Introduction

Third- and fourth-generation cephalosporins are broad-spectrum β-lactam antimicrobial agents classified as critically important for human medicine by the World Health Organization due to their use in the management and treatment of serious infections caused by multidrug resistant (MDR) Gram-negative and Gram-positive bacteria1,2. Third-generation cephalosporins were introduced for medical use in the early 1980s, with fourth-generation cephalosporins introduced in 19943; since then, resistance to these drugs has increased4. Resistance to these extended-spectrum cephalosporins (ESCs) is widespread worldwide in Enterobacterales from both humans and animals, and is mediated predominantly by two main groups of inactivating enzymes: extended-spectrum beta-lactamases (ESBLs) and plasmid-mediated AmpC β-lactamases (AmpCs)5,6.

There are differences in the distribution of ESC resistance (ESC-R) genes between host species and geographical origins6. ESBLs were initially detected in Europe; in 1983, the first ESBL—a SHV variant—was identified in Germany in Klebsiella pneumoniae and Serratia marcescens isolates7. Until the end of the 1990s, most of the ESBLs detected were variants of SHV and TEM8. Since then, the main types in Europe have changed, with CTX-M β-lactamases becoming the dominant ESC-R enzymes in Enterobacterales in humans and farm animals9. For example, the ESBL gene blaCTX-M-1 is broadly disseminated among Escherichia coli from animals in Europe6. In human patients, the frequency of ESBL-producing E. coli and K. pneumoniae isolates was higher in Europe compared to North America between 2004 and 20068,10. In Canada, a national study reported the AmpC gene blaCMY-2 in E. coli isolates from patients of tertiary-care hospitals in 200011. Subsequently, this variant was also reported in E. coli isolates from intensive care units (ICUs) between 2005 and 200612. The first reports of blaCMY-2 in Enterobacterales from food-producing animals in Canada were between 1995 and 199913, and the gene has been commonly found in E. coli and Salmonella enterica from food-producing animals and food since then14,15,16. ESBLs in Canada were first observed in 2000 in E. coli and Klebsiella spp. from human clinical samples17. More recently, there have been reports of increasing occurrence of CTX-M-type β-lactamases in Enterobacterales in farm and companion animals from Canada18,19. In contrast, blaCMY-2 was rarely reported in E. coli from humans in Europe20,21, but blaCMY-2 isolates from livestock production have been reported6. A recent study from Germany found blaCMY-2 in isolates from humans, livestock and meat products22, and it has also been reported in E. coli from companion animals in France23. International travel and trade have likely influenced the global dissemination of ESC-R genes24. Comparison among ESBL and AmpC-producing E. coli isolates of human and animal origins from Germany, the Netherlands and the United Kingdom showed that human isolates were more closely related to each other than to isolates from animals25. In another recent study in The Netherlands, the majority of community-acquired ESBL and AmpC carriage among E. coli was found to be due to human-to-human transmission, with much less attributed to food, animal and environmental sources26.

ESBL and AmpC genes are typically located on mobile genetic elements27 and their spread has been linked to horizontal gene transfer via plasmids28,29,30. However, the expansion of bacterial clones carrying ESC-R genes also plays a role31. For example, the global spread of blaCTX-M-15 in humans is associated with the E. coli epidemic lineage sequence type (ST) 13131,32. Although numerous studies have been conducted examining the types and putative sources of ESC-R genes9,10,11,12,13,14,15,16,19,21,27, these mostly investigated isolates from single host species or geographical regions. To gain a broader understanding of the emergence and spread of resistance to these important drugs, diverse sources must be examined. In addition, the relative importance of horizontal gene transfer and clonal expansion in the spread of ESC-R between ecological niches and biological compartments is unclear but may shed light on where to target interventions to reduce the burden of antimicrobial resistance (AMR). In this study, we used genomic and evolutionary approaches with an extensive collection of Enterobacterales isolates to determine the dynamics and evolution of ESBL and AmpC over time in different ecological compartments (country and source) and identify potential factors contributing to the spread and persistence of ESC resistance. Thus we (i) identified similarities and differences in ESC-R genes between multiple Enterobacterales species, host species and geographical regions; (ii) analysed clonal lineages and putative plasmids able to spread across compartments; and (iii) examined the temporal trend of ESC-R genes in different compartments.

Results

Bacterial isolates and epidemiological data

A total of 1930 genomes were used in this study (Supplementary Table S1) for which extensive epidemiological metadata were available. These included E. coli, K. pneumoniae, S. enterica, Enterobacter hormaechei, Klebsiella oxytoca, Proteus mirabilis, Klebsiella michiganensis, Citrobacter europaeus, Citrobacter freundii and Raoultella ornithinolytica (Supplementary Table S2). Of the 1818 E. coli, 1617 were ESC-R isolates collected between 2008 and 2016, 79 ESC-R isolates before 2008 or after 2016 and 122 ESC-susceptible isolates mainly from 2008 to 2016. The set also included 112 non-E. coli, comprising 108 ESC-R genomes from 2008 to 2016, and four ESC-susceptible non-E. coli (Supplementary Figs. S1–S3).

The isolates derived from diverse sources, which were classified into eleven categories: human, cattle, chicken, food (porcine, bovine and avian meat), pig, turkey, dog, horse, wastewater (including wastewater from human activity as well as from animal and meat processing plants), cat and other avian sources (bird and duck) (Supplementary Table S3 and Source Data file). Overall, the majority of isolates were of human (n = 817, 42.3%), cattle (n = 372, 19.3%) and chicken (n = 263, 13.6%) origins (Supplementary Fig. S4). Assessment of the potential inclusion of outbreak-associated isolates, which may introduce bias into the analyses, identified only two ST10 and six ST177 genomes sharing ≤10 single nucleotide polymorphisms (SNPs) in their core genome. For ST131, there were only 36 genomes (mostly with blaCTX-M-15) sharing ≤10 SNPs, but these were from different years (from 2008 to 2016) and countries. Overall, this suggests outbreak isolates were not included in this study.

MLST and plasmid replicon identification

Assembly summary statistics are presented in the Source Data file. There were 285 known STs and 13 unknown STs identified among the 1818 E. coli isolates, the most common being ST131 at 19.7% (n = 359), ST10 at 6.9% (n = 125) and ST117 at 5.6% (n = 102) (Supplementary Fig. S5A). These three common STs were identified in all four countries (Canada, France, Germany and the UK) (Supplementary Fig. S5B) and multiple host species, with ST131 mainly associated with humans, ST10 mainly with cattle and ST117 mainly with poultry (Supplementary Fig. S5C–F). The distribution of isolates showed relatively even coverage of the main study time period between 2008 and 2016 for Canada, France and Germany. Relatively higher numbers of isolates were included in specific years for the UK (2015) and Germany (2012), but this was due to the sample selection in the previously published studies22,32,33 (Supplementary Figs. S6A, B, S7A–D). Regarding identification of putative plasmids, 60 replicons were identified in the E. coli isolates, representing 19 plasmid incompatibility (Inc) types (Supplementary Table S4, Supplementary Figs. S8, S9).

Genes conferring resistance to third generation cephalosporins

By study design, the majority of isolates selected for inclusion were phenotypically resistant to ESCs and so do not represent the prevalence of ESC-R in the general populations. Among the 1696 ESC-resistant E. coli isolates, 39 different ESC-R genes were identified. There were 55 combinations, or profiles, of ESC-R genes, some of which comprised more than one ESC-R gene. The most frequent types and profiles of ESC-R genes were: blaCTX-M-1 (n = 414) and blaCTX-M-15 (n = 418) with 24.4% and 24.6%, respectively, followed by blaCMY-2 19.2% (n = 326), blaCTX-M-14 10.8% (n = 183), blaSHV-12 8.3% (n = 140), blaCTX-M-27 3.5% (n = 60), blaCTX-M-55 1.2% (n = 20), blaCTX-M-2 0.9% (n = 16), and blaCMY-2 + blaSHV-12 0.7% (n = 12). The remaining 46 ESC-R profiles were rarely observed, each with 10 or fewer isolates. The non-E. coli isolates shared 11 ESC-R profiles with E. coli, while there were 44 ESC-R profiles unique to E. coli and 16 ESC-R profiles unique to the non-E. coli isolates (Supplementary Fig. S10A, B).

The major ESC-R genes in E. coli, blaCTX-M-1, blaCTX-M-15, blaCMY-2, blaCTX-M-14 and blaSHV-12, were globally distributed, though differences in frequency were found between countries, sources and compartments (Fig. 1a, b, Supplementary Fig. S11 and Supplementary Tables S5, S6). To note, this analysis did not include the isolates from Pietsch et al.22 (n = 148 ESC-R genomes) as the design of that study focused specifically on German blaCMY-2 isolates; after excluding these 148 genomes, this resulted in 1548 ESC-R E. coli genomes which were used for further comparative analysis. In this collection of 1548 ESC-R E. coli genomes, the blaCTX-M-1 gene was more frequent in European countries (France, Germany and the UK), while blaCTX-M-15 and blaCMY-2 were more common in North America (Canada) (Fisher’s Exact test, adjusted p-value < 0.0001; Fig. 1a). The blaCTX-M-1 gene was more frequent in isolates of animal (cattle, chicken, pig, turkey, dog, horse and cat) than of human origin, whereas blaCTX-M-15 was more frequent in human isolates (Fisher’s Exact test, adjusted p-value < 0.0001). The frequency of blaCMY-2 was similar between isolates from humans and animals (Fisher’s Exact test, adjusted p-value = 0.1) (Fig. 1b).

Fig. 1: Frequency and diversity of ESC-R genes in E. coli by country and source.
figure 1

a Percentage of ESC-R gene by country and b source (total n = 1,660 ESC-R + ESC-susceptible genomes*). The bars are coloured by ESC-R gene, as shown in the inset legend. c Diversity indices (SR: species richness; SE Shannon entropy, SD Simpson diversity, and reciprocal BP Berger-Parker) of ESC-R profile by compartment (country + source), where only compartments with minimum sample size of n = 30 were included in the analysis (total n = 1361 ESC-R genomes*). The shape of the dots is linked with the source and the colours with the country. d Common and unique ESC-R genes among countries and e compartments (total n = 1548 ESC-R genomes*). Heatmaps are coloured as per inset scale bars. *Genomes from Pietsch et al.22 (n = 158; 148 ESC-R genomes + 10 ESC-susceptible genomes) were excluded. Source data are available in the Source Data file. ESC-R: extended-spectrum cephalosporin resistance.

Full size image

Fifteen ESC-R profiles were observed at least in two countries (Fig. 1d). Six profiles were found only in Canada, in human (n = 5), chicken (n = 1) and cattle (n = 1) isolates. Three profiles were unique to UK isolates from human (n = 1), food (n = 1) and chicken (n = 10) sources (Fig. 1e). Twenty-three ESC-R profiles were unique to France, mainly in human isolates (Fig. 1e). Highly diverse ESC-R profiles were observed in isolates of French human origin (Fig. 1c), which was not the case for the ST and plasmid replicon profiles (Supplementary Fig. S12A, B).

Location of ESC-R genes in plasmid and chromosome contigs

Ascertainment of ESC-R gene locations in plasmids or chromosomes was obtained through MOB-suite34 and RFPlasmid35 (Supplementary Table S7). Initial use of MOB-suite alone identified an unexpectedly high proportion of ESC-R genes on chromosome-classified contigs, particularly for the blaCTX-M-15 and blaSHV-12 genes. Plasmid-classified contigs with ESC-R genes were further categorised into typeable plasmids (presence of replicon sequences in the same contig) and non-typeable plasmids (no identified replicon sequences in the same contig). In contrast, application of the RFPlasmid tool classified the majority of all ESC-R genes examined as located on plasmids, the majority of which were non-typeable plasmids. To resolve these disagreements and validate the location of ESC-R genes, we randomly selected 20 isolates—five each with the four most common ESC-R genes identified, blaCTX-M-1, blaCTX-M-15, blaCMY-2 and blaCTX-M-14—where prediction of the genomic location of the ESC-R gene differed between MOB-suite and RFPlasmid and performed long-read sequencing to obtain complete contiguated genome sequences. In all 20 long-read genome sequences, the RFPlasmid predictions were correct, suggesting MOB-Suite over-predicted the number of chromosome-located ESC-R genes. Using the RFPlasmid results (Supplementary Table S7), the majority of each of the five main ESC-R genes were located on plasmids, but to different degrees. The vast majority of blaCMY-2 genes were predicted to be plasmid-borne (95.7%), in contrast to blaCTX-M-14 (78.1%) and blaCTX-M-15 (79.9%). Overall, 313 typeable plasmids linked with ESC-R genes (hereafter ESC-R plasmids) were recovered from the short-read data (blaCTX-M-1 n = 136, blaCTX-M-15 n = 23, blaCMY-2 n = 88, blaCTX-M-14 n = 40 and blaSHV-12 n = 26), representing 27.6% of the predicted plasmids associated with ESC-R genes. Of these, IncI1 and IncF were the most frequently identified plasmid Inc types associated with ESC-R genes. blaCTX-M-1 and blaCMY-2 were predominantly associated with IncI1 plasmids, blaCTX-M-15 and blaSHV-12 with IncF plasmids, and blaCTX-M-14 with both IncI1 and IncF plasmids (Supplementary Table S8). Analysis of the entire genetic content (plasmidome) of the 313 typeable and a subset of 79 non-typeable ESC-R-carrying plasmid contigs showed clustering by ESC-R genes, plasmid replicon types and subtypes (Supplementary Fig. S13).

Inferring the mechanisms of transmission and fitness of main ESC-R genes in E. coli

Simpson’s diversity indices of STs and plasmid replicon profiles in the subset (n = 1343) of the genomes carrying at least one of the five major ESC-R genes were calculated and compared; we refer to this as the ‘main ESC-R gene’ dataset hereafter. This analysis was also performed on the subset of 313 typeable ESC-R plasmid contigs as a validation step (Table 1). The ST diversity indices for most ESC-R genes were similar, ranging from 17.94 to 33.90 in the main ESC-R gene dataset and from 10.46 to 23.79 in the typeable plasmid subset. The exception was for genomes with blaCTX-M-15, where the ST diversity was 3.24 in the main ESC-R gene dataset and 3.66 in the typeable plasmid subset. The plasmid replicon profile diversity for genomes carrying most ESC-R genes were also similar in the main ESC-R gene dataset, with indices ranging from 72.88 to 81.88; the exception was genomes with blaCMY-2, where the index was 145.71. In the typeable plasmid subset, the range of plasmid replicon profile diversity was lower, between 1.20 and 5.83. To infer the relative importance of clonal expansion versus putative plasmids for dissemination of ESC-R genes in E. coli, the ratio of plasmid-to-ST diversities for each major ESC-R gene was calculated. For blaCTX-M-1, blaCTX-M-14 and blaSHV-12 in the main ESC-R gene dataset, the ratio was slightly <3 (range 2.42–2.88); for blaCMY-2 the ratio was 8.12, and for blaCTX-M-15 the ratio was 22.49. In the typeable plasmid subset, the ratio was 0.05 and 0.10 for blaCTX-M-1 and blaCTX-M-14, respectively, 0.17 for blaCMY-2, 0.56 for blaSHV-12 and 1.59 for blaCTX-M-15. Plasmid replicon profile and ST Simpson’s diversity indices were also calculated for each major ESC-R gene by individual compartment in the full dataset (Supplementary Fig. S14).

Table 1 Characteristics and Simpson’s diversity (SD) of sequence types (STs) and plasmid replicon profiles of E. coli (n = 1,343) and E. coli with typeable ESC-R gene-carrying plasmids (n = 313) harbouring the five major ESC-R genes
Full size table

Trends over time of ESC-R genes by compartment in E. coli

In the two countries from which both human and animal data were available, the dynamics of AmpCs and ESBLs changed over the study period (Fig. 2a, b). In farm animals (mainly chickens) in Canada, the main ESC-R gene type shifted from AmpC to ESBL (Wald z-statistic, p < 0.0001). Conversely, in humans from France, the proportion of ESBLs declined over time while that of AmpCs increased (Wald z-statistic, p < 0.0001); nevertheless, ESBLs remained the most common gene type (Supplementary Table S9). There were no significant changes in AmpC and ESBL frequency over time in the other examined compartments, humans from Canada (Wald z-statistic, p = 0.2), nor in farm and companion animals from France (Wald z-statistic, p = 1.0) (Supplementary Table S9). When evaluated by individual ESC-R genes, contrasting trends within individual compartments were observed. For example, over the study period blaCMY-2 increased in human isolates from France, while this ESC-R gene decreased in chickens from Canada and was replaced by blaCTX-M-1 (Fig. 2c, Supplementary Table S9).

Fig. 2: Distribution and trend over time of ESC-R genes in E. coli by compartment.
figure 2

a Trends in percentage and b count for ESC-R types over time in Canada and France across humans and farm animals (cattle, chickens, pigs and turkeys). The lines and bars are coloured by ESC-R types. c Trends over time of the percentage of the major ESC-R gene in main compartments (colour key is in inset legend). Source data are available in the Source Data file. ESC-R extended-spectrum cephalosporin resistance, ESBL extended-spectrum β-lactamases.

Full size image

Resistance to other antimicrobial classes

In addition to the ESC-R genes, an additional 137 acquired AMR genes conferring resistance to other antimicrobial agents were identified, with up to 22 AMR genes and up to 12 plasmid replicons in a single genome (Supplementary Figs. S15, S16). There was variable presence of determinants (genes or SNPs) conferring resistance to different antimicrobial classes between compartments (Fig. 3a, Supplementary Tables S10, S11). Logistic regression analysis identified that determinants conferring resistance to aminoglycosides, sulfonamides, tetracycline, beta-lactams, diaminopyrimidines and quinolones-fluoroquinolones (due to mutations in gyrA, parC and parE; Supplementary Fig. S17 and Table S12) decreased over time in human isolates from France (Wald z-statistic, p-value < 0.00125), while in chickens from Canada the presence of genes conferring resistance to sulfonamides and tetracycline increased over time (Wald z-statistic, p-value < 0.00125) (Fig. 3b). Diversity indices showed highly diverse AMR profiles in isolates of French human origin and animals from France and Germany (Supplementary Fig. S18).

Fig. 3: Presence and dynamics of AMR determinants in E. coli.
figure 3

a Presence of resistance determinants by antimicrobial class and compartment (country + source) for n = 1189 E. coli isolates. Two-sided Fisher’s Exact Test was used to compare the proportion of AMR class between compartments, and the significance levels are indicated with asterisks: *p < 0.01, **p < 0.001 and ***p < 0.0001. The bars are coloured by country and the pattern by source. b Odds ratios and 95% confidence intervals (coloured as per inset legend) demonstrating change in AMR determinant presence over the study period by compartment; significance cut-off is 0.00125 after Bonferroni adjustment for multiple hypothesis testing. Source data are available in the Source Data file. Q-FQ quinolone-fluoroquinolone.

Full size image

Core-gene SNP-based phylogeny for E. coli

A maximum likelihood core-gene tree was constructed for 1818 E. coli genomes (n = 1696 ESC-R and n = 122 ESC-susceptible). The length of the core-gene alignment was 2300,993 bp (with 365,784 SNPs), which represents 45% of the reference genome EC958 (Refseq: NZ_HG941718.1). In the phylogenetic tree, there were eight major clusters largely concordant with phylogroup identification36 (Fig. 4). Within phylogroup A, CC/ST10 was predominant, in phylogroup G the main ST was ST117 and in phylogroup B2 it was CC/ST131 (Fig. 4). These results were consistent with the pangenome network analysis, where clusters largely matched the population structure of the core-gene tree (Supplementary Fig. S19). There were some exceptions, particularly with phylogroup D genomes, that were phylogenetically clustered in the core-gene tree but fell into distinct clusters in the gene content network (Supplementary Fig. S19A).

Fig. 4: Core-gene phylogenetic tree for 1818 E. coli isolates.
figure 4

Maximum likelihood tree with eight main clusters linked with phylogroup. Bootstrap support values > 95 are represented by blue dots for main clusters. Metadata are represented next to the tree by four rings: phylogroup, ESC-R gene, source and country (coloured as per inset legend). The number of AMR genes per genome is plotted on the outside as a bar plot (light green colour).*ST117 cluster contains mainly ST117 but also other closely related STs. Source data are available in the Source Data file. ESC-R extended-spectrum cephalosporin resistance.

Full size image

Genomes from different countries and sources were distributed across the different phylogroups, with ESC-susceptible genomes placed in diverse clusters. Previously identified host associations were observed, such as ST131 with humans and ST117 with chickens. Similarly, the ESC-R genes were widely distributed in the different clusters, with the exception of the previously identified association of blaCTX-M-15 with ST131. Separate phylogenetic trees were constructed for genomes carrying blaCTX-M-1, blaCTX-M-15, blaCMY-2, blaCTX-M-14, blaSHV-12, blaCTX-M-27, blaCTX-M-55 or blaCTX-M-2 (Supplementary Figs. S20–S22). Overall, the main clusters displayed in each tree correlated with phylogroups.

The 360 CC/ST131 genomes were classified in five clades based on marker alleles for the type 1 fimbriae fimH (H) adhesin: clade A (H41, n = 13) with half of the isolates carrying blaCTX-M-14, the remaining carrying other ESC-R genes; clade B (H22, n = 47; H30, n = 2; H54, n = 1) mostly carrying blaCMY-2; clade B0 (H27, n = 1) with a single blaSHV-12 isolate; clade C1, also known as H30-R (H30, n = 70), with more than half of isolates carrying blaCTX-M-27; and clade C2, also known as H30-Rx (H30, n = 220; H35, n = 6), mostly harbouring blaCTX-M-15. Similar results were previously reported22,32,37,38,39,40,41. All genomes in Clades C1 and C2 also had double mutations in the gyrA and parC genes, which are associated with high-level resistance to fluoroquinolones42.

Discussion

Resistance to ESC drugs is a global public health concern due to its widespread nature and the use of these drugs for difficult to treat infections. Detailed within-country investigations have been conducted, which have identified patterns of transmission between host species. For example, in The Netherlands ESBL/AmpC gene profile distribution from livestock and food-associated E. coli were different from those in E. coli from humans43, and a further study revealed humans as the main source of community-acquired ESBL and AmpC carriage in E. coli from humans based on distinguishable patterns of ESBL and AmpC gene occurrence in different sources26. Similarly, in the UK, whole genome sequencing (WGS) demonstrated distinct populations of E. coli, including those with ESBL and AmpC genes, in livestock and retail meat compared with those isolated from bloodstream infections in humans33. However, given the global nature of food supplies, animal movement and human travel, examining patterns from multiple countries and sources can provide additional insight and comparison of patterns and transmission of AMR. In particular, the use of WGS applied in this context provides valuable information for monitoring trends of specific ESBL and AmpC genes, as well as other AMR genotypes and potential linkage to particular plasmids, as was recently demonstrated in a collection of pig isolates from the UK44 and in calf isolates from France45. Here, a genomic and evolutionary approach was used to determine the distribution and dynamics over time of ESBL and AmpC in E. coli from different geographical origins and sources and to infer the mechanisms underlying the transmission of the major ESC-R genes.

In comparing Canada and France, countries for which both human and animal data were available, two trends are clear. First, in farm animals in Canada (represented predominantly by chickens in this dataset), the majority of isolates carried AmpC genes at the beginning of the study period, but by the end, the majority of ESC-R was due to ESBL genes. In contrast, the main determinant of ESC-R in French farm animals (represented predominantly by cattle in this dataset) consistently was ESBL genes. Second, in humans from Canada both AmpC and ESBL genes were found over the entire study period, with ESBL genes consistently found at a higher prevalence. In humans from France, ESC-R was initially dominated by ESBL genes, and then part way through the study period the AmpC gene blaCMY-2 was identified and increased over time, although ESBLs remained as the main resistance mechanism. These observations agree with previous work describing the emergence and increase of ESBLs in Europe8,9,46, where blaCTX-M variants have become major determinants in humans and animals in different enteric bacterial species9, with later emergence of blaCMY-222. The results presented are also consistent with previous work from North America, where the AmpC gene blaCMY-2 was identified in isolates from humans, animals and foods in the early to mid 2000s12,14, while more recent studies have detected the presence of blaCTX-M-type variants in animals from Canada19.

The major ESC-R genes in E. coli identified in this study were blaCTX-M-1, blaCTX-M-15, blaCMY-2, blaCTX-M-14 and blaSHV-12. These genes are also found in other enteric bacterial species, like those included in this study, and across the globe6. Despite their ubiquitous nature, there was variation in their distribution and frequency between geographical origins and sources. In these data, blaCTX-M-1 was more frequent in Europe compared to North America, where blaCMY-2 and blaCTX-M-15 were more common. With respect to source (host species), blaCTX-M-1 was more frequent in animals, blaCTX-M-15 in humans, and blaCMY-2 was frequent in both humans and animals. The link between specific sources and these ESC-R genes has been previously reported6,22,32,33. ESC-susceptible genomes were distributed throughout different clusters and compartments (country + source), thereby indicating that, in general and with the exceptions mentioned, ESC-R genes were not associated with particular chromosomal backgrounds or ecological compartments.

The first report of blaCTX-M-15 identified its location on large plasmids isolated from E. coli, K. pneumoniae and E. aerogenes isolates from urinary and pulmonary specimens of patients hospitalized in New Delhi, India, in 199947. From August 2000 to April 2002 in Canada, a major outbreak of multidrug-resistant and ESBL-producing E. coli isolates was identified in long-term care facilities48. A following study identified a plasmid with the blaCTX-M-15 gene associated with the outbreak strain, and this plasmid was genetically very similar to the one isolated in India49. In the UK, blaCTX-M-15 isolates were initially isolated from clinical samples in 200150; subsequently, the spread of blaCTX-M-15 was associated with the expansion of the human-associated clone ST13132. In this study, a higher proportion of blaCTX-M-15 was identified in Canada and the UK compared to France, related to the higher proportion of ST131 in these countries. In France, more diverse STs were associated with blaCTX-M-15, suggesting plasmids—particularly IncF plasmids—may play a role in the dissemination of this gene in that setting. Although ST131 and, therefore, blaCTX-M-15 was strongly associated with human-derived isolates, it was also observed in isolates of dog, pig and wastewater origin. Transmission between humans and pets has been observed previously6,51. In addition to the blaCTX-M-15/ST131/human association, blaCTX-M-27 was also observed to occur mainly in ST131 clade C1 human isolates, again with multiple IncF replicons. In 2006, ST131 C1/H30R with blaCTX-M-27 was detected in Japan52, followed by a reported rise of blaCTX-M-2746. The expansion of the ST131 C1/H30R clade was found to be responsible for the increase of this ESBL gene40, and was subsequently observed in France53 and Germany54. Moreover, our study also found blaCMY-2 predominantly in ST131 clade B in humans and chickens from Canada, an association previously reported elsewhere22,41.

The AmpC gene blaCMY-2 has been reported frequently in North America and to be increasing in Europe22. Here we confirm those results in our temporal analysis of genomes from France, from which both human- and animal-derived genomes are available. We also revealed that this significant increase was observed only among isolates from humans and not from farm animals (here, predominantly cattle). We observed similar frequencies between humans from Canada and France, but temporal trend analysis showed the frequency increased over time in French human isolates. In contrast to the relatively constant presence of the gene in Canadian human-derived genomes, blaCMY-2 decreased over time in Canadian chickens, and was replaced by blaCTX-M-1. IncA/C plasmids have previously been associated with blaCMY-2 in isolates from cattle faeces and wastewater in Canada55; in this study, IncA/C2, IncB/O/K/Z and IncI1 replicons were observed frequently in blaCMY-2-carrying E. coli from chickens in Canada. In Germany, the spread of blaCMY-2 was attributed to transmission of IncI1 and IncK plasmids in isolates of human origin22. Here we observed that human French blaCMY-2-carrying isolates frequently harboured IncI1 replicons, which is consistent with a previous French study of isolates from companion animals23.

Recovery of fully contiguated plasmid sequences from short-read data is very challenging due to the frequent occurrence of repeat elements that impede assembly of complete plasmid or chromosome sequences56. Despite this limitation, diverse bioinformatic tools are available for plasmid detection, classification and reconstruction34,57,58, with a recent study identifying MOB-suite as the only tool to reconstruct correctly the majority of E. coli plasmids in a test set56. Using MOB-suite, we were able to recover typeable ESC-R plasmids, but it over-represented the proportion of ESC-R genes carried on the chromosome. This was particularly true for blaCTX-M-15 and blaSHV-12 genes, which may be because these genes are associated with IncF plasmids. IncF plasmids are highly diverse with multiple copies of insertion sequences (ISs)59, which will limit successful plasmid assembly. The chromosome overrepresentation was resolved with RFPlasmid35, which identified a much larger proportion of ESC-R genes carried on plasmids and was later confirmed by long-read sequence data of a subset of isolates. Through this plasmid prediction approach, we were able to infer the proportions of ESC-R genes located on plasmids and the chromosome: blaCMY-2, blaSHV-12 and blaCTX-M-1 were frequently located on plasmids with 95.7%, 93.6% and 85.4%, respectively, while the proportion of plasmid-borne blaCTX-M-14 and blaCTX-M-15 was 78.1% and 79.9%, respectively. A total of 313 typeable ESC-R plasmids were recovered by MOB-suite and RFPlasmid, representing 27.6% of all predicted plasmid contigs associated with ESC-R genes. Although relatively low, this percentage is comparable to that obtained in a recent study56; the remaining plasmid contigs were not typeable as they did not contain replicon sequences.

As part of this study, the predominant mode of transmission of each of the five main ESC-R genes was assessed using the diversities of bacterial backgrounds (STs), plasmid replicons and their ratio in two datasets: (1) in the main ESC-R gene dataset of 1343 E. coli ESC-R genomes carrying at least one of the five main ESC-R genes, and (2) in the subset of 313 genomes containing typeable ESC-R plasmids (Table 1). From the analysis of the main ESC-R gene dataset, two main trends were apparent. First, in four of the five ESC-R genes, ST diversity was similar; the exception was blaCTX-M-15, which had lower ST diversity. This was to be expected, given the predominance of ST131 representing the chromosomal backgrounds of isolates with the blaCTX-M-15 gene. In this study, 55% of genomes with blaCTX-M-15 were ST131; for the other ESC-R genes, the most common ST (either ST10, ST117 or ST131) represented only 9–17% of genomes. Second, the diversity of plasmid profiles, representing unique combinations of replicons, was similar for four out of the five main ESC-R genes. In this case, the exception was blaCMY-2, with a Simpson’s diversity index of 145.7 compared to the 72.9–81.9 observed with the other ESC-R genes. However, when the diversity of plasmid replicons in genomes carrying blaCMY-2 was assessed in individual compartments (Supplementary Fig. S14), the diversities within each compartment appeared to be similar to those for other ESC-R genes. The reason for this apparent contradiction is that the Simpson’s diversity index evaluates the number and abundance of unique combinations of plasmid replicons; it does not assess whether or not those plasmid replicons are the same. Here, each compartment had approximately the same diversity of plasmid replicon profiles with the blaCMY-2 gene, but these profiles were different between compartments, explaining the higher diversity when all compartments were considered together, as in Table 1. Given the difficulty in reconstructing full plasmid sequences with short-read genome data and, therefore, definitively linking ESC-R genes with plasmid types, we repeated this analysis on a subset of 313 genomes containing typeable plasmids, where the plasmid contig contained both the ESC-R gene and a replicon sequence. Similar to the main ESC-R gene dataset of 1343 genomes, this approach identified blaCTX-M-15 with the lowest ST diversity and highest ratio of plasmid replicon to ST diversity. In contrast to the main ESC-R gene dataset where blaCMY-2 genomes had the highest plasmid replicon diversity, in the typeable plasmid subset blaCTX-M-15 and blaSHV-12 had the highest plasmid replicon diversity. This discrepancy is likely because of the differential recovery of typeable plasmids; due to the multiple ISs in IncF plasmids, IncF plasmids represented a lower proportion of plasmids in the typeable plasmid subset (26.8%) than in the main ESC-R gene dataset (36.7%) (Supplementary Table S13). As IncF plasmids are highly diverse60, we used the replicon sequence typing (RST) scheme for IncF plasmids61 in these analyses to subdivide them beyond IncF. Therefore, the lower proportion of IncF plasmids in the typeable plasmid subset would have an even greater effect on plasmid replicon diversity compared to the main ESC-R gene dataset. Conversely, IncI1 plasmids were overrepresented in the typeable plasmid subset (62.9%) vs the main ESC-R gene dataset (15.3%). Furthermore, as specific ESC-R genes have associations with specific plasmid Inc types, such as blaCTX-M-15 with IncF (Supplementary Table S8), the typeable plasmid subset is not representative of the main ESC-R gene dataset.

The comparative use of ST and plasmid replicon diversities for different ESC-R genes can shed light into their evolutionary history and epidemiology. Using blaCMY-2 from the main ESC-R gene dataset as an example, the ST diversity was much higher in genomes derived from French humans, where this ESC-R emerged more recently than in Canada and, therefore, may not have yet undergone the same process of purifying selection. The ratios of ST to plasmid diversity for each ESC-R gene also identified differences in how these genes spread. A higher ratio could be due to two factors: (1) the ST diversity is relatively low, as was the case for blaCTX-M-15 in both the main ESC-R gene dataset and typeable plasmid subset, reflecting clonal expansion of lineages with blaCTX-M-15 carried on either plasmids or the chromosome as the predominant mechanism of spread, or (2) the plasmid replicon diversity is relatively high, as was the case for blaCMY-2 in the main ESC-R gene dataset, suggesting the gene is highly mobile between plasmids. While acknowledging the typeable plasmid subset is not representative of the main ESC-R gene dataset, examining those plasmid contigs with both ESC-R genes and plasmid replicons provides further insight into the mechanisms of ESC-R spread. This revealed dominant plasmid types associated with blaCTX-M-1 and blaCTX-M-14—IncI1 plasmid sequence type (pST) 3 plasmids with blaCTX-M-1 and IncF (subtype F2:A-:B-) plasmids with blaCTX-M-14—suggesting the spread of these genes is mainly due to dominant plasmids across different genetic backgrounds. The higher ST:plasmid diversity ratio of blaSHV-12 in the typeable plasmid subset is a result of relatively low ST and high plasmid diversity, which suggest that both ST clonal expansion (here, mainly ST117) and plasmids contribute to its spread. In the 26 genomes with blaSHV-12 typeable plasmids, ST117 comprised 30.8% of the STs harbouring blaSHV-12 IncF plasmids, while other STs carried IncF and IncI1 plasmids. When comparing the plasmid subtypes between E. coli and non-E. coli, we found three common plasmid subtypes (IncN pST1 blaCTX-M-1, IncI1 pST12 blaCMY-2 and IncI1 pST26 blaSHV-12), which suggest that some plasmid subtypes are crossing species borders and mobilising the ESC-R genes between bacterial species (Supplementary Table S14). This result supports previous observations where plasmid replicons62 and ESC-R genes were detected in diverse Enterobacterales species14,18.

It is important to consider the limitations of these data when interpreting the results. The bacterial collections from which isolates were selected for sequencing were generated for a variety of purposes, including surveillance and research projects, and as such utilised different inclusion criteria and methods as outlined in the Supplementary Material. The human isolates from France and Canada and the animal isolates from Germany and France were from cases of clinical disease, whereas animal isolates from Canada were from clinical cases and asymptomatic carriage. To reduce potential bias, isolates collected over the same time period were randomly selected from the different collections proportionately by year and from a diverse assortment of host species. Importantly, the identification of ESC-R was based on phenotypic testing, thus allowing comparison of the identity and diversity of ESC-R genes between compartments and was thus the focus of this study. To ensure the most robust inference from this dataset, we also restricted statistical comparisons to those ecological compartments with good representation, such as farm animals and humans from Canada and France when analysing the temporal trends in ESC-R genes. For the same reason, in the diversity analyses we took multiple random subsets of ecological compartments with larger sample sizes to comparable sizes of ecological compartments with smaller sample sizes. One striking result was the comparatively high diversity of ESC-R genes in genomes derived from French humans; these genomes were collected over the entire study period and from various parts of the country. Although the sampling strategy cannot be ruled out as a causative factor in the results, the underlying causes for this greater diversity of ESC-R in E. coli from humans in France warrants further investigation. Through the plasmidome analysis, we showed that the included 79 non-typeable plasmids were closely related to the typeable plasmids, demonstrating their likely Inc type but also the limitations of assembly with short-read data. Thus, large scale long-read sequence data are necessary to identify reliably the location of ESC-R genes to the chromosome or to specific plasmids, as we demonstrated with the 20 genomes for which we used long-read sequence data to validate the ESC-R gene location. With such data, a greater resolution of the plasmidome network and transmission can be obtained.

The differing trends and types of ESC-R in the countries examined could be due in part to different amounts of antimicrobial use (AMU). In Europe, the use of antimicrobials agents as growth promoters in animals has been stopped since 2006, and overall sales of veterinary antimicrobial agents have decreased, including in France63, Germany64 and the UK64. In comparison, Canada stopped the use of antimicrobials for growth promotion in 2016 and had higher sales for use in animals, with the predominant antimicrobial agents sold being tetracyclines, beta-lactams, macrolides and diaminopyrimidines-sulfonamides65. This contrast is perhaps reflected in the results of this study identifying an increase in resistance to sulfonamides and tetracyclines over the study period in chicken isolates from Canada. A recent study on AMU and AMR in broiler chickens in Canada during 2013-2019 showed that resistance to some antimicrobial classes decreased due to an overall reduction in AMU, but other antimicrobial classes, such as aminoglycosides, increased; this could be associated with the increased use of aminoglycosides in water to treat disease66. There can be differences in AMU between humans and animals by antimicrobial class, and between different species of animals63,67, which may also influence the prevalence of resistance to specific antimicrobials. However, there are limited data available on antimicrobial consumption and there are other factors, such as genetic linkage, which make understanding the causal relationships of use and resistance complex. Combining the high-resolution data conferred by WGS to understand genetic linkage of AMR genes along with greater available of AMU data will facilitate greater understanding of the relative contributions of different drivers of AMR.

In conclusion, this study confirmed the changes over time of the relative importance of ESBL and AmpC genes causing resistance to ESC drugs in Europe and North America. There was variation in the major ESC-R gene frequency between compartments (country + source), and in the main mechanism, clonality and/or plasmids, responsible for the dissemination of these genes. In addition, the co-occurrence of ESC-R genes with other AMR determinants was higher in some compartments in different countries. The changing patterns of antimicrobial use may be associated with the changing trends of resistance, but more data are required; given the demonstrated differences between compartments, the collection/addition of AMU data to WGS in future studies will enable greater understanding of this link. Utilising WGS to obtain high-resolution data not only to identify the type and diversity of ESC-R genes present, but also linking this information to bacterial host and plasmid diversity, offers deeper insight into the epidemiology of resistance to these critical drugs. Further in-depth study of these plasmids and the role of other mobile genetic elements, such as transposons, in the transmission and global epidemiology of ESC-R is required, and will be facilitated by the increasing feasibility of large scale long-read genome sequencing to generate fully contiguated genome sequences. In line with the conclusions of other studies26, these results highlight the need for long-term internationally harmonised monitoring in order to understand the dynamics and relative importance of different sources to the burden of AMR over time.

Methods

Bacterial isolates

All ESC-R isolates included in this study were selected independent of the genetic mechanism causing the resistance. The majority (84%) of isolates in this study were selected to include representatives from different ecological compartments (human, animal, food/meat) of ESC-R E. coli isolated from 2008 to 2016. These isolates were randomly selected from available culture collections across three countries: France, Germany and Canada. The remaining 16% of isolates were selected to provide context, namely: (1) ESC-R E. coli isolates obtained prior to 2008 or after 2016 (4%); (2) ESC-R isolates of bacterial species other than E. coli from 2008 to 2016 (5.6%); and (3) ESC-susceptible E. coli (6.5%). With this strategy, 1524 isolates were selected: 718 isolates from Canada, 607 from France and 199 from Germany (Table S1). Bacterial collection and isolation methods for each subset are described in the Supplemental methods, section 1.

In addition, genomes from previous studies were included. This included 158 from Germany22 and 248 from the UK32,33; these genomes were selected following a similar approach of sample selection as described above, where all genomes available between 2008 and 2016 were included and a small number of genomes collected prior to 2008 and/or genomes of ESC-susceptible isolates were selected.

Whole genome data, MLST and typing of fimH allele and phylogroups

Paired-end short-read whole genome sequence data were generated on the Illumina platform for the 1524 isolates selected above, and fastq data available from the previous studies for the German (study accession PRJEB23663)22 and UK genomes (study accessions PRJEB4681, PRJEB8774 and PRJEB8776)32,33 were downloaded from the European Nucleotide Archive (http://www.ebi.ac.uk/ena). The methods for DNA extraction and WGS are detailed for each subset in the Supplemental method, section 1. Raw reads were processed with Trimmomatic v0.3368 to remove adapters and poor quality bases. The trimmed reads were used to assemble draft genomes using Spades v3.11.169; summary statistics of the assemblies were obtained using Quast v4.570, and assemblies with a high number of contigs (>900 contigs) were excluded. Genome quality was also assessed using CheckM v1.1.271 where genomes with completeness below 95.0% and/or contamination in the assembly with more than 15.0% were excluded. Centrifuge 1.0.3-beta72 and Kleborate v2.0.473 were used for the species assignment. In cases where species assignment differed between Centrifuge and Kleborate, FastANI v1.3274 was applied and compared with NCBI genome references. BWA v0.7.1775, Samtools v1.976 and Bcftools v1.877 were used to map reads against their corresponding assemblies for the largest four contigs to estimate sequencing mean chromosomal depth coverage.

The Achtman sequence typing scheme78 was applied using ARIBA v2.13.579 to assign the multilocus sequence type (MLST) for each genome. The sequence type (ST) was allocated to clonal complexes (CC) using the E. coli database available from the Enterobase website80. The fimH alleles for E. coli were downloaded from the Center for Genomic Epidemiology (http://www.genomicepidemiology.org/) and used with ARIBA for typing the alleles. Phylotyping for E. coli was performed using the ClermonTyping v1.4.136 tool.

Assessment of potential inclusion of outbreak isolates

To determine whether or not highly similar outbreak isolates were included in the dataset, which could affect the proportion of specific ESC-R genes and temporal analysis, pairwise SNP distances were analysed within the core genome for each main ST: ST131, ST10 and ST117. To obtain the core genome, reads were mapped to the relevant reference genome using Snippy v4.3.681. The reference genomes were: EC958 (NZ_HG941718.1; https://www.ncbi.nlm.nih.gov/nuccore/NZ_HG941718.1) for ST131, K-12 MG1655 (NC_000913.3; https://www.ncbi.nlm.nih.gov/nuccore/NC_000913.3) for ST10 and MDR_56 (NZ_CP019903.1; https://www.ncbi.nlm.nih.gov/nuccore/NZ_CP019903.1) for ST117. Pairwise SNP differences were obtained using snp-dists v0.7.0 (https://github.com/tseemann/snp-dists). High similarity between pairs of isolates was considered to be differences ≤10 SNPs.

In silico identification of AMR determinants and plasmid replicons

ARIBA v2.13.579 was used to identify acquired AMR genes and plasmid replicons with the ResFinder82 and PlasmidFinder61 databases, respectively. ESBL and AmpC genes were identified from these databases. Mutations in gyrase genes (gyrA and gyrB) and topoisomerase IV genes (parC and parE) associated with quinolone resistance were also identified using ARIBA v2.13.579 with the wild-type sequences as the input databases. The frequency and distribution of resistance determinants to ESC and other antimicrobial classes were analysed by country, by source (i.e., host species) and by compartment, which is defined as the unique combination of country and source. A similar approach was used for plasmid replicons and mutations in topoisomerase genes. The numbers of acquired AMR genes and plasmid replicons for each sample were compiled and compared by analysing the distribution of their frequencies among groups (country + ESC-R genes) for each source.

Prediction of ESC-R gene location to plasmid or chromosome contig in de novo assembly

The prediction of contigs as plasmid or chromosomal origin was first achieved using MOB-suite v2.034 with the mob_recon option with default setting. In addition, RFPlasmid35 was also used; a cut-off of ≥0.6 was used to identify the plasmid contigs. Resfinder82 and Plasmidfinder61 in Abricate v0.9.8 (https://github.com/tseemann/abricate) were used to identify ESC-R genes and plasmid replicon sequences in contigs. The plasmid subtyping was done using Blast+ v2.9.083 and the plasmid MLST (pMLST) scheme (https://bitbucket.org/genomicepidemiology/pmlst_db/). Where the prediction of chromosome vs plasmid origin differed between MOB-suite and RFPlasmid, a subset of five genomes for the four most common ESC-R genes, for a total of 20 isolates, was randomly selected for long-read sequencing (see Supplemental Method, section 2, for further information on long-read data), to confirm the localization of the ESC-R gene either to a plasmid or the chromosome. The plasmidome analysis using PANINI84 was performed in the set of 313 typeable plasmids plus a subset of 79 non-typeable plasmids (identified with MOB-suite) carrying ESC-R genes recovered from short-read data (See Supplemental Method, Section 3).

Diversity analysis of ST, ESC-R, AMR and plasmid replicon profiles

The diversities of STs, ESC-R genes, AMR profiles and plasmid replicon profiles were estimated as previously described for Salmonella Typhimurium85 using the vegan v2.5-686 R package. Four diversity measures were calculated, covering a range of weightings of richness and abundance: SR (species richness), SE (Shannon entropy), SD (Simpson’s diversity) and reciprocal BP (Berger-Parker)85,87. The ESC-R profile was defined as the combination of unique ESC-R genes within a single genome. The AMR profile comprised the presence of all AMR determinants, including ESC-R genes and mutations in the topoisomerase genes. The diversity analysis was performed by compartment (country + source). As sample size varied, compartments with larger sample sizes were randomly subsampled to the size of the smaller compartment. The subsampling consisted of 10,000 permutations without replacement and the mean with 95% confidence intervals of diversity indices were calculated. The calculated values for SE, SD and BP were converted into the effective number of profiles prior to plotting the results85.

In addition, Simpson’s diversity (SD) for ST and plasmid replicon profiles were further investigated in two datasets: (1) in the main ESC-R gene dataset of 1343 E. coli genomes with the five main ESC-R genes, where we investigated diversity (i) within genomes carrying each major ESC-R genes, and (ii) between main compartments within each major ESC-R gene; (2) in the subset of 313 typeable ESC-R plasmids, where SD assessment of plasmid replicon profiles and their respective bacterial ST across ESC-R genes were performed; in this analysis, contigs predicted to be plasmids (identified through both MOB-suite and RFPlasmid) that were carrying both a main ESC-R gene and a replicon sequence were used. The diversity estimates were computed through the sample-size-based rarefaction and extrapolation method implemented in the iNEXT v2.0.2088 R89 package.

Phylogenetic analysis

The draft assemblies were annotated using Prokka v1.1390 and the pangenome was analysed with Roary v3.12.091; core genes were identified with 90% sequence similarity and presence in 95% of isolates. SNPs were extracted from the core gene alignment using snp-sites v2.5.192 and used to construct a maximum likelihood tree using the HKY and discrete GAMMA nucleotide model including constant sites information in IQ-tree v1.6.1193. A phylogenetic tree was constructed for the entire E. coli (n = 1839) genome collection; due to the large sample size, the ultrafast bootstrap approximation94 -B 1000 was used. Separate phylogenies were also constructed for genomes containing each main ESC-R gene (blaCTX-M-1, blaCTX-M-15, blaCMY-2, blaCTX-M-14 and blaSHV-12) using subsets of the same core gene alignment. For the separate phylogenies, the standard nonparametric bootstrap95 -b 100 was applied. Metadata were plotted against the trees using the ggtree v2.0.296 R89 package. The pangenome network analysis was also performed using PANINI84; details are provided in the Supplemental Method, Section 4.

Statistical analysis

Fisher’s Exact Test from the rstatix’s v0.5.097 R89 package was used to compare the proportions of the major ESC-R genes between host species (human vs animal) and continents (Europe vs North America), and similarly, between main compartments within each major ESC-R gene and for overall AMR class. A Bonferroni correction was applied to account for multiple hypothesis testing.

Kruskal-Wallis tests, also available in rstatix’s v0.5.097, were used to compare the numbers of acquired AMR genes between groups constituted by country and ESC-R gene; groups with similar (statistically not significant) distributions were combined. The medians of the number of acquired AMR genes between the groups were compared through Mann–Whitney U tests.

To test for temporal trends in antimicrobial resistance, a logistic regression was applied, where the binary outcome variable was the presence or absence of a specific AMR determinant and the predictor variable was the year of isolation. To note, this analysis did not include a number of genomes from humans in France and Canada: 20 genomes that had both AmpC and ESBL genes and four AmpC-carrying genomes, as they were additional samples that were not originally selected with the same study selection strategy. Logistic regression analysis was done individually by compartment (country + source) for each ESC-R gene type (all AmpC genes, including blaCMY variants and blaDHA-1; and all ESBL genes, including variants of blaCTX-M, blaTEM, blaSHV and blaVEB-1), main ESC-R genes and other antimicrobial classes (aminoglycosides, sulfonamides, tetracyclines, other beta-lactams [non-ESBL and non-AmpC genes], diaminopyrimidines, macrolides, phenicols and quinolones/fluoroquinolones due to mutations in gyrA, parC and parE). The glm(y ~ x, family = “binomial”) function available in R89 was used to obtain the estimates, p-values based on z-scores (Wald statistic), odds ratio (OR) along with the 95% confidence intervals. A Bonferroni correction was applied to account for multiple hypothesis testing, as described above.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.