Clinical Application of Whole-Genome Sequencing To Inform Treatment for Multidrug-Resistant Tuberculosis Cases

Abstract

The treatment of drug-resistant tuberculosis cases is challenging, as drug options are limited, and the existing diagnostics are inadequate. Whole-genome sequencing (WGS) has been used in a clinical setting to investigate six cases of suspected extensively drug-resistant Mycobacterium tuberculosis (XDR-TB) encountered at a London teaching hospital between 2008 and 2014. Sixteen isolates from six suspected XDR-TB cases were sequenced; five cases were analyzed in a clinically relevant time frame, with one case sequenced retrospectively. WGS identified mutations in the M. tuberculosis genes associated with antibiotic resistance that are likely to be responsible for the phenotypic resistance. Thus, an evidence base was developed to inform the clinical decisions made around antibiotic treatment over prolonged periods. All strains in this study belonged to the East Asian (Beijing) lineage, and the strain relatedness was consistent with the expectations from the case histories, confirming one contact transmission event. We demonstrate that WGS data can be produced in a clinically relevant time scale some weeks before drug sensitivity testing (DST) data are available, and they actively help clinical decision-making through the assessment of whether an isolate (i) has a particular resistance mutation where there are absent or contradictory DST results, (ii) has no further resistance markers and therefore is unlikely to be XDR, or (iii) is identical to an isolate of known resistance (i.e., a likely transmission event). A small number of discrepancies between the genotypic predictions and phenotypic DST results are discussed in the wider context of the interpretation and reporting of WGS results.

INTRODUCTION

In 2013, of the 6.1 million new tuberculosis (TB) cases reported worldwide, 480,000 were defined as multidrug resistant (MDR) (1). MDR strains of Mycobacterium tuberculosis are defined as those that are resistant to the key first-line drugs, rifampin (RIF) and isoniazid (INH). Such cases need to be treated using a more complex and flexible drug selection algorithm than the standard, with a combination of at least four drugs from an available panel classified according to their mode of action, efficacy, and clinical experience, and should be further modified if other resistances become apparent (2).

Extensively drug-resistant (XDR) strains of M. tuberculosis are defined as MDR strains that in addition to being RIF and INH resistant are also resistant to the two most effective remaining bactericidal drug classes, at least one fluoroquinolone (FLQ), and one second-line injectable agent (amikacin [AMI] or kanamycin [KAN], or capreomycin [CAP]). The XDR designation reflects the concern that these are even more difficult to treat. Globally, in 2014, 9.0% of MDR cases were reported as being XDR (1); these figures are likely to be underestimates, because testing and reporting practices are not universal.

The overall rates of TB in the United Kingdom have been stable, at ∼14 per 100,000 since 2005, with London having the highest regional incidence (41/100,000) (3). The proportion of MDR cases is rising slowly but steadily, and XDR cases are now appearing. Between 1995 and 2012, there were a total of 24 XDR cases in the United Kingdom (6 in 2011 and 2 in 2012).

The clinical management of XDR-TB is difficult and lengthy, the drain on health care resources is significant, and there is a real public health risk of transmission to the wider population due to ineffective treatment, thus amplifying the problem. To give an idea of the problem, treatment is recommended for 8 months with five drugs, followed by 12 months with oral drugs only, leading to a minimum of 20 months of treatment (1, 4). All of the drugs have considerable side effects, and patient care in an isolation unit is often required during the early months while the patient is still infectious.

Individualized treatment for MDR- or XDR-TB has traditionally depended on phenotypic drug susceptibility testing (DST) for M. tuberculosis. The results of a full DST workup take several weeks, leading to difficult drug treatment decisions, the opportunity for further resistance to arise if inadequate regimens are prescribed, and potential patient exposure to unnecessary drugs and side effects during the interim period.

Molecular diagnostics have enabled the rapid identification of potential resistance-associated mutations; PCR-based methods, such as the GeneXpert MTB/RIF (Cepheid) (5) and GenoType MTBDRplus (Hain Lifescience) (6) assays, have become routine in many laboratories to identify MDR-TB strains by detecting the key mutations in genes associated with resistance to RIF (rpoB) and INH (inhA and katG). Extended PCR-based assays, such as the GenoType MTBDRsl test (Hain Lifescience) (7), also exist to identify additional mutations associated with XDR-TB. However, as these assays are limited to key commonly described mutations known to cause antibiotic resistance and do not allow the detection of new mutations, whole-genome sequencing (WGS) provides a more comprehensive alternative to these targeted approaches. Although the precise link between many of the mutations and phenotypic DST has yet to be established, WGS already has the potential to speed up and augment individualized treatment decisions compared to the use of DST alone, which has been shown to have reliability issues (8).

WGS has been used to retrospectively investigate M. tuberculosis transmission networks (9–12) and drug susceptibility (13); however, its use in informing patient therapy in real time has yet to be reported. In this study, WGS was performed on 16 isolates from six patients with suspected XDR-TB isolated at a London teaching hospital. The aim was to investigate the potential to provide comprehensive genotypic information using WGS in a clinical setting within a short time frame to inform treatment decisions, and also to evaluate the practical issues in providing this support within a large London hospital. This report describes an analysis of the WGS data from these isolates, how the relevant information was relayed to clinicians, and how the information has the potential to impact successful treatment decisions.

MATERIALS AND METHODS

Isolates and DNA preparation.

The patient isolates were cultured using the BacT/Alert microbial detection system (Organon Teknika Corp., Durham, NC) (14), according to standard manufacturer protocols. Cells were harvested after 15 to 17 days (1 to 2 days after being flagged positive, where tested). The bacterial loads in the cultures were 4 × 10⁶ to 10 × 10⁶/ml. The pelleted cells were heat inactivated at 90°C for 45 min and the DNA extracted using the UCP pathogen minikit (Qiagen, Germany), following mechanical disruption using Pathogen lysis tubes with small beads (Qiagen).

Drug susceptibility testing.

Drug susceptibility testing, using the proportion of the resistance ratio method (15), was provided by the Public Health England National Mycobacterial Reference Laboratory (NMRL), London, United Kingdom.

Sequencing.

WGS was performed using the Ion Torrent personal genome machine (Ion PGM) in conjunction with Ion Torrent workflow reagents (Life Technologies, Paisley, United Kingdom). For each sample, 100 ng of genomic DNA was fragmented by sonication using a BioRuptor UCD-200 (Diagenode, Belgium), and the library was prepared using the Ion Plus fragment library kit, according to the manufacturer's instructions. Briefly, fragmented DNA was end repaired, ligated to Ion-compatible adapters, and nick translated. To enable the multiplexing of samples on one sequencing chip, barcoded adaptors were used (Ion Xpress barcode adapters 1–16 kit). Each library was size selected using a 2% E-Gel SizeSelect agarose gel (Life Technologies) to produce a median fragment size of approximately 330 bp and then quantified by quantitative PCR (qPCR) (Ion library quantitation kit). The barcoded libraries were pooled in equimolar amounts before being clonally amplified on Ion Sphere particles (ISPs) using the Ion OneTouch instrument. The template-positive ISPs were then enriched using the Ion OneTouch ES before being sequenced on the Ion 316 chip using a 200-bp sequencing kit.

Sequence analysis.

The sequence reads were mapped to the M. tuberculosis strain H37Rv reference genome (GenBank accession no. NC_000962.3) with TMAP version 4.0.3 (Ion Torrent, USA), alignments were sorted with Picard version 1.76 (http://broadinstitute.github.io/picard/), and site statistics were generated with SAMtools mpileup version 0.1.18 (16). The variant sites were filtered based on the following criteria: mapping quality (MQ) of >30, site quality score (QUAL) of >30, ≥4 reads covering each site with ≥2 reads mapping to each strand but with a maximum depth of coverage of 200×, ≥75% of reads supporting the site (DP4), and an allelic frequency (AF1) of 1. The sites that failed these criteria in any strain were removed from the analysis. Phylogenetic reconstruction was performed using RAxML version 7.4.2 (17) with a general time-reversible (GTR) model of nucleotide substitution and a Gamma model of rate heterogeneity; the branch support values were determined using 1,000 bootstrap replicates. The branch single nucleotide polymorphism (SNP) counts were estimated using ancestral sequence reconstruction performed with PAML version 4 (18). The presence or absence of specific regions of difference (RD) was deduced by mapping reads to the RD sequences using BWA-MEM (19) and the alignments analyzed manually. Drug resistance SNPs were identified as above and compared to the TB Drug Resistance Mutation Database (TBDReaMDB) (20).

Nucleotide sequence accession number.

The sequence data have been deposited in the European Nucleotide Archive with the study accession no. ERP006124.

RESULTS

Description of isolates.

Sixteen isolates from six clinical cases were subjected to WGS (Table 1). The cases were:

• Case 1. Two sputum isolates were received 5 months apart from a case that was previously reported (21) and for which a change in PZA resistance (PZA^s to PZA^r) was reported on DST.

• Case 2. Two sputum isolates were received 9 months apart, for which a change in ethambutol (EMB) resistance (EMB^s to EMB^r) was reported on DST, and the patient had inadequate adherence to treatment during the interim period.

• Case 3. Five sputum isolates were taken over a 3-week period (days 1, 7, 12, 15, and 18).

• Case 4. One sputum isolate was received from a known contact of case 3.

• Case 5. Three nonsputum isolates were collected in total, with two isolates taken on a recurrent presentation of spinal TB after a treatment default and one isolate from a chest abscess from a previous admission prior to a default.

• Case 6. Three sputum isolates were taken during a 1-week period.

TABLE 1.

Details of cases and isolates

Case	Isolate	Date of isolation (DD/MM/YYYY)	Sourcea	Coverageb
1	1a	16/07/2008	SPT	33.2
	1b	10/11/2008	ISPT	39.6
2	2a	02/07/2010	SPT	55.6
	2b	03/04/2011	SPT	53.7
3	3a	03/05/2013	SPT	36.3
	3b	10/05/2013	SPT	35.7
	3c	15/05/2013	SPT	39.3
	3d	17/05/2013	SPT	41.9
	3e	21/05/2013	SPT	167.8
4	4a	29/07/2013	SPT	52.3
5	5a	14/09/2013	LN	41.2
	5b	20/09/2013	PUS	47.9
	5c	08/07/2013	LN	55.7
6	6a	01/01/2014	SPT	56.3
	6b	03/01/2014	SPT	51.6
	6c	06/01/2014	SPT	41.7

^aSPT, sputum; ISPT, induced sputum; LN, lymph node.

^bFold sequence coverage of the H37Rv reference genome.

All patients, except case 5, who defaulted, either completed or are still on treatment. The sputum samples were repeatedly culture negative, and all imaging has improved relative to pretreatment.

Sequence analysis for potential resistance SNPs.

The sequence reads were mapped to a reference genome, and SNPs and indels were identified upstream or within the coding frame of any gene reported to be associated with antibiotic resistance in TBDReaMDB (20); these are shown in Table 2, alongside the DST results, where tested. An analysis carried out according to the drug class revealed the following.

TABLE 2.

Correlation of SNPs with phenotypic DST by antibiotic class

^aThe E. coli coordinates are used for RpoB as is standard.

^bThe DST results are presented for each case together with the SNPs identified in genes listed as being associated with drug resistance, as defined in TBDReaMDB (20). S, sensitive; R, resistant; H, highly resistant; +, presence of SNP. Samples are listed in date order.

^cAGs/Ps, aminoglycosides/peptides.

^d-, single-base deletion.

^e*, stop codon.

(i) INH.

All isolates were resistant to INH by DST, and all carried the high-confidence (as defined in TBDReaMDB) Ser315Thr mutation in katG.

(ii) RIF.

All isolates were resistant to RIF by DST. High-confidence mutations were seen in rpoB, with Leu430Pro in case 1, and Ser450Leu in the other five (Escherichia coli numbering is used as the standard nomenclature for RpoB; E. coli residues 430 and 450 refer to M. tuberculosis residues 511 and 531, respectively).

(iii) PZA.

The PZA results were more complicated to interpret because of known issues with DST accuracy (8) and the large numbers of potential mutations associated with different resistance mechanisms against this drug. Five out of the six cases were resistant to PZA by DST. PZA resistance is due to a number of different mutations in the pncA gene that lead to a decrease or ablation in PncA (pyrazinamidase) enzyme activity by introducing frameshifts, stop codons, or amino acid substitutions. Therefore, it is not surprising that apart from the likely transmission cases (3 and 4), the isolates from each case carried a different mutation from those in other cases:

• Case 1. As discussed previously (21), the first isolate was reported by DST as susceptible and the second as resistant. As both isolates had a frameshift mutation introducing a stop codon, causing a loss of 40% of the protein, it is likely in this instance that the discrepancy between the DST and WGS data was due to the inaccuracy of the phenotypic testing.

• Case 2. Both isolates were reported by DST as being PZA^r. However, the WGS indicated no mutation in pncA for isolate 2a, and there was a Gly132Asp substitution in isolate 2b. As the Gly132Asp mutation in pncA has been reported a number of times as being associated with resistance, and PZA DST is particularly difficult, this might explain the discordance.

• Cases 3 and 4. These isolates were reported by DST as resistant, and both have a stop codon, introducing a mutation in pncA at residue 64.

• Case 6. These isolates were reported by DST as resistant, and a pncA mutation (His57Arg) was identified, which has been reported elsewhere (22).

(iv) EMB.

All isolates were reported as resistant to EMB by DST, except for the first isolate, 2a, of case 2. Four previously reported SNPs were observed in embB and two in the promoter region of embA, and all were concordant within the cases. As EMB also presents difficulties with DST, the discordance in case 2 may well be due to inaccuracies in DST, although this was not rechecked. All except one embA SNP have been reported before, and the novel SNP occurred in strains that also had an embB SNP.

(v) FLQ.

All tested isolates from cases 1 to 5 were reported by DST as resistant to both ofloxacin (OFX) and moxifloxacin (MOX). All possessed mutations within the quinolone resistance-determining region (QRDR) of gyrA (GyrA residues 88 to 94), except for the first isolate for case 2. The case 6 isolates were reported by DST to be susceptible to OFX and MOX.

• Case 1. This isolate carried an Ala90Val mutation, which was reported by Malik et al. (23) to produce resistance to all quinolones but lower resistance to levofloxacin (LEV) (2 μg/ml) and MOX (1 μg/ml) than to OFX (4 μg/ml). This knowledge, which was later confirmed by MIC phenotypic testing, allowed MOX to be used at a higher dose (21).

• Case 2. Both isolates were reported by DST to be resistant to both OFX and MOX. Curiously, by WGS, isolate 2b had an Asp94Gly mutation that causes quinolone resistance, but isolate 2a did not, and an analysis of the loci for both isolates confirmed that the sequence call was confident, with no evidence of heterozygosity. However, a previous gyrA-specific PCR had shown the presence of the Asp94Gly mutation in isolate 2a. The other SNPs in gyrA (encoded by Glu21Gln and Gly688Asp) are not thought to be associated with resistance (see below).

• Cases 3 and 4. These isolates also carried the high-confidence quinolone resistance mutation, Asp94Gly.

• Case 5. These isolates carried a Gly88Cys mutation, which is reported to cause resistance (24).

• Case 6. These isolates were sensitive by DST and contained no resistance-associated mutations in gyrA QRDR.

All isolates contained three mutations leading to SNPs in gyrA that are not associated with resistance and are located outside the QRDR: Ser95Thr (25), Glu21Gln, and Gly688Asp (26).

(vi) STR.

Streptomycin (STR) is not recommended for use in MDR-TB treatment (27), and the only case isolates with an STR DST result (cases 1, 2, and 3) were resistant. The isolates in cases 1, 2, 3, 4, and 6 contained well-known mutations in rpsL; the case 2 isolate had a Lys88Arg substitution, while the others carried Lys43Arg. The case 5 isolates had no mutations in rpsL and two in gid (Glu92Asp and Ala138Glu) that appear not to be related to resistance (28), and we would predict that these isolates would be susceptible to STR; however, no DST result was available for comparison.

(vii) AMI.

The case results for AMI are as follows:

• Case 2. These isolates were resistant by DST to AMI, KAN, and CAP. Both isolates possessed a 1401A→G SNP in the rrs gene, which was previously reported to cause cross-resistance to these three antibiotics (29) and was evaluated as a high-confidence resistance SNP in a systematic review (30).

• Case 3. These isolates were resistant to AMI and KAN but susceptible to CAP. The only SNP in a relevant gene not present in the susceptible isolates was an eis promoter SNP (-14C→T). The systematic review by Georghiou et al. (30) supports this resistance profile.

• Case 4. This isolate was reported to be resistant to KAN but susceptible to AMI and CAP, a more unusual phenotype. As it had the same eis -14C→T SNP as that in case 3 and no other suspect SNPs in known genes, we suggest that the WGS result is more reliable.

The isolates from all six patients contained one or more SNPs in rrs, eis, or tlyA that were not previously shown to be associated with resistance.

(viii) ETH.

The case results for ethionamide (ETH) are as follows:

• Cases 1, 3, and 4. These isolates were resistant to prothionamide (PRO) and contained mutations in the ethA gene, a nonsense and a frameshift mutation, respectively. As with pncA and PZA, a functional ethA gene is required to activate the prodrug, and any mutation inactivating the encoded monooxygenase enzyme will lead to resistance.

• Case 5. These isolates contained a single-base deletion in ethA that is not known to be associated with resistance. The SNP would only cause the last 15% of the protein to be lost, and it would not be unusual for a protein to retain activity under these circumstances.

(iv) PAS.

The isolates from cases 4 and 5 were reported by DST as being resistant to para-aminosalicylic acid (PAS). The resistance mechanisms to PAS are less well defined, with only thyA identified as being involved in resistance (31). Any mutation inactivating the gene or reducing the encoded enzyme activity will lead to resistance. Therefore, although the thyA Thr26Pro allele has not been described before, the presence of this allele in case 5 is potentially indicative of resistance. In contrast, no SNP was seen in this gene in the isolates from cases 3 or 4, and the case 4 isolate was PAS^r (no result was given for case 3), which suggests that other loci may be involved in resistance.

Phylogenetic reconstruction.

Human M. tuberculosis has been found to fall into six large clades that have evolved in different geographical areas (32). In the context of 258 publicly available sequences (33, 34), phylogenetic reconstruction using 27,352 high-confidence SNP sites confirmed that the isolates were part of the East Asian (Beijing) lineage (see Fig. SA1 in the supplemental material), which was further confirmed by the identification of the lineage-defining mutation A→C at H37Rv position 1834177 (35) and the absence of the RD105 region (36).

A previous study used a cutoff of >12 SNPs to determine that isolates were not epidemiologically linked (9). An analysis using only the study isolates based on 410 high-confidence SNP sites (Fig. 1) showed that these isolates from the same hospital differed by between 33 and 297 SNPs and were therefore unlikely to form a local transmission network, except for the known contact cases of cases 3 and 4. Furthermore, the study isolates appear to be quite diverse within the East Asian clade (see Fig. SA1 in the supplemental material).

FIG 1.

Phylogenetic reconstruction of 410 high-confidence SNP sites from the study isolates. (A) Estimated numbers of SNPs are labeled on each branch. The clusters in which no SNP distances are shown are 0; however, due to the nature of the SNP calling algorithm, more sites are excluded as more isolates are added. We therefore reanalyzed the case isolates separately, and the resulting SNP distances were: case 1, 0; case 2, 16; case 3, 0; cases 3/4, 1; case 5, 1; and case 6, 0. Ref, reference. (B) The presence/absence of each RD region is aligned to each isolate. These regions of difference are genomic deletions that were previously reported within the M. tuberculosis complex (47). The RD105 deletion has been reported as being specific for East Asian (Beijing) strains (36).

The SNP sites were reanalyzed, comparing only the isolate data within the same case (except for cases 3 and 4, which were grouped). Cases 1, 3, and 6 still showed no SNP differences between the respective isolates, the case 5 isolates differed by one SNP, and the case 4 isolate differed from those of case 3 also by one SNP. Meanwhile, the isolates from case 2 differed by 16 SNPs, as determined using a pairwise comparison of the SNP differences (see Table SA1 in the supplemental material). The discrepancy with the 8 SNPs different for the case 2 isolates, indicated in Fig. 1, is likely due to the presence of extra SNPs in transposase or PE_PGRS genes included in a pairwise analysis, which are not likely to be present in all isolates and hence were filtered out in the SNP collating process for a larger number of isolates.

Effect of WGS on clinical information.

For case 1, as described previously (21), targeted amplification and sequencing of the gyrA, gyrB, and pncA genes had been undertaken previously. These data had indicated an incorrect initial DST report for PZA and suggested that high-dose MOX treatment might be effective based on the GyrA Ala90Val substitution identified, despite the breakpoint-resistant phenotype. WGS confirmed these previous data and allowed other resistance SNPs to be identified, but it was not in itself clinically useful, because this isolate was sequenced retrospectively.

For case 2, two isolates were analyzed, one taken during the first treatment course, and one during the second. Between these samples, the patient had a single recurrent positive sputum culture associated with a period of excess alcohol intake and missing treatment doses. DST suggests that EMB resistance (EMB^r) had evolved between the treatment courses; however, WGS suggested that the known EMB^r genotypic markers had not changed. The explanations include that other genes are responsible for resistance to EMB, the initial DST result was incorrect, or the initial isolate contained a mixed population with a low-level EMB^r isolate. A decision had to be made using data from both DST and WGS as to whether to continue EMB or stop. It was continued, taking into account the observation that EMB can have a synergistic effect with other antibiotics, including aminoglycosides/peptides and FLQs, even in the presence of EMB^r (37).

In the case 3 isolates, the presence of a frameshift mutation in pncA leading to a truncated gene product helped confirm that the reported DST result was correct, which is clinically useful, as there is a rate of false resistance even in liquid culture DST (38). The WHO guidelines suggest the use of PZA, irrespective of the DST results, given the challenges of testing (27). The WGS results therefore increased our confidence in the DST result and aided the decision to stop the drug.

Case 4 was a contact of case 3, and treatment had been started immediately, with an assumption that the isolates had the same resistance profile. WGS showed that the isolate from this case was almost identical to the case 3 isolates, with only one SNP difference detected; this confirmed that a transmission event had occurred directly or indirectly, thus allowing further confidence in the continuation of the treatment started on this basis.

Case 5 was a spinal case of TB who had been treated for MDR-TB four times before this admission. We requested an earlier sample from the reference laboratory from a previous treatment session. WGS allowed us to confirm that no further resistance had been acquired during the last treatment period, thus aiding the decision to treat with an MDR rather than an XDR regimen before the full DST result was available.

Case 6 presented with pulmonary smear-positive TB. A PCR performed on the sputum samples had revealed RIF resistance, and the country of origin put the patient at risk for XDR-TB. The WGS revealed no known mutations associated with FLQ or aminoglycoside resistance. Thus, this information combined with the GenoType MTBDRsl test results were used as evidence to treat with an MDR-TB regimen rather than an extended XDR-TB regimen until the phenotype was further confirmed by DST.

Reporting.

A key issue we addressed was how to sensibly report and condense the relevant findings from WGS into an easily accessible form for use by clinical colleagues. An example of a representative anonymized report produced for case 3 is presented in Fig. 2. This short factual report provided a summary of isolates analyzed, an indication of the predicted resistance profiles with supporting evidence for review, and a wider context of the isolates in comparison to the others previously analyzed in the hospital.

FIG 2.

An example report returned to the clinicians. Page 1 (left) summarizes the overall phenotypic and genotypic profiles of the isolates for a patient. Page 2 (right) lists more detail describing the mutations identified and their potential phenotypic consequences. The full report for case 3 is available in the supplemental material.

The front page of the report provided a patient-centered summary of the existing information available for isolates analyzed by WGS, including hospital identifiers, sample type, and isolation date, plus any known drug resistance, if available from routine laboratory testing. The genomic profiles of the isolates, based on the presence or absence of genomic regions of difference (RD), were provided as a low-resolution visual pattern of isolate similarity and enabled an assignment to be made to a known lineage of M. tuberculosis. Most importantly, a summary of the predicted resistance profile was included on the front page, together with an indication of the confidence in the supporting evidence for review, highlighting the requirement for further interpretation. The second page of the report provided a summary of this supporting evidence for the predicted resistance profiles.

As a conceptual starting point, in the absence of agreed phenotypic-genotypic correlations for many of the antibiotics and the diverse findings in the literature, TBDReaMDB was used as a reference database for the genes and mutations associated with resistance, providing a fixed reference for a high-confidence association with resistance and links to the supporting literature. If a high-confidence mutation was detected, this was reported and thus expected to confer resistance. If no high-confidence mutations were detected, any mutations in the genes associated with resistance were reported. This enabled further follow-up through a review of the limited or recent literature available, in the context of other clinical information available for each case to inform treatment decisions. Clearly, the aim of the report was to remain factual, rather than interpretative, and to avoid any direct diagnostic calls or recommendations for treatment.

An appendix to the WGS report was included to add further context for the reported isolates compared to the other analyzed local hospital isolates. This would immediately help to identify contacts of the case or potential transmission events and also demonstrates the power of such analyses to be able to compare to expanding local, national, or international WGS databases.

DISCUSSION

Sixteen isolates from six patients with suspected XDR-TB treated at a London teaching hospital between 2008 and 2014 were analyzed by WGS in order to identify the genotypic mutations that may predict potential antibiotic resistance and reveal phylogenetic relationships. This provided a useful case study for the application of WGS in a clinical setting, involving a team with a mix of clinical, microbiology, and genomics expertise. In several cases, it was shown that the speed of generation and the comprehensive nature of the WGS data were useful for informing clinical decisions in this focused cohort of suspected XDR-TB, but they also raised a number of issues for consideration and further discussion.

Concordance of genotypic prediction and phenotypic resistance.

The WGS data generally demonstrated good concordance with the DST results. Any discrepancies were mainly drug class dependent and may be explained by known issues with current DST methodology; some drugs are particularly difficult to test for, including PZA and EMB, leading to an international program to improve accuracy (29). In practice, knowledge of the genotype increased the confidence of clinicians in continuing or discontinuing treatment with a particular drug, especially when considered alongside other information available to them and possible alternative treatment options.

However, these discrepancies could also be explained by the sampling of different representative members of a heterogeneous population or, similarly, due to the potential analysis of a mixed population. In particular, the discordant DST/WGS results for isolate 2a, which was FLQ^r by DST while having been confirmed to have a Asp94Gly mutation (by gyrA-specific PCR), had no known resistance-associated mutations in gyrA by WGS. This might reflect an analysis of distinct members of a closely related but heterogeneous population evolving in the host, a conclusion possibly further supported by the higher-than-expected number of SNPs (a total of 16) estimated between the two isolates collected 9 months apart. However, while this might in theory represent the acquisition of multiple SNPs, the use of cutoffs defined in previous studies (five or fewer SNPs for epidemiologically linked cases [9] and six or fewer SNPs to differentiate relapse from cases of reinfection [33]) would define these two isolates as two closely related but separate populations. But it is important to note that due to the nature of the SNP calling process, in which sites are discarded if they fail the specified criteria in at least a single isolate, SNP differences are likely to be lower when larger numbers of isolates are compared, thus explaining why the number of SNP differences between the case 2 isolates increases from 8 to 16 when using a pairwise analysis. Furthermore, this must also be considered when comparing SNP cutoffs between data sets.

One of the benefits of WGS is that with enough depth, it can resolve mixed infections through the identification of populations of heterogeneous SNP calls, although there was no indication of heterozygosity in any sequence analyzed in this study. An optimal approach in order to obtain a more representative sampling in complicated cases would be to sequence multiple purified colonies from the same clinical sample, especially where different colony morphologies were observed. The natural history of an organism, such as M. tuberculosis, with long periods of latent host infection and active replication during disease with multiple independent foci of infections (liquefying granulomas) replicating at different rates and genetically diverging, emphasizes the need to consider closely related heterogeneous populations and therefore carefully record sample handling procedures prior to sequencing.

Prediction of drug resistance from genotypic data.

A fundamental issue underlying the application of WGS is how to interpret SNPs in relation to phenotypic antibiotic resistance. This is far from straightforward, even with well-known mutations, because there are no agreed-upon standards by which genotype-phenotype associations are assessed; different types and levels of evidence are available for different SNPs. An issue that is particularly important for clinicians treating patients with complex TB cases is to know whether an SNP causes the same level of resistance to all analogues of a drug. For example, the RpoB Ser450Leu mutation causes high-level cross-resistance to different rifamycins, but other SNPs have different phenotypes (39). Similarly, the likely MICs for strains with a particular SNP are useful, as it may be possible to overcome resistance with increased dosage, e.g., the GyrA Asp94Gly mutation has a lower MIC for MOX than for OFX or LEV (23).

Our experience with WGS in the clinical context of drug-resistant TB highlights the urgent need for an internationally recognized database for standardized genotype-phenotype correlation interpretation that can be added to and interrogated by clinicians and microbiologists who will be increasingly accessing WGS platforms at the local hospital level. Currently, the most robust approach to the interpretation of SNP data would be to rely on an up-to-date systematic review of the literature for a particular mutation; for example, with the study by Georghiou et al. (30), although even this review is time and data dependent. In this study, we accessed the recent literature and relied on TBDReaMDB (20), a database of resistance mutations that we have been found to be extremely useful and is used as the type of reference resource or knowledge base that is required in a routine diagnostic setting. An example on which to model a TB diagnostic database would be the Stanford University HIV Drug Resistance Database (40), an effective tool that is routinely used in clinical decisions. We have set up an exploratory wiki (http://tbSNPs.info/) as a publicly viewable option to generate discussion on this subject. In TBDReaMDB, some SNPs are flagged as high confidence, and we have used this in the reports, but even the most robust resistance-associated mutation will have the probability to cause resistance rather than yield a definitive answer, because of errors in DST (27) and the potential presence of compensatory mutations. A more robust genetic approach is to carry out site-directed mutagenesis, which could in theory be systematically performed for important SNPs.

Clinical impact of WGS for suspected XDR-TB cases.

Through this case study, we explored how relevant information from WGS can be extracted and reported to clinicians in a format and time frame that helps inform clinical decisions, particularly with regard to selecting the most effective drug treatment. We have demonstrated the clear clinical utility of WGS within a clinical context and that it is complementary to other information required for clinical decision-making in order to better inform treatment choices or even to modify doses of drugs already in use. In addition, when a TB case (case 4) was confirmed clinically in a contact group of a previous case (case 3), WGS confirmed the transmission event. Thus, the confidence of a clinician to continue the empirical treatment regimen for the case contact, based on that used for the index case, was increased. In this way, clinically timely WGS proved to have both public health and patient treatment impacts.

WGS would have a greater impact if routinely applied as soon as there was an indication of an MDR-TB case from the GeneXpert test result following admission. Much of clinical decision-making is based upon evaluating levels of uncertainty, and we would argue that WGS can significantly help aspects of that process. In addition to providing information that is part of the decision-making process, the WGS data allow informed retrospective reflection on treatment and allow others to evaluate and audit the service provided. DST and WGS were thus both seen by clinicians as being useful, with different advantages and disadvantages, providing some similar but also some complementary information, and with different timings.

For this study, a prototype report was developed (see Fig. 2) as one possible approach to help generate debate with respect to useful content and a consideration of how such reports should be presented or supplied in the future. Developing best practice is difficult without an open discussion, and we have not found other published examples of such reports for comparison. The aim was to provide a good balance of data and interpretation, together with clarity and consistency of the presentation format, and to invite feedback from end-user clinicians. An issue of central importance is how much interpretation should be included. The forms we present were factual, leaving all interpretation to the clinical team, except for the reporting of high-confidence mutations, as defined by TBDReaMDB. This reemphasizes the need for a more objective and authoritative reference resource available to all parties, providing an association of mutations with resistance and links to a knowledge base of supporting evidence.

Toward routine use of WGS for MDR or XDR-TB.

The impact of WGS in routine settings has been discussed by Köser et al. (41), and our study supports the use of this approach in cases of highly drug-resistant TB. As observed in this study, implementing WGS within current diagnostic pipelines would enable a comprehensive detection of the presence, or indeed the absence, of mutations strongly associated with resistance. Although, as discussed elsewhere, the reliability of predicting resistance phenotype from the genotype is variable, this will greatly improve with the completion of and public access to larger association studies. Evidence of transmission and the confirmation of contact cases are other advantages of the routine implementation of WGS, as illustrated by the results for an epidemiologically confirmed contact in our study. Placing individual isolates in the wider context of a local hospital, the national or international collection of isolates would also be invaluable and easily achievable to allow transmission to be observed on a larger geographic scale.

The use of lineage-specific deletions and SNPs allows the isolates to be placed within the six major geographic M. tuberculosis lineages (35). This is useful, because these data are stable and therefore give information for comparisons between strains and strain collections. While it is unclear that these lineage designations have any clinical relevance, there have been reports that Beijing strains are more virulent (42) or acquire antibiotic resistance more easily (43). A more detailed phenotype-genotype analysis will eventually identify subgroups that have properties that are clinically relevant; for example, Casali et al. (12) recently described two Beijing subclades (clades A and B) that were prevalent in southern Russia, with clade A appearing to be spreading particularly rapidly. Other molecular typing data can be generated using WGS, most notably spoligotype data, using a de novo assembly of the direct repeat (DR) region (44). While this is less discriminatory or informative than SNPs, it allows interrogation and linkage to spoligotype databases. Mycobacterial interspersed repetitive unit (MIRU) typing, a form of variable-number tandem-repeat (VNTR) typing specific for M. tuberculosis strains, is more problematic for WGS, because the ability to determine the numbers of repeats of all informative MIRUs used for typing (45) will depend on increased read length.

The utility of WGS in a clinical context will depend not only on the information provided but also on the timing of results in relation to other data available. Figure 3 gives an overview of the current timelines in our hospital for standard testing. Sputum microscopy is routine, and the GeneXpert MTB/RIF test is used on all suspected TB cases, providing rapid confirmation that M. tuberculosis is present and an indication of whether the detected strains are MDR. While a GenoType MTBDRsl test may indicate resistance to FLQs, the most important second-line drugs for clinical management, this is not routinely undertaken locally due to costs relating to the small number of positives and the expiration of unused kits. Clearly, both the GeneXpert MTB/RIF and GenoType MTBDRsl tests are rapid diagnostics; however, these identify only a small number of resistance mutations. The more comprehensive nature of WGS has the advantage of identifying all possible mutations as well as providing the added value of phylogenetic surveillance. Routinely, liquid cultures are sent for DST at the reference laboratory as soon as they are culture positive (usually 10 to 20 days after inoculation) and are returned after a further 2 to 3 weeks with breakpoint information as to whether the cultures are susceptible or resistant to 13 antibiotics. As this is 4 to 6 weeks after a patient has been admitted, there is a significant window for the provision of clinically useful information before DST results are available. Figure 3 indicates that WGS already has the potential to provide useful information some weeks before DST results are reported, or even earlier if WGS results can be achieved directly from sputum samples. We are not suggesting that DST be curtailed, but rather that with critical cases, any delay in obtaining a broader resistance profile has clinical importance. One area in which it already could play a major role is for the drugs EMB and PZA, for which phenotypic DST is less reliable (27). PZA has a high rate of hepatotoxicity, especially in older patients, and so a WGS result showing a stop mutation in pncA might help support a DST result and enable a clinician to stop this treatment sooner.

FIG 3.

Expected time frame for receiving results for each test following sample collection.

The cost of WGS has plummeted over recent years, and while the cost per sample will vary depending on the throughput and equipment, the overall cost for WGS is probably within the range of that of other standard tests carried out in a large hospital. With the potential for WGS to provide useful information 4 weeks before DST results are available, there is a clear potential cost benefit for a disease, such as TB, that has significant public health risks and is extremely expensive to manage using multiple drugs over a long time period. By ensuring effective treatment at the earliest possible stage, any reduction in use of hospital isolation units or length of drug treatment would save considerably more than the cost of utilizing WGS. However, we have only anecdotal data to support this intuitive opinion, and it would be interesting to factor it into cost-benefit analyses, such as that by Resch et al. (46), who studied the cost effectiveness of different approaches to treating MDR-TB.

In conclusion, in the six cases we addressed, we provide evidence that WGS can inform and affect clinical decision-making, both by increasing confidence in a course of treatment and also by adapting medication on the basis of this information. DST and WGS can both play important and complementary roles in identifying antibiotic resistance patterns, and WGS can provide this information several weeks earlier than DST. Clearly, WGS will play an important and cost-effective role if used routinely for highly resistant cases of TB in a hospital environment in the United Kingdom.