is -1. After loading the TA 491

count files (replicates for cholesterol and glycerol) and the gene annotation file into TRANSIT, and 492

running the Gumbel method with default parameters, we obtained an output file with essentiality 493

results (Table S9 and S10). Uncertain (U) and too small (S) genes were not taken into account for the 494

in silico essentiality analysis. Minato and colleagues used the same statistical method for classify 495

essential genes [57]. Conversely, DeJesus and colleagues [56] used a Hidden Markov Model based 496

statistical method for classifying genes into four essentiality states: essential (ES), growth defect 497

(GD), nonessential (NE), and growth advantage (GA). In order to evaluate the performance of the 498

Mtb GSMNs to predict gene essentiality data, we use only binary classifiers, therefore we reclassify 499

these genes just in two groups as follows: NE genes included NE and GA genes, and ES genes 500

included GD and ES genes. 501

For the in silico gene essentiality analysis, we set the simulation conditions (asparagine, phosphate, 502

sodium, ammonium, citrate, sulfate, zinc, calcium, chloride, Fe3+, Fe2+, and glycerol or cholesterol) 503

according to Griffin minimal medium [54], 7H9 OADC medium, and “MtbYM” medium and a FBA-504

based gene essentiality analysis was performed in the eight Mtb models using the “single gene 505




26 of 39

deletion” function of Cobra Toolbox. Default maximization of biomass objective function was used 506

to predict growth in all models. If a specific growth rate of no more than 5% of the wild-type is 507

obtained, the given gene was considered as essential (in silico), otherwise was considered non-508

essential. 509

Percentage of in silico gene essentiality predictions were categorized as: true-positive, false-positive, 510

true-negative, and false-negative when the in silico data were compared with experimental 511

essentiality data. 512

TP (true-positive): model simulation predicts no growth when essential genes are deleted. 513

FP (false-positive): model simulation predicts no growth when not essential genes are deleted. 514

TN (true-negative): model simulation predicts growth when not essential genes are deleted. 515

FN (false-negative): model simulation predicts growth when essential genes are deleted. 516

For evaluate the performance of the Mtb GSMNs we used sensitivity, specificity, accuracy, and 517

Mathews correlation coefficient (MCC) metrics: 518

𝒔𝒆𝒏𝒔𝒊𝒕𝒊𝒗𝒊𝒕𝒚 =𝑻𝑷

𝑻𝑷 + 𝑭𝑵 (𝟓) 519

520

𝒔𝒑𝒆𝒄𝒊𝒇𝒊𝒄𝒊𝒕𝒚 =𝑻𝑵

𝑻𝑵 + 𝑭𝑷 (𝟔) 521

522

𝑨𝒄𝒄𝒖𝒓𝒂𝒄𝒚 =(𝑻𝑷 + 𝑻𝑵)

(𝑻𝑷 + 𝑭𝑷 + 𝑻𝑵 + 𝑭𝑵) (𝟕) 523

𝑴𝑪𝑪 =(𝑻𝑷 ∗ 𝑻𝑵) − (𝑭𝑷 ∗ 𝑭𝑵)

√(𝑻𝑷 + 𝑭𝑷)(𝑻𝑷 + 𝑭𝑵)(𝑻𝑵 + 𝑭𝑷)(𝑻𝑵 + 𝑭𝑵) (𝟖) 524

525




27 of 39

Utilization of carbon and nitrogen sources 526

The methodology for modeling the effect of different carbon sources and nitrogen sources on Mtb 527

growth was adapted from Lofthouse and colleagues [14]. The biology Phenotype MicroArray 528

experiments classification used were obtained from Lofthouse and colleagues, 2013. They classified 529

growth and no-growth in different carbon and nitrogen sources from the original Biolog data of Khatri 530

and colleagues and the Roisin’s minimal media [14,65]. In addition, we used Roisin’s minimal media 531

data that also were obtained by Lofthouse and colleagues. 532

To model the carbon source experiment, we simulated the media as a modified form of Roisin’s 533

minimal media containing unlimited quantities of ammonia, phosphate, iron, sulfate, carbon dioxide 534

and a Biolog carbon source influx of 1 mmol/gDW/h. Similarly, the nitrogen source experiment was 535

simulated using a modified form of Roisin’s media, where ammonia was replaced with 1 mmol/g 536

DW/h of the Biolog nitrogen source and pyruvate was used as a carbon source (influx at 1 mmol/g 537

DWt/h). 538

For comparing the utilization of carbon and nitrogen sources in all Mtb models with experimental 539

data, we used Mathew’s correlation coefficient metrics (equation 8). 540

In silico growth predictions in carbon and nitrogen sources also were categorized as: true-positive, 541

false-positive, true-negative, and false-negative. 542

TP (true-positive): model simulation predicts growth while growth is observed experimentally (or 543

respiration rate is observed in biolog phenotype microarrays) in presence of a unique carbon or 544

nitrogen source. 545

FP (false-positive): model simulation predicts growth while no growth is observed experimentally in 546

presence of a unique carbon or nitrogen source. 547

TN (true-negative): model simulation predicts no growth while no growth is observed experimentally 548

in presence of a unique carbon or nitrogen source. 549




28 of 39

FN (false-negative): model simulation predicts no growth while growth is observed experimentally 550

in presence of a unique carbon or nitrogen source. 551

Supporting Information 552

S1 Appendix. Overview of the eight constraint-based models of Mtb used in this study. A 553

document providing additional detail about the eight Mtb GSMNs used in this study. (DOCX) 554

S2 Appendix. Curation of additional Thermodynamic Infeasible Cycles in sMtb2.0. A document 555

providing detailed description of the curation of TICs in sMtb2.0. (DOCX) 556

S3 Appendix. Curation of additional Thermodynamic Infeasible Cycles in iEK1011_2.0. A 557

document providing detailed description of the curation of TICs in iEK1011_2.0. (DOCX) 558

559

S1 Fig. ROC curves of gene essentiality predictions for Mtb GSMNs. a Receiver operating 560

characteristic curve for the gene essentiality predictions in cholesterol minimal medium, b 561

Receiver operating characteristic curve for the gene essentiality predictions in glycerol minimal 562

medium, c Receiver operating characteristic curve for gene essentiality predictions in 7H9 563

Middlebrook OADC medium, d Receiver operating characteristic curve for gene essentiality 564

predictions in MtbYM medium. (TIF) 565

S2 Fig. central role of succinate dehydrogenase in oxidation of odd-chain substrates. (TIF) 566

S1 Table. Set analysis of genes from all the eight Mtb GSMNs. A table with pairwise comparisons, 567

unions, and intersections of genes annotated in the eight GSMNs of Mtb. (XLSX) 568

S2 Table. Metabolic pathways associated to all the intersected gene sets of Mtb GSMNs. (XLSX) 569

S3 Table. Genes and metabolic pathways shared between GSMNs of Mtb. (XLSX) 570

S4 Table. Network topological properties of Mtb GSMNs. (XLSX) 571




29 of 39

S5 Table. List of unbalanced reactions and Metabolites without formulas in the Mtb GSMNs. 572

(XLSX) 573

S6 Table. List of blocked reactions and dead-end metabolites. (XLSX) 574

S7 Table. List of Thermodynamically infeasible cycles identified for the eight Mtb GSMNs. 575

(XLSX). 576

S8 Table. Transposon sequencing analysis of Mtb genes required for growing on minimal 577

medium plus cholesterol. A list of genes of Mtb classified as essential, non-essential, and uncertain 578

for growing in cholesterol minimal medium, the essentiality analysis was obtained by applying the 579

Bayesian/Gumbel Method incorporated into the software TRANSIT. (XLSX) 580

S9 Table. Transposon sequencing analysis of Mtb genes required for growing on minimal 581

medium plus glycerol. A list of genes of Mtb classified as essential, non-essential, and uncertain for 582

growing in glycerol minimal medium, the essentiality analysis was obtained by applying the 583

Bayesian/Gumbel Method incorporated into the software TRANSIT. (XLSX) 584

S10 Table. Predictive power of Mtb GSMNs for classify essential and non-essential genes on 585

cholesterol minimal medium. Genes whose in silico knockouts give growth rate values lesser than 586

5% of the maximum growth rate are classified as essential, otherwise are classified as non-essential. 587

(XLSX) 588


glycerol minimal medium. Genes whose in silico knockouts give growth rate values lesser than 5% 590

of the maximum growth rate are classified as essential, otherwise are classified as non-essential. 591

(XLSX) 592

S12 Table. Predictive power of Mtb GSMNs for classify essential and non essential genes on 593

7H9 OADC medium. Genes whose in silico knockouts give growth rate values lesser than 5% of the 594

maximum growth rate are classified as essential, otherwise are classified as non-essential. (XLSX) 595




30 of 39


cholesterol minimal medium. Genes whose in silico knockouts give growth rate values lesser than 597

5% of the maximum growth rate are classified as essential, otherwise are classified as non-essential. 598

(XLSX) 599

S14 Table. List of common False Positive and False Negative Genes of all the Mtb GSMNs 600

during gene essentiality predictions. (XLSX) 601

S15 Table. Predictive power of Mtb GSMNs growing on sole carbon sources. (XLSX) 602

S16 Table. Predictive power of Mtb GSMNs growing on sole nitrogen sources. (XLSX) 603

S17 Table. New added reactions into the sMtb and iEK1011 models. (XLSX) 604

S18 Table. List of reactions with modified gene annotation in sMtb2.0. (XLSX) 605

S19 Table. List of reactions that participate in Thermodynamically Infeasible Cycles. (XLSX) 606

S20 Table. Null Space of the stoichiometric matrix formed by unbounded reactions of sMtb2.0 607

and iEK1011_2.0. (XLSX) 608

S21 Table. Gibbs free energy change of Unbounded Reactions of updated Mtb GSMNs. (XLSX) 609

S22 Table. Gene Essentiality Predictions for iEK1011_2.0 and sMtb2.0 on four Mtb media. 610

(XLSX) 611

S23 Table. Growth Phenotypes of iEK1011_2.0 and sMtb2.0 on unique carbon and nitrogen 612

sources. (XLSX) 613

S1 File. Updated Mtb models sMtb2.0 and iEK1011 in .mat and ,xlsx format. (ZIP) 614

S2 File. Matlab script for checking the charge and mass balance of Mtb GSMNs. (ZIP) 615

S3 File. Matlab scripts for identifying unbounded reactions and thermodynamically infeasible 616

cycles in Mtb GSMNs. (ZIP) 617




31 of 39

S4 File. Matlab scripts for identifying Mtb GSMNs with Energy Generating Cycles. (ZIP) 618

S5 File. Matlab scripts for gene essentiality analysis of Mtb GSMNs on four media conditions. 619

(ZIP) 620

Author Contributions 621

Conceptualization: Víctor A. López-Agudelo, Dany JV Beste., Rigoberto Rios-Estepa. 622

Data curation: Víctor A. López-Agudelo 623

Formal analysis: Víctor A. López-Agudelo, Emma Laing, Tom Mendum, Dany JV. Beste, Rigoberto 624

Rios-Estepa. 625

Funding acquisition: Dany JV. Beste, Rigoberto Rios-Estepa. 626

Investigation: Víctor A. López-Agudelo, Emma Laing, Tom Mendum, Andres Baena, Luis F. 627

Barrera, Dany JV. Beste, Rigoberto Rios-Estepa. 628

Methodology: Víctor A. López-Agudelo, Dany JV. Beste, Rigoberto Rios-Estepa. 629

Project administration: Rigoberto Rios-Estepa, Dany JV. Beste. 630

Resources: Víctor A. López-Agudelo, Emma Laing, Tom Mendum, Andres Baena, Luis F. Barrera, 631

Dany JV. Beste, Rigoberto Rios-Estepa. 632

Software: Víctor A. López-Agudelo, Tom Mendum, Dany JV. Beste. 633

Supervision: Dany JV. Beste, Rigoberto Rios-Estepa. 634

Validation: Tom Mendum, Dany JV. Beste, Rigoberto Rios-Estepa. 635

Visualization: Víctor A. López-Agudelo 636

Writting – original draft: Víctor A. López-Agudelo, Andres Baena, Luis F. Barrera, Dany JV. Beste, 637

Rigoberto Rios-Estepa 638

Writting – review & editing: Víctor A. López-Agudelo, Tom Mendum, Dany JV. Beste, Rigoberto 639

Rios-Estepa. 640

641




32 of 39

642

Funding 643

This work was supported by grants from COLCIENCIAS (1115-5693-3520 and the Medical Research 644

Council (MR/K01224X/1). 645

646

Acknowledgments 647

Víctor A. López-Agudelo acknowledges the economic support provided by the Administrative 648

Department of Science, Technology and Innovation - COLCIENCIAS, Colombia, during his Ph.D. 649

studies (National Ph.D. scholarship Conv. 727-2015). 650

References 651

1. World Health Organisation. Global Health TB Report. 2018. 652

2. Deng M, Lv X-D, Fang Z-X, Xie X-S, Chen W-Y. The blood transcriptional signature for 653

active and latent tuberculosis. Infect Drug Resist. 2019;12: 321. 654

3. MacLean E, Broger T, Yerlikaya S, Fernandez-Carballo BL, Pai M, Denkinger CM. A 655

systematic review of biomarkers to detect active tuberculosis. Nat Microbiol. 2019;4: 748. 656

4. Cumming BM, Addicott KW, Adamson JH, Steyn AJC. Mycobacterium tuberculosis induces 657

decelerated bioenergetic metabolism in human macrophages. Elife. 2018;7: e39169. 658

5. Shi L, Jiang Q, Bushkin Y, Subbian S, Tyagi S. Biphasic Dynamics of Macrophage 659

Immunometabolism during Mycobacterium tuberculosis Infection. MBio. 2019;10: e02550-660

18. 661

6. Kalia NP, Shi Lee B, Ab Rahman NB, Moraski GC, Miller MJ, Pethe K. Carbon metabolism 662

modulates the efficacy of drugs targeting the cytochrome bc1:aa3 in Mycobacterium 663

tuberculosis. Sci Rep. 2019;9: 8608. doi:10.1038/s41598-019-44887-9 664

7. Howard NC, Marin ND, Ahmed M, Rosa BA, Martin J, Bambouskova M, et al. 665

Mycobacterium tuberculosis carrying a rifampicin drug resistance mutation reprograms 666

macrophage metabolism through cell wall lipid changes. Nat Microbiol. 2018;3: 1099. 667

8. Murima P, McKinney JD, Pethe K. Targeting bacterial central metabolism for drug 668

development. Chem Biol. 2014;21: 1423–1432. 669




33 of 39

9. Bald D, Villellas C, Lu P, Koul A. Targeting Energy Metabolism in Mycobacterium 670

tuberculosis , a New Paradigm in Antimycobacterial Drug Discovery . MBio. 2017;8: e00272-671

17. doi:10.1128/mbio.00272-17 672

10. Beste DJ V, Hooper T, Stewart G, Bonde B, Avignone-Rossa C, Bushell ME, et al. GSMN-673

TB: a web-based genome-scale network model of Mycobacterium tuberculosis metabolism. 674

Genome Biol. 2007;8: R89. 675

11. Jamshidi N, Palsson BØ. Investigating the metabolic capabilities of Mycobacterium 676

tuberculosis H37Rv using the in silico strain iNJ 661 and proposing alternative drug targets. 677

BMC Syst Biol. 2007;1: 26. 678

12. Chindelevitch L, Stanley S, Hung D, Regev A, Berger B. MetaMerge: scaling up genome-679

scale metabolic reconstructions with application to Mycobacterium tuberculosis. Genome 680

Biol. 2012;13: r6. 681

13. Fang X, Wallqvist A, Reifman J. Modeling phenotypic metabolic adaptations of 682

Mycobacterium tuberculosis H37Rv under hypoxia. PLoS Comput Biol. 2012;8: e1002688. 683

14. Lofthouse EK, Wheeler PR, Beste DJV, Khatri BL, Wu H, Mendum TA, et al. Systems-based 684

approaches to probing metabolic variation within the Mycobacterium tuberculosis complex. 685

PLoS One. 2013;8: e75913. doi:10.1371/journal.pone.0075913 686

15. Rienksma RA, Suarez-Diez M, Spina L, Schaap PJ, dos Santos VAPM. Systems-level 687

modeling of mycobacterial metabolism for the identification of new (multi-) drug targets. 688

Seminars in immunology. Elsevier; 2014. pp. 610–622. 689

16. Brahmachari SK, Bhat AG, Kushwaha S, Bhardwaj A. Systems level mapping of metabolic 690

complexity in Mycobacterium tuberculosis to identify high-value drug targets. J Transl Med. 691

2014;12: 263. doi:10.1186/s12967-014-0263-5 692

17. Garay CD, Dreyfuss JM, Galagan JE. Metabolic modeling predicts metabolite changes in 693

Mycobacterium tuberculosis. BMC Syst Biol. 2015;9: 57. 694

18. Ma S, Minch KJ, Rustad TR, Hobbs S, Zhou S-L, Sherman DR, et al. Integrated modeling of 695

gene regulatory and metabolic networks in Mycobacterium tuberculosis. PLoS Comput Biol. 696

2015;11: e1004543. 697

19. Xavier JC, Patil KR, Rocha I. Integration of biomass formulations of genome-scale metabolic 698

models with experimental data reveals universally essential cofactors in prokaryotes. Metab 699

Eng. 2017;39: 200–208. 700

20. López-Agudelo VA, Baena A, Ramirez-Malule H, Ochoa S, Barrera LF, Ríos-Estepa R. 701

Metabolic adaptation of two in silico mutants of Mycobacterium tuberculosis during infection. 702

BMC Syst Biol. 2017;11. doi:10.1186/s12918-017-0496-z 703

21. Rienksma RA, Schaap PJ, Dos Santos VAPM, Suarez-Diez M. Modeling host-pathogen 704




34 of 39

interaction to elucidate the metabolic drug response of intracellular mycobacterium 705

tuberculosis. Front Cell Infect Microbiol. 2019;9. doi:10.3389/fcimb.2019.00144 706

22. Rienksma RA, Schaap PJ, Martins dos Santos VAP, Suarez-Diez M. Modeling the metabolic 707

state of Mycobacterium tuberculosis upon infection. Front Cell Infect Microbiol. 2018;8: 264. 708

23. King ZA, Lu J, Dräger A, Miller P, Federowicz S, Lerman JA, et al. BiGG Models: A platform 709

for integrating, standardizing and sharing genome-scale models. Nucleic Acids Res. 2015;44: 710

D515–D522. 711

24. Kavvas ES, Seif Y, Yurkovich JT, Norsigian C, Poudel S, Greenwald WW, et al. Updated and 712

standardized genome-scale reconstruction of Mycobacterium tuberculosis H37Rv, iEK1011, 713

simulates flux states indicative of physiological conditions. BMC Syst Biol. 2018;12: 25. 714

25. Gu C, Kim GB, Kim WJ, Kim HU, Lee SY. Current status and applications of genome-scale 715

metabolic models. Genome Biol. 2019;20: 121. doi:10.1186/s13059-019-1730-3 716

26. Machado D, Herrgård MJ, Rocha I. Stoichiometric representation of gene–protein–reaction 717

associations leverages constraint-based analysis from reaction to gene-level phenotype 718

prediction. PLoS Comput Biol. 2016;12: e1005140. 719

27. Hädicke O, Klamt S. EColiCore2: a reference network model of the central metabolism of 720

Escherichia coli and relationships to its genome-scale parent model. Sci Rep. 2017;7: 39647. 721

28. Madigan CA, Martinot AJ, Wei J-R, Madduri A, Cheng T-Y, Young DC, et al. Lipidomic 722

analysis links mycobactin synthase K to iron uptake and virulence in M. tuberculosis. PLoS 723

Pathog. 2015;11: e1004792. 724

29. Chao A, Sieminski PJ, Owens CP, Goulding CW. Iron Acquisition in Mycobacterium 725

tuberculosis. Chem Rev. 2018;119: 1193–1220. 726

30. Pereira R, Nielsen J, Rocha I. Improving the flux distributions simulated with genome-scale 727

metabolic models of Saccharomyces cerevisiae. Metab Eng Commun. 2016;3: 153–163. 728

doi:10.1016/j.meteno.2016.05.002 729

31. Huss M, Holme P. Currency and commodity metabolites: Their identification and relation to 730

the modularity of metabolic networks. IET Syst Biol. 2006;1: 280–285. doi:10.1049/iet-731

syb:20060077 732

32. Thiele I, Palsson BØ. A protocol for generating a high-quality genome-scale metabolic 733

reconstruction. Nat Protoc. 2010;5: 93. 734

33. Heirendt L, Arreckx S, Pfau T, Mendoza SN, Richelle A, Heinken A, et al. Creation and 735

analysis of biochemical constraint-based models using the COBRA Toolbox v. 3.0. Nat 736

Protoc. 2019;14: 639. 737

34. Chan SHJ, Cai J, Wang L, Simons-Senftle MN, Maranas CD. Standardizing biomass reactions 738




35 of 39

and ensuring complete mass balance in genome-scale metabolic models. Bioinformatics. 739

2017;33: 3603–3609. doi:10.1093/bioinformatics/btx453 740

35. Kumar VS, Dasika MS, Maranas CD. Optimization based automated curation of metabolic 741

reconstructions. BMC Bioinformatics. 2007;8: 212. 742

36. Thiele I, Vlassis N, Fleming RMT. fastGapFill: efficient gap filling in metabolic networks. 743

Bioinformatics. 2014;30: 2529–2531. 744

37. Benedict MN, Mundy MB, Henry CS, Chia N, Price ND. Likelihood-based gene annotations 745

for gap filling and quality assessment in genome-scale metabolic models. PLoS Comput Biol. 746

2014;10: e1003882. 747

38. Yousofshahi M, Ullah E, Stern R, Hassoun S. MC 3: a steady-state model and constraint 748

consistency checker for biochemical networks. BMC Syst Biol. 2013;7: 129. 749

39. Gopinath K, Moosa A, Mizrahi V, Warner DF. Vitamin B12 metabolism in Mycobacterium 750

tuberculosis. Future Microbiol. 2013;8: 1405–1418. doi:10.2217/fmb.13.113 751

40. Gopinath K, Venclovas Č, Ioerger TR, Sacchettini JC, McKinney JD, Mizrahi V, et al. A 752

vitamin B12 transporter in Mycobacterium tuberculosis. Open Biol. 2013;3: 120175. 753

doi:10.1098/rsob.120175 754

41. Minias A, Minias P, Czubat B, Dziadek J. Purifying selective pressure suggests the 755

functionality of a vitamin B12 biosynthesis pathway in a global population of mycobacterium 756

tuberculosis. Genome Biol Evol. 2018;10: 2326–2337. doi:10.1093/gbe/evy153 757

42. Young DB, Comas I, de Carvalho LPS. Phylogenetic analysis of vitamin B12-related 758

metabolism in Mycobacterium tuberculosis. Front Mol Biosci. 2015;2: 6. 759

doi:10.3389/fmolb.2015.00006 760

43. Noor E. Removing both Internal and Unrealistic Energy-Generating Cycles in Flux Balance 761

Analysis. arXiv Prepr arXiv180304999. 2018. 762

44. Schellenberger J, Lewis NE, Palsson B. Erratum: Elimination of thermodynamically infeasible 763

loops in steady-state metabolic models (Biophysical Journal (2010) 100 (544-553)). Biophys 764

J. 2011;100: 1381. doi:10.1016/j.bpj.2011.02.005 765

45. Price ND, Famili I, Beard DA, Palsson BØ. Extreme pathways and Kirchhoff’s second law. 766

Biophys J. 2002;83: 2879–2882. 767

46. Beard DA, Liang S, Qian H. Energy balance for analysis of complex metabolic networks. 768

Biophys J. 2002;83: 79–86. 769

47. Maranas CD, Zomorrodi AR. Optimization Methods in Metabolic Networks. Optimization 770

Methods in Metabolic Networks. John Wiley & Sons; 2016. doi:10.1002/9781119188902 771




36 of 39

48. Fleming RMT, Thiele I, Provan G, Nasheuer HP. Integrated stoichiometric, thermodynamic 772

and kinetic modelling of steady state metabolism. J Theor Biol. 2010;264: 683–692. 773

49. Fritzemeier CJ, Hartleb D, Szappanos B, Papp B, Lercher MJ. Erroneous energy-generating 774

cycles in published genome scale metabolic networks: Identification and removal. PLoS 775

Comput Biol. 2017;13: e1005494. doi:10.1371/journal.pcbi.1005494 776

50. Basler G. Computational prediction of essential metabolic genes using constraint-based 777

approaches. Methods in Molecular Biology. Springer; 2015. pp. 183–204. doi:10.1007/978-778

1-4939-2398-4_12 779

51. Sassetti CM, Rubin EJ. Genetic requirements for mycobacterial survival during infection. Proc 780

Natl Acad Sci. 2003;100: 12989–12994. doi:10.1073/pnas.2134250100 781

52. Lamichhane G, Zignol M, Blades NJ, Geiman DE, Dougherty A, Grosset J, et al. A 782

postgenomic method for predicting essential genes at subsaturation levels of mutagenesis: 783

Application to Mycobacterium tuberculosis. Proc Natl Acad Sci. 2003;100: 7213–7218. 784

doi:10.1073/pnas.1231432100 785

53. Zhang YJ, Ioerger TR, Huttenhower C, Long JE, Sassetti CM, Sacchettini JC, et al. Global 786

Assessment of Genomic Regions Required for Growth in Mycobacterium tuberculosis. PLoS 787

Pathog. 2012;8: e1002946. doi:10.1371/journal.ppat.1002946 788

54. Griffin JE, Gawronski JD, DeJesus MA, Ioerger TR, Akerley BJ, Sassetti CM. High-789

resolution phenotypic profiling defines genes essential for mycobacterial growth and 790

cholesterol catabolism. PLoS Pathog. 2011;7: e1002251. doi:10.1371/journal.ppat.1002251 791

55. DeJesus MA, Zhang YJ, Sassetti CM, Rubin EJ, Sacchettini JC, Ioerger TR. Bayesian analysis 792

of gene essentiality based on sequencing of transposon insertion libraries. Bioinformatics. 793

2013;29: 695–703. doi:10.1093/bioinformatics/btt043 794

56. DeJesus MA, Gerrick ER, Xu W, Park SW, Long JE, Boutte CC, et al. Comprehensive 795

Essentiality Analysis of the Mycobacterium tuberculosis Genome via Saturating Transposon 796

Mutagenesis . MBio. 2017;8: e02133-16. doi:10.1128/mbio.02133-16 797

57. Minato Y, Gohl DM, Thiede JM, Chacón JM, Harcombe WR, Maruyama F, et al. Genome-798

wide assessment of Mycobacterium tuberculosis conditionally essential metabolic pathways. 799

BioRxiv. 2019;4: 534289. 800

58. VanderVen BC, Fahey RJ, Lee W, Liu Y, Abramovitch RB, Memmott C, et al. Novel 801

Inhibitors of Cholesterol Degradation in Mycobacterium tuberculosis Reveal How the 802

Bacterium’s Metabolism Is Constrained by the Intracellular Environment. PLoS Pathog. 803

2015;11: e1004679. doi:10.1371/journal.ppat.1004679 804

59. DeJesus MA, Ambadipudi C, Baker R, Sassetti C, Ioerger TR. TRANSIT - A Software Tool 805

for Himar1 TnSeq Analysis. PLoS Comput Biol. 2015;11: e1004401. 806




37 of 39

doi:10.1371/journal.pcbi.1004401 807

60. Chen WH, Lu G, Chen X, Zhao XM, Bork P. OGEE v2: An update of the online gene 808

essentiality database with special focus on differentially essential genes in human cancer cell 809

lines. Nucleic Acids Res. 2017;45: D940–D944. doi:10.1093/nar/gkw1013 810

61. de Carvalho LPS, Fischer SM, Marrero J, Nathan C, Ehrt S, Rhee KY. Metabolomics of 811

Mycobacterium tuberculosis reveals compartmentalized co-catabolism of carbon substrates 812

(Maybe Helpful Article about M. tuberculosis Untargeted Metabolomic of Profiling of C13-813

Labelle

A systematic evaluation of Mycobacterium tuberculosis ...Dany JV Beste 3,O and Rigoberto Rios-Estepa...

Documents

Transcript of A systematic evaluation of Mycobacterium tuberculosis ...Dany JV Beste 3,O and Rigoberto Rios-Estepa...