Microarray Evaluation of the Listeria monocytogenes Infection and Amoxicillin Treatment in Mice

Abstract

By using Affymetrix Mouse Genome Arrays and 20 biological replicates per experimental condition, the predictive value of liver and blood gene expression profiles previously identified was validated as predictive of Listeria monocytogenes infection severity (lethal and nonlethal infection). The ability of these genes to predict the outcome of antibiotic treatment was also assessed. Lethally infected BALB/c mice were treated with amoxicillin at 10 or 20 mg/kg; only the higher dose prevented death. The liver genes predicted that 70% of the animals treated at 10 mg/kg, but only 25% of the mice treated at 20 mg/kg, belonged to the lethal infection group, and this prediction was similar to the ultimate mortality outcome. These results confirm the value of microarrays as tools to predict host response to infection and efficacy of antibacterial therapy. These results might lead to applications that would help clinicians to adjust antibiotic dosages for efficient treatment but yet without toxicity.

Recent advances in genomics have created an impetus for using DNA microarrays for the identification of biomarkers. These gene expression profiles, or markers, find application in several disciplines, including cancer biology, toxicology, and infectious disease research. In cancer biology, microarrays have allowed the identification of genes that could predict leukemia type (Golub et al. 1999), the prognosis in breast cancer patients (van de Vijver et al. 2002), the outcome in mature B-cell malignancies (Dave 2006), and the development of cancer in ulcerative colitis patients (Watanabe et al. 2007). In toxicology, microarrays are used to understand the molecular mechanisms of toxicity and thus provide greater definition than current indicators for classifying toxicants according to health risk (Suter, Babiss, and Wheeldon 2004). In addition, toxicogenomics has been used successfully to identify compound-specific gene expression changes that might serve as markers of drug toxicity, such as nephrotoxicity (Amin et al. 2004) and hepatotoxicity (Minami et al. 2005).

Microarrays are also valuable tools not only for direct pathogen identification (Wang et al. 2002; Palmer et al. 2006), but also for indirect pathogen recognition. In the latter case, pathogen-specific host gene expression signatures are identified that can be used as surrogate markers for diagnostic purposes (Campbell and Ghazal 2004). This strategy recently allowed the identification of biomarkers able to distinguish between human immunodeficiency virus (HIV)-1–seropositive and –seronegative patients (Ockenhouse et al. 2005) and between the lethal and the nonlethal blood stage of Plasmodium yoelii malaria infections (Schaecher et al. 2005). With the growing threat of bioterrorism, it has become vital to quickly identify individuals who are ill as a result of a bioagent and to predict the severity of the illness (Lin et al. 2003). Because of their ability to simultaneously monitor the expression levels of thousands of genes, microarrays can be used for comparing clinically relevant groups (e.g., healthy versus diseased people), uncovering new subclasses of diseases, and predicting clinically important outcomes, such as the response to therapy and survival (Tarca, Romero, and Draghici 2006).

We have previously identified, in the liver and the blood of Listeria monocytogenes (L. monocytogenes)–infected mice, gene expression profiles (markers) able to predict the severity of the infection (Ng et al. 2005). L. monocytogenes is a gram-positive bacterium and the causative agent of the food-borne illness listeriosis. This disease, characterized by severe gastroenteritis, central nervous system (CNS) infections, and mother-to-child transmission, has an overall mortality rate of 30% (Roberts and Wiedmann 2003). L. monocytogenes possesses a variety of molecular determinants (Dussurget, Pizarro-Cerda, and Cossart 2004) and virulence mechanisms, including subversions of several cellular host functions. (Pizarro-Cerda and Cossart 2006), that allow successful establishment of the infection. Within 10 min of systemic administration, the majority of the L. monocytogenes accumulate in the liver and spleen. In the liver, hepatocytes are the primary site of multiplication of the bacteria (Gregory et al. 1992). L. monocytogenes is also one of the most well-studied bacterial pathogens; the mouse model of listeriosis has been widely used to understand the general mechanism of infection and to gain insights into the biology of the mammalian immune system (Garifulin and Boyartchuk 2005).

In the study described here, we confirmed the predictive ability of these liver and blood genes using a newer version of the Affymetrix Mouse Genome Array and a sample size of 20 animals per condition. In addition, we evaluated whether these genes could also predict the outcome of antibiotic treatment in infected mice. Our data showed that toxicogenomics approach can be useful for biomarker identification, prediction of drug treatment and toxicity.

MATERIAL AND METHODS

Bacterial Strain Culture Condition

Listeria monocytogenes strain 19303, obtained from Dr. Kimber White (Medical College of Virginia, Virginia Commonwealth University, Richmond, VA, USA), was grown in brain-heart infusion broth (BHI) (BD, Franklin Lakes, NJ, USA) and cultured overnight at 35°C under agitation. Bacterial concentration was determined by measuring the optical density at 600 nm. Culture was subsequently adjusted to 2 × 10³ and 7 × 10⁴ colony-forming units (CFU)/200 μl and aliquoted into individual syringes for immediate administration to the animals.

Animals

Six- to 8-week-old female BALB/c mice (Charles River Laboratories, Wilmington, MA, USA) were maintained on Purina Certified Rodent Chow 5002 (Richmond, IN, USA) and purified tap water ad libitum in microisolator cages under controlled lighting (12-h light/dark cycle). All animals were treated in accordance with a protocol approved by the Institutional Animal Care and Use Committee (IACUC).

Experimental Designs

Groups of mice were infected intravenously (IV) (tail vein injection) with 200 μl of a nonlethal (NL; 2 × 10³ CFU) or a lethal (L; 7 × 10⁴ CFU) dose of L. monocytogenes. Control mice were administered an equal volume of BHI. After 24 h, lethal-dose Listeria– infected mice were orally administered amoxicillin (AMX) at 10, 20, 50, and 100 mg/kg (Sigma, St Louis, MO, USA) or an equivalent volume of phosphate-buffered saline (PBS). Mice were sacrificed on day 3, 24 h after AMX administration, and 48 h after Listeria infection.

For microarray analysis, livers were cut into 20- to 50-mg pieces, immediately flash frozen in liquid nitrogen, and stored at –80°C until processed for RNA extraction. Blood was collected via retro-orbital sinus on EDTA, and 275 μl of blood was then transferred into 700 μl PAXgene (Qiagen, Valencia, CA, USA), mixed, and incubated for 2 h at room temperature before being stored at –80°C until RNA extraction.

For infection burden determination, the left lobe of the liver collected at sacrifice was weighed and placed in a 5-ml homogenizing tube kept on ice until further processing. The livers of five animals per group were processed for CFU determination.

Infection Burden Determination

Liver samples were homogenized using a Heidolph homogenizer with Potter-Elvehjem Tissue Grinders. The homogenate was diluted and plated in duplicate onto trypticase soy agar plates containing 5% sheep blood (BD). After 24 h of incubation at 37°C, bacterial colonies were counted and expressed as CFU/mg liver.

Gene Expression Analysis

Total RNA Extraction

Approximately 20 to 50 mg of each of the 100 liver tissue samples was transferred into 1 ml Trizol reagent (Invitrogen, Carlsbad, CA, USA) along with a stainless steel bead (Qiagen) and homogenized with a TissueLyser (Qiagen). Total RNA was extracted with 200 μl chloroform, precipitated with 500 μl 100% ethanol, and transferred to a RNeasy mini spin column (Qiagen) for purification according to the manufacturer’s instructions.

Total RNA extraction from the 100 blood samples in PAX gene reagent was performed according to the manufacturer’s instructions (Qiagen) with the following modifications: washing the blood cell pellet with 500 μl RNase-free water and using 180 μl buffer BR1, 150 μl buffer BR2, and 20 μl proteinase K. In both cases, total RNA was stored at –70°C, and integrity was examined by agarose gel electrophoresis and spectrophotometric analysis.

Affymetrix GeneChip Protocols

cDNA was prepared and cleaned from 1 to 2 μg total RNA using the One-Cycle cDNA Synthesis and Clean-up module kits (Affymetrix, Santa Clara, CA, USA). The cDNA was further processed to biotin-labeled cRNA using the IVT labeling kit (Affymetrix). The cRNA was fragmented before hybridizing to the GeneChip Mouse Genome 430 2.0 Array (Affymetrix) and scanning (scaling and normalization factors = 1.0). All procedures were performed according to the manufacturer’s instructions, as described in the Affymetrix Genechip Expression Analysis Technical Manual, revision 5 (http://www.affymetrix.com/support/technical/manual/expression_manual.affx). From the 100 animals used in this experiment, 100 and 73 liver and blood samples were successfully processed.

Data Analysis

The Affymetrix CEL files were loaded into GeneSpring GX version 7.3.1 (Agilent Technologies, Santa Clara, CA, USA) and preprocessed using GenChip Robust Multi-Array analysis (GC-RMA) (Wu and Irizarry 2004). Two independent experiments were conducted; one containing the 100 liver samples and one containing the 73 blood samples. The per-gene normalization was applied across all of the samples of each experiment to normalize the expression level around the value of 1. In each case, probes with reliable measurements were selected from the 45,101 original probes by performing four successive levels of quality control evaluations (Table 1). This approach allowed the identification of 5399 and 1271 well-measured probes in the liver and blood samples, respectively.

Using the software GeneSpring GX, the following analyses were performed: (1) clustering analysis on conditions and genes using Pearson correlation and average linkage clustering; (2) supervised class prediction analysis (cross-validation and test set prediction) with the support vector machine (SVM) method using the polynomial dot product (order 3) as Kernel function, and a diagonal scaling factor = 3; and (3) principal component analysis on conditions with mean centering and scaling.

RESULTS AND DISCUSSION

The Previously Identified Gene Markers Predict the Severity of Listeria Infection

We previously identified 8 genes in the liver and 14 genes in the blood of Listeria-infected mice that were able to predict the severity of the infection (Ng et al. 2005). We undertook to confirm the predictive ability of these genes using a new version of the Affymetrix mouse GeneChip microarray. For this purpose, groups of 20 BALB/c mice were intravenously infected with nonlethal (NL; 2 × 10³ CFU) or lethal (L; 7 × 10⁴ CFU) doses of L. monocytogenes. Control mice were administered sterile bacterial culture medium. A large number of biological replicates is known to increase the ability to detect biologically significant gene expression differences (Wei, Li, and Bumgarner 2004) and the power of the statistical analysis (Jorstad, Langaas, and Bones 2007).

To first assess the global gene expression profile in response to infection, we performed a hierarchical clustering analysis using the 5399 and 1271 well-measured probes (see Material and Methods) across 60 and 43 samples of liver and blood, respectively (Figure 1). For the liver, all 20 samples from each experimental group clustered together except for a single NL and a single control sample, which clustered with the samples from the control group and the NL group, respectively (Figure 1A ). For the 1271 well-measured probes identified in the 43 blood samples, the clustering analysis also revealed high reproducibility and consistency in the host response to the infection, and low animal-to-animal variation (Figure 1B ). Although all 10 samples from lethally infected mice clustered together, only 2 (of 16) NL samples and 2 (of 17) control samples clustered with samples from a different treatment group. Our hierarchical clustering analysis could thus provide an accurate estimate of the (very low) biological variability within the groups, an essential factor for determining whether the between-group differences are meaningful (Wei, Li, and Bumgarner 2004).

Of the 8 previously identified predictive genes in the liver, 7 were present among the 5399 well-measured probes: Cxcl9, Cyp17a1, Ifi47, Tgtp, Asl, Saa3, and 2310042G06rik/Slm02. Two probes representing Cxcl9 were present on the arrays, and both were included in our analysis, for a total of eight probes. To validate the ability of these eight probes to predict the severity of a Listeria infection, cross-validation analysis was performed using the 60 samples from the control, NL, and L treatment groups as training and test sets. This analysis returned 58 correct predictions (97%) and 2 incorrect predictions (3%). As described in Table 2, the incorrect predictions were one control sample (5%) assigned as NL and one NL sample (5%) assigned as control. Both these samples were previously shown to cluster outside their actual respective administration group (Figure 1). Hierarchical clustering analysis using these eight predictive probes demonstrated that the 60 liver samples presented a clustering pattern similar to that found using the global gene expression profiles (Figure 2A ).

Similar results were obtained with the 3 (of 14) previously identified blood predictive genes present among the 1271 well-measured probes in the blood samples: Anxa2, Plac8, and Igfbp4. The Igfbp4 gene was represented by four probe sets of similar normalized intensity value on the microarrays, and we chose two representative probe sets of Igfbp4 in our analysis. Using the cross-validation analysis described above, we found that these four probes (three genes) allowed accurate classification of 81% (35/43) of the samples into the appropriate experimental group. The results, presented in Table 2, showed that the analysis accurately predicted the classification of 82%, 69%, and 100% of the 17 control samples, 16 NL samples, and 10 L samples, respectively. When we performed hierarchical clustering analysis, these four probes allowed most of the 43 samples to cluster efficiently in the proper administration group (Figure 2B ).

Our earlier study (Ng et al. 2005) was performed on samples collected 6 h after the infection, whereas our current study involved samples collected 2 days after the infection. Despite this difference, our predictions of the severity of the Listeria infection were equally accurate (97% and 81% for the liver and blood samples, respectively), indicating that these predictive genes are consistent, even across different experiments using different array platforms.

Gene Markers Predict Antibiotic Treatment Efficacy in Lethally Infected Mice

The ability of these markers to efficiently discriminate animals infected with a lethal or nonlethal dose was based on differences in the host response to the infection, as shown in Figure 1. The best predictor genes have large between-group variability, but small within-group variability. This between-group variation was not only observed among the predictive genes, but also among the 5399 well-measured liver genes: 91% (4932/5399) of the probes were significantly differentially expressed (p < .05) between the nonlethal and the lethal treatment groups (data not shown). We hypothesized that this differential host response was associated with the different number of colonizing bacteria. Differential host responses to dose escalation have been reported in vitro in human peripheral blood mononuclear cells stimulated with bacteria or bacterial products at various concentrations (Boldrick et al. 2002). The authors of this study concluded that the graded nature of the response implied that the number of microbial cells used to stimulate immune cells determines the level or probability of response in the latter. If this hypothesis is true, it should be possible to predict the outcome of the treatment of the Listeria infection: when the number of colonizing bacteria is reduced with antibiotic therapy, the transcriptional profile of the host would shift and become similar to the profile of a host infected with low (nonlethal) doses of bacteria or none at all. To test this hypothesis, we treated lethally infected mice with amoxicillin (AMX), and assessed whether the gene markers able to predict the severity of Listeria infection would also have a value in predicting the efficacy of antibacterial treatment. AMX was chosen because it is one of the first-line antibiotics used against Listeria infection in humans (Hof 2003). In addition, another penicillin, ampicillin, was used to successfully treat Listeria infection in mice (Porter and Harty 2006; Williams and Bevan 2004).

Groups of 20 BALB/c mice were infected with a lethal dose of Listeria and, 24 h later, treated with AMX at 10 or 20 mg/kg (L + AMX10 and L + AMX20) or PBS as a control. These AMX concentrations were chosen based on the analysis of the effects of 10, 20, 50, and 100 mg/kg doses on the clinical outcome of a lethal Listeria infection (Table 3). All control mice died on day 3 (2 days postinfection). Treatment with AMX at 10 mg/kg delayed lethality for up to a week, but did not prevent infection-induced death. However, mice receiving AMX at 20 to 100 mg/kg survived until their scheduled sacrifice on day 21. These results suggest that treatment with AMX at 10 and 20 mg/kg differentially affected the clinical outcome, even though the difference in dose concentration was only twofold. We therefore analyzed the host gene expression response associated with such differential clinical outcomes after AMX treatment in these dose groups.

Gene Expression Profiles Correlate with Changes in Bacterial Load

To first obtain a general appreciation of the effect of AMX treatment on the global host transcriptional response, we used principal component analysis (PCA) to decompose the complex gene expression profiles of the 5399 well-measured liver probes for all five experimental conditions (control, NL, L, L + AMX10, and L + AMX20) (Figure 3). PCA confirmed that gene expression profiles for animals in the control, NL, and L groups were quite distinct from each other, and within-group variation was low. It also revealed that most of the L + AMX10 samples displayed a transcriptional profile more similar to that of the L samples (without AMX treatment), whereas the L + AMX20 samples constituted a heterogeneous group in which some samples were more similar to samples for the NL treatment condition.

We next attempted to determine the number of probes in which expression levels differed by at least 1.5-fold between treatment groups (NL, L, L + AMX10, L + AMX20) and the control group. The results, presented in Figure 4A , demonstrated that as the dose of AMX increased, there was a decrease in the total number of probes whose level of expression changed. Moreover, the number of treatment-related changes in the probes gradually decreased as the severity of the infection decreased from L to L + AMX10, to L + AMX20, to NL. We also observed a good correlation (r = .95) between the number of changed probes (Figure 4A ) and the number of colonizing bacteria in the liver (Figure 4B ) (data not shown). Across the three lethally infected conditions (L, L + AMX10, and L + AMX20), most of the changed probes (≥92%) were consistently up- or down-regulated (data not shown). Altogether, these results suggest that the host transcriptional response was associated with the changes in the bacterial loads as a result of administration of AMX at different concentrations.

If that observation is correct, we would expect gradual changes, both increases and decreases, in gene expression across samples from L to L + AMX10 to L + AMX20 to NL to control. To verify this hypothesis, we evaluated how many probes, among the ones that differed by at least 1.5-fold between L + AMX20 and L, also differed, following the same trend, in control animals. Among the 1839 probes differentially regulated between L + AMX20 and L conditions, 1504 were gradually up- or down-regulated across all conditions (Figure 4C ). Because 82% (1504 out of 1839) of the probes displayed such an expression profile, our results offer solid evidence for a strong association between the intensity of the gene expression profile and the number of colonizing bacteria.

Our results thus suggest that by decreasing the bacterial load, the antibiotic treatment brought the host gene expression profiles of infected mice to levels similar to those in uninfected animals. This hypothesis was supported by two arguments. First, the number of probes whose expression changed in AMX-treated groups versus control animals decreased to a level similar to the levels observed in mice infected with a nonlethal dose of Listeria. Second, there was a good correlation (r = .95) between the number of changed probes and the CFU/mg of liver in the different treatment groups.

Liver and Blood Gene Markers Predict Antibiotic Treatment Efficacy

The ability of the previously described predictive genes to classify the L + AMX10 and L + AMX20 samples into the L, NL, or control groups was assessed. Using liver predictive genes, 70% (14/20) of the L + AMX10 samples were predicted to belong to the lethal infection group, while only 25% (5/20) of the L+AMX20 samples were classified in the lethal group (Table 4). On the other hand, 75% (15/20) of the L + AMX20 samples were predicted to belong to the NL administration group, compared with only 25% (5/20) of the L + AMX10 samples.

Similar observations were made using the blood samples (data not shown). The three blood predictive genes classified 64% (9/14) and 12.5% (2/16) of the L + AMX20 and L + AMX10 samples, respectively, as belonging to the NL treatment group. They predicted that 75% (12/16) and 36% (5/14) of the L + AMX10 and L + AMX20 samples, respectively, would belong to the lethal group.

Taken together, these results demonstrate that the previously identified and here confirmed gene markers can accurately predict the order of magnitude of the infection based on the correlation between the gene expression profiles and the bacterial burden.

CONCLUSIONS

Using a newer generation of mouse array (GeneChip Mouse Genome 430 2.0) and larger biological replicates (20 animals per experimental condition), we confirmed our previous findings that gene expression profiles are predictive for the severity of L. monocytogenes infection (Ng et al. 2005). In addition, our study demonstrated a correlation between the intensity of the gene expression profiles and the bacterial burden. That correlation allowed the gene markers to efficiently predict the efficacy of antibacterial therapy. These results might find important clinical applications by helping clinicians to adjust the dose of antibiotics needed for effective treatment and yet without toxicity.