In our main experiment, we compared the consequences of infection of eight different natural C. elegans
strains with 5 different S. marcescens
strains plus one control (heat-killed bacteria of S. marcescens
Db11, a strain for which the genome sequence is now complete). The C. elegans
strains were isolated from Münster, in Northwest Germany, and belong to four different microsatellite genotypes [18
]. The S. marcescens
strains originate from different locations around the world. The interaction between the two species was examined with the help of a survival assay, in which the survival of individual worms was monitored in the presence of a defined concentration of bacteria [20
]. The survival assay was performed in 96-well plates on five occasions (runs). During each run, all possible bacterial and worm strain combinations were assayed in parallel, resulting in a total of 16 data points per factor combination per run and 80 data points per factor combination in total.
80 out of a total 3840 cases (2.08%) had to be excluded because of errors during automated worm-transfer (either no worm or more than one worm per well), resulting in between 75 and 80 usable data points per combination of worm and bacterial strains. The number of valid cases did not differ significantly among these factor combinations (likelihood ratio test [LRT], χ2 = 0.904, d.f. = 35, P > 0.999). In the control (worms with dead S. marcescens Db11), only 12 out of 625 were not found in the category "alive" (1.92%). Of these, 9 were morbid and 3 were dead. The recorded number of live worms per strain did not differ significantly from 100% (LRT, χ2 = 0.455, d.f. = 7, P > 0.999). It also did not differ significantly among the worm strains (LRT, χ2 = 0.416, d.f. = 7, P > 0.999). These results show that the experimental set-up itself does not cause significant levels of dead or morbid worms and that it does not have a different effect on different worm strains.
The different C. elegans strains show substantial differences as to their ability to survive in the presence of pathogenic S. marcescens (Fig. ). In general, the strains MY6 and MY18 were most resistant, whereas MY14 and MY15 were most susceptible. Moreover, the strains with identical microsatellite genotypes generally produce similar but not identical levels of resistance. This suggests that these strains bear additional genetic differences, which were not resolved by microsatellite genotyping. At the same time, the different S. marcescens strains differ considerably in their effect on C. elegans (Fig. ). Here, strain Sm2170 was most virulent, whereas strains Sma3 and Sma13 generally produced the fewest cases of mortality and morbidity. Since S. marcescens strains were grown under identical conditions and since some of them are already known to differ in phenotype (e.g. red pigmentation), the observed differences are most likely determined genetically. Most interestingly, the interaction between specific worm and bacterial strains seems to differ across the table. For instance, C. elegans strain MY10 is more susceptible to S. marcescens strain Sma13 than to ATCC274, whereas the opposite is true for C. elegans strain MY20 (Fig. ). Similarly, host strain MY15 is more susceptible to pathogen strain ATCC274 than to strain Db11, whereas the pattern is reversed for almost all other host strains (Fig. ).
Figure 1 Treatment response for the different bacterial and worm strain combinations of the main experiment. The response is expressed as host condition (values for the whole experiment), such that the black area refers to the proportion of dead worms, grey to (more ...)
In general consistency with these observations, ordinal logistic regression (OLR) analysis indicates a significant effect of the factors bacterial strain, worm strain or genotype, the interaction between the two, and also experimental run on the treatment response (Table ). The two respective models employed are significantly better than models without any predictors (model including worm strain as factor: LRT, χ2 = 1285.63, d.f. = 199, P < 0.0001; model including worm microsatellite genotype as factor: LRT, χ2 = 854.55, d.f. = 99, P < 0.0001). However, they are both significantly worse than the respective saturated models (model including worm strain: LRT, χ2 = 491.32, d.f. = 199, P < 0.0001; model including worm genotype: LRT, χ2 = 442.94, d.f. = 99, P < 0.0001). The latter test examines whether the model employed considers a sufficient number of factors or factor combinations to explain the variation found in the data. The results suggest that the model is not sufficiently complex. We decided against employing more complex models (e.g. consideration of host genotype nested in host strain in a single model), because the response variable is ordinal with only three categories (alive, morbid, dead), such that a larger number of predictor variables in the model would most likely lead to highly increased random error in the regression analysis. Thus, as an alternative, we analysed the data using association tests.
Ordinal logistic regression analysis of the importance of different factors in the main experiment.
Two-way associations were analysed with the LRT. The results show a significant effect of either of the different factors on worm condition (Table ). The relevance of these associations was further examined by taking into account a second predictor variable using the Cochran-Mantel-Haenszel (CMH) test of conditional independence. All previously identified associations remained significant, irrespective of the second predictor variable considered (Table ). The only exception refers to the case where the factor worm strain was corrected by the factor worm genotype, suggesting that the observed variation among C. elegans strains is due to differences in genotypic composition. The remaining results indicate that the significant effect from one of the factors on the treatment response is independent of the significant effect from one of the other factors. This finding is consistent with the presence of an interaction effect from the factors bacterial strain and nematode strain/genotype, as above suggested by OLR.
Association analysis of the impact of different factors on worm condition in the main experiment .
In the second experiment, we specifically addressed the presence of an interaction between two bacterial strains (Db11, ATCC274) and four host strains (MY8, MY10, MY14, MY15), the latter belonging to two different host genotypes. For this experiment, all factor combinations were included in each 96-well plate and in one experimental run. Only 7 out of 384 cases had to be excluded for the reasons given above (1.82%), resulting in 46 to 48 data points per factor combination. Again, the number of valid cases did not differ among factor combinations (LRT, χ2 = 0.093, d.f. = 3, P = 0.996). In the control treatment of this experiment, all animals were alive.
The second experiment confirmed the presence of variation in host resistance and pathogen virulence, although the overall level of virulence was lower than in the main experiment (Fig. ). Subsequent OLR revealed a significant effect from the factor worm strain or worm genotype, and also the interaction between the bacterial strain and either worm strain or genotype. The effect of bacterial strains was significant before Dunn-Sidák adjustment of significance levels (due to multiple testing), but insignificant afterwards (Table ). For these OLR analyses, the models employed were significantly better than a model without any predictors (model including worm strain as factor: LRT, χ2 = 62.29, d.f. = 7, P < 0.0001; model including worm microsatellite genotype as factor: LRT, χ2 = 59.12, d.f. = 3, P < 0.0001). Moreover, they were not significantly worse than the respective saturated models (model including worm strain: LRT, χ2 = 4.62, d.f. = 7, P = 0.7060; model including worm genotype: LRT, χ2 = 1.32, d.f. = 3, P = 0.7238), suggesting that they contained sufficient details to explain the observed variation.
Treatment response for the different bacterial and worm strain combinations of the second experiment. The black area denotes the proportion of dead worms, grey the proportion of morbid, and white the proportion of live worms.
Ordinal logistic regression analysis of the importance of different factors in the second experiment.
Subsequent performance of association tests generally corroborated the OLR analyses: The different predictor variables had a significant effect on the treatment response (LRT analysis in Table ). With two exceptions, this was still true after correcting for one of the other predictors (CMH tests in Table ). One of the exceptions refers to the factor bacterial strain, which no longer produced a significant effect if corrected by any of the other factors. This is in agreement with results from the OLR analysis. The other case shows that the factor worm strain becomes insignificant if corrected by worm genotype, which confirms the findings for the main experiment (see above). Consequently, the results clearly demonstrate that there are significant differences among host genotypes and, most importantly, that there are significant strain- or genotype-specific interactions between the two species.
Association analysis of the impact of different factors on worm condition in the second experiment.