Encoded metabolic remodeling amplifies drug resistance in Mycobacterium tuberculosis

Abstract

Antibiotic pressure causes pathogens to evolve many forms of altered drug susceptibility. In addition to target or activator mutations conferring canonical drug resistance, mutations can serve as steppingstones to or enhancers of resistance. In clinical strains of Mycobacterium tuberculosis (Mtb), we find that idsA2, which encodes an isoprenyl pyrophosphate synthase involved in the synthesis of precursors for essential components of the cell wall and electron transport chain, is undergoing diversifying selection in Lineage 4, and that these mutations are associated with the acquisition of first-line antibiotic resistance. By engineering isogenic Mtb strains to express clinically prevalent variants of idsA2, we show that idsA2 variants alter the inhibitory concentrations of first-line drugs, most significantly increasing the inhibitory concentration of ethambutol by two-fold. Targeted lipid analyses reveal that disrupting IdsA2 function redirects limited resources in the isoprenoid synthesis pathway, leading to increased production of decaprenyl phosphate species. This suggests that idsA2 and ubiA mutations share an ethambutol resistance mechanism. Where idsA2 mutations arise after embB mutations, they maintain their multiplicative effect on the inhibitory concentration of ethambutol such that the variants together result in high-level resistance. As a result, identification of idsA2 mutations can be utilized to improve the specificity of genotypic ethambutol susceptibility testing. Together, this work shows how idsA2 mutations remodel bacterial metabolism and augment ethambutol resistance.

Author summary

Tuberculosis is the deadliest infectious disease in the world and becomes more difficult to treat with the acquisition of drug resistance. Defining the mechanisms and genetic basis of drug resistance will allow for improved screening and drug regimen optimization for treatment success. To identify previously unrecognized mechanisms of altered drug susceptibility, we have combined population genomics and experimental genetics approaches, focusing on Mycobacterium tuberculosis genes evolving in clinical strains. We identified idsA2 as a target of frequent mutations that are associated with drug resistance. Experimental data indicate that idsA2 variants decrease the susceptibility of the bacteria to multiple antibiotics through isoprenoid synthesis remodeling, with the strongest effect on ethambutol resistance. IdsA2 mutations often occur after embB mutations to multiplicatively increase ethambutol resistance. These data suggest that inclusion of idsA2 variants in drug resistance testing could improve the specificity of genotypic detection of ethambutol resistance.

Introduction

While traditionally thought of as monomorphic, the importance of the genetic variation of Mycobacterium tuberculosis (Mtb) has become increasingly evident. This variation has evolved over both short and long timescales, resulting in heterogeneity within a single patient as well as distinct phylogenetic lineages of Mtb across the globe [1]. Drug resistance mutations are some of the most prevalent variants in Mtb [2,3], illustrating the strength of antibiotics as a selective pressure on the Mtb genome. Although the best understood mutations selected by antibiotic pressure cause high-level drug resistance, drug pressure also selects for other classes of mutations that give Mtb a fitness advantage during treatment. Drug tolerance mutations allow Mtb to survive longer in the face of drug stress [4]; resilience mutations shorten the period of post-antibiotic growth arrest [2]; and compensatory mutations mitigate the fitness defects caused by high-level resistance mutations [5]. Furthermore, while resistance mutations are typically found in drug activators or targets [6], low-level resistance mutations that indirectly but meaningfully alter antibiotic action have been associated with treatment failure and enable acquisition of additional drug resistance variants [7,8].

The complexity of bacterial evolution under drug pressure can be effectively dissected by coupling population genomic and experimental approaches. Here we leverage this strategy to identify the impact of clinically prevalent mutations in idsA2, a common target of diversifying selection in the Mtb genome with variants that have been associated with drug resistance [9,10]. IdsA2 encodes an isoprenyl pyrophosphate synthase (IdsA2) that catalyzes the synthesis of precursors required for essential processes such as building the cell wall and bacterial energetics [11]. As determined by transposon insertion screening, idsA2 itself is predicted to be nonessential in vitro, putatively due to enzymatic redundancy [12–14]. Isoprenoid synthesis itself is not targeted by any antimycobacterial agents.

In this work, we leverage our Mtb variant dataset derived from over 55,000 clinical strains to show that idsA2 is mutating for loss of function specifically in Lineage 4 Mtb strains, and these variants are associated with drug resistance. When engineered into Mtb, idsA2 variants increase bacterial survival in the face of treatment with both ethambutol and isoniazid as a result of different consequences of isoprenoid synthesis pathway remodeling. Our phylogenomic and experimental analyses suggest that the effect on ethambutol susceptibility is clinically most relevant, where idsA2 mutations combine with embB mutations to multiplicatively contribute to increased ethambutol resistance beyond critical concentrations [15]. The second-step nature of idsA2 mutations could allow them to be utilized clinically on top of embB variants to improve the specificity of ethambutol resistance predictions. In total, we present idsA2 variants as previously unappreciated genetic contributors to first-line drug resistance discovered through large-scale study of Mtb evolution.

Results

idsA2 is under selection for loss of function

Antibiotics are a profound selective force on the Mtb population and many known drug resistance genes are under diversifying selection even over short time frames in a single individual. We previously identified idsA2 as a frequent target of within-host selection [2]. Here we extended this work to determine if idsA2 variants are under diversifying selection at the population scale, leveraging our variant dataset derived from the whole genome sequencing data of ~55,000 clinical isolates (S1 Table); we refer to this dataset going forward as the 55K dataset [3]. Across all strains, the ratio of nonsynonymous to synonymous mutational events (pN/pS) in idsA2 was 0.99; this was in the top 6% of pN/pS values across the Mtb genome (Fig 1A). While a pN/pS of 1 is usually interpreted as evidence of neutral selection [16], it could also reflect heterogeneous selective pressures across different lineages, which have different mutational backgrounds (S2 Table). Indeed, in Lineage 4, the pN/pS of idsA2 was 1.29 which was within the top 3% of pN/pS values across the genome (Fig 1A). In Lineages 1, 2, and 3, the pN/pS values were closer to the median (S1 Fig), suggesting that the benefit or ability of idsA2 to diversify is unique to Lineage 4.

Fig 1. idsA2 is under selection for loss of function.

(A) Histograms showing the number of genes per pN/pS ratio across the 55K dataset (left) and Lineage 4 (right). Median values indicated by the dotted lines are both 0.69 and idsA2 pN/pS values indicated by the arrows are 0.99 and 1.29. (B) The number of strains per codon from Lineage 4 of the 55K dataset with disruptive in-frame deletions or insertions, frameshift, deleterious or tolerated missense, or nonsense mutations in idsA2. Strains with mutations introduced into H37Rv by recombineering are indicated. (C) Pie chart of the number of strains per idsA2 variant type from Lineage 4 in the 55K dataset. (D) Histogram showing the number of genes per rate of frameshifts and nonsense variants across Lineage 4 of the 55K dataset. The median value indicated by the dotted line is 12.7 mutations/kb and the number of frameshifts and nonsense mutations in idsA2 is 46.3 mutations/kb.

We next sought to predict the effect of idsA2 mutations. In Lineage 4 strains, idsA2 mutations occurred across the gene body, and 61% of strains were predicted to have a deleterious variant in idsA2, annotated as a frameshift, nonsense, or missense mutation computationally predicted to be deleterious based on sequence homology and the physical properties of amino acids (Fig 1B-1C) [17]. To contextualize this genetic signal of loss of function, we compared the burden of nonsense and frameshift mutations in idsA2 to that of other genes in the Mtb genome (S3 Table). IdsA2 was in the top 6% of genes for frequency of frameshift and nonsense mutations (Fig 1D), suggesting that in Lineage 4, idsA2 is under selection for loss of function mutations.

In clinical strains, idsA2 mutations are associated with drug resistance

We then assessed the association between idsA2 mutations and antibiotic sensitivity using data collected by the CRyPTIC Consortium, which performed whole-genome sequencing and measured minimum inhibitory concentrations (MICs) for first- and second-line drugs across more than 10,000 Mtb strains (S2 Fig) [18]. For each drug, we calculated the average difference in MIC between Lineage 4 strains with idsA2 mutations and their phylogenetic nearest neighbor(s), assessing the significance of this difference through permutation testing. IdsA2 mutational events were significantly associated with increases in the MIC of rifampicin, isoniazid, and ethambutol (Figs 2 and S3). The average difference in MIC was further increased when limited to only predicted-deleterious idsA2 mutational events (S4 Fig). We did not identify a statistically significant association between idsA2 mutational events and increases in the MICs of rifampicin, isoniazid, or ethambutol in Lineages 1, 2, or 3 (S5 Fig). Due to the concurrent use of these drugs for treatment of tuberculosis, mutations altering drug susceptibility are genetically linked (S6 Fig). Therefore, these data demonstrate that idsA2 mutations are associated with first-line drug resistance mutations, but not whether idsA2 mutations themselves shift the MIC of any of these drugs. Thus, we next sought to experimentally define the effect of idsA2 mutations.

Fig 2. idsA2 mutations are associated with drug resistance in clinical strains.

Percent changes in minimum inhibitory concentration (MIC) of 13 drugs associated with idsA2 mutational events in Lineage 4 data from the CRyPTIC consortium. Each dot represents the MIC difference attributed to an idsA2 mutational event relative to its nearest neighbor strains, normalized using the maximum possible difference in MIC for the drug. The blue line indicates the average MIC difference across all idsA2 mutational events. A permutation test was performed to assess significance of the average difference associated with idsA2, and multiple test correction was performed with an FDR of 1%. ** q < 0.01.

idsA2 mutations alter sensitivity to first-line antibiotics

We used oligo recombineering to introduce clinically identified idsA2 variants into H37Rv, a Lineage 4 reference strain [19]. We chose two variants: the nonsense variant, S115X, representing an overt loss of function of IdsA2, and the missense variant, L204R, which was the most common idsA2 mutation across Lineage 4 in our 55K dataset (Fig 1B). Because the causality of chromosomal single nucleotide variants cannot be established through complementation, we generated multiple independent clones for each allele and used whole genome sequencing to identify potential secondary site mutations. We also isolated recombineered clones without idsA2 variants to use as wildtype controls.

We then assessed the impact of the engineered idsA2 mutations on Mtb fitness in the presence and absence of drug. In standard axenic culture conditions, idsA2 mutant strains had a minor, although statistically significant, growth defect relative to wildtype strains (S7 Fig). The strains’ antibiotic sensitivity was assessed for four first- and second-line antibiotics: ethambutol, isoniazid, rifampicin, and ofloxacin, quantitating the drug concentration at which 90% of bacterial growth was inhibited (IC90). Mutations in idsA2 caused a roughly 2-fold increase in the IC90 of ethambutol, a more modest increase in the IC90 of isoniazid, and a small but consistent decrease in the IC90 of rifampicin (Figs 3 and S7). We validated these results in a growth-based competition assay, which demonstrated that the idsA2 mutant strains have a significant fitness advantage over wildtype Mtb in ethambutol and isoniazid at drug concentrations equivalent to the IC90s of the wildtype strain (S8 Fig). These data also align with data from transposon-insertion and CRISPRi knockdown libraries under drug treatment, again showing that idsA2 mutants have a fitness advantage in ethambutol and isoniazid [20,21]. Mutations in idsA2 did not increase drug tolerance to any of the tested agents as measured by time-kill assays (S9 Fig).

Fig 3. Isogenic idsA2 mutants have altered sensitivity to first-line antibiotics.

(A) Fold change in IC90 of four tuberculosis antibiotics, ethambutol (EMB), isoniazid (INH), rifampicin (RIF), and ofloxacin (OFX), as measured by AlamarBlue reduction assay. IC90 was calculated from nonlinear regression curves for each clone replicate from two experiments. Differences between each mutant and wildtype were tested by performing ordinary one-way ANOVA with Dunnett’s multiple comparisons test on IC90 values (S7 Fig). (B-E) Nonlinear regression curves of growth in varying concentrations of antibiotics. Points come from two independent experiments, and each represents the average of two (wildtype) to three technical replicates for each clone replicate.

IdsA2 loss of function alters isoprenoid synthesis

To increase our confidence in the relationship between idsA2 variants and these subtle but reproducible antibiotic effects, we next sought to define the metabolic consequences of idsA2 disruption. Previous in vitro and structural analyses showed that purified IdsA2 catalyzes short-chain isoprenoid synthesis, subsequently allowing the synthesis of essential components such as cell wall precursors and menaquinone. Specifically, IdsA2 mediates the conversion of dimethylallyl pyrophosphate (DMAPP) through the intermediate molecule geranyl pyrophosphate (E-C₁₀), to farnesyl pyrophosphate (E-C₁₅) (Fig 4A) [11,22]. It is also capable of producing the downstream molecule geranylgeranyl pyrophosphate (E-C₂₀) [23]. At each of these steps however, previous work suggests there is functional redundancy of IdsA2 with other prenyl diphosphate synthases such as GrcC2, IdsA1, and IdsB [23–25]. It was not clear to what extent these enzymes would compensate for IdsA2 disruption.

Fig 4. IdsA2 functionality impacts isoprenoid synthesis and bacterial energetics.

(A) Model of isoprenoid synthesis feeding into menaquinone (MK) production and the electron transport chain [11,26]. Made in BioRender. Frey, AM. (2026) https://BioRender.com/pz1qsl9 [27]. (B) Concentrations of C₅ (DMAPP) and E-C₂₀ from wildtype and S115X mutant Mtb as measured by LCMS against standards added quantitatively. Dots are a combination of two clone replicates each with three culture replicates. Different shapes indicate different clone replicates. Differences between S115X and wildtype were measured by Mann-Whitney test for each molecule separately. (C) Total menaquinone as measured by LCMS. Dots are a combination of three clone replicates each with three culture replicates. Different shapes indicate different clone replicates. Differences in menaquinone levels determined by Kruskal-Wallis test with Dunn’s multiple comparisons test. (D) Ratio of NAD⁺ to NADH determined by fluorometric assay. Data combines two independent experiments. Differences in NAD⁺/NADH ratio determined by ordinary one-way ANOVA with Tukey’s multiple comparisons test. (E) Fold change in IC90 of bedaquline (BDQ) and pretomanid (PA-824) as measured by AlamarBlue reduction assay. IC90 was calculated from nonlinear regression curves for each clone replicate from one representative experiment for BDQ. IC90 was calculated from nonlinear regression curves for each clone replicate from two experiments for PA-824. Differences between each mutant and wildtype were tested by performing ordinary one-way ANOVA with Dunnett’s multiple comparisons test on IC90 values (S12 Fig).

To directly assess how idsA2 disruption impacts this pathway, we sought to quantify the steady-state abundance of the predicted substrates and products of IdsA2 in the cell. To this end, we lysed the idsA2 nonsense mutant (S115X) and wildtype strains in acetonitrile, methanol, and water and detected metabolites using liquid chromatography mass spectrometry (LCMS). In the mutant strain, DMAPP concentrations were significantly increased and E-C₂₀ concentrations were significantly decreased compared to the wildtype strain (Fig 4B). E-C₁₀ and C₁₅ were not detected, presumably because their previously described rapid turnover prevents their accumulation as detectable pools [11,28]. These data show that loss of IdsA2 activity alters the short-chain isoprenoid synthesis pathway.

Since E-C₂₀ is the precursor to menaquinone, we next used targeted lipid LCMS with collisional mass spectrometry to assess the levels of menaquinone in idsA2 mutant and wildtype strains [29]. We found that total menaquinone levels in the missense mutant strain were significantly decreased compared to the wildtype strain, with the nonsense mutant trending in the same direction (Figs 4C and S10-S11). Since menaquinone and NAD redox are coupled by NADH dehydrogenase activity [26,30,31], we then used a fluorescence assay to measure NAD⁺ and NADH concentrations and found that the ratio of NAD⁺ to NADH similarly trended down in both idsA2 mutant strains (Figs 4D and S11). In previous experimental and genomic studies, decreases in NAD⁺/NADH ratios have been implicated in isoniazid resistance [32–36]. Therefore, the observed alteration in NAD⁺/NADH redox in the idsA2 mutant strains represents a plausible mechanism for the small increase in the IC90 of isoniazid (Fig 3).

Isoniazid activity is impacted by, but does not directly target, bacterial energetics. However, newer World Health Organization (WHO)-recommended second-line antibiotics such as bedaquiline and pretomanid do. Therefore, we asked whether idsA2 mutations also impact the sensitivity to these drugs. Interestingly, idsA2 mutant strains were significantly more sensitive to bedaquiline, demonstrating the potential for collateral sensitivity caused by these mutations (Figs 4E and S12). The mechanism behind this collateral sensitivity is unclear. In a CRISPRi screen, it was observed that both bedaquiline and rifampicin sensitivity increase in strains with increased permeability [21]. We tested the hypothesis that idsA2 mutants also have increased permeability by measuring calcein accumulation in the bacterial cell. However, our findings were inconsistent between the idsA2 mutant strains (S13 Fig). Further work is required to determine the mechanisms of these sensitivity phenotypes.

idsA2 mutants have increased levels of cell wall precursors

We next asked whether the synthesis of cell wall precursors is also impacted by idsA2 mutation. On this arm of the isoprenoid synthesis pathway, geranyl pyrophosphate (E-C₁₀) is converted by Rv1086 into the Z-isomer of farnesyl pyrophosphate (Z-C₁₅) which is then directly converted into decaprenyl phosphate, a crucial carrier lipid for synthesis of multiple constituents of the mycobacterial cell wall (Fig 5A) [37–39]. Directed analysis of the total abundance of decaprenyl phosphate species in idsA2 mutant and wildtype strains found that steady-state levels of decaprenyl phosphate species were significantly increased in both the idsA2 nonsense and missense mutant strains (Figs 5B and S14). These data suggest a straightforward metabolic rerouting whereby IdsA2 impairment causes increased E-C₁₀, which is converted into Z-C₁₅ – presumably by Rv1086 – and this excess Z-C₁₅ is in turn used for synthesis of decaprenyl phosphate [37,38].

Fig 5. idsA2 mutants have increased synthesis of decaprenyl phosphate species.

(A) Model of isoprenoid synthesis feeding into arabinogalactan and lipoarabinomannan (LAM) synthesis pathway [11,40]. Made in BioRender. Frey, AM. (2026) https://BioRender.com/7o2xeht [41]. (B-C) Total decaprenyl phosphate (DP) species and decaprenylphosphoryl pentose (DPP) as measured by LCMS. Dots are a combination of three clone replicates each with three culture replicates. Different shapes indicate different clone replicates. Differences in DPP levels determined by Kruskal-Wallis test with Dunn’s multiple comparisons test.

Notably, one decaprenyl phosphate species with significantly increased levels in the idsA2 mutant strains was decaprenylphosphoryl pentose (DPP) (Figs 5C and S14). Previous work has found that mutation or upregulation of the decaprenylphosphoryl-5-phosphoribose (DPPR) synthase, encoded by ubiA, also increased both DPP levels and ethambutol MIC [42–44]. Since DPP is the combined measure of decaprenylphosphoryl arabinose (DPA) – the carrier molecule for arabinose [45,46] – and the precursor to DPA, decaprenylphosphoryl ribose (DPR), it was proposed that the increased quantity of DPA allowed arabinose to outcompete ethambutol for binding to its drug target, the arabinosyltransferase EmbB [44,47,48]. In support of this model, we found that in addition to increased levels of DPP, idsA2 mutants trended towards higher levels of lipoarabinomannan (LAM), reflecting an increase in product abundance downstream of DPA synthesis, and suggesting that proportional changes in DPP are a reasonable readout to estimate proportional changes in DPA levels (S15 Fig). Together, our genomic, functional, and metabolic data support a model in which idsA2 variants increase the inhibitory concentration of ethambutol through increased production of DPA.

idsA2 mutations have multiplicative effects with embB mutations

Since mutation of EmbB, the arabinosyl transferase targeted by ethambutol, is the most common ethambutol resistance mechanism [49], we next sought to determine if idsA2 mutations interact with mutations in embB. Using the Lineage 4 strains in the 55K dataset, we performed phylogenetic reconstruction to assess whether embB mutations occurred before or after idsA2 mutational events. Excluding nodes in which ordering could not be determined, 90% of the time, embB mutations arose before idsA2 mutations (Fig 6A). EmbB mutations arose before predicted-deleterious idsA2 mutations 97% of the time (S16 Fig). This argues that idsA2 variants are not primarily steppingstone mutations for acquisition of embB variants and instead suggests that they confer an additional advantage on top of existing ethambutol resistance.

Fig 6. idsA2 mutations confer multiplicative drug resistance with embB mutations.

(A) Ordering of 216 idsA2 mutational events with embB mutations within Lineage 4 of the 55K dataset. (B) IC90s of ethambutol (EMB) calculated from nonlinear regression curves for each clone replicate from two experiments. (C) Nonlinear regression curves of growth in varying concentrations of ethambutol for each strain group. Points come from two independent experiments, and each represents the average of two (idsA2 wildtype) to three technical replicates for each clone replicate. (D) Growth curves in standard 7H9 media. Differences between mutants and wildtype were tested by performing ordinary one-way ANOVA with Dunnett’s multiple comparisons test for each day. * p < 0.05, ** p < 0.01, *** p < 0.001.

To experimentally test the effects of idsA2 variants in combination with embB mutations, we selected for ethambutol resistance mutations in both the idsA2 mutant and control strains. Representative strains from each idsA2 strain background were grown on ethambutol, candidate colonies were whole-genome sequenced, and two independent clones with the ethambutol resistance-associated embB Q497R mutation from each strain background were selected [49]. There were no additional mutations shared between strains (S4 Table). We then assessed the effects of the combined mutations on ethambutol resistance. In the absence of additional ethambutol resistance mutations, idsA2 mutations increased the ethambutol IC90 by 0.5 µg/mL, resulting in an IC90 of 1 µg/mL; this is a two-fold shift from baseline (Figs 3 and 6B-6C). In the presence of embB Q497R, idsA2 mutations increased the ethambutol IC90 by 4 µg/mL, resulting in an IC90 of 8 µg/mL (Fig 6B-6C). Therefore, idsA2 mutations maintained the ability to shift the inhibitory concentration of ethambutol by two-fold, even when starting at a higher baseline ethambutol IC90. An ethambutol IC90 of 8 µg/mL is sufficient for a strain to be considered clinically resistant [50]. Thus, idsA2 mutations can combine with embB mutations to elevate ethambutol resistance to clinically relevant concentrations.

An alternative explanation for embB and idsA2 variant ordering could be that idsA2 variants compensate for a loss of fitness due to embB mutation. However, the growth defect caused by the idsA2 variants was maintained in the presence of the embB mutation, while embB Q497R itself does not alter replicative fitness in axenic culture (Fig 6D). Together, these data demonstrate that idsA2 mutations augment ethambutol resistance.

Presence of idsA2 mutations can increase the specificity of ethambutol susceptibility testing

There is a need to improve genotype-based ethambutol resistance diagnostics [51]. Currently, the only gene with mutations associated with ethambutol resistance according to the WHO is embB, and there remain gaps in both the sensitivity and specificity of WHO-verified embB mutations for detecting phenotypic ethambutol resistance [49]. Knowing the effect of idsA2 mutations on ethambutol resistance, we asked if there is clinical utility in identifying idsA2 mutations for genotypic ethambutol susceptibility testing.

Returning to the CRyPTIC dataset, we first assessed the sensitivity and specificity of predicting ethambutol resistance using only WHO-verified ethambutol resistance mutations in embB. Since lineage typing would not typically occur at diagnosis, we used all strains regardless of lineage for this analysis. EmbB-based diagnosis of ethambutol resistance – here defining resistant strains as having an ethambutol MIC ≥ 4 µg/mL and sensitive strains as having an MIC of ≤2 µg/mL – had a sensitivity of 85% and a specificity of 95% (Fig 7A). Since idsA2 mutations often arise after embB mutations, inclusion of idsA2 variants was not expected to increase sensitivity. However, specificity is also an important metric for diagnostics, as a high false positive rate can lead to unnecessary escalation to more toxic drug regimens. We hypothesized that checking for the existence of an idsA2 mutation in an embB mutant strain may improve the specificity of detecting ethambutol resistance. Indeed, the specificity of using the combined presence of a WHO-verified embB mutation and an idsA2 mutation for detection of ethambutol resistance was 99.93%, or a less than 1% false positive rate (Fig 7B).

Fig 7. Second-step ethambutol resistance mutations improve diagnostic specificity.

(A) Distribution of ethambutol MICs across all strains in the CRyPTIC dataset with WHO-verified embB mutations in pink. 11% of strains with WHO-verified embB mutations have an MIC ≤ 2 µg/mL meaning they would be false positives for diagnosing ethambutol resistance. (B) Distribution of ethambutol MICs across WHO-verified embB mutant strains with strains with idsA2 mutations labeled in purple. 8% of strains with a WHO-verified embB mutation and mutation in idsA2 have an MIC ≤ 2 µg/mL meaning they would be false positives for diagnosing ethambutol resistance. (C) Distribution of ethambutol MICs across WHO-verified embB mutant strains with second-step mutations in purples. 5% of strains with a WHO-verified embB mutation and a second step mutation in idsA2, ubiA, or embB, have an MIC ≤ 2 µg/mL meaning they would be false positives for diagnosing ethambutol resistance.

Since idsA2 mutations were present in <2% of embB mutant strains in our dataset, we next asked if we could improve the frequency at which this two-step approach would be effective by detecting mutations in other genes representing a second step to ethambutol resistance. Mutations in ubiA and the presence of more than one mutation in embB have both been associated with higher ethambutol MICs [44,52]. Repeating our analysis with these variants, we found that in both cases the specificity was approximately 99.7% for prediction of ethambutol resistance, again indicating a less than 1% false positive rate (S1 Data). Furthermore, approximately 27% of embB mutant strains had additional mutations in at least one of these three genes (Fig 7C). Therefore, 27% of the time, the specificity in diagnosing ethambutol resistance can be improved to over 99% by identifying secondary-site mutations in strains with WHO-verified ethambutol resistance variants.

Discussion

The pleiotropic effects of idsA2 mutation on drug susceptibility

In this study, we demonstrate that idsA2 is under selection in a lineage-specific manner and that, in Lineage 4 Mtb strains, loss of IdsA2 enzymatic function is associated with first-line drug resistance. IdsA2 loss of function remodels the isoprenoid synthesis pathway, resulting in wide-ranging impacts on essential processes from cell wall synthesis to bacterial energetics. This pathway remodeling is also linked to its most clinically relevant effect, amplifying ethambutol resistance after embB mutation.

This work demonstrates the complex ramifications of metabolic rewiring on drug susceptibility and the importance of epistatic interactions on the presentation of those effects. For example, we observed that idsA2 mutations are sufficient to increase the sensitivity of rifampicin in isogenic strains. However, in the clinical strains assayed in the CRyPTIC dataset, idsA2 mutations often co-occur with mutations in rpoB, which cause high-level rifampicin resistance, seemingly abolishing any enhanced susceptibility that the idsA2 variant may have conferred. We note that the mechanistic basis of the changes in rifampicin and bedaquiline susceptibility remain unclear; we considered changes in cell wall permeability which has been previously associated with a similar set of drug phenotypes [21], but did not find consistent evidence of this in a calcein accumulation assay.

In general, evolution favors antibiotic escape. Thus, while the collateral sensitivity to bedaquiline is a clinically encouraging finding, it is more concerning that some new drugs may suffer from collateral resistance. Quabodepistat, a drug that targets DprE1, one of the two essential enzymes involved in the conversion of DPR to DPA, is currently undergoing Phase 3 clinical trials for the treatment of drug-resistant TB as part of a combination therapy [53]. Although not identified in studies of Quabodepistat, studies of other DprE1 inhibitors have identified ubiA mutations as causing low-level resistance [54,55]. Therefore, it is plausible that increased synthesis of DPR could outcompete DprE1 inhibitors from binding to their targets, and that this new drug class could therefore be less effective against strains with ubiA or idsA2 mutations. Additionally, idsA2 mutations spontaneously developed in the presence of a new thieno[2,3-b]pyridin-6(7H)-one derivative [56]. Together, this illustrates the importance of testing new antimycobacterial compounds on strains with existing drug resistance.

Validating IdsA2 functionality in whole-cell assays

In vitro studies of isoprenoid synthesis enzymes were essential to our understanding of this pathway; however, they incompletely account for the complex whole-cell environment. Here, our analysis of the steady-state abundance of isoprene-based molecules in whole-cell lysate from wildtype and idsA2 mutant bacteria allowed us to infer the role of IdsA2 in the context of functionally redundant enzymes and validate previous findings [11]. Aligning with structural evidence [22], the simplest explanation for the significant accumulation of DMAPP in the IdsA2 loss-of-function strain is that IdsA2 uses DMAPP as a substrate in the whole-cell context. Furthermore, although we could not directly detect high turnover intermediates E-C₁₀ or E-C₁₅, we can infer that IdsA2 uses E-C₁₀ as a substrate, as decreased competition for E-C₁₀ would allow for metabolic rerouting and the observed increase in decaprenyl phosphate species in the idsA2 mutant strains. Finally, previous work has found that IdsA2 can synthesize E-C₂₀ in vitro [11,24]. We found that E-C₂₀ levels were decreased in our IdsA2 loss-of-function strain; however, it is unclear whether this is due to loss of IdsA2 as a redundant E-C₁₅ to E-C₂₀ synthesizing enzyme or due to decreased E-C₁₅ substrate availability. Overall, our work confirms previous findings that IdsA2 is a DMAPP/E-C₁₀ to E-C₁₀/E-C₁₅ synthase, with a potential role in E-C₂₀ synthesis.

Selection on the isoprenoid synthesis pathway

Ethambutol is of moderate potency as an antimycobacterial agent, and it is perhaps not surprising that the cell has discovered multiple ways to evade ethambutol pressure. As discussed above, although mutating the drug target embB is the most common way to acquire ethambutol resistance, there is evidence that ubiA mutations also contribute to ethambutol resistance by increasing DPA levels [44]. In this work, we provide evidence that idsA2 mutations contribute to ethambutol resistance through the same mechanism, and we propose that they may not be the only other mutations to do so. Interestingly, published work assessing signals of selection in the Mtb genome across the four major lineages found modest signals of diversifying selection in grcC2 and idsB – genes encoding enzymes with functional redundancy to IdsA2, acting at the DMAPP to E-C₁₀ and E-C₁₅ to E-C₂₀ steps respectively [23,24] – in Lineage 2 [3]. Further work will be required to determine if mutations in these genes or others phenotypically replicate idsA2 mutants.

Further work is also required to understand the reason for these lineage-specific signals of selection. One possibility is that there may be a preference for acquisition of a mutation in one lineage over another due to an epistatic interaction with a lineage-defining mutation. For idsA2, a variant private to Lineage 4 in rv1086, which encodes the protein that converts E-C₁₀ to Z-C₁₅, is a plausible explanation for the Lineage 4-specific idsA2 selection [37]. It should also be noted that, consistent with the literature [1,57], in our datasets there is a higher rate of embB mutations conferring ethambutol resistance in Lineage 2 strains (45% in the 55K dataset) compared to the other lineages (8–15% in the 55K dataset). This could contribute to the observed lineage-specific selective pressure on idsA2 or could be an additional example of epistasis. Overall, while we have established that there is an increased rate of idsA2 mutation in Lineage 4, it remains to be tested whether an idsA2 mutation introduced to a strain from another lineage would recapitulate the phenotypes we observed in Lineage 4.

The challenge of genomically predicting ethambutol resistance

This work shows how both lineage-specific selection and small MIC changes can obfuscate the importance of a variant in drug resistance, especially where genetic interactions may result in larger shifts in MIC than would be predicted by the effects of each genetic variant alone. These issues likely contribute to the challenge of identifying genetic markers of ethambutol resistance, where there remain gaps in both the sensitivity and specificity of known variants to detect phenotypic ethambutol susceptibility [49,58]. These gaps may be due in part to the continuous nature of ethambutol resistance. The field has traditionally sought out single variants to explain and detect high-level drug resistance. This approach is well suited for drugs such as isoniazid where the MICs have a bimodal distribution due to the substantial jump in MIC caused by single variants. However, ethambutol MICs distribute unimodally with a long tail skewing up towards resistance [50]. This and other work show that this is not only due to embB mutations conferring varying levels of resistance, but due to the stepwise increases of MIC on top of embB variants caused by mutations in idsA2, ubiA, and additional embB mutations [44,52]. Importantly, we show that by leveraging an understanding of combinatorial ethambutol resistance, the specificity of predicting ethambutol resistance can be increased. As the field improves modeling for predicting phenotypic drug resistance, lineage and sub-lineage effects as well as multi-step evolution should be considered.

Conclusion

In this work, we show that idsA2 variants have pleiotropic effects on the cell, of which the elevated abundance of decaprenyl phosphate species provides the most direct connection to the ability of idsA2 mutants to grow in higher ethambutol concentrations. Notably, we show that idsA2 mutations also augment resistance on top of embB mutations, elevating resistance to the clinical breakpoint. This adds idsA2 to a group of second-step mutations contributing to ethambutol resistance including mutations in ubiA and second-site mutations in embB. Detection of these second-step mutations can increase the specificity of embB mutation-based ethambutol resistance prediction. These results not only help close the gap in our genotypic understanding of phenotypic drug susceptibility in Mtb, but also emphasize the importance of considering the genetic interactions that might create lineage-specific or combinatorial drug phenotypes that might be relevant to patient care.

Methods

Genomic analyses

The two datasets used in this study are referred to as the 55K dataset and the CRyPTIC dataset. We used a table of pN/pS values calculated from the 55K dataset for previously published work [3], but excluding reversions (S1 Table). Variant calling and tree building for the 55K dataset was performed for a previous study [3]. Resulting phylogenetic trees, ancestral reconstruction, and variant calling were deposited to Mendeley (https://doi.org/10.17632/h9dd6hsvc2.1). The CRyPTIC Consortium has generously made their dataset publicly available. We downloaded the CRyPTIC_reuse_20231208.csv and used the ENA_RUN information to pull whole genome sequences from the Sequence Read Archive (SRA) and perform variant calling using the same pipeline as used for the 55K dataset with H37Rv as a reference [3]. Resulting phylogenetic trees, ancestral reconstruction, and variant calling were deposited to Mendeley (https://doi.org/10.17632/cy4dy352kj.1). Included in this pipeline, annotations and effects for each variant were called with snpeff and sift4g, and ancestral reconstruction was performed using pastml with DOWNPASS parsimony [17,59,60]. Whenever we refer to variants in a specific gene, we mean any variants that are annotated as being within the gene and meeting the following criteria: disruptive inframe deletion or insertion, frameshift, missense, or stop variants. When we refer specifically to predicted-deleterious idsA2 mutations, those were variants annotated as frameshift or stop variants by snpeff or deleterious missense variants by sift4g. tbprofiler was used to assign lineages for each strain; only Lineage 4 strains were used for all analyses with the exception of Figs 1A and 7 [61,62]. Code used for the following analyses is available in the associated Github (https://github.com/abigailmfrey/IdsA2_Paper_Code). All analyses were performed using the Harvard FAS Research Computing (FASRC) cluster.

Variant plotting

Variants within the idsA2 coding region (Gene ID Rv2173) and meeting the variant requirements described above were pulled from the annotated event calls for the whole dataset of interest. Lineage 4 strains with these variants were identified. The number of strains with each variant was calculated, and then variants were grouped by codon per annotation type. Missense variants were designated as deleterious or tolerated by sift4g [17]. Disruptive inframe insertions and deletions were combined, and any combination of annotations including a frameshift was counted as a strain with a frameshift variant. The results were plotted in lollipop form across the gene or in pie chart form to demonstrate abundance of variant type using GraphPad Prism 10.

Loss of function prediction analysis

To compare loss of function signals across the genome, the rate of frameshift and nonsense mutations per kilobase per gene was calculated using the 55K dataset. Frameshift and stop variants were pulled from the event calling, removing variants with a miss rate of 5% or higher. Variants were also removed if they fell in a region of the genome deemed unreliable for variant calling or had more than one reversion [63]. Genes must have had at least 80% of the gene callable and have at least 100 bases. The number of frameshifts and stop variants per gene was then divided by the number of kilobases in the callable region of the gene resulting in events per kilobase.

Variant association with MIC in CRyPTIC dataset

To assess the association of idsA2 variants with MIC changes across the 13 drugs from the CRyPTIC dataset, we sought to determine the difference in MIC between strains resulting from idsA2 mutational events and their nearest neighbors. First, we performed filtering on the MIC data. Only MIC values defined as high quality for the drug being analyzed were used. Due to differences in plate set-up, the extreme values were adjusted to reflect only the values available on both plates [50]. For example, UKMYC5 has an ethambutol range of 0.06-8 mg/L with minimum MIC values labeled as <=0.06 mg/L and maximum >8 mg/L. UKMYC6 has an ethambutol range of 0.25-32 mg/L with minimum MIC values labeled as <=0.25 mg/L and maximum >32 mg/L. For our analysis, all values less than or equal to 0.25 (<=0.06, 0.12, <=0.25, 0.25) from both plates were labeled as 0.25 while all values greater than or equal to 8 (8, > 8, 16, 32, > 32) were labeled as 8.

Next, we calculated the difference in MIC for each strain and node relative to their respective near neighbors on the phylogenetic tree. Variant calling, tree building, annotation, and variant settings are described above. The CRyPTIC tree was pruned using dendropy to include only strains in the lineage of interest with high quality MIC data for the drug of interest [64]. Across the whole tree, we calculated the difference in MIC between each target and sibling set, otherwise referred to as the nearest neighbor(s). By calculating a value for the difference in MIC for each strain and node, we could then use this dataset to both identify nodes with idsA2 mutational events and permute randomly across the tree.

We identified the nodes with idsA2 mutational events using the variant calling per our standard settings coupled with the ancestral reconstruction. Each point plotted in Fig 2 is data from a single idsA2 mutational event. We averaged the differences in MIC across idsA2 events to acquire the associated change in MIC for idsA2 mutational events, visualized by the blue line in Fig 2. There are three instances in which the nearest neighbor strains to an idsA2 event node have an idsA2 mutant strain, however, we chose to retain these events in the dataset for consistency with the permutation test, in which we averaged the calculated differences across the same number of random nodes as idsA2 events in that tree, 10,000 times. Significance was determined by how many times the test returned an average difference greater than the idsA2 events. Multiple testing was corrected by two-stage linear step-up procedure of Benjamini, Krieger and Yekutieli in Graphpad Prism 10 with a 1% false discovery rate.

Mycobacterium tuberculosis isogenic mutant strain construction

IdsA2 point mutants were constructed using oligo-mediated recombineering as previously described [2,19]. Specifically, a strain of H37Rv was transformed with two plasmids, pKM427 (a chromosomal site L5 integrating vector carrying a zeocin resistance cassette and a hygromycin cassette that is inactivated by a premature stop codon) and pKM402 (an episomal vector carrying the phage recT recombinase gene under an ATc inducible promoter). The resulting strain was grown to early log phase, induced with ATc, then transformed with two single-strand DNA oligos. One oligo confers hygromycin resistance, and the other oligo carries the idsA2 mutation of interest (S115X or L204R). The transformants were recovered for 2 days in 7H9 (7H9 salts with 0.2% glycerol, 10% OADC, 0.05% Tween-80) then plated on 7H10 (7H10 media with 0.5% glycerol, 10% OADC, 0.05% Tween-80) containing hygromycin (50 μg/mL). Colonies were screened for idsA2 mutations by Sanger sequencing. Three colonies were retained for each mutant of interest and three hygromycin-resistant colonies with no idsA2 mutations were retained for use as wildtype controls. All strains were cured of the pKM402 plasmid. The resulting strains were whole genome sequenced (200 Mbp Illumina DNA sequencing, SeqCenter). Whole genome sequencing reads have been deposited onto the SRA (BioProject ID PRJNA1356913). Variants relative to H37Rv were identified using breseq [65]. Variants found in all strains, in masked regions, or found to be inconsistently called within our dataset were not included in S4 Table [63]. There were no additional variants shared between clone replicates (S4 Table). However, a single base pair deletion in papA5, a gene involved in PDIM biosynthesis, was found in one L204R clone. Due to the loss of PDIM enhancing bacterial growth rate, this clone was dropped from the study [66]. Unless otherwise indicated, experiments were conducted using three independently derived clones of wildtype and S115X strains, and two independently derived clones of the L204R strain.

Growth curves

Mid-log phase cultures grown in 7H9 at 37°C shaking were diluted to an OD₆₀₀ 0.001 in duplicate in 15 mL 7H9. OD₆₀₀ was measured on days 3–7. Differences between each mutant and wildtype were tested by performing ordinary one-way ANOVA with Dunnett’s multiple comparisons test. Two independent growth curves were performed.

Antibiotic resistance measurement

Antibiotic resistance against four commonly used anti-tuberculosis drugs (EMB, INH, RIF, and OFX) and two newer second line antibiotics (BDQ and PA824) was assessed using AlamarBlue assays as previously described with the following modifications [2,8]. Strains were grown to mid-log phase in 7H9 and diluted to a final OD₆₀₀ of 0.0015 in in 200 µL 7H9 media containing antibiotics at indicated concentrations in a 96-well plate. After five days of incubation at 37°C with constant agitation, AlamarBlue reagent was added at 10% of the culture volume. The plates were then incubated at 37°C for one day before measuring absorbance. The amount of AlamarBlue reduction was estimated by subtracting background OD₆₀₀ from OD₅₇₀. The percentage of bacterial growth was calculated on a row-by-row basis by subtracting the relative absorbance measure of the highest drug concentration well from the relative absorbance measure of the no-drug control well. Strains were plated in duplicate (WT) or triplicate (mutant) rows. The antibiotic concentration at which growth is inhibited by 90% (IC90) for each clone replicate was approximated by performing nonlinear regression on data from two independent experiments performed with different starting concentrations using GraphPad Prism 10. For bedaquiline, calculations were performed for the two experiments independently and one representative experiment is shown. The three wildtype clones’ IC90s were averaged and used to calculate the IC90 fold-change for each clone of each strain per drug. P-values represent differences between each mutant and wildtype IC90 as tested by performing ordinary one-way ANOVA with Dunnett’s multiple comparisons test. Nonlinear regression curves are shown for data where replicate rows were averaged, and regression was performed for all clone replicates together for visual clarity.

Competition assay

Competition assays in the presence of ethambutol and isoniazid were performed as previously described with the following modifications [8]. Strains were grown to mid-log phase in 7H9. Cultures were normalized to an OD₆₀₀ of 0.4. Clone replicates were then pooled in equal ratio to reduce any background effect. The wildtype pool was then pooled in a 1:1 ratio with either the S115X pool or the L204R pool. A sample of each input pool was collected and gDNA was extracted to measure technical error. The input pools were then added into fresh 7H9 at final OD₆₀₀ 0.005 in technical triplicate. Drugs were added to the relevant flasks at the wildtype IC90 concentration as determined by the AlamarBlue assays (0.55 µg/mL ethambutol and 0.04 µg/mL isoniazid). Cultures were grown for 5 days shaking at 37°C. On day 5, the OD was taken, and the remaining bacteria was spun down, resuspended in Tris-EDTA-NaCl buffer (10 mM Tris pH 8, 1 mM EDTA, 25 mM NaCl) and gDNA was extracted using phenol:chloroform:isoamyl alcohol (25:24:1) with bead beating. A 656 base pair region of idsA2 was amplified from purified gDNA with the following primers: GACCAGCTGCGGAGGTATTT and CCGAAATGCTCGAAGATGGC. The resulting PCR product was gel purified and Sanger sequenced. The relative proportion of the mutant strain in each sample was determined based on the nucleotide peak heights. P-values represent the differences between the proportion of the mutant strain in the input, ethambutol, and isoniazid conditions compared to the no drug control as tested by performing ordinary one-way ANOVA with Šídák's multiple comparisons test. Experiment was repeated three times with similar results.

Antibiotic tolerance measurement

Antibiotic tolerance against four commonly used anti-tuberculosis drugs (EMB, INH, RIF, and OFX) was assessed using a time-kill assay as previously described with the following modifications [2]. Strains were grown to mid-log phase in 7H9 and diluted to a final OD₆₀₀ of 0.05 in 10 mL 7H9 media in a 50 mL inkwell. Antibiotics were added to final concentrations of 100x the calculated MIC of our wildtype strain (55, 3, 0.6, and 34 µg/mL respectively). The cultures were then incubated at 37°C with constant agitation and sampled at day 1, day 3, and day 5. On each day, a 1 mL aliquot of each antibiotic-supplemented culture was taken, and the cells were washed and resuspended in 1mL 7H9 without antibiotic. After washing off the antibiotics, a series of dilutions determined by pilot experiments were prepared in a 96-well plate using fresh 7H9 media. 5μL of each dilution was spotted on a 7H10 agar plate with no antibiotics. Colony enumeration was performed at 24 days after plating.

Metabolite extraction

Two clone replicates of wildtype and S115X were grown in technical triplicate to mid-log phase in 7H9. 10 mL of culture was spun down at 3000 rpm for 10 minutes at 4°C then the pellet was resuspended in 1 mL of chilled 2:2:1 acetonitrile:methanol:water. Bacteria were then lysed by bead beating 0.5 min x 6 resting on ice in between cycles. Tubes were then spun down at 14,000 rpm for 10 minutes at 4°C. Supernatant was removed and filtered twice for removal from the BL3.

LC-MS/MS analysis of isoprenoid pyrophosphates

Isoprenoid pyrophosphates were analyzed using an LC–MS/MS system equipped with a Phenomenex Luna C18 column (150 × 2 mm) by the Harvard Center for Mass Spectrometry. Mobile phase A consisted of 10 mM ammonium bicarbonate, and mobile phase B was methanol. Standards were purchased in 1 mg/mL concentrations in methanol from Sigma (D4287, G6772, F6892, G6025) and run at known concentrations to generate a standard curve. 10 µL aliquots of standards and samples were injected for each run. The flow rate was maintained at 0.3 mL/min. The LC gradient program was as follows: 0 min, 0% B; 6 min, 99% B; 10 min, 99% B; 10.1 min, 0% B; 13 min, 0% B. Detection was performed in negative ionization mode using multiple reaction monitoring (MRM). MRM transitions and corresponding parameters for each analyte were as follows:

• DMAPP (C5): m/z 245 → 78.9 (fragmentor 80 V, CE 17 V, cell voltage 4 V) and m/z 245 → 158.8 (fragmentor 80 V, CE 9 V, cell voltage 4 V)

• GPP (E-C10): m/z 313.1 → 78.9 (fragmentor 90 V, CE 21 V, cell voltage 5 V) and m/z 313.1 → 294.9 (fragmentor 90 V, CE 4 V, cell voltage 4 V)

• FPP (C15): m/z 381.1 → 78.9 (fragmentor 110 V, CE 21 V, cell voltage 13 V) and m/z 381.1 → 158.8 (fragmentor 110 V, CE 13 V, cell voltage 4 V)

• GGPP (E-C20): m/z 449.2 → 79 (fragmentor 140 V, CE 33 V, cell voltage 4 V) and m/z 449.2 → 158.8 (fragmentor 140 V, CE 13 V, cell voltage 4 V)

Source parameters were set as follows: gas temperature, 350 °C; gas flow, 12 L/min; nebulizer pressure, 35 psi; sheath gas temperature, 400 °C; sheath gas flow, 12 L/min.

Concentrations of DMAPP and GGPP were normalized based on protein concentration of the sample as determined by microBCA (Thermo Scientific). Difference between S115X and wildtype were measured by Mann-Whitney test for each molecule separately.

Lipid extraction

Lipid extraction was done as previously described [29]. Cultures were grown in 7H9 to mid-log phase. They were then spun down and washed with detergent-free media then diluted to an OD₆₀₀ of 0.0075 in 45 mL 7H9 without detergent in triplicate. Cultures were then grown to an OD₆₀₀ of 0.55-0.7. Cultures were spun at 4000 rpm for 10 minutes and then washed twice with 10 mL Optima LCMS grade water. Pellet was resuspended in 1 mL methanol. Resuspended solution was then added to a glass vial containing 3 mL methanol and 2 mL chloroform creating a final ratio 1:2 chloroform:methanol (C:M). Samples were left for 30 minutes at room temperature to sterilize before removal from the BL3. Samples were then incubated on a rotator for 1 hour at room temperature and then were spun down at 3000 rpm for 10 minutes at room temperature and supernatant was transferred to a new glass vial. Pellet was resuspended in 1:1 C:M and incubated for 1 hour on rotator at room temperature. Spin was repeated and supernatant was collected and added to previous supernatant. Supernatant was then evaporated using Genevac. Dried lipids were resuspended in 4 mL 1:1 C:M. Lipid suspension was sonicated at 60 sonics/minute for 3 minutes, shaken throughout. Samples were again spun down and then supernatant was transferred to a pre-weighed amber glass vial. Lipids were dried using Genevac. Vials were then weighed again to acquire final total lipid mass. All weighing was performed using a Mettler Toledo XP205 analytical balance with an EN 8 SLC deionizing power supply. Lipids were resuspended to 1 mg/mL 1:1 C:M and stored at -20°C.

Lipid LCMS and collisional mass spectrometry

Lipid HPLC-MS and collision-induced dissociation mass spectrometry (CID-MS) were carried out using a reversed phase method [67]. Lipid preparations at 1 mg/mL in 1:1 C:M were evaporated under nitrogen and dissolved to the same volume in solvent A (95:5 methanol:water with 2 mM ammonium formate). 10 μg (10 μL) lipid was injected per sample, using an Agilent 1260 infinity HPLC with an Agilent Poroshell 120 EC-C18 column (1.9 μm particle size, 3.0 x 50 mm, 699675–302), connected to an Agilent 6530 electrospray ionization quantitative time-of-flight (ESI-QTOF) mass spectrometer. A flow rate of 0.15 mL/minute was used, with the following ratios of solvent A to solvent B: 0–4 minutes, 100% A; 4–10 minutes, gradient to 100% B; 10–15 minutes, 100% B; 15–20 minutes, gradient to 100% A; followed by a 5-minute postrun with 100% A. Solvent B was 90:10:0.1 1-propanol:cyclohexane:water with 3 mM ammonium formate. Ionization was maintained at 325°C gas temperature, 8 L/minute flow rate of drying gas, 35 PSI nebulizer pressure, and 3500 V. Spectra were collected at 2 spectra/second from 100 to 3200 m/z, in both the positive and negative ion modes, over separate runs. CID-MS was performed on the same instrument with the same conditions, detecting a 40–3000 m/z range with collision energy of 20–50 V, using a medium (4 m/z) mass window, starting from the chosen ion m/z rounded down to the nearest whole number minus 0.3 (e.g., for ion m/z 909.6378, CID-MS was carried out from 908.7-912.7 m/z).

Lipid measurement

Raw data and analysis files for lipid HPLC-MS can be found in the Zenodo repository, at https://doi.org/10.5281/zenodo.21362678. All analysis of lipid mass spectral data was performed using Agilent Masshunter Qualitative Analysis software. Menaquinones were detected in the positive mode, and decaprenyl species in the negative mode, by searching for exact m/z matches. The single ionized adduct with the highest intensity was chosen for each lipid. Individual chromatographic peaks were visualized and structurally validated by CID-MS and integrated in a sample-blinded manner, in profile mode, with a 10 ppm m/z window and no smoothing. Figs 4C and 5B,5C show fold change values calculated for each strain’s clone and culture replicates relative to an average value taken across all wildtype samples. There is significantly more variability within culture replicates than between clone replicates as shown in S14 Fig, allowing us to plot clone and culture replicates together. This unbiased HPLC-MS method for total Mtb lipids cannot distinguish between decaprenylphosphoryl ribose and decaprenylphosphoryl arabinose, nor decaprenylphosphoryl glucose and decaprenylphosphoryl mannose. Therefore, we report detection of decaprenylphosphoryl pentose (putatively a combination of ribose and arabinose) and decaprenylphosphoryl hexose (putatively a combination of glucose and mannose). Total decaprenyl phosphate species values were calculated by summing the intensities of each decaprenyl phosphate species we could detect: decaprenyl phosphate, decaprenylphosphoryl pentose and decaprenylphosphoryl hexose. Similarly, total menaquinone levels were calculated by summing intensities of menaquinones and dihydromenaquinones 8 and 9. P-values were determined by Kruskal-Wallis test with Dunn’s multiple comparisons test.

NAD⁺/NADH measurements

NAD⁺/NADH measurements were performed with the Fluorescent NAD/NADH Detection Kit (Cell Technology). Briefly, cultures were grown to mid-log phase in 7H9, normalized by OD₆₀₀, and 1 mL or 3 mL of culture were taken for NAD or NADH measurements respectively. Samples were pelleted by centrifugation, washed twice with PBS, and resuspended in extraction buffer. Lysis buffer and bead beating lysed the bacteria and samples were incubated at 60°C for 1.5 hours. On ice, reaction and the opposite extraction buffer were added. The plate setup was performed as described with half the volume and after 1 hour incubation was read at 550 nm excitation and 595 nm emission.

Calcein accumulation assay

Permeability was assessed using Calcein AM (Invitrogen) as previously described [68]. Cultures were grown to mid-log phase. 1.8 mL was pelleted and washed twice in PBST (PBS pH 7.4 with 0.05% Tween-80). Cells were resuspended in PBST to a final OD 0.4 and 100 µL was aliquoted to a black bottom plate in technical triplicate. Plates were incubated at 37°C with shaking for 30 minutes. 1 µg/mL Calcein AM was then added to each well. Plates were incubated at 37°C shaking and read at 488 nm excitation 520 nm emission every minute for 1 hour. RFU/min was estimated by performing a simple linear regression on the data from 20 minutes on, to allow uptake to reach an equilibrium. The data across three experiments was combined and an ordinary one-way ANOVA with Tukey’s multiple comparisons test determined the difference between strains.

Detection of LAM by western blot

The method for cell glycan extraction was modified from previous work [69]. 15 mL mid-log phase culture was pelleted, spun down, resuspended in 2 mL PBST, and heat killed in 1 mL aliquots. Bacteria were then lysed by bead beating 1.5 min x 5 resting on ice in between cycles. Samples were spun at 20,000 G for 3 minutes and supernatant was removed and set aside. Cells were washed with 400 µL PBST and this was added to the supernatant. Collected supernatant was treated with 1: 2000 DNase benzonase and incubated at 37°C for 30 minutes. A 20 µL aliquot was treated with proteinase K and 10% SDS and incubated at 37°C for another 30 minutes. Samples were then treated as previously described [3]. Samples were incubated for 10 minutes at 95°C in 6X Laemmli SDS-Sample Buffer (Boston BioProducts). SeeBlue Plus2 Pre-stained Protein Standard (Invitrogen) and samples were loaded onto NuPAGE 4–12%, Bis-Tris gels (Invitrogen) and run in NuPAGE MES SDS Running Buffer (Invitrogen) for 30–45 minutes at 170 V. Protein was transferred onto nitrocellulose membranes (Thermo Scientific) using 10X Transfer Buffer (Boston Bioproducts) diluted in methanol and water (1:1:8), on a Trans-Blot Turbo Transfer System (BioRad) for 30 minutes at 25 V and 1 A. Membranes were rinsed with PBS and blocked for two hours at room temperature using Blocking Buffer for Fluorescent Western Blotting (Rockland). Monoclonal Anti-Mycobacterium tuberculosis LAM antibody, Clone CS-35, obtained through BEI Resources (NIAID, NIH: Monoclonal Anti-Mycobacterium tuberculosis LAM, Clone CS-35 (produced in vitro), NR-13811) was diluted 1:1000 in blocking buffer supplemented with 0.1% Tween-20. Membranes were incubated in primary antibody overnight at 4°C. Membranes were washed with PBS with 0.05% Tween-20. IRDye 800CW Goat anti-Mouse IgG Secondary Antibody (LI-COR) was diluted 1:20000 in blocking buffer with 0.1% Tween-20 and 0.0025% SDS. Membranes were incubated in secondary antibody for 1 hour at room temperature. Membranes were washed with PBS with 0.05% Tween-20 and left in PBS before imaging. Images were acquired with LI-COR Odyssey CLx Imager. Blot quantifications were done in Empiria Studio using automatic lane finder and background adjustments and manually adjusted band finding. Display adjustments for figure visualizations were also done in Empiria Studio. Intensities were normalized by protein concentration as determined by microBCA (Thermo Scientific). Fold changes were calculated relative to wildtype separately for each experiment, and then all data was plotted in S15 Fig, with shape fills indicating the different experiments.

Ordering analysis

To determine the phylogenetic order of idsA2 and embB mutations, we utilize the ancestral reconstruction of idsA2 mutational events meeting the indicated variant criteria and occurring in Lineage 4 of the 55K dataset. However, embB repeatedly acquires the same variants making the ancestral reconstruction of embB events more likely to have multiple parsimonious solutions. To avoid this uncertainty, we instead perform a local parsimony analysis for embB around idsA2 mutational events. To do so, at each idsA2 mutation node, we asked which of the strains descending from this event had embB variants. First, if none of the strains had any embB mutation, the event was labeled as having only idsA2. Next, if a certain embB variant was called as a “miss” in all strains under the idsA2 event, indicating poor variant calling, the variant was thrown out for this event. Next, if all strains were found to have the same embB variant, we then went back to the parent node and asked if all strains under that node have the same embB variant indicating that the embB variant happened at the parent node or earlier. If all strains under the parent node had the embB variant, the event was called as embB happening first. If there was any other combination of calls within the parent strains (no variant, variant, and/or misses) the embB event is assumed to occur at the same node as the idsA2 event. Finally, if only some strains resulting from the idsA2 event had an embB variant, this is called as the embB event(s) coming second.

Acquisition of ethambutol resistance mutations

A representative clone for each of our strains of interest (idsA2 WT, S115X, L204R) were grown on ethambutol at 4 or 8 µg/mL. Several colonies were then picked, grown up in 7H9, and gDNA was extracted. WGS was performed (200 Mbp Illumina DNA sequencing, SeqCenter) and reads were mapped to H37Rv using breseq [65]. Two strains from each background with embB Q497R mutations were used for further experiments. There were no additional mutations shared between strains (S4 Table).

Sensitivity and specificity of ethambutol resistance predictions

Analysis was performed using strains of all lineages in the CRyPTIC dataset. Variant calling criteria and MIC handling is described above. Strains were considered sensitive to ethambutol if they had an MIC at ≤2 µg/mL, and resistant if they had an MIC of ≥4 µg/mL. WHO-verified embB mutations are those defined as associated with resistance to ethambutol in the second edition of the drug resistance catalog [49]. After identifying the presence of WHO-verified embB mutations, our variant calling criteria were then used to determine whether each strain had mutations in idsA2, ubiA, or any additional embB mutation. Specificity was calculated for each variant condition of interest.

Supporting information

S1 Fig. Distribution of pN/pS ratios for L1, L2, and L3.

Histograms of pN/pS ratios for all genes in Mtb from the Lineages 1, 2, and 3 from the 55K dataset. Median values indicated by dotted line are 0.64, 0.75, and 0.67 respectively. idsA2 pN/pS values indicated are 0.61, 0.81 and 0.87 respectively.

(PDF)

S2 Fig. IdsA2 variants in the CRyPTIC dataset.

(A) The number of strains per codon from Lineage 4 of the CRyPTIC dataset with disruptive in-frame deletions or insertions, frameshift, deleterious or tolerated missense, or nonsense mutations in idsA2. (B) Pie chart of total percentages of each variant type among Lineage 4 strains in the CRyPTIC dataset.

(PDF)

S3 Fig. Association between idsA2 mutational event and MIC change in the CRyPTIC dataset.

Difference in MIC of 13 drugs associated with idsA2 mutational events in Lineage 4 data from the CRyPTIC consortium. Each dot represents the MIC difference attributed to an idsA2 mutational event relative to its nearest neighbor strains. The blue line indicates the average MIC difference across all idsA2 mutational events. A permutation test was performed to assess the significance of the average difference associated with idsA2, and a multiple test correction was performed with an FDR of 1%.

(PDF)

S4 Fig. Association between predicted-deleterious idsA2 mutational event and MIC change in the CRyPTIC dataset.

Difference in MIC of 13 drugs associated with predicted-deleterious idsA2 mutational events in Lineage 4 data from the CRyPTIC consortium. Each dot represents the MIC difference attributed to an idsA2 mutational event relative to its nearest neighbor strains. The blue line indicates the average MIC difference across all idsA2 mutational events. A permutation test was performed to assess the significance of the average difference associated with idsA2, and a multiple test correction was performed with an FDR of 1%.

(PDF)

S5 Fig. Association between idsA2 mutational event and MIC change in Lineages 1, 2, and 3 of the CRyPTIC dataset.

Difference in MIC of rifampicin (RIF), isoniazid (INH), and ethambutol (EMB) associated with idsA2 mutational events in Lineage 1, 2, and 3 data from the CRyPTIC consortium. Each dot represents the MIC difference attributed to an idsA2 mutational event relative to its nearest neighbor strains. The blue line indicates the average MIC difference across all idsA2 mutational events. A permutation test was performed to assess the significance of the average difference associated with idsA2, and a multiple test correction was performed with an FDR of 1% across each lineage.

(PDF)

S6 Fig. MIC distribution and presence of drug resistance mutations in idsA2 mutant strains from the CRyPTIC dataset.

Rifampicin (RIF), isoniazid (INH), and ethambutol (EMB) MICs for idsA2 mutant strains in Lineage 4 data from the CRyPTIC consortium. Each dot is an idsA2 mutant strain. The dot is pink if there is a mutation in the indicated gene associated with drug resistance for the plotted drug (rpoB for RIF, katG for INH, and embB for EMB) in that strain.

(PDF)

S7 Fig. Growth curve and IC90 statistics.

(A) Growth curves of wildtype and idsA2 mutant strains in standard 7H9 media. Differences between each mutant and wildtype were tested by performing ordinary one-way ANOVA with Dunnett’s multiple comparisons test for each day. * p < 0.05, ** p < 0.01, *** p < 0.001 **** p < 0.0001. (B) Plots of IC90 values for each strain separated by drug: ethambutol (EMB), isoniazid (INH), rifampicin (RIF), and ofloxacin (OFX). Differences between each mutant and wildtype were tested by performing ordinary one-way ANOVA with Dunnett’s multiple comparisons test.

(PDF)

S8 Fig. Competition assay.

The proportion of mutant DNA in a wildtype to single mutant competition assay either at initial pooling (input), or after 5 days of growth in 7H9 without drug, 0.55 µg/mL ethambutol, or 0.04 µg/mL isoniazid. S115X is shown to the left, L204R is shown to the right. Dotted line at input. Differences between the no drug condition and other conditions were tested by performing ordinary one-way ANOVA with Šídák's multiple comparisons test.

(PDF)

S9 Fig. Kill curve assay.

(A) Minimum duration to 90% cell death (MDK90) for wildtype and idsA2 mutant strains in four tuberculosis antibiotics: ethambutol (EMB), isoniazid (INH), rifampicin (RIF), and ofloxacin (OFX). The concentration of each drug was 55, 3, 0.6, and 34 µg/mL. MDK90 was calculated using a semilog curve equation determined using the line segment through which Fraction Surviving = 0.01 fell. Differences between each mutant and wildtype for each drug were tested by ordinary one-way ANOVA with Dunnett’s multiple comparisons test. (B-F) Fraction of surviving bacteria over time calculated from colony enumeration.

(PDF)

S10 Fig. Representative chromatograms and collision spectra for menaquinone species.

Red boxes on extracted ion chromatograms (EICs) indicate peaks with lipid identity confirmed by CID-MS via the highlighted fragments. For each lipid, displayed prenyl stereochemistry and position of saturated groups is assumed from literature patterns and may differ across chromatographic peaks. Blue diamonds in spectra indicate the lowest targeted m/z value for CID-MS.

(PDF)

S11 Fig. Supporting data for menaquinone and redox analysis.

(A) Intensity measured in arbitrary units (counts) for measured menaquinone species. (B) The concentrations of NAD⁺ and NADH as measured by fluorometric assay. Data combines two independent experiments.

(PDF)

S12 Fig. IC90 statistics and nonlinear regression curves for second-line antibiotics.

(A,C) Plots of IC90 values for each strain separated by drug: bedaquiline (BDQ) and pretomanid (PA-824). Differences between each mutant and wildtype were tested by performing ordinary one-way ANOVA with Dunnett’s multiple comparisons test. (B,D) Nonlinear regression curves of growth in varying concentrations of antibiotics. Points each represents the average of two (wildtype) to three technical replicates for each clone replicate. For BDQ, the plot is representative of two experiments. For PA-824, points come from two independent experiments.

(PDF)

S13 Fig. Calcein accumulation assay.

(A) Rate of calcein accumulation following 20 minutes of equilibration, as measured by a simple linear regression on the calcein accumulation curve. Data combines three experiments, indicated by shading of the dots while dot shape indicates clone replicates. Differences between each strain are measured by ordinary one-way ANOVA with Tukey’s multiple comparisons test. (B) Representative curves of calcein accumulation over time.

(PDF)

S14 Fig. Supporting data for polyprenyl analysis.

(A-C) Representative chromatograms and collision spectra for decaprenyl phospholipids. Red boxes on EICs indicate peaks with lipid identity confirmed by CID-MS via the highlighted fragments. For each lipid, displayed prenyl stereochemistry and position of saturated groups is assumed from literature patterns and may differ across chromatographic peaks. Blue diamonds in spectra indicate the lowest targeted m/z value for CID-MS. (D) Intensity measured in arbitrary units (counts) for measured decaprenyl phosphate species. (E) Fold change in DPP levels as measured by LCMS where each bar is a clone, and dots are culture replicates. Different shapes indicate different clone replicates corresponding with Fig 5C. Differences in DPP levels determined by Kruskal-Wallis test with Dunn’s multiple comparisons test.

(PDF)

S15 Fig. Lipoarabinomannan (LAM) measurement by western blot.

(A) Fold change in LAM levels as measured by western blot with anti-LAM CS-35 antibody. Data from two experiments indicated by dot fill while dot shapes indicate clone replicates. Differences in LAM levels determined by Kruskal-Wallis test with Dunn’s multiple comparisons test. (B) Representative western blot.

(PDF)

S16 Fig. Ordering of predicted-deleterious idsA2 mutational events.

Ordering of 158 predicted-deleterious idsA2 mutational events within Lineage 4 of the 55K dataset.

(PDF)

S1 File. Raw Gel: Raw images for LAM western blot data.

Two western blots were performed. Ladder is SeeBlue Plus2 Pre-stained Protein Standard. Primary antibody is Monoclonal Anti-Mycobacterium tuberculosis LAM antibody, Clone CS-35. Secondary antibody is IRDye 800CW Goat anti-Mouse IgG Secondary Antibody. Images were taken on a LI-COR Odyssey CLx Imager and processed and analyzed using Empiria Studio. Data from both experiments is reported in S15A Fig, and the second blot is shown as a representative image in S15B Fig.

(PDF)

S1 Table. pN/pS calculations without reversions.

pN/pS values were generated into a histogram for Figs 1A and S1.

(XLSX)

S2 Table. Lineage 4-defining variants.

(XLSX)

S3 Table. Number of nonsense and frameshift events per gene per lineage.

Lineage 4 events per kb were generated into a histogram for Fig 1D.

(XLSX)

S4 Table. Strain list.

Includes strain names and variants identified by WGS after filtering.

(XLSX)

S1 Data. Data underlying plots.

Excludes Fig 1A and S1 (S1 Table) and Fig 1D (S3 Table).

(XLSX)

Data Availability

All data are available in the main text or supplementary materials. Sequencing data can be found in the Sequence Read Archive database (Bioproject: PRJNA1356913). Phylogenetic trees, ancestral reconstruction, and variant calling can be found on Mendeley (55k dataset DOI: 10.17632/h9dd6hsvc2.1, CRyPTIC dataset DOI: 10.17632/cy4dy352kj.1). Code can be found at https://github.com/abigailmfrey/IdsA2_Paper_Code. Raw data and analysis files for lipid HPLC-MS can be found in the Zenodo repository, at https://doi.org/10.5281/zenodo.21362678.

Funding Statement

This work was supported by grants from the National Science Foundation (https://www.nsf.gov/): DGE 2140743 (AMF), National Institutes of Allergy and Infectious Diseases (https://www.niaid.nih.gov): Harvard Graduate Program in Tropical Infectious Diseases Training Grant T32 AI049928 (AMF), F32 AI174653 (PHC), U19 AI142793-01 (SMF), P01 AI143575 (SMF), U19 AI162584 (DBM), R01 AI165573 (DBM), and DP2 AI192739 (QL). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.