Mutation rates and adaptive variation among the clinically dominant clusters of Mycobacterium abscessus

Significance

Mycobacterium abscessus (Mab) is an emerging public health threat that causes severe lung disease in affected individuals. Recent genomic studies have revealed that the Mab evolutionary tree is characterized by extremely dense clusters of isolates sampled from diverse geographies. While Mab was historically thought to be acquired by patients independently from the environment, these results raise the possibility of patient-to-patient transmission. In this study we provide evidence that the emergence of Mab clusters has coincided with a slowing of the mutation rate, resulting in the appearance of dense clusters. We propose that changes in the mutation rate result from beneficial mutations in DNA repair machinery. These results challenge the transmission model of cluster emergence in Mab.

Abstract

Mycobacterium abscessus (Mab) is a multidrug-resistant pathogen increasingly responsible for severe pulmonary infections. Analysis of whole-genome sequences (WGS) of Mab demonstrates dense genetic clustering of clinical isolates collected from disparate geographic locations. This has been interpreted as supporting patient-to-patient transmission, but epidemiological studies have contradicted this interpretation. Here, we present evidence for a slowing of the Mab molecular clock rate coincident with the emergence of phylogenetic clusters. We performed phylogenetic inference using publicly available WGS from 483 Mab patient isolates. We implement a subsampling approach in combination with coalescent analysis to estimate the molecular clock rate along the long internal branches of the tree, indicating a faster long-term molecular clock rate compared to branches within phylogenetic clusters. We used ancestry simulation to predict the effects of clock rate variation on phylogenetic clustering and found that the degree of clustering in the observed phylogeny is more easily explained by a clock rate slowdown than by transmission. We also find that phylogenetic clusters are enriched in mutations affecting DNA repair machinery and report that clustered isolates have lower spontaneous mutation rates in vitro. We propose that Mab adaptation to the host environment through variation in DNA repair genes affects the organism’s mutation rate and that this manifests as phylogenetic clustering. These results challenge the model that phylogenetic clustering in Mab is explained by person-to-person transmission and inform our understanding of transmission inference in emerging, facultative pathogens.

Mycobacterium abscessus (Mab) is an emerging, multidrug-resistant pathogen that is increasingly responsible for opportunistic pulmonary disease, most commonly in patients with underlying structural lung conditions such as cystic fibrosis (CF) (1). Mab is divided into three subspecies: M. a. abscessus, M. a. massiliense, and M. a. bolletii, the former two being the most common clinical subspecies (2). Infections with Mab are associated with accelerated decline in lung function and can result in severe disseminated disease and/or loss of eligibility for lung transplantation. Mab infections are difficult to treat, requiring a prolonged antibiotic regimen with high rates of treatment failure (3).

Until recently, Mab infections were thought to be acquired independently by each patient from environmental reservoirs in soil or water. Large-scale analyses of whole-genome sequencing (WGS) of Mab clinical isolates have revealed that a large proportion of clinical isolates fall into several phylogenetically characterized clusters, often termed dominant circulating clones (4). Isolates within these clusters have high core genome similarity that is not explained by point source outbreaks as they were sampled from diverse geographic origins. These observations have been interpreted as supporting widespread recent transmission, either through direct contact or indirectly through fomites, but a number of epidemiological studies have refuted person-to-person transmission as a major mode of Mab dissemination (5–11). Reconciling these conflicting lines of evidence is important as they dictate public health practice and allocation of resources to protect vulnerable populations from infection. However, it is still unclear what role transmission plays in the emergence of Mab clusters.

While widespread transmission is one proposed explanation for the phylogenetic structure observed in Mab, it is important to consider other factors in the species’ evolutionary history that may contribute. For example, transient niche switching between environmental sources and the human host may introduce bottlenecks resulting in low diversity. Another explanation is a change in mutation rates, either through changes in bacterial generation time or through genetic variation. A third consideration is clinical sampling, where particular Mab lineages that cause more frequent, severe and/or persistent disease are more likely to be isolated and sequenced. Consistent with the latter hypothesis, previous work has shown that isolates within clusters are more likely to contain point mutations associated with antibiotic resistance and are associated with greater bacterial burden and lung inflammation in mice (4).

Interpreting tree structures in the context of population history is aided by phylogenetic dating, which translates genetic distances into time scale. Previous studies (4, 9, 12, 13) have estimated molecular clock rates in Mab but these have been limited to subpopulations within clusters. Here, we report on molecular clock rate estimation for the deep branches of the M. a. abscessus phylogeny and for within-host isolates belonging to clusters. We provide evidence for an evolutionary slow-down coincident with the expansion of clinically dominant clusters (DCs) and propose that the emergence of dense phylogenetic clustering in Mab may result from genetically encoded changes to the mutation rate. These results indicate that transmission is not the only plausible explanation for the dense clustering observed in the Mab phylogeny, and that more research is needed to clarify the transmission dynamics of this emerging pathogen.

Results

Phylogenetic Analysis Confirms the Presence of Large, DCs Spanning Multiple Geographies.

We identified 1,629 M. abscessus isolates with publicly available WGS data and associated dates of collection (SI Appendix, Table S1). We excluded isolates with likely contamination or low-quality assemblies (Methods). Among the 1,461 isolates passing quality filters, the majority (80.1%) were derived from in vitro culture of pulmonary samples (SI Appendix, Table S2). Of the 1,181 pulmonary isolates, 98.3% (1,161/1,181) were collected from patients with CF, while the rest were collected from patients with no documented diagnosis. The dataset represented 483 patients with unique patient identifiers, with 59.6% (288/483) of patients contributing a single isolate, and the remaining contributing between 2 and 24 isolates sampled over time. Isolates originated from nine countries and were collected between 1998 and 2017, except for the reference strain ATCC 19977 which was collected in 1957 (Fig. 1B).

Fig. 1.

DCs of M. abscessus and the presence of temporal signal in internal branches. (A) Species-wide phylogeny of M. abscessus representing one isolate per patient to avoid within-patient clustering of samples. Each subspecies tree was constructed independently using a subspecies-specific reference genome. The four largest phylogenetic clusters are highlighted. Tree scale represents SNPs/site. (B) Histogram showing the distribution of collection dates of all clustered isolates v. all unclustered isolates in M. a. abscessus. (C) Pruned M. a. abscessus phylogeny representing 95% of the original subspecies tree diversity. (D) Root-to-tip regression showing evidence of positive temporal signal in the pruned tree shown in C.

The majority (91.7%) of samples had ≥98% average nucleotide identity (ANI) with one of three reference genomes (SI Appendix, Table S3) representing each of the three Mab subspecies, consistent with the commonly used ANI threshold for bacterial subspecies. Isolates with <98% ANI with any subspecies were excluded from downstream analysis. The resulting dataset overrepresented clinically relevant subspecies of M. abscessus, with 93.9% belonging to either M. a. abscessus or M. a. massiliense. To better resolve true evolutionary relationships among samples, we defined and restricted analysis to the core genome for each subspecies, then further excluded predicted recombination events (Methods). We performed maximum likelihood (ML) phylogenetic analysis for each of the three Mab subspecies, using only the first isolate from each patient to focus on between-host diversity. The resulting subspecies trees confirmed the presence of multiple, globally distributed clades with dense phylogenetic clustering (Fig. 1A). We excluded clustering due to potential point source outbreaks by limiting to clusters containing isolates from more than one country and a minimum of three isolates for downstream analysis. In total, we confirmed 16 clusters (SI Appendix, Fig. S3) including four clusters with >50 isolates, which we refer to here as dominant clusters (DCs). Isolates within clusters comprised 76.4% of isolates included in the phylogenetic analysis. The median pairwise distance in M. a. abscessus DC 1 was 68 SNPs. For comparison, the median pairwise distance across the subspecies trees were 4,119, 5,093, and 8,924 SNPs for M. a. abscessus, M. a. massiliense, and M. a. bolletii, respectively. All four DCs contained isolates originating from three continents (North America, Europe, and Australia). Clustered isolates also spanned the full range of isolation times, including ATCC 19,977, which was placed within M. a. abscessus DC 1 despite having been collected at least 40 y before any other isolates in the tree (Fig. 1 A and B).

Coalescent Analysis Reveals a Faster Long-Term Evolutionary Rate Than Observed in DCs.

To aid in transmission inference and to reconstruct the evolutionary history of Mab, we performed a coalescent analysis using our SNP alignment. Previous work has reported molecular clock rates within Mab clusters, but to our knowledge no study has successfully dated any of the full Mab subspecies phylogenies. To evaluate the presence of temporal signal, we performed root-to-tip regression for each subspecies tree but in each case did not observe a positive correlation between sampling date and root-to-tip distance. We reasoned that the high degree of dense clustering in the Mab phylogeny may obscure the genetic-temporal signal. To test this hypothesis, we pruned each subspecies tree until it contained the minimal number of samples required to capture 95% of the genetic diversity in the original tree (Methods). Following pruning of the M. a. abscessus subspecies tree, root-to-tip regression demonstrated weakly positive temporal signal (Fig. 1 C and D and SI Appendix, Fig. S4C). Because the data supports a relaxed clock model (Methods), and root-to-tip regression assumes a strict clock, we further tested for evidence of temporal signal using a Bayesian evaluation of temporal signal (BETS). This comparison yielded a Log₁₀(Bayes’ Factor) = 277.4 (SI Appendix, Table S4) which strongly favors the presence of positive temporal signal (14).

We used a Bayesian implementation of coalescent analysis, BEAST (15), to estimate the molecular clock rate of the pruned M. a. abscessus subspecies tree assuming a relaxed clock model, which allows each branch of the tree to have its own clock rate. To explore the effect of the tree prior on molecular clock estimation, we ran analyses using two different tree priors, assuming either a constant coalescent model, which assumes constant population size over time, or a Bayesian skyline coalescent model, which allows the population size to change over time. While the Bayesian skyline coalescent had a slightly better marginal likelihood, we found no significant difference between the clock rates estimated by the two models (Table 1 and Fig. 2C). Our estimates for the clock rate in the long, internal branches of the tree were approximately 10-fold faster than nearly all previously reported estimates of the clock rate within Mab clusters (Fig. 2C). The clock rate estimate we obtain is sensitive to which samples are selected by Treemmer (SI Appendix, Fig. S5C). We therefore performed multiple replicates of the pruning procedure and repeated the coalescent analysis for each subset of isolates retained by Treemmer. Here, we present the most conservative (lowest) estimate we obtained from Treemmer sampling. Clock rate estimates were also robust to the algorithm choice or stringency of the recombination search procedure (SI Appendix, Fig. S5B). We next attempted to date the full (unpruned) M. a. abscessus phylogeny using the faster rate inferred from our coalescent analysis of the pruned tree. These chains suffered from poor convergence and sampled trees with a distorted topology, elongating the short branches and eliminating the densely clustered structure observed in the ML phylogeny (SI Appendix, Fig. S6). We therefore hypothesized that the lack of temporal signal in the full M. a. abscessus dataset may result from a difference between the long-term clock rate and the more recent evolutionary rate within phylogenetic clusters.

Table 1.

Coalescent analysis results and model comparison

Model	Marginal log likelihood (SD)	UCLD Rate (SNPs/site/year)	UCLD SD
Coalescent constant population	−7,144,301.64 (21.42)	2.82 × 106 [1.33 × 106, 4.83 × 106]	0.182 [0.135, 0.234]
Bayesian skyline coalescent	−7,144,118.76 (19.93)	2.63 × 106 [1.16 × 106, 4.0 × 106]	0.181 [0.138, 0.232]

Fig. 2.

Longitudinally sampled within-host isolates and coalescent analysis support a slower clock rate within DCs. (A) Clock rate estimation using isolate pairs sampled from the same patient over time. (Left) Distribution of SNP distances between all possible within-host isolate pairs compared to SNP distances between all possible isolate pairs within M. a. abscessus DC 1. The orange line represents the threshold of 20 SNPs used to exclude pairs representing independent infections. (Right) Regression of SNP distance on the time between isolates for each isolate pair. The slope of the regression line was 2.51 SNPs/y [1.24 to 3.78 95% CI], or roughly 5.93 × 10⁻⁷ SNPs/site/y [2.92 × 10⁻⁷ to 8.9 × 10⁻⁷ 95% CI]. (B) UCLD clock rate estimates (dots) for the three models used in this study (two coalescent models shown in red and one regression model shown in blue) compared to mutation rate estimates from the within-host regression (blue) and published clock rates (gray). Lines represent the 95% HPD interval, except in the case of the within-host regression in which case the line represents the 95% CI of the regression shown in A.

Analysis of Longitudinally Sampled Within-Host Isolates Supports a Slower Clock Rate in DCs.

To further investigate the differences in clock rate between long-term historical time scales and short-term contemporary time, we used pairs of isolates sampled from the same patient overtime to estimate the mutation rate. We focused on estimating a within-host clock rate for M. a. abscessus to compare directly with the estimate from our coalescent analysis and because we had the most longitudinal samples for this subspecies. All isolate pairs belonged to one of the M. a. abscessus DCs. To avoid including isolate pairs representing polyclonal infections, excluded isolate pairs with a SNP distance >20, yielding 48 isolate pairs. This threshold was selected by comparing the distribution of all possible within-host isolate pairs to the distribution of pairwise SNP distances within Mab DCs (Fig. 2A and SI Appendix, Fig. S7). Regressing the SNP distance between isolate pairs on the time between sample pairs yielded a mutation rate of 2.21 SNPs/y [1.06 to 3.37 95% CI], or 5.17 × 10⁻⁷ SNPs/site/y [2.48 × 10⁻⁷ to 7.92 × 10⁻⁷ 95% CI]. This value is more similar to previously reported molecular clock rate estimates within Mab clusters, which range from 1.58 × 10⁻⁷ to 8.25 × 10⁻⁷ SNPs/site/y (4, 9, 12, 13), than to those estimated in the long branches of the phylogeny (Fig. 2C).

Simulated Ancestries Support Mutation Rate Slow-Down Over Population Size Effects as the Major Driver of Phylogenetic Clustering.

To test whether sampling from clades with low genetic diversity is sufficient to explain the degree of clustering observed in the ML phylogeny, or whether changes in the molecular clock rate are a better explanation for the observed phylognetic data, we performed coalescent simulations of ancestries with one clustered subpopulation (A) and one unclustered subpopulation (B). Under a range of effective population sizes (N_e) and mutation rates (μ), we simulated ancestries through time with 1,000 replicates per simulation. For each replicate, we assessed the degree of phylogenetic clustering by measuring d_A/d_B where d_i is the mean pairwise SNP distance within subpopulation i. The total N_e (N_e^T), number of samples, and sampling dates for subpopulations A and B were drawn from the true values for M. a. abscessus DC 1 and from all “unclustered” isolates in our dataset, respectively (Fig. 3A). We first tested whether a smaller population size for the clustered subpopulation A (N_e^A) could explain the degree of phylogenetic clustering observed. We estimated a range of realistic values for N_e^A using BEAST by assuming a range of possible clock rates (SI Appendix, Table S5). These rates spanned a published clock rate estimate for M. a. abscessus DC1 to a clock rate that is 50-fold higher, far exceeding the upper 95% highest posterior density (HPD) interval for the published value (13). A schematic outlining this procedure is available as SI Appendix, Fig. S8. Under the simulated conditions of equal mutation rates between subpopulations A and B, we only observed phylogenetic clustering to the extent observed in the clinical isolate phylogeny in the scenario where N_e^T >> N_e^A (Fig. 3B and SI Appendix, Fig. S9). We next tested whether a change in the mutation rate at the root of supopulation A could better explain the observed phylogenetic structure in Mab. The simulated ancestries where N_e^A and N_e^B are fixed and the mutation rate slowdown is between 10- and 20-fold recapitulated the observed phylognetic structure (Fig. 3C). This result was robust to changes in N_e^A and N_e^T across the 95% HPD for each parameter (SI Appendix, Fig. S10). These results indicate that for person-to-person transmission to drive phylogenetic clustering in Mab, the total N_e must be very large compared to the N_e of the clustered subpopulation, an unlikely scenario given the observed genetic data and sampling times. By contrast, a mutation rate change of 10- to 20- fold explains the degree of clustering observed in Mab under more realistic population sizes given these data.

Fig. 3.

Simulated ancestries support a mutation rate slow-down coincident with DC emergence. (A) Example neighbor joining (NJ) tree from the simulation shown in B. The number of samples and sampling dates for subpopulations A and B are drawn from real data from M. a. abscessus cluster 1 and from all unclustered samples, respectively. The dotted line crosses the common ancestor node for subpopulation A, and all branches descended from this node (shown here in red) are subject to the change in mutation rate. (B) Boxplot showing the degree of phylogenetic clustering in subpopulation A relative to subpopulation B over a range of effective population sizes of subpopulation A (N_e^A). The simulation assumes that N_e^T = 8,000. (C) Boxplot showing the degree of phylogenetic clustering in subpopulation A relative to subpopulation B over a range of mutation rate changes. The simulation assumes that N_e^T = 8,000 and N_e^A = 718. In both (B and C) the degree of clustering is estimated with metric log${\displaystyle \left(\raisebox{1ex}{${\mathit{d}}_{\mathit{A}}$}\!\left/ \!\raisebox{-1ex}{${\mathit{d}}_{\mathit{B}}$}\right.\right)}$  where d_i is the mean pairwise SNP distance among all sample pairs in subpopulation i. The red lines represent the clustering metric estimated using the observed ML phylogeny. Boxes extend from the lower to upper quartile values with a line at the median. Whiskers show the range of the data.

UvrD/Rep Family Helicases Are Enriched in DCs in M. a. abscessus.

To identify genetic variants under positive selection in Mab clusters, we searched for evidence of phylogenetic convergence (https://www.nature.com/articles/ng.2747) in the core genome of M. a. abscessus with regions of recombination excluded. Briefly, we performed ancestral sequence reconstruction for the internal nodes of the subspecies tree and counted the number of independent arisals of each SNP (Methods). Of the 65,202 SNPs identified in M. a. abscessus, 34,813 (53.4%) of mutations occurred only once and were excluded. For the remaining SNPs, we ranked each by its enrichment in DCs relative to other regions of the M. a. abscessus tree (Fig. 4A). The most significantly enriched variants in the M. a. abscessus DCs included homologs of several known or putative virulence factors as well as a number of cluster-enriched mutations in UvrD-like helicases MAB_1054, MAB_3515c, and MAB_3516c (Fig. 4A and SI Appendix, Table S6). While the function of these genes is not well characterized in M. abscessus, all three genes share homology with the UvrD/Rep family of helicases. This family of genes is involved in DNA repair and has been associated with growth and persistence in vivo (16, 17).

Fig. 4.

DCs are associated with a decrease in the spontaneous mutation rate in vitro. (A) Subspecies phylogeny for M. a. abscessus (Left) and corresponding heatmap representing the presence of the derived allele (blue) or ancestral allele (gray) for each nonsynonymous variant. Variants are ranked in ascending order by their P-value derived from Fisher’s exact tests for enrichment in the three largest clusters. The top 50 hits are shown. The genes where each variant is located are denoted. (B) Subspecies phylogeny for M. a. massiliense. The branches where mutations have occurred in homologs of Rep and UvrD2 are shown with asterisks. Probability of acquiring an amikacin resistance mutation per generation in (C) M. a. abscessus strains (D) M. a. massiliense strains, or (D) in all strains tested that are part of a dominant cluster or not. Mean ± SD is displayed. In C and D, n = 3 biological replicates. In E, each point represents the average mutation probability for a single strain. The P-value derived from a one-sided Mann–Whitney U test.

Isolates in Mab DCs Are Associated with a Lower Spontaneous Mutation Rate.

To test the association between phylogenetic clustering and spontaneous mutation rate, we selected M. a. abscessus and M. a. massiliense strains that were grouped within DCs or that were not present in those clusters. Each of the strains present in the DCs has the derived (minor) allele at four sites in the UvrD/Rep family genes MAB_1054, MAB_3515c, and MAB_3516c, and the strains outside of DCs possess the ancestral (major) allele at those sites (Table 2 and Fig. 4A). The strains used and their genotypes are described in SI Appendix, Tables S7. The spontaneous mutation rate was estimated in all strains using a fluctuation assay (18) adapted for use in M. abscessus. The fluctuation assay counts cells gaining resistance to the antibiotic amikacin during growth in culture without antibiotic. By normalizing the number of mutational events to the population size of the culture, the fluctuation assay estimates the spontaneous mutation rate per generation per the number of bases that can be mutated to confer resistance to amikacin. The strains chosen display similar growth rates (SI Appendix, Fig. S11A) and very low baseline resistance to amikacin (SI Appendix, Table S8) which allows them to be compared using a fluctuation assay. Among M. a. abscessus strains, those present in DCs with the derived (minor) allele at the four UvrD/Rep loci had lower spontaneous mutation rates than those with the ancestral (major) allele (Fig. 4C). The trend among M. a. massiliense strains was less clear (Fig. 4D), potentially reflecting the effects of differences in strain background in this subspecies. Nonetheless, the mutation rate was significantly lower in isolates present in DCs across both subspecies (Fig. 4E), consistent with the hypothesis that the emergence of dense clustering may be driven in part by changes in mutation rate.

Table 2.

Summary of top 10 nonsynonymous variants enriched in DCs of M. a. abscessus

Gene	Gene symbol	Description	Comments	Enrichment P-value (FDR adjusted)	Number of clustered taxa containing variant	Number of unclustered taxa containing variant
MAB_3334c	gatB *	Aspartyl/glutamyl-tRNA (Asn/Gln) amidotransferase subunit B*,†		6.70 × 1062	240	1
MAB_3244		Mercury resistance transport protein/cytochrome c biogenesis protein*		5.56 × 1061	189	3
MAB_3480		Putative short chain dehydrogenase/Reductase*		1.55 × 1060	232	13
MAB_4141 ‡		PE/PPE family protein*		1.70 × 1059	189	2
MAB_3242	fni *	Isopentenyl-diphosphate delta-isomerase*,†		7.47 × 1059	189	3
MAB_3481	fadE *	Probable acyl-CoA dehydrogenase*	FadE mutations in Mtb result in decreased growth and virulence in macrophages and in mice (19)	1.69 × 1058	232	15
MAB_4057c	mshA *	D-inositol 3-phosphate glycosyltransferase†	Catalyzes biosynthesis of mycothiol, an important compound in resisting oxidative stress and detoxifying harmful agents such as antibiotics (20, 21)	1.80 × 1057	189	2
MAB_3581c		Hypothetical protein		3.59 × 1057	169	3
MAB_1054 §	pcrA*/uvrD1†	ATP-dependent DNA helicase*,†	Involved in nucelotide excision repair (NER), mismatch repair (MMR), homologous recombination (HR), and rolling circle replication (22)	1.45 × 1055	216	7
MAB_3515c	rep * , †	ATP-dependent DNA helicase*,†	Involved in NER, MMR, HR and rolling circle replication (22)	1.46 × 1055	189	7

^*Mycobrowser annotation.

^†Prokka annotation.

^‡Multiple variants in MAB_4141 have been collapsed and the lowest P-value is shown. See SI Appendix for full list of homoplasic SNPs ranked by their enrichment in clusters.

^§Gram-negative bacteria possess UvrD and a closely related homolog, Rep, while gram-positive bacteria possess a single homolog called PcrA (22).

Discussion

In this study, we have pooled a large global sample of Mab genome sequences and examined evolutionary rates and positive selection with the goal of understanding the emergence of Mab clusters. We present evidence supporting a slower evolutionary rate within DCs compared to the long internal branches of the phylogeny and further data associating genetic variation in DNA repair genes with the slower evolutionary rate. We posit that changes to the mutation rate underlie the dense clustering in Mab, and that while other effects such as population or transmission bottlenecks and sampling bias may contribute to the low genetic diversity within clusters, they are not sufficient to explain the observed phylogenetic structure given the available data. Phylogenetic clustering is often used to infer transmission, although previous work has indicated that clustering can depend on many other factors (23, 24). Taken together with epidemiological evidence refuting widespread person-to-person transmission, our findings challenge the model of phylogenetic clustering in Mab as evidence of person-to-person transmission as a major mode of Mab dissemination.

Previous studies have restricted molecular clock rate estimation to Mab clusters, most likely because of the lack of temporal signal at the subspecies-level. By pruning the Mab tree while preserving the majority of the subspecies diversity, we were able to measure a molecular clock rate for Mab spanning the time scale from the common ancestor of the subspecies to the present. Estimates of evolutionary rates are recognized to depend on the time scale of measurement. In such examples, the clock rate usually decreases with increasing time depth (25). We observe the opposite effect where the clock rate increases with time depth. The pruned Mab tree most likely reflects an evolutionary process taking place predominantly in environmental reservoirs, in contrast to phylogenetic clusters where terminal branches reflect evolution within the human lung environment. We also note that our measurement of within-host mutation rate in Mab was independent of phylogenetic inference and the full temporal signal in our sample. Nevertheless, the measured rate within-host was comparable to the clock rate measured previously using coalescent analysis within clusters (Fig. 2C).

The molecular clock rate is influenced by both the spontaneous mutation rate and the generation time of an organism. Changes in generation time could result from physiological constraints of the human lung environment, if for example, differences in nutrient availability change the growth rate in the lung compared to those living in environmental reservoirs. Generation times and spontaneous mutation rates can also be modified by genetic changes, for example to DNA replication or repair machinery. The observation that some lineages have acquired phenotypes that are beneficial in the host environment (4) raises the possibility that genetic changes that underlie or associate with this adaptation affect mutation rate, either directly by affecting DNA replication or repair machinery, or indirectly by affecting the growth rate. Ruis et al. reported differences in mutational signatures along internal branches of the subspecies phylogeny compared to internal branches of DCs, supporting a change in the mutational process within some clades of the tree (13).

We scanned the core genomes for variants enriched in the DCs in M. a. abscessus, the subspecies for which we have the most WGS data and for which there are the highest number of clusters. Notably, we observed several cluster-enriched mutations in UvrD-like helicases. UvrD is a superfamily one helicase that is involved in DNA repair and is widely conserved in gram-negative bacteria. UvrD1 in M. tuberculosis plays an important role both in nucleotide excision repair (NER) and in pathogenesis and persistence because its DNA repair activity confers tolerance to reactive oxygen intermediates and reactive nitrogen intermediates (16). Mutations that promote the activity of UvrD1 and its homologs may therefore increase resistance to genotoxic stress in the context of the human lung environment, with the additional consequence of decreasing the spontaneous mutation rate. We also observed mutations in UvrD-like helicases occurring along the cluster-defining branch of the large M. a. massiliense cluster, further supporting the role of the NER pathway in Mab pathogenesis.

Our study has several limitations. Our dataset overrepresented clinical isolates from the UK where routine WGS is part of Mab surveillance. Oversampling of some lineages and geographies may mean that we have not adequately captured the global M. abscessus diversity and that we may have focused on large phylogenetic clusters because of their geographic distribution and not because of their clinical significance. While our results support a faster molecular clock rate in the long internal branches of the phylogeny compared to clustered lineages, we are unable to confirm the underlying mechanism for a change in the clock rate. Our in vitro assays indicate that changes in the clock rate are at least in part determined by variation in the spontaneous mutation rate. However, we note that the spontaneous mutation rate measured for DC v. non-DC isolates of M. a. abscessus differed by roughly fivefold, while our coalescent analysis and simulation studies support a slightly higher (between fivefold and 15-fold) clock rate differential. It is therefore plausible that additional factors such as generation time contribute to changes in the molecular clock rate. Differences in growth conditions across environments are difficult to quantify as Mab is ubiquitous in both natural and built environments, but nutrient availability likely differs greatly between natural reservoirs such as soil and water compared to man-made reservoirs such as showerheads. A healthy lung supports relatively little bacterial growth, but inflammation increases nutrient availability by promoting mucus production and vascular leakage into the airway (26). Our analysis of variants associated with phylogenetic clustering is also limited in its ability to identify true causal genotypes or those that are enriched in clusters due to hitchhiking to other adaptive variants.

It is possible that our recombination removal procedure does not detect every mutation imported by recombination, and that unpurged recombination events may contribute to the high clock rate we measure in the pruned phylogeny. We believe any effect of recombination is likely minimal for several reasons: 1) Recombination rates vary widely across branches, making unpurged recombination events more likely to erode temporal signal than to cause a systematic error in clock rate estimates (SI Appendix, Fig. S5A) and 2) our clock rate estimates were robust to choice of algorithm or to the stringency of recombination search (SI Appendix, Fig. S5B). Moreover, the algorithms we used to purge recombination events were also used in previous studies that published M. abscessus phylogenies (4, 9, 12, 13). If these algorithms fail to purge enough recombination events to result in extreme distortion of clock rate estimates, they must also create distortion in phylogenies and the inference that follows from them. This strengthens the argument that the presence of phylogenetic clustering is not sufficient to infer widespread transmission.

While our study argues against transmission as the main driver of Mab dissemination and the emergence of DCs, these findings do not address the geographic distribution of Mab within clusters. It is possible that DC-associated variants such as the UvrD/Rep mutations described in this study are acquired ancestrally to the emergence of each DC, then circulated via contaminated household or medical equipment. This is similar to the model proposed by Bryant and colleagues, in which variation acquired along the branches ancestral to DCs creates large phenotypic changes, conferring a sudden increase in intracellular survival, tolerance to nitric oxide, and amikacin resistance (27). It is likely that emergence of clades with higher virulence potential involves acquisition of multiple genotypes including core genome mutations and horizontal gene transfer. Convergent evolution can also act in concert with the Bass Becking hypothesis that “everything is everywhere but the environment selects.” This could generate DCs that have a common core genome across many genetic loci and are prevalent in human infection across a wide range of geographies. Currently there is little publicly available sequencing data from environmental isolates. Isolating and sequencing M. abscessus from the environment could aid in understanding why clustered Mab isolates are detected in distant geographic locations. Future work could also compare genotypes and mutation rates in environmental samples compared to clinical isolates.

In summary, we present evidence for changes in the molecular clock rate that contribute to the dense phylogenetic clustering observed in Mab. These results argue against person-to-person transmission as the primary driver of clustering in the Mab phylogeny. Continuing genomic surveillance is needed to understand how Mab may be adapting to human infection and to characterize the risk of transmission to at-risk patients. We argue further that the phylogeographic structure of Mab is influenced by its complex ecological history as an emerging facultative pathogen that should be considered carefully in future analysis of Mab genomic data. Future work should extend these analyses by examining differences in mutation rates and phenotypes between clinical isolates and environmental strains for evidence of host adaptation.

Methods

Data Sources and Description.

WGS data were curated from published studies (4, 7, 10, 12, 28–30) and additional data were downloaded from the NCBI Sequence Read Archive (SRA). A search of “abscessus” was conducted on October 4, 2019 using the filters “DNA” and “Illumina”. Sequences were selected that had an associated date of collection with the year at minimum. Collection dates were obtained either from the literature, SRA metadata table, or provided by the authors. A summary of all samples used in this study is provided in SI Appendix, Table S1. Raw sequencing reads were downloaded using the NCBI SRA toolkit v2.9.6 (31).

Quality Control of Illumina Reads.

Reads were trimmed and filtered with trimmomatic v0.36 (32). To detect contaminated samples, the mean GC content of the reads were compared to a modelled normal distribution of GC content using fastqc v0.11.8 (33). If the sum of the deviations from the normal distribution represented more than 30% of the reads, the sequencing run was excluded from further analysis.

Subspecies Assignment and Genome Assembly.

Multiple sequencing runs corresponding to one biological sample were combined into one file. Reads were then assembled de novo using SPAdes v3.11.1 (34). Assemblies were compared to a reference sequence for each of the three subspecies (SI Appendix, Table S3) and a whole-genome-based average nucleotide identity (gANI) score was calculated using fastANI v1.2 (35). The resulting alignments were assigned to a subspecies based on having gANI of at least 98% with exactly one reference strain (SI Appendix, Fig. S2). 21/1,572 isolates that did not have at least 98% ANI with any reference strain were excluded. No isolates matched more than one reference genome. After assigning isolates to a subspecies, reads were mapped to the corresponding reference genome using BWA MEM v0.7.17 (36). After alignment, isolates with <80% coverage of the reference genome at a depth of at least 20× or with missing data at >30% of sites were excluded.

Variant Calling and SNP Filtering.

Variants were called using Pilon v1.23 (37). We excluded ambiguous calls as wells as indels and MNVs. We also excluded calls with low coverage using a minimum depth of either 10% of the mean coverage, or five, whichever was greater (default settings for Pilon –mindepth). We used bcftools v1.10.2 (38) to exclude calls with mapping quality or base quality scores <20.

Core Genome Inference.

Unlike the professional mycobacterial pathogens M. tuberculosis or M. leprae, which have small accessory genomes, M. abscessus has a large, open accessory genome (39). To define the core genome for phylogenetic analysis and to exclude loci that are highly divergent from the reference sequence, we masked all loci for downstream analysis where read depth was <20× in more than 5% of the isolates in our sample set. The remaining loci were compared to an estimate of the core genome calculated by Roary (40), and the two methods yielded nearly identical results. We further excluded regions with average mapping quality scores or average-based quality scores <20. Finally, we used GenMap v1.1 (41) to calculate mappability scores across each reference genome as defined in reference (42). We then calculated a mappability pileup score for each base position as described in reference (43) and excluded loci where the mappability pileup score was <95% from downstream analysis. In total, we excluded 14.4%, 13.5%, and 15.1% of the M. a. abscessus, M. a. massiliense, and M. a. bolletii reference genome length, respectively.

Phylogenetic Tree Inference.

A full-length sequence alignment was created using a custom script, masking loci excluded by our filtering criteria described above. Gubbins v2.4.1 (44) was used to predict and mask regions of variation produced by recombination from the full-length alignment. A recombination-free SNP alignment generated by Gubbins was then used to build a phylogeny for each subspecies using IQ-TREE v1.6.12 (45), using the ModelFinder Plus (-mfp) option and 1,000 bootstrap replicates. Clusters were defined as previously described (4) using a one-dimensional scanning statistic to identify clades with a distribution of branch lengths that is shorter than the distribution of branch lengths in the rest of the tree (46).

Coalescent Analysis.

Temporal signal was assessed using root-to-tip regression. In both M. a. abscessus and M. a. massiliense subspecies trees and all three major clusters, a slightly negative slope was observed, indicated lack of temporal signal. We recognized that the high degree of clustering and short time span over which the clustered isolates were sampled may confound the genetic temporal signal. To address this, we pruned the trees down to taxa that preserved 95% of the relative tree length of the original trees using Treemmer v0.3 (47). A threshold of 95% eliminated most of the dense clustering, but preserved the long branches of the original tree (Fig. 2B and SI Appendix, Fig. S4B). Using the pruned tree, we observed evidence of temporal signal in the MAB subspecies by regressing the root-to-tip distance for each sample against the date of collection. The linear regression was performed using the statsmodels package and the Pearson correlation coefficient was calculated using the pingouin package in Python. We generated a SNP alignment including only the taxa present in the pruned M. a. abscessus tree and used this as input for coalescent analysis using BEAST v2.6.3 (15). ModelTest-NG v0.1.6 (48) was used to select the site model GTR+4. MegaX (49) was used to perform a ML test to determine that a relaxed clock model is best suited for our data. This choice was confirmed using the nested sampling algorithm to compare marginal likelihoods for a strict v. relaxed clock model. We then used nested sampling to compare various tree priors. We used BETS to confirm the presence of temporal signal by comparing marginal likelihoods when the tip dates were used to those when all samples were assumed to be isochronous (SI Appendix, Table S4). Based on marginal likelihood values, we ran both a coalescent constant population model and a Bayesian skyline coalescent model, assuming a relaxed clock and including the sampling dates for both models. We used the ML phylogeny of the pruned dataset as a starting tree. For each model we ran two independent chains for at least 100 million states to ensure convergence.

To test the effects of different recombination search procedures, we used Gubins v3.2.1 using 1) the default parameters and 2) more stringent search parameters by decreasing the minimum number of SNPs to call a recombination to two, increasing the significance threshold to P = 0.10, and raising the trimming ratio to 1.5 to disfavor trimming of recombination edges. We also implemented recombination search using an alternative algorithm, ClonalFrameML (50). We then used BEAST to estimate clock rates using each alignment, assuming a constant population tree prior (SI Appendix, Fig. S4B). To test the effects of which isolates are sampled by Treemmer, we ran Treemmer 10 times to obtain 10 independent samples. We then repeated our analysis using BEAST, again assuming a constant population tree prior (SI Appendix, Fig. S4C).

Phylogenetic Dating.

To attempt to date the full MAB phylogeny, we ran both coalescent constant population and Bayesian skyline coalescent models assuming a relaxed clock without sampling dates. We set the mean UCLD clock rate to the values estimated from our coalescent analysis of the pruned tree. We ran each chain for at least 100 million states.

Estimation of Mutation Rate Within-Host.

We identified 90 patients with two or more M. a. abscessus isolates. Because coinfection with multiple M. abscessus clones is common in CF patients, we first sought to exclude any isolate pairs that represented distinct clonal infections. We selected a threshold of 20 SNPs by comparing the distribution of all possible within-host isolate pairs to the distribution of pairwise SNP distances within Mab clusters. We further excluded any isolates where >20% of the genome is predicted to be rearranged by recombination. For the remaining isolate pairs we regressed the pairwise SNP distance between the first and last isolates sampled from each patient on the time between the collection of each sample. The linear regression was performed using the Python statsmodels package v0.12.2. The slope of the regression line was used as an estimate of the number of SNPs accumulated over time.

Ancestry Simulations.

We used msprime v1.0.2 (51) to simulate ancestries to model two subpopulations: 1) a “clustered” subpopulation (A) and 2) an “unclustered” subpopulation (B). We assumed a constant population size and initially set the total effective population size (N_e) to 8,000, the N_e estimated in our coalescent analysis of the pruned Mab tree. We chose a range of realistic N_e for subpopulation A, N_e^A by running a separate coalescent analysis of M. a. abscessus cluster 1 using BEAST, assuming a constant population size, relaxed clock, and a range of UCLD clock rates spanning 1.7 × 10⁻⁷ SNPs/site/y, as previously reported for this cluster by ref. (13) to 8.5 × 10⁻⁶ SNPs/site/y, a value well above the 95% HPD interval for our estimate of the clock rate in the pruned tree (SI Appendix, Table S5). We assumed the N_e of subpopulation B to be: N_e^B = N_e^total−N_e^A. The number of samples and sampling dates supplied to the model for subpopulations A and B are drawn from the true values for Mab cluster 1 and from all “unclustered” isolates in the tree, respectively. We assumed an initial mutation rate of 1 × 10⁻⁹ mutations/generation with a sequence length of 5 Mb and a divergence time of 30,000 generations between subpopulations A and B. We also assumed the nucleotide subsitution model estimated using BEAST.

We first simulated a scenario where the mutation rate remains constant, varying N_e^A over the range shown in SI Appendix, Table S5. For each simulated ancestry we overlaid mutation events, supplying the model with a starting substitution rate of 1 × 10⁻⁹ mutations/generation, a nucleotide substitution model estimated using BEAST, and a sequence length of 5 Mb. We then extracted the simulated genotypes and constructed a neighbor joining (NJ) tree using the Phylo module of the Python package biopython v1.79. The NJ tree was used to exclude any simulations where the two subpopulations did not exhibit reciprocal monophyly. We used the extracted genotypes to estimate the degree of clustering in each tree by comparing the mean pairwise SNP distance in subpopulation A (d_A) to the mean pairwise SNP distance in subpopulation B (d_B). For each simulated condition, we performed 1,000 replicates. To simulate ancestries with a mutation rate change, we repeated the above procedure conservatively assuming N_e^A = 718. We modeled a per generation mutation rate change starting at the common ancestor node for subpopulation A and ranging from onefold (no change) to 1/25-fold spanning a range of reasonable rate changes based on the range of clock rate estimates observed in this study and reported in the literature (Fig. 2B). For all simulated conditions, we the simulation over a range of values for N_e^total spanning the 95% HPD intervals for the N_e estimates using the Treemmer tree under both the constant coalescent and coalescent skyline tree priors.

Scan for Variants Enriched in Clinically Successful Clades.

To search for variants associated with DCs, we first inferred the number of independent arisals in the MAB phylogeny for each SNP using SNPPar v1.0 (52). Briefly, SNPPar uses TreeTime to perform an ancestral sequence reconstruction at each node, then infers mutation events arising on each branch of the tree. We quantified the number of independent mutation events for each SNP. We considered only mutation events where the derived call was the minor allele and the ancestral call was the major allele. Because of the extreme phylogenetic clustering in Mab, we defined the major and minor alleles using the pruned alignment generated by Treemmer. For all homoplasic SNPs (occurring two or more times independently in the tree), we used Fisher’s exact test to quantify the enrichment of that variant within phylogenetic clusters and reported FDR-adjusted P-values to correct for multiple testing. We conducted two separate analyses: a) focusing on the three largest clusters in the M. A. abscessus tree (described in the main text), and b) considering all clusters identified by this approach. In each case we obtained similar results. For both analyses, full lists of variants ranked by their enrichment scores are provided in SI Appendix, Supplementary Data. For each variant reported in Table 2, we manually examined SNPPar output to confirm that the variant was acquired independently along branches ancestral to at least two separate clusters. We also used SNPPar to identify all mutation events occurring on the branch ancestral to M. a. massiliense cluster 1.

M. abscessus Culture.

M. abscessus clinical isolates used for fluctuation assays were isolated from patients in Taiwan and in Boston, MA, USA. All M. abscessus strains were cultured in Middlebrook 7H9 broth (271,310, BD Diagnostics, Franklin Lakes, NJ, USA) with 0.2% (v/v) glycerol, 0.05% (v/v) Tween-80 (P1754, Sigma-Aldrich, St. Louis, MO, USA), and 10% (v/v) OADC (90,000-614, VWR, Radnor, PA, USA). Proliferation rates were determined by inoculating 5 mL media with 500,000 colony forming units (cfu), then measuring OD600 at 24 and 48 h postinoculation. Proliferation rate was calculated as follows: proliferation rate in doublings per day = log₂(OD600 at 48 h/OD600 at 24 h).

Fluctuation Assay.

Fluctuation assays were performed as described previously (19) with adaptations for M. abscessus. M. abscessus strains were grown to saturation, then diluted to an OD600 < 0.01 and grown overnight until cultures reached an OD600 = 0.6 to 0.8. These cultures were used for fluctuation assays as well as determination of baseline amikacin sensitivity. Baseline amikacin sensitivity was tested by spreading dilutions of 10 million, 100 million, or 1 billion cfu on agar 7H10 (262,710, BD Diagnostics) + OADC (90,000-614, VWR) + 300 µg/mL amikacin sulfate (A2324, Sigma-Aldrich) plates. For fluctuation assays, 5,000 cfu were plated in 1 mL media in each of 21 wells of a 24-well culture plate (10,861-558, VWR), and initial cultures were plated for cfu on 7H10 + OADC to confirm number of cells in initial culture. Then, 24-well plates were incubated shaking at 150 rpm for 6 d at 37 °C. 18 wells of each strain were plated onto six-well plates containing 5 mL 7H10 agar + OADC + 300 µg/mL amikacin sulfate in each well. Amikacin was used as the antibiotic selection for the assay because all tested strains had low baseline resistance (SI Appendix, Table S8), which is required for the validity of the fluctuation assay. The remaining three wells from each strain were diluted and plated on 7H10 + OADC plates to enumerate average final cfu. Colonies on all agar plates were counted after 5 d incubation at 37 °C. Mutation rate per generation per total number of bases that can be mutated to confer amikacin resistance was calculated using the Shinyflan application for the flan R package v0.9 (19) using the ML estimation method, assumption of unknown fitness effects of mutations, the exponential (LD model) distribution of mutant lifetime, and a Winsor parameter of 1,024. One strain (BWH-E) was excluded from analysis due to highly variable results across replicates (SI Appendix, Fig. S11B). Mutation rates in clustered and unclustered samples were compared using a two-tailed, unpaired t test with GraphPad Prism v9.3.1.

Supplementary Material

Appendix 01 (PDF)

Dataset S01 (CSV)

Dataset S02 (CSV)

Dataset S03 (CSV)

Dataset S04 (CSV)

Dataset S05 (CSV)

Data, Materials, and Software Availability

Notebooks, snakemake piplines, custom scripts, conda environments, and BEAST XML files are available at: https://github.com/nicolettacommins/mab_mutation_rates_2022. Accession IDs for all publicly available dataused in this study are available in the supplementary file: mabsc_all_sample_metadata.csv. All study data are included in the article and/or supporting information.