Ethical considerations

Each contributing study obtained local ethical approval, details of ethics committees and approval numbers are listed in S1 Table. The Baby GERMS-SA [29] and Malawi–Liverpool–Wellcome (MLW) Biobank [22] studies were granted consent waivers from local ethics committees for the use of routine diagnostic specimens/isolates and routine clinical data for research; all other studies obtained written informed consent from the parents of participating neonates. Approval for the cross-study analysis presented here was granted by the Observational/ Interventions Research Ethics Committee of the London School of Hygiene and Tropical Medicine (ref #29931). Anonymised data from each primary study were shared for analysis, including date of specimen collection and hospital site identifier (where the study included more than one site), along with pathogen WGS data for the corresponding bacterial isolate.

Bacterial isolates and sequence analysis

Whole-genome sequences of K. pneumoniae isolated from neonates in LMICs in Asia and Africa were sourced from prospective clinical studies of neonatal sepsis: Burden of Antibiotic Resistance in Neonates from Developing Societies (BARNARDS) [5], Sepsis Prevention in Neonates in Zambia (SPINZ) [30–32], Baby GERMS-SA [29,33], Mortality from Bacterial Infections Resistant to Antibiotics (MBIRA) [34], Indian District Hospitals (DH) [35], Group B Strep Correlates of Protection Study (GBS-COP) [36], Neonatal Infections & MicroBIome (NIMBI-plus) [37], NeoBAC [38]; and long-term prospective surveillance of bloodstream infection at Queen Elizabeth Central Hospital in Blantyre, Malawi (MLW Biobank) [22] and Kilifi County Teaching and Referral Hospital (KCTRH) carried out by the Kenya Medical Research Institute (KEMRI)/Wellcome Trust Research Programme (KWTRP) [38]. All included studies were undertaken as observational or surveillance studies aiming to capture all blood culture-positive sepsis cases during a defined period, not specifically instigated to investigate or respond to outbreaks. These studies and the corresponding whole-genome sequence data are a subset of those included in a meta-analysis of K and O antigen prevalence amongst K. pneumoniae causing neonatal sepsis in LMICs published in 2026 [39]. That study adjusted for within-unit clustering to minimise the impact of localised transmission events on the estimation of regional antigen prevalence, but did not investigate transmission as a phenomenon, which is the goal of the present study. Details of samples included from each study are given in Table 1, and characteristics of each study site are given in Table 2 (this represents information provided by co-authors from each study team). The prospective analysis plan for this study has been published elsewhere (see S1 Protocol in [39]). We used the United Nations Statistics Division M49 standard to assign countries to geographical sub-regions (Asian countries) and intermediate regions (African countries) [40].

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Summary of isolates included in the study.

https://doi.org/10.1371/journal.pmed.1005077.t001

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Characteristics of facilities included in the study.

https://doi.org/10.1371/journal.pmed.1005077.t002

For the prospective studies (BARNARDS, SPINZ, Baby GERMS-SA, MBIRA, DH, GBS-COP, NIMBI-plus, NeoBAC), blood culture was performed for all neonates with clinically suspected sepsis, and all blood culture isolates identified as Klebsiella that could be later revived for DNA extraction were included in sequencing and this analysis. Baby GERMS-SA, GBS-COP, KWTRP, and MLW also included Klebsiella cultured from cerebrospinal fluid (CSF) of participants. For the long-term prospective surveillance studies (KWTRP, MLW), isolates from all positive blood or CSF cultures were stored for future research. At KWTRP, all neonates had blood culture on admission and again if their clinical condition deteriorated. At MLW, cultures were performed for all neonates with clinical signs of sepsis or meningitis, temperature >37.5°C, or other signs of clinical deterioration.

For all studies, neonatal isolates identified as Klebsiella, which could later be revived for DNA extraction, were sequenced, and those sequences that passed quality filters were included in this analysis (see inclusion criteria below). Details of microbiology, sequencing, and bioinformatics methods used in each study, including the quality-control criteria applied prior to sharing genome sequences for this analysis, have been reported previously [39] and are summarised in S2 Table. Briefly, all studies utilised Illumina platforms and assembled genomes using either SKESA v2.3.0 (MBIRA) or SPAdes (all other studies). Genome assemblies were analysed using Kleborate v3.0 [41] to confirm species, identify multi-locus sequence types (STs), AMR and hypervirulence determinants, and to identify capsular (K) loci (KL) and O types using Kaptive v3.0 [42].

Inclusion and exclusion criteria

Inclusion criteria for individual samples from the studies were: K. pneumoniae isolated from the blood or CSF of a neonate (defined as 0–28 days post birth), between 2013 and 2023 inclusive. CSF isolates were included from three studies (KWTRP surveillance, n = 5; GBS-COP, n = 2; MLW Biobank, n = 31). Repeat K. pneumoniae isolates from the same participant with the same KL were excluded, ensuring that each genome represents a unique infection from an individual neonate (with the exception of MLW Biobank, for which archived isolates were not associated with patient identifiers thus precluding the exclusion of repeat isolates). Genome sequences identified as species other than K. pneumoniae or failing to meet the quality control criteria for assemblies (≤1,000 contigs and genome size 5–6.2 Mbp, as reported by Kleborate v3) were excluded. Additionally, study sites contributing fewer than 10 high-quality K. pneumoniae genomes that met the inclusion criteria were excluded (54 genomes from 12 sites), as it is not meaningful to estimate a fraction when the denominator is < 10. Specifically, data were excluded from 12 sites that participated in multicenter studies (MBIRA, BARNARDS, DH, GBS) but yielded fewer than 10 high-quality K. pneumoniae genomes. Seven of these sites had <5 high-quality K. pneumoniae genomes each (no clusters as per the definition below); five sites had n = 6–8 isolates each, 3 of which had a cluster. The reasons for low counts vary but are likely associated with differences in patient numbers as well as diagnostic stewardship and sensitivity. A flow diagram detailing sample inclusion is shown in S1 Fig, including available information on the number of samples that were not stored, could not be revived for DNA extraction, failed sequencing, or did not meet genome quality control thresholds for each contributing study.

Identification of transmission clusters

Pairwise single-nucleotide variant (SNV) distances were calculated using Pathogenwatch, which enumerates differences in a library of 1,972 core genes (totalling 2,172, 367 bp) [43]. Transmission clusters were defined within each study site based on both the pairwise genetic distances (≤10 SNVs) and temporal distances (≤4 weeks based on date of specimen collection) between isolates. The igraph v1.4.1 R package [44] was used to identify single-linkage transmission clusters by constructing an undirected graph from an edge list of isolate pairs. Each edge in the undirected graph represented a pairwise connection between isolates that were genetically similar (pairwise distance less than or equal to the specified genetic distance threshold), sampled within a specified temporal window (up to the specified temporal distance threshold), and isolated from the same site. Distinct transmission clusters were then defined as connected components of the resulting graph. This results in single-linkage clusters, where each isolate is within the thresholds of both genetic and temporal distance from at least one other isolate in the cluster, but pairs of isolates within the same cluster may be separated by distances greater than the threshold. Therefore clusters can include isolates cultured over any period, as long as the distances between consecutive isolates in the cluster fall below the temporal threshold.

Estimation of cluster and transmission proportions

We estimated two key parameters—cluster proportion and transmission proportion—as indices of the contribution of transmission to the burden of K. pneumoniae neonatal disease in each neonatal care setting. The cluster proportion estimate was calculated as the number of K. pneumoniae isolates in clusters, divided by the total number of culture-confirmed and sequenced K. pneumoniae, for each site. The transmission proportion, on the other hand, was a conservative estimate of the proportion of K. pneumoniae isolates that were attributable to transmission. To estimate this, we excluded the first neonate in each cluster as a potential index patient, and assumed that all other infections within clusters were due to onward transmission in the unit. This almost certainly results in an underestimation of infections attributable to transmission, as the index patient may have also acquired K. pneumoniae from a contaminated source within the unit or an asymptomatically colonised neonate. The transmission proportion estimate was thus calculated as below:

Sensitivity analyses

We conducted a sensitivity analysis to determine the impact of the choice of genetic distance and temporal distance thresholds on the estimates of cluster proportion and transmission proportion. The estimates were calculated and compared across a range of genetic distance thresholds (range: 0–25 SNVs) and temporal distance thresholds (range: 1–52 weeks). We also assessed how using different methods for calculating pairwise genetic distances between all isolate pairs could impact the obtained estimates. To do this, we generated reference-free SNV alignments of all genomes in the BARNARDS and SPINZ datasets using the split k-mer analysis (SKA) method with SKA2 v0.4.1 [45,46]. A SKA index was first built from the input assemblies using a k-mer size of 31, after which pairwise distances between all genomes were calculated using the “distance” subcommand with a minimum frequency threshold of 0.9 to include only core-genome SNVs. The cluster and transmission proportion estimates were re-calculated using these SKA-derived SNV distances and compared to those obtained using the Pathogenwatch core gene SNV distances.

Statistical analysis

All analyses and visualisations were performed using R v4.2.3 [47].

The methods for clustering infections, quantifying transmission, and exploring the sensitivity of the estimates to selected thresholds (as described above) are implemented in a species-agnostic Shiny web application (https://klebsiella.shinyapps.io/transmission_estimator/), with source code available at https://github.com/klebgenomics/transmission_estimator (DOI: https://doi.org/10.5281/zenodo.17593948).

To assess heterogeneity of estimates between sites, for the multicentre studies BARNARDS and MBIRA, we performed meta-analysis of proportions using a generalised linear mixed model (GLMM) with a logit transformation to pool estimates of cluster proportion and transmission proportion across all study sites. A random effects model was applied to account for between-site variability, and 95% confidence intervals were computed using the Clopper–Pearson method. Heterogeneity between sites was assessed visually using Forest plots and quantitatively using the I² statistic. The meta-analysis was conducted using the metaprop function of the meta v7.0.0 R package [48].

To identify facility characteristics associated with transmission, we performed pairwise two-tailed binomial tests with Bonferroni correction, with facility-level data as categorical (facility size, number of neonatal beds, availability of piped water) or binary (onsite availability of neonatal surgical facilities) variables.

The proportions of isolates with extended-spectrum beta-lactamase (ESBL) and carbapenemase genes were compared between clustered and non-clustered cases using the Chi-squared test. To model different bacterial features (ESBL gene carriage, carbapenemase gene carriage, ST, O type, K locus) and institutional factors (hospital site) as potential predictors of transmission, we fitted logistic regression models in which each introduction was treated as a single observation, with the outcome recorded as transmission (cluster size ≥2) or no evidence of transmission (unclustered singleton case). For each cluster, each predictor was assigned the consensus value for that cluster (e.g., if 3 of 4 cases had ESBL gene/s detected, the cluster was recorded as ESBL-positive). Presence of ESBL (Kleborate score ≥1) or carbapenemase (Kleborate score ≥2) genes were encoded as binary variables. ST was encoded as a categorical variable, with one category per ST for commonly transmitted STs (occurring in ≥3 clusters and ≥2 sites), and the reference category being the group of all 158 ′other STs’ (49.2% in the reference category). O type was encoded as a categorical variable, with the most common type, O1 (51.6%), as the reference category. K locus was encoded as a categorical variable, with the top 10 K loci compared against ‘other K loci’ as the reference category (n = 70, 87.5% in the reference category). Models were fitted using Firth’s bias reduction method, implemented in the logistf package in R (v1.26.0). All alternative models were compared using the likelihood ratio test, using the anova function in base R. This study is reported as per the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guideline (S1 Checklist).

Estimating transmission proportion

The study included 1,523 isolates from 27 sites in 13 countries, including Malawi (n = 389), Zambia (n = 304), Kenya (n = 233), Ethiopia (n = 125), Tanzania (n = 58) and Rwanda (n = 15) in Eastern Africa; Nigeria (n = 50) and Ghana (n = 25) in Western Africa; South Africa (n = 214) and Botswana (n = 29) in Southern Africa; India (n = 34), Pakistan (n = 30) and Bangladesh (n = 17) in South Asia (Table 1 and S2 Fig). Sample size per site ranged from 12 to 362 genomes, with dates of isolation spanning 26–537 weeks per site. Six sites included data from only one year. A majority of the isolates were from Eastern Africa (n = 1,124), followed by Southern Africa (n = 243), South Asia (n = 81), and Western Africa (n = 75). See Table 2 for characteristics of neonatal care facilities included in the study.

A total of 156 transmission clusters were identified across all sites, comprising 1,035 unique infections (68.0%) (S3 Fig). Excluding the first infection in each cluster as a potential index case, we estimate at least 879 (57.7%) infections were acquired via nosocomial transmission. The identified clusters spanned between 1 and 249 days duration between the first and last cases (median = 14 days; interquartile range, IQR: [6, 28 days]), with a median of 3 infections per cluster (range: [2, 188]; IQR: [2, 5]) (S4 Table). Almost half (73/156) of the clusters comprised only 2 infections. Excluding these 73 clusters as potential coincidental isolation of similar but independently introduced strains rather than nosocomial transmission, there were still 889 (58.4%) infections involved in 83 clusters affecting at least 3 patients each (median cluster size, 5 cases). Sixteen clusters overlapped either the beginning (n = 7; median cluster size = 3 [range: 2, 188]) or the end (n = 9; median cluster size = 3 [range: 2, 32]) of the sampling periods in the respective sites, suggesting that transmission may have been ongoing before sampling began or continued after sampling ended.

The median estimates, across 27 sites, for cluster and transmission proportion were 53.3% and 33.5%, respectively (Fig 1). Sensitivity analyses showed these estimates were robust to the choice of thresholds and genetic distance estimation methods (S4 Fig and S1 Appendix). Cluster and transmission proportion estimates varied widely by site, ranging from 0% to 93% and 0% to 87%, respectively (Fig 1). In 10 of the 27 sites (37%), at least half of all infections were attributable to transmission, and in 14 sites (52%) at least a third of cases were attributable to transmission. Across all sites, the median number of clusters detected per site per year was three (IQR: [2, 4 clusters]; range: [0, 13 clusters]). Considering only clusters with ≥3 patients, we identified median one cluster per site per year (IQR: [1, 3 clusters]; range: [0, 8 clusters]). Transmission proportions were heterogeneous within regions, although higher values were estimated in Eastern Africa (median = 65.0% [range: 29.6, 93.5%]; n = 11 sites) compared with Western Africa (40% [range: 37.5, 41.2%]; n = 3 sites) and South Asia (43.3% [range: 23.5, 88.2%]; n = 3 sites). Intermediate values were estimated in Southern Africa (median = 54.8% [range: 0, 77.3%]; n = 10 sites).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Estimates of cluster and transmission proportion across sites.

(A) Cluster proportion (B) Transmission proportion. Plots show the point estimates (boxes) and 95% confidence intervals (horizontal bars) for each site. The median estimate is represented by the broken vertical line.

https://doi.org/10.1371/journal.pmed.1005077.g001

To determine whether differences in study design and sampling methodology contributed to the heterogeneity observed between sites, we conducted two secondary meta-analyses using data from two multi-site studies in our dataset (MBIRA and BARNARDS), each of which used a consistent study protocol across multiple sites [5,34]. The pooled estimate of transmission proportion across the seven BARNARDS sites was 26.1% (95% CI: [11.5, 49.0%]; range: [0, 79%]) and there was substantial heterogeneity between sites (I² = 88.6% [95% CI: 79.0, 93.8%]). The findings with the MBIRA dataset were similar, with a pooled transmission proportion estimate of 37.0% (95% CI: [15.8, 64.8%]; range: [8, 86%]) and significant heterogeneity (I² = 89.0% [95% CI: 78.8, 94.3%]) was observed across the six sites. Notably, the lowest cluster proportion observed in both these studies was at the same neonatal unit, sampled under different protocols at different time periods (BARNARDS 15A, 2016–2017 and MBIRA 15B, 2021), but yielding similar results each time (0 clusters out of 12 isolates, and 1 clustered pair out of 13 isolates, respectively). This suggests that the observed heterogeneity in the full analysis is unlikely to be influenced by primary study protocol differences alone, but may instead reflect variation in unmeasured factors such as demographic, geographic and/or site-specific variations.

S5 Fig shows a comparison of cluster proportion across different sites based on descriptive characteristics of the health facilities (summarised in Table 2). Notably, facility size and neonatal bed count were not associated with the number of K. pneumoniae isolates per site (R² < 0.01, p = 0.993 and R² = 0.14, p = 0.053, respectively, using linear regression). For example, the lowest cluster proportions were observed at site 15, which is one of the largest neonatal units but as noted above had very few cases or clusters, potentially because it is one of the better-resourced facilities (located in Cape Town, South Africa with relatively strong microbiological diagnostics, surveillance and IPC capacity). Facilities with only occasional availability of piped water (two in Ethiopia, one in Zambia) had a higher proportion of cases in clusters (87.7% [range: 64.5, 89.4%]) than those where piped water was reported as being available most of the time (n = 6; 42.5% [range: 23.5, 74.0%]; p < 0.001 using Binomial test) or always (n = 18; 52.9% [range: 0, 93.5%]; p < 0.001 using Binomial test) (S5 Fig). District hospitals with no surgical facilities onsite (two in Kenya, one in India) also had a higher proportion of cases in clusters (88.2% [range: 65.0, 93.5%]) than hospitals with surgical facilities available onsite (n = 24; 49.9% [range: 0, 89.4%], p < 0.001 using Binomial test).

Genomic characteristics of transmitted K. pneumoniae

A total of 172 distinct K. pneumoniae STs were identified amongst the included isolates (S5 Table). Of these, 57 STs were associated with at least one transmission cluster. We defined commonly transmitted STs as those that were identified in ≥3 transmission clusters in two or more sites. Fourteen STs met this definition (Fig 2), and these made up 63.8% (n = 972/1523) of all infections, including 71.8% (n = 743/1035) of clustered infections and 46.9% (n = 229/488) of unclustered infections. Twelve of the 14 commonly transmitted STs were detected in ≥3 countries each, however, ST152 and ST25 were each identified in just one country (South Africa and Malawi, respectively). Most of the 14 STs were associated with clusters at Malawi site 6B (see Fig 2B), which is not unexpected given this is the largest dataset and originates from seven years of routine blood culture surveillance; however, besides ST25, all common STs associated with clusters at this site were also associated with clusters at multiple other sites. The 14 commonly transmitted STs were responsible for most large clusters with ≥15 cases (n = 6/10 clusters; 60%; Fig 2D), although four large clusters were caused by rarely transmitting STs (ST6775, n = 29; ST37, n = 26; ST2004, n = 20; and ST1414, n = 17). Fig 3 shows, for each ST, the number of unique introductions (treating each cluster or singleton as a unique strain ‘introduction’ in a neonatal ward) versus the number of infection clusters. Across all STs, the mean fraction of introductions resulting in detected infection clusters was 15.3%, and this fraction was significantly higher among the 14 commonly transmitted STs compared to the other 158 STs (mean 33.1% versus 13.8%, Chi-squared test p < 0.001).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Commonly transmitted K. pneumoniae sequence types (STs).

(A) Transmission profile showing number of clusters, number of singleton cases, total number of sites the STs were detected, number of sites where the STs were part of transmission clusters, their adjusted odds ratios compared to other STs in a logistic regression model to assess association with transmission, and corresponding p-values (columns from left to right). P-values for STs with a significant association with transmission are formatted in bold (see Methods for details of the logistic regression analysis). (B) Site distribution of cluster isolates. Points are coloured to indicate geographical region and sized to indicate number of clustered isolates per site. (C) Resistance distribution (proportion) of cluster isolates. Resistance scores are as reported by Kleborate based on the presence or absence of extended-spectrum beta-lactamase genes (ESBL), carbapenemase genes (Carb), and/or colistin resistance determinants (Col). (D) ST composition of large clusters. Points are coloured to indicate the 14 commonly transmitted STs and sized to indicate the number of isolates, as per the figure legend.

https://doi.org/10.1371/journal.pmed.1005077.g002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Transmission probability of individual sequence types (STs).

(A) Distribution of the fraction of strain introductions resulting in clusters, for commonly transmitted STs (defined as STs that were identified in ≥3 transmission clusters in two or more sites) and other STs. The distributions are coloured to indicate commonly transmitted STs significantly associated with transmission in the logistic regression model. (B) Number of unique introductions vs. number of clusters, for individual STs. Points are sized to indicate the number of sites where clustered isolates were detected for each ST, and coloured to indicate commonly transmitted STs significantly associated with transmission in the logistic regression model. The dashed line shows y = x, along which the number of unique introductions is equal to the number of clusters. The solid line represents a linear regression line of best fit through the origin (slope 0.28, R² = 0.920). (C) Fraction of clusters and singletons with ESBL and/or carbapenemase. Points are plotted for the 57 STs involved in clusters, and coloured to indicate commonly transmitted STs significantly associated with transmission. ST, sequence type; ESBL/CP, extended-spectrum beta-lactamase or carbapenemase gene carriage (Kleborate resistance score ≥ 1).

https://doi.org/10.1371/journal.pmed.1005077.g003

The majority of isolates (n = 1384/1523, 90.9%) carried an ESBL or carbapenemase gene. Clusters were more likely to carry ESBL (94.2% of clusters (n = 147/156) versus 81.8% of singletons (n = 399/488), p < 0.001) or carbapenemase genes (19.9% of clusters (n = 31/156) versus 10.2% of singletons (n = 50/488), p = 0.0025) (S6 Table). In a logistic regression model for transmission, treating each cluster or singleton as a unique strain ‘introduction’ in a neonatal ward, and defining the outcome for each introduction as either evidence of transmission (i.e., cluster of size ≥2) or no evidence of transmission (i.e., a singleton unclustered case), transmission was significantly positively associated with carriage of ESBL genes (adjusted odds ratio, aOR = 2.84 [95% CI: 1.45, 5.57], p < 0.001) or carbapenemase genes (aOR = 1.81 [95% CI: 1.10, 2.98], p = 0.02) (S7 Table). Including Site as a covariate improved model fit (p = 0.05 using likelihood ratio test), but did not substantially alter the effect estimates for ESBLs (aOR = 2.48 [95% CI: 1.26, 4.90], p = 0.006) or carbapenemases (aOR = 2.08 [95% CI: 1.04, 4.14], p = 0.04).

It is challenging to tease apart the effect of AMR from the effect of lineage, as certain lineages may appear more likely to cause detectable infection clusters due to their resistance genes (if AMR is a driver of transmission and/or infection), or strains with resistance genes might appear more likely to cause detectable infection clusters due to their association with particular genetic lineages (if non-AMR-related lineage-associated differences drive transmission and/or infection). The 14 commonly transmitted STs showed a significantly higher rate of ESBL gene carriage (amongst singleton infections or deduplicated clusters) compared to the other 158 STs (94.2% (n = 308/327) versus 75.1% (n = 238/317), p < 0.0001), although there was no significant difference in carbapenemase gene carriage rates (14.1% (n = 46/327) versus 11% (n = 35/317), p = 0.30) (S8 Table). Amongst the 14 most common transmission STs, nearly all (n = 728/743; 98%) of the clustered isolates carried at least one ESBL or carbapenemase gene, compared with 93% (n = 213/229) amongst singleton isolates of these STs (Fig 3C). Adding ST (as a categorical variable with each of the 14 common STs as unique categories, compared to all other STs as the reference category) to the logistic regression model with ESBL, carbapenemase and Site did not improve the model (p = 0.226), and had little impact on the estimated effects of ESBL (aOR = 2.16 [95% CI: 1.08, 4.32], p = 0.03) and carbapenemase (aOR = 2.03 [95% CI: 1.01, 4.08], p = 0.05), and only two STs were significantly independently associated with transmission (ST17, aOR = 2.24 [95% CI: 1.11, 4.52], p = 0.03 and ST1741, aOR = 20.9 [95% CI: 1.2, 363.7], p = 0.01).

As K and O antigens have been proposed as potential vaccine targets, we explored their prevalence (based on genomic predictions) and association with transmission in our dataset. Six unique O serogroups were predicted (see S9 Table), the majority being O1 (n = 332/644, 51.6%) or O2 (n = 158/644, 24.5%), which were common amongst both clusters and singletons (51.3% versus 51.6%, respectively, for O1, 26.9% versus 23.8% for O2). Including O types in the transmission model with ESBL and carbapenemase genes, with or without ST and/or Site, did not improve model fit (p > 0.2 using likelihood ratio test), and no O types were significant in any models (see S7 Table).

Eighty distinct K loci were detected. Forty-three K loci were found in isolates associated with transmission clusters, and 39 of these were also identified in singleton isolates (S10 Table). Distinct K loci were associated with specific STs, making it difficult to tease apart the effect of K antigens from strain background. For example, KL102, which was the most common K locus associated with transmission, was closely associated with ST307: 95% (n = 241/253) of KL102 isolates from transmission clusters were ST307, and ST307 accounted for 83% (n = 15/18) of KL102 clusters detected. Similarly, 79% (n = 42/53) of KL25 cluster isolates belonged to either ST17 (n = 33, 62%) or ST607 (n = 9, 17%). Including the top 10 most common K loci in the transmission model together with ESBL and carbapenemase genes, with or without ST and/or Site, did not improve the model fit (p > 0.38 using likelihood ratio test), although KL62 was significantly positively associated with transmission (aOR 4.77 [95% CI: 1.30, 17.49], p = 0.025) in the model with ESBL and carbapenemase genes, ST, and Site. KL62 was associated with 12 clusters (n = 26 isolates) out of 34 introductions (n = 6 different STs) across 16 sites.