Abstract
Tuberculosis (TB) is the leading cause of death by an infectious disease, and global TB control efforts are increasingly threatened by drug resistance in Mycobacterium tuberculosis (M. tb). Unlike most bacteria, where lateral gene transfer is an important mechanism of resistance acquisition, resistant M. tb arises solely by de novo chromosomal mutation. Using whole genome sequencing data from two natural populations of M. tb, we characterized the population genetics of known drug resistance loci using measures of diversity, population differentiation, and convergent evolution. We found resistant sub-populations to be less diverse than susceptible sub-populations, consistent with ongoing transmission of resistant M. tb. A subset of resistance genes (“sloppy targets”) were characterized by high diversity and multiple rare variants; we posit that a large genetic target for resistance and relaxation of purifying selection contribute to high diversity at these loci. For “tight targets” of selection, the path to resistance appeared narrower, evidenced by single favored mutations that arose numerous times on the phylogeny and segregated at markedly different frequencies in resistant and susceptible sub-populations. These results suggest that diverse genetic architectures underlie drug resistance in M. tb, and combined approaches are needed to identify causal mutations. Extrapolating from patterns observed in well-characterized genes, we identified novel candidate variants involved in resistance. The approach outlined here can be extended to identify resistance variants for new drugs, to investigate the genetic architecture of resistance, and, when phenotypic data are available, to find candidate genetic loci underlying other positively selected traits in clonal bacteria.
Importance Mycobacterium tuberculosis (M. tb), the causative agent of tuberculosis (TB), is a significant burden on global health. Antibiotic treatment imposes strong selective pressure on M. tb populations. Understanding causative and compensatory mutations for drug resistance in M. tb is important for treatment of TB infections and controlling the increasing prevalence of drug resistance. Whole genome sequencing (WGS) can be used to identify novel loci mediating drug resistance and predict resistance patterns in the clinic faster than traditional methods. We have used WGS from natural populations of drug resistant M. tb to characterize the effect of positive selection on patterns of diversity at known resistance mediating loci. These methods can be used to identify novel mutations under positive selection, including resistance loci, in M. tb and other clonal pathogens.
Introduction
Mycobacterium tuberculosis (M. tb), the causative agent of tuberculosis (TB), is estimated to have caused 1.4 million deaths in 2015, making it the leading cause of death due to an infectious disease. The proportion of TB due to MDR (M. tb resistant to first line anti-tuberculosis drugs isoniazid and rifampin) is increasing (1), which poses a major threat to global public health. Unlike most bacteria, M. tb evolves clonally, so resistance cannot be transferred among strains or acquired from other species of bacteria: drug resistance in M. tb results from de novo mutation within patients and transmission of drug resistant clones (2–4). The relative contributions of de novo emergence and transmitted drug resistance varies across sampling locations (5–9). Seven distinct lineages have been identified among globally extant populations of M. tb. Among these, lineage 2 (L2) has been associated with relatively high rates of drug resistance, and it has been postulated that the acquisition of resistance is a result of higher rates of mutation in this lineage (10). Studies of M.tb evolution within hosts with TB have shown that emergence of drug resistance is associated with clonal replacements that lead to reductions in genetic diversity of the bacterial population (11, 12).
Many of the methods developed to identify advantageous mutations, such as those conferring antibiotic resistance, depend on recombination to differentiate target loci from neutral variants (13). However, in clonal organisms like M. tb, neutral and deleterious mutations that are linked to advantageous variants will evolve in tandem with them. Linkage among sites can also cause competition between genetic backgrounds with beneficial mutations, decreasing the rate of fixation for beneficial alleles, while deleterious alleles are purged less efficiently (14–16).
While the majority of the M. tb genome is subject to purifying selection that purges variation (3), antibiotic pressure exerts strong positive selection on advantageous variants that confer resistance. M. tb drug resistance has been the focus of extensive investigation, and a variety of resistance mutations have been characterized for commonly used anti-tuberculosis drugs (17). Drug resistance mutations can be associated with fitness costs (18–20), and compensatory mutations that ameliorate these fitness costs have been identified in the context of rifampicin resistance (21, 22). Resistance mutations found to have lower fitness costs in vitro - as measured by competition assays - are found at higher frequencies among M. tb clinical isolates and appear to be transmitted at higher rates relative to mutations with high in vitro fitness costs (18, 23). Candidate loci involved in resistance and compensation for its fitness effects have been identified previously by screening for homoplastic variants that are significantly associated with drug resistant phenotypes (24) and genes with higher dN/dS in resistant compared to sensitive isolates (25). Application of these methods to whole genome sequence data from M. tb clinical isolates has recovered known drug resistance loci, as well as loci associated with cell surface lipids and biosynthesis, DNA replication, and metabolism.
The goal of the present study was to use patterns of genetic diversity at known drug resistance loci to identify the population genomic signatures of positive selection in natural populations of M. tb. Using whole genome sequence data from two populations for which phenotypic resistance data were available, we have identified several distinct signatures associated with these loci under selection. Based on these results, we propose methods of identifying loci under positive selection, including novel drug resistance loci, in clonal bacteria such as M. tb.
Results
We inferred the phylogeny of 1161 M. tb isolates from Russia and South Africa (see Methods, Supplementary Table 1) using the approximate maximum likelihood method implemented in FastTree2 (Figure 1). The majority of the isolates belong to L2 (n = 667) and lineage 4 (L4, n = 481). Overall diversity of L2 was lower than L4 in our sample (Figure 2, p < 2.2 × 10−16). Seven hundred and sixty two of the isolates in our sample (66%) are resistant to one or more anti-tuberculosis drugs (Table 1).
Drug resistant TB can be acquired as a result of de novo mutations within a patient or by infection with a resistant strain. When resistance is primarily mediated by de novo mutations, diversity should be similar in susceptible and resistant populations as resistance will arise on multiple genetic backgrounds. By contrast, if resistance develops primarily via transmission of resistant strains, the resistant sub-population should be less diverse than the susceptible sub-population. We compared the diversity of susceptible and resistant sub-populations and found genome wide estimates of nucleotide diversity to be higher in isolates susceptible to a range of drugs for which phenotyping data were available (paired t-test, p = 0.029). In comparisons of gene-wise diversity in susceptible and resistant populations, we found that resistant isolates had a greater number of genes with no diversity, but levels of diversity within genes in which it was measurable were similar between resistant and susceptible populations (Figure 3).
One hundred nine genes out of 3,162 unmasked genes were invariant across all isolates in this sample, likely a result of strong purifying selection. An additional 136 genes harbored variation in the drug susceptible populations but were invariant across all of the drug resistant populations. We did not observe any genes that were invariant in susceptible isolates specifically, further supporting the conclusion from genome wide diversity estimates that resistant isolates represent a subset of the diversity found in susceptible populations. In order to investigate potential biases in the genes that lose diversity in the setting of resistance and evaluate whether the observed pattern of shared zero diversity genes was likely to arise by chance, we performed weighted random sampling of genes. This was weighted based on diversity in susceptible populations, assuming that genes with low diversity in susceptible populations are more likely to be invariant in resistant populations. After sampling each drug resistant population 1000 times, we found that no samples contained shared genes amongst all populations of isolates resistant to first and second line drugs. This suggests that specific genes tend to lose diversity in the setting of drug resistance, which could result from purifying selection specific to this setting. However, in our data set, drug resistant populations are not independent and isolates are often resistant to multiple drugs. We therefore repeated the sampling with only first line drugs and found that the maximum number of sampled genes shared across all populations was 2 (median 0). While these results suggest that certain genes are more likely to lose diversity as drug resistance evolves, we cannot completely rule out the possibility that the pattern arose as a result of overlapping membership in resistant populations.
We compared diversity of drug resistance associated genes (Table 2) with the rest of the genome. We found gid, rpsL, and pncA to be in the top 5th percentile of gene-wise π and/or θ values. rrs and ethA are in the top 5th percentile of θ, but not π. Surprisingly, despite being a target of multiple drug resistance mutations (Table 2), we did not identify extreme levels of diversity in katG (80th and 82nd percentile of π and θ, respectively).
We also examined gene-wise diversity values within each lineage to look for lineage specific high diversity genes. In both L2 and L4, rpsL, pncA, gid, ethA, and thyA were in the top 5th percentile of diversity (π and/or θ). In L2, rpoB, embB, Rv1772, and folC were additionally in the top 5th percentile of gene-wise π and/or θ values. In L4, Rv0340 was in the top 5th percentile of gene-wise π and/or θ. While rpoB and embB were not in the top 5th percentile of gene-wise θ in L4, they still had high diversity (91st and 82nd percentile, respectively). The lineage specific differences in diversity of Rv1772, folC, and Rv0340 suggest that there are interactions between these loci and loci that differentiate L2 and L4.
We used gene-wise estimates of Tajima’s D to investigate gene specific skews in the site frequency spectrum that could result from selection. We previously identified a relationship between gene length and gene-wise estimates of Tajima’s D for M. tb (26), and this finding was corroborated here (R2 = 0.3 of TD with gene length after log2 transformation). In order to identify genes with extreme values of TD - out of proportion with their length - we performed linear regression on log2 transformed gene lengths and Tajima’s D values and identified genes with the largest residuals (Figure 4). pncA, ethA, and embC all had Tajima’s D values much lower than expected based on their length (5th percentile of residual values). This indicates that these genes contain an excess of rare variants compared to other genes in the genome.
We calculated the ratio of π and θ of resistance associated genes in isolates susceptible and resistant to first line drugs and identified genes with markedly different diversities in resistant and susceptible sub-populations (Figure 5A). Among resistance genes in the top 5th percentile of gene-wise π and θ overall, diversity of pncA and ethA is relatively high among resistant isolates, whereas diversity of gid is similar in resistant and susceptible populations. We also examined differences in this ratio between isolates in L2 and L4 (Figure 5B). Rv1772 and embR were more diverse in resistant isolates in L2, and kasA and tlyA were more diverse in resistant isolates in L4.
We used FST outlier analysis to identify SNPs and indels that exhibited extreme differences in frequency between susceptible and resistant populations. Our a priori expectation was that variants mediating resistance would be at markedly higher frequency in the drug resistant sub-population and that genes harboring variants with high FST would be enriched for drug resistance genes. After removing SNPs in regions corresponding to indels and variants at sites missing data for greater than 5% of isolates, the highest FST SNPs in comparisons of resistant and susceptible sub-populations to first line drugs are in katG (2155168, FST = 0.89, INH), rpoB (761155, FST = 0.72, RIF), and rpsL (781687, FST = 0.37, STR). These SNPs were also FST outliers in the lineage specific analyses. We used a randomization procedure to assess the significance of observed FST values and found the maximum FST values after randomly assigning resistant and susceptible designations to be 0.023 for INH, 0.019 for RIF, and 0.018 for STR. In addition to SNPs within known drug resistance associated genes, we identified FST outliers in genes that may be novel targets for drug resistance (Table 3).
Homoplastic SNPs – i.e. SNPs that evolve more than once on a phylogeny – are candidate loci under positive selection and have previously been used to identify resistance associated mutations in M. tb (24). We identified homoplastic SNPs and indels in our sample. Of 235 genes with homoplastic SNPs, 13 are known to be associated with drug resistance (Figure 6). pncA had the largest number of homoplastic SNPs of any gene in the genome (n = 27 distinct SNPs that appear > 1 on the phylogeny). Drug resistance genes were significantly enriched among genes with homoplastic SNPs (Fisher’s Exact Test, p = 1.2 × 10−4). The SNPs identified in FST analysis were also identified as homoplastic (Figure 6). Our results suggest that complementary approaches based on homoplasy and FST outlier analysis can be used to identify SNPs associated with a trait of interest (in this case drug resistance). In addition to genic SNPs, we observed homoplastic SNPs that are also FST outliers in intergenic regions upstream of drug resistance associated genes (Table 3). These are candidate resistance and compensatory mutations with a regulatory mechanism of action.
In our analyses of indels, we controlled for the possibility that indels affecting the same gene may not be called in exactly the same position by considering indels within same gene as identical. We identified four drug resistance associated genes with homoplastic indels: gid, ethA, rpoB, and pncA. All of these genes except rpoB harbored indels with identical starts and stops. FST values for the deletion in gid were in the top 5th percentile for CAP, EMB, Et, K, OFL, and PZA resistant populations, but, interestingly, the deletion was not associated with STR resistance (FST = 0.04). Unlike homoplastic SNPs, homoplastic indels were not significantly enriched for drug resistance associated loci (p = 1).
We recovered 20 out of 40 known drug targets by identifying genes with extreme values of diversity, homoplastic SNPs, or SNPs that are FST outliers in comparisons of resistant and susceptible subpopulations. All genes with both extremely high diversity (top 5th percentile) and homoplastic mutations were drug resistance associated (i.e. gid, ethA, pncA, and rpsL). We identified 67 genes with high diversity and Tajima’s D values more negative than expected based on gene length; only two of these were associated with drug resistance (i.e. ethA and pncA). Twenty out of of 51 homoplastic SNPs that are also FST outliers fall within or upstream of known drug resistance associated genes. The remaining SNPs may be false positives or novel drug resistance mutations.
Discussion
Highly virulent bacterial pathogens such as M. tb, Yersinia pestis (27), Francisella tularensis (28), and Mycobacterium ulcerans (29) appear to evolve clonally, i.e. with little to no evidence of lateral gene transfer. Positively selected loci in these organisms are likely to be associated with important phenotypes such as drug resistance, heightened transmissibility, or host adaptation. However, few methods are available for identifying loci under positive selection in the setting of limited recombination. We adopted an empirical approach to this problem, and used natural population data to characterize patterns of diversity at loci known to be under positive selection in M. tb.
In this analysis of M. tb diversity among isolates from settings with endemic drug resistance, we found M. tb nucleotide diversity to be similar to previous estimates from a globally distributed sample (26) and confirmed that diversity of M. tb is low relative to other bacterial pathogens (4). We identified lineage-specific patterns in overall diversity, with L4 having higher diversity than L2 (πL2: 3.6 × 10−5, πL4: 1.5 × 10−4). Previously published analyses of whole genome sequence data from L2 indicate that the majority of L2 isolates worldwide belong to a sub-lineage that has undergone relatively recent expansion (30). In our sample from Russia and South Africa, the majority of L2 isolates belong to this sub-lineage, while the L4 isolates are associated with deeper branching sub-lineages. This likely contributes to the observed patterns of diversity.
We found genome-wide diversity to be higher in susceptible M. tb sub-populations relative to those resistant to first- and second-line drugs, except in comparison with PRO and MOX resistant isolates. The observation of higher diversity in drug susceptible populations is consistent with a significant role for transmitted resistance in the propagation of drug resistant M. tb. A recent study of extensively drug resistant (XDR) M. tb infection in South Africa concluded that XDR cases result primarily from transmission of resistance, rather than de novo evolution of resistance mutations during infection (9). The primary studies for the sequence data analyzed here also identified clusters of drug resistant isolates (5, 6), suggesting that resistant isolates were being transmitted. Our results, along with these previously published observations, suggest that the fitness of drug resistant isolates is high enough to allow them to circulate in endemic regions. These populations have a high burden of drug resistant TB, and the role of transmitted drug resistance may differ in other settings.
An alternative – but not mutually exclusive – explanation for the observation of higher diversity in susceptible populations is that drug resistant M. tb is under distinct evolutionary constraints that reduce average genome-wide levels of diversity. In support of this hypothesis, we identified a specific subset of genes that were invariant across drug resistant populations and found that this pattern was unlikely to have arisen by chance. Interestingly, while average diversity was lower for resistant sub-populations, the gene-wise diversity distributions had heavier tails, indicating there were more genes with extreme levels of diversity.
We identified a specific subset of genes conferring resistance that were more diverse in resistant compared with susceptible populations, which could result from diversifying selection on drug targets. For example, a large genetic target for resistance would potentially be reflected in high levels of diversity in associated genes. Relaxation of purifying selection may also contribute to high levels of diversity in resistance genes. Extreme patterns of diversity were not exclusive to antibiotic targets, and high diversity in other genes could reflect diversifying selection for compensatory mutations in drug resistant populations, and/or relaxation of purifying selection in the setting of drug resistance.
We found the genetic architecture of resistance to vary among targets, and resistance-associated genes tended to fall within categories that we term “sloppy”, “tight” and “hybrid” targets of selection (the latter has a combination of tight and sloppy features and applies to rpsL, embB, and rpoB). “Sloppy” resistance genes are characterized by high levels of diversity. Genes associated with PZA, EMB, Et, and STR resistance (i.e. pncA, gid, rpsL, rrs, ethA) have high levels of diversity; some also had an excess of rare variants (pncA, ethA, embC). The finding that these genes accumulate multiple, individually rare mutations implies that there is a large genetic target for resistance and/or compensatory mutations within the gene. pncA also contains the highest number of homoplastic SNPs (27 SNPs appeared more than once on the phylogeny) of any gene in the data set. Among the 62 non-synonymous pncA mutations in our dataset, 55 have been previously reported in association with drug resistance (TB Drug Resistance Mutation Database (31)). The newly described SNPs may mediate drug resistance or compensation for the fitness effects of other variants. Relaxed purifying selection could play a role in accumulation of diversity in pncA and other sloppy targets. An M. tb strain harboring a deletion in pncA conferring resistance to PZA was estimated to be endemic in Quebec by 1800, long before the use of PZA for the treatment of TB (32–34). This suggests that purifying selection on pncA is relatively weak, which could contribute to its exceedingly high diversity and broaden the adaptive paths to resistance.
In contrast to pncA, gid, which is associated with low level STR resistance (35), does not appear to have the signatures of a “sloppy” target for resistance despite its high diversity. We identified just three homoplastic SNPs within gid, and previous studies have found that STR resistant isolates do not encode the same gid mutations (36). This could indicate that a multitude of mutations within gid confer resistance, but levels of diversity in the gene were similar in resistant and susceptible isolates. Previous studies of sequence polymorphism in gid have identified high diversity in this gene in both resistant and susceptible isolates (36–38). gid appears to be subject to relaxed purifying selection in the presence and absence of antibiotic pressure.
We found some drug targets to be highly diverse in resistant sub-populations of either L2 or L4 (but not both), suggesting that there are interactions between resistance mutations in these genes and the genetic background. FST outliers specific to each lineage may also have lineage specific roles in drug resistance (Table 3). Epistatic interactions between drug resistance mutations and M. tb lineage have been reported previously: for example, specific mutations in the inhA promoter have been associated with the L1 and M. africanum genetic backgrounds (39, 40).
In contrast to “sloppy” targets, we discovered individual homoplastic SNPs associated with drug resistant sub-populations (i.e. with high FST) representing “tight” targets of selection in genes conferring resistance to INH, RIF, and STR. Numerous resistance mutations have been described in katG, rpoB, and rpsL, embB, and gyrA, but we find drug resistant sub-populations to be defined by a specific subset of mutations in these genes. This suggests that certain mutations are strongly favored relative to others conferring resistance to the same drugs when M. tb is in its natural environment. Antibiotic resistance can impose fitness costs on M. tb during in vitro growth, with the range of fitness costs varying among mutations, even within the same gene (18). Mutations can also have different fitness effects depending on the genetic background, but the most fit mutants were the same across M. tb lineages in a study of RIF resistance (18).
In our analyses, we found the dominant INH resistance mutation in katG to affect the serine at position 315. This change reduces affinity to INH but preserves catalase activity (41), and is associated with lower fitness costs than other katG mutants, both in vitro and in a mouse model (42, 43). This mutation was recently shown to precede mutations conferring resistance to other drugs during accumulation of resistance in evolution of multi-drug resistant M. tb (44). The dominant mutations we identified in rpoB (codon 450) and rpsL (codon 43) have also been found to have lower fitness costs in vitro compared to other mutations conferring resistance to RIF and STR in these genes (18, 43, 45). These results suggest that many of the findings regarding the relative fitness costs of M. tb resistance mutations in vitro and in animal models are relevant to the pathogen’s natural environment.
While the fitness effects of mutations in gyrA (codon 94) and embB (codon 306) have not been measured, based on our homoplasy and FST results, we hypothesize that they are likely to have lower fitness costs than other mutations in these genes and that they represent “tight” targets of selection. Mutations at gyrA codon 94 were previously found to be the most prevalent in a survey of gyrA and gyrB mutations in fluoroquinolone resistant M. tb clinical isolates (46). Interestingly, the mutation in embB codon 306 has been previously associated with acquisition of multiple resistances (47), and we find that this position is an FST outlier for all first line drugs in L4. This mutation is not an FST outlier in L2 (i.e top 5th percentile), with percentiles for FST values ranging from 0.07-0.68 for first line drugs in this lineage. Our results suggest that the genetic background affects interactions among resistance mutations, and that embB 306 is important for acquisition of multidrug resistance in L4 but not L2.
We searched for indels with the signature of a “tight” target, i.e. homoplastic mutations segregating at markedly different frequencies in drug susceptible and resistant sub-populations. Unlike the pattern observed with SNPs, genes associated with drug resistance were not significantly enriched among those harboring homoplastic indels. We identified one homoplastic indel that was also an FST outlier - a deletion in gid that causes a frameshift. Patterns of variation in gid are complex and suggest a role for relaxation of purifying selection (i.e. in the accumulation of excess SNPs in both resistant and susceptible isolates) and perhaps a tight target associated with multi-resistance (i.e. this homoplastic/FST outlier deletion that was associated with resistance to CAP, EMB, Et, K, OFL, and PZA).
Isolates harboring the SNPs that define “tight” targets in rpoB, katG, and rpsL have previously been found to compete successfully with wild type strains during both in vitro and in vivo competition assays (42, 43, 48). Our finding that, save for the frameshift mutation in gid, indels in resistance genes do not have the signature of “tight” targets suggests that they are generally associated with higher fitness costs than SNPs. Fifteen drug targets are in genes found in transposon mutagenesis experiments to be essential for M. tb growth in vitro, such as rpoB and rpsL; deletions in these genes are likely to interrupt important functions (49). Deletions in non-essential genes could also have fitness costs. Deletions in katG, which is non-essential, can result in INH resistance but they are not observed as frequently in clinical isolates as the KatG S315 SNP, particularly among transmitted INH-resistant strains (23).
There are several limitations to our study. Resistance to multiple drugs was common in our sample, and in some cases it was difficult to identify patterns of diversity and population differentiation that were specific to individual drugs. Our results are also limited by the accuracy with which drug resistance phenotypes were determined and limited phenotypic data for some drugs (particularly second line drugs). For example, misclassification could affect gid, which mediates low level resistance to streptomycin: isolates with gid mutations may have been mis-classified as susceptible when they in fact harbor low level resistance. The approaches described here could be extended using data sets enriched for resistance to particular drugs of interest, in order to characterize the genetic basis of resistance to a broader range of TB treatments.
Our sample was heavily skewed to lineages 2 and 4, and the results are not necessarily applicable to other M. tb lineages. Finally, the data analyzed here were generated with short sequencing read technologies, and we were thus limited to characterizing diversity in regions of the M. tb genome that can be resolved with these methods. Masked portions of the genome may include unknown resistance targets.
We were not able to recover all drug resistance associated genes using the analyses performed here. For some genes, this is likely a result of limited phenotypic data (i.e. thyA and folC, which are associated with aminosalicylic acid (PAS) resistance). Our list of known drug targets was dominated by genes associated with INH resistance, and signatures in these genes may not be as obvious due to the high frequency of the KatG S315 mutation in drug resistant populations.
We identified 31 homoplastic SNPs that are also FST outliers in at least one drug resistant population that do not fall within the list of known drug resistance genes. These SNPs may be novel resistance determinants; notably, all non-synonymous SNPs fall within genes linked with drug resistance in other studies (i.e. efflux pumps, differentially regulated in resistant isolates or in response to the presence of drug, potential drug targets, connected to drug targets or resistance determinants) (50–54).
Here we have used drug resistance loci in M. tb to identify the signatures of positive selection in a clonal bacterium. We found these loci to be associated with distinct patterns of diversity that likely reflect differing genetic architectures underlying the traits under selection. The evolutionary path to resistance is broad for some drugs with “sloppy targets”, whereas for drugs with “tight targets” the means of acquiring resistance appear more limited. This is likely due to fitness effects of resistance mutations in M. tb’s natural environment, as numerous resistance mutations have been identified in tight target genes. We also found evidence suggesting that there are important interactions among loci during the evolution of resistance. Our results suggest that purifying selection on a subset of genes intensifies in the setting of resistance, which could reflect epistatic interactions and/or a response to the metabolic milieu imposed by antimycobacterial agents. The results presented here can be used to create more realistic models of resistance evolution in M. tb and to develop novel strategies of preventing or mitigating the acquisition of resistance. For example, the narrow path to resistance for drugs with tight targets reveals potentially exploitable vulnerabilities, as does the finding of interdependencies among specific loci and the genetic background in the evolution of resistance and multi-resistance. As new TB drugs become available for clinical use, the approach outlined here can be extended to understand their architectures of resistance.
Efforts are underway to sequence and perform drug susceptibility testing on thousands of M. tb isolates with the goal of creating an exhaustive catalogue of drug resistance mutations and eventually using WGS to diagnose drug resistance in clinical settings (CRyPTIC project, http://modmedmicro.nsms.ox.ac.uk/cryptic/, last accessed: May 24, 2017). We found that loci under positive selection can be identified using relatively simple methods: “tight” targets are highly differentiated in their allele frequencies across phenotypic groups (i.e. FST outliers) and appear as homoplasies on the phylogeny; “sloppy” targets are characterized by high diversity and/or low Tajima’s D, as well as homoplasies. Extrapolating from patterns observed among known resistance variants, we have discovered new candidate regulatory and genic resistance variants. The methods used in this study are widely available and should scale to analysis of the large collections of genomic and phenotypic data that are currently being generated. This approach can be extended to identify novel resistance loci in bacteria for which drug susceptibility phenotypes are defined, as well as other positively selected loci in clonal bacterial populations.
Intriguingly, we found lipid metabolism genes to be enriched in the list of genes harboring homoplastic SNPs (p = 0.013). We’ve previously shown that these genes have extreme values of diversity in a global sample of M. tb isolates and within individual hosts (26), suggesting that lipid metabolism genes are also under positive selection in M. tb populations. In the case of genes associated with lipid metabolism, the results presented here could be extended by phenotypic characterization of lipid profiles and identification of homoplastic variants that are at markedly different frequencies in isolates with distinct lipid profiles.
Methods
Reference guided assembly
We downloaded sequencing read data from two large surveys of drug resistant M. tb in Russia (5) and South Africa (6). We used FastQC (55) and TrimGalore (56) for quality assessment and adaptor trimming of the reads. Trimmed reads were mapped to M. tb H37Rv (NC_000962.3) using BWA-MEM v 0.7.12 (57). We used Samtools v 1.2 (58) and Picard Tools (https://broadinstitute.github.io/picard/) for sorting, format conversion, and addition of read group information. Variants were identified using Pilon v 1.16 (59). A detailed description of the reference guided assembly pipeline is available at https://github.com/pepperell-lab/RGAPepPipe. We removed isolates with mean coverage less than 20X, isolates with percentage of the genome covered at 10X less than 90%, isolates where a majority of reads did not map to H37Rv, and isolates where greater than 10% of sites were unknown after mapping. The final dataset contains 1161 M. tb isolates. The alignment was masked to remove repetitive regions including PE/PPE genes.
Phylogenetic analysis
We estimated the phylogeny using the masked alignment from reference guided assembly with FastTree-2.1.9 (60). We compiled FastTree using the double precision option to accurately estimate branch lengths of closely related isolates. We used FigTree (http://tree.bio.ed.ac.uk/software/figtree/) for tree visualization.
SNP annotation
A VCF of single nucleotide variants was created from the masked alignment using SNP-sites v 2.3.2 (61). SNPs were annotated using SnpEff v 4.1j (62) to identify synonymous, non-synonymous, and intergenic SNPs.
Indel identification
Insertions and deletions were identified during variant calling with Pilon. We used Emu (63) to normalize indels across multiple isolates. We used a presence/absence matrix for the normalized indels for further analyses of indel diversity.
Population genetics statistics
Whole genome and gene-wise diversity (π and θ) and neutrality (Tajima’s D) statistics were calculated using Egglib v 2.1.10 (64) for whole genome alignments and genewise alignments. Isolates were further divided by lineage and drug resistance phenotype. Sites with missing data due to indels or low quality base calls more than 5% of isolates in the alignment were not included in calculation of statistics. Values of Tajima’s D showed a correlation with gene length in our sample. To find genes with extreme values of Tajima’s D, we performed linear regression in R (65) on log transformed Tajima’s D values and gene length and identified genes with large residual values. Weir and Cockerham’s FST (66) was calculated using populations of resistant and susceptible isolates for each drug using vcflib v1.0.0-rc0-262-g50a3 (https://github.com/vcflib/vcflib). For non-biallelic SNPs, we calculated FST for the two most common variants.
Homoplasy
We used TreeTime (67) to perform ancestral reconstruction and place SNPs and indels on the phylogeny., we identified homoplastic SNPs and indels.
Data availability
Unless otherwise noted, all data and scripts associated with this study are available at https://github.com/pepperell-lab/mtbDrugResistance.
Funding information
This material is based upon work supported by the National Science Foundation Graduate Research Fellowship Program under Grant No. DGE-1256259 to TDM. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation. TDM is also supported by National Institutes of Health National Research Service Award (T32 GM07215). CSP is supported by National Institutes of Health (R01AI113287). Funding for this project was provided by the University of Wisconsin Madison School of Medicine and Public Health from the Wisconsin Partnership Program.
Acknowledgments
We thank members of the Pepperell Lab for their input on analyses and data visualization.