Complete genome assemblies were prepared or publicly available for 11 named or provisional species in the genus Bordetella and were compared in the present study to identify rearrangements of gene order. A neighbor-joining phylogenetic reconstruction of these species, plus two distantly related Bordetella isolates (H567 and J329), rooted with outgroup Achromobacter xylosoxidans, clustered the classic bordetellae apart from the otherwise disparate species branches (Fig. 1), consistent with previous studies (10). The number of single nucleotide polymorphisms (SNPs) observed in each species varied. Isolates of B. bronchiseptica derived primarily from two distinct clades, complex I and complex IV, which differ by ∼35,000 SNPs (Fig. S1A and B in the supplemental material). The distant B. bronchiseptica strain I328 (37) did not belong to either complex (Fig. 1) and was separated from other B. bronchiseptica isolates by ∼60,000 SNPs. Isolates of B. pertussis (complex II), B. parapertussis (complex III), and B. holmesii were much more similar, exhibiting <100 SNP differences between most isolates in each (Fig. S1C). Complete genomes from isolates of B. hinzii (n = 3) and Bordetella trematum (n = 2) differed by ∼9,600 and ∼11,000 SNPS, respectively.

Among the various Bordetella species, only B. holmesii and B. parapertussis displayed signs of chromosome structure variation in alignments of the isolate genomes analyzed here (Fig. 2). Genomes of B. holmesii (n = 63) comprised 11 discrete chromosome structures (Fig. S2A). The most abundant structure was observed in 48 isolates (cluster BH-01), and six other structures each differed from it by a single, unique inversion (Fig. 2A). All observed inversions were symmetric, encompassing the replication origin-terminus axis, and flanked by opposing copies of IS elements. Other structures differed by small deletions, not rearrangements, including cluster BH-03 (Data Set S1). Genomes of B. parapertussis (n = 79) exhibited two predominant chromosome structures of comparable levels of abundance, varying by a symmetric inversion between copies of IS1001, which largely followed phylogenetic boundaries (Fig. 2B). In both species, unique structures were also observed in only one isolate (Fig. S2B, singletons). Predicted protein-coding gene contents adjacent to select inversion breakpoints observed in B. holmesii and B. parapertussis are listed in Data Set S1. These results indicated that inversions like those frequently observed in B. pertussis also occur in at least a few other Bordetella species.

The chromosome of B. pertussis displays structural fluidity (36), and the 469 complete genomes of B. pertussis included here were aligned in 109,746 pairwise comparisons to identify rearrangements and group isolates with colinear chromosomes. These genomes exhibited 107 unique chromosome structures, including 32 clusters of two or more isolates (e.g., cluster BP-01 and cluster BP-02) and 75 singletons observed in only one isolate (Fig. S2C). The abundances of unique structures varied, with the largest cluster including 95/469 (20%) of the isolates; the distribution suggested that structural diversity remains undersampled, as observed previously (36). The data set here sampled years 2000 to 2016 (n = 370) most densely, and a rarefaction curve of unique structures recovered during this time period did not approach saturation, as expected (Fig. S2B). Links between clusters of unique structures reported here and those reported previously are listed in Table S2.

Circulating isolates of B. pertussis exhibited little gene sequence diversity (13, 38), but the distribution of pairwise SNP distances between isolates within the largest clusters varied. While cluster BP-01 and cluster BP-05 appeared relatively clonal, exhibiting <15 SNPs between isolates, SNP distances within others, such as cluster BP-02, cluster BP-03, cluster BP-04, cluster BP-06, and cluster BP-08, measured up to 32 SNPs (Fig. S3A). These differences largely reflected the temporal spread of isolates within a cluster. Isolates in clusters with fewer pairwise SNPs were recovered during narrower time frames than those with more SNPs and wider temporal distributions (Fig. S3B). These data illustrate that certain chromosome structures predominate due to their stability (e.g., cluster BP-02), while the abundances of others can be attributed to recent expansion (e.g., cluster BP-01).

Pairwise alignments between a subset of genomes representing each of the unique structures in B. pertussis, B. holmesii, and B. parapertussis were investigated further to identify pairs that differed by either a single inversion (Fig. 3A) or insertion/deletion (i.e., a gap of >1,500 bp) but were otherwise colinear. Eighty-two pairs were identified among the 107 B. pertussis structures, of which 49 differed by symmetric inversion, 11 by asymmetric inversion, and 22 by insertion/deletion. Symmetric inversions of varied sizes were observed near both the replication origin (oriC) and terminus (dif) and were larger than asymmetric inversions on average (Fig. 3B).

To investigate relationships among discrete B. pertussis chromosome structures, the identified pairs were used to construct an undirected network that connected 71 unique structures (nodes) by inferred single inversion or insertion/deletion events (Fig. 3C, edges). Sixty-two structures were connected to a single graph, representing 407 isolates (87%) in the data set. The structures in vaccine and laboratory reference strains, such as strain 134 (B202 [accession no. CP016338]), were not connected to the larger network of recent clinical isolates but instead formed small groups of 2 to 4 related structures. The remaining 36 structures, representing 49 isolates that included Tohama I, could not be connected to the network or each other.

Some of the largest structural clusters were located near the network core and connected through loops of parallel, symmetric inversions. Four of these clusters also exhibited the highest degrees of centrality, indicating their significant connectivity to other structures in the network (Fig. S4A). In contrast, cluster BP-01, which was the most abundant, appeared closer to the periphery, following a path of successive rearrangements that diverged from the network core, and therefore, it deviated from many other abundant clusters. Singletons also formed peripheral nodes, often crowded around an abundant cluster, suggesting they represent rare aberrations from predominant chromosome structures. Network edges representing insertion/deletion and asymmetric inversion were also largely responsible for connecting small, peripheral nodes to the network, while edges at the core comprised symmetric inversions. Similar patterns were observed in smaller networks constructed of chromosome structures observed in B. holmesii and B. parapertussis (Fig. S5). Taken together, these results demonstrated the extent to which symmetric inversion has shaped and continues to shape the genome compared to the relative rarity of asymmetric and insertion/deletion events.

The phylogenetic history of discrete chromosome structures observed in B. pertussis was further explored to infer the divergence of specific structures and assign putative directionality to network edges shown in Fig. 3C, where possible. A time-scaled Bayesian reconstruction of 438 isolates from the ptxP3-prn2-ptxA1 background was calculated using 908 variable, core nucleotides (Fig. S6) with an estimated mutation rate of 1.30 × 10⁻⁷ substitutions per site per year. The resulting tree topology separated isolates into two large clades corresponding to the fimH1 (fim3-1) and fimH2 (fim3-2) alleles, consistent with previous studies (13, 36), and large structural clusters were restricted to a single clade.

Within the fimH1 clade, there was insufficient sampling to discriminate the ancestral order of cluster BP-03 and cluster BP-04. Genomes exhibiting both structures appeared near the tree root and in isolates dating back to 2000 (Fig. S6; Table S1). However, inspection of more contemporary subclades revealed evidence for multiple inversion events between the two clusters, consistent with the wide range of pairwise SNPs in each (Fig. S3). The divergence dates of some chromosome structures could be estimated from the phylogeny (e.g., cluster BP-01 and cluster BP-07), while multiple events were predicted for others (e.g., cluster BP-10) (Fig. S6). Likewise, the fimH2 allele background included the large central cluster BP-02, and multiple inversion events were predicted that gave rise to unrelated isolates with colinear chromosomes (e.g., cluster BP-06 and cluster BP-08) (Fig. S6). While some inversions have occurred repeatedly or reversibly, most chromosome structures exhibited stable phylogenetic linkage, suggesting that rearrangement events are either rare or under selection.

Observed inversions among genomes of B. pertussis, B. parapertussis, and B. holmesii were primarily flanked by IS elements, which appeared in multiple copies throughout each genome. All repetitive sequence content was quantified independently of gene annotation by counting unique 15-bp sequences (15-mers) in a representative isolate genome from each of the species analyzed here. Genomes of B. pertussis and B. holmesii each included a set of highly abundant 15-mers, while B. parapertussis, B. petrii, B. bronchialis, Bordetella sp. strain H567, B. bronchiseptica I328, B. pseudohinzii, and B. flabilis included decreased numbers of abundant 15-mers (Fig. S7A). Representatives from all other species included in Fig. 1 did not exhibit any abundant 15-mers. A similar pattern resulted from comparison of predicted protein-coding gene sequences (Fig. S7B), and abundant genes corresponded to transposases encoded in various IS elements in B. pertussis, B. holmesii, B. parapertussis, B. petrii, B. bronchialis, Bordetella sp. H567, B. bronchiseptica I328, B. pseudohinzii, and B. flabilis (Table 1). The complements and copy numbers of the IS elements detected varied, with B. pertussis carrying the greatest number. While all species considered here included a few duplicated protein-coding genes of other functions, those absent from Table 1 did not encode any, including transposases, with >2 copies. These results support the expected correlation between genome repeat contents, specifically IS elements, and the observation of rearrangement.

Most observed inversions were symmetric, and their “balance” was assessed by calculating the ratio of breakpoint distances relative to either the replication origin (oriC) or terminus (dif), whichever was closer, according to equation 1. The size of replichores, oppositely replicated chromosome halves, in most of the 107 unique B. pertussis chromosome structures were balanced, as the natural logs of their ratios were close to zero (Fig. 4A), particularly among those representing the largest clusters, and most of the deviations were observed in small clusters or singletons (Fig. S4B). Replichore sizes were similarly balanced in B. holmesii (Fig. S4C). Although replichore sizes in B. parapertussis appeared less balanced, the two predominant structures produced similar results (Fig. S4D). Many of the 49 symmetric inversions observed in B. pertussis were also balanced (Fig. 4B), maintaining similar relative replichore sizes before and after inversion. Some unbalanced inversions were also observed, and these appeared more common near the replication terminus than near the origin.

Linear regression of the right and left relative breakpoint distances in 55 symmetric inversions observed in B. pertussis and B. holmesii further illustrated the bias toward balance (Fig. 4C). The resulting linear model was used to predict all possible symmetric inversions in a representative isolate genome from each species based on the positions of detected IS elements (Table 1). As expected, the number of predicted inversions increased with IS element copy number and was greater than the number observed in the current data set of B. pertussis, B. holmesii, and B. parapertussis genomes (Fig. 5). Additional breakpoints were determined from alignment of B. pertussis J549 (cluster BP-04) with representatives of all 106 other structures but were still fewer than those derived from predicted inversions (Fig. 5A). Many individual copies of IS481 in J549 were predicted to engage in multiple inversions, pairing with as many as 16 other insertions, reflecting the particularly high density of this IS element and varied balance points of observed inversions. Only one genome sequence each was available for B. petrii and Bordetella sp. H567, but symmetric inversions were predicted based on the locations of their IS elements (Table 1). No inversions were predicted between IS elements in genomes of B. bronchiseptica I328, B. bronchialis, B. flabilis, or B. pseudohinzii, as suitable pairs could not be matched within the prediction window. This generalized model of symmetric inversion revealed considerable rearrangement potential, much of which remains unobserved. The model also illustrated that inversion potential does not simply depend on IS element copy number but on relative position and orientation as well.

The Centers for Disease Control and Prevention’s (CDC’s) collection includes U.S. B. pertussis isolates collected through surveillance and during outbreaks. In total, 469 assembled genomes were included in the current study based on availability, and most were selected for sequencing as part of previous studies (Table S1 in the supplemental material). One set (n = 170) was selected to capture potential geographic diversity among 28 U.S. states in 2000 to 2013 (71). Another selection (n = 108) focused on isolates obtained through the Enhanced Pertussis Surveillance/Emerging Infection Program Network (EPS) (72) in 2011 to 2014, which prioritized isolates from hospitalized cases and forced sampling of every possible combination of year, state, vaccination status, and age group. Additionally, 133 isolates were selected prospectively from sporadic or EPS submission in 2014 to 2016; 31 epidemic isolates were sequenced previously (35); and 27 others, such as vaccine reference strains and novel immunogen-deficient genotypes, were also included (73–76). U.S. isolates of other Bordetella species were selected from the CDC collection for sequencing across as many years as available, but with additional emphasis on those recovered after 2000 from patients presenting respiratory illness. A few isolates sequenced previously were also included (37, 77, 78). The present study featured new genome assemblies for 125 B. pertussis, 11 B. bronchiseptica, 60 B. holmesii, 78 B. parapertussis, 2 B. trematum, 1 B. hinzii, and 1 other Bordetella species isolate as outlined in Table S1.

Isolates were cultured on Regan-Lowe agar without cephalexin for 72 h at 37°C. Genomic DNA isolation and purification was performed using the Gentra Puregene yeast/bacteria kit (Qiagen, Valencia, CA) with slight modifications. Briefly, two aliquots of approximately 1 × 10⁹ bacterial cells were harvested and resuspended in 500 μl of 0.85% sterile saline and then pelleted by centrifugation for 1 min at 16,000 × g. Recovered genomic DNA was resuspended in 100 μl of DNA hydration solution. Aliquots were quantified using a Nanodrop 2000 (Thermo Fisher Scientific, Inc.; Wilmington, DE). Additional chloroform purification was performed to remove polysaccharides from preparations of all species except B. pertussis (79).

Whole-genome shotgun sequencing was performed using a combination of the PacBio RSII (Pacific Biosciences, Menlo Park, CA), Illumina HiSeq/MiSeq (Illumina, San Diego, CA), and Argus (OpGen, Gaithersburg, MA) platforms as described previously (35). Genomic DNA libraries were prepared for PacBio sequencing using the SMRTbell template preparation kit 1.0 and polymerase binding kit P4 or P6, while Illumina libraries were prepared using the NEB ultra library preparation kit (New England Biolabs, Ipswich, MA). De novo assembly was performed using either the Hierarchical Genome Assembly Process (HGAP) (version 3; Pacific Biosciences) (80) or FLYE (version 2.4.2) (81). The resulting consensus sequences were circularized using either circlator (version 1.5.1) (82) or gepard (version 1.30) (83). Where necessary, hybrid de novo assembly of PacBio and Illumina sequences was performed using Unicycler (version 0.4.0) (84). All assemblies were reordered to start at the coding region for glucose-inhibited cell division protein A (gidA). Assemblies were confirmed by comparison to restriction digest optical maps using the Argus system (OpGen) with MapSolver (version 2.1.1; OpGen) and further polished by mapping Illumina reads using CLC Genomics Workbench (version 10) (CLC bio, Boston, MA). Assemblies were annotated using the National Center for Biotechnology Information (NCBI) Prokaryotic Genome Annotation Pipeline (PGAP). Genome sequence-based allele characterization of B. pertussis molecular typing loci (prn, ptxP, ptxA, ptxB, and fimH) was assigned by genome alignment to a curated set of wild-type and deficient alleles using high stringency.

A neighbor-joining phylogenetic reconstruction of the genus Bordetella was determined from pairwise Mash (version 2.0) (85) distances using mashtree (version 0.29) (86). Further phylogenetic reconstruction within species B. holmesii and B. parapertussis was calculated using kSNP3 (87), with a kmer size of 23 bp. Maximum-parsimony trees and pairwise SNP distances were calculated from core variable positions. Trees were annotated with iTOL (version 4) (88). The phylogeny of B. pertussis isolates from the predominant ptxP3-prn2-ptxA1 background was reconstructed from SNPs determined relative to the reference strain C734 (accession number CP013078) using snippy (version 3.1) (https://github.com/tseemann/snippy) after masking all IS elements with N’s. Isolates with <75% coverage breadth were removed (n = 6). A maximum-likelihood phylogeny was estimated from the core SNP alignment using RAxML (version 8.1.16) (89), and a temporal correlation of 0.5231 was determined using TempEst (version 1.5.1) (90). Bayesian divergence time estimation from the core alignment was performed using BEAST (version 1.8.3) (91) with an HKY substitution model, four-category gamma site heterogeneity, and strict clock. Markov chain length was 100,000,000 with parameters sampled every 1,000 states. Model convergence and parameter expected sample sizes were evaluated with tracer (version 1.6) (92), and a maximum clade credibility tree was calculated after a burn-in of 10,000,000 states using TreeAnnotator (version 1.10.0) (91) with common ancestor heights. Tree visualization and annotation were performed with the ggtree R package (version 1.10.5) (93).

Repeat content in each genome was quantified by counting the abundance of unique 15-mers with jellyfish (version 2.2.6) (94). Protein-coding genes within each genome were clustered using CD-HIT (version 4.6) (95) with default parameters, except for match cutoffs, which were set at 95% nucleotide sequence identity and 90% length difference. Detected IS elements were annotated using ISfinder (96), and copy numbers were quantified by BLASTn query of a representative genome from each species.

Genomes from each species were aligned in all pairwise combinations using progressiveMauve (97) with optimized parameters (–seed-weight = 16, –hmm-identity = 0.85) and clustered according to their structures as described previously (36). To improve the accuracy of B. pertussis whole-genome alignments, adjacent insertions of IS481 with the same orientation and separated only by their 6-bp target sequences were first collapsed to a single insertion using a custom Perl script. A rarefaction curve of structures observed among B. pertussis isolates collected during 2010 to 2016 was calculated using the vegan R package (version 2.3.0) (98). Alignments were further analyzed using custom Perl scripts to identify pairs of B. pertussis genome clusters differing by only a single inversion or single insertion/deletion of >1,500 bp, set to exceed the length of individual IS elements, and to characterize inversion symmetry measured relative to the positions of oriC and dif (99, 100). The resulting list was used to construct an undirected network of rearrangement and insertion/deletion events (edges) between genome clusters (nodes) using the network (version 1.13.0) (101) and sna (version 2.4) (102) R packages. The centralization of each node in the network was computed as degree of centrality using sna.

The balance of symmetric inversions was calculated from the left and right breakpoint distances, measured from the nearest replication origin or terminus as follows:(1)balance=| ln(left/right) |

The replichore balance of discrete chromosome structures was determined by the same equation using the distances between the replication origin and terminus. Symmetric inversions observed among 49 B. pertussis alignment pairs were used to fit a linear model to the balance of breakpoint distances in the right and left replichores. The distribution of natural log ratios (balance) was used to prune extremely unbalanced inversions (outliers). The remaining 35 breakpoint distances were used to fit a linear model using the stats R package (version 3.4.4). The resulting model was used to predict all possible inversions between IS elements located within the corresponding 95% prediction intervals in a representative genome from each species. Observed and predicted inversions were visualized using the circlize R package (version 0.4.4) (103).

Source code for custom scripts developed under the present study is available at https://github.com/mikeyweigand/Bordetella_species. The whole-genome shotgun sequences have been deposited at DDBJ/EMBL/GenBank under the accession numbers listed in Table S1. The versions described in this paper are the first versions. Raw sequence data are available from the NCBI Sequence Read Archive, organized under BioProject accession numbers PRJNA279196 and PRJNA287884.