A retrospective genomic analysis of drug-resistant strains of M. tuberculosis in a high-burden setting, with an emphasis on comparative diagnostics and reactivation and reinfection status

While global rates of incidence of tuberculosis (TB) have been falling at an average of 2% per year between 2000 and 2014, the incidence of drug resistant (DR) TB remains steady [1]. Many Eastern European countries have high rates of multidrug-resistant (MDR) and extensively drug-resistant (XDR) tuberculosis and are included among the WHO High-burden countries lists [1]. MDR tuberculosis is defined as infections resistant to the first-line antibiotics rifampicin and isoniazid. XDR tuberculosis is defined as an infection that is MDR and additionally resistant to at least one drug in both of the two classes used to treat MDR: fluoroquinolones and the second-line injectable drugs (amikacin, capreomycin or kanamycin).

High burden and poor clinical or public health control of TB may also influence the nature of recurrent TB cases. Recurrent TB, defined as TB that occurs after a patient has been considered cured by standard TB treatment, may arise due to exogenous reinfection or endogenous reactivation [2]. Endogenous reactivation could be due to ineffective TB treatment and/or the emergence of drug-resistance within an individual [2, 3]. Reinfection from another individual is an example of exogenous reinfection. Recurrent TB cases are reported as “relapse” cases (i.e., retreatment cases after treatment success) by WHO standard definitions, which does not distinguish true relapses from reinfection cases. Understanding the extent to which either scenario (reinfection or reactivation/relapse) is occurring in recurrent TB infections allows for application of appropriate control measures.

Molecular approaches can help differentiate between reactivation or reinfection among recurrent TB cases [4, 5]. In combination with epidemiological, clinical, and laboratory data, they can also reveal insight into transmission dynamics of TB, including DR-TB. Whole genome sequencing of Mycobacterium tuberculosis (Mtb) isolates has an advantage over genotyping in that it can evaluate variation across the entire genome. It also provides a more comprehensive and precise analysis of phylogenetic relationships and transmission dynamics that can help determine whether recurrent TB cases are due to reinfection or reactivation.

The Republic of Moldova is a small Eastern European country with a high MDR-TB burden [1]. National drug resistance surveys between 2006 and 2017 showed that MDR-TB prevalence among new cases increased from 5.0% in 2000 to 26% in 2017, and among previously treated patients, increased from 33.2% up to 56% (all data from the National TB registry of Moldova) [6, 7]. Various economic and cultural reasons contribute to this situation [8‐10] which is exacerbated by long-term hospitalization and inadequate infection control measures in these facilities [11].

To understand the genomic basis of DR-TB in Moldova, we carried out whole-genome sequencing of Mtb isolates from 190 MDR (non-XDR), XDR, and drug-sensitive TB patients, identified and collected as part of the TB Portals Program [12]. Paired longitudinal samples from 87 of these patients with recurrent cases were used to characterize the nature of Mtb reactivation or reinfection in these cases.

We performed genomic analysis on 278 Mtb isolates (239 drug-resistant, 39 drug-sensitive, as determined by standard clinical drug susceptibility tests) from 190 TB patients. The majority of the isolates were from two spoligotypes in nearly equal distribution (41% H3 and 38% Beijing). All H3 spoligotypes had the 4.2.1 numeric SNP barcode and all Beijing spoligotypes had the 2.2.1 numeric SNP barcode. Not all 4.2.1 lineage samples had the H3 spoligotype. The remaining samples were distributed among nine other spoligotypes, all present at a frequency of less than 10% (Fig. 1a). An extremely similar spoligotype distribution, with the H3/4.2.1 and Beijing/2.2.1 spoligotypes being dominant, is seen in the drug-resistant samples (Fig. 1b). There is much more uniform spoligotype variation among drug-sensitive samples (Fig. 1c).

Bayesian phylogenetic analysis of genomic SNPs demonstrated differing evolutionary structures between Beijing/2.2.1 and H3/4.2.1 samples (Fig. 2). The majority of the H3/4.2.1 samples clustered into one shallow, strongly supported (clade posterior probability > 0.95) group (mean Hamming distance 17.1 SNPs, range 0–40 SNPs). Beijing/2.2.1 samples mostly clustered into four well-supported (P > 0.95) clades (Fig. 2) which were each more divergent than the main H3/4.2.1 clade (mean pairwise Hamming distances 40.2 (range 0–53), 30.0 (0–53), 35.3 (3–68), and 33.1 (21–31) SNPs). All H3/4.2.1 samples had a smaller mean pairwise diversity (mean Hamming distance 74.4 SNPs) than the Beijing/2.2.1 samples (mean Hamming distance 144.5 SNPs) but a greater range of individual pairwise diversity (Hamming distance range 0–495 for H3/4.2.1, 0–233 for Beijing/2.2.1).

Whole genome variant analysis (WGVA) was performed to reveal molecular patterns of drug resistance within our sample set. The presence or absence of drug-resistance SNPs, as determined by data in the ReSeqTB database (Table 1) [21], were mapped onto the tips of this phylogeny and are presented as a heat map to the right of the tree in Fig. 2. Our analysis shows the rpoB S450 L (rifampicin resistance), inhA promoter c(− 15)t (isoniazid resistance), and katG S315 T (isoniazid resistance) mutations broadly distributed across DR samples in both the H3/4.2.1 and Beijing/2.2.1 clades. The isolates in the H3/4.2.1 clade do not appear to have any mutations in the 16S ribosomal RNA gene rrs that would lead to resistance to second-line injectable drugs (amikacin, kanamycin, or capreomycin). Several widely phylogenetically distributed Beijing/2.2.1 samples have the rrs a1401g mutation. This distribution implies the rrs a1401g mutation has arisen independently multiple times among the divergent Beijing/2.2.1 isolates.

Susceptibility to first and second-line drugs was determined using phenotypic drug susceptibility testing (pDST) assays. Standard LPAs of drug resistance (Hain MTBDRplus and GenoType MTBDRsl and Cepheid GeneXpert) also were performed on many of these samples. All samples with a LPA also had a phenotypic test, and no LPA results were more severe than phenotypic DST results. There were 12 samples in which pDST-determined drug resistance did not match what was predicted by WGVA (Table 2). In 10 of these cases the WGVA drug resistance profile was “susceptible” while that from pDST was mono-resistant, MDR, or XDR. Two patients with drug susceptible Mtb infection, as determined by pDST, had WGVA profiles concordant with mono-resistance.

Seven of these samples were from paired samples. In two cases (Patient IDs 1247 and 1612) both samples in the pair exhibited genotype/phenotype mismatches. For the other three samples in all cases that later isolate had the genotype/phenotype mismatch. For Patient ID 1248 the paired samples had different spoligotypes while for the other two (Patient IDs 740 and 1242) the spoligotypes were the same.

This discrepancy between pDST and WGVA results could be an indication of mixed infection when there is evidence of low-level DR variants. For the 10 samples with a WGVA drug susceptible profile and a pDST resistant profile, the full genomic read data were examined to determine if low-frequency variants were present at the ReSeqTB DR SNP sites. Analysis with LoFreq [22] found no statistically significant low-frequency variation in these samples. Four of these samples (SRR3743495, SRR5153913, SRR3743493, and SRR3743482) had low level variation (greater than 1% of reads but not statistically significant) at several ReSeqTB DR loci.

The sample with the most low-level variation was SRR5153913 (Patient ID 740). This sample was classified as MDR by the BACTEC and Hain DSTs but it had no SNP calls at ReSeqTB DR SNP sites. This sample had low-frequency variation greater than 1% of reads at rpoB S450 L (4/107 reads), katG S315 T (4/142), inhH promoter c(− 15)t (4/158), rpsL K88R(4/195), and pncA G97D (2/139). These variants would allow subpopulations of the pathogen to grow in cultures containing rifampicin, isoniazid, streptomycin, and pyrazinamide (respectively), corresponding to the MDR profile.

We also examined depth of coverage as a possible explanation for the discrepancy between pDST and WGVA results. Three of the 10 samples with WGVA susceptible/pDST resistant profiles had low coverage (average number < 100 reads) across all of the ReSeqTB DR SNP sites. Due to sampling effects low coverage at DR SNP sites could lead to under-reporting of low-frequency DR SNPs that may be present in the patient. The low frequency variation found at ReSeqTB DR SNP sites in these three samples was on the order of 1% of reads or less.

Two samples (patient/isolate IDs 1242/SRR5153868 and 1248/SRR5153901) were identified as being WGVA-resistant and pDST sensitive (Table 2). Isolate SRR5153868 had the rpsL K43R variant for streptomycin resistance and isolate SRR5153901 had the gyrA A90V variant for ofloxacin resistance. Additionally, SRR5153868 had three DR variants present at very low frequency: rpoB L452P (2/177), pncA Q10P (2/159), and pncA H57D (2/192). Isolate SRR5153901 had one low frequency variant in the rpoB RRDR that was not a recognized DR variant. There are several potential sources for this disagreement, including differences in diversity between WGVA samples and pDST samples due to mixed infection in the original samples [28] and localized failure of individual DSTs.

We analyzed a range of drug-susceptible and resistant samples from 190 Moldovan TB patients to characterize the local structure of this epidemic. Our analysis presents three main findings: 1) phylogenetic evidence and network analysis suggest a high degree of person-to-person transmission of drug-resistant tuberculosis, 2) one strain is so prevalent that there is extremely little diversity among the patient samples, and 3) low frequency variants in several cases where there is disagreement between phenotypic drug resistance results and whole genomic analysis suggest a mixed infection.

Genomic analysis of local TB outbreaks have typically found one or more circulating strains with an extreme paucity of diversity [20, 30, 31]. These “clones” [20] are interpreted as evidence of rapid person-to-person transmission of a single, typically drug-resistant, lineage rather than the in vivo development of drug resistance over the course of treatment. This pattern of clonality occurs independently of the TB lineages that are present in the local outbreak, consistant with lineage not being a causal factor in this pattern. In Moldova, the H3/4.2.1 samples demonstrated this very shallow phylogenetic structure while the Beijing/2.2.1 samples had a more complex and deeper structure. These patterns are consistent with multiple Beijing/2.2.1 sublineages being responsible for infection in Moldova over the sampling period, while one main H3/4.2.1 sublineage predominated during this time. There were three smaller groups within the Beijing/2.2.1 samples that also were very shallow and closely related, consistent with limited person-to-person transmission of DR TB.

This shallow, highly unresolved phylogenetic structure in the H3/4.2.1 clade prompted a network analysis of these samples (Fig. 4). Several small, two sample, networks were found, in addition to two larger networks. The smaller of these consisted of five samples (two identical) in a straightforward reticulate network. The larger network was very complex with many reticulate paths among nodes. This pattern is consistent with a small number of closely related H3/4.2.1 strains leading to many of the H3/4.2.1 infections in Moldova. This pattern also makes the determination of reinfection or reactivation in the paired samples difficult to ascertain. For four of these pairs the potential of reinfection would not be obvious, as the length of time between samples and the number of SNPs were consistent with the Mtb genome substitution rate under selection pressure from antibiotic therapy [29]. In the paired samples with relatively fewer genomic changes (but still greater than neutral expectations), it is also possible that increased mutation rate due to selection from drug treatment and host immune pressure could result in more SNPs between samples than expected. Eldholm et al. [29] observed this among longitudinal samples from an infection that began as susceptible to first-line drugs and progressed to XDR.

The circulation of highly interrelated H3/4.2.1 sublineages in Moldova makes the determination of reactivation or reinfection cases problematic [3, 11, 32, 33]. Our network analysis demonstrates that a sample taken later in time, while having the same spoligotype and numeric SNP barcode, can be more closely related to samples from other individuals than the initial sample from the same patient. Conversely, paired samples with the same spoligotype and numeric SNP barcode but a large number of genomic SNPs between the samples are also evidence of reinfection that was previously classified as “relapse”. Together, our observations of high genomic similarity between circulating strains and large genomic variation between paired samples with the same lineage classification are evidence that person-to-person transmission of DR TB is probably much higher than currently recognized. To a limited extent similar patterns were found in Beijing/2.2.1 sublineages as well, indicating that the effect of circulating strain genealogical structure is not lineage specific.

A portion of samples exhibited discrepancy in drug resistance determination between pDST vs WGVA results, most having “susceptible” results by WGVA despite having been determined “resistant” by pDST. This discrepancy could be due to technical or biological causes. In four cases, low-frequency variants were found by WGVA at DR sites, though this variation was not statistically significant by LoFreq [22] analysis. Depth of coverage could also influence whether low-frequency DR variants were able to be detected at all in our analysis, as a portion of discordant samples had low coverage at DR SNP sites. It is also possible that drug resistance in these samples may be conferred by not-yet identified loci or variants. For example, several pDST-determined XDR samples do not carry SNPs at the rrs locus, so resistance may be due to secondary loci which have not been identified as having a strongly significant effect of second-line drug resistance, such as tlyA for capreomycin [34] and the eis promoter for kanamycin [35]. Our analysis also found low-frequency DR variants in one of the two cases with pDST-sensitive profiles. As these samples had one recognized DR variant each this suggests that drug resistance was not detected by culture-based methods due to growth and/or selection conditions. Finally, mixed infections could also lead to the pattern of lineage divergence found among the longitudinal samples. If treatment removes the susceptible majority lineage, subsequent samples could primarily consist of the minority lineage which would appear to be reinfection but would actually be a persistent infection.

Our study sought to better understand the impact of genomic diversity, evolution, and epidemiology of DR-TB on the ability to determine recurrence status of infections. The nature of the current DR TB outbreak in Moldova was well-suited for this analysis. Whole genome sequencing provided detailed information about the population structure of local circulating strains and the variety of strains infecting individual patients. Genomic sequencing provided strong evidence of a widespread clonal strain with very low diversity among these samples. For this population structure it is just as likely that individuals are being reinfected by a circulating closely related strain as it is that their initial infection was not entirely eradicated by treatment. Genomic sequencing also provided strong evidence that cases previously classified as relapse (based on locus-specific genotyping) were, in fact, reinfection due to the large genomic diversity between the paired longitudinal samples. While specific to our samples, these results provide some actionable insights into local TB epidemics and ways to control them. As more evidence accumulates that a significant portion of existing cases of DR TB are the result of reinfection, either by very similar or divergent (but having the same classification) strains, additional efforts must be put into reducing transmission, especially person-to-person transmission in clinical settings. Our results showing mixed infection in some patients demonstrates the importance of both phenotypic and genotypic methods for TB diagnosis and drug resistance testing. Genomic sequencing at greater depth will provide stronger evidence of existing low-level mixed infections. These data will give us a better understanding of the frequency of mixed TB infection and its impact on TB outcomes.