Genomic analysis of diversity, population structure, virulence, and antimicrobial resistance in Klebsiella pneumoniae, an urgent threat to public health

Whole-Genome Analysis Supports KpI, KpII, and KpIII as Distinct Species K. pneumoniae, K. quasipneumoniae, and K. variicola.

Within each phylogroup the mean pairwise nucleotide divergence between genomes was ∼0.5%, whereas nucleotide divergence between phylogroups was 3–4% (calculated across the core genes). The two KpII-A isolates were 1.8–1.9% divergent from KpII-B and were 3.2–3.7% divergent from KpI and KpIII. As the split network indicates (Fig. 1A), there was very little evidence of homologous recombination between phylogroups, with the exception of a single human gut carriage isolate from Vietnam (Fig. 1A and SI Appendix, Fig. S4). Further, principal components analysis (PCA) on accessory gene content clearly distinguished the four phylogroups (Fig. 1C). These data provide whole-genome support for the proposal that KpI, KpII, and KpIII are distinct species by demonstrating these phylogroups constitute discrete bacterial populations that are evolving independently, with limited homologous recombination between groups (Fig. 1). Between-phylogroup nucleotide conservation, at 96%, is at the level commonly used as a cutoff for species differentiation in taxonomic analysis (32), and differences in gene content (Fig. 1C) further support the proposition that KpI, KpII, and KpIII are separately evolving populations that can be considered as separate species.

The observed speciation into genetically distinct phylogroups indicates that there are barriers to gene flow among these closely related populations. These barriers could arise through ecological separation into distinct niches, mechanistic barriers to homologous recombination, or adaptive selection against hybrid genotypes (33). There are no obvious mechanistic barriers to homologous recombination between KpI, KpII, and KpIII; indeed the observation of a large recombination between KpI and KpII (SI Appendix, Fig. S4) shows that homologous recombination is possible, although the rarity of this event (1 out of >300 genomes) suggests there could be selection against such hybrids. Although our sampling of K. pneumoniae isolates was blind to the distinction of KpI from KpII and KpIII, the characteristics of isolates falling into each phylogroup were quite distinct (SI Appendix, Fig. S5), suggesting that their speciation is likely driven by long-term separation in distinct ecological niches.

Our isolate collection, which focused on human and bovine-associated bacteria but also contained isolates from nonhuman primates and marine mammals, was comprised mainly (87%) of K. pneumoniae KpI. Because a number of different criteria, mainly unrelated to core phylogeny, were used to select the isolates included in this study, there is unlikely to have been a sampling bias with respect to phylogroup. Therefore, we hypothesize that this preponderance of KpI is associated with the bias of our collection toward mammalian-associated infection isolates. Other studies of human clinical K. pneumoniae isolates report similarly high rates of KpI and low rates of KpII or KpIII (34, 35), and all the sequence types reported in the literature as being linked to hospital outbreaks or pyogenic liver abscess belong to KpI (including CC258 and CC23). Notably, all the publicly available genomes of K. pneumoniae clinical isolates that we analyzed, including the K. pneumoniae subsp. rhinoscleromatis reference genome, clustered within KpI (15).

Although both KpII (K. quasipneumoniae) and KpIII (K. variicola) are capable of causing infections in humans, they appear to be less pathogenic than KpI, being associated more frequently with carriage (SI Appendix, Fig. S5). KpII was found almost exclusively in humans but was generally associated with colonization (50%) or HA infection (25%), consistent with low virulence and opportunistic infection (SI Appendix, Fig. S5A). No KpII or KpIII isolates in our collection were linked to either liver abscess or the death of a patient. We detected no KpII among the bovine isolates. In contrast, almost half of our KpIII isolates were of bovine origin, compared with 20% of KpI isolates [odds ratio (OR) 5.2; P = 0.001; Fisher’s exact test) (SI Appendix, Fig. S5). The KpIII phylogroup was proposed in 2004 to be a distinct plant-associated nitrogen-fixing species, K. variicola, based on DNA–DNA hybridization and gene-sequence analysis (18). It has been isolated frequently from a wide range of plants (18, 36) and also has been shown to be an important nitrogen-fixing symbiont of leaf-cutter ants (37). Consistent with these reports, in our analysis the two public reference genomes of plant-associated K. pneumoniae belonged to KpIII (SI Appendix, Fig. S5). It is likely that the high number of bovine-derived KpIII isolates compared with human-derived KpIII isolates reflects bovine consumption of raw plant matter rather than any particular adaptation of KpIII to colonize or infect bovine hosts. Consistent with this notion, seven of the nine bovine KpIII were fecal carriage isolates, and only two were associated with infection. Importantly, our data show that the nif nitrogen-fixing operon (36) was present in all KpIII (K. variicola) genomes, supporting its identification as a nitrogen-fixing species. In contrast, nif was detected in only one KpI genome (a bovine mastitis isolate) and in half of the KpII-B genomes. This finding strongly supports the ecological separation of KpIII from KpI, with KpIII occupying a niche in which the ability to fix nitrogen is essential and KpI occupying a niche in which such ability is unnecessary and possibly disadvantageous and selected against. The intermediate frequency of nif in KpII is intriguing; because all our KpII isolates originated from humans, we hypothesize that nitrogen fixing is important in environmental-source populations of KpII but the nif operon is lost rapidly upon colonization of humans, possibly through negative selection.

Population Structure and Dynamics of K. pneumoniae KpI.

We identified a total of 91,898 core genome SNPs among 283 KpI genomes (247 newly sequenced and 36 publicly available genome sequences) and inferred from these SNPs an ML phylogeny (Fig. 2A) and neighbor-joining split network (SI Appendix, Fig. S6A). These revealed a deep branching, star-like population structure, suggesting an early radiation of K. pneumoniae KpI into hundreds of distinct equally distant lineages (Fig. 2A). The deep branching structure, which was supported by genome-specific and lineage-specific SNPs (SI Appendix, Fig. S6B), is polytomous at the root with low bootstrap support for sequential branching patterns (Fig. 2A). Differences in gene content provided further support for the inferred population structure (SI Appendix, Fig. S3).

We divided the KpI genomes into 157 distinct phylogenetic lineages based on analysis of the core gene ML tree using RAMI (Fig. 2A) (38). Median divergence between lineages was 0.46% (range 0.04–0.61%), whereas genomes within the same lineage differed by a median of 0.02% (range 0–0.08%) and generally shared the same MLST sequence type. We used fineSTRUCTURE (39) to investigate the relationships between KpI genomes and to identify clusters in a phylogeny-independent manner. These analyses identified some recombination within KpI, but the observed relationships were consistent with the ML tree and the phylogenetically defined lineages (Fig. 2A and SI Appendix, Fig. S6C). The lineage accumulation curve for each geographical location (Fig. 2B) indicated that further sequencing of many hundreds of additional isolates would be needed to capture the full diversity of the broader K. pneumoniae KpI population. The maintenance of so many distinct lineages of KpI in each geographical area studied is intriguing and may be driven by adaptive selection or may be simply a product of genetic drift in a large population. However, these factors likely play out outside of any association with humans; hence our dataset, which is heavily biased toward animal-associated isolates, is not well suited to distinguish these possibilities. The inclusion of isolates from wider environmental sources will be important to extend the population framework further, against which we can begin to separate clones associated with disease from those causing sporadic infections that, although acutely important to individuals, can be seen at the wider population level as background epidemiological noise.

Genetic Determinants of CA Invasive Infection in Humans.

K. pneumoniae KpI is best known as an opportunistic cause of HA infections, likely depending more on host factors such as compromised immunity than on specific pathogenicity factors in the bacterium. However, our human isolate collection included both HA (isolated >48 h after admission to hospital) and CA (isolated within 48 h of admission to hospital) KpI (Fig. 3). Further, 38 isolates came from CA invasive infections (defined as isolation of K. pneumoniae from a normally sterile site such as blood, CSF, intraocular, pleural, pericardial, or joint fluids, or deep-seated tissue abscesses), in which bacterial factors are likely to play a role in infection. Most (74%) of the 157 KpI phylogenetic lineages were observed only once within our collection; however, those that were represented by multiple isolates came from a diversity of specimen types (e.g., respiratory tract, urinary tract, digestive tract, blood) and from a mixture of carriage, invasive infection, and noninvasive infection (SI Appendix, Fig. S7). Thus, the ability of K. pneumoniae to cause invasive CA infections is not determined by lineage per se but may be associated with a specific virulence gene profile acquired horizontally and accessible through the pangenome.

Several genetic loci have been identified as virulence factors in K. pneumoniae on the basis of murine models of infection. These include gene clusters associated with the synthesis of siderophore systems yersiniabactin, aerobactin, colibactin, salmochelin (40–43), or microcin (44); the “regulators of mucoid phenotype” rmpA and rmpA2, which can up-regulate capsule production (45, 46); an allantoinase gene cluster (47); the ferric uptake operon kfuABC (48); and the two-component regulator kvgAS (49). Most of these genes were detected only in KpI, except for kfuABC (found in all KpII-B, 75% of KpIII, and 20% of KpI) and allantoinase (found in 50% of KpII-B and 5% of KpI). To investigate whether these genes also are associated with the ability cause disease in humans, we examined their distribution among human KpI isolates associated with invasive infection, noninvasive infection, and asymptomatic gut or throat carriage (Fig. 3A). Genes rmpA and rmpA2 and the siderophore clusters were significantly associated with invasive human infection, compared with noninvasive or carriage isolates (ORs 3–15) (Fig. 3A).

Yersiniabactin, whose synthesis is encoded by the ybt, irp1, irp2, and fyuA genes which form the Yersinia high-pathogenicity island, HPI (40), was the most prevalent virulence-associated locus, present in one third of the KpI human isolates (Fig. 3A). Despite this high prevalence it was a strong predictor of infection vs. carriage in humans, with an OR of 7.4 [95% confidence interval (CI), 2.2–40; P = 0.0001; Fisher’s exact test] and a positive predictive value of 95%. This effect was not dependent on chromosomal background, because yersiniabactin was significantly associated with infection in a logistic regression model that included phylogenetic lineage (OR 1.3; P = 0.003). Similarly, all K. pneumoniae encoding aerobactin (iucABCD, iutA), colibactin (clbA-R), salmochelin (iroN, iroBCD), and rmpA or rmpA2 were isolated from human infections. These gene clusters were in strong linkage with each other as well as with yersiniabactin, and each was detected in 9–16 different KpI lineages (SI Appendix, Figs. S7 and S8). The combination of salmochelin, aerobactin, and rmpA was frequently, but not always, linked to the presence of genes from the known K. pneumoniae virulence plasmids pLVPK and pK2044 (43, 50) (SI Appendix, Table S1).

To extend these observations, we performed a PGWAS, screening each gene in the KpI pangenome for association with infection in humans. The strongest associations, reaching pangenome-wide significance after correcting for multiple testing, were the rmpA/2 and siderophore genes and five additional predicted iron-metabolism genes (OR 8–10; 95% CI, 3–35) (SI Appendix, Fig. S9), also present on the virulence plasmid pK2044. Next we assessed whether the siderophores and predicted iron-metabolism genes could explain CA invasive infections (Fig. 3B). They were very common among CA invasive infections (75% carried one or more, and 60% carried three or more) but were rare among other classes of isolates (<10%) (Fig. 3B). These results are striking and suggest that access to iron may be central to the ability of K. pneumoniae to cause invasive disease in immunologically competent human hosts. Iron is essential for cell growth and replication in the human body, and nearly all free iron is bound by host proteins. Therefore bacteria must compete with host systems for iron, and siderophores are known to be critical to the ability of many bacteria to grow and replicate during colonization and infection of hosts (51). In addition, siderophores also can play other roles in the interaction of K. pneumoniae and other Enterobacteriaceae with hosts, including modulating immune responses via the host protein lipocalin 2, binding noniron metal ions, and protecting against reactive oxygen species (51).

Because serotyping is technically challenging and unreliable, we assessed capsular variation based on allelic diversity of the wzi gene. We identified 136 alleles, including 118 in KpI. There was strong evidence of horizontal transfer: 27 wzi alleles were found in two or more different KpI lineages, and 20 of the 37 lineages with two or more genomes contained two or more wzi alleles (SI Appendix, Fig. S7). It is unclear what the drivers of capsular diversity are in K. pneumoniae. K. pneumoniae is an opportunistic pathogen that does not depend on colonization or infection of humans and other animals for survival; therefore it seems unlikely that capsular exchange within clones is associated with selection for host immune evasion, as is evident in human-adapted bacteria such as Streptococcus pneumoniae (52) or Neisseria meningitidis (53). K. pneumoniae capsule types K1, K2, and K5 have been proposed as virulence factors associated with pathogenicity in humans and in mouse models; however, in our dataset, all CA invasive infection isolates with the K1, K2, or K5 wzi alleles also carried siderophore genes that potentially could explain their virulence, and these wzi alleles also were found in isolates from asymptomatic carriage and HA infections.

The Emergence of Virulent K. pneumoniae Clones.

We identified several K. pneumoniae clones that were significantly enriched for siderophores and/or rmpA genes (compared with the rest of KpI using Fisher’s exact test) (SI Appendix, Fig. S7A). The best-known virulent K. pneumoniae clone is ST23, which is able to cause severe disease in apparently healthy individuals (13, 54) and typically carries all four acquired siderophore systems as well as rmpA and rmpA2. The 13 ST23 isolates in our collection were isolated from invasive infections more frequently than were other KpI (69% of ST23 were invasive vs. 35% of other KpI; OR 4.4, P = 0.01; Fisher’s exact test) and were associated solely with CA infections. Our population genomic framework highlights how unusual these characteristics of ST23 are in the context of the broader population of KpI and also shows that each of the yersiniabactin, salmochelin, aerobactin, colibactin, rmpA, rmpA2, and microcin gene clusters of ST23 were associated with severe human disease in other KpI lineages with global distributions and high rates of invasive disease (SI Appendix, Fig. S7B). A feature shared by these clones that distinguishes them from the other noninvasive lineages is the presence of yersiniabactin, salmochelin, and rmpA in various combinations: (i) ST65, which also carried colibactin and rmpA2, was associated with lethal infections in humans and marine mammals [and has been associated with liver abscesses in Taiwan (55)]; (ii) ST592, which also carried aerobactin and rmpA2, was associated with a liver abscess and sepsis; (iii) ST25, which sometimes also carried aerobactin and rmpA2, was associated with liver abscesses in Vietnam and sepsis in Laos; (iv) ST60 was isolated from sepsis and also from an abdominal abscess in a monkey. ST231, which carried yersiniabactin and sometimes aerobactin, was isolated from cases of CA lethal pneumonia and abscesses and was multiply drug resistant.

Using the K. pneumoniae population framework, we also were able to make finely detailed comparisons between closely related isolates of the same lineage(s) that were differentiated by recorded clinical outcome. These data revealed the emergence of virulence within individual KpI clones, whereby invasive isolates were differentiated from noninvasive isolates of the same lineage by the presence of rmpA and siderophores. These clones include ST43 (yersiniabactin, salmochelin, aerobactin, and rmpA) and ST36 (yersiniabactin, colibactin, aerobactin, and rmpA), both linked to bacteremia. Within ST1, ST14, ST15, ST35, and ST48, the acquisition of yersiniabactin was linked to bacteremia and sepsis (SI Apppendix, Fig. S7B). These data build on previous observations made from a restricted number of isolates (47, 56, 57) by showing that the effect on virulence of acquiring iron-scavenging genes is not dependent on strain background or on specific combinations of siderophores and that the acquisition of iron-scavenging systems is central to the ability of K. pneumoniae to cause invasive disease in nonimmunocompromised patients. Importantly, this finding suggests that the acquisition of any one of the siderophore clusters increases the risk of severe infection in humans.

Genetic Determinants of Mastitis in Cows.

K. pneumoniae is a frequent cause of bovine mastitis in dairy herds, but no known pathogenicity factors have been reported. Here we found that there was no association between observed mastitis and K. pneumoniae lineage(s). Moreover, genes that were associated with invasive infection in humans were rare among bovine isolates and were not associated with mastitis. PGWAS analysis for mastitis vs. carriage among bovine KpI isolates identified three genes significantly associated with KpI mastitis in cows (OR 30; 95% CI, 5.7–247; P < 1 × 10⁻⁶, adjusted for multiple testing P = 0.03). These genes formed a lac operon cluster encoding a transcriptional regulator (LacI), β-d-galactosidase (LacZ), and lactose permease (LacY) that is distinct from the lactose operon conserved in the chromosome of other KpI and KpIII isolates. The acquired lac operon was present in 18/20 KpI clinical mastitis isolates, 8/10 KpI subclinical mastitis isolates, and 3/19 KpI bovine fecal isolates, indicating association with both clinical and subclinical infection. This finding suggests that the utilization of lactose from the cow using this second lactose operon may confer an important selective growth advantage to K. pneumoniae isolates associated with mastitis. Similar associations have been made in Streptococcus agalactiae (58).

The acquired lac operon was common in K. pneumoniae, present in 50% of the KpI (SI Appendix, Fig. S10), 60% of the KpII, and 16% of the KpIII, but was not associated with infection or specimen type among human isolates. Notably, the acquired lac operon was located adjacent to a copy of the fec iron-enterobactin operon, which also showed a positive association with mastitis (OR 20; P < 1 × 10⁻⁴) and bovine host (OR 2; P = 0.0005). The combination of fec and lac operons likely provides K. pneumoniae the ability to invade via the udder and to thrive within mammary epithelial cells. The 30 bovine isolates carrying these operons were found in 23 different KpI lineages as well as in a KpII, confirming they are subject to extensive horizontal transfer (SI Appendix, Fig. S10). Moreover the fec and lac operons are known to be colocated together on numerous unrelated K. pneumoniae plasmids (pKPN3, pKN_LS6, pKPN_IT, pKP007, pKPN_CZ, pK29, and JM45_p1), indicating that the operons themselves are mobile and in strong genetic linkage within K. pneumoniae.

Antimicrobial Resistance Genes in K. pneumoniae.

Given the clinical importance of AMR in K. pneumoniae, we performed a targeted analysis of all known AMR genes within our genomic dataset. We detected 84 AMR genes among the newly sequenced isolates (SI Appendix, Fig. S8). Our data confirm that the SHV, OKP, and LEN β-lactamases are core chromosomal genes of KpI, KpII, and KpIII, respectively (SI Appendix, Fig. S8). FosA and oqxAB, which confer low-level resistance to fosfomycin and quinolones, respectively, and have been detected as horizontally acquired AMR determinants in Escherichia coli, were core to all three phylogroups (SI Appendix, Fig. S8) and likely originated in K. pneumoniae. A further 78 AMR genes were detected in our collection, in 150/288 isolates (SI Appendix, Fig. S8). It was evident that the distribution of these AMR genes varied substantially among locales, with many resistance genes associated with particular countries or regions but not with lineage, presumably reflecting differing local antimicrobial use or availability (SI Appendix, Fig. S11) and most likely resulting from horizontal transfer of AMR genes among locally circulating bacterial populations. We refer to these genes as “acquired AMR genes” to differentiate them from the core β-lactam resistance genes described above.

Acquired AMR genes were found commonly in the human-associated lineages KpI and KpII-B. In KpII-B, the human carriage-linked phylogroup, the median number of observed AMR genes was five per isolate. In KpI, acquired AMR genes were associated with humans (median 5 vs. 0; P = 0.002; Wilcoxon rank test) and were more common in carriage isolates than in infection isolates (median 9–10 vs. 4; P < 1 × 10⁻⁸) (Fig. 4A). Similarly, extended-spectrum β-lactamases (ESBLs), including ESBL alleles of SHV, were strongly associated with human isolates overall (detected in 40% of human isolates vs. 2% of bovine isolates; P < 1 × 10⁻⁶; Fisher’s exact test), especially in the human-associated KpII-B group (50%). ESBL genes were not found in our KpIII isolates (SI Appendix, Fig. S5).

Nosocomial carriage isolates had more acquired AMR genes than community carriage isolates (median 13 vs. 7; P = 0.08) (Fig. 4C), implying that AMR genes have a role in opportunistic HA infection within the K. pneumoniae population. These patterns also were evident at the level of individual genes: Among human K. pneumoniae, most AMR genes were more common among HA isolates than among CA isolates (P < 1 × 10⁻¹⁶; paired Wilcoxon test) and also were more common among carriage isolates than among isolates associated with infection (P < 1 × 10⁻⁶) (Fig. 4B). Each of the individual AMR genes was detected at lower rates in bovine isolates than in human isolates, with the exception of tet(B),which was found in 10% of bovine isolates and in only 1% of human isolates.

We were unable to link resistance loci reliably to specific plasmids because of the constraints inherent in using short-read Illumina data to generate whole-genome assemblies and the often repetitive nature of mobile elements that frequently disrupt those assemblies. However, screening against the PlasmidFinder database (59) identified 28 known plasmid replicons in 69/150 K. pneumoniae isolates with acquired AMR genes. These included six colicin plasmid replicons but also included a large range of replicons associated with large known conjugative AMR plasmids (SI Appendix, Table S1).

These data show that hospital isolates and carriage isolates accumulate AMR genes, likely through selection from antimicrobial exposure during long-term carriage or during hospital treatment. The more “successful” HA infection-associated clones, which lack virulence genes but appear multiple times in our collection, were MDR and had the greatest numbers of AMR genes (SI Appendix, Fig. S7). However, the precise complement of AMR genes differed among isolates of the same lineage (SI Appendix, Fig. S7), indicating distinct and potentially quite frequent gene-acquisition events. This finding suggests that the accumulation of resistance genes in nonvirulent K. pneumoniae clones is a consequence, rather than a driver, of the successful spread of such clones within the human population, which is achieved mainly through effective asymptomatic colonization. This notion is consistent with the characterization of K. pneumoniae HA infections as resulting from opportunistic growth in nongut replicative niches that is selected for by the use of antimicrobials and can overwhelm compromised host immunity.

The Emerging Threat of Highly Pathogenic XDR K. pneumoniae.

Given the existence of multiple virulent and MDR KpI clones that have access to a diverse mobile pool of virulence and AMR genes of high penetrance, and given the strong selective pressures imposed on bacteria in the hospital setting, there is potential for the emergence of an extremely drug-resistant (XDR), hypervirulent K. pneumoniae clone capable of causing severe, untreatable infections in healthy individuals. Indeed, convergence of specific virulence genes with AMR in K. pneumoniae already is beginning to occur. Many isolates of the epidemic KPC-producing ST258/ST11 clonal complex (CC258) already have acquired yersiniabactin (SI Appendix, Fig. S7). Experimental infection models show that yersiniabactin in CC258 enhances the bacterium’s ability to colonize the respiratory tract and cause pneumonia (60). The presence of yersiniabactin in CC258, and also in the ESBL clones ST14 and ST15, is worrying, because our data show that it not only is strongly associated with infection in humans but also appears to be a frequent first step in the acquisition of additional siderophores that augment the ability of KpI clones to cause invasive non–hospital-related infections. There also is a high risk that current hypervirulent clones may acquire AMR, and MDR ST23 already has been reported in Korea, Vietnam, China, Madagascar, and Brazil (12, 61–63).

Now that the associations have been recognized, new efforts can be more focused on the task of reconstructing the mobile elements that carry the critical virulence genes, ideally using long-read sequencing approaches that have proven successful for unraveling MDR plasmids (64). However, given the plasticity of the K. pneumoniae pangenome and the broad distribution of key virulence and AMR genes, we need to place all K. pneumoniae within a wider population framework to target known epidemic clones and increase the likelihood of identifying new and emerging clones. Clinically, AMR phenotypes are monitored routinely in most hospital laboratories; however, our study indicates that it also will be crucial to perform active surveillance for key virulence genes and to determine clonal background. Genomic surveillance is being used increasingly to monitor KPC-producing K. pneumoniae, especially ST258, and the emergence of NDM-1 and colistin-resistant (XDR) isolates (27, 28, 65–67). Crucially, our study shows that we can augment these surveillance efforts by using key virulence genes as strong predictors of invasive disease in humans, and by determining clonal background, so that we can identify and track XDR hypervirulent clones as they emerge. To facilitate future genome-based surveillance, the genomes presented here have been deposited in the K. pneumoniae BIGSdb database (bigsdb.web.pasteur.fr) (12), which includes a core genome MLST scheme as well as key K. pneumoniae accessory genes encoding critical determinants of AMR, virulence, and capsule type. The data presented here provide a new genomic framework with which a new and deeper understanding of the K. pneumoniae population can be developed (SI Appendix, Fig. S12). Together with the K. pneumoniae BIGSdb database, this framework will provide a critical foundation and practical support for future studies investigating ecological niche adaptation, pathogenicity, and lineage diversification in K. pneumoniae and will facilitate more deeply informed genomic tracking and surveillance for the emergence and convergence of virulence and AMR in this increasingly important pathogen.

Bacterial Isolates and Sequencing.

A total of 288 bacteria isolates, sampled to maximize diversity, were contributed from coauthors in six countries (SI Appendix, Supplementary Text). Genomic DNA was extracted and sequenced via Illumina HiSeq; data have been deposited in the European Nucleotide Archive under accession no. ERP000165. All K. pneumoniae genome data publicly available in the PATRIC database in March 2013 also were included in the analysis; all isolates and metadata are given in Dataset S1.

Clinical Definitions.

Infection vs. carriage status was defined on the basis of available clinical data associated with each sample, that is, whether the isolate was considered, at the time of isolation, to be the cause of an infection. Invasive infections were defined as those associated with the isolation of K. pneumoniae from a normally sterile site (blood, CSF, intraocular, pleural, pericardial or joint fluids, or deep-seated tissue abscesses). The remaining infections, classed as noninvasive, were pneumonia, urinary tract infections, or wound infections with no recorded bacteremia. Bacteria isolated from samples taken >48 h after admission to hospital were classified as nosocomial; those isolated within 48 h of admission to hospital were classed as CA. This information was unavailable for 31 human KpI isolates, which were excluded from all analyses of mode of acquisition.

Variant Detection and Phylogenetic Analysis.

SNPs were identified via mapping of Illumina reads to a reference genome (K. pneumoniae strain NTUH-K2044, NC_006625.1). Core genes were defined as the 1,743 NTUH-K2044 chromosomal genes with 100% coverage in all 328 genomes; sites outside these genes were excluded, leaving 175,120 SNPs for phylogenetic analysis. The alleles at these sites were concatenated to form a multiple alignment of SNPs for phylogenetic analyses. KpI lineages were defined based on patristic distances in the ML tree using RAMI (38), supported by a phylogeny-free approach using fineSTRUCTURE (39). STs were assigned to each genome according to the K. pneumoniae MLST database (13) by mapping to known alleles using SRST2 (68). Full details of all analyses are provided in SI Appendix, Supplementary Text.

Gene Content Analysis.

De novo assemblies of Illumina reads were generated using Velvet and were combined with publicly available genomes to generate a nonredundant set of pangenome sequences. Taxonomic assignment was performed using MG-RAST v3.2 (69). A gene content matrix, indicating coverage of each gene in each genome, was generated by mapping to the annotated pangenome sequence. Read sets also were screened for known alleles of important genes using a read mapping approach with SRST2 (68). Gene databases analyzed were AMR alleles [ARG-Annot (70)], plasmid replicons [PlasmidFinder (59)], and virulence and wzi alleles [K. pneumoniae BIGSdb (12, 71)]. Full details of all analyses are provided in SI Appendix, Supplementary Text.

Statistical Analyses.

Statistical analyses were performed in R as detailed in SI Appendix, Supplementary Text. Briefly, accumulation curves were generated using the vegan package. PCA was performed using the prcomp function. Pairwise gene content distances were calculated as Jaccard distances. PGWAS was done using Fisher’s exact test and Benjamini–Hochberg correction of P values. Differences in the total number of acquired AMR genes among various isolate classes were assessed using the Wilcoxon test. Median nucleotide divergence within and between lineages or phylogroups was calculated from pairwise distances between genomes, obtained for each pair by dividing the total number of variant SNPs by the total length of the core genes analyzed (1,475,502 bp).