Table 1 lists the 12 H. influenzae clinical strains and the reference strain Rd, a largely non-pathogenic strain, used in the comparative genomic studies described herein, their NCBI locus tags, the location where the sequencing was performed, and their clinical origins. Nine of the clinical strains were sequenced using 454 LifeSciences novel pyrosequencing technology 25. The number of sequencing runs, the extent of genomic coverage, and the number of contigs resulting from first and in some cases second pass assemblies are tabulated (Table 2).

Gene clustering parameters for the grouping of homologs were empirically determined by minimizing the change in the number of clusters per change in the parameters (Figure 1). We hypothesize that this minimum point coincides with the best estimate threshold for distinguishing true orthologs from functionally distinct homologs. Some homologs will be more similar than 70%, while some orthologs will be more divergent than 70%, but as a uniform criterion, the threshold is optimized. Visual inspection of the clusters reveals that most clusters are reasonable. Mosaic genes were particularly difficult to cluster due to high levels of rearrangement. In the remainder of the paper, genes in the same cluster are considered to be the same gene.

We identified 2,786 gene clusters among the 13 strains (Table 3). Of these, 52% were found in every strain (core genes) and 19% were found in only a single strain (unique genes). The remaining 29% of genes were found in some combination of two or more strains, but not all (distributed genes; Figure 2). The number of clusters found per strain varied from 1,686 in PittEE to 1,878 in PittII (Table 4). All strains possessed some unique genes not seen in any of the other strains. A pair-wise comparison was performed among all possible strain pairs, which determined the mean number of genic differences between any two strains was 395 with a standard deviation of 94 (Figure 3). This analysis also identified minimal and maximal genic differences of 81 and 577, respectively, for the strain pairs 2866:PittII and 2866:PittAA. The number of coding sequences identified per genome by AMIgene did not correlate strongly with genome size. This is likely due to the presence of split open reading frames (ORFs) in the 454 sequenced genomes as an analysis of the 4 completed genomes showed a linear relationship between gene number and genome size with an R² = 0.910. In contrast, the correlation between total gene clusters and genome size is 0.86, implying that the number of distinct genes found on the genome is linearly related to the genome size.

A dendrogram based on non-core genic differences (Figure 4a) demonstrates the diversity in the NTHi population. A typical strain differs from its nearest neighbor by more than 200 genes. The strains collected from otitis media with effusion (OME) patients at Children's Hospital in Pittsburgh (designated as Pitt strains) show that a genetically diverse population can be isolated contemporaneously from a single geographic location from patients with similar indications. In contrast, two pairs of strains, PittEE/R2846 and PittII/R2866 are relatively similar despite geographically distinct points of isolation. Interestingly, the laboratory strain Rd KW20 is not an outlier among the clinical strains. For comparison, a maximum likelihood tree was generated using sequence from seven multi-locus sequence typing (MLST) housekeeping genes for the same set of 13 strains (Figure 4b). The topology of the trees is significantly different, both in terms of pairwise groupings and overall structure.

The identified number of new genes and core genes found per addition of each genome (as determined by incremental clustering of the 13 strains) shows an exponentially decaying trend in both cases (Figures 5 and 6). Qualitative inspection suggests a diminishing return on new genes found in future sequences, though it is expected that approximately 40 new gene clusters will be found in each of the next few genomes that are sequenced. The number of core genes appears to trend towards a horizontal asymptote near 1,450 genes. A quantitative analysis of these results is developed below in the section 'Mathematical development of a finite supragenome model'.

Whole genome alignments were generated between Rd and each of the 12 clinical strains to quantify genomic insertions and deletions independently of gene identification (Table 5). On average, each of the clinical strains had 127 genomic insertions (>90 base-pairs (bp) in length) that did not correspond to any Rd KW20 sequence. Similarly, each clinical strain contained, on average, 147 genomic deletions (>90 bp) when compared to the Rd KW20 strain. The average total length of non-matching sequences between the 12 clinical strains and Rd was 321 kb, approximately 18% of the genome. The quantity of non-matching sequences reasonably accounts for the average of 390 genic differences between strain pairs. Figure 7 shows a genomic region in which two different forms of an insert, homologous to the plasmid ICEhin, have integrated into the same site of two different genomes, but which is wholly absent from the other strains in the alignment. Similarly, a 40 kb contiguous region in Rd shows extensive deletional diversity among seven of the clinical strains, with only two of the clinical strains demonstrating the same local genomic organization (Figure 8). Interestingly, the two strains, PittAA and PittEE, that are similar in this region are highly divergent overall (Figure 3). Genic diversity also exists on a smaller scale. Figure 9 displays a 20 kb region from 7 clinical strains that shows 5 different combinations of possession and loss of the lic2C gene, the NTHI0683 gene, and the UreABCEFGH operon.

Global genomic alignments of PittEE against R2846 and R2866 were performed (Figures 10 and 11). PittEE and R2846 are very similar at the global level and this is reinforced by the gene cluster analysis, which revealed only 96 genic differences. In contrast, R2866 has a large inversion and several large insertions and deletions with respect to PittEE. This diversity at the global level corresponds to the 377 genic differences identified between these two strains by cluster analysis (Figure 3). Global alignments were not visualized for most strains since the ordering of the contigs had not been determined.

The codon usage of each gene cluster was compared to the typical H. influenzae codon usage pattern by the epsilon-score calculated by CodeSquare 26. A low epsilon score indicates that a gene's codon usage is similar to typical patterns of the organism, while a high score indicates atypical codon usage. Since the epsilon score is partially dependent on the length of a coding sequence, all scores were normalized by length. The average normalized score is 0 and low values continue to indicate typical codon usage. Figure 12 is a scatter plot of the normalized epsilon scores versus the number of strains in which the gene was found. The range of normalized epsilon values is similar for core, distributed, and unique genes, though the median values are slightly higher for distributed and unique genes (Tables 6 and 7). The Mann Whitney U-test was employed to determine the significance of this difference. To eliminate any remaining length bias, only genes with lengths of 200-300 amino acids were analyzed. The median normalized-epsilon value of core genes is significantly smaller than the medians of distributed and unique genes, and as a consequence, these non-core genes are more likely to have foreign origins. Interestingly, there is no significant difference between distributed and unique genes and most of these non-core genes display typical H. influenzae codon usage.

Phage insertion is a common origin of genomic diversity. The influence of phage was quantified by a homology search between all gene clusters and the NCBI NT database. A gene cluster was said to be 'phage associated' if one of the top ten significant matches was annotated as a sequence of phage origin. Overall, 9.3% of gene clusters were phage associated. The distribution of these genes is not uniform among core and non-core genes. Only 0.3% of core genes were phage associated, while 14.6% and 25.8% of distributed and unique genes, respectively, were phage associated (Table 8).

The comparative genomic data presented above are supportive of the DGH and reinforces the concept that, at the species level, there is an H. influenzae supragenome that is much larger than the genome of any single individual strain, and hence many strains must be sequenced to generate an accurate picture of the species supragenome. Among the questions we may ask about the supragenome, the most obvious is, how many strains must be sequenced to observe the entire (or nearly all) of the supragenome?. The problem is similar to determining the read coverage necessary to sequence an entire individual genome using a random shotgun library approach. Lander-Waterman statistics provide an answer in the latter case by using the assumption that reads are independently and randomly sampled from the genome with equal probability. Previously, Tettelin et al. 27 developed a supragenome model for S. agalactiae that, like Lander-Waterman statistics, is based on the assumption that contingency genes are independently sampled from the supragenome with equal probability, except in the case of rare genes, which are modeled as unique events that appear only once in the entire global population. The model requires four parameters: the number of core genes, the number of contingency genes, the probability of finding a contingency gene, and the expected number of 'unique' genes found per strain. This model predicted that the supragenome of S. agalactiae is infinite in size (that is, the expected number of unique genes found in each strain is non-zero). While the model is an insightful attack on the problem, we question the assumption that contingency genes are sampled in the population with equal probability. It is important to compare the existing model against a new model that does not rely on this assumption.

The Supragenome is represented here by a generative model that emits genomes according to a set of probabilistic rules. The supragenome contains N genes that are modeled as Bernoulli random variables with 'success' probabilities that correspond to the population frequency of each gene. A genome is generated by observing the Bernoulli variables: a gene is present if the corresponding trial is a success and otherwise absent. Each gene variable is assumed to be independent of all other genes. This assumption is sometimes violated in real H. influenzae genomes. For example, genomic islands are sets of genes that are not independent. However, we proceed with this assumption since it significantly reduces the complexity of the model and is reasonable in many cases.

The true population frequencies are, in general, unknown. Therefore, population frequencies are also treated in a probabilistic fashion. It is assumed that there are K discrete classes of genes. Each class k has an associated population frequency, μ_k. All genes in class k will have population frequency μ_k. Each of the N genes is assigned to a class according to a probability distribution given by the vector π, where π_k is the probability that a gene is assigned to class k. Conceptually, π_k is the percentage of genes in the supragenome that have population frequency μ_k. The assignment of a gene to a class is independent of all other gene assignments.

The complete model is depicted in plate notation in Figure 13. 'Z' is the hidden class variable in which z_n corresponds to the class of gene n. 'X' is the observed gene variable, where x_n,s corresponds to the presence or absence of gene n in strain s. The outer plate represents the supragenome, while the inner plate represents instances of specific genomes. The model requires 2 × K + 2 parameters: N, K, a mixture coefficient π_k for each class, and a Bernoulli probability μ_k for each class. The number of gene classes, K, and their associated Bernoulli probabilities, μ_k, are fixed in advance. Care must be taken to choose classes that represent low and high population frequencies. Seven classes were selected for this study (K = 7) with associated probabilities μ = <0.01, 0.1, 0.3, 0.5, 0.7, 0.9, 1.0>. The class with probability 1.00 represents 'core' genes that appear in all strains.

The remaining parameters, N and π_k, are selected under a maximum likelihood scheme. Suppose that |S| genomes have been sequenced and a particular gene from class k was observed in n of the |S| strains. The probability of this observation is given by a binomial probability since this result is the sum of independent Bernoulli variables. As a function of π_k and N, the probability is given by:

P ( x = n | z = k , μ k ) = | S | ! n ! ( | S | − n ) ! μ k n ( 1 − μ k ) | S | − n MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8tuQ8FMI8Gi=hEeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqadeqadaaakeaacaWGqbWaaeWaaeaacaWG4bGaeyypa0JaamOBaiaacYhacaWG6bGaeyypa0Jaam4AaiaacYcacqaH8oqBdaWgaaWcbaGaam4AaaqabaaakiaawIcacaGLPaaacqGH9aqpdaWcaaqaamaaemaabaGaam4uaaGaay5bSlaawIa7aiaacgcaaeaacaWGUbGaaiyiamaabmaabaWaaqWaaeaacaWGtbaacaGLhWUaayjcSdGaeyOeI0IaamOBaaGaayjkaiaawMcaaiaacgcaaaGaeqiVd02aa0baaSqaaiaadUgaaeaacaWGUbaaaOWaaeWaaeaacaaIXaGaeyOeI0IaeqiVd02aaSbaaSqaaiaadUgaaeqaaaGccaGLOaGaayzkaaWaaWbaaSqabeaadaabdaqaaiaadofaaiaawEa7caGLiWoacqGHsislcaWGUbaaaaaa@5F3B@

However, we do not know the true gene class, so we must consider a mixture of binomial probabilities:

P ( x = n | π → , μ → ) = ∑ k = 1 K P ( x = n | z = k , μ k ) ⋅ P ( z = k | π k ) = ∑ k = 1 K π k | S | ! n ! ( | S | − n ) ! μ k n ( 1 − μ k ) | S | − n MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8tuQ8FMI8Gi=hEeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=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@84D9@

Now consider the complete set of genes. Let c = <c₀, c₁, ..., c_S>, where c_n is the number of genes observed that appear in exactly n of |S| strains. The probability of the total observation is given by a multinomial distribution:

P ( c → | N , π → , μ → ) = N ! c 0 ! c 1 ! ⋯ c s ! ∏ n = 0 | S | p ( x = n | π → , μ → ) C n = N ! c 0 ! c 1 ! ⋯ c s ! ∏ n = 0 | S | ( ∑ k = 1 K π k | S | ! n ! ( | S | − n ) ! μ k n ( 1 − μ k ) | S | − n ) C n MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8tuQ8FMI8Gi=hEeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=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@9EF7@

The parameters N and π can be determined by maximizing the log-likelihood of the observation c:

log ⁡ P ( c → | N , π → , μ → ) = log ⁡ N ! − ∑ n = 0 | S | log ⁡ ( c n ! ) + ∑ n = 0 | S | c n log ⁡ ( ∑ k = 1 K π k | S | ! n ! ( | S | − n ) ! μ k n ( 1 − μ k ) | S | − n ) MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8tuQ8FMI8Gi=hEeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=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@8D94@

The log-likelihood function was maximized by fixing N and maximizing with respect to π. The maximization was performed using the MATLAB function fmincon with the constraint:

∑ k = 1 K π k = 1 MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8tuQ8FMI8Gi=hEeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqadeqadaaakeaadaaeWbqaaiabec8aWnaaBaaaleaacaWGRbaabeaaaeaacaWGRbGaeyypa0JaaGymaaqaaiaadUeaa0GaeyyeIuoakiabg2da9iaaigdaaaa@3D89@

and requiring that the coefficients are between 0 and 1. The maximization was performed for values of N starting at the minimum possible value (the number of genes actually observed) to 6,000. The combination of N and π that maximized the overall log-likelihood was selected as the best parameter estimate.

The model was validated by training the supragenome parameters using only the first 8 sequenced genomes and comparing the predictions with the observed results for 13 strains. The maximum likelihood number of genes was 3,078. Of these genes, 1,423 are core genes, 417 are contingency genes with population frequency >0.1, and 1,238 are contingency genes with 0.1 population frequency. No genes were predicted in the 0.01 population frequency class. Predictions for the 0.01 class may be inaccurate due to the small sample of 8 genomes. The 1/100 maximum likelihood confidence interval for total genes ranged from 2,975 to 3,681. Figure 14 shows the distribution of the genes among the seven classes.

Figure 5 compares model predictions based on 8 strains to actual observations of core genes (shared among the first N strains) and total genes found after sequencing the 9th through 13th strains. In both cases the model predictions follow the observed trends. Figure 6 compares predictions to observations of the number of new genes found in the Nth sequenced strain. Again the model predictions follow the observed trend. Figure 2 compares the best-fit gene distribution (based on 8 strain models) to the observed distribution of genes found in exactly N of 13 strains. Overall, the predicted trends follow the observed distribution; however, the predictions were too low for genes seen in 1 of 13 strains, and too high for genes seen in 2 of 13 strains. This bias may be due to the small sample size (eight strains) used to train the supragenome model. Predictions for genes seen in four to seven strains were also somewhat lower than observed.

The supragenome model predicted an average of 1,776 genes per strain with a standard deviation of 14 genes. Of the 13 strains, the average number of genes was 1,793 with a standard deviation of 62 genes. The model predicted an average of 373 different genes when comparing any two strains with a standard deviation of 17 genes. Among the 13 sequenced strains, the average was 395 with a standard deviation of 91 genes. In both cases the model predication for average was reasonable, while the standard deviation was underestimated by about four-fold. This suggests that the assumptions used for the supragenome model may omit important sources of variation. Genomic islands and other genes that appear together in the genome likely contribute to the total variance.

Altogether, the above results show that the supragenome model generates reasonable predictions for the average properties of the supragenome. To obtain improved predictions, the model was re-trained on all 13 strains. The supragenome class distribution for the extended model is shown in Figure 14. The results are similar to the model trained on 8 strains, except that the class with population frequency 0.01 is now predicted to contain 2,609 genes, while the 0.10 frequency class was reduced in size to 590 genes. This large change is due to improved resolution of rare genes in the 13 strain training set. The model now predicts 5,230 genes, with a 1/100 likelihood interval ranging from 4,425 to 6,052 (Table 9). Nearly all of the increase over the eight strain model is due to the class of rarest genes. Of these genes, 1,437 are core genes, 594 are contingency genes with population frequency >0.1, and 3,199 genes are rare contingency genes with population frequency <0.1. Figures 15 and 16 show the prediction trends for total, core, and new genes observed after sequencing N strains (up to 30 strains).

Complete or nearly complete genomic sequences of 11 unique clinical strains of H. influenzae were generated and used in comparative genomic analyses with the two published NTHi genomes 3233 in the development of a supragenome model. Genomic sequence of nine clinically isolated NTHi strains was generated at The Center for Genomic Sciences by the 454 Life Sciences GS-20 sequencer using standard protocols 25. Strains were sequenced to a depth of 16×, or greater, and assembled de novo by the 454 Newbler assembler to 81 contigs, on average. Lander-Waterman statistics predict that greater than 99.9% of each genome was sequenced. Regions of duplicated sequence caused most of the assembly gaps. Informal comparison between high-quality Sanger reads and 454 data suggest an error rate of less than 1 in 1,000 bases. Most base call errors are single base insertions or deletions in homonucleotide repeats that can result in frame-shift artifacts. The other two clinical NTHi isolates (R2846 and R2866) included in the comparison were sequenced at the University of Washington Genome Center (Alice Erwin, personal communication). The complete genomic sequences of H. influenzae strain Rd KW20 and 86-028NP and the incomplete sequences of strains R2846 and R2866 were accessed through the Microbial Genomes Database of NCBI.

The most recent versions of the genome assemblies were deposited with GenBank, with the following accession numbers for the indicated strains: CP000671 (CGSHiEE); CP000672 (CGSHiGG); AAZD00000000 (CGSHi22121); AAZJ00000000 (CGSHi22421); AAZF00000000 (CGSHi3655); AAZG00000000 (CGSHiAA); AAZH00000000 (CGSHiHH); AAZI00000000 (CGSHiII); and AAZE00000000 (CGSHiR3021).

The 454-assembled PittEE strain genomic contigs were scaffolded against all four of the completed H. influenzae genomes using Nucmer 34, which indicated the greatest similarity to strain 86-028NP. Using a maximum parsimony approach, the PittEE genome was reduced to 12 contigs by a combination of: sequencing PCR amplicons targeted to fill gaps between neighboring contigs, as inferred by the scaffolding; and sequencing a 4 kb clone library and searching for clones that spanned gaps in the 454 sequence. Gap closure experiments were designed by a custom Perl script, and PCR primers were designed by Primer3 35. Similarly, PittAA was reduced to 47 contigs by sequencing of PCR amplicons generated following scaffolding. Clones and PCR amplicons were assembled along with 454 contigs by a modified Phred-Phrap-Consed pipeline where 454 contigs were converted to PHD format files and input to Phrap as long reads 36373839.

Coding sequences for all 13 strains, including those previously annotated, were identified by the AMIgene microbial gene finder adjusted to low-GC parameters and trained on the Rd KW20 genome 40. AMIgene builds three Markov models to identify coding sequences with different codon usage statistics. This provides increased sensitivity for genes of possible foreign origin. Prior to gene calling, all contigs were artificially stitched together using a linker (NNNNNCATTCCATTCATTAATTAATTAATGAATGAATGNNNNN) that provided start and stop codons in all six reading frames, permitting the identification of genes that extend past the ends of a contig 27.

Each pair of genes was examined for protein homology by alignment of six-frame nucleotide translations to predicted protein sequences. Alignments were generated by tfasty34, part of the Fasta v3.4 package 41. Six-frame translations were employed to minimize the impact of frame-shift artifacts. Each gene was also aligned against the full nucleotide sequence of the 13 genomes by fasta34 (also part of the Fasta package): Fasta34 parameters, fasta34 -H -E 1 -m 9 -n -Q -d 0; Tfasty34 parameters, fasty34 -H -E 1 -m 9 -p -Q -d 0. Genes were clustered based on homology using a single-linkage algorithm. A link was defined by a significant tfasty match between genes that exceeded an identity threshold of 70% and covered at least 70% of the shorter gene (a detailed discussion of parameter selection is found in the supplementary materials at 42). The asymmetric length criterion was chosen to insure that fragmented genes would cluster with the full length version of the gene. A side-effect of this criterion is that multi-domain proteins may fuse with proteins that are composed of a subset of those domains. Significant fasta matches between genes and genomic sequence were used to identify sequence conservation between a gene cluster and a strain. In the event of a significant match (70% identity/70% length), the matching genome was considered to possess the gene cluster for purposes of quantifying the number of strains that contain the gene cluster. See supplementary materials for a comparison of our clustering methods and the COG method 42.

Multi-alignments were generated for each cluster using poa (partial order alignment) in order to visually and computationally verify the integrity of the clusters 43. If the multi-alignment of a cluster was less than 120 bp in length, the cluster was filtered as a likely false-positive gene. Finally, an attempt was made to split false clusters formed by multi-domain proteins by searching for point of partition in the multi-alignment that divided the majority of genes into two non-overlapping sets. The algorithm was implemented using a custom Perl script.

Two types of dendrograms were generated and compared. A gene possession-based phylogenetic tree of the 13 NTHi strains was constructed by defining the distance between a pair of genomes i and k to be:

∑ n | g n , i − g n , k | MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfeBSjuyZL2yd9gzLbvyNv2Caerbhv2BYDwAHbqedmvETj2BSbqee0evGueE0jxyaibaiKI8=vI8tuQ8FMI8Gi=hEeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciGacaGaaeqabaqadeqadaaakeaadaaeqbqaamaaeiaabaWaaqqaaeaacaWGNbWaaSbaaSqaaiaad6gacaGGSaGaamyAaaqabaGccqGHsislcaWGNbWaaSbaaSqaaiaad6gacaGGSaGaam4AaaqabaaakiaawEa7aaGaayjcSdaaleaacaWGUbaabeqdcqGHris5aaaa@41C2@

where g_n,i = 1 if gene n is present in strain i and 0 otherwise. The strains were clustered based on the distance metric by the unweighted group average method implemented in the Phylip package 444546. A tree was also generated using sequence alignments of seven housekeeping genes used in multi-locus sequence typing 47. The tree was constructed using the maximum likelihood method implemented in fastDNAml as part of the Phylip package 4849.

Whole genome alignments were generated by Nucmer and visualized by Mummerplot 34. MUMmer parameters were set to -maxmatch -l 16 -o. The order of PittEE contigs was inferred from optical restriction fragment maps generated by Opgen (Madison, WI, USA) 50. Whole genome alignments were not built for most strains since the ordering of the contigs was not determined.

Inserted and deleted genomic sequence, in comparison to the Rd KW20 genome, was identified by maximal sequence matching performed by Nucmer 34 with the settings -maxmatch -l 16 -o. Non-matching sequence was identified and quantified by a custom Perl script.

Multistrain local sequence alignments against reference sequences (86-028NP or Rd KW20) were generated using BLASTn 51 by querying the reference sequence against a database containing the genomic sequence of all 13 strains. Alignments were then visualized using BioPerl scripts. By the nature of this alignment procedure, sequence that is present only in non-reference strains is not visualized. Gene annotations for reference strains were obtained from GenBank.

Phage derived gene clusters were identified by selecting a representative sequence from each gene cluster to use as a BLASTx query against the NCBI NR (non-redundant) protein database. GenBank records of the top ten significant protein matches with e-value >1e-8 were queried for the keyword 'phage'. If the keyword was identified among the matches, the gene cluster was flagged as 'phage derived'.

The codon usage of a representative sequence from each cluster was analyzed by CodeSquare using Rd KW20 mean codon usage as a reference 26. The epsilon statistic reported by CodeSquare was normalized for ORF length dependence using a best-fit power function for the mean and variance (as a function of length). Gene clusters were divided into three categories: core (gene found in all 13 strains), contingency (2-12 strains), and unique (1 strain). To minimize length bias, codon usage analyses were limited to genes with lengths between 200 and 300 amino acids. Significant differences in the median epsilon statistic were calculated using the non-parametric Mann-Whitney U test.