A data compendium of Mycobacterium tuberculosis antibiotic resistance

The Comprehensive Resistance Prediction for Tuberculosis: an International Consortium (CRyPTIC) presents here a compendium of 15,211 Mycobacterium tuberculosis global clinical isolates, all of which have undergone whole genome sequencing (WGS) and have had their minimum inhibitory concentrations to 13 antitubercular drugs measured in a single assay. It is the largest matched phenotypic and genotypic dataset for M. tuberculosis to date. Here, we provide a summary detailing the breadth of data collected, along with a description of how the isolates were collected and uniformly processed in CRyPTIC partner laboratories across 23 countries. The compendium contains 6,814 isolates resistant to at least one drug, including 2,129 samples that fully satisfy the clinical definitions of rifampicin resistant (RR), multi-drug resistant (MDR), pre-extensively drug resistant (pre-XDR) or extensively drug resistant (XDR). Accurate prediction of resistance status (sensitive/resistant) to eight antitubercular drugs by using a genetic mutation catalogue is presented along with the presence of suspected resistance-conferring mutations for isolates resistant to the newly introduced drugs bedaquiline, clofazimine, delamanid and linezolid. Finally, a case study of rifampicin mono-resistance demonstrates how this compendium could be used to advance our genetic understanding of rare resistance phenotypes. The compendium is fully open-source and it is hoped that the dataset will facilitate and inspire future research for years to come.


Introduction 30
Tuberculosis (TB) is a curable and preventable disease; 85% of those afflicted 31 can be successfully treated with a six-month regimen. Despite this, TB is the world's 32 top infectious disease killer (current SARS-CoV-2 pandemic excepted) with 10 million 33 new cases and 1.2 million deaths estimated in 2019 alone (1). Furthermore, drug 34 resistant TB (DR-TB) is a continual threat; almost half a million cases resistant to the 35 first-line drug rifampicin (RR-TB) were estimated, with three quarters of these 36 estimated to be multidrug-resistant (MDR-TB, resistant to first-line drugs isoniazid and 37 rifampicin) (1). Worryingly, only 44% of DR-TB cases were officially notified and just 38 over half of these cases were successfully treated (57%) (1). 39

40
To address these issues, the World Health Organisation (WHO) is encouraging 41 the development of better, faster and more targeted diagnostic and treatment 42 strategies through its EndTB campaign (1). Of particular interest is universal drug 43 susceptibility testing (DST). Conventionally, DST relies on lengthy (4 weeks minimum) 44 culture-based methods that require strict biosafety conditions for Mycobacterium 45 tuberculosis. The development of rapid genetics-based assays has decreased 46 diagnostic time to as little as 2 hours through the detection of specific resistance 47 conferring mutations e.g. the Cepheid Xpert® MTB/RIF test (2,3). However, assay bias 48 towards specific genic regions can result in misdiagnosis of resistance, the 49 prescription of ineffective treatment regimens and subsequent spread of multi-drug 50 resistant disease, as seen during an MDR outbreak in Eswatini (4-6). Furthermore, 51 detection of rifampicin resistance is used to infer MDR-TB epidemiologically as 52 rifampicin resistance tends to coincide with resistance to isoniazid (7). While this 53 modus operandi is successful at pragmatically identifying potential MDR cases quickly 54 and incubated for 14 days at 37 o C. Quality control runs were performed periodically 130 using M. tuberculosis H37Rv ATCC 27294, which is sensitive to all drugs on the plates. 131 132

Minimum Inhibitory Concentration (MIC) measurements 133
Minimum inhibitory concentrations (MICs) for each drug were read after 134 incubation for 14 days by a laboratory scientist using a Thermo Fisher Sensititre™ 135 Vizion™ digital MIC viewing system (10). The Vizion apparatus was also used to take 136 a high contrast photograph of the plate with a white background, from which the MIC 137 was measured again using the Automated Mycobacterial Growth Detection Algorithm 138 taken. MICs were then classified as high (at least two methods concur on the MIC), 147 medium (either a scientist recorded a MIC measurement using Vizion but did not store 148 the plate picture, or Vizion and AMyGDA disagree and there is no BashTheBug 149 measurement), or low (all three methods disagree) quality. 150 To ensure adequate data coverage for this study, we took the MIC from the 151 Vizion reading provided by the trained laboratory scientist if it was annotated as having 152 medium or low quality. 153 154

Binary phenotype classification 155
Binary phenotypes (resistant/susceptible) were assigned from the MICs by 156 applying epidemiological cut-off values (11); samples with MICs at or below the 157 ECOFF are, by definition, wild-type and hence assigned to be susceptible to the drug 158 in question (11). Samples with MICs above the ECOFF are therefore classified as 159 resistant (Fig. S1, Table S1). Please see (11) for the body of work supporting the use 160 of the ECOFF relative to the compendium isolates and supplemental Table S1 for the 161 ECOFFs for each drug tested.  based partially on these isolates (26)). 208 The resulting VCF file for each isolate (see "Genomic data processing and variant 209 calling" section above) was compared to the genetic catalogue to determine the 210 presence or absence of resistance-associated mutations for eight drugs: rifampicin, 211 isoniazid, ethambutol, levofloxacin, moxifloxacin, amikacin, kanamycin and 212 ethionamide. We did not apply the approach used in (7) to make a prediction if a novel 213 mutation was detected in a known resistance gene, as we simply wanted to measure 214 how well a pre-CRyPTIC catalogue could predict resistance in the compendium. 215 These results (found in PREDICTIONS.csv, see "Data availability" section for access) 216 were then compared to the binary phenotypes (see "Binary phenotype classification" 217 section for how these were defined) with the following metrics calculated: TP: the 218 number of phenotypically resistant samples are that correctly identified as resistant 219 ("true positives"); FP, the number of phenotypically susceptible samples that are 220 falsely identified as resistant ("false positives"); TN, the number of phenotypically 221 susceptible samples that are correctly identified as susceptible ("true negatives"); FN, 222 the number of phenotypically resistant samples that are incorrectly identified as 223 susceptible ("false negative"); VME, very major error rate (false-negative rate), 0-1; 224 ME, major error rate (false-positive rate), 0-1; PPV, positive predictive value, 0-1; NPV, 225 This directory contains the data used for multiple CRyPTIC project publications 262 referenced throughout this manuscript. As stated above, each project has taken 263 slightly different subsets of this data as documented in those papers. For example, 264 see how tables such as "MUTATIONS.csv" and "GENOTYPES.csv" were used and 265 filtered, (along with others) in this study to obtain the reuse file 266 "CRyPTIC_reuse_table_20221019.csv" in Figure 1. Again, for optimal use of 267 CRyPTIC data in your own project, please refer to 268 "CRyPTIC_reuse_table_20221019.csv" in the "reuse" directory. All data for this study 269 were analysed and visualised using either R or python3 libraries and packages. See

15,211 M. tuberculosis clinical isolates 281
The CRyPTIC compendium contains 15,211 isolates for which both genomic 282 and phenotypic data was collected by 23 of the partner countries across the continents 283 of Asia, Africa, South America and Europe. An overview of the processing of the 284 isolates is presented in Figure 1, and for a full description please see Methods.

312
Where the origin of an isolate was not known, the collection site identity was assigned (this  show the proportion of isolates amongst the different lineages (Table S2) and sub-322 lineages (Table S3)  MICs to 13 antitubercular drugs, regular quality assurance checks detected problems 331 with plate inoculation and reading in two laboratories. Therefore, 2,922 isolates were 332 removed from the dataset, leaving a total of 12,289 isolates with matched phenotypic 333 and genotypic data for further analysis (Fig.1)

343
"medium" quality (either Vizion and AMyGDA disagree, or the scientist recorded a MIC 344 measurement using Vizion but did not store the plate picture) or "low" quality (all three MIC 345 measurements methods disagree) phenotype classifications as described in Methods. Drug

Resistance classification and distribution 351
Unsurprisingly, given its size and bias toward the collection of resistant isolates, 352 resistance to each of the 13 drugs is represented within the compendium (Fig. 3a). 353 The drugs with the highest percentage of resistant isolates are the first line drugs 354 isoniazid and rifampicin (49.0% and 38.7% respectively). Of the second line drugs, 355 levofloxacin had the highest proportion of resistant isolates in the dataset (17.6%) and 356 amikacin the lowest (7.3%). A low proportion of isolates were resistant to the NRDs, 357 bedaquiline (0.9%), clofazimine (4.4%), delamanid (1.6%) and linezolid (1.3%). 358

359
Of the 12,289 isolates with matched phenotypic and genotypic data, 6,814 360 (55.4%) were resistant to at least one drug (Fig. 3b). For the purpose of describing 361 five broader resistance categories present in the dataset, we assumed that all MICs 362 that could not be read had susceptible phenotypes. These five resistance categories 363 comprise: isoniazid and rifampicin susceptible with resistance to another 364 antitubercular drug, isoniazid resistant but rifampicin susceptible, RR/MDR, pre-XDR 365 (RR/MDR + fluoroquinolone resistance), and XDR (RR/MDR + fluoroquinolone 366 resistance + resistance to a group A agent: bedaquiline or linezolid)). Consequently, 367 the calculated prevalence of MDR, XDR etc. in the dataset (Fig. 3) are likely under-368 estimates. Of the isolates resistant to one or more drugs, 22.8% were resistant to 369 isoniazid and not rifampicin, 68.8% were either RR or MDR and 8.4% were resistant 370 to at least one antitubercular drug, but not isoniazid or rifampicin (Fig. 3b). Of the 371 RR/MDR isolates, 38.8% were pre-XDR and 3.0% were XDR. Two of the XDR isolates 372 returned a resistant phenotype to all 13 of the drugs assayed (Table S5) and therefore 373 could be reasonably described as totally drug resistant (TDR). One such isolate 374 belonged to L4 and was contributed by South Africa, and the other belonged to L2 with 375 an unknown country of origin contributed by Sweden. 376

377
The proportion of drug susceptible isolates collected differed between 378 countries, as not all laboratories oversampled for resistance (Fig. 3c). In countries that 379 contributed more than 100 resistant isolates, each of the broad phenotypic resistance 380 categories in Fig. 3b were seen except for Peru, Vietnam and Nepal which did not 381 contribute any XDR isolates (Fig. 3c). Vietnam and Brazil sampled a high proportion 382 of non-MDR/RR resistant phenotypes; 73.9% and 55.1% of resistant isolates 383 contributed by these countries, respectively, were neither MDR nor RR. For Nepal and 384 India, an especially high proportion of the MDR/RR isolates contributed were 385 fluoroquinolone resistant (92.9% and 69.8% respectively), which has been previously 386 observed for this geographical region (32,33). 387

388
Of the 6814 resistant isolates, 256 were from L1, 2886 were from L2, 625 were 389 from L3 and 3043 were from L4. All five broader categories of resistance were 390 represented in the four major M. tuberculosis lineages (Fig. 3d). We note that the 391 relative proportions of resistance categories will have been influenced by the different 392 local sampling approaches since lineage distributions are typically geographically 393 distinct (Fig. S2). Bearing this in mind, we observe that in the compendium, L3 isolates 394 contained the most MDR/RR isolates as a proportion of resistant isolates (77.6%), L2 395 isolates contained the most pre-XDR isolates as a proportion of MDR/RR isolates 396 (54.2%) and L2 contained the most XDR isolates as a proportion of MDR/RR isolates 397

Co-occurrence of drug resistance amongst the CRyPTIC isolates 414
As we measured MICs to 13 drugs in parallel, we can ask whether, and how 415 often, co-occurrence of drug resistance occurs amongst the isolates. We found 416 isolates with all possible two-drug resistant combinations in this dataset (Fig. 4a, Table  417 S6). With the exception of correlations between drugs in the same class (rifabutin vs 418 rifampicin, moxifloxacin vs levofloxacin), Isoniazid resistance was the most strongly 419 associated with resistance to each of the other drugs. Resistance to any of the drugs 420 was also strongly associated with resistance to rifampicin. Of the second line drugs, 421 levofloxacin and moxifloxacin were more commonly seen as a second resistant 422 phenotype than the injectable drugs kanamycin and amikacin. were most likely to also be resistant to isoniazid, followed by rifampicin and rifabutin. 438 The NRDs were less commonly seen as a second resistance phenotype and the 439 smallest proportional resistance combinations involved the NRDs (e.g. 1.5% of 440 isoniazid resistant isolates were bedaquiline resistant). Within the NRDs however, co-441 occurrence of resistance was proportionally higher; bedaquiline, linezolid and 442 delamanid resistance was commonly seen with clofazimine resistance (52.4%, 34.2% 443 and 26.3% of isolates having co-resistance with clofazimine respectively). 444 445

Additional antibiotic resistance in isolates with non-MDR or MDR phenotypic 446 backgrounds 447
To further investigate drug resistance patterns amongst the isolates, we 448 examined in more detail the correlation structure of phenotypes by conditioning on 449 different resistance backgrounds including isoniazid and rifampicin susceptible, 450 isoniazid resistant and rifampicin susceptible, rifampicin resistant and isoniazid 451 susceptible, MDR, pre-XDR and XDR ( Fig. 4b-f). We found that a greater proportion 452 of isolates that were susceptible to isoniazid and rifampicin were resistant to the 453 second line drugs levofloxacin (24.1%), kanamycin (18.1%), moxifloxacin (13.7%), 454 and amikacin (8.9%) than the first line drug ethambutol (3.8%) (Fig. 4b). The proportion 455 of isolates resistant to clofazimine or levofloxacin was particularly high (32.9% and 456 24.1%, respectively), and more isolates were resistant to these two drugs than 457 ethambutol in an isoniazid resistant and rifampicin susceptible background but not in 458 MDR/RR isolates (Fig. 4c-f). 459 460 MDR/RR isolates were most commonly resistant (excluding rifabutin) to the first 461 line drug ethambutol (46.3%), closely followed by levofloxacin (41.4%). As expected, 462 the proportion of fluoroquinolone resistance was higher in MDR/RR isolates than non-463 MDR isolates (37) and we found a greater proportion of isolates were resistant to 464 levofloxacin than moxifloxacin, a pattern seen in all other backgrounds (Fig. 4c-f) For isolates with an XDR phenotype, a higher proportion were resistant to 470 linezolid than bedaquiline (66.7% compared to 44.6%) and 11.3% of XDR isolates 471 were resistant to both bedaquiline and linezolid (Fig. 4f). XDR isolates were also 472 resistant to the other NRDs, clofazimine (41.3%) and delamanid (18.8%). In non-XDR 473 backgrounds the most common NRD resistance seen was clofazimine (Fig. 4b-e).

Genetic-based predictions of resistance 491
To establish a baseline measure of how well resistance and susceptibility could 492 be predicted based on the state of the art prior to the CRyPTIC project, we compared 493 genetic-based predictions of susceptibility and resistance to the binary phenotypes 494 derived from MICs for eight drugs and for all isolates in this compendium (Table 2). 495 Since these data were not collected prospectively or randomly, but indeed are 496 enriched for resistance, the calculated error rates are not representative of how well 497 such a method would perform in routine clinical use. The results were broadly in line 498 with prior measurements on a smaller (independent) set (23). The hybrid catalogue 499 does not make predictions for rifabutin, linezolid, bedaquiline, delamanid or 500 clofazimine; indeed, this is one of the main aims of the consortium and new catalogues 501 published by CRyPTIC and the WHO will begin to address this shortcoming (Fig. 5)

Resistance to new and re-purposed drugs 556
As previously stated, relatively few isolates are resistant to the NRDs, 557 bedaquiline (n = 109), clofazimine (n = 525), delamanid (n = 186) and linezolid (n = 558 156). South Africa contributed the greatest number of isolates resistant to bedaquiline, 559 clofazimine and linezolid (Fig. 5a), while China and India contributed the most isolates 560 resistant to delamanid. Since the collection protocol differed between laboratories it is 561 not possible to infer any differences in the relative prevalence of resistance to the 562 NRDs in these countries. The results of a survey of all non-synonymous mutations in 563 genes known or suspected to be involved in resistance to these four drugs (e.g.  For this case study, we defined RMR as any isolate that is rifampicin resistant 597 and isoniazid susceptible, and discounted isolates with no definite phenotype for either 598 drug. Of the 4,655 rifampicin resistant isolates in the compendium that also had a 599 phenotype for isoniazid, 302 (6.5%) were RMR. These isolates were contributed by 600 12 different countries, and we found South African and Nigerian contributions had a 601 significantly higher proportion of RMR isolates than that of the total dataset at 17.5% 602 (p <0.00001) and 27.3% (p = 0.00534) respectively (Fig. 6a) compared with 6.5% for 603 the total dataset. We note that these proportions are influenced by sampling strategies 604 but the higher contribution of RMR isolates from South Africa is consistent with 605 previous studies (43). 606 607

Rifampicin mono-resistance is incorrectly predicted by current diagnostics 609
A widely used, WHO-endorsed diagnostic tool, the Xpert® MTB/RIF assay, 610 uses a proxy whereby any SNP detected in the "rifampicin-resistance determining 611 region" (RRDR) of rpoB results in a prediction of MDR. However, the suitability of the 612 proxy is dependent upon prevalence of RMR in the population (43). We tested the 613 reliability of this on the 4,655 rifampicin resistant isolates in our dataset that had a 614 phenotype for isoniazid (Fig. 6b).

There are genetic differences between rifampicin mono-resistant and 633 multidrug resistant isolates 634
We have analysed our matched phenotypic and genotypic data to examine 635 whether there were any differences in the genetic determinants of rifampicin 636 resistance between RMR and MDR isolates as was seen in a recent study of South 637 African isolates (46). The proportion of RMR isolates with no rpoB mutation (5.3%, Fig.  638 6b) was significantly higher than that of MDR isolates (1.8%, p <0.00001). This 639 suggests that non-target-mediated resistance mechanisms, such as upregulation of 640 rifampicin specific efflux pumps, could be more influential in providing protection 641 against rifampicin in RMR isolates than in MDR isolates. 642

643
The majority of RMR and MDR isolates contained one or more SNPs in rpoB, 644 with the most having at least one mutation in the RRDR. To date, several non-645 synonymous RRDR mutations have been found in RMR M. tuberculosis isolates, 646 including the resistance conferring mutations S450L, H445D and D435Y, which are 647 also seen in MDR isolates (47,48). For both RMR and MDR isolates in this dataset, 648 the most common rpoB RRDR mutation seen was S450L (63.6% and 41.1% of 649 isolates respectively, Fig. 6c). Five mutations were present in RMR isolates that were 650 not seen in MDR isolates: S428G, S441A, S441V, S450M and S450Q, however these 651 were seen at low prevalence (< 2%) of RMR isolates. We found more RMR isolates 652 had His445 mutated than MDR isolates (27.8% of RMR and 9.5% of MDR, p 653 <0.00001), and mutations at Ser450 and Asp435 were more prevalent in MDR isolates 654 than RMR isolates (43.7% of RMR and 65.8% of MDR (p <0.00001), and 9.3% of RMR 655 and 15.5% of MDR (p = 0.00328) respectively). 656

657
In RMR isolates we observed 27 different rpoB mutations that fall outside the 658 RRDR; 11 were found in RMR but not MDR isolates and all were seen at < 2% 659 prevalence (Fig. S6). The most common non-RRDR mutation in both MDR and RMR 660 isolates was a cytosine to thymine mutation 61 bases upstream of the rpoB start codon 661 (10.1% and 8.6% of isolates respectively). The resistance conferring mutations, rpoB 662 I491F, V695L and V170F, were seen at low proportions (< 2% of isolates) with no 663 significant difference between MDR and RMR isolates.  resulting in improved quality of care for patients. However, relying solely on these 698 diagnostic methods has several drawbacks. Aside from the Xpert® MTB/RIF assay 699 potentially increasing false positive MDR diagnoses as discussed earlier in the RMR 700 case study, the assay assumes isoniazid resistance upon detection of rifampicin 701 resistance. Thus, less is known about the prevalence of mono isoniazid resistance or 702 'true' cases of MDR (confirmed rifampicin and isoniazid resistance) (1) and with large 703 datasets such as this compendium, we can further investigate these important and 704 rarer clinical phenotypes (like that of RMR in our case study). Another example of a 705 rarer phenotype is that of isoniazid-resistant and rifampicin-susceptible (Hr-TB) 706 isolates; a greater number of these were contributed by CRyPTIC countries than RMR 707 isolates (n = 1470 versus n = 302), a pattern also recently observed in a global 708 prevalence study (49). A modified 6-month treatment regimen is now recommended 709 Encouragingly, CRyPTIC isolates with a Hr-TB background exhibited relatively low 713 levels of resistance to other antitubercular drugs, including those in the augmented 714 regimen (Fig. 4c). However, without appropriate tools to assess and survey this, we 715 will continue to misdiagnose and infectively treat these clinical cases. In 2018, 716 CRyPTIC and the 100,000 Genomes project demonstrated that genotypic prediction 717 from WGS correlates well with culture-based phenotype for first-line drugs, which is 718 reflected in our summary of the genetic catalogue applied to this dataset (Table 3)

(7). 719
While predictions can be made to a high level of sensitivity and specificity, there is still 720 more to learn, as exemplified by the isolates in the compendium that despite being 721 resistant to rifampicin and isoniazid could not be described genetically (Table 2). This fluoroquinolones than second line injectable drugs (Fig. 4a). This could be due to more 730 widespread use of fluoroquinolones as well as their ease of administration and hence 731 them being recommended over injectables for longer MDR treatment regimens (1). 732 Concerningly, we found that resistance to levofloxacin and moxifloxacin, and 733 kanamycin and amikacin, were more common than resistance to the mycobacterial 734 specific drug ethambutol in an isoniazid and rifampicin susceptible background (Fig.  735  supports this recommendation as we saw more resistance to kanamycin than amikacin 748 in all phenotypic backgrounds. For fluoroquinolones, more isolates were resistant to 749 levofloxacin than moxifloxacin in all phenotypic backgrounds suggesting moxifloxacin 750 may by the most appropriate fluoroquinolone to recommend, although we note this 751 conclusion is critically dependent on the validity of the cutoff, here an ECOFF, used to 752 infer resistance. However, the amenability of drugs to catalogue-based genetic 753 diagnostics is also an important consideration, and our data suggest levofloxacin 754 resistance could be predicted more reliably than moxifloxacin, with fewer false 755 positives predicted (Table 2). Testing for fluoroquinolone resistance using molecular 756 diagnostic tests remains limited. Global data from the past 15 years suggests that the 757 proportion of MDR/RR TB cases resistant to fluoroquinolones sits at around 20%, with 758 these cases primarily found in regions of high MDR-TB burden (1). While recently 759 approved tools, such as the Cepheid Xpert® MTB/XDR cartridge, will permit both 760 isoniazid and fluoroquinolone testing to be increased, the same pitfalls are to be 761 encountered regarding targeted diagnostic assays (52). In contrast, the genetic survey 762 in this study demonstrates the potential of WGS for genetic prediction of resistance to 763 second-line drugs and studies within the consortium to investigate this are underway. 764

765
The CRyPTIC compendium has facilitated the first global survey of resistance 766 to NRDs. Reassuringly, prevalence of resistance to the NRDs was substantially lower 767 than for first-and second-line agents in the dataset (Fig. 3a), and resistance to the 768 new drugs bedaquiline and delamanid was less common than the repurposed drugs 769 clofazimine and linezolid in an MDR/RR background (Fig. 4c). However, the presence 770 of higher levels of delamanid and clofazimine resistance than ethambutol resistance 771 in the isoniazid and rifampicin susceptible background does suggest some pre-existing 772 propensity towards NRD resistance (Fig. 4b). 773 774 Co-resistance between NRDs was seen in isolates in the compendium, the 775 most common being isolates resistant to both bedaquiline and clofazimine. This link is 776 well documented and has been attributed to shared resistance mechanisms such as 777 non-synonymous mutations in Rv0678 which were found in both clofazimine and 778 bedaquiline resistant isolates in the compendium (42) (Fig. 5b,c). Increased 779 clofazimine use could further increase the prevalence of M. tuberculosis isolates with 780 clofazimine and bedaquiline co-resistance, limiting MDR treatment options including 781 using bedaquiline as the backbone of a shorter MDR regimen (53). Therefore, 782 proposed usage of clofazimine for other infectious diseases should be carefully 783

considered. 784 785
The WHO recommends against the use of bedaquiline and delamanid in 786 combination to prevent the development of co-resistance, which could occur relatively 787 quickly (54); the rate of spontaneous evolution of delamanid resistance in vitro has 788 been shown to be comparable to that of isoniazid, and likewise bedaquiline resistance 789 arises at a comparable rate to rifampicin resistance (55). In this compendium, 12.9% 790 of bedaquiline resistant isolates were resistant to delamanid and 7.1% of delamanid 791 resistant isolates were resistant to bedaquiline. Several scenarios could account for 792 this, including the presence of shared resistance mechanisms. For example, as 793 bedaquiline targets energy metabolism within the cell, changes to cope with 794 energy/nutrient imbalances upon the acquisition of resistance-associated ATPase 795 pump mutations may result in cross resistance to delamanid in a yet unknown or 796 unexplored mechanism (12). it is imperative that genetic determinants of resistance 797 are fully explored for the NRDs, as these are our current treatments of last resort, with 798 special attention given to those mechanisms that could be shared with other agents.