Clinical characteristics

There were 3394 E. coli isolated from 264,692 blood culture and CSF tests over the period from 2000 to 2021. The number of cases of E. coli per year for all ages ranged from 88 to 233, for neonates ranged from 4 to 41 (Fig. 1a). The number of blood culture or CSF samples taken per year for all ages ranged from 2796 to 26230 with an average of 11029 in a year. The number of blood culture or CSF samples taken per year for neonates ranged from 77 to 3211 with an average of 1228 per year. The positivity rate per 1000 blood culture or CSF samples for all age groups was highest in 2006 and lowest in 2013, with an average positivity rate of 14.1/1000 samples/year. For neonates, it was highest in 2004 and lowest in 2015, with a similar average positivity rate of 13.7/1000 samples (Fig. 1b).

a Numbers of E. coli cases per year at QECH. Bars represent the crude frequency of E. coli infection for each year from 2000 to 2021, with the different colours representing the different age groups of the patients. b Blood culture and CSF positivity rate (per 1000 blood culture or CSF samples) of E. coli in neonates and the entire patient population (including neonates). c Age range of neonates in the current study, colours highlighting early (<72 h of life) or late (>72 h of life) onset infection.

We identified 207 E. coli isolated from 203 individual samples (four had two separate E. coli morphologies identified) in the period from September 2012 to March 2021 (Fig. 2c); 95/200 (47.5%) were female, with a median age of 3 [IQR 2–8] days (Fig. 1c). EOS accounted for 111/203 (54.7%) of cases, with late onset sepsis accounting for the rest (Fig. 1c). Of these isolates 163/201 (81.1%) were cultured from blood and 38/201 (18.9%) from CSF. There were 110/200 (55%) cases from the Chatinkha nursery, 62/200 (31%) from paediatric nursery with the rest coming from other wards (Supplementary Fig. 1). Of these 207 isolates, 169 could be regrown and passed all QC steps, and were included in our analyses (Fig. 2a).

a Number of E. coli isolates selected for WGS for this current study. 2012 and 2021 were years for which isolates were only selected from part of the year. b Frequency of different STs in the collection, most frequently isolated STs are colour coded, previously untyped STs highlighted in teal. c Frequency of the different STs by year, colour code as in (b).

Population structure

There were 71 different STs represented in the collection (Fig. 2a, b; 60 previously type and 11 previously untyped, though these have now been assigned an ST). The most frequently isolated STs were ST69 with 20/169 (11.8%) isolates, ST131 with 18/169 (10.7%) isolates, ST10 with 15/169 (8.9%) isolates and ST410 with 8/169 (4.7%) isolates. There were 13/169 (7.7%) isolates from 11 STs which were previously unreported using MLST (Fig. 2b, c). Over half of the observed STs were only represented by a single isolate (38/71, 53.5%) showing that neonates tested here were exposed to and infected by a highly diverse pool of E. coli that span the species phylogeny. ST410 was disproportionately found in the CSF rather than blood culture samples (6/8 [75%]), compared to ST69 (2/20 [10%]), ST131 (2/18 [11%]) and ST10 (2/14 [14%]) which were found primarily in blood culture samples.

Importantly, the ST diversity was also highly variable over time. 50/71 (70.4%) STs were only found in a single year, and 7/71 (9.9%) STs were found in only two years, with each year showing a similar pattern of high diversity during the entire study period. Frequently occurring STs were also prominent in different years (e.g. ST410 in 2016 and 2017), only ST69 was consistently isolated and was the most frequently isolated or joint most frequently isolated ST in 5 out of 7 of the years with more than ten isolates. Even the most prevalent STs like ST69 or ST131 however fluctuated over time, with numbers between 1/29 (3%) and 6/35 (17%) for ST69, and 1/20 (5%) and 5/29 (17%) for ST131, respectively, with no clear trend over time observable for any of the main STs. We also note that previously unreported STs are derived from a range of years, including the most recent data. This indicates that these are not representing older lineages that might not be covered well in databases consisting mainly of recent samples, but indicating a high undescribed diversity circulating at the present time. Four samples were polymicrobial. They both contained two different colony morphologies and two different STs. Two of these samples contained ST10 along with another ST and one contained ST69 along with another ST.

O-antigen and H-antigen diversity

There were 63 O-types found in the collection, none of which were identified in more than 10% of the isolates. The most frequently isolated were O15 with 15/169 (8.9%) isolates, O25B with 15/169 (8.9%) isolates and O8 with 13/169 (7.7%) isolates (Fig. 3a). These same O-types (O15, O25B and O8) were also the only ones found in greater than 75% of the years (6 out of 7 or more) that had more than 10 isolates (Fig. 3b). Like ST there was no sign of the population becoming increasingly dominated by any O-type over time, the composition between years differed strongly (Fig. 2b). There were no years in which any O-type represented greater than 20% of the isolates, the largest proportion of isolates belonging to a single O-type per year were O11 and O8 which were both associated with 3/16 (18.8%) of all cases in 2016 (Fig. 3b).

a A bar chart showing the frequency of the different O-types. b A bar chart showing the proportion of isolates per year that had different O-types. The colours are the same as those represented in (a). c A bar chart showing the frequency of the different H-types. Where an isolate had more than one H-type gene, this was counted twice. d A bar chart showing the proportion of isolates per year that had different H-types. The colours are the same as those represented in (c).

We performed long-read sequencing for 14 isolates which either had no O-type call or had multiple O-type calls, to determine the genomic region between galF and gnd where the O-antigen type locus is usually found in E. coli (Supplementary Fig. 2). Having no O-antigen is highly unusual and leads to increased susceptibility to antimicrobial stress, and thus seems unlikely to be present in clinical isolates. Ten different O-antigen loci were revealed in these 14 isolates, with four of them represented by two isolates each. Two isolates were confirmed as O178 and one appeared to be a hybrid of O8 and O160, whilst seven of these ten O-types were so far undescribed (Supplementary Fig. 2) and one showed similarity to the OX-13 gene in Salmonella (BKRHXR).

There were 34 H-types found in the collection, of which 4 H-types were each identified in more than 10% of isolates. These were H4 with 29/173 (16.8%), H18 with 24/173 (13.9%), H5 with 18/173 (10.4%) isolates and H9 with 17/173 (9.8%) isolates (Fig. 3c). Five H-types were found in at least 75% of the years (6 out of 7 or more) that had more than 10 isolates (H4, H18, H5, H9 and H7). H4 and H18 were the only H-types responsible for greater than 20% of the isolates in any year with more than 10 isolates (in two years each) but no H-type was identified in over 50% of isolates in a single year. The largest proportion of isolates belonging to a single H-type in a single year was H4 with 5/14 (35.7%) of cases in 2013 (Fig. 3d). There were 4/169 (2.4%) isolates that had more than one H-type and may be able to undergo phase variation for immune escape, hence the denominator of 173 above. ST410 isolates with O8 and H9 were more likely to be found in CSF (6/33 [18.2%]) than bloodstream (2/133 [1.5%]; χ² = 16.03, p = 0.0015).

Excluding STs for which there was just one isolate, we examined whether multiple O-types or H-types were found in isolates of a single ST. The median number of O-types per ST for the STs that met this criterion was 2, with 10/23 (43%) encoding for just a single O-type and 13/23 (57%) encoding for multiple O-types. The median number of H-types per ST for the STs that met our criteria was 1 with 14/26 (54%) encoding for just a single H-type and 12/26 (46%) encoding for multiple H-types. Three of the most frequently occurring STs encoded for multiple O-types and H-types. ST10 showed the highest diversity of O-types and H-types, with 10 different O-types and 7 different H-types. ST131 covered 3 different O-types (O25B, O11, and O16) and 2 different H-types (H4 and H5), which occurred from 2013 to 2020 and often with multiple O-types or H-types in the same year. ST69 isolates included 4 different O-types and 2 different H-types. ST410 on the other hand occurred frequently from 2016 to 2020 but all isolates were encoded for only a single O-type (08) and a single H-type (H9).

Predicted vaccine coverage

The EXPEC9V conjugate vaccine (which covers the O1A, O2, O4, O6A, O15, O16, O18A, O25B, and O75) might be expected to confer immunity to up to 64/169 (37.9%) of these cases, the original 10 V composition (including O8) would have covered up to 77/169 (45.6%), demonstrating a loss of 7.7% by the removal of just one O-antigen of high prevalence in our setting (Fig. 4a and Supplementary Fig. 3). The EXPEC4V vaccination (which covers O1A, O2, O6A, and O25B) would cover have covered 29/169 (17.2%) of cases (assuming the O6 isolates are of the O6A subtype, something that we have not been able to confirm; Fig. 4a, b; ‘Methods’). Analyzing the data by year (including only years with 10 isolates or greater) the EXPEC9V vaccine covered fewer than 50% of the isolates’ O-types in every year, and 2 out of 7 years covered less than 30% of the vaccine O-types, with the lowest coverage in 2013 where only 23% of the isolates were from vaccine O-types, and importantly we observe no major change in coverage over time (Fig. 4c). The EXPEC4V vaccine covered less than 30% of the isolates’ O-types in every year, and 5 out of 7 years covered less than 20% of the vaccine O-types, with the lowest coverage in 2017 where only 5% of the isolates were from vaccine O-types, with fluctuations over time that do not indicate any improvement in coverage in future (Fig. 4c). Regarding the isolates which were resistant to first- and second line antimicrobial therapy (benzylpenicillin, gentamicin and ceftriaxone), the EXPEC9V vaccine would be expected to confer immunity to 13/34 (38.2%) of these cases and the EXPEC4V vaccine would be expected to confer immunity to 9/34 (26.5%) of these isolates.

a Rarefaction curve showing the theoretical protection given against vaccines covering the most frequently isolated H-types and O-types, as well as the potential protection given by the EXPEC9V and EXPEC4V. The horizontal line shows the point at which 80% of isolates would be covered. For isolates with more than one H-type both were counted, there were multiple isolates with more than one H-type so the line for H-type goes above 1. Supplementary Fig. 3 shows the same graph but where isolates had more than one H-type called they were only counted once. b A bar chart showing the frequency of the different O-types with the O-types covered by EXPEC4V and the additional coverage on top of this offered by EXPEC9V highlighted. c A bar chart showing the proportion of isolates per year that had different O-types. The colours are the same as those represented in (b).

To cover 80% of cases, an O-antigen based vaccine would need to offer protection against the top 30 O-types. In contrast, to cover the top 80% of these cases, a H-type vaccine would need to cover the top 12 H-types (Fig. 4a) if a protein-based vaccine were considered. If the four most frequently occurring O-types in our setting were selected from our cohort for a vaccine (O15, O25B, O8, and O17) this vaccine would cover 50/169 (29.6%) of cases, ranging from 21 to 40% per year, and the nine most frequently occurring O-types (O15, O25B, O8, O17, O18A?, O11, O16, O1A and O45) would represent just under half of the isolates 81/169 (47.9%), ranging from 33 to 67% per year, with numbers fluctuating over our study period showing no indication that there would be an increase in coverage in future.

Antimicrobial resistance and plasmid replicons

At the time of the study the first line treatment for neonatal sepsis and meningitis in QECH was benzylpenicillin and gentamicin, with second line treatment ceftriaxone. E. coli isolates with resistance to all three antibiotics were therefore difficult to treat (42/194 [21.6%]). There was occasional but limited use of amikacin or meropenem for neonates with proven or high suspicion of ceftriaxone resistance or who were very unwell. The use of meropenem and amikacin increased over the study period. Isolates were frequently multi-drug resistant (MDR; resistant to antimicrobials in four or more antimicrobial categories; 60/200 [30%]) and extensively-drug resistant (XDR; resistant to antimicrobials in all but two or fewer antimicrobial categories tested; 29/200 [14.5%]).

We identified AMR genes against all major classes of antibiotics and several efflux pump systems, in line with the phenotypic resistances detected (Supplementary Fig. 5; Supplementary Data 5). The number of AMR genes varied by ST, with ST410 (mean 24.0, SD 1.9) and ST131 (mean 20.6, SD 4.8) having the greatest number of average AMR genes per isolate. ST10 had a lower number of AMR genes per isolate (mean 8.9, SD 1.6), whilst ST69 was intermediate (mean 12.9, SD 1.5). ST410 was present only from 2016 onwards which may partly explain the higher number of resistance genes, whilst the other STs, including ST131 were present throughout the study period.

The number of E. coli isolates that were resistant to ampicillin was 143/190 (75.3%) and was stable over the period (−0.3% change per year; S.E = 6.7%; p = 0.96; Fig. 5a). bla_EC genes, which are chromosomally encoded in E. coli, were found in all 169 isolates (bla_EC-15, bla_EC-5, bla_EC-18 and bla_EC-8). The most frequently occurring plasmid-encoded beta-lactamase penicillinase genes included bla_TEM-1 found in 121/169 (71.6%) isolates, and bla_OXA-1 found in 21/169 (12.4%) of isolates (Supplementary Fig. 4). Whilst only 49/197 (24.9%) were resistant to gentamicin, this however showed a temporal trend, increasing from 3/15 (20%) in 2013 to 16/39 (41%) in 2020 (22.8% change per year; S.E. = 7.8%; p = 0.0035; Fig. 5a). Gentamicin resistance mechanisms were mainly variants of the gene aac(3), aac(3)-IId found in 23/169 (13.6%) isolates and the aac(3)-IIe gene found in 12/169 (7.1%) isolates (Supplementary Fig. 5).

a First-line antibiotics for neonatal infection (n = 190 for ampicillin, n = 197 for gentamicin). b Second-line antibiotics for neonatal infection (n = 198). c Occasionally used antibiotics for neonatal infection (n = 67 for amikacin, n = 39 for meropenem). d Antibiotics not used in neonates, but used elsewhere in the hospital or in the community (n = 195 for chloramphenicol, n = 199 for ciprofloxacin, n = 200 for co-amoxiclav, n = 200 for co-trimoxazole). Trend lines represent linear regression (solid line) with 95% confidence intervals (shaded area).

Ceftriaxone was the second-line treatment at the time of study, and 55/198 (27.8%) were resistant to ceftriaxone. This increased over the period from 0/7 (0%) in 2012 to 18/39 (46.2%) in 2020 (31.8% change per year; S.E. = 8.3%; p = 0.00014; Fig. 5b); and 42/55 (76.4%) of the isolates resistant to ceftriaxone were also resistant to ampicillin and gentamicin, hampering the effectiveness of all first- and second-line treatments. The increase in ceftriaxone resistance was due to the widespread ESBL gene bla_CTX-M-15, which was detected in 39/169 (23.0%) isolates (Supplementary Fig. 5), whilst other alleles, bla_CTX-M-14 and bla_CTX-M-27, could be identified in only 1/169 (0.6%) isolate each, both from 2018 (Supplementary Fig. 3). ESBL genes were frequent in ST410 (8/8 [100%]) and ST131 (9/18 [50%]) as observed in other studies^43,44, and infrequent in ST69 (3/20 [15%]) and ST10 isolates (1/15 [7%]). The proportion of isolates from late onset infection resistant to ceftriaxone was 16.9% (95% CI 3.5–30.2%) higher than from early onset cases (37.2% vs 20.4%, χ² = 6.2; p = 0.012); the same pattern was observed for gentamicin with 15.8% (95% CI 2.8–28.7%) more resistant isolates in late onset infections compared to early onset cases (33.6% vs 17.9%, χ² = 5.8; p = 0.016), whereas similar proportions were observed for ampicillin with 78.5% vs 73.2% in late and early, respectively (χ² = 0.5; p = 0.50).

Alternatives for isolates resistant against all of the above antimicrobials are amikacin or carbapenems, and so far only a small proportion, 7/67 (10.4%), were resistant to amikacin. This increased from 1/6 (17%) in 2016 (amikacin was not routinely tested until 2016) to 7/21 (33%) in 2019 (52.7% change per year; S.E. = 25.9%; p = 0.042; Fig. 5c). The main gene identified was the amikacin resistance gene aac(6’)-Ib-cr5 in 21/170 (12.4%) isolates (Supplementary Fig. 5). Of the 42 isolates resistant to all first and second-line agents 11/42 (26.2%) were causing meningitis and a further 5/42 (11.9%) were resistant to amikacin, leaving only meropenem as effective treatment for these (amikacin does not reliably penetrate the blood-brain barrier). In line with the low levels of carbapenem resistance identified in other studies from this setting, no isolates showed phenotypic meropenem resistance.

Other antimicrobials are not regularly used on the neonatal unit but are tested for routinely for E. coli. Fluoroquinolones are still widely used in other wards, and 45/199 (22.6%) were resistant to ciprofloxacin which increased over the period from 0/7 (0%) in 2012 to 12/39 (30.8%) in 2020 (20.0% change per year; S.E. = 8.0%; p = 0.013; Fig. 5d). Chloramphenicol resistance has been observed in other isolates in this setting to decrease, and in line with this we observed 41/195 (21%) isolates resistant to chloramphenicol with a decreasing trend, from 2/7 (28.6%) in 2012 to 3/39 (7.7%) in 2020 (−30.5% change per year; p = 0.000027; Fig. 5d). Overall 54/67 (80.6%) were resistant to co-amoxiclav (not used on the neonatal unit and infrequently used in other hospital wards) and the proportion decreased slightly over the period (−20.7% change per year; S.E. = 21.1%; p = 0.33; Fig. 5d). Resistance to co-trimoxazole (used as prophylaxis against Pneumocystis pneumonia in HIV patients) was high over the period at 183/200 (91.5%) and increased slightly (14.9% change per year; S.E. = 9.8%; p = 0.13; Supplementary Fig. 5). Details on resistance mechanisms against these are provided in the Supplementary Materials.

There were 34 different plasmid replicons found in the dataset. The most frequently identified were the commonly detected IncF plasmid replicons that frequently carry resistance cassettes, with IncFIB_AP001918 found in 107/169 (63.3%) of isolates, IncFI found in 91/169 (53.8%) of isolates and IncFII_p found in 46/169 (27.2%) of isolates. Multiple different plasmid replicons were also found of the Col, IncH, and IncX types, with other types found less frequently.