Isolates

A total of 1243 C. jejuni and 616 C. coli isolates collected between 2002 and 2012 in Switzerland were included in the main analyses. For one C. jejuni and three C. coli isolates no unambiguous flaB sequence was available and they were excluded from analyses involving these sequences. The C. jejuni were isolated from dogs (159 isolates), chickens (435 isolates) and human campylobacteriosis cases (649 isolates), and have mostly been described previously [5,11,15,24,25,27]. The C. coli were isolated from pigs (360 isolates), chickens (175 isolates), and human clinical cases (81 isolates) and have also been characterized previously [5,15,25-27]. Briefly, chicken and pig isolates were collected at slaughterhouses in the framework of monitoring programs, while dog isolates stemmed from fecal samples submitted to our in-house diagnostics department, with the exception of 23 isolates from 2002/3 and 19 samples from 2012, which were taken from healthy dogs. Human case isolates were provided by the Swiss National Centre for Enteropathogenic Bacteria (NENT), in the framework of collaborative studies.

For an additional source attribution analysis 23 available Swiss C. jejuni isolates taken from healthy dairy cows in 2001/2 were included [15]. This group was only used in the source assignement by MLST to get an idea if cattle are possibly important. They were not considered in the main analyses due to the bias of the small sample size and limited time span of only one year. A list of all included isolates is provided in the supplementary material (Table S1, Table S2).

Sampling has been approved by the Federal Office of Public Health for the human isolates and by the Federal Veterinary Office for the animal isolates. All patient data as well as data on farms and slaughterhouses where samples originated were anonymized.

Statistical Analyses

The PSI as a measure for the overlap of two frequency distributions [28] was used to compare the MLST sequence types (ST) and flaB types of isolates from different sources. PSI ranges from zero to one, where one indicates that the two groups are identical and zero means they share no types. It was calculated according to: $PSI=1-0.5{\sum }_1^i|p_i-q_i|$where p_i and q_i are the proportion of isolates from group p and q, respectively, belonging to type i. Confidence intervals (CI) were estimated with the bootstrap method implemented in C++ as follows: The standard deviation (SD) of 1000 resampling rounds was calculated and the 95% CI estimated with PSI±2*SD.

Analysis of molecular variance (AMOVA) and calculation of fixation indices (F_ST) were performed with the Arlequin software [29]. Briefly, AMOVA partitions overall molecular variance into variance between individuals within populations, among populations within groups and among groups of populations. F_ST is a measure for genetic differentiation between populations ranging from zero to one where zero means that the populations are not separated. Both F_ST calculations and AMOVA were done using either flaB fragment sequences or concatenated sequences of the seven MLST alleles.

The STRUCTURE software [20] was used to assign the human isolates to their most probable source. Briefly, the most likely origin of the isolates is estimated based on the similarity in allele frequencies with the potential source populations. For the MLST data the program was run with 100,000 burn-in steps and 100,000 repeats applying the no-admixture model and assuming uncorrelated gene frequencies. The POPFLAG was set to zero for the human isolates and USEPOPINFO as well as STARTATPOPINFO turned on. This way, the program incorporated information on the population of origin for the isolates of the different potential sources, while the origin of the human isolates was presumed unknown. To test for the influence of single loci on the assignments, we ran jackknife analyses where each locus was left out once and the program was run with the remaining six loci as described above, except that burn-in steps and repeats were reduced to 10,000 to moderate run time.

Bootstrap confidence intervals for source attribution were estimated using the following procedure implemented in Perl: STRUCTURE was run 100 times each time using a new subset of 100 human isolates randomly drawn with replacement from the whole set. For the resulting 100 assignment proportions, the mean and standard deviation were calculated and the 95% CI estimated with mean±2*SD. For the bootstraps, burn-in steps and repeats were also reduced to 10,000 to moderate run time, while the other parameters were left as before.

Self-attribution probabilities for the isolates from different animal hosts were calculated like the bootstrap confidence intervals for assignment of the human isolates, now taking a random sample of 100 draws with replacement of each isolate group and assigning them to a separate population. The POPFLAG for this test population was set to zero and the program was run as described above, calculating the probability of assignment to the correct source population.

As an attempt to assign the human isolates to their most probable source according to their flaB sequence, STRUCTURE was run using all polymorphic sites (153 in our dataset) in the flaB SVR. For this analysis, C. jejuni and C. coli were not separated because flaB-types are shared between both species, in contrast to STs, which are species-specific. However, to account for the larger variability of flaB over time (see results section) and avoid bias due to the unequal temporal distribution in the different sources, only isolates from 2008 to 2012 were included. Thereby, no cattle isolates were available for the flaB analyses. Parameters were set as described for the MLST data, except that burn-in steps and repeats were again reduced to 10,000.

Graphical displays of STRUCTURE results were created with the DISTRUCT software [30]. Derivation of binominal exact confidence intervals as well as standard logistic regression and penalized maximum likelihood logistic regression were performed with STATA IC 12.1 (StataCorp LP, Tx, USA).

PSI

The proportion similarity indices (Table 1) for MLST types between C. jejuni from different sources were generally high with the greatest similarity between human and chicken isolates. The overlap between human and dog isolates was significantly lower, while dog and chicken isolates showed a high similarity. Comparable results were obtained for flaB types (Table 1).

PSIs were generally lower between C. coli from different sources both for the MLSTs and flaB types (Table 1). Nevertheless, the overlap between human and chicken isolates was significantly higher than between human and pig isolates, while the similarity of chicken and pig isolates lay somewhere in between.

In all comparisons the overlap of flaB types was higher than of MLSTs though never significantly so.

F_ST

Fixation indices were calculated between the same isolate groups as the PSIs (Table 1) using both the concatenated sequences of the seven MLST alleles and the flaB sequences alone. With this method C. jejuni isolates from humans were again determined to be closer to those from chickens than to those from dogs. Chicken and dog isolates were found to be very close with an F_ST of 0.010 (p=0.02) for the MLST sequences and an F_ST not significantly different from zero for the flaB sequences (p=0.18). In this instance the result differs from that obtained with the PSI, where the correspondence between human and chicken isolates was approximately the same as between dog and chicken isolates.

For C. coli human and chicken isolates were also found to be very closely related (Table 1). In contrast, there was a marked difference between human and pig isolates. As observed with the PSI, the difference between chicken and pig isolates was slightly lower.

Source assignment

For the MLST data the STRUCTURE program was run separately for C. jejuni and C. coli because the human samples did not in all years represent an unbiased species distribution from all human cases; in 2008, for example, analyses focused on C. jejuni [24]. For C. jejuni 76.8% (95% CI: 64.7%-85.0%) of the human isolates were assigned to the chicken reservoir while 23.2% (95% CI: 15.0%-35.3%) were assigned to the dog reservoir. Our jackknife analyses showed that these STRUCTURE source attributions were robust to the exclusion of single alleles (Table S3). When this analysis was repeated including the 23 cattle isolates, 69.3% (95% CI: 57.5%-72.3%) of the human isolates were attributed to chicken, 9.5% (95 %CI: 8.8%-19.0%) to dogs and 21.2% (95% CI: 13.7%-28.7%) to cattle. For C. coli 86.4% (95% CI: 81.1%-92.3%) of the human isolates were estimated to stem from chickens and 13.6% (95% CI: 7.7%-18.9%) from pigs. Considering that 91% of human cases are caused by C. jejuni and 9% by C. coli [5], this suggests that in total 77.5% of the human campylobacteriosis cases can be attributed to the chicken reservoir, 21.1% to the dog reservoir and 1.4% to the pig reservoir, if cattle isolates are not considered. With the inclusion of the cattle isolates still 70.9% of the cases are attributed to chickens, but only 8.6% are attributed to dogs, 19.3% to cattle and 1.2% to pigs (Figure 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. STRUCTURE source attribution of human C. jejuni and C. coli isolates.

Single isolates are represented by vertical lines; the probability of attribution to the different source clusters is indicated by the color. Attribution to the chicken cluster is shown in blue, cattle cluster in green, dog cluster in red and pig cluster in yellow.

https://doi.org/10.1371/journal.pone.0081796.g001

Self-attribution was highest for C. coli in chickens with 79.8% (95% CI: 74.0%-85.6%) and slightly lower for C. coli in pigs with 70.1% (95% CI: 63.2%-77.0%). C. jejuni chicken isolates showed a self-attribution of 62.5% (95%CI: 53.7%-71.3%) while it was only 38.1% (95% CI: 30.3%-45.9%) in dog isolates and 26.1% (95% CI: 17.6%-34.5%) in cattle isolates.

When STRUCTURE was run with the flaB sequences analyzing C. jejuni and C. coli together (since they share alleles), similar results were obtained. In this analysis 80.5% of the cases were assigned to the chicken reservoir, 17.7% to the dog reservoir and 1.8% to the pig reservoir. However, for flaB the separation between dog and chicken isolates was even less clear and dog isolates were mainly attributed to the chicken cluster.

Campylobacter population turnover

To analyze the genotype variation over the years, AMOVA was performed with the concatenated MLST allele sequences as well as the flaB sequences (Table 2) for samples of 10 and more per year and host. For C. jejuni 4.6% of overall variance in MLST sequences was due to variance among hosts while only 1.6% was due to variance among years within hosts. The opposite was true for flaB sequences where only 0.6% of overall variance was due to variance among hosts and 2.0% was due to variance among years within hosts.

In C. coli variance among years within hosts was much more pronounced, accounting for 10.6% of overall variance in MLST sequences and 24.6% in the flaB sequences (Table 2). The ST shift occurred somewhat in parallel between hosts. For example the frequency of ST-825 and ST-827 increased in both human and chicken isolates over the observed time period while there was a decrease in the frequency of ST-854 which was also observed in the pig isolates (Table S4).

Comparison of quinolone resistant and susceptible isolates

AMOVA was performed for C. jejuni and C. coli separately to determine if MLSTs and flaB-types differ between quinolone resistant and susceptible strains (Table 3). In C. jejuni 5.9% of overall variance between the concatenated MLST allele sequences was due to differences between resistant and susceptible isolates within hosts, while there was no significant difference between hosts. Similar results were obtained for flaB sequences, where 8.0% of variance was due to variance between resistant and susceptible isolates within hosts. In C. coli the difference between quinolone susceptible and resistant isolates was smaller though it was still significantly different from zero (p <0.001; Table 3) while there was again no significant difference between hosts. When looking at the isolates from all hosts together, the F_ST for MLST between quinolone resistant and susceptible isolates of 0.067 (p<0.001; 95%CI: 0.058-0.076) suggests considerable differences in sequence types for C. jejuni, whereas this was significantly lower for C. coli (F_ST =0.017; p<0.001; 95%CI: 0.011-0.023).

To detect possible associations between ST and quinolone resistance in C. jejuni a logistic regression model was fitted including host and ST as risk factors (Table 4). The ten most frequent STs, together representing 61.9% of all isolates, were considered separately and compared to the remaining 157 pooled STs (Figure 2). Since all isolates with ST-45 (n=77) were quinolone susceptible and all ST-464 isolates (n=27) were resistant, the standard logistic regression model could not be applied. Instead, Firth’s penalized likelihood approach to separation was used to account for the problem of separation [31]. Isolates from humans had significantly higher odds of being resistant, than those from chickens or dogs. Eight of the ten most frequent STs were associated with significantly lower quinolone resistance, with the lowest OR for ST-45. ST-464 was associated with significantly higher odds of quinolone resistance.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Frequency and origin of the ten most frequent C. jejuni and C. coli sequence types (ST) used in the regression analyses.

The remaining 157 C. jejuni and 109 C. coli types were pooled in the category ‘other ST’. Numbers on the y-axis indicate the number of isolates.

https://doi.org/10.1371/journal.pone.0081796.g002

For C. coli the ten most frequent STs, accounting together for 59.7% of all isolates (Figure 2), were also used to fit a logistic regression model for quinolone resistance but here the standard model could be applied. Chicken was specified as the baseline host and the remaining pooled STs (109 different types) were used as baseline (Table 5). Again human isolates had significantly higher odds of resistance than chicken isolates but also pig isolates had higher odds. Five of the ten most common STs were associated with significantly lower odds of resistance.

A temporal analysis showed a tendency towards increasing quinolone resistance over time for dog and chicken isolates, while there was little change among human and pig isolates (Table S5).