Background

Mycobacterium tuberculosis is the causative agent of tuberculosis (TB), which is most often spread person-to-person via cough aerosols. Although many individuals who are exposed to M. tuberculosis never develop active disease, the World Health Organization (WHO) estimated 10 million new cases of active TB and 1.3 million deaths in 2017 alone [1].

Since its initial documentation in the 1940s [2], drug-resistant TB has threatened public health control efforts. In 2016, there were an estimated 490,000 new cases of multidrug-resistant (MDR) TB, which is defined by phenotypic resistance to both isoniazid and rifampicin [3]. Approximately 10% of MDR-TB cases globally can be classified as extensively drug-resistant (XDR), indicating that there is concomitant resistance to quinolones (such as the fluoroquinolones, levofloxacin, and moxifloxacin) and to a second-line injectable agent (amikacin, kanamycin, or capreomycin) [3]. As expected, drug-resistance patterns predict treatment outcome; in 2015, TB treatment success overall was 83%, whereas the success rate was 54% for MDR-TB or rifampicin-resistant-TB (RR-TB) and only 30% for XDR-TB [4].

Culture-based techniques remain the current reference standard for both diagnosis and drug-susceptibility testing of TB, but these processes are time-consuming and require specialized laboratory capacity. More recently, the use of rapid molecular tests for the diagnosis of TB has increased globally, particularly the use of Xpert MTB/RIF (Cepheid, Sunnyvale, CA), a PCR-based assay that simultaneously detects the presence of M. tuberculosis and resistance to rifampicin.

Current recommendations for the treatment of drug-susceptible TB include a 6-month course of a multi-drug regimen of rifampicin, isoniazid, pyrazinamide, and ethambutol. Historically, treatment of MDR- or XDR-TB involved the long-term use of second-line drugs, including injectable agents [5]. More recently, the MDR-TB treatment landscape has changed with the introduction of multiple novel second-line drugs that can be administered orally (Table 1). In 2012, bedaquiline, a diarylquinolone, became the first TB drug from a novel drug class to receive US Food and Drug Administration (FDA) approval in over 40 years [48, 49] (Table 1). Another oral agent, delamanid, a nitro-dihydro-imidazooxazole derivative, has also shown promise for TB treatment [50, 51].

Table 1 Antitubercular drug-resistance mechanismsa
Full size table

In 2018, the WHO published updated treatment guidelines for MDR/RR-TB [47], recommending fully oral MDR regimens for many patient groups. Recommended treatment strategies include both shorter, standardized MDR regimens (for 9 to 12 months) and longer, individualized treatment regimens (for 18 to 20 months). The updated guidelines group antitubercular drugs on the basis of how they should be combined to create individualized, longer MDR-TB regimens [47] (Table 1).

Despite advances in both diagnostics and therapeutics for TB, challenges remain. Obstacles for rapid M. tuberculosis diagnosis include: (i) the imperfect sensitivity of molecular tests for the detection of this pathogen, particularly in the case of paucibacillary TB (where there is a lower bacterial burden); (ii) lack of comprehensive molecular assays due to incomplete knowledge of all resistance mutations in TB; and (iii) technical limitations to the numbers of mutations that can be included on diagnostic molecular platforms. Furthermore, the rollout of rapid diagnostic platforms to low-resource settings has been a challenge. Remaining treatment challenges include: prolonged treatment courses, leading to greater drug exposure, toxicity, and non-compliance; unacceptable side-effect profiles; logistics of drug access; and re-infection [52].

The dawning of the new age of genome sequencing began to revolutionize our approach to human diseases, including TB. In 1998, Cole et al. [53] reported the complete genome sequence of the M. tuberculosis reference strain H37Rv, which was approximately 4.41 million base pairs in length and encoded approximately 4000 genes. The first sequencing of a clinical reference strain, CDC1551, quickly followed [54]. An accompanying editorial optimistically stated: “After several decades in the slow lane of classical microbiology, M. tuberculosis is once again at the cutting edge of science” [55]. However, even at the time of these breakthroughs, there was recognition that translating these genomic data into clinical benefit would prove challenging [55]. Despite these challenges, it is clear, more than 20 years later, that M. tuberculosis genomic data have been remarkably useful in improving our understanding of how drug-resistant TB evolves and spreads and in helping to inform diagnostics and therapies.

In this review, we discuss the molecular epidemiologic and diagnostic advances made by sequencing M. tuberculosis, with a focus on drug-resistant TB. We do not review the practice of whole-genome sequencing (WGS) of M. tuberculosis as this has been reviewed recently [56]. Key findings that are discussed include the use of WGS to identify drug-resistance determinants in M. tuberculosis and to elucidate the evolution and spread of drug-resistant organisms, and the clinical applications of this technology (Table 2).

Table 2 Spotlight on whole-genome sequencing studies of drug-resistant M. tuberculosis
Full size table

Identifying M. tuberculosis drug-resistance determinants

Drug resistance in M. tuberculosis is the result of chromosomal mutations in existing genes that are passed along through vertical descent, that is, passed from mother to daughter cells. Unlike many other bacterial pathogens, M. tuberculosis rarely recombines via lateral exchange of DNA [83] and also lacks plasmids. Many of the resistance determinants were discovered before the sequencing of the M. tuberculosis genome was completed. By 1998, resistance mechanisms had already been discovered for classical first- and second-line TB drugs including isoniazid (alterations in genes katG and inhA); rifampicin (in rpoB); streptomycin (in rrs and rpsL); pyrazinamide (in pncA); ethambutol (in embB); quinolones (in gyrA); and kanamycin (in rrs) (reviewed in Ramaswamy and Musser [84]) (Table 1). However, the targeted amplification and sequencing of known or suspected resistance genes revealed that these mechanisms were insufficient to explain all phenotypic resistance [85, 86], and resistance mechanisms for several newer drugs—including pretomanid, bedaquiline, and delamanid—were discovered over the next eight years during a period when WGS was becoming routine. Together, in the past 20 years, WGS-based approaches, focused on both laboratory-derived and naturally circulating populations of drug-resistant M. tuberculosis, have provided a more complete account of the genomic features that cause treatment resistance, enabling the identification of novel resistance mechanisms for existing drugs, and the determination of the mechanisms of action of newly discovered drugs.

Identifying resistance determinants in laboratory-derived mutants

Drug-resistant mutants can be derived in vitro by growing drug-susceptible M. tuberculosis strains in drug-containing media, and selecting for mutants that are able to grow in the presence of the drug. Sequencing laboratory-derived resistant mutants has played a critical role in identifying both the mechanism of action of new TB drug classes, including diarylquinolines (for example, bedaquiline) [19] and nitroimidazopyrans (for example, PA-824) [19, 29], and rare resistance mechanisms for established antitubercular drugs, including ethambutol [15], pyrazinamide [13], carbapenems [30], cycloserine [87], clofazimine, and bedaquiline [20]. For example, WGS of laboratory mutants identified drug efflux as a mechanism of resistance to clofazimine and bedaquiline [20,21,22], and this approach continues to be a mainstay for identifying the mechanism of action of compounds that are in development for TB [88].

Although laboratory-derived mutants are helpful in elucidating novel resistance mechanisms, mutations that have evolved in laboratory settings do not always match those in clinical isolates of drug-resistant M. tuberculosis [89, 90], for reasons that are largely unknown. Studies by Ford et al. [91, 92] suggested that these mismatches could not be explained by differences in the mutation rate in these settings, because the in vitro mutation rate of M. tuberculosis correlates well with the in vivo mutation rate in both humans and in a macaque model. Differences in the relative fitness of specific mutants grown in in vitro compared to in vivo conditions are probably responsible for these mismatches, but more work is needed. Regardless of the reason, if the goal is to identify a full complement of resistance mutations on which to base molecular diagnostics, isolates from clinical collections must be studied as these bacteria have evolved their resistance within the host.

Quantifying and identifying resistance determinants in clinical strains

Among the larger studies exploring resistance in natural populations, Walker et al. [58] analyzed the genomes of 3651 drug-resistant and -susceptible M. tuberculosis isolates for associations between phenotypic resistance to eight first- and second-line drugs, and then predicted genotypic resistance on the basis of a compiled catalog of 232 resistance mutations in 23 candidate resistance genes. Resistance to most drugs could be predicted accurately, with a mean sensitivity of 92% and specificity of 98%, suggesting that the majority of resistance—particularly for first-line drugs—is explained by known mechanisms and mutations (Table 1). Numerous other studies have found similar results using smaller datasets [7, 25, 57, 69, 93, 94]. This result was echoed in a more recent study by the Comprehensive Resistance Prediction for Tuberculosis (CRYPTIC) Consortium and the 100,000 Genomes Project that focused solely on first-line drugs, which included analysis of 10,209 globally diverse M. tuberculosis isolate genomes against a database of mutations identified in a literature search [60]. Notably, predictions for mutations that are associated with resistance to pyrazinamide were greatly improved over earlier predictions; this study achieved 91.3% sensitivity in predicting resistance to this drug compared to 57% sensitivity in Walker et al. [58]. Although the news has been encouraging with respect to completing the catalog of mutations that cause resistance to first-line drugs, few studies have attempted to predict resistance to second-line drugs [95]. Some of these drugs, such as D-cycloserine, pyrazinamide, and para-aminosalicylic acid (PAS), are more difficult to assay because they have been reported to have variable drug phenotypes in clinical microbiology laboratories [96] (discussed later).

To fill gaps in the catalog of drug-resistance mechanisms, genome-wide association study (GWAS) approaches, originally designed for use on human genomic data, have been adapted for non-recombining microbes such as M. tuberculosis and used to predict novel resistance mechanisms [97, 98] (Table 3). The majority of GWAS predictions remain experimentally unverified, but several new resistance-associated genotypes have been validated. Farhat et al. [7] sequenced 116 M. tuberculosis genomes and developed a phylogenetic convergence test, ‘PhyC’, to identify resistance associations. They identified a mutation in ponA1 (c.1095G>T) and showed that it conferred a minimum inhibitory concentration (MIC) to rifampicin that was twofold higher than that of wild-type bacteria. Zhang et al. [57] sequenced 161 genomes from China and searched for genes that appeared to be under positive selection and more frequently mutated in drug-resistant isolates. Resistance-associated polymorphisms in two intergenic regions upstream of the known resistance genes thyA-Rv2765 and thyX-hsdS.1 were shown to cause increased gene expression of a lacZ construct in Mycobacterium smegmatis, suggesting that these mutations may mediate PAS resistance through the overexpression of downstream genes.

Table 3 Publicly available software packages implementing microbial GWAS methods for identifying drug-resistance-associated genetic variants in bacteria
Full size table

Desjardins et al. [25] used a combination of the correlated evolution test [104] (to test for correlated evolution of genotype and phenotype) and a simple GWAS framework to search for novel drug-resistance mechanisms in 498 genomes from South Africa and China. Of note, they combined all variants within each gene that were predicted to inactivate gene function, and used these combinations as the input into the association test to increase statistical power in the detection of genomic features that are associated with resistance. They found that loss-of-function mutations in ald (Rv2780), which is predicted to encode an alanine dehydrogenase, correlated with unexplained resistance [25]. They also confirmed experimentally that these mutations conferred increased resistance of laboratory and clinical isolates to D-cycloserine [25], a key drug in MDR- and XDR-TB regimens that has severe psychiatric and central nervous system toxicities.

Hicks et al. [105] used the algorithm phyOverlap to perform a GWAS on 549 clinical isolates from China, in which they identified mutations that disproportionately occurred in isoniazid-resistant isolates. In addition to known resistance and compensatory mutations for first- and second-line drugs, they identified an association with prpR (Rv1129c). They then went on to characterize prpR as a transcriptional regulator of propionate metabolism which, instead of drug resistance, confers tolerance to multiple antibiotics in a macrophage model of infection.

In one of the largest GWAS published to date, Coll et al. [59] combined PhyC with a GWAS approach within a mixed-regression framework to detect determinants of resistance to 14 drugs in a large dataset of 6465 global M. tuberculosis clinical isolates. Although no functional experiments were performed to validate the predictions, new resistance-associated mutations were identified, including new codons in ethA (a gene known to activate ethionamide, which is a prodrug) that are associated with ethionamide resistance, and mutations in the thyX promoter associated with PAS resistance. Mutations in the promoter of thyX have been previously shown to upregulate thyX [43, 57, 106].

Predicting susceptibility and drug resistance in M. tuberculosis

As the list of suspected resistance determinants grows, there has been a need to establish well-curated databases of drug-resistance single nucleotide polymorphisms (SNPs) [107]. Initially, SNP databases, including TBDB [108] and PATRIC [109], were created to bring together genome annotation data and other functional data. Unfortunately, some of the pioneering databases of drug-resistance-associated mutations in M. tuberculosis, including TBDReamDB [110], have not been maintained to include emerging data.

Software and web-based tools have also been developed to enable the community to infer TB drug resistance from WGS data. These tools include CASTB [111], KVarQ [112], MyKrobe Predictor TB [113], PhyResSE [114], TBProfiler [115], and TGS-TB [116]. Studies have compared the sensitivity and specificity of these tools in predicting drug resistance [117,118,119], and have found that they tend to perform quite well for first-line drugs but less optimally for second-line drugs. In addition to tools, there have been improvements to databases, including the creation of the Relational Sequencing TB Database Platform (ReSeqTB) [120, 121] and efforts from the CRyPTIC Consortium [122], which seeks to develop a curated database of clinically relevant drug-resistance mutations.

Continued refinement of these drug-resistance databases and prediction tools is necessary. Miotto et al. [123] performed a systematic review in which they assigned a confidence level to associations of individual and groups of mutations with phenotypic drug resistance. Importantly, they identified that certain mutations that are included in current commercial diagnostic tests, including eis c-2a, do not have a convincing association with drug resistance. Input from ongoing large sequencing projects will be needed to optimize the inference of resistance phenotypes from sequence data, especially for mutations that are present at low frequency in natural populations.

Challenges in uncovering the remaining resistance elements

Although WGS approaches have been successful in identifying resistance mechanisms, there are computational and experimental challenges that hamper efforts to complete the catalog of TB drug resistance. For example, for non-recombining organisms such as M. tuberculosis, interpretation of GWAS output can be complicated because non-causal variation can be tightly linked to causal variation [124]. Furthermore, as a result of frequent multidrug resistance, resistance mutations for one drug can appear to be highly associated with phenotypic resistance to multiple drugs [25], and confirmatory wet lab studies, which are non-trivial in M. tuberculosis, are often necessary to identify causal resistance mutations correctly. In addition, genotype–phenotype associations are largely dependent on accurate phylogenies, and phylogenetic reconstruction can be challenging in M. tuberculosis because of its slow rate of evolution [92, 125,126,127,128], which gives rise to relatively few SNPs in clinical isolates.

When defining phenotypic resistance, different studies often use different drug concentration cutoffs and test in different media, complicating the meta-analysis of multiple datasets. In addition, phenotypic resistance testing of some antitubercular drugs, including pyrazinamide and D-cycloserine, is notoriously challenging and unreliable [129], introducing phenotypic inaccuracies that can confound analyses. Furthermore, the dichotomous classification of phenotypic resistance as ‘resistant’ or ‘susceptible’ will fail to identify drug-resistance mutations that result only in minimal increases in MIC, and there is emerging evidence that such mutations may be clinically relevant. TB relapse following treatment has been found to occur more commonly in individuals who harbored M. tuberculosis isolates that were susceptible to, yet had minimally increased MIC values for, either isoniazid or rifampicin [130]. Future study designs that address phenotypic resistance as a spectrum, rather than a binary value, will be needed to identify such mutations.

Heteroresistance, defined as the coexistence of pathogen populations that have differing nucleotides at a specific drug-resistance locus [131], can also confound genotype–phenotype comparisons [132,133,134]. A bacterial culture in which only a small fraction of the population is resistant can appear to be resistant when tested on media containing a drug, yet when grown on drug-free media for genome sequencing, the sensitive fraction can dominate, resulting in a genotypic prediction of sensitivity [132]. The problem of heteroresistance seems to be particularly common with fluoroquinolone resistance [135].

Last, innate characteristics of the M. tuberculosis genome—namely, highly repetitive DNA sequences and the high guanine-cytosine (GC) content of the genome (65.6%) [53]—present technical difficulties for both WGS and bioinformatic analyses. GC-rich regions can be troublesome for library PCR amplification and sequencing, and reads that represent highly repetitive regions of the genome can confound alignments by mapping to multiple regions of the genome and hampering accurate de novo assemblies. In addition, approximately 10% of the coding regions in M. tuberculosis are dedicated to two repetitive protein families that are unique to mycobacteria (the PE and PPE families), which have conserved Pro-Glu (PE) and Pro-Pro-Glu (PPE) motifs [53]. Even with WGS investigation [136], the function of the PE and PPE genes has remained elusive, although recent studies have suggested that they may play a role in virulence [137]. Their association with drug resistance remains largely unexplored because bioinformatic studies of M. tuberculosis often exclude these genes [138, 139]. In the future, long-read sequencing technology may allow these regions to be sequenced successfully in order to assess if they have a role in drug resistance.

Understanding the evolution and spread of drug resistance in M. tuberculosis

Prior to WGS, the diversity and epidemiology of resistant M. tuberculosis were examined using DNA fingerprinting techniques, like IS6110 restriction fragment length polymorphism (RFLP) typing [140], spoligotyping (spacer oligonucleotide typing, a method of typing strains according to the distinct hybridization patterns of their spacer oligonucleotides) [141], and mycobacterial interspersed repetitive units-variable number of tandem repeats (MIRU-VNTR) typing [142,143,144,145]. These techniques enabled assessments of the diversity of resistant strains in specific geographic regions [146,147,148,149] and, when combined with the genetic profiling of resistance mutations, allowed strain-level monitoring of patients on TB therapy [150].

The dramatic increase in resolution afforded by WGS has extended the sensitivity and resolution with which the diversity and evolution of drug-resistant M. tuberculosis can be assessed. This has resulted in the more confident identification of cases of recent transmission [151] and re-infection [152], and has provided insights into the evolution of resistance within individual patients and across populations. WGS has also enabled more sensitive differentiation of de novo acquisition of resistance (where resistance mutations emerge within a host) from person-to-person transmission of resistance, a critical capability given that these two scenarios require different health-system responses in order to stem resistance.

Within patient evolution of drug resistance

Despite the slow evolutionary rate of M. tuberculosis, estimated at 0.3–0.6 SNPs/genome/year [69, 92, 125,126,127,128], experimental data suggest that drug resistance can evolve within an individual patient during TB treatment. Eldholm et al. [61] described the first documented case of XDR evolution of M. tuberculosis from a fully susceptible ancestor within a single patient, by sequencing nine serial isolates collected over a 42-month period. During this time, seven known resistance mutations emerged in a stepwise fashion after the clinical use of each corresponding drug, revealing how TB drug pressures can rapidly shape M. tuberculosis populations in vivo.

However, the evolution of drug resistance within a host is not always linear, and instead can involve a complex interplay of heterogeneous M. tuberculosis populations [153, 154]. In particular, transient genetic diversity can exist before a dominant clone emerges. In addition, as the size of the transmission bottleneck (the number of bacteria transmitted during an infection event) in M. tuberculosis is not well understood [155], it is difficult to estimate the relative contribution of diversity that is transmitted to the patient versus diversity that evolves within the patient. Numerous WGS studies, performed either on isolates or directly on DNA extracted from serially collected sputum samples, have revealed substantial transient genetic diversity in pathogen populations within patients, particularly within resistance genes [61, 62, 106, 156,157,158,159]. This diversity was observed to endure months before a single variant became fixed in the population (the situation when only a single variant remains). In the study by Eldholm et al. [61] mentioned above, the seven resistance-conferring mutations that eventually dominated were from amongst 35 mutations observed in total throughout the sampling period [61, 160]. They joined eight other mutations that were not resistance-associated but that also became fixed in the population, probably as the result of a phenomenon called ‘hitchhiking’ in which non-adaptive mutations are selected because of their linkage and physical proximity to consequential mutations.

The relative fitness cost of drug-resistance mutations often determines which mutations become fixed within a host. While multiple mutations that confer resistance to a specific drug can evolve repeatedly, mutations conferring no or little fitness cost are typically selected, resulting in fixed dominant mutations [61, 156]. Compensatory mutations (discussed in more detail later), which serve to counterbalance the deleterious effects of acquired resistance, have also been shown to emerge during treatment [156].

WGS has also revealed how combination chemotherapy effectively prevents the emergence of drug resistance during treatment for TB. In a study of very deep WGS of serial sputum specimens from patients receiving treatment for TB, Trauner et al. [62] demonstrated that the combined action of multiple active drugs prevented transient mutants from fixing within a population and becoming dominant. The fewer the drugs that were applied, the more likely it was that resistance would develop and become fixed.

Population views of drug-resistance evolution

A number of careful WGS studies have empirically established SNP-based criteria to discriminate cases of recent transmission from unrelated infections—usually using the criterion that recently transmitted strains differ by fewer than 6–12 total SNPs across the M. tuberculosis genome [63, 125, 126, 161]. In a 2016 review, Nikolayevskyy and colleagues [63] systematically compared WGS to fingerprinting techniques for detecting transmission, including a meta-analysis of 12 studies published between 2005 and 2014. They concluded that results from WGS studies not only have higher discriminatory power, but they also enable more sensitive detection of transmission events that may have been missed by epidemiologic methods.

Although traditional spoligotyping analyses suggested that drug-resistant strains were diverse, WGS of clinical isolates began to reveal the full breadth of diversity in resistant M. tuberculosis. The TB epidemic in South Africa over the past two decades has been well-studied in this regard. In an early WGS investigation, Ioerger et al. [64] examined 14 phenotypically diverse strains within the Beijing lineage and showed that resistance mutations arose independently multiple times, and that XDR isolates may be less fit and less able to transmit. WGS studies across larger sets of strains from the same region in South Africa suggested that, although de novo resistance is indeed common, highly resistant strains (including MDR and XDR strains) have the ability to spread broadly by person-to-person transmission. This includes the ongoing transmission of a circulating XDR clone in South Africa that is linked to the infamous Tugela Ferry XDR outbreak [162] that brought XDR-TB to the world stage in 2005. A more recent large-scale study confirmed that XDR strains have been broadly transmitted person-to-person in KwaZulu-Natal [65].

The patterns observed in South Africa hold for many other parts of the world. Recent studies have shown that patterns of both de novo evolution and person-to-person spread of drug resistance in M. tuberculosis also occur in Belarus, Russia, England, and Malawi [73, 139, 159, 163, 164]. In a composite analysis of over 5000 M. tuberculosis isolates from patients from around the globe, Manson et al. [66] confirmed that both de novo evolution and person-to-person transmission are important factors for the rise and spread of drug-resistant TB worldwide. The emergence of MDR and XDR M. tuberculosis was found to be a frequent occurrence that is distributed fairly evenly across the globe [66]. This analysis also predicted that 37% of MDR isolates in this study had spread person-to-person, which is probably a vast underestimate of how frequently MDR is transmitted once evolved [66].

Geographic movement of people is also an important consideration with regard to person-to-person transmission. Further examination of the MDR clades from Manson et al. [66] revealed that they included widespread international, and even intercontinental, dissemination of strains that were separated by as few as four SNPs, probably due to spread via international travel [67]. Even within a single province in South Africa, Nelson et al. [68] showed, using genomic sequence data and global positioning system coordinates, that many cases of person-to-person transmission (with ≤ 5 SNPs) of XDR-TB occur between people living a median of 108 km apart, pointing to migration between urban and rural areas as a driver of TB spread. Collectively, these studies reinforce the idea that the geographic movement of people must be taken into consideration in any strategy for controlling the spread of TB resistance.

Ordering of the acquisition of resistance and compensatory mutations

Recent WGS studies have helped to illuminate the steps or ‘fitness landscape’ through which M. tuberculosis develops and compensates for drug resistance. Several studies [66, 69, 70] have shown that the order of acquisition of drug-resistance mutations in complex resistance cases is partly constrained in clinical M. tuberculosis. For example, in MDR-TB, isoniazid resistance (most often involving a katG S315T mutation) overwhelmingly evolves prior to resistance to rifampicin and second-line drugs. This was first shown using regional datasets from South Africa [69] and Argentina [70], and recently confirmed by Manson et al. [66] using a global dataset of 5310 strains. In the study by Manson et al. [66], this ordering was shown to hold true over 95% of the time, even for distinct global regions and time frames, including times when both rifampicin and isoniazid were in use, suggesting that the earlier introduction of isoniazid in the 1950s was not the major contributor to this effect. It was also shown that inhA promoter mutations that confer isoniazid resistance (such as those observed by Perdigão et al. [165] in Portugal) were acquired earlier than rifampicin mutations, although the number of samples harboring these mutations was much smaller. Further studies are necessary to determine whether isoniazid preventive monotherapy, which is one of the treatments for latent tuberculosis, may account for some of this effect, as this could result in a background level of increased isoniazid monoresistance.

Compensatory mutations that potentially ease fitness effects caused by resistance often occur after the evolution of primary resistance. This phenomenon was reviewed by Fonseca et al. [71], and examples include mutations in the ahpC promoter region and the rpoC/rpoA genes, which act as compensatory mutations for isoniazid and rifampicin resistance, respectively. Newer WGS work has pointed to several novel compensatory mutations in M. tuberculosis, particularly for rifampicin resistance. Comas et al. [72] identified a set of compensatory mutations in the rpoB gene that conferred high competitive fitness in vitro and were also found frequently in clinical populations. In a large-scale analysis of 1000 strains from Russia, Casali et al. [73] examined strains with primary resistance mutations in rpoB and identified accompanying compensatory mutations in rpoA and rpoC. Cohen et al. [69] identified putative rifampicin compensatory mutations that are present in South African strains by searching for rpoA, rpoB, and rpoC mutations that evolved only after or concurrent with rifampicin resistance-conferring mutations. A recent study of highly resistant M. tuberculosis strains from Central Asia confirmed that the presence of compensatory mutations, particularly those compensating for the fitness cost of mutations that confer rifampicin resistance, is associated with transmission success and higher drug-resistance rates [74]. Beyond rifampicin resistance compensation, Coll et al. [59] identified mutations in pncB2 that may compensate for pyrazinamide resistance conferred by pncA, and similarly, mutations in thyX-hsdS.1 (the thyX promoter) that may compensate for PAS resistance conferred by thyA; however, experimental validation of these potential compensatory relationships is needed. Even fewer studies have identified stepping-stone mutations in M. tuberculosis, which emerge prior to higher-level resistance mutations. Cohen et al. [69] found that ubiA mutations emerge in a stepping-stone fashion prior to more classic embB mutations that confer ethambutol resistance. Safi et al. [15] also showed in vitro that multi-step selection involving ubiA, aftA, embB, and embC is required to achieve the highest levels of ethambutol resistance.

The challenge of mixed infections

Although WGS approaches have great sensitivity in detecting cases of recent transmission, reconstructing the details of transmission networks with WGS [166,167,168] can be difficult. Transmission network mapping is highly dependent on sampling density and studies rarely, if ever, comprehensively sample an outbreak or the extent of within-host diversity. It is also becoming clear, from the prevalence of very close relationships between isolates from patients who have no other direct epidemiological connections, that transmission may largely result from casual contact in community settings [169]. In addition, the phylogenetic reconstruction of transmission networks can be especially challenging, particularly because of the very close relationships between strains and the slow rate of evolution of M. tuberculosis [92, 125,126,127,128].

Mixed infections represent a major challenge for understanding drug-resistance evolution within patients [153, 158, 159]. It can be straightforward to disambiguate co-infections of strains from different lineages, but mixed infections involving strains that have few genetic differences can also occur, making these strains difficult to distinguish. Köser et al. [75] used WGS for rapid drug-susceptibility testing of a patient with XDR-TB, and determined that the patient carried two different XDR-TB Beijing strains with differing resistance mutations. In a study by Liu et al. [76], three dominant sub-clones differing by 10–14 SNPs were detected within a single patient, each with different resistance patterns and probably different anatomical distributions. Also, co-infection by strains with differing resistance patterns may yield misleading composite views of resistance; for example, co-infection with two MDR-TB strains—one with quinolone resistance and the other with aminoglycoside resistance—may be mistaken for infection with a single XDR-TB strain.

Furthermore, newer data suggest that there can be genetic heterogeneity among M. tuberculosis isolates from different parts of the body, potentially leading to incomplete views of drug resistance within a patient (Fig. 1). In a study by Lieberman et al. [77], the authors observed evidence for both within-host evolution and mixed infection by piecing together the genetic variation observed among M. tuberculosis isolates from multiple post-mortem biopsies from the same patient. Another recent study by Dheda et al. [78] showed that drug concentrations at seven body sites were inversely correlated with the MIC of the bacteria isolated from these sites. Sequencing and comparison to pre-treatment and serial sputum samples suggested ongoing acquired resistance and differential evolution across sites [78]. These findings underscore the limitations of diagnosing or studying the evolution of drug-resistant M. tuberculosis using a single patient specimen. However, they also show the promise of WGS for informing interventions related to drug delivery, dosing, and diagnostics, thereby helping to prevent the development of acquired resistance within a patient. More research in this area is needed to determine the breadth and scope of mixed infections among patients with active TB, their contribution to changing drug-resistance patterns over time, and the role of spatial heterogeneity in the evolution of drug resistance.

Fig. 1
figure 1

Challenges to predicting drug resistance accurately from clinical specimens using current culture-dependent molecular diagnostics. The left panel depicts an expectorated sputum sample, which may not accurately represent the microbiologic diversity within the source patient. Culturing this sample (center panel) introduces further biases between faster- and slower-growing strains, such that faster-growing strains are over-represented within the cultured sample. Genomic DNA that is isolated and sequenced is input to computer algorithms that determine the genomic content, including the identification of drug-resistance mutations. However, disambiguating samples that contain mixed strains or detecting heteroresistance remains a computational challenge. The left panel was adapted from Ford et al. [170], with permission from Elsevier

Full size image

From bench to bedside: promise and challenges

Given that the failure to identify and treat patients who have drug-resistant TB leads to increased mortality, spread of resistant strains, and gain of additional drug resistance [171], there is a critical need to diagnose resistant M. tuberculosis in patients rapidly. Several important molecular diagnostic platforms have been established for the identification of M. tuberculosis and drug resistance within this organism, but they are limited to the identification of a defined subset of resistance mutations [172], do not always include the earliest-arising mutations that precede MDR [66], and do not provide knowledge that is useful in determining whether a patient has been re-infected, whether the patient has a recurrent or mixed infection, or whether a particular infection represents a transmission event. WGS holds significant potential to modernize the TB laboratory and improve upon TB management [173], and this topic has been reviewed previously [173, 174]. To date, WGS has been primarily applied as a clinical tool to achieve two goals: first, to detect M. tuberculosis within a clinical sample, and second, to detect resistance mutations and predict resistance patterns so that appropriate treatment can be provided. In order to provide clinically useful information, a diagnostic platform must be rapid. Historically, WGS has relied upon an input of pure mycobacterial cultures, which is time-consuming (requiring multiple weeks) and therefore of less clinical utility. Several investigations have attempted to address this issue by using earlier culture inputs or by attempting culture-independent, direct sequencing from clinical specimens [80, 82, 175]. In a rapid, yet still culture-dependent method, Pankhurst et al. [80] prospectively compared real-time WGS of “early positive liquid cultures” to routine M. tuberculosis diagnostics, and found that WGS achieved a faster time to diagnosis at a lower cost.

Although the advances achieved using WGS are promising, several hurdles must be overcome before it can be put into practical use in the clinic (Fig. 1). Requirements for costly equipment, technical expertise, and substantial computational resources present challenges to implementation [173]. Direct sequencing of patient samples has revealed that the vast majority of DNA present is from the patient or from non-mycobacterial prokaryotes, with variable quantities of mycobacterial DNA present. Doughty et al. [81] performed a pilot study demonstrating the feasibility of direct sequencing using a benchtop sequencer (Illumina MiSeq, San Diego, CA) and sputum samples from eight patients. Although they were able to identify the presence of M. tuberculosis, the low depth of sequencing coverage of the genome (0.002 to 0.7x) prevented drug-susceptibility prediction. Separately, Brown et al. [176] performed an enrichment step with biotinylated RNA baits prior to direct sequencing of sputum, resulting in higher quality data (> 20x depth and > 90% coverage) that allowed the identification of resistance mutations.

Using a targeted DNA enrichment strategy to study 43 individuals with active pulmonary TB, Doyle et al. [177] compared WGS directly from sputum with mycobacterial growth indicator tube (MGIT) WGS. Although direct sputum sequencing was able to identify drug resistance much more rapidly than MGIT WGS, only 74% of sputum samples yielded interpretable WGS data (vs 100% from MGIT); thus, additional optimization of these methods is needed to increase the sensitivity of this approach. Similarly, in a recent study, the use of pyrosequencing of a concentrated sputum sediment (rather than from sputum directly), dramatically shortened the time to initiation of an MDR-treatment regimen [178].

One promising technology that could change clinical WGS is long-read sequencing using the Oxford Nanopore Technologies (ONT; Oxford, UK) platform. An advantage of ONT is the ability to allow sequencing to continue until sufficient coverage of the genome has been obtained, potentially solving the problem of low or variable amounts of M. tuberculosis in clinical samples [82]. Early ONT studies have shown promise in identifying antimicrobial-resistance genes in different bacterial species [179]. Unfortunately, at present, both the high error rate of ONT MinION and potential difficulties with GC-rich regions limit the utility of this technology; thus, improvements in accuracy are necessary to enable the identification of resistance associated with point mutations [179]. ONT metagenomic sequencing has been successfully applied to improve pathogen detection and antimicrobial-resistance testing in other clinical settings [180]; however, to date, applications of this technology to M. tuberculosis have been limited to pre-clinical research [82].

Despite these challenges, WGS offers several advantages over the technologies that are currently employed for diagnosis and epidemiological monitoring of TB. Using WGS directly on patient sputum could reduce the turnaround time for diagnosis and determination of antibiotic-resistance status from weeks to hours [61, 159], and would prevent the introduction of culture-induced biases. The depth of information provided by WGS could be used to identify whether an individual harbors multiple co-infecting strains [106, 160] and to distinguish recurrent infection as either relapse or re-infection [174, 181]. In addition, WGS could provide real-time epidemiological information that could be useful for understanding patterns of drug resistance and for establishing chains of transmission [174]. Encouragingly, the high levels of concordance observed between the genotypes and phenotypes of clinical samples indicate that WGS can provide high accuracy for both diagnosing TB and informing treatment options [113]. Finally, WGS of patient samples would provide a high level of convenience, by combining diagnosis, resistance profiling, and epidemiological analysis into a single test [85]. Given these advantages, the WHO has recently published a technical guide for the implementation of next-generation sequencing (NGS) technologies for the detection of drug resistance in M. tuberculosis [182].

Routine whole-genome sequencing of mycobacterial isolates

In 2017, England became the first nation to launch routine WGS of all prospectively identified M. tuberculosis clinical isolates [183]. Sponsored by Public Health England (PHE), prospective WGS is being performed on all positive mycobacterial cultures. Within 5–7 days of receipt of the culture from the reference lab, data will be provided on the mycobacterial species, the predicted drug susceptibility, and the molecular epidemiology of the strains. If, from the sequence analysis, a strain is predicted to be fully susceptible to first-line antitubercular drugs, phenotypic drug-susceptibility testing (DST) will no longer be performed routinely. However, if drug resistance to any first-line drug is identified, then phenotypic DST will follow. Beyond drug-susceptibility prediction, these efforts will have profound implications for TB control because WGS data can be used for real-time molecular epidemiology in this context.

Given the high sensitivity of WGS in detecting drug resistance to first-line TB drugs [60], similar algorithms utilizing WGS to predict susceptibility (rather than to identify drug resistance) for first-line drugs, in lieu of phenotypic DST, have been endorsed in the Netherlands and in New York [60]. It seems highly likely that these kinds of efforts would be helpful in higher-burden TB settings than those mentioned here, but the feasibility of this approach has not yet been established, from either a practical or economic standpoint, in settings where the numbers of drug-resistant TB cases are high.

Conclusions and future directions

Since the first applications of WGS to M. tuberculosis in 1998, WGS techniques have greatly accelerated our understanding of drug-resistance mechanisms in this pathogen. Importantly, WGS studies now indicate that, for many drugs, the vast majority of resistance is explained by known mutations. The increasing availability of whole-genome sequences from phenotypically diverse M. tuberculosis, combined with improved GWAS algorithms, is enabling the discovery of the remaining determinants of unexplained resistance. In addition, WGS has provided valuable insight into how resistance mutations evolve and spread. It is clear that both de novo acquisition of resistance mutations and clonal transmission are critical factors in the spread of drug-resistant TB.

Furthermore, WGS investigations have revealed that there is a specific order in which drug-resistance mutations are acquired: isoniazid resistance is almost always acquired before rifampicin resistance, which has significant implications for the design of diagnostic tests. Within individual patients, WGS studies have highlighted that mixed infections are common, and often represent important intermediates in the evolution of drug resistance.

WGS also holds great promise for revolutionizing the rapid clinical diagnosis of TB in the future. Although there are still substantial technical hurdles, WGS can be used to diagnose the presence of M. tuberculosis rapidly, as well as to pinpoint appropriate antibiotic treatment regimens by identifying the complement of M. tuberculosis drug-resistance mutations that are present within a clinical sample. Indeed, improvements in the prediction of drug susceptibility with WGS may obviate the need for phenotypic culture methods, especially for first-line drugs.

Although WGS offers many benefits, targeted NGS, in which sequence data are obtained from only a focused panel of genes or genetic regions rather than from the entire genome, is gaining momentum [184]. One of the advantages of targeted NGS over WGS is that it can be performed directly on clinical specimens and is, therefore, faster than culture-based WGS. Other advantages include reduction in both labor and computational efforts and reduced costs. The potential offered by the application of targeted NGS to the prediction of drug resistance from genomic data is self-evident. Nevertheless, it seems that WGS would have greater discriminatory power than targeted NGS for molecular epidemiology purposes.

Ultimately, the use of WGS is expected to continue to advance our understanding of M. tuberculosis drug resistance. Furthermore, its practical use in clinical settings holds great potential to improve public health through real-time molecular epidemiology tracking, to identify global hotspots of drug-resistance emergence, and to facilitate the development of improved approaches for the diagnosis and treatment of drug-resistant TB.