Parasites impose a permanent threat for hosts. As a consequence, immune defenses are important for host fitness. However, the immune response can also produce self-damage and impair host fitness if not properly regulated. Effectors that up- and downregulate the immune response should, therefore, evolve in concert, and be under the action of correlational selection. To address this issue, we assessed the shape of the selection operating on pro- and anti-inflammatory effectors following an inflammatory challenge in laboratory mice. We found that selection acts on the combination of these two traits as individuals that produced large amount of pro-inflammatory cytokines could achieve relatively high fitness (survival) only if also producing a large amount of anti-inflammatory effectors. To our knowledge, this is the first study providing evidence for correlational selection on immunity.

Parasites and pathogens are permanent threats for their hosts, and as such have promoted the evolution of defense mechanisms, including the immune system. Once infected, organisms deploy a complex array of physiological effectors whose ultimate function is to clear the infection (Kaufmann et al. 2002). Nevertheless, immune protection does not come without costs. During the last decade, the study of costly immune defenses has flourished. Ecological immunology has put immune defenses into the classical economical context that has been successfully applied to life-history traits (Sheldon and Verhulst 1996). Because immune functioning confers benefits in terms of parasite clearance but also incurs costs, natural selection is expected to minimize the ratio between costs and benefits (Zuk and Stoehr 2002; Schmid-Hempel 2003; Viney et al. 2005). The nature of the cost of immune defenses has been largely explored (Schmid-Hempel and Ebert 2003). Several studies have attempted to measure the metabolic cost of immune activation with somehow mixed results (e.g., Martin et al. 2003; Eraud et al. 2005), whereas others have looked at the trade-off between immune activation and other fitness components, such as reproduction or growth (e.g., Bonneaud et al. 2003; Soler et al. 2003).

Recently, however, it has become clear that, in addition to these metabolic costs, the immune response can potentially impair host fitness if it targets host cells and tissues. Even though immunopathology and autoimmunity are the object of a main focus in biomedical research (Sell and Max 2001), their ecology and evolution has been mostly overlooked (Graham et al. 2005). In this context, the immune system can be seen as a double-edged sword. One edge allows to get rid of the parasite and contribute to health. The other edge is the consequence of an uncontrolled immune response that could partly damage and ultimately kill the host or reduce its life span (Belloni et al. 2010).

Because of its nonspecific nature, the inflammatory process is particularly relevant as a source of immunopathology. Indeed, several diseases arise as collateral undesirable short-term or delayed outcomes of acute and chronic inflammation, especially in advanced age (Coussens and Werb 2002; see also Sorci and Faivre 2009, for a recent review). Inflammation is a general defense mechanism that involves the recruitment of cells (e.g., leucocytes) at the site of injury. Upon recognition of pathogen-associated molecular patterns (PAMPs) by pattern recognition receptors expressed on the surface of immune cells, the inflammatory response is initiated with the production of pro-inflammatory cytokines (i.e., TNF-α, INF-γ, IL-6) that further contribute to the recruitment and activation of macrophages and granulocytes (Sell and Max 2001). The overproduction of these effectors can be a threat to organismal homeostasis, because cytokine storms are the main determinants of symptoms associated with many infectious diseases and septic shock (Annane et al. 2005; Graham et al. 2005). To keep the inflammatory response at bay, monocytes and regulatory T-cells produce anti-inflammatory cytokines (i.e., IL-10) whose function is to resolve inflammation (Ouyang et al. 2011).

Pro- and anti-inflammatory effectors have been extensively studied in the context of health sciences (Antonelli 1999); their role during sepsis and particularly their ability to predict survival in response to an infectious threat have already been emphasized (Dinarello 1997; Osuchowski et al. 2006; Yende et al. 2008). However, to the best of our knowledge, there is no study that has attempted to measure the shape of selection acting on these immune traits. When facing an infection, the final outcome in terms of host fitness is likely to depend on the balance between pro- and anti-inflammatory responses, rather than on single traits. An imbalance toward an anti-inflammatory response would insufficiently control the pathogen, whereas an imbalance toward a pro-inflammatory response would lead to harmful immunopathological effects. This leads to the clear-cut prediction that selection should operate on character combinations, instead of acting on each trait separately. Selection for optimization of character combination is often referred to as correlational selection (Blows and Brooks 2003).

We conducted an experiment on mice to assess the shape of selection acting on pro- and anti-inflammatory cytokines produced in response to a severe inflammatory insult. We challenged mice with Escherichia coli lipopolysaccharide (LPS). LPS is a PAMP recognized by toll-like receptor 4, which leads to a strong but short and nonprogressive systemic inflammation (Saito et al. 2003). This experimental design allows assessing the fitness consequences of immune effectors produced in response to the challenge without the potential confounding effect of replicating parasites. This procedure has been often adopted in ecological immunology studies (Bonneaud et al. 2003; Velando et al. 2006; Palacios et al. 2011).

Material and Methods


One hundred and fifty adult virgin SWISS mice (50 males and 100 females) were used for this study. Fifty of them were purchased from Janvier (Laval, France), whereas the remaining mice were born in the animal facility at the Université de Bourgogne, produced by mice gathered from the same provider, and issued from 26 full-sibling families. Male mice were housed individually, whereas females were housed in groups of four in Plexiglas cages and kept under standardized conditions (temperature 21 ± 1°C, relative humidity 60 ± 10%) with a 12-h light/dark cycle. Pellet food and tap water were provided ad libitum.


When three-month old, mice were exposed to an inflammatory challenge. At day 0, in the morning, mice were weighed (± 0.1 g) and a blood sample (100–150 μl) was collected by retroorbital puncture under isoflurane anesthesia and kept at 4°C. Blood was rapidly centrifuged (4000 rpm, 15 min, 4°C) and plasma was stored at −80°C until cytokine assay (see next section). At day 1, female mice received an intraperitoneal injection of E. coli LPS (16.7 mg/kg in 100 μl of PBS) (LPS, serotype 055:B5, Sigma, St. Louis, MO) (Tateda et al. 1996). Given that males are known to be more sensitive to septic shock (Marriott and Huet-Hudson 2006), we injected them with a slightly smaller dose (13.9 mg/kg in 100 μl of PBS). Three hours after the LPS injection (corresponding to the peak of circulating IL-6 and IL-10; Tateda et al. 1996), mice were weighed and bled again. Blood sampling took place in the morning as for day 0. Blood samples were centrifuged and plasma stored as mentioned above. Survival was monitored every 12 h and body mass measured every day during one week postinjection.

To make sure that any mortality and increase in cytokine production was due to the inflammatory challenge, we used an additional group of 24 mice as a control. These individuals were treated as the experimental mice with the exception that they were injected with 100 μl of PBS, instead of LPS.

Three mice that did not respond to the LPS challenge (both in terms of change in body mass and circulating levels of IL-6 and IL-10), possibly because of injection failure, were excluded from the analyses.

The experiment has been conducted in compliance with and has received the agreement of the Animal Care and Ethical Committee of the Université de Bourgogne, Dijon (protocol B1510).


We quantified plasma levels of two cytokines, IL-6 and IL-10. IL-6 is a good marker of infectious stress and is particularly important during severe septic shock. It is also often used as a marker for systemic activation of the inflammatory response (Barton 1997). IL-10 is an anti-inflammatory cytokine involved in the resolution of inflammation (Barsig et al. 1995; Couper et al. 2008). Circulating IL-6 and IL-10 were quantified by flow cytometry (Demas et al. 2011) using the Cytometric Bead Array (CBA) Mouse Cytokine Flex Set kit (BD Biosciences, San Diego, CA) according to manufacture instructions and following previous work (Prunet et al. 2006). IL-6 and IL-10 kits consist of beads (diameter: 7.5 μm; excitation and emission wavelengths at 488 and above 600 nm [FL3], respectively) dyed to two different fluorescence intensities. Each particle is coupled to antibodies binding a specific cytokine (IL-6 or IL-10) and represents a discrete population, unique in its FL3 intensity. The captured cytokines are detected via a direct immunoassay using a specific antibody coupled to PE emitting at 585 nm (FL2). Data were acquired with FlowMax software (Partec, Munster, Germany) on a GALAXY flow cytometer (Partec) equipped with a laser emitting at 488 nm, and analyzed using BD CBA software (BD-Biosciences). Forward versus side scatter gating was employed to exclude any sample particle other than the 7.5 μm beads. Data were displayed as two color dot plots (FL2 [PE]: band pass 580 ± 10 nm vs. FL3 [beads]: long pass 665 nm) so that the discrete FL3 microparticle dye intensities were distributed along the y-axis. Ten point standard curves ranging from 20 to 5000 pg/mL were obtained by serial dilution of the reconstituted lyophilized standard. The lower limits of detection were 1.4 pg/mL for IL-6 and 9.6 pg/mL for IL-10.


We used a well-established statistical framework to estimate the strength and mode of selection (linear and nonlinear selection) and to visualize it (Brodie et al. 1995; Blows and Brooks 2003). This multiple regression approach has been extensively used in evolutionary ecology studies (Kingsolver et al. 2001).

Following Lande and Arnold (1983), we regressed relative fitness (the individual survival divided by the mean survival of the LPS-injected population) on standardized log-transformed cytokines (mean = 0 and standard deviation = 1). Lande and Arnold (1983) defined the selection gradients as the “slope of the straight line that best describes the dependence of relative fitness on character zi, after removing the residual effects of other character on fitness.” The slopes of the linear multiple regression provides the selection gradients, whatever the distribution of the fitness measurement and the measured traits (Lande and Arnold 1983).

To assess selection gradients, we used a mixed model that included linear, quadratic, and interaction terms as described in equation (1) (Blows and Brooks 2003)


where w is relative fitness, α is the intercept, zi are metric traits, and βi, γii, γij are linear, quadratic, and correlational gradients. Quadratic selection gradients were doubled as recommended by Stinchcombe et al. (2008). In addition to the two cytokines, the initial model also included body mass at day 0, sex, and the origin of the mice (locally born or purchased from a provider), whereas family identity was declared as a random effect.

In many cases, however, fitness measurements are based on data with nonnormal distribution (binomial for survival or mating success, Poisson for lifetime reproductive success) and the P-values associated with the slopes of the linear regression are unreliable. For this reason, it is a common procedure to use models with the appropriate distribution of errors to infer the statistical significance of the selection gradients provided by the linear regression model (see for instance Råberg and Stjernman 2003). We therefore used a generalized linear mixed model with a binomial distribution of errors, a logit link function, and family identity declared as a random factor, to assess the P-values associated with the selection gradients. Nonsignificant terms (sex, initial body mass, origin of the mice) were removed until reaching the minimal adequate model. We used Wald Z and likelihood ratio test to infer the statistical significance of fixed and random effects, respectively (Bolker et al. 2009). We also checked for the overdispersion because Wald Z should be used only if data are not overdispersed. We computed the ratio between residual deviance and residual degrees of freedom (Crawley 2007) and found a ratio close to one (1.07), suggesting that the data were not overdispersed. Nevertheless, we also ran the model using a quasi-binomial distribution of errors and found consistent results, further suggesting that data overdispersion did not bias the statistical significance. In addition, to further assess if the error distribution of the data was correctly modeled, and to detect any departure from the model assumptions, we used a diagnostic procedure described by Zuur et al. (2009). The procedure consists in the graphical examination of the quantile–quantile plot (Zuur et al. (2009). The quantile–quantile plot is the graph of quantiles of residuals assuming the fitted model is the true model, against the actual quantiles of the residuals from the fitted model (Zuur et al. 2009). In the absence of major deviations from the model assumptions, fitted and simulated quantiles should lie on the equality slope (1:1). In agreement with this, we found that fitted and simulated quantiles were very close to the 1:1 line and within the simulated 95% pointwise confidence interval (see Fig. S1). This suggests that our logistic mixed model correctly fitted the data with no major departures from the model assumptions.

Changes in body mass during the course of the experiment were analyzed using a mixed model with a normal distribution of errors. Time postinjection, squared time, treatment (LPS vs. PBS), and sex were included as fixed factors, whereas mice identity and family were declared as random factors. This analysis only included mice that were still alive at the end of the experiment (one week postinjection) to avoid a possible effect of selective disappearance.

All the statistical analyses were performed using R software version 2.13.1 (R Development Core Team 2011). Three-dimensional visualization of the fitness surface was done using a cubic spline fitting function with the software STATISTICA (StatSoft 1999).


Over the course of the experiment 51% of the mice died. Eighty four percent of the total mortality occurred during the 72 h that followed the LPS injection (Fig. 1).

Figure 1.

Time course of mouse mortality after an LPS challenge (solid line). Dashed lines around the solid line represent 95% confidence interval. Most of the mortality occurred during the first three days after the inflammatory challenge. Mice injected with PBS (broken line) had a 100% survival during the same time period, showing that LPS was the specific cause of the observed mortality.

Baseline levels of IL-6 and IL-10 were very low and below the detection threshold of the flow cytometer, indicating that none of the individuals used for this study suffered from inflammatory disorders prior to the immune challenge. Three hours postinjection mean (± SE) plasma levels of IL-6 and IL-10 were 139.94 ± 4.36 ng/mL and 188.4 ± 25.2 pg/mL, respectively. IL-6 titers closely approximated a normal distribution, whereas the distribution of IL-10 was skewed to the right (Fig. 2). Controls mice that received an injection of PBS did not show any increase in IL-6 and IL-10 as the concentrations 3 h postinjection were all below the detection threshold. None of the PBS-injected mice died during the course of the experiment (Fig. 1).

Figure 2.

Frequency distribution of IL-6 (A) and IL-10 (B) titers.

The analysis of selection gradients indicated that relative fitness was a negative linear function of IL-6 production (β± SE =−0.40 ± 0.076, P < 0.001, Table 1). Neither the linear nor the quadratic gradients for IL-10 reached statistical significance (all Ps > 0.1, Table 1). However, the correlational selection gradient was statistically significant, suggesting that selection operates on the combinations of the two traits (γ′± SE = 0.15 ± 0.078, P= 0.039; Table 1). To visualize the correlational selection, we drew the cubic spline fitness surface for the two cytokines (Fig. 3).

Table 1.  Generalized linear mixed model with a binomial distribution of errors exploring the association between survival after an LPS challenge and pro- (IL-6) and anti-inflammatory (IL-10) cytokines (linear, squared, interaction terms), sex, body mass at the day of LPS challenge and the origin of the mice. Family identity was included as a random factor. The table reports the initial model as well as the simplified one after removal of the nonsignificant sex, body mass, and origin terms. The significance of the random factor was assessed using a likelihood ratio test. Sample size is 147 mice.
Source of variationdfWald Z P
Initial model
 IL-6 1 −4.51 <0.001
 Squared IL-6 1 −0.006 0.995
 Squared IL-1010.660.511
 IL-6 × IL-10 1 2.17 0.030
 Body mass 1 −0.62 0.536
 Mice origin1−1.060.270
 Random effect df χ2 P
Final model
 IL-10 1 0.73 0.466
 Squared IL-61−0.370.710
 Squared IL-10 1 1.57 0.116
 IL-6 × IL-1012.060.039
 Random effect df χ2 P
Figure 3.

Adaptive landscape for IL-6 and IL-10 produced in response to an LPS challenge in mice. (A) Raw data; (B) the fitness surface (cubic spline) illustrates the nonlinear, correlational selection acting on the two cytokines.

LPS-injected mice lost body mass during the first days postinjection (mean variation in body mass ± SE =−16.30%± 2.5%) compared to PBS-individuals (mean ± SE =−2.70%± 0.48%). However, this effect was transitory since at the end of the experimental one-week period LPS-mice had almost recovered to their initial body mass (mean ± SE =−3.86%± 0.85%). This resulted in a highly significant interaction between treatment (LPS vs. PBS) and squared time postinjection (mixed model: t1,563=−10.09; P < 0.001). There was also a highly significant interindividual variation in body mass change (likelihood ratio test, χ21= 520.91, P < 0.0001), whereas the family effect only approached significance (likelihood ratio test, χ21= 2.95, P= 0.086).


This study aimed at experimentally assessing the shape of the selection acting on immune traits involved in the promotion and resolution of inflammation. To this purpose, we measured both pro- (IL-6) and anti-inflammatory (IL-10) effectors and investigated the fitness (survival) consequences of variable investment into the two functions in mice exposed to an inflammatory challenge.

Our experimental design was successful in eliciting an inflammatory response since values for both cytokines raised 3 h after the LPS challenge, whereas for control mice injected with PBS, cytokine levels remained low (below the detection threshold of the flow cytometer). Incidentally, it should be noted that baseline (preinjection) cytokine levels were virtually equal to zero, further indicating that mice did not suffer from any inflammatory disorder before the challenge. In addition to this, the majority of the mortality occurred within the first three days following the exposure, in agreement with the view that mortality was due to cytokine storms and septic shock (Remick et al. 2002; Xiao et al. 2006). Overall, these findings corroborate our assumption that mice exposed to LPS mounted an inflammatory response comparable to the one observed during a severe bacterial infection causing sepsis (Osuchowski et al. 2006).

LPS-injected mice lost body mass during the first days postinjection, whereas PBS controls did not show any time-related change in body mass. Body mass loss was, however, transitory since by the end of the one-week experimental period, LPS mice had almost recovered their initial body mass. Because this analysis was restricted to animals that were still alive at the end of the experiment, it cannot be due to selective mortality. Anorexia and weight loss are well-known symptoms induced by exposure to LPS and contribute to the definition of the so-called sickness behavior (Inui 2001).

Using the statistical framework put forward by Lande and Arnold (1983), we found that relative fitness was a negative linear function of the pro-inflammatory IL-6. Mice that produced high amount of IL-6 had a very poor survival prospect. This result confirms previous work conducted on humans and rodents showing that plasma IL-6 is a very good predictor of survival during sepsis (Casey et al. 1993; Remick et al. 2002; Osuchowski et al. 2006; Bozza et al. 2007). Interestingly, the linear selection gradient for IL-10 was not statistically significant, neither were the quadratic gradients for the two cytokines. These results suggest that selection does not operate on IL-10 independently of other traits and that intermediate levels of IL-6 and IL-10 are not associated with improved fitness (absence of stabilizing selection), under the specific conditions used for our experiment. On the contrary, we found evidence for a correlational selection gradient. The visual inspection of the adaptive landscape, based on a spline function, revealed first that maximum fitness was achieved for minimum values of IL-6 and IL-10. Second, when IL-10 production was low, any increase in IL-6 dramatically reduced fitness. A valley in the adaptive landscape was indeed reached for individuals with high IL-6 response and low IL-10 production. Third, the condition to achieve high fitness when overproducing IL-6 was to massively produce IL-10 as well. Therefore, under the specific conditions of our experimental design, our results suggest that selection might operate as to optimize the combined response of pro- and anti-inflammatory effectors.

For the sake of simplicity, we used LPS as a proxy of bacterial infection. Of course, using replicating parasites would make the picture more complex, because a robust pro-inflammatory response could confer a benefit in terms of parasite clearance. In this case, maximum fitness would not be achieved for the lowest IL-6 level. Using live parasites might, however, also confound the pattern because in addition to the mortality due to a disregulated inflammatory response, live pathogens might have induced some immune-independent mortality (due to a direct spoliation effect). Obviously, the next step would be to replicate the present study using live bacteria. It is also important to fully acknowledge that the use of LPS is not the only artificial aspect of our study. We used laboratory mice kept under controlled conditions and assessed fitness as the short-term survival following the LPS challenge. Future work should definitely extend this approach to natural populations of both short- and long-lived species, integrating both viability and reproductive fitness. As mentioned above, because we did not use live pathogens, we could not assess the benefits of the inflammatory response in terms of parasite clearance. Nevertheless, we believe that our study still fits into a cost–benefit framework, as we investigated the cost of the inflammatory response and the benefits of a proper immune regulation.

In spite of using laboratory mice under controlled environmental conditions, we found some interindividual variation in the response variables. SWISS mice are relatively outbred compared to other laboratory mouse strains and it is therefore possible that standing genetic variation contributes to explain some of this phenotypic variation. In addition to genetic sources, other environmentally based sources of variation can persist even though temperature, food quality, and availability were identical for all individuals. This is a general remark that applies to all studies involving congenic strains of mice where interindividual variation in phenotypic traits persists in spite of presumed low genetic variation. Incidentally, even though our experiment was not designed to explore the sources of phenotypic variation, we found that the response to the LPS injection had a highly significant family effect for survival and a marginally nonsignificant effect for changes in body mass. Because mice were produced by full-sibling family and shared the same maternal environment, we cannot identify the precise sources of this among-sibling resemblance.

Currently available evidence is in agreement with the results reported here and suggests that correlational selection probably acts on pro- and anti-inflammatory effectors, even though such evidence mostly comes from studies on humans. Previous work on human malaria already supported the idea that cytokine balance, and particularly the balance between IL-6 and IL-10 (Day et al. 1999), is important for resolving infections without severe pathology (Othoro et al. 1999; Angulo and Fresno 2002; Dodoo et al. 2002;Artavanis-tsakonas et al. 2003). Other clinical studies also reported a link between high ratio of pro- and anti-inflammatory effectors and poor outcomes in patients with systemic inflammatory response syndrome (Taniguchi et al. 1999; Loisa et al. 2003). Recently, signs of selection acting on cytokine genes have been reported by Kuningas et al. (2009). They looked at the occurrence of IL10 haplotypes across age classes in a population living in an agricultural area of the Upper-East region of Ghana. This area is endemic for several infectious diseases (including malaria, typhoid fever, and helminth infections), and therefore selection for parasite resistance is supposed to be strong. In agreement with this prediction, the authors found that IL10 haplotypes associated with a high pro- to anti-inflammatory cytokine ratio were enriched among elderly, suggesting a positive selection for pro-inflammatory defenses in a context of serious infectious threat. Interestingly, however, people who had access to safe water sources had an opposite trend, with pro-inflammatory IL10 haplotypes being less represented in old age classes. Overall, these lines of evidence are in agreement with the idea that selection acts in concert on pro- and anti-inflammatory effectors, with the environmental infectious risk having a major modulatory effect.

Correlational selection is supposed to be a widespread force acting on metric traits, even though it has been a difficult task to assess it in natural settings, the need of large sample size to have enough statistical power to detect interactive effects being probably one of the major constraints (Blows and Brooks 2003). Previous studies that have addressed the shape of selection acting on immunity have focused on single traits and have overlooked the complex interactions between up- and downregulatory effectors. Interestingly, the majority of these studies have reported significant positive directional selection on immune traits: antibody production (Råberg and Stjernman 2003), encapsulation rate (Rantala et al. 2011), phenoloxidase activity (Rolff and Siva-Jothy 2004), and E. coli bacterial killing capability (Wilcoxen et al. 2010). Evidence for stabilizing selection has been reported by a single study on blue tits (Cyanistes caeruleus) where individuals with intermediate primary antibody responsiveness to a diphtheria vaccine had the best survival prospect (Råberg and Stjernman 2003).

A particularly interesting result has been reported by Svensson et al. (2001) who measured antibody responsiveness in female side-blotched lizards (Uta stansburiana) immunized against the tetanus toxoid. They found that the shape of the selection on antibody production did depend on the phenotype of the female, with the immune response of yellow-throated females being positively and directionally selected, whereas the directional selection gradient for orange-throated was negative. Svensson et al. (2001) reported therefore significant correlational selection between color morphs and antibody responsiveness. Given that color morphs in this lizard species are associated with a wide range of behavioral, physiological, and life-history traits (Sinervo et al. 2001), it is possible that this correlational selection also involves other immune traits.

Correlational selection should promote the evolution of linkage disequilibrium and epistatic interactions, finally favoring the emergence of genetic correlations between traits (Sinervo and Svensson 2002). Current evidence, indeed, suggests that epistatic interactions exist between IL6- and IL10-related single-nucleotide polymorphism on the incidence of inflammatory diseases (i.e., Alzheimer's disease) (Combarros et al. 2009). Correlational selection can thus be a powerful engine maintaining genetic variation at loci involved in immune regulation, and perhaps can contribute to explain some of the long-lasting unsolved questions in immunology, such as the evolution of the Th1/Th2-polarized cells.

Associate Editor: A. Read


We are very grateful to Line Prezioso for managing the animal husbandry. We are also grateful to the associate Editor A. Read and two anonymous referees for their valuable comments and suggestions. This work has received financial support from the Région Bourgogne to AAB, BF, and GS and from the ANR (Program EVOREGIM) to BF and GS.