Contribution and Interaction of Shiga Toxin Genes to Escherichia coli O157:H7 Virulence

Escherichia coli O157:H7 is the predominant cause of diarrhea-associated hemolytic uremic syndrome (HUS) worldwide. Its cardinal virulence traits are Shiga toxins, which are encoded by stx genes, the most common of which are stx1a, stx2a, and stx2c. The toxins these genes encode differ in their in vitro and experimental phenotypes, but the human population-level impact of these differences is poorly understood. Using Shiga toxin-encoding bacteriophage insertion typing and real-time polymerase chain reaction, we genotyped isolates from 936 E. coli O157:H7 cases and verified HUS status via chart review. We compared the HUS risk between isolates with stx2a and those with stx2a and another gene and estimated additive interaction of the stx genes. Adjusted for age and symptoms, the HUS incidence of E. coli O157:H7 containing stx2a alone was 4.4% greater (95% confidence interval (CI) −0.3%, 9.1%) than when it occurred with stx1a. When stx1a and stx2a occur together, the risk of HUS was 27.1% lower (95% CI −87.8%, −2.3%) than would be expected if interaction were not present. At the population level, temporal or geographic shifts toward these genotypes should be monitored, and stx genotype may be an important consideration in clinically predicting HUS among E. coli O157:H7 cases.


Covariate Adjustment
We adjusted our final risk difference (RD) and relative risk (RR) estimates for age, blood in the stool, vomiting, and fever. We considered age to confound the association between stx genotype and hemolytic uremic syndrome (HUS) (Figure 1). There is a relationship between age and stx genotype, with younger children more likely to be reported with stx2a-only and stx2a2c genotypes than stx1a2a and other genotypes. The stx2a-only and stx2a2c genotypes may be more capable of infecting and causing disease in young children, or conversely, worse at infecting and causing disease in older children and adults, thereby skewing the distribution of younger cases. There may also be differences in exposures. Particular E. coli O157:H7 strains may be disproportionally associated with some transmission vehicles than others [1], and transmission vehicles are possibly non-randomly distributed across ages. For example, cases infected with cattle-biased lineages of E. coli O157:H7 tend to be older, likely reflecting the greater contact between working-age individuals and cattle. Children <10 years, and especially those <5 years, are also at a substantially increased risk of HUS, thus causing confounding.
We adjusted for symptoms that may have been likely to alter how a patient was treated. Particular treatments may potentiate (e.g., antibiotics) or protect against (e.g., early intravenous fluids) the development of HUS. Some symptoms could make these treatments more or less likely. For example, bloody diarrhea may signal to a clinician a likely bacterial infection and they may treat with antibiotics. There is some evidence, including the current study, that symptoms like bloody diarrhea are more common with stx1a2a genotype strains. Therefore, there may be E. coli O157:H7 cases who receive antibiotics because of their bloody diarrhea, triggering progression to HUS, and this causal sequence may be more common in cases with the stx1a2a genotype ( Figure 1). Lacking full data on interventions used, we adjusted for symptoms to block this alternate causal pathway.
A potential limitation of this approach is that our adjustment for symptoms may only be partially blocking the indirect causal pathway between genotype and HUS. We believe this likely had minimal impact on our estimates. Although bloody diarrhea cannot be explained entirely by stx genotype among non-O157 Shiga toxin-producing E. coli (STEC) [2], in our cohort bloody diarrhea was more commonly experienced by cases with the stx1a2a genotype. Conversely, vomiting was more commonly experienced by cases with the stx2a genotype. Hematochezia is often associated with bacterial agents and vomiting with viral agents [3], and therefore these symptoms may impact the likelihood of antibiotic treatment. Thus, the indirect path may be accounting for a larger portion of the stx1a2a genotype's than the stx2a genotype's HUS risk; fully controlling for it would be expected to increase RD and RR estimates.

Interaction
Stabilized inverse probability weights were calculated using a multinomial model that regressed genotype on age and symptoms. The probabilities p01, p10, and p11 were then calculated as the sum of the weights of HUS cases of the stx2a-only, stx1a-only, and stx1a2a genotypes, respectively, over the sum of the weights of all cases of the given genotype. The probabilities were then combined using the formula 11 − 10 − 01 + 00 (1) with p00 set to 2/1,000,000 to obtain the amount of additive interaction. This entire process was bootstrapped, with 10,000 iterations, and bias-corrected and adjusted 95% confidence intervals were determined. We did not assess multiplicative interaction. The risk of HUS in E. coli O157:H7 cases without any of the stx genes we tested for is near 0. With p00 = 0, all relative risks (RRs) are undefined and the relative excess risk due to interaction cannot be calculated.

Loss of stx Genes
Because of the potential for the loss of stx-carrying bacteriophages after isolation, we conducted a sensitivity analysis in which isolates with an atypical stx genotype for their pulsed field gel electrophoresis (PFGE) pattern and phylogenetic lineage [4] were reclassified to a different stx genotype. More than half of the isolates with a particular PFGE pattern needed to have the same stx genotype for us to reassign genotype. However, if the isolate had been typed to determine its phylogenetic lineage and it fell in a different lineage than the majority of isolates with the PFGE pattern, its stx genotype was not reclassified. For example, there were 11 isolates with the XbaI PFGE pattern EXHX01.0248. All were either typed to lineage Ib (n = 8) or did not undergo lineage typing (n = 3). Nine of these isolates had the stx1a2a genotype, one had the stx1a-only genotype, and one had the stx2a-only genotype. The latter two isolates were reclassified as having the stx1a2a genotype for the sensitivity analysis.

Missing Data
Of the 1,160 culture-confirmed cases reported to the department of health during the study period, 194 had no bacterial isolate in Washington State University's specimen bank or could not be revived. Over half of these (n = 125) were reported in the first three years of the study, 2005-2007. Missing predominantly the oldest isolates suggests that the most probable associated mechanisms are degradation and being misplaced. Neither of these is likely to be a function of genotype or HUS status. We believe they are largely random processes and satisfy criteria for missing completely at random.
We conducted supplemental analyses to determine if other factors appeared to affect isolate missingness. In univariate analysis, we found associations between missingness and presence of diarrhea, HUS status, hospitalization, and year. In multivariable analysis with these variables, only year [odds ratio (OR) 1.36; 95% confidence interval (CI) 1.27, 1.47] appeared to be associated with having an isolate available for testing. The point estimate for diarrhea was relatively high (OR 3.60), but the 95% CI (0.84, 13.9) suggested statistical uncertainty. Given the very small proportion of E. coli O157 cases without diarrhea, we do not believe this would have impacted our results.
We also examined patterns of missingness for HUS status. There were 28 cases with a genotyped isolate and 10 without a genotyped isolate for whom we could not determine HUS status. Case report forms for these cases indicated they were hospitalized, presenting the possibility that they could have experienced HUS. However, the indicated hospital either did not have a record for the patient during the time of the E. coli O157 episode (n = 36) or the record was incomplete and HUS status could not be determined (n = 2).
Among cases with a genotyped isolate, in univariate analysis, fever, having an underlying illness, year, and age were associated with missingness. After multivariable analysis, only year (OR 1.23; 95% CI 1.03, 1.48) and fever (OR 0.32; 95% CI 0.18, 0.83) were associated with having ascertained HUS status. A year in which illness occurred is likely a predictor of missingness, because older records may be more likely to have been lost, particularly for hospitals still using paper systems at the time. As with isolates, this was likely a random process. Cases with a fever were more likely to have missing HUS status; 1% (8/545) of cases without a fever were missing HUS status vs. 4% (13/348) of those with a fever. These missing cases were evenly distributed across the genotypes. Given the similar distribution of fever across genotypes among known HUS status cases, these missing cases likely had little impact on the results.

Interaction
The small number of stx1a-only isolates compromised and yielded an estimate of additive interaction for stx1a and stx2a that lacked precision. The bootstrap generated a bimodal distribution, with the two peaks centered at −0.0396 and −0.270 ( Figure S2). This suggests there may be heterogeneity but is more likely due to the small sample of stx1a-only isolates. There was one HUS case with a stx1a-only isolate, and in that case's stratum, defined by age and symptoms, there were no other stx1a-only isolates. Bootstrap samples that include this case are likely to overestimate the HUS risk associated with the stx1a-only genotype, making the interaction appear sub-additive to a very great degree. If the risk of HUS were 0 in cases with the stx1a-only genotype, we would expect the interaction to be the same as the RD between stx1a2a and stx2a; the −0.0396 peak is much closer to the RD, suggesting that the inclusion of the stx1a-only HUS case in a given bootstrap sample is driving the bimodal distribution. Only 152 of the 10,000 estimates of the interaction were ≥0.

Hospitalized Cases
We restricted this sensitivity analysis to cases who had been hospitalized. By definition, this removed only non-HUS cases, because non-hospitalized cases were assumed to not have HUS. Thus, HUS incidence increased for all genotypes. HUS incidence was 15.8% (28/177) for stx1a2a, 25.5% (24/94) for stx2a, and 18.8% (16/85) for stx2a2c hospitalized cases.
Given the much smaller sample size, RD and RR estimates were not adjusted for blood in stool and vomiting to avoid over-parameterizing. Both symptoms highly predicted antibiotic exposure, providing reassurance that adjusting for antibiotic use would block the indirect path. However, measured fever did not predict antibiotic use, though it did predict HUS. It may therefore be working through a different pathway and was left in the model. Figure S1.     Symptom adjusted models are adjusted for blood in stool, vomiting, and fever. Fully adjusted models are adjusted for age, blood in stool, vomiting, and fever. Abbreviations: CI, confidence interval; RD, risk difference; RR, relative risk