Among 713 isolates, 330 MLST genotypes were identified. Two isolates (ST-284 and ST-327) had 4 alleles typical of C. jejuni and 3 typical of C. coli. All others had >5 typical C. jejuni alleles. Table 3 shows assignment accuracy when we used the whole dataset and a leave-one-out strategy to assign strains to 3 host species (cow, sheep, and chicken) on the basis of 7 alleles, the clonal complex, the ST, and combinations thereof. Because random assignment would be correct one third of the time, how much improvement genotype information showed compared with random assignment is more informative than the percentage correct, i.e., what proportion of the gap between 33% correct expected by using random assignment and 100% correct with perfect prediction has been closed. Assignment by using the 7 alleles closed 37% of this gap compared with 10% for ST and 13% for clonal complex. Prediction did not improve substantially when ST or clonal complex information was added to allele information. These overall results emphasize the limits in using an ST or clonal complex as a summary of MLST when predicting host of origin. We therefore used alleles in all further analyses and explored the basis for the better accuracy of this approach.

Prediction of host of origin to 3 host sources on the basis of alleles is shown in Table 4. The method showed higher accuracy for distinguishing chicken strains from cow or sheep strains than for distinguishing between strains from the 2 bovid species. When we performed analysis restricted to cattle and sheep isolates, we obtained an assignment accuracy of 58% compared with 50% expected by chance and thus explained only 16% of remaining uncertainty. This additional analysis showed little detectable host association for these 2 closely related host species. Further comparison of chicken isolates with a combined population from cattle and sheep showed improved resolution and allowed correct prediction 80% of the time (60% of uncertainty removed), which indicated substantial host association.

Given the nature of the dataset, we must consider possible confounding factors such as differences in time or location of sampling, which may lead either to completely spurious associations or to overestimates of their magnitude. Indeed, there was evidence for modest time and geographic effects within our dataset. For example, in a comparison of UK chicken isolates in 1997 or earlier and in 1998 or later (Table 5), 66% could be assigned to the population of the correct period based on allelic profile. Similarly, when UK and Dutch chicken isolates were considered, 69% were assigned to the correct country. We therefore performed additional analyses in which host was negatively associated with time, space, or both (Table 5). Late UK chicken isolates (1998–2003) were assigned by using early UK chicken (1997 or earlier) and late UK bovid isolates (1998–2003) as training sets, which indicated 77% assignment to chickens. UK chicken isolates were assigned by using non-UK chickens and UK bovid isolates as training sets, which indicated 64% assignment to chickens. These analyses showed that host effect is stronger than that of time or space and that our findings are not the result of confounding by these factors.

To explore the mechanism underlying the better performance observed for allele-based assignment and to better understand the biologic processes that produce this host signature in the bacterial genome, we investigated assignment within the ST-21 complex. This clonal complex comprises a substantial proportion of isolates and is highly diverse (5,10,17,18). Our sample contained 252 ST-21 complex isolates. Of these, 188 were not ST-21 but differed at 1 to 3 alleles from the central genotype. We assigned these 188 isolates to chicken or bovid hosts on the basis of only alleles at which they differed from ST-21 by using all non–ST-21 complex isolates as the training set. A total of 66% of isolates were assigned to the correct host. This finding suggests that ST-21 complex isolates acquire alleles that are characteristic of the host population. To demonstrate that this deviation from 50% is not a sampling artifact or chance effect, we restricted analysis to the 88 unique ST-host combinations, which largely eliminates the possible effects of clonal expansion within host, and performed a permutation test to assess the possible role of chance. Of these combinations, 67% were correctly assigned, which was a higher proportion than observed in any of 10,000 iterations in a permutation test (Figure, panel A).

The overall accuracy of host assignment based on acquired alleles is limited because many of these alleles are each too rare to enable accurate estimation of their frequency in particular host gene pools. Acquired alleles that are frequently observed give more accurate host prediction. To illustrate this visually (Figure, panel B), we used as predictors only those alleles that are found in >10 different ST-host combinations in the non–ST-21 complex isolates and substantially differentiated between chicken and bovid populations (on the basis of a 65% cut-off value). All 4 isolates with 2 alleles, both suggestive of either chicken or bovid origin, were from the predicted source. In 1 instance, 2 potentially informative alleles gave conflicting information; 1 suggested bovid origin and 1 suggested chicken origin. Isolates with this ST came from both sources. Of the 24 STs with only 1 informative allele, 18 were correctly assigned; only 4 were incorrectly assigned. The remaining 2 STs were isolated from chicken and bovid sources.

Our analyses confirm the association of C. jejuni genotypes with host species, and demonstrate a clear distinction between isolates obtained from chickens and those obtained from bovids, when alleles are considered independently in statistical analysis. This finding was robust to sampling differences in time and place and suggested that host effects were stronger than geographic and temporal effects, which must be considered if these associations are to be used in epidemiologic investigations. Moreover, because populations of C. jejuni in farm animals such as bovids and chickens may show greater similarity than those from other hosts (5,9), the approach may be more accurate when considering C. jejuni from a more diverse host range. The distinction between cow and sheep isolates is much weaker. Differentiation between these species might be demonstrable if more genetic information was available. However, the minor differences observed may be a sampling artifact with these species sharing a common gene pool.

The allele-based method we have used helps alleviate the problem of excess discrimination in Campylobacter typing. Many alleles show differences in frequency between hosts. These alleles provide useful information on source for STs that are too rare to allow estimates of their frequency in different hosts (e.g., because they are absent from training sets).

Figure

Figure. Prediction of source of origin within the sequence type ST-21 complex. A) Observed accuracy of prediction (arrow) compared with distribution of values obtained by permuting host labels so that the alleles...

Our approach has some limitations that must be considered in any more extensive application. The current accuracy estimate of 80% in distinguishing chicken isolates from bovid ones may be optimistic if sampling effects are quantitatively important. Sampling effects would include the nature of the sample (feces, meat), as well as time and place. For example, the dominant Campylobacter types found in processed carcasses have been shown to differ from those found in live chickens entering the processing plant (19). Nonetheless, we have shown that easily identifiable sampling effects are overwhelmed by the host effect. Moreover, because analysis within the ST-21 complex (Figure) is robust to identified and unidentified sampling effects, we do not believe this to be a major problem.

An additional limitation of our allele-based application of STRUCTURE is that it assumes allelic independence, which is clearly violated for the dataset analyzed. Assignment accuracy can be estimated into 2 different ways. The first, which we have used throughout this report, is a holdout procedure whereby source of origin of strains for which the actual origin is known is predicted by using the rest of the sample as a training set. This method provides an unbiased empiric measure of accuracy. To predict isolates for which the source is unknown, this procedure is not possible. Thus, it would be desirable to use estimates of accuracy that the algorithm provides. Because STRUCTURE assumes each allele is independent, its estimate of the accuracy with which it estimates the frequency of a particular multilocus genotype frequency is often overconfident. For example, in differentiating chicken isolates from those originating in cattle and sheep, STRUCTURE predicts an accuracy of 91%, but empirically it achieves an average accuracy of 80%. A better estimate of uncertainty would be necessary for predictive purposes. More sophisticated genetic models that reflect dependence among loci should achieve more accurate assignment as well as better estimates of statistical uncertainty.

Despite these limitations, this approach demonstrates the ability to assign isolates probabilistically to populations. When broad reference populations from the full range of possible sources are available, groups of isolates, such as those affecting a human population over a period of time, can be apportioned to their sources. However, precision in the attribution of C. jejuni may be less than that of Salmonella, in which different animal and bird species appear to host more differentiated populations (3). Prediction is possible with individual isolates, in some instances to 1 source, although prediction may suggest a range of populations rather than 1 population. For example, 2 of the sequenced C. jejuni genomes are from known sources, 1 from a chicken (isolate RM-1221) (20) and 1 from a human with campylobacteriosis who had drunk raw milk (isolate 81-176) (17,18). Assigning these isolates on the basis of reference datasets we used in this report predicted their origin as chicken (99% probability) and cattle/sheep (97% probability), respectively.

The broad host range of C. jejuni, spanning a variety of mammalian, avian, and other species, makes it a good model for studying features that may be informative of the ecology of multihost pathogens. C. jejuni acquires genome fragments estimated to be a few hundred bp in length (21) from of other members of this species. Our analysis of the ST-21 complex shows that isolates in this complex have acquired genetic material prevalent in the population of Campylobacter carried by their host species (Figure). This observation implies that there is persistent differentiation in allele frequencies between different host species and that many ST-21 isolates represent lineages that have persisted within the same host species long enough to acquire a substantial number of alleles.

We surveyed 7 loci and found an average of 0.32 host-specific alleles in 81 STs other than ST-21 that were members of ST-21 complex, i.e., ≈5% of the alleles in this analysis. The acquired genes were approximately evenly distributed between these types, with 6 of the 7 loci having >1 import. The MLST loci were chosen because they represent core metabolic functions of C. jejuni (10) and are not obvious candidates for host adaptation. Therefore, we are likely observing the neutral level of acquiring genetic information. Extrapolating linearly from these 7 loci to 1,654 gene-coding sequences in the C. jejuni genome (22) gives an estimate of 76 genes with alleles typical of a particular host species within each ST-21 complex isolate. This estimate is rough because it is based on fairly limited data and recombination and selection at other genes may be different. However, this approximation shows the potential for substantial adaptation to the most recent host by homologous recombination. Homologous recombination may be an important factor in allowing a bacterial species to colonize a wide range of host species stably while adapting to some extent to each host.

In conclusion, a population genetic approach has allowed host assignment for C. jejuni for which host-specific markers are unavailable but host species populations are differentiated by allele frequency at a range of loci. Host association appears stronger than temporal and geographic effects. Homologous recombination generates a host signature in the C. jejuni genome and analyses that use this signal have improved accuracy of host prediction. The inherent standardization and portability of sequence typing in combination with the availability of such improved assignment techniques support the application of this approach to clarify aspects of C. jejuni epidemiology on a global scale and to study other suitable microbes.