Article | Open | Published:

Phenotypic differentiation of Streptococcus pyogenes populations is induced by recombination-driven gene-specific sweeps

Scientific Reports volume 6, Article number: 36644 (2016) | Download Citation


Genomic recombination plays an important role in driving adaptive evolution and population differentiation in bacteria. However, controversy exists as to the effects of recombination on population diversity and differentiation, i.e., recombination is frequent enough to sweep through the population at selected gene loci (gene-specific sweeps), or the recombination rate is low without interfering genome-wide selective sweeps. Observations supporting either view are sparse. Pathogenic bacteria causing infectious diseases are promising candidates to provide observations of recombination. However, phenotype-associated differentiations are usually vague among them due to diverse disease manifestations. Here we report a population genomic study of the group A Streptococcus pyogenes (GAS), a human pathogen with highly recombining genomes. By employing a genome-wide association study on single nucleotide polymorphisms (SNPs), we demonstrate a phenotypic differentiation of GAS, represented by separate clustering of two sublineages associated with niche-specific infections, i.e., skin infection and pharyngitis-induced acute rheumatic fever. By quantifying SNPs associated with the differentiation in a statistical and phylogenetic context, we propose that the phenotype-associated differentiation arose through recombination-driven gene-specific sweeps, rather than genome-wide sweeps. Our work provides a novel paradigm of phenotype-associated differentiation induced by gene-specific sweeps in a human pathogen and has implications for understanding of driving forces of bacterial evolution.


A systematic understanding and characterization of the genotype-phenotype correlations in infectious diseases remains a fundamental goal in physiological studies and clinical practices. The correlations are often complicated by multiple host-pathogen interactions, mainly involving versatile adaptive strategies of pathogens through gaining genetic variations to achieve rapid adaptation and prolonged persistence in distinct host environments. The genetic variations include, but are not limited to, point mutations, recombinations, and copy number variations, which in combination result in a spectrum of disease phenotypes for a given pathogen species, leading to clinical complications and difficult diagnostic interpretations.

Over the past two decades, genomic recombination has been found to have a major impact on bacterial evolution, including phenotypic differentiation, ecological cohesion and microorganism speciation1,2. Multiple models have been proposed to explain the observed patterns of genetic diversity in light of mutation, recombination, and natural selection3,4,5,6,7,8. When recombination events introduce advantageous changes in alleles or genes favored by positive selection, the changes may sweep through the population resulting in reduced sequence diversity within the affected population5. The recombination events may also allow neutral or deleterious alleles in proximity to the favorable alleles to “hitchhike” at a high frequency. The sweeps may occur at selected gene loci without affecting the sequence diversity in other regions of the genome, described as gene-specific sweeps5,6,9. When recombination events are frequent across the genome, different loci will support distinct phylogenetic trees, with little support for a single clonal genealogy. Alternatively, in highly clonal bacteria with relatively low rates of recombination, diversity is purged at all loci, described as genome-wide sweeps10. Ultimately, the sweeping process at the gene-specific level or the genome-wide scale will imprint specific patterns in the genome, which can be detected by population genomic studies.

The availability of whole-genome information has made it possible to investigate the genetic outcomes of recombinations and the possible evolutionary mechanisms at the level of population genomics based on genome-wide variations. Numerous population genomic studies on genome-wide single nucleotide polymorphisms (SNPs) have demonstrated the profound impact of genomic recombination on driving adaptation and accelerating evolution, such as resistance to antibiotics in Streptococcus pneumoniae11, association with epidemic outbreaks in Legionella pneumophila12, and serotype switching in Chlamydia trachomatis13. Remarkably, genomic recombination has also been shown to induce ecological differentiation via gene-specific selective sweeps in aquatic bacteria, such as Vibrio cyclitrophicus14, Prochlorococcus15, and Synechococcus16, and in a soil bacteria population Mesorhizobium17.

Group A Streptococcus pyogenes (GAS) is an obligatory human pathogenic bacterium with complicated colonization environments, ranging from skin surfaces, throat, pharynx, to various deeper internal organs of human body. Clinical manifestations of GAS infection range from benign clinical conditions, e.g., pharyngitis and impetigo, to more lethal outcomes that include necrotizing fasciitis (NF), streptococcal toxic shock syndrome (STSS), and acute rheumatic fever (ARF)18. Various clinical strains have been isolated from individual patients or epidemic outbreaks of GAS infection. Many of the causative isolates reemerged with improved fitness or enhanced virulence by acquiring novel genes or large segments of foreign DNA contents via genomic recombination19,20,21,22,23,24. GAS has shown remarkable capabilities in rapidly adapting to diverse human body niches and gaining tropism for specific niche environments. The wide range of infection-niche tropism and versatile adaptive capability of GAS have long been recognized to be largely associated with the highly plastic and transformable genome of this organism, as evidenced by frequent genomic recombination events observed in GAS genomes25,26. However, neither evolutionary mechanisms of the adaptations, nor conclusive associations between the observed infection phenotypes and genotypic properties have been well defined.

Several kinds of genetic markers have been employed to delineate those associations, but have displayed limitations. The emm gene locus encoding the surface M protein has been commonly used to serotype GAS strains. Tissue tropism has been based on the peptidoglycan sequences, which define the gene subfamily patterns of the emm-like gene locus in the chromosome27. However, this method has not been sufficient to account for many of the observed GAS disease phenotypes, including the complex phenotypes of M1 strains28, and the multi-tissue tropism exhibited by M3 and M59 strains24,29. Although the GAS genomes have been shown to be highly recombining25, population genomic studies have heretofore not been rigorously applied to GAS in such a manner as to investigate the role of recombination on shaping its mosaic genome architecture and driving its rapid evolution.

In the current study, we conducted a comprehensive population genomic study of natural GAS isolates to elucidate the association of phylogenetic divergence with the infection phenotypes, and to identify possible evolutionary mechanisms underlying the association. We demonstrate a phenotypic differentiation of GAS strains associated with their clinical manifestations, i.e., skin infection and acute rheumatic fever (or rheumatogenicity). Using a genome-wide population genomic approach, we further propose that the phenotypic differentiation is induced by gene-specific sweeps driven by genomic recombination. For the first time, we establish an association between genotypes and phenotypes of GAS in causing skin diseases and acute rheumatic fever, and identify the molecular basis of the association in a model of recombination-driven gene-specific sweeps.


SNP ascertainment in the core genome of GAS strains

We performed the comparative genomic analysis of the core genomes of 44 distinct GAS strains with known genomic information. As a result, we identified 71,558 SNPs, corresponding to an average of 43 SNPs/kb. This shows a high level of genomic variation compared with the clonal pathogens, such as Bacillus anthracis30 (0.23 SNPs/kb) and Mycobacterium tuberculosis31 (6 SNPs/kb). The multi-allelic SNPs were excluded for further study, thus leaving us 68,892 (96.3%) bi-allelic SNPs. Among them, 58,057 (84%) fall into coding regions, and 20,701 (30%) are non-synonymous substitutions causing amino acid changes. The calculation of Watterson’s θ32 provided a genome-wide average mutation rate of 0.011. This is greater than the average pair-wise nucleotide difference of 0.009, and results in an negative Tajima’s D value33 of −0.831. This suggests a departure from the neutral evolution expectation of the population of GAS34.

Phylogenetic structure inference reveals population differentiation associated with infection phenotypes of GAS

In order to investigate the evolutionary history of GAS genomes, we applied the maximum likelihood method to infer the genome-wide phylogeny based on the detected SNPs. We observed two distinct highly-supported sublineages, with clustered isolates associated with different disease phenotypes depending on the infection niches, i.e., skin infection (SI) and acute rheumatic fever (ARF; or rheumatogenicity), respectively (shaded in blue and green in Fig. 1). ARF is a post-infection sequela typically developed following throat or pharyngeal infection, mainly affecting connective tissues, such as muscles and joints. The clustering of the strains is further evidenced by the lower within-cluster nucleotide diversity (0.007) than the population-wide average (0.009) (Fig. 1). This demonstrates a phenotype-associated differentiation of GAS revealed by divergent phylogenetic clusters.

Figure 1: The phylogeny of the core genome of GAS and the heatmap of pair-wise nucleotide differences of the 44 known genomes.
Figure 1

The phylogenetic tree was inferred using Maximum Likelihood method with 1000 bootstraps. A grey scale for the nucleotide differences is given on the right side of the heatmap. The population differentiation is clearly observed with high support values, evidenced by the lower nucleotide differences within the clustered strains (~0.007) than the population-wide average (~0.009). The branches of the clustered strains are shaded in color: light blue for strains associated with SI, and light green for those associated with ARF (or rheumatogenicity). The strains from serotype M12 and M28, or M1 are also clustered (in grey shading). However, no further analysis was performed on them due to the limited number of serotypes in those clusters insufficient for statistical tests.

GAS has been known to undergo frequent recombination, which interrupts the clonal phylogenetic signals in the genome25. Therefore, we examined its haplotype structure by examining the linkage disequilibrium (LD) of the SNP loci using Haploview35. Only a limited number of short regions are observed with evidence of strong LD, in comparison with the long-range continuous blocks of LD in the highly clonal species B. anthracis30 and M. tuberculosis36 (Supplementary Fig. S1). The highly recombining property of the GAS genomes makes it feasible to perform a genome-wide association study (GWAS) to identify variants associated with niche-specific infections, i.e., SI and ARF. In total, we found 895 and 1,638 SNPs associated with SI and ARF, respectively (p-value ≤ 10−3). Those SNPs cause 221 and 390 amino acid changes in the two sublineages, respectively. The correction for population stratification was performed by excluding the strains with redundant M serotypes from the test due to the close phylogenetic relationship between them. The Q-Q plots show relatively weak inflation in the association test (the inflation factor λ = 1.07~1.34) (Supplementary Fig. S2). Further analysis with the Cochran-Mantel-Haenszel (CMH) test does not improve the inflation factor, implying that the influence of population stratification on the association test has been well controlled by excluding redundant M serotype strains from the test. However, we note that the permutation testing for multiple corrections only leaves no more than 50 candidates with a corrected p-value ≤ 0.05, probably due to the small sample size and unequal proportion of associated and non-associated samples. Therefore, for further study, we chose SNPs with significant non-corrected p-values ≤ 10−3 and high odds ratios ≥ 28 by fitting the overall distribution of the p-value and odds ratio (Supplementary Fig. S3). Due to the genomic inflation, some of these SNPs are likely false positives. Therefore, we did not focus on any particular SNPs in detail, but rather focus on the whole set of associated SNPs that are also enriched in true positive associations.

Phenotype associated SNPs predominantly arise from gene-specific sweeps

We propose that the phenotype-associated differentiation of the GAS population is induced by gene-specific sweeps, based on four lines of evidence. First, the non-random spatial distribution of the associated SNPs forms concentrated clusters at several genes, gene operons, or neighboring genes, based on a sliding window analysis (Fig. 2A,B). The clustering of the SNPs indicates the dependence among neighboring SNPs, which could possibly be brought about by recombination. To test the dependence between SNPs, we compared the distribution of the observed inter-SNP distances against those that are expected assuming a random independent occurrence of SNPs across the genome. The fitting curve for all genome-wide SNPs is highly consistent with the expected exponential distribution (Fig. 2C). In contrast, the fitted distributions for SNPs associated with SI or ARF deviate from those expected, manifested by an excess of closely located phenotype-associated SNPs (Fig. 2D,E). This reveals a pattern predicted by gene-specific sweeps, with frequent recombination and positive selection on the phenotype-associated SNPs.

Figure 2: Highly concentrated distributions of the SNPs associated with niche-specific infection.
Figure 2

(A,B) Spatial density of SNPs associated with SI and ARF, respectively. The gene loci harboring the most significant clustering of SNPs are indicated and represented as high peaks in the density. (CE) Distribution of inter-SNP distances for genome-wide SNPs, SNPs associated with SI, and those associated with ARF, respectively. The observed distribution was fitted and compared with the expected exponential distribution, assuming a random independent occurrence of the SNPs. The fitting curve for the genome-wide SNPs is consistent with the expected distribution. However, the fitting distribution for SNPs associated with SI or ARF deviates from the expected distribution, revealing excess of closely located SNPs. The inset in (D,E) shows the zoomed curve of the expected distribution.

The second line of evidence is based on the significance of the clustering and the composition of the clusters. Under a model of gene-specific sweeps, non-synonymous SNPs under selection would be surrounded by clusters of non-selected SNPs (i.e., synonymous SNPs and inter-genic SNPs), which may hitchhike in the same recombined tract of DNA. In order to assess the significance of the SNP clustering, we first quantified the associated SNPs into spatially discrete clusters (≥2 SNPs per cluster) by simulating the process of gene-specific sweeps. To accomplish this goal, we employed an anchor-extension strategy based on a non-synonymous SNP-centric concept (Fig. 3A): anchoring an initial cluster at non-synonymous SNPs (assuming that the non-synonymous SNPs are more likely to be positively selected) and extending the initial cluster on both sides within a range of the estimated recombination tract length, L. The significance of the clustering was subsequently assessed by calculating the probability of the SNPs in each cluster located within the L range under a model of independent and random distribution of SNPs across the genome. Ultimately, the majority of the non-synonymous SNPs (>89%) fall into 86 (SI) and 159 (ARF) well-defined clusters characteristic of gene-specific sweeps. Among the defined clusters, 68 (79%; SI; p-value ≤ 0.01) and 114 (72%; ARF; p-value ≤ 0.05) are significant (Fig. 3B and Supplementary Dataset 1). These clusters are composed of non-synonymous SNPs and predominantly synonymous SNPs in the same genes/gene operons (or neighboring genes), but are also partly inter-genic SNPs of nearby regions. This suggests hitchhiking of synonymous or inter-genic mutations with positively selected non-synonymous mutations. Alternatively, each cluster of synonymous, non-synonymous, and inter-genic mutations could have been introduced at the same time by a single recombination event from a distant relative, which subsequently swept through the sublineages.

Figure 3: Assessment of clustering of SNPs associated with niche-specific infection.
Figure 3

(A) Illustration of the anchor-extension strategy for categorizing the associated SNPs in clusters based on a model of gene-specific sweeps. A non-synonymous SNP is first anchored as an initial cluster and the initial cluster is extended by including the neighboring SNPs within a range of estimated recombination tract length L. If the SNPs fall into a single gene or gene operon of length longer than L, all the associated SNPs in that gene or gene operon were included in the target cluster. The extension of the target cluster is indicated by color coded ovals, with purple for a cluster spanning a distance shorter than L or within a gene operon, and orange for a cluster spanning a distance longer than L or beyond a gene operon. The inclusion of ambiguous SNPs on the boundaries was determined by minimizing the normalized root-mean-square of inter-SNP distances NRMSD. (B) Manhattan plots of the significance of SNP clusters. The significance is evaluated against a model of independent and random distribution of SNPs across the genome (top panel for SI and bottom panel for ARF). The red dotted lines show the threshold for statistical significance (p-value ≤ 0.01 for SI; and p-value ≤ 0.05 for ARF).

Our third line of evidence for gene sweep-induced differentiation is suggested by the phylogenetic incongruence of different genes across the genome. By definition, the phenotype-associated SNPs result in a gene tree supporting the separate grouping of the strains associated with SI and ARF. Under a genome-wide sweep model, the rest of the genome, including genes flanking the associated SNPs, would follow the same clonal genealogy. Under the gene-specific sweep model, different flanking regions would support distinct phylogenies, possibly with SI and ARF grouped separately. To distinguish between the two models, we selected 9 and 10 of the most significant SNP clusters associated with SI and ARF, respectively. We next extracted the neighboring SNP loci and constructed the phylogenetic trees from these loci. All of the reconstructed trees show topological deviation from the core genome phylogeny. This supports monophyletic clustering of the strains associated with SI and ARF (Fig. 4A and Supplementary Fig. S4). Specifically, only three out of the 19 trees support the clustering of strains associated with SI and six of the trees support the clustering of strains associated with ARF, and, remarkably, none of the trees supports the separate clustering of the two groups.

Figure 4: Phylogenetic incongruence of phenotype-associated gene loci and alterations of nucleotide diversity of those gene loci.
Figure 4

(A) Phylogenetic trees reconstructed for SNP loci flanking the associated SNPs at different gene loci. Nineteen of the gene loci with the most significant SNP clustering were selected and the trees for ten of them are shown (see Supplementary Fig. S4 for the full set of trees). All trees exhibit topological deviation from the core genome phylogeny and each tree is topologically distinct. The nodes in the trees are indicated in color: blue for the sublineage associated with SI, green associated with ARF, and magenta for the remaining strains in the population. The support values of the trees were derived by bootstrapping 1000 replicates. Only the topology of the trees is shown by ignoring the branch lengths. (BD) Alterations of genetic diversity of the associated SNPs (top panel in black bars) in comparison with those flanking the associated SNPs (bottom panel in grey bars). Three quantities are shown: nucleotide diversity π between sublineages and within sublineages (B), nucleotide divergence Da between sublineages (C), and population differentiation Fst between sublineages (D). The ratio of inter-sublineage versus within-sublineage nucleotide diversity shows substantial decrease for the flanking non-associated SNPs compared with the associated SNPs. Similar decrease is also observed for nucleotide divergence and population differentiation.

Furthermore, we also observed tight linkage of the phenotype-associated SNPs in clusters, as represented by the localized continuous LD blocks, indicating the simultaneous introduction of the SNPs within clusters via recombination or hitchhiking with gene sweeps (Supplementary Fig. S5). We observe that the tight linkage is interrupted by the nearby SNPs surrounding the clusters, which is manifested by the shorter and fragmented LD blocks (Supplementary Fig. S5). This is again consistent with phenotype-associated SNPs spreading through the populations via gene-specific sweeps, resulting in linkages within the associated SNP clusters but little linkage of these clusters and the rest of the genome.

Under the gene-specific sweep model, the phenotype-associated genes should be highly differentiated between sublineages with different phenotypes, and should contain relatively low diversity within sublineages. Consistent with the gene-specific model, we observed that the associated SNP loci show high differentiation between the sublineages and low diversity within each sublineage (Fig. 4B–D). Specifically, the inter-sublineage nucleotide diversity is on average 4.7-fold higher than the within-sublineage when considering the associated SNP clusters. For those nearby SNPs surrounding the clusters, the difference is decreased to 1.9-fold (Fig. 4B). The inter-sublineage divergence (Da) and differentiation (Fst) also show substantial decreases when considering the nearby SNPs surrounding the clusters by an average of 4.1-fold and 2.2-fold, respectively (Fig. 4C,D). These patterns are consistent with the gene-specific sweeps, which purge the genetic variations at specific gene loci under positive selection, while maintaining the diversity in other regions of the genome.

Recent recombination events in GAS genomes are more prominent within sublineages than between sublineages

In order to quantify the impact of recombination events, we first assessed the intragenic conversions of DNA fragments within each coding gene between pair-wise strains using geneconv37 and the Phi test38. A total of 106 out of the 767 genes included in the analysis (containing at least 10 SNPs per gene) carry significant signals of recombination between at least one pair of strains (Table 1). Two features are observed for the recombination signals among the 106 genes: (i) A shorter mean fragment length is observed for intragenic recombination between sublineage SI and ARF (389 bp) than within each of the sublineages (554 bp and 525 bp, respectively) (Table 1); (ii) Significantly more recent recombination events were detected within the sublineage SI than between the sublineages SI and ARF (p-value = 0.0007; Fisher’s exact test) by comparing the number of recombination fragments within sublineages and between sublineages (Table 1 and Supplementary Fig. S6). The significance test suggests that the recent intragenic recombination is more prominent within sublineages than between sublineages, probably due to the higher sequence divergence between sublineages6,14. The recombination between sublineages probably represents more ancient events that occurred before the separation of the sublineages, and has eroded gradually, resulting in shorter and fewer recombination fragments to be detected. Figure S7 shows the recombination events in two representative genes valS (valyl-tRNA synthetase) and lacE (lactose-specific PTS system IIB/IIC component) with more recombinations within the sublineage SI than between the sublineages SI and ARF. We generally observed that the occurrence of recombination is negatively correlated with sequence divergence between the sublineages.

Table 1: Statistics of intragenic recombination events within and between sublineages.

Relative contributions of recombination and mutation in shaping the population structure

Our findings thus far strongly suggest that phenotypic differentiation of the GAS genomes occurred through acquisition of infection niche-specific SNPs predominantly via genetic recombination. Next, we sought to quantify the relative contributions of recombination and point mutation (r/m) to differentiation in the GAS genomes. We estimated the ratio r/m by comparing the number of non-synonymous associated SNPs acquired through recombination against point mutations. We only used the non-synonymous SNPs for the estimation of r/m considering that the synonymous SNPs or non-coding SNPs are probably more likely to be hitchhiked and less likely to be selected, and therefore may bias the estimation, although we also note that recent studies showed that the sequence variations in non-coding regulatory regions may also result in fitness effects39,40. The non-synonymous point mutation candidates were identified as singleton associated SNPs, which cannot be categorized to any SNP cluster and are unlinked to the neighboring SNP clusters. We obtained 10 and 12 such point mutations associated with SI and ARF, respectively (Supplementary Table S1), corresponding to the ratio r/m of 22~34. This approach represents a conservative estimate, considering that the SNPs outside the clusters may also arise from recombination events, which are nevertheless difficult to unambiguously identify. This empirically derived ratio is comparable to that estimated by ClonalFrame41, which gave a median r/m of 34, from averaging over 15 randomly selected genomic regions (Supplementary Fig. S8 and Table S2). This suggests that more than 90% of the polymorphisms in GAS genomes may have been introduced by recombination and clearly demonstrates the remarkable impact of genetic recombination on driving adaptation and diversification of the genetic repertoire.

Functional implications of the phenotype associated SNPs

We found high representation of the associated SNPs in genes involving functions of basic molecular metabolism and membrane transport (Table 2). These functions reflect the deterministic factors in adaptation for GAS in human niches, i.e., nutrient utilization and molecular metabolism, for optimizing intracellular growth and survival. A similar result was also reported in the analysis of the population variations in a human cavity-causing bacterium, S. mutans, where the genes associated with adaptation to, or benefits from, a human oral niche are mainly involved in carbohydrate metabolism, metal and peptide translocation, oxidative stress response, and others42. Interestingly, we observed that the functional category, resistance to antibiotics and toxic compounds, is more represented in strains associated with ARF (p-value ≤ 0.05; Fisher’s exact test). Such allele changes may reflect the molecular response of the bacteria in combating the specific challenges they encounter in causing ARF. The SNPs associated with SI and ARF are both enriched in the gene locus comprising the recA and mutL-mutS system, which are responsible for genetic recombination and DNA mismatch repairs. Multiple non-synonymous mutations were found in the gene cinA within the locus, which is proposed to be specifically required at some stage of genetic transformation. This further supports the importance of the high rate of genetic recombination in GAS in driving phenotypic differentiation and evolution.

Table 2: Functional classification of SNP clusters associated with SI or ARF.

Phylogenetic structure of the accessory genome differs from the core genome phylogeny

We constructed the phylogeny tree for the accessory GAS genomes by identifying the dispensable genes, which are partially encoded by GAS genomes (Fig. 5). Neither the overall topology of the tree, nor the ecological differentiation associated with niche-specific infection, is maintained as strongly as observed for the core genome (Fig. 1). The alteration of the phylogeny indicates the distinct evolutionary history of the accessory genome from the core, probably shaped by recombination via horizontal gene transfer. It is further evidenced by the substantial sequence mosaicism and divergence in the accessory genes. As much as 89% of the accessory genes are completely absent in at least one of the genomes, and 15% of them are specific for individual strains. Specifically, the strains from the sublineage SI and ARF cluster together, forming a monophyletic branch. Furthermore, three of the strains associated with ARF (or rheumatogenicity), i.e., M18 MGAS8232, M5 Manfredo, and M23ND, are more closely related with the sublineage SI (shaded in purple in Fig. 5). These findings indicate that gene content exchanges in the accessory genome between the two sublineages occurred. Therefore, we performed gapped multiple sequence alignment of the accessory genome and calculated pair-wise nucleotide diversity between the GAS genomes. The calculation shows that there is no significant difference in the sequence diversity within and between the sublineages SI and ARF (p-value > 0.07, t-test), indicative of the absence of complete genetic barriers to horizontal gene transfer between the two sublineages. A prominent example for horizontal gene transfer between the sublineages SI and ARF is the FCT locus (encoding Fibronectin-binding proteins, Collagen-binding proteins and T-antigens), where several strains associated with SI and ARF share the same pattern of gene organization43. The surface proteins encoded by the locus facilitate the adherence of bacteria to host cells, and the host cell surfaces in turn may provide venues for the bacteria to disseminate gene contents between the strains.

Figure 5: The phylogenetic tree of the accessory genome of GAS and the clustering of accessory genes.
Figure 5

The tree was derived based on a hierarchical clustering analysis of the pair-wise homology of accessory genes between the GAS genomes. The tree for the GAS strains is shown on the left and the clustering for the accessory genes on the top. The heatmap indicates the homology of pair-wise flexible genes between GAS genomes. A color key is shown on the top left. The accessory gene tree is topologically different from the core genome phylogeny. Three of the strains associated ARF or rheumatogenicity (M18 MGAS8232, M5 Manfredo, and M23ND) cluster together with the strains associated with SI (shaded in purple).


With the widespread availability of genomic information, one of the fundamental goals has been to delineate the associations between genetic characteristics and clinical phenotypes. The associations are complex and in many cases are vague, especially in infectious diseases caused by bacterial pathogens, where phenotypic and genotypic heterogeneity make such systematic inferences difficult. In the versatile human pathogen, Group A Streptococcus pyogenes (or GAS), the heterogeneity is largely facilitated by the highly recombining genomes of the organism, where stretches of DNA fragments are exchanged or transferred within the bacterial community or across distant species. A population genomic study of GAS genomes based on genome-wide SNPs demonstrated an explicit differentiation in the phylogenetic structure associated with niche-specific infections, i.e., skin diseases and pharyngeal-induced acute rheumatic fever (or rheumatogenicity). Although ecologically-associated population differentiation has been observed in aquatic and soil bacterial organisms14,15,16,17, ours is the first such report demonstrating gene-specific sweeps driving infection-niche tropism in a human pathogen such as GAS, a species with multiple clinical outcomes and diagnostic complications. We observed that the phylogenetic clustering from SNPs in our study is different from that based on the commonly used genetic markers, such as the emm pattern. For example, serotype M3 and M14 strains have long been classified as emm pattern A-C, which is proposed to be more closely related with pharyngeal infection. In contrast, these strains are clustered together with the skin-tropic strains with emm pattern D in the current study. This may explain why the isolates from M3 and M14 also caused skin-associated diseases, such as necrotizing fasciitis or streptococcal toxic shock syndrome. It is possible that the two types of strains have changed their serotypes by undergoing distinct recombination in the emm gene locus, or they have become rapidly adapted to skin niches by acquiring allele changes at specific gene loci.

Further statistical tests of the SNPs associated with the infection phenotypes reveal that those associated variations predominantly arise from gene-specific selective sweeps via recombination. This is supported by: (i) the associated SNPs are not randomly distributed, but clustered at several gene loci, with the fitted distribution significantly deviating from the expected distribution in a model of random independent occurrence of the SNPs. (ii) Most of the associated SNPs fall into statistically significant clusters, comprising of non-synonymous SNPs, and predominantly synonymous SNPs in the same genes/gene operons (or neighboring genes), but also partly inter-genic SNPs of nearby regions, characteristic of hitchhiking of synonymous or inter-genic SNPs with positively selected non-synonymous SNPs. (iii) Incongruence of the phylogenetic trees is observed for the nearby SNPs not associated with infection phenotypes compared with the associated SNPs, suggesting the distinct evolutionary processes in the swept regions and the rest of the genome. (iv) There is substantial difference in the population diversity parameters for the associated SNPs as compared with the non-associated nearby SNPs, i.e., higher ratio of inter-sublineage diversity to within-sublineage diversity (from 4.7 fold to 1.9 fold), signifying the effects of gene-specific sweeps on purging the variability out of the subpopulation at specific gene loci while maintaining the diversity in other regions of the genome. Overall, the genetic characteristics strongly suggest the phenotype-associated differentiation is predominantly induced by recombination-driven gene-specific sweeps. Compared with the low rate of point mutations, the rate of recombination is sufficiently high (22–34:1) such that the gene sweeps occur rapidly for the bacteria to become adapted to diverse host environments. The selected gene loci were homogenized by the sweeps within the associated sublineages, resulting in a cohesively adaptive population while maintaining the population-wide divergence3,4,5. The homogenization of gene sweeps within sublineages has two effects in the GAS genomes: (i) The recent recombination events occur more prominently within sublineages than between sublineages (p-value = 0.0007), due to reduced sequence divergence within sublineages; (ii) There is a shorter recombination length between sublineages than within sublineages. This implies that the recombination events occurring between sublineages are more ancient than those within sublineages and has eroded during the course of its evolution. The recombination events that occurred within sublineages, which have largely shaped the current population structure, may have just recently emerged. It is consistent with the concept that GAS is an obligatory human pathogen, and its adaptation to the human host is relatively recent on the scale of its evolutionary development.

Even with the overwhelming contribution of recombinations versus point mutations to genomic variation (>90%), the true level of recombination in GAS has been very likely underestimated due to the limited sample size, or the undetectable events, involving the sequences from distant taxa outside of the population. Moreover, the use of SNP data for the study of recombination prevents us from detecting gene gain and gene loss events, such as phage acquisition, which has been extensively studied in GAS21,44,45,46,47,48. On the other hand, the inherent nature of hitchhiking in the gene sweeps implies that not all of the allelic changes introduced from gene sweeps will carry benefits for the adaption and survival of specific lineages. Therefore, the actual number of selective variations relevant to niche-specific infection could be different from that identified in the recombined regions mediated by gene-specific sweeps.

The associated SNPs are highly represented in genes involving functions of basic molecular metabolism and membrane transport for both sublineages associated with SI and ARF. This may reflect the common functional aspects in the micro-adaptation process for the bacteria to circumvent diverse challenges in human environments. With regard to the differential representation of the functions between the two sublineages, we observed that the SNPs involved in resistance to antibiotics and toxic compounds are significantly more represented in ARF (p-value ≤ 0.05). Interestingly, we note that the isolates causing ARF emerged shortly after the use of antibiotics for treatment of ARF49 or in a recurrent epidemics of ARF50. Recently, reports of ARF development following macrolide treatment failure in GAS pharyngitis due to macrolide resistance have been recorded, suggesting a possible link between our observation and clinical diagnoses, though further studies are needed in this regard51.

Our study provides the first paradigm of disease phenotype-associated differentiation in a human pathogen causing clinical complications, which to date has lacked robust identification of genotype-phenotype associations. Using a population genomic approach, we reveal a phenotypic differentiation of GAS genomes associated with two niche-specific infections, i.e., skin diseases and acute rheumatic fever (or rheumatogenicity). Based on statistical tests, we propose the molecular basis of the differentiation attributed to recombination-driven gene-specific sweeps. However, it should be noted that our study is based on a limited sample size, and may have lacked power to identify the genotype-phenotype associations in a comprehensive manner. Additional sampling of strains with contrasting phenotypes from multiple geographical regions and time points would likely increase the power to detect the associations.

We observed that the mechanism of the phenotypic differentiation of GAS in the current study parallels that of the ecological differentiation of a marine bacterium Vibrio cyclitrophicus14 and soil bacteria Mesorhizobium17 with respect to the gene-specific sweeps. It is intriguing that the genetic mechanism for ecological differentiation is shared between environmental bacteria and human adapted pathogens, suggesting that the gene-specific sweep may represent a common mechanism of niche adaptation for bacterial organisms.


SNP detection and estimation of genetic parameters

The genomic sequences for the GAS strains were obtained from the NCBI Genbank database. The core genome was created by aligning the fragmented genomes against one of the complete genomes AP53, which was sequenced and investigated by our group previously. The mosaic regions with ambiguity in determining SNPs were excluded, including the repetitive regions, phage elements, and transposons. The polymorphisms were detected using VAAL52. The Watterson’s θ was calculated based on the segregating site estimator as defined by Watterson32. The average pair-wise nucleotide difference was calculated by counting the number of nucleotide differences for each pair of genomes normalized to the core genome length. Tajima’s D test of neutrality was calculated as defined by Tajima33. The calculation of Watterson’s θ and Tajima’s D was performed on 21 strains containing non-redundant M serotypes to avoid biased estimation due to the low sequence diversity between strains with the same serotypes.

Genome-wide association analysis of SNPs associated with phenotypic differentiation

The genome-wide association test was performed using the Chi-squared test. In order to reduce the influence of the population structure on the association test, we excluded the strains with redundant M serotypes from the analysis due to the close relationships between them. The inflation factor, λ, of the association test was calculated as the ratio of the observed median Chi-square and the expected median Chi-square under the hypothesis of no association. Correction for population stratification was performed using the Cochran-Mantel-Haenszel (CMH) test with no improvement on the inflation factor, implying that the influence of population stratification has been well controlled by excluding strains with redundant M serotypes from the association test. The multiple testing corrections using permutation testing only leaves no more than 50 candidates with a corrected p-value ≤ 0.05, and therefore we chose the SNPs based on the overall distribution of non-corrected p-value and odds ratio (Supplementary Fig. S3). The final set of significant SNPs has p-values ≤ 10−3 and odds ratios ≥ 28.

Sliding window analysis of the spatial SNP density

The SNP density was calculated by counting the number of SNPs in a sliding window of 2000 bp. The density was normalized to every 1000 bp. Several regions show significantly higher SNP density than the average.

Fitting the distributions of inter-SNP distances

Assuming that the position of mutational events is a random variable, the distribution of SNPs can be described using a Poisson process model and the distance between SNPs should follow the exponential distribution53:

where λ is the mutation rate. When the recombination rate is non-zero, the distribution of SNP distances can be generalized to the gamma distribution54:

where α determines the shape of the gamma distribution and β is the mutation rate. The observed distribution of distances for the genome-wide SNPs was well fitted to a gamma distribution with α = 1, β = 0.037, which is coincident with the exponential distribution with a mutation rate of λ = 0.037. However, the distance distribution for SNPs associated with SI and ARF are fitted to a variable of the exponential distribution:

where for SI-associated SNPs, and for ARF-associated SNPs, deviating from the expected exponential distribution with , and , for SI and ARF respectively. The parameters of the distribution were determined based on Chi-squared minimization.

Clustering of SNPs associated with phenotypic differentiation and the significance test

The SNPs identified to be associated with ecological differentiation were grouped into spatially discrete clusters in order to examine the presence of inter-dependence of SNPs characteristic of gene-specific sweeps. The grouping was performed with an anchor-extension strategy based on a non-synonymous SNP-centric strategy: a non-synonymous SNP with association was first anchored as an initial cluster and then the initial cluster was extended by scanning the neighboring associated SNPs within a range of 1,781 bp, which is the estimated recombination tract length L. If the SNPs fall into a single gene or gene operon of length longer than L, the scanning of the neighboring SNPs was performed in the whole range of the gene or gene operon. The inclusion of ambiguous SNPs on the boundaries was determined by minimizing the normalized root-mean-square of the inter-SNP distances in the target cluster and manually inspected:

where di is the ith inter-SNP distance, k is the total number of SNPs in the target cluster, and l is the spanning length of the SNPs in the cluster. Each cluster contains at least 2 SNPs. The significance of the SNP clustering was evaluated conditioning on a null hypothesis that the mutations are independently and randomly distributed across the chromosome. Under the null hypothesis, the probability that distinct SNPs clustered within a distance of l can be estimated based on a gamma distribution given an average mutation rate of μ55:

Calculation of genetic diversity and population differentiation

The calculation of nucleotide diversity π (the average number of nucleotide differences per site between two individuals) and nucleotide divergence Da (the average number of net nucleotide differences per site between two sub-populations) was determined according to the definition by Nei and Li56. The analysis of population differentiation Fst was based on an estimator by Hudson et al.57. In order to remove the bias induced from the closely related strains with the same M serotypes, the strains with redundant M serotypes were excluded from the calculation.

Reconstruction of the phylogeny structure

The core genome phylogenetic structure was inferred using Maximum Likelihood method implemented in MEGA58 with 1000 bootstraps based on the concatenated alleles at all SNP loci and coding SNP loci, respectively. The trees from both sets of SNPs show no differences. The accessory genome tree was constructed using hierarchical clustering analysis of pair-wise homology of accessory genes between GAS strains. The accessory genes were identified based on reciprocal BLAST comparison with homology lower than 50% for any compared orthologs. The sequence diversity of the accessory genome was estimated as the number of nucleotides absent/present between each pair of genomes based on multiple gapped alignment of the GAS genomes with the core genome removed using Mauve59. In order to avoid the overwhelming influence of phage elements and transposons on the accessory genome quantification, the phage sequences and transposons were also removed in the construction of phylogenetic trees and sequence diversity calculation. For constructing trees for protein-coding genes or SNP loci flanking the SNP clusters, the non-informative SNPs (those SNPs that only occur in one of the strains) were excluded. The topological conflicting between trees was measured using the path difference metric, which calculates the differences in the lengths of the paths for each pair of tree leaves. The calculation was implemented in the R environment.

Detection of intragenic recombination and identification of recent recombination events

The intragenic recombination in protein-coding genes was detected with geneconv37 and PhiPack38 based on a gene conversion model of recombination. Only the genes containing at least 10 SNP loci were considered, leaving us a total of 767 qualified genes. The multiple sequence alignments for each of the 767 genes among the GAS genomes were constructed and subsequently subject to recombination detection. The recombination fragments were considered to be significant if the Bonferroni-corrected KA p-value ≤ 0.05 in geneconv and the p-value ≤ 0.05 for at least two of the three methods (Phi permutation, Neighbor similarity score and Max χ2) implemented in PhiPack. The potential recent recombination events are defined as those that involve gene fragments that have a tree distance shorter than a specified threshold compared to the core genome phylogeny. The threshold was determined based on the overall distribution of the tree distance assuming that recent recombination should not severely affect the phylogeny structure shaped by selective sweeps. The significance test of the occurrence of recent recombination events was performed within the sublineage SI, but not within ARF due to the smaller number of samples in this sublineage.

Estimation of the relative contribution of recombination to mutation

The empirical estimate of the recombination to mutation ratio r/m was obtained by comparing the non-synonymous SNPs acquired through recombinations relative to point mutations. The SNPs acquired through point mutations are defined as the singleton SNPs associated with SI or ARF, which cannot be categorized to any SNP clusters and have a low level of linkage disequilibrium (D′<0.7, with D′ defined by Lewontin60) with the neighboring SNPs within SNP clusters. A more thorough estimation was also performed using ClonalFrame41 by averaging over 15 randomly selected aligned genomic regions of lengths 2000–9000 bp (Supplementary Fig. S8 and Table S2).

Additional Information

How to cite this article: Bao, Y.-J. et al. Phenotypic differentiation of Streptococcus pyogenes populations is induced by recombination-driven gene-specific sweeps. Sci. Rep. 6, 36644; doi: 10.1038/srep36644 (2016).

Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.


  1. 1.

    & Impact of recombination on bacterial evolution. Trends Microbiol. 18, 315–322 (2010).

  2. 2.

    et al. Re-evaluating prokaryotic species. Nat Rev Microbiol. 3, 733–739 (2005).

  3. 3.

    & Ordering microbial diversity into ecologically and genetically cohesive units. Trends Microbiol. 22, 235–247 (2014).

  4. 4.

    et al. Identifying the fundamental units of bacterial diversity: A paradigm shift to incorporate ecology into bacterial systematics. Proc Natl Acad Sci USA 105, 2504–2509 (2008).

  5. 5.

    Bacterial speciation: genetic sweeps in bacterial species. Curr Biol. 26, R112–R115 (2016).

  6. 6.

    , & Horizontal gene transfer and the evolution of bacterial and archaeal population structure. Trends Genet. 29, 170–175 (2013).

  7. 7.

    & A systematics for discovering the fundamental units of bacterial diversity. Curr Biol. 17, R373–R386 (2007).

  8. 8.

    & On the origin of prokaryotic species. Genome Res. 19, 744–756 (2009).

  9. 9.

    , , & Looking for Darwin’s footprints in the microbial world. Trends Microbiol. 17, 196–204 (2009).

  10. 10.

    et al. Genome-wide selective sweeps and gene-specific sweeps in natural bacterial populations. ISME J. 10, 1589–1601 (2016).

  11. 11.

    et al. Rapid Pneumococcal evolution in response to clinical interventions. Science 331, 430–434 (2011).

  12. 12.

    , , & Recombination drives genome evolution in outbreak-related Legionella pneumophila isolates. Nat Genet. 46, 1205–1211 (2014).

  13. 13.

    et al. Whole-genome analysis of diverse Chlamydia trachomatis strains identifies phylogenetic relationships masked by current clinical typing. Nat Genet. 44, 413–419 (2012).

  14. 14.

    et al. Population genomics of early events in the ecological differentiation of bacteria. Science 336, 48–51 (2012).

  15. 15.

    et al. Single-cell genomics reveals hundreds of coexisting subpopulations in wild Prochlorococcus. Science 344, 416–420 (2014).

  16. 16.

    , , & Fine-scale diversity and extensive recombination in a quasisexual bacterial population occupying a broad niche. Science 348, 1019–1023 (2015).

  17. 17.

    , , , & Association mapping reveals novel serpentine adaptation gene clusters in a population of symbiotic Mesorhizobium. ISME J. doi: 10.1038/ismej.2016.88 (2016).

  18. 18.

    , , & The global burden of group A streptococcal diseases. Lancet Infect Dis. 5, 685–694 (2005).

  19. 19.

    et al. Evolutionary origin and emergence of a highly successful clone of serotype M1 Group A Streptococcus involved multiple horizontal gene transfer events. J Infect Dis. 192, 771–782 (2005).

  20. 20.

    et al. Complete genome of Acute Rheumatic Fever-associated serotype M5 Streptococcus pyogenes strain Manfredo. J Bacteriol. 189, 1473–1477 (2007).

  21. 21.

    et al. Progress toward characterization of the group A Streptococcus metagenome: Complete genome sequence of a macrolide-resistant serotype M6 strain. J Infect Dis. 190, 727–738 (2004).

  22. 22.

    et al. Genome sequence and comparative microarray analysis of serotype M18 group A Streptococcus strains associated with acute rheumatic fever outbreaks. Proc Natl Acad Sci USA 99, 4668–4673 (2002).

  23. 23.

    et al. High incidence of invasive Group A Streptococcus disease caused by strains of uncommon emm types in Thunder Bay, Ontario, Canada. J Clin Microbiol. 54, 83–92 (2015).

  24. 24.

    et al. Genome sequence of a serotype M3 strain of group A Streptococcus: Phage-encoded toxins, the high-virulence phenotype, and clone emergence. Proc Natl Acad Sci USA 99, 10078–10083 (2002).

  25. 25.

    et al. Recombination within natural populations of pathogenic bacteria: Short-term empirical estimates and long-term phylogenetic consequences. Proc Natl Acad Sci USA 98, 182–187 (2001).

  26. 26.

    , , & Influence of recombination and niche separation on the population genetic structure of the pathogen Streptococcus pyogenes. Infect Immun. 70, 1971–1983 (2002).

  27. 27.

    & Tissue tropisms in group A streptococcal infections. Future Microbiol. 5, 623–638 (2010).

  28. 28.

    , , & Molecular insight into invasive group A streptococcal disease. Nat Rev Microbiol. 9, 724–736 (2011).

  29. 29.

    et al. Full-genome dissection of an epidemic of severe invasive disease caused by a hypervirulent, recently emerged clone of Group A Streptococcus. Am J Pathol. 180, 1522–1534 (2012).

  30. 30.

    et al. Genetic variation and linkage disequilibrium in Bacillus anthracis. Sci Rep. 1, 169 (2011).

  31. 31.

    et al. Genomic analysis identifies targets of convergent positive selection in drug-resistant Mycobacterium tuberculosis. Nat Genet. 45, 1183–1189 (2013).

  32. 32.

    On the number of segregating sites in genetical models without recombination. Theor Popul Biol. 7, 256–276 (1975).

  33. 33.

    Statistical method for testing the neutral mutation hypothesis by DNA polymorphism. Genetics 123, 585–595 (1989).

  34. 34.

    Evolutionary rate at the molecular level. Nature 217, 624–626 (1968).

  35. 35.

    , , & Haploview: analysis and visualization of LD and haplotype maps. Bioinformatics 21, 263–265 (2005).

  36. 36.

    & The advent of genome-wide association studies for bacteria. Curr Opin Microbiol. 25, 17–24 (2015).

  37. 37.

    Statistical tests for detecting gene conversion. Mol Biol Evol. 6, 526–538 (1989).

  38. 38.

    , & A Simple and Robust Statistical Test for Detecting the Presence of Recombination. Genetics 172, 2665–2681 (2006).

  39. 39.

    et al. Transfer of noncoding DNA drives regulatory rewiring in bacteria. Proc Natl Acad Sci USA 111, 16112–16117 (2014).

  40. 40.

    et al. Combined Analysis of Variation in Core, Accessory and Regulatory Genome Regions Provides a Super-Resolution View into the Evolution of Bacterial Populations. PLoS Genet. 12, e1006280 (2016).

  41. 41.

    & Inference of bacterial microevolution using multilocus sequence data. Genetics 175, 1251–1266 (2007).

  42. 42.

    et al. Evolutionary and population genomics of the cavity causing bacteria Streptococcus mutans. Mol Biol Evol. 30(4), 881–893 (2012).

  43. 43.

    et al. Genomic characterization of a pattern D Streptococcus pyogenes emm53 isolate reveals a genetic rationale for invasive skin tropicity. J Bacteriol. 198, 1712–1724 (2016).

  44. 44.

    Bacteriophage genomics. Curr Opin Microbiol. 11, 447–453 (2008).

  45. 45.

    , , , & Phage-associated mutator phenotype in group A streptococcus. J Bacteriol. 190, 6290–6301 (2008).

  46. 46.

    et al. Genome sequence of an M3 strain of Streptococcus pyogenes reveals a large-scale genomic rearrangement in invasive strains and new insights into phage evolution. Genome Res. 13, 1042–1055 (2003).

  47. 47.

    et al. Genome analysis of an inducible prophage and prophage remnants integrated in the Streptococcus pyogenes strain SF370. Virology 302, 245–258 (2002).

  48. 48.

    et al. Unique genomic arrangements in an invasive serotype M23 strain of Streptococcus pyogenes identify genes that induce hypervirulence. J Bacteriol. 196, 4089–4102 (2014).

  49. 49.

    , , & Penicillin and the marked decrease in morbidity and mortality from Rheumatic Fever in the United States. N Engl J Med. 318, 280–286 (1988).

  50. 50.

    , & Epidemiologic analysis of Group A Streptococcal serotypes associated with severe systemic infections, Rheumatic Fever, or uncomplicated Pharyngitis. J Infect Dis. 166, 374–382 (1992).

  51. 51.

    , & Macrolide treatment failure in streptococcal pharyngitis resulting in acute rheumatic fever. Pediatrics 129, e798–e802 (2012).

  52. 52.

    et al. Sensitive, specific polymorphism discovery in bacteria using massively parallel sequencing. Nat Methods 6, 67–69 (2009).

  53. 53.

    , & Simple models of genomic variation in human SNP density. BMC Genomics 8, 1–7 (2007).

  54. 54.

    et al. Detection and characterization of DNA variants in the promoter regions of hundreds of human disease candidate genes. Genomics 87, 704–710 (2006).

  55. 55.

    , , , & A scan statistic for identifying chromosomal patterns of SNP association. Genet Epidemiol. 30, 627–635 (2006).

  56. 56.

    & Mathematical model for studying genetic variation in terms of restriction endonucleases. Proc Natl Acad Sci USA 76, 5269–5273 (1979).

  57. 57.

    , & Estimation of levels of gene flow from DNA sequence data. Genetics 132, 583–589 (1992).

  58. 58.

    , , , & MEGA6: Molecular Evolutionary Genetics Analysis version 6.0. Mol Biol Evol. 30, 2725–2729 (2013).

  59. 59.

    , , & Mauve: Multiple alignment of conserved genomic sequence with rearrangements. Genome Res. 14, 1394–1403 (2004).

  60. 60.

    The Interaction of selection and linkage. I. General considerations; Heterotic models. Genetics 49, 49–67 (1964).

Download references


This work was supported by the Kleiderer-Pezold Chair (to FJC) and a Canada Research Chair (to BJS). The calculation in the current study was performed in high performance computing facilities in the Center for Research Computing, University of Notre Dame.

Author information


  1. W.M. Keck Center for Transgene Research, University of Notre Dame, Notre Dame, IN 46556, USA

    • Yun-Juan Bao
    • , Shaun W. Lee
    • , Victoria A. Ploplis
    •  & Francis J. Castellino
  2. Département de Sciences Biologiques, Université de Montréal, Montréal, QC H3C 3J7, Canada

    • B. Jesse Shapiro
  3. Department of Biological Sciences, University of Notre Dame, Notre Dame, IN 46556, USA

    • Shaun W. Lee
  4. Department of Chemistry and Biochemistry, University of Notre Dame, Notre Dame, IN 46556, USA

    • Victoria A. Ploplis
    •  & Francis J. Castellino


  1. Search for Yun-Juan Bao in:

  2. Search for B. Jesse Shapiro in:

  3. Search for Shaun W. Lee in:

  4. Search for Victoria A. Ploplis in:

  5. Search for Francis J. Castellino in:


F.J.C. and V.A.P. supervised the study and provided the data source material. Y.-J.B. performed the study. Y.-J.B., B.J.S. and S.W.L. performed the analysis and interpreted the data. Y.J.B. drafted the manuscript. All authors wrote and reviewed the manuscript.

Competing interests

The authors declare no competing financial interests.

Corresponding author

Correspondence to Francis J. Castellino.

Supplementary information

About this article

Publication history





Further reading


By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.