Pre-Existing Isoniazid Resistance, but Not the Genotype of Mycobacterium Tuberculosis Drives Rifampicin Resistance Codon Preference in Vitro

Both the probability of a mutation occurring and the ability of the mutant to persist will influence the distribution of mutants that arise in a population. We studied the interaction of these factors for the in vitro selection of rifampicin (RIF)-resistant mutants of Mycobacterium tuberculosis. We characterised two series of spontaneous RIF-resistant in vitro mutants from isoniazid (INH)-sensitive and -resistant laboratory strains and clinical isolates, representing various M. tuberculosis genotypes. The first series were selected from multiple parallel 1 ml cultures and the second from single 10 ml cultures. RIF-resistant mutants were screened by Multiplex Ligation-dependent Probe Amplification (MLPA) or by sequencing the rpoB gene. For all strains the mutation rate for RIF resistance was determined with a fluctuation assay. The most striking observation was a shift towards rpoB-S531L (TCG→TTG) mutations in a panel of laboratory-generated INH-resistant mutants selected from the 10-ml cultures (p<0.001). All tested strains showed similar mutation rates (1.33×10−8 to 2.49×10−7) except one of the laboratory-generated INH mutants with a mutation rate measured at 5.71×10−7, more than 10 times higher than that of the INH susceptible parental strain (5.46–7.44×10−8). No significant, systematic difference in the spectrum of rpoB-mutations between strains of different genotypes was observed. The dramatic shift towards rpoB-S531L in our INH-resistant laboratory mutants suggests that the relative fitness of resistant mutants can dramatically impact the distribution of (subsequent) mutations that accumulate in a M. tuberculosis population, at least in vitro. We conclude that, against specific genetic backgrounds, certain resistance mutations are particularly likely to spread. Molecular screening for these (combinations of) mutations in clinical isolates could rapidly identify these particular pathogenic strains. We therefore recommend that isolates are screened for the distribution of resistance mutations, especially in regions that are highly endemic for (multi)drug resistant tuberculosis.


Introduction
The emergence, spread and persistence of drug resistance inhibits the successful treatment and control of tuberculosis (TB).
In contrast to many other bacterial pathogens, the etiological agent of tuberculosis, Mycobacterium tuberculosis, does not acquire antimicrobial resistance via horizontally transferred plasmids or other mobile genetic elements, but almost exclusively via the acquisition of point mutations or, occasionally, through genomic deletions [1,2]. M. tuberculosis is thus a genetically isolated and clonal organism [1,3] and recently acquired mutations are passed on to the progeny, resulting in the accumulation of mutations in the genome. Whether these mutations are sustained in the bacterial population depends on chance, their relative fitness and subsequent selective sweeps.
Genetic characterisation of drug-resistant clinical isolates indicates that only a small range of mutations is responsible for the majority of resistance in clinical isolates; screening for only 3 to 5 mutations in the rpoB gene detects more than 80% of all clinical M. tuberculosis isolates with resistance to rifampicin (RIF), a critical component of any successful anti-TB treatment and a marker for multidrug resistance [4,5,6,7]. The majority of these mutations are located within an 81-bp region of rpoB, the gene that encodes the b-chain of RNA polymerase [8]. This mutational hotspot is therefore often referred to as the rifampicin resistance determining region. Virtually the same distribution of mutations conferring RIF resistance is seen in in vitro mutants [9,10,11,12,13]. RIF resistance is therefore often used as a proxy to study antibiotic resistance in the laboratory.
Not all rpoB mutations are equivalent; the implications for the bacteria, such as the level of RIF resistance or the consequential fitness, can vary considerably. Some mutations, like rpoB-S531L are found more often than others and detection of unusual or rare mutations in clinical isolates may therefore be indicative of distinctive circumstances.
In an earlier study, we found that acquisition of rifamycin resistance by a RIF-resistant, but rifabutin-susceptible strain, carrying an rpoB-S522L (TCGRTTG) mutation, was associated with a shift in the mutational spectrum outside of the rifampicin resistance determining region [9]. A possibly related effect was observed by others in a strain that caused an outbreak in London, which was resistant to isoniazid and became resistant to RIF via unusual mutations such as rpoB-V176F [14]. Also, a preference for specific drug resistance mutations in clinical isolates belonging to certain M. tuberculosis genotypes or strains obtained from specific geographical locations has been put forward [15,16,17,18].
Together these data suggest that the genetic background of the M. tuberculosis bacteria, such as the genotype or pre-existing drug resistance, influence the optimal evolutionary route of the bacteria [19,20,21]; certain combinations of mutations may be lethal or lead to very unfit organisms, requiring adaptive mutations to restore their fitness. Experiments performed in vitro suggest that rpoB mutations S531L (TCGRTTG) and H526D (CACRGAC) confer the lowest fitness deficit on the bacteria [8,10,11,19,22,23]. However, methods by which ''fitness'' is measured in different studies, and even how fitness is defined, are not standardised [23,24].
The predominance of certain rpoB mutations, in isolates of a specific bacterial genotype, is thus presumably due to a combination of the likelihood of each mutation occurring and its subsequent ability to survive and disseminate. Here we attempt to study the contribution and interaction of these two factors by determining the spectrum of spontaneous mutations in a wellcharacterised panel of laboratory and clinical strains using two different selection strategies. This panel includes strains representing genotypes from different geographical locations as well as several isoniazid-resistant isolates.
Spontaneous rifampicin-resistant mutants from genetically characterised strains were selected from bacterial populations obtained from two different culture strategies; Method 1 ( Figure 1) allowed us to observe the emergence of each targeted mutation with minimal competition between mutants, and with Method 2 ( Figure 2) we determined the frequency, and thereby the ''success'', of each mutation over time in a larger population. Mutants were characterized by Multiplex Ligation-dependent Probe Amplification (MLPA [25]), enabling detection of the RIFresistance conferring mutations in rpoB V176F (GTCRTTC), S522L (TCGRTTG), H526Y (CACRTAC), H526D (CACR GAC) and S531L (TCGRTTG). Resistant mutants obtained without these mutations had two regions of their rpoB gene sequenced using methods previously published [9,25].
Our results indicate that the (rapid) accumulation of drug resistance mutations can significantly reduce the subsequent spectrum of mutations. If so, these constrained pathways may For each strain 25 1-ml cultures were inoculated with approximately 1000 bacteria from a 10-ml starting culture after which they were incubated at 36uC in a shaking incubator. When the bacteria reached the mid-logarithmic phase (ca. three weeks), as determined by addition and colour development of the growth indicator resazurin, bacteria were transferred to rifampicin-containing solid medium to select for resistant mutants; the total contents of all 1-ml cultures were each plated on a single well of a 25-well plate, so that all 1-ml cultures from a single strain were plated on a single plate. The plates were then sealed and incubated at 36uC until sufficient bacterial growth of the mutants was visible (ca. four weeks). From each well, a DNA extract was made, so that one DNA sample contained DNA from any colonies that grew on the corresponding well. Finally, DNA samples were screened by MLPA [25] or sequencing of the rifampicin resistance determining region in rpoB. doi:10.1371/journal.pone.0029108.g001 also be applicable in vivo, which could facilitate the rapid detection of specific bacterial pathogens by molecular screening methods.

Results
We set out to study the influence of the genotype of M. tuberculosis on the emergence and distribution of subsequent adaptive, spontaneous resistance mutations. Besides the influence of different genotypes, we also studied the influence of pre-existing isoniazid resistance, by comparing the spectrum of mutations acquired by isoniazid-resistant strains to the spectrum acquired by their (pan)susceptible parental strains. We used two different methods to map these de novo mutant distributions, which allowed us to study two distinct factors of acquisition and accumulation of mutations in vitro. In addition, we determined the mutation rate of all strains selected for this study.

Recently acquired isoniazid resistance caused constrained genetic pathways in vitro
Six isoniazid (INH)-resistant strains, carrying different drugresistance mutations, and their two parent strains were used for these experiments; laboratory-generated mutant strains H15, H26,  H48, H71, H103 and their parent strain MTB72, and clinical  isolate 2001-2184 and its isogenic INH-resistant clinical isolate  2001-2185 (Table 1). The number and nature of rpoB-mutants observed for each strain are depicted in Tables 2 (method 1) and 3 (method 2). No resistant mutants were obtained for strains H15, H26, H48 and H71 using method 1.
It was assumed that no rpoB-mutants were present at the time of inoculation and approximately 10 7 bacteria were present at the time of plating, however, multiple mutational events occurred in some of the 1-ml cultures (indicated by the presence of more than one rpoB mutation identified by MLPA/sequencing). For pragmatic reasons, these were scored as two separate ''counts'', therefore the number (n) in Table 2 does not represent the number of independent cultures where the specific mutation arose. In stead, it indicates the number of times the mutation was found and it should be noted that with our method it could not be determined whether all mutants in one culture carrying the same mutation were the result of one or more mutational events. We have shown previously that mixtures of genotypes can be detected with MLPA, to at least 1:10 ratios [2] The (Pearson's) X 2 test was used to determine the probability of two hypotheses: h 0a : the rate of each of the four most prevalent mutations (S531L, H526D, H526Y and S522L) is equal and h 0b : the distribution of rpoB mutations is unaffected by the presence of INH-resistance.
The results of this test are reported in Tables 4 and 5. Hypotheses were rejected at a p-value of 0.05 or lower.
With method 1 the proportions of the four targeted rpoBmutations were not significantly different and h 0a was accepted for For method 2 the same non-selective starting cultures as for method 1 were used to inoculate one 10-ml culture for each strain used in this study. These cultures were incubated at 36uC in a shaking incubator until mid-logarithmic phase (ca. three weeks). Then four aliquots of 0.5 ml from each culture were plated on rifampicin-containing solid medium to select for resistant mutants. The plates were sealed and incubated at 36uC until sufficient bacterial growth of the mutants was visible (ca. four weeks) When mutant colonies were visible on the plates, DNA was extracted: one mutant colony corresponded to one DNA sample. Finally, DNA samples were screened by MLPA [25] or sequencing of the rifampicin resistance determining region in rpoB. doi:10.1371/journal.pone.0029108.g002 Strains are identified by the name given by either the RIVM (numerical codes) or the KIT (letter+number). DST: drug susceptibility profile, H: isoniazid, R: rifampicin, S: susceptible, R: resistant. The genotype of the strains indicated in the table is determined by spoligotyping, mutations in parentheses are characteristic genotypic mutations identified by MLPA and confirmed by sequencing (ogt, mutT2, mutT4, Ag85C). Mutations in katG confer resistance to isoniazid; deletions were initially picked up by two different PCR reactions, amplifying either the region that covers the drug resistance mutations at codon 315 or the genotypic mutation at codon 463. The notation in this table indicates that the PCR fragment in question was absent and that therefore the region was deleted in the specific strain [2]. doi:10.1371/journal.pone.0029108.t001 Table 2. Spectrum of spontaneous rpoB-mutations obtained by method 1 (multiple parallel 1-ml cultures).  Table 4). However, for strain 2001-2185 the h 0a null-hypothesis was rejected (p a ,0.05, Table 4), indicating that the proportions of the four rpoB-mutants were not equal in this strain. The results obtained by method 2 are depicted in Table 3. The S522L mutation was not observed in any of the strains, except for MTB72 (2/65 (3%)), while this mutation was found in 6/38 (16%, MTB72), 1/19 (5%, H103) and 2/18 (11%, 2001-2184) samples tested with method 1 (Table 2). Instead, most of the RIF-resistant mutants carried the rpoB-S531L mutation when method 2 was used; for all strains, except 2001-2185, which had acquired an equal number of H526Y mutants (7/17 (41%)), this was the category with the highest number of representatives (Table 3). For strains MTB72, H48, H71, H103 and 2001-2185 this shift in the mutational spectrum was significant; the h 0a was rejected for these strains, but not for H26, H15 and 2001-2184.
We also compared the mutant distributions obtained for the INH-resistant strains to the distributions obtained for the wildtype parent strains (Table 5). In comparison to their susceptible parent strains MTB72 and 2001-2184, strains H103 and 2001-2185 did not show a significantly different spectrum using method 1; p bvalues were 0.60-0.50 for H103 and 0.09-0.08 for 2001-2185 (Table 5) and therefore h 0b was accepted. However, with method 2 the distribution of laboratory-generated mutants from H71 and H103 were dramatically different than that of parent strain MTB72; this difference was significant and h 0b was rejected (Table 3, p b ,0.001 Table 5).
Although the null hypotheses, h 0a and h 0b , were accepted or rejected on the basis of the p-values, none of the strains, except MTB72 with method 1 (Table 2), met the criteria needed to ensure sufficient statistical power to the X 2 test, possibly leading to a type II error (incorrectly accepting the null hypothesis). To investigate whether the shift in the spectrum of rpoB-mutations towards rpoB-S531L mutations was detectable with method 2 in both INH sensitive and all resistant strains, we decided to group the results obtained for all INH-susceptible strains (MTB72 and 2001-2184) and all INH-resistant strains (H15, H26, H48, H71, H103 and 2001-2185). We then compared the spectrum of mutations obtained by each of these two groups (mutants vs wildtype) between method 1 and method 2. Because we were only interested in the shift of the proportion of rpoB-S531L mutations, we reduced the groups of observed mutations to two: ''rpoB-S531L'' and ''others''. The X 2 test was then performed to test the similarity between these four mutant distributions obtained. Since there was only one degree of freedom (there are only 2 categories), Yates' correction factor was used (26). These results are depicted in Figure 3.
With method 1 30% (n = 17) of the mutants derived from the INH-susceptible parent strains had acquired an rpoB-S531L mutation. For the INH-resistant strains this was 26% (n = 11) of the total number of mutants. These proportions were almost the Table 3. Spectrum of spontaneous rpoB-mutations obtained by method 2 (single colonies from 10-ml cultures). same and the difference between the wildtype and the INHresistant strains was therefore not significant (p.0.60, Figure 3). In contrast, the spectrum of mutations obtained with method 2 resulted in 41% (n = 31) of rpoB-S531L mutants for the wildtype and 69% (n = 68) for the INH-resistant strains, which was a significant shift towards rpoB-S531L after acquisition of INH resistance (p,0.001, Figure 3). We then compared the proportion of rpoB-S531L mutations acquired by INH-resistant strains via method 1 and method 2 (26% vs 69% respectively, Figure 3) and also found that the increased contribution of rpoB-S531L mutations was significant between the two methods (p,0.001, Figure 3). The mutation distribution of the wildtype pools had also significantly shifted towards rpoB-S531L (0.03,p,0.04), but to a lesser extent: 30% (n = 17) with method 1 and 41% (n = 31) with method 2, Figure 3).

RIF -resistance conferring mutation (codon change) in rpoB
These results suggest that pre-existing INH-resistance can considerably influence the spectrum of subsequent rpoB mutations, with a preference for rpoB-S531L mutations in the 10-ml cultures, but not in the 1-ml cultures.
Two of the Latin-American Mediterranean (LAM) strains, the paired isolates 2001-1669 and 2001-1670, carried an INHresistance conferring mutation in the katG gene; S315T (AG-CRACC, Table 1). Resistance was, however, not acquired in the patient, since the mutation was already present at the moment of diagnosis (primary resistance). Both strains were already resistant to INH, therefore we could not determine the influence of the katG-S315T mutation on the spectrum of rpoB-mutations. In stead, we compared the spectrum of these two INH-resistant strains to the spectrum obtained by strain 9900098, the susceptible LAM strain in our test-panel. This strain showed a propensity for rpoB-S522L with method 1 (6/18 (33%), Table 2) and a preference for rpoB-H526D/Y with method 2 (13/21 (62%), Table 3). The spectrum obtained by strain 2001-1669 was significantly different from this spectrum, by both methods (p,0.001 in both cases, Table 5), but the spectrum obtained by 2001-1670 was only significantly different when method 2 was used (p,0.001, Table 3).
The observed in vitro spectrum of mutations was not significantly correlated with the genotype of the strain In addition to the influence of pre-existing INH-resistance we also assessed the spectrum of RIF-resistance conferring mutations of representative genotypes. We selected four strains with the Beijing genotype, two ''atypical'' Beijing strains (17583 and 2002-1640) and two ''typical'' Beijing strains (2002-1585 and 9500592) [27,28,29], and three LAM strains (9900098, 2001-1669 and 2001-1670), of which two were INH-resistant ( Table 1). The same two methods, 1 and 2, for culture and selection of mutants were used as described earlier. Results are depicted in Tables 2 (method 1) and 3 (method 2). Although the rpoB-V176F was not included in the analyses with the X 2 test, a separate column was included in the tables to indicate the proportion of this mutation that is only seldom identified in clinical isolates [23,30]. The rpoB-V176F mutation is located outside of the 81-bp hotspot, where up to 96% of RIF-resistance conferring mutations appear in resistant strains, both in vivo as in vitro [4,5,8]. A higher prevalence of this mutation may indicate extraordinary circumstances [9].
The rpoB-V176F was not observed in any of the RIF-resistant mutants derived from the other strains.  The probability (p a ) of hypothesis h0 a (the targeted rpoB-mutations (S531L, H526D, H526Y and S522L) have an equal chance of occurring) was determined by X 2 for each strain for both method 1 and method 2. For all tested strains, except in the first experiment with MTB72, the statistical power of the X 2 test was reduced due to one or more mutations occurring less than five times. NA: not available, mutants were not obtained for these strains. Bold: p a ,0.05, therefore hypothesis is rejected. doi:10.1371/journal.pone.0029108.t004 Another striking observation was that none of the 16 mutants derived from strain 9500592 obtained with method 1 had acquired an rpoB-S531L mutation, which is generally associated with the least fitness deficit and is most often seen in RIF-resistant clinical isolates. However, with method 2, 2/7 (29%, Table 3) mutants carried this mutation; the five other mutants carried uncommon mutations (Table 3). This shift away from rpoB-S531L observed with method 1 was significantly different from an equal distribution (p,0.02, Table 4) between the four most common rpoB-mutations, but the spectrum obtained with method 2 was not (0.10,p,0.15, Table 2).
As can be seen, there was a considerate amount of variation in the mutant distribution within as well as between the genotypes (Tables 2 and 3). Replication of the experiment with MTB72, a susceptible Haarlem strain, included in every experiment as a control, showed that the results between experiments were already quite different for a single strain (Tables 2 and 3), both with method 1 and method 2. Therefore, it was not possible to compare the mutant distributions between the different genotypes; combining the data per genotype would not have resulted in a representative distribution.
Furthermore, our data do not seem to show a clear systematic relationship between the amount (proportion), nature and spectrum of uncommon mutations in rpoB (''other rpoB'', Tables 2 and 3) and the genotype; any shift towards or inclusion of atypical rpoB mutations of a strain, appeared to be strain specific rather than genotype specific, as was the case for strains 2001-1669, 9900098 (method 1, Table 2), MTB72, H103, 95000592 and 2002-1640 (method 2, Table 3).
In our study no reproducible correlation between genotype and a specific mutation spectrum was observed.
Clinical INH resistance did not lead to a significantly higher rifampicin mutation rate A fluctuation assay was performed to estimate the mutation rate to rifampicin resistance in the strains included in this study. Two independent experiments were performed for each strain, but, due to contamination and technical problems, duplicate results were only obtained for MTB72 and 2002-1585. Results are depicted in Figure 4.
Most of the mutation rates were measured at around 10 28 , which is comparable to data we obtained previously [2] and to data published by others [13,30]. Strains 17583 and 2002-1585 showed a moderately increased mutation rate (1.21610 27 and 2.49610 27 , respectively; Figure 4), however, well within the range of natural or experimental variation; Werngren and Hoffner [13] have seen comparable variation when testing multiple strains and upon repetition of the experiment with strain 2002-1585 we observed a lower mutation rate (4.50610 28 , Figure 4). Although still within the limits, the isogenic strains 2001-2184 and 2001-2185, which represent the T1 spoligotype-family, have a somewhat lower mutation rate (2.85610 28 and 1.33610 28 , Figure 4). However, strain H66, which obtained INH-resistance by a partial deletion of the katG gene, showed an increased mutation rate, when compared to its wildtype parent strain MTB72; under the same experimental conditions the increase was 10.5 times (5.71610 27 versus 5.46610 28 , Figure 4). In contrast, for the strains which had acquired INH resistance in the patient, via katG-S315T (2001-2185, 2001-1669 and 2001-1670, Figure 4), we did not find an increased mutation rate.

Discussion
The predominance of a mutation is the product of the probability of the mutation occurring in the genome and the probability of the altered microorganism surviving in a given environment. Thus, what ultimately matters for the generation of drug resistance is not the mutation rate, but the substitution rate; the rate at which a bacterial strain can produce viable, adequately adapted offspring. In this study we have shown that the in vitro mutation rate was within the same range for all but one M. tuberculosis strain tested and that there were no consistent significant differences in the spectrum of rpoB-mutations between the various M. tuberculosis genotypes we tested. However, the substitution rate may be different for the different strains; under more rigorous selective conditions, for example in the host, the type of resistance mutation in combination with the genetic background of a strain may have a more dramatic impact on which and how many mutants will survive and be transmitted to new hosts [21,31,32].
Previously, researchers have shown that drug resistance mutations can have varying consequences on the fitness of the bacteria carrying them [19,20,22,23,33,34,35]. This partly depends on the magnitude of the fitness deficit of each de novo resistance mutation, but can also be influenced by the genetic background of the bacterial strain and by the ability/opportunity to acquire compensatory mutations [19,20,36]. Most studies have shown that the S531L mutation in rpoB is the most prevalent rifampicin-resistance conferring mutation found in vitro and in vivo, presumably because it confers the least fitness deficit [8,10,11,19,22,23], at least before transmission or adaptive mutation has occurred [37].
However, other studies have reported an altered distribution of mutations in rpoB among RIF-resistant isolates [9,11,15,16,17,18], indicating that other resistance mechanisms can also be successful under certain circumstances.
In this study we hypothesized that the genetic background does not only influence the fitness of the bacteria, but also impacts the probability of specific mutations accumulating in a population. In addition, we hypothesized that an altered spectrum of mutations may also lead to a changed mutation rate due to a changed adaptation strategy [9].
We tested these hypotheses by selecting spontaneous RIFresistant mutants using two distinct selection strategies which allowed us to investigate the likelihood of specific mutations in rpoB occurring (method 1) versus the likelihood of surviving in the population (the in vitro fitness, method 2). We made use of wellcharacterised M. tuberculosis strains with different genetic backgrounds. Mutations were identified by MLPA [25] or by sequencing of two regions in rpoB [9,38]. To assess the effect of genetic background on the mutation rate we performed a fluctuation assay, using the same concentration of RIF (8 mg/ml) to select for resistant mutants [2].
Comparison of the rpoB mutant distribution obtained from the wildtype parent strains to the rpoB mutant distribution obtained from the INH-resistant daughter strains, showed that acquisition of INH-resistance resulted in a dramatic shift toward rpoB-S531L mutations ( Figure 3). This effect was most evident when method 2 was used to select for RIF-resistant mutants. The mutant distribution of this group was significantly shifted compared to the wildtype INH-susceptible pool with method 2 and to the INHresistant pool with method 1 (Figure 3). The laboratory-generated INH-resistant mutants (H15, H26, H48, H71, H103 (Table 3).
This difference in rpoB mutant spectrum may find its origin in the different INH-resistance conferring mutations carried by the two populations; the laboratory-generated rpoB-mutants carry INH-resistance-conferring mutations that have rarely or even never been observed in clinical isolates [2], whereas the clinical strains in our study (2001-2185, 2001-1669 and 2001-1670) all carry the most prevalent mutation seen in INH-resistant patient isolates, katG-S315T [4,5]. This mutation is thought to confer only a very small fitness deficit [20,34,37,39,40] and the strains carrying this mutation therefore would have a higher ''baseline fitness'' than the strains carrying uncommon INH-resistance mutations. It is assumed that strains with mutations that confer little fitness costs are not in need of restoration of fitness and therefore not likely to acquire adaptive mutations [41]. As a consequence of the higher ''baseline fitness'', katG-S315T mutants will be more able to withstand or overcome the deleterious effects associated with certain rpoB-mutations, allowing for a wider spectrum of RIFresistance conferring mutations. Strains that already have a low ''baseline fitness'', such as potentially our in vitro INH-mutants, can probably only survive acquisition of ''fit'' rpoB-mutations, such as rpoB-S531L. Thus, in our model the relative fitness cost of the different rpoB mutations is probably more apparent in strains that are already quite unfit.
The higher baseline fitness of the clinical INH-resistant mutants could, in part, also be a result of adaptive or compensatory Description of the strains can be found in Table 1. Red bars indicate replicates from a second, independent experiment. doi:10.1371/journal.pone.0029108.g004 mutations, acquired within the patient. A higher baseline fitness could have also allowed for the many rare mutations in rpoB, including deletions and insertions, we observed for almost all strains with both methods (Tables 2 and 3); most of these mutations are rarely or never found in RIF-resistant strains isolated from (immunocompetent) patients, where severe (sequential) bottlenecks have probably selected against most of these mutations.
Since multiple mutants are present with method 2 there is more competition between mutants than with method 1 [11], we therefore would have expected a reduced spectrum of rpoBmutations. In addition, other researchers have reported rpoB-S531L to be the most frequently observed mutation in similar experiments [12]. In those experiments 5-ml cultures were used and there was no opportunity to identify multiple mutations in a whole culture, as we did; in stead one or sometimes 10 colonies per plate were analysed, perhaps biasing towards ''fitter'' and therefore bigger colonies to be picked.
Our results indicate that probably more severe bottlenecks or strong selective sweeps have to be applied in vitro to mimic the selection processes bacteria are constantly subjected to when living inside the host [36]. However, prolonged incubation of bacteria can already lead to a decreased spectrum of mutations [11], suggesting that despite an initial burst of multiple rpoB-mutations, even in vitro conditions can induce rather constrained evolutionary pathways [21].
We have shown that, regardless of genotype, many more drug resistance mutations occur than would be expected on the basis of clinical screening. Particularly codon 526 can be highly variable, as was observed by others [4,5,12,22], and we detected many uncommon mutations (Tables 2 and 3). We also found a relatively high degree of insertions and deletions, which did not seem to be specific for any strain or method and there were some inclinations towards certain mutations, in particular for the rare mutation V176F. However, we did not detect a difference in the spectrum of mutations between the various genotypes. We did not detect a specific M. tuberculosis genotype-related in vitro effect that determines which mutations predominate.
Others have performed similar studies, albeit on a smaller scale and with more focus on Beijing/non-Beijing [10,13,24], and also did not find a genotype-related effect. We observed that the mutational spectrum can vary substantially between experiments. Researchers have previously reported that the spectrum of RIFresistance conferring mutations in rpoB can be dependent on the growth conditions of the bacteria, such as the pH of the growth medium or even the age of a bacterial culture [11]. Genotyperelated characteristics have been observed in vivo [15,16,17,18] but to our knowledge structural differences between M. tuberculosis genotypes have yet to be demonstrated in vitro.
We have focused here on differences in evolutionary pathways of the bacteria, but are aware that differences observed in vivo to a certain extent may be attributable to host characteristics; reports of ''favoured mutations'' have often been restricted to a single geographical location, where it is likely that hosts share genotypic characteristics [42,43]. It has furthermore been discovered that host factors, such as the Toll-like receptor 2 allele, can partly influence the manifestation of disease and the susceptibility of humans to certain bacterial genotypes [44].
In a recent study sequential, increasingly drug-resistant isolates were obtained from a single patient and it was shown that the first (susceptible) and the last (multidrug resistant) isolate only differed by the two drug resistance-conferring mutations, as determined by whole genome sequencing [32]. This homogeneity for INH and RIF resistance conferring mutations in sequential isolates from a single patient was corroborated by others [45]. These reports imply that the selection pressure exerted by antibiotic usage can significantly reduce the spectrum of viable mutants and evolutionary pathways are probably restricted as a result [21,46]. However, host-to-host transmission probably has an even more severe selective and thus constraining effect [47,48]. In addition, Schürch et al [49] have retrospectively followed a chain of transmission of TB in the Netherlands over a period of more than a decade; the strain in question only acquired six mutations, of which five had been acquired within one patient.
In areas where transmission is rapid, successful strains [50,51] may have very different characteristics to strains which are successful in more slowly evolving epidemics [48]. These two forms of transmission could lead to very different evolutionary routes; in areas with a high transmission rate it is likely that clusters of drug-resistant strains will be clonal and acquisition of drug resistance happened only once (before the spread -secondary resistance). On the contrary, in regions where the epidemic moves more slowly, drug-resistant strains are, at least initially, not expected to be highly clustered since drug resistance is more likely to be established within the patient (primary resistance) and adaptive evolution appears to continue even during latent infection [52].
If bacteria with specific mutations have a significant advantage in a disseminating population, the same drug resistance mutations will have a higher chance of occurring repeatedly, such as rpoB-S531L in our katG mutants. Thus, such a cluster of primary resistant strains would be indistinguishable from a truly clonal cluster of secondary resistant strains by current genotyping methods.
This situation is analogous to the high prevalence of rpoB-S531L mutants in our 1-ml experiments (method 1) which are not clonal, whereas rpoB-S531L mutants from the 10-ml experiment (method 2) may (or may not) represent clonal expansion of a singe mutational event. Misinterpretation of these two scenarios may result in incorrect measures being taken and failure to control the spread of resistance.
It has been suggested that sequencing of complete bacterial genomes should be used for the standard screening of clinical isolates [53]. Whole genome sequencing of seemingly clonal drug resistance clusters could reveal additional (''piggy-back'') mutations, indicating if drug resistance was indeed spread or acquired within the host.
A recent article [6] reports that the sensitivity of a molecular drug resistance test depended on the resistance profile of the bacteria tested, varying from close to 91% for MDR-TB to only 56%/70% for strains with monoresistance to INH or RIF. These results are supported by Van Rie et al. who analysed multiple strains from South Africa and concluded that by targeting just three mutations it was possible to detect up to 90% of MDR-TB [7].
As most molecular tests target only the most prevalent drug resistance conferring mutations, the increased sensitivity for MDR strains may be a direct result of a constrained pathway for the bacteria: only a few combinations of drug resistance conferring mutations can lead to a ''fit'' MDR strain, whereas a monoresistant strain can have a much wider spectrum of mutations. Indeed, in a study performed by Iwamoto it was shown that as strains accumulate multiple drug resistance mutations, the spectrum of mutations is considerably reduced [54].
These reports, in addition to our data, indicate that sign epistasis can play a significant role in the development of multidrug resistance in M. tuberculosis, as was suggested by Trindade et al. [55] and that the routes to a successful MDR or even XDR strain tend to become more restricted with each mutation that is acquired [21].
We are entering an era in which we will recognise the importance of certain mutations and we can use them to build early warning systems or optimise treatment [31]. Although for a genetically stable organism such as M. tuberculosis only a fraction of the data generated will be informative, data acquired with whole genome sequencing will be invaluable for the design of molecular tests with a much higher discriminatory ability than current methods, by targeting informative single nucleotide polymorphisms scattered throughout the genome [2,25,53,56,57,58,59].

M. tuberculosis strains used
Fifteen strains were used during the course of this study, including clinical isolates with and without isoniazid (INH) resistance, and a laboratory strain (MTB72) and its spontaneous INH-resistant mutants. This collection of strains represents the M. tuberculosis genotypes Haarlem [60], Beijing ('typical' and 'atypical' sublineage [28,29]), Latin-American Mediterranean (LAM) [61] and T1 [62] Table 1); for this reason they are assumed to have the same clonal origin. The same is true for strains 2001-1669 and 2001-1670.
Strain MTB72 (ATCC 35801) is a pansusceptible laboratory strain belonging to the Haarlem genotype. Strains with a HXX code are spontaneous INH-resistant mutants, derived from strain MTB72 in our laboratory as described previously [2].
The characteristics of the strains and their origins are described in Table 1.

Bacterial culture
Bacteria were cultured in Middlebrook 7H9 medium (Difco, BD, Sparks, MD, USA), supplemented with oleic acid/albumin/ dextrose/catalase (OADC Enrichment, BD, Sparks, MD, USA), in a shaking incubator at 37uC. For all strains, a liquid starting culture was made by inoculating pure colonies from Löwenstein-Jensen or Coletsos slopes into 10 mL of liquid culture medium. When these cultures reached the logarithmic growth phase (circa 3 weeks) they were mixed vigorously to homogenize the bacterial suspension. Clumped cells were allowed to settle for 3 min and aliquots of the cell suspension were then transferred to fresh nonselective liquid medium to make multiple parallel cultures of 1 mL (method 1) or a final culture of 10 mL (method 2) as described below.
Method 1 (Figure 1). For each strain 25 1-ml cultures were inoculated from the 10-ml non-selective liquid starting culture by transferring 1 uL of this cell suspension (,1000 bacteria) to a 2 mL screwcap tube containing 1 mL of MB7H9 medium + OADC with a sterile inoculation needle. Two to three sterile glass beads were added to each culture to ensure mixing. The 1 mL cultures were incubated in a shaking incubator at 36uC for approximately three weeks. Bacterial growth was monitored each week by adding 30 mL of a 0.02% (wt/vol) resazurin solution (Sigma), a growth indicator, to additional 1-ml cultures incubated simultaneously. These additional cultures were used only to monitor growth and resazurin was not added to the cultures from which the RIF-resistant mutants were analysed. After addition of the resazurin, all cultures were wrapped in aluminum foil and after 24 hours incubation at 36uC the color development was determined. After sufficient growth, when the reference cultures turned a bright pink, the target cultures were centrifuged at 5000 g for 8 min and 850 mL of the supernatant was discarded. For each culture the remaining 150 mL was resuspended and plated in one well of a square 25-well replica plate (Greiner, Germany) containing 3 mL of MB7H11 + OADC supplemented with 8 mg/L rifampicin (Sigma-Aldrich Chemie) to select for RIFresistant mutants.
The 25-well square plates were allowed to dry in a biosafety laminar flow cabinet, until all liquid was absorbed into the solid medium. The plates were then sealed in plastic bags and incubated at 36uC. After growth was clearly visible in all or most wells, total populations of mutants were isolated from each well and analysed for mutations in rpoB (see below).
Method 2 ( Figure 2). For each strain a single 10-ml culture was made with a 0.5 ml inoculum from the 10-ml, non-selective liquid starting culture described above. Bacterial growth was monitored by determining the turbidity of the cultures at regular intervals. When sufficient growth was established, from each strain four 0.5 ml aliquots were taken from each strain and plated on four separate 8 mg/L RIF-containing plates to select for resistant mutants. Plates were wrapped and sealed separately in plastic bags to minimise cross-contamination and incubated at 36uC for 3-4 weeks, until clear mutant colonies could be observed. DNA was then isolated from single colonies by the method described below and samples were further analysed to screen for mutations in rpoB.

Screening/characterisation of mutants
Isolation of DNA. For method 1 200 mL lysis buffer (10 mM Tris-HCl/1 mM EDTA pH 8.0 containing 1% Triton X-100 (BDH Laboratory Supplies, Poole, England)) was placed in each well of the 25-well square plates. Resuspension of the mutant colonies was ensured by pipetting the buffer up and down with a micropipette. With a clean pipette filtertip the suspension containing the total mutant population from a single well, was transferred to a microcentrifuge tube and then heated at 95uC in a heat block for 30 minutes. After lysis, cells were centrifuged at 5000 g for 3 minutes and 130 mL of the supernatant was collected.
For method 2 individual mutant colonies were picked from the RIF-containing plate and each suspended in 150 mL lysis buffer after which the same procedure as described above for method 1 was followed to extract the DNA.
Sequence analysis rpoB. For all RIF-resistant mutants for whom no mutation in rpoB was identified by MLPA, two fragments of the rpoB gene were sequenced. The same DNA extracts used for MLPA were used for PCR and subsequent sequencing analysis. PCR of rpoB clusters I and III was carried out as previously described [9], using primer pairs rpoB-2F/2R and rpoB-7F/7R respectively [25]. The PCR products were sequenced in both directions using the dideoxy chain termination method with the Big Dye Terminator cycle sequencing Kit (Applied Biosystems). Sequence analysis was performed on a 310 Genetic Analyzer (Applied Biosystems). All rpoB codon numbers, except V176, are reported using the E. coli numbering system.

Fluctuation assay
The mutation rates (i.e. the chance of a mutation occurring per generation) for rifampicin resistance were estimated with a fluctuation assay. Our method has been previously published [2] and was based on the p0-method described by Luria and Delbrück [64].

Statistical analysis
For all analyses our null hypothesis (h 0 a) was that under each condition tested each mutation in rpoB was equally likely to occur and secondly (h 0 b) that the distribution of specific mutations would be unaffected by the presence of INH resistance. To test this hypothesis Pearson's X 2 test was used. Hypotheses were rejected at a p-value of 0.05 or lower. We used the X 2 test to determine if the frequency of each of the four most prevalent mutations (S531L, H526D, H526Y and S522L) was equal when using each selection method and secondly, to determine whether the distribution of rpoB mutations in each of the INH-resistant mutants was equal to that of the INH susceptible parent. The p-values of the distributions of the individual strains are depicted in Tables 4 and 5. In most of the cases the statistical power of the X 2 test was reduced, since not all of the criteria for the X 2 -test were met (i.e. in all the observed categories the minimum sample size should be five and none of the observed frequencies should be zero). Therefore we grouped all strains which were INH-resistant and all strains which were INH-susceptible and reduced the categories to ''rpoB-S531L'' and ''other mutations''. To adjust for this reduction in number of categories, we used Yates' correction for continuity [26].

GenBank accession numbers
The majority of the rifampicin resistance-conferring mutations in rpoB we found in this study have been previously observed and reported. Those mutations that were not deposited to GenBank (http://www.ncbi.nlm.nih.gov/Genbank/) or uploaded to the M. tuberculosis drug resistance mutations database TBDReaMDB (http://www.tbdreamdb.com/) were regarded as novel and were deposited to GenBank. Only mutations found in DNA samples derived from single mutant colonies (method 2) and not from mixed cultures (method 1) were considered. Mutations in rpoB AAC519AAA (GenBank JN819066), CGA529CTA (GenBank JN819067), indel D512-519 ins ATC (GenBank JN819068) and ins 514-515 AAATTC (GenBank JN819069) were deposited to the GenBank database. In addition, the isoniazid resistanceconferring mutation katG ACT271ATT in strain H103 was deposited to GenBank, under the accession number JN819065.