Connective molecular pathways of experimental bladder inflammation

Igor Dozmorov1, Marcia R. Saban2, Nicholas Knowlton1, Michael Centola1 and Ricardo Saban2

1 Oklahoma Medical Research Foundation, Arthritis and Immunology Research Program, Microarray Core Facility
2 Department of Physiology, The University Oklahoma Health Sciences Center, Oklahoma City, Oklahoma 73104


    ABSTRACT
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 GRANTS
 REFERENCES
 
Inflammation is an inherent response of the organism that permits its survival despite constant environmental challenges. The process normally leads to recovery from injury and to healing. However, if targeted destruction and assisted repair are not properly phased, chronic inflammation can result in persistent tissue damage. To better understand the inflammatory process, we recently introduced a profiling methodology to identify common genes involved in bladder inflammation. The method represents a complementation to the classic quantification of inflammation and provides information regarding the early, intermediate, and late events in gene regulation. However, gene profiling fails to describe the molecular pathways and their interconnections involved in the particular inflammatory response. The present work introduces a new statistical technique for inferring functional interconnections between inflammatory pathways underlying classic models of bladder inflammation and permits the modeling of the inflammatory network. This new statistical method is based on variants of cluster analysis, Boolean networking, differential equations, Bayesian networking, and partial correlation. By applying partial correlation analysis, we developed mosaics of gene expression that permitted a global visualization of common and unique pathways elicited by different stimuli. The significance of these processes was tested from both biological and statistical viewpoints. We propose that connective mosaic may represent the necessary simplification step to visualize cDNA array results.

cluster analysis; connective mosaics; partial correlation


    INTRODUCTION
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 GRANTS
 REFERENCES
 
INFLAMMATION underlies all major bladder pathologies including malignancy and represents a defense reaction to injury caused by physical damage, chemical substances, microorganisms or other agents (20). During acute inflammation, activation of specific molecular pathways leads to an increased expression of selected genes whose products attack the insult but ultimately should protect the tissue from the noxious stimulus. However, once the stimulus ceases, gene expression should return to basal levels to avoid tissue fibrosis, chronic inflammation, and loss of function as indicated by a reduced bladder capacity in patients with chronic cystitis (13).

Although sensory nerve (7, 23) and mast cell products (22) are known to be key parts of the inflammatory puzzle, other key molecules such as bacterial toxins are constantly being described to have a role in bladder inflammation (17). Therefore, as the database describing the repertoire of inflammatory mediators implicated in bladder inflammation increases, the central mechanisms by which injury can induce inflammation, cell damage, and repair often becomes less rather than more clear (20). To make sense of the vast knowledge of the genes involved in inflammatory response may require analysis of the patterns of change and the elucidation of gene networks far more than definition of additional members of inflammatory cascades (7, 19). To better understand the inflammatory process, we recently introduced a gene profiling methodology to identify common genes involved in bladder inflammation (19). The method used cluster analysis, more precisely, self-organizing maps, to determine the time course of gene regulation. However, cluster analysis does not include negatively correlated genes and fails to describe the different molecular pathways involved in the inflammatory response.

In the present work, we used three different stimuli known to induce a bladder inflammatory response: substance P (SP) (2), antigen challenge of sensitized animals (19), and Escherichia coli lipopolysaccharide (LPS) (17). SP, LPS, and antigen cause inflammation that is morphologically indistinguishable (19). However, each stimulus starts the inflammatory cascade by activating a distinct mechanism: SP activates G-protein-coupled neurokinin (NK) receptors, antigen challenge is dependent on immunoglobulin E (IgE) (8), and LPS seems to be dependent on toll-like receptors and CD14 (1). Therefore, based on the premise that the three stimuli have a different initiating cascade but share points of convergence and final common outcome, we used these paradigms to test a new statistical method applied to cDNA arrays and to determine the bladder molecular inflammatory pathways.


    MATERIALS AND METHODS
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 GRANTS
 REFERENCES
 
Animals
Female C57BL/6J mice, 10 to 12 wk old, were used in these experiments that were performed in conformity with the "Guiding Principles for Research Involving Animals and Human Beings" (OUHSC Animal Care and Use Committee protocol 00-109).

Antigen Sensitization Protocol
One group of mice in this study was sensitized with intraperitoneal injections of 1 µg of dinitrophenol-human serum albumin (DNP4-HSA) in 1 mg alum on days 0, 7, 14, and 21. This sensitization protocol induces sustained levels of IgE antibodies up to 56 days postsensitization (9), and its specificity resides in the fact that this response is abolished by antibodies to IgE (9). One week after the last sensitization, cystitis was induced (see Induction of Cystitis, below) by intravesical challenge with antigen DNP4-ovalbumin (DNP-OVA) to induce bladder mast cell degranulation (16, 21, 23).

Induction of Cystitis
Acute cystitis was induced as we described previously (1517, 21, 23). Briefly, female mice were anesthetized (40 mg/kg ketamine and 2.5 mg/kg xylazine; ip), then transurethrally catheterized (24 gauge, 3/4 inch, Angiocath; Becton-Dickson, Sandy, UT), and the urine was drained by applying slight digital pressure to the lower abdomen. The urinary bladders were instilled with 300 µl of one of the following substances: pyrogen-free saline, SP (10 µM), E. coli LPS strain 055:B5 (Sigma, St. Louis, MO; 100 µg/ml), or antigen DNP-OVA (1 µg/ml) in actively sensitized mice. Substances were infused at a slow rate to avoid trauma and vesico-ureteral reflux (16). To ensure consistent contact of substances with the bladder, infusion was repeated twice within a 30-min interval and a 1-ml syringe was maintained in the catheter end for 1 h. After that the catheter was removed and mice were allowed to void normally. One, four, and twenty-four hours after instillation, mice were killed with pentobarbital (100 mg/kg ip) and bladders were removed rapidly.

Alterations at Histological Level
The urinary bladder was evaluated for inflammatory cell infiltrates, mast cell numbers, and the presence of interstitial edema. A semi-quantitative score using defined criteria of inflammation severity was used to evaluate cystitis (5, 19). A cross section of bladder wall was fixed in formalin, dehydrated in graded alcohol and xylene, embedded in paraffin, and cut serially into four 5-µm sections (8 µm apart) to be stained with hematoxylin and eosin and Giemsa. Histology slides were scanned using a Nikon digital camera (model DXM1200) mounted on a Nikon microscope (model Eclipse E600). Image analysis was performed using a MetaMorph Imaging System (Universal Imaging, West Chester, PA). The severity of lesions in the urinary bladder was graded as follows: 1+, mild (infiltration of a 0–10 neutrophils/cross section in the lamina propria, with little or no interstitial edema); 2+, moderate (infiltration of 10–20 neutrophils/cross section in the lamina propria, with moderate interstitial edema); 3+, severe (diffuse infiltration of >20 neutrophils/cross section in the lamina propria, with severe interstitial edema) (5).

Sample Preparation for cDNA Expression Arrays
We used the same sample preparation technology as described previously (7, 18, 19, 21). Briefly, three bladders from each group were homogenized together in Ultraspec RNA solution for isolation and purification of total RNA. Mouse bladders were pooled to ensure sufficient RNA for gene array analysis. The justification for this approach is that there is not sufficient RNA in a single mouse bladder for performing cDNA array experiments, and the purification step reduces the yield of total RNA. By pooling three bladders together, we avoid using RNA amplification, but it needs to be taken in consideration that the variance is artificially low due to pooling of data.

RNA was treated with DNase according to manufacturer instructions (Clontech Laboratories, Palo Alto CA), and 10 µg RNA was evaluated by denaturing formaldehyde/agarose gel electrophoresis. This procedure was repeated using additional three bladders in each experimental group. Therefore, two pools of RNA were generated per experimental group for a total of six mice and two separate hybridizations per group.

Mouse cDNA Expression Arrays
cDNA probes prepared from DNase-treated RNAs obtained from each of the experimental groups were hybridized simultaneously to membranes containing Atlas Mouse 1.2 arrays (catalog no. 7853-1; Clontech, Palo Alto, CA). A complete list of genes present in this array can be found at http://www.clontech.com/atlas/genelists/index.html. Briefly, 5 µg of DNase-treated RNA was reverse-transcribed and labeled with [{alpha}-32P]dATP, according to the manufacturer’s protocol (Clontech). The radioactively labeled complex cDNA probes were hybridized overnight to a mouse cDNA expression arrays (Clontech) using ExpressHyb hybridization solution with continuous agitation at 68°C. After high- and low-stringency washes, the hybridized membranes were exposed overnight at room temperature to a ST Cyclone phosphor screen.

Quantification of Gene Expression
The phosphor imaging screen contains phosphor crystals that absorb the energy emitted by the radioactivity of the sample and reemit that energy as a blue light when excited by a red laser. Results are presented as digital light units (DLU). Spots on the arrays were quantified using grid analysis provided by OptiQuant Image Analysis Software (Packard BioScience, Downers Grove, IL). Quantification of each detectable spot was performed by measuring the digital light units generated by OptiQuant. The results were placed into a Microsoft Excel spreadsheet.

Database Submission of Microarray Data
The microarray data was prepared according to "minimum information about a microarray experiment" (MIAME) recommendations, has been deposited in the Gene Expression Omnibus (GEO) database, and can be accessed at http://www.ncbi.nlm.nih.gov/geo/. The samples can be retrieved with GEO accession number GSE597.

Outline of Statistical Analysis
Normalization.
Normalization was conducted using an iterative nonlinear curve-fitting procedure as described (6). This procedure assumes that intensities corresponding to mRNA not expressed by the tissue will be normally distributed and computes the mean and standard deviation (SD) (typically around 70% of all genes presented on each array). Next, we normalized each expression profile to its own background (defined by adjusting a mean = 0 and SD = 1 of the distribution of nonexpressed genes). For further analysis, data obtained after normalization of each profile to its own background were log-transformed with substitution of negative values by the minimal logarithmic value obtained within positive values.

Robust regression analysis of expressed genes.
This analysis was based on the fact that, in a linear regression analysis between two compared samples, the majority of genes are equally expressed and, therefore, randomly distributed around the regression line with a small portion of differentially expressed "outliers." The contribution of outliers to the regression analysis was down-weighted in an iterative manner. All expression profiles of both control and experimental groups were then rescaled to a common standard: the averaged profile of the control group. Our procedure for outlier exclusion was based on the selection of equally expressed genes with close to normally distributed residuals (measured as deviations from the regression line).

Selection of "hypervariable genes" (HV genes).
The next step was identification of a group of similarly expressed genes from control samples, denoted "the reference group." This group was used for selection of HV genes in experimental samples using the F-clustering procedure. The reference group was composed of genes expressed above background in control samples with a low variability of expression (as determined by an F test) and whose residuals approximate a normal distribution (based on the Kolmogorov-Smirnov criterion). The variability of expression of this reference group was due to technical variation. This value was used to identify genes that vary due to experimental conditions in a statistically significant manner.

F-means cluster analysis of HV genes coexpression.
This clustering procedure consisted of the following steps. 1) Gene expression normalization, log transformation, and rescaling as described above. 2) Identification of, and limiting subsequent analyses, genes expressed above background (3 SD above background noise), in at least one time point. 3) Identification and limiting of subsequent analysis to genes with expression levels that vary among time points (based on comparison with reference group variability by F criterion). These genes are denoted as "hypervariable." 4) Determination of a parameter, termed connectivity, for each of these hypervariable genes. Connectivity was defined as the number of genes whose expression behavior along the time varied from a given gene expression within ranges of the reference group (based on the F test). The appropriate threshold for the F test was used to diminish the portion of false-positive selections. Although genes could be associated with multiple clusters, inclusion was limited to the clusters of highest connectivity, such that the broadest biologic phenomenon, that is those involving the largest number of genes, could be distinguished. 5) HV genes were sorted by their connectivity, and the gene of highest connectivity was used as a "parent." All genes whose deviations from its expression not higher than variability of the reference group comprised cluster 1. The next gene of highest connectivity, not belonging to the first cluster, was used as a parent for the cluster 2, and the process continued until all genes were analyzed. It was also taken in consideration that a gene associated with two clusters may represent a functional link between these clusters. Genes that had zero connectivity did not belong to any cluster. Matrices of correlation coefficients were calculated for these clusters and were represented in a graphical output termed a connectivity mosaic such that patterns of correlated and noncorrelated genes could be identified by visual inspection.

Networking of the genes by partial correlation.
Correlation coefficients characterized coexpression within clusters. However, not every pair of highly correlated genes was functionally interconnected, as their coexpression may be due to independent influences. A partial correlation coefficient, which excludes third part influences, better characterized the functional interconnection. The partial correlation of variables x and y with respect to z was defined as an x-y correlation after removing the effects of z, or prxy,z = (rxy - rxzryz)/[(1 - rxz2)(1 - ryz2)]1/2. This analysis was performed in a pair-wise manner for all genes shown to be related by cluster analysis. A Monte Carlo simulation study was used to define the statistical threshold (t) below which partial correlation coefficients were likely to be due to chance. Sets of genes were considered to be interconnected if their respective partial correlation coefficients were greater than t. This refinement of assessing gene interrelationships provided a means of identifying genes that were likely to be directly related (and therefore functionally related), as opposed to cluster and correlation analysis, which identifies genes that are simply correlated in their behavior.

Discriminant function analysis (DFA).
DFA is a method that identifies a subset of genes whose expression values can be linearly combined in an equation, denoted a root, whose overall value is distinct for a given characterized group. DFA therefore, allows the genes that maximally discriminate among the distinct groups analyzed to be identified (14). In the present work, a variant of the classic DFA, named the "forward stepwise analysis," was used for selection of the set of genes whose expression maximally discriminates among experimentally distinct groups. The forward stepwise analysis was built step-by-step. Specifically, at each step all variables were reviewed to identify the one that most contributes to the discrimination between groups. This variable was included in the model, and the process proceeds to the next step. The statistical significance of discriminative power of each gene was also characterized by partial Wilk’s lambda coefficients (3), which are equivalent to the partial correlation coefficients generated by multiple regression analyses. The Wilk’s lambda coefficient used a ratio of within-group differences and the sum of within plus between group differences. Its value ranged from 1.0 (no discriminatory power) to 0.0 (perfect discriminatory power).


    RESULTS
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 GRANTS
 REFERENCES
 
Morphological Analysis
Intravesical instillation of SP, LPS, or antigen (OVA) induced bladder inflammation characterized by vasodilation, edema, and inflammatory cell infiltrate (Fig. 1, AD). However, morphometric analyses were not sufficiently sensitive to determine specific alterations induced by each pro-inflammatory stimulus. On the contrary, based on classic analysis, the conclusion is that, regardless of the stimulus, the urinary bladder maintains a unique set of responses leading to inflammation. However, we deliberately chose three stimuli that start the inflammatory cascade by activating different receptors and second messengers. Next, we used cDNA technology combine with a stringent statistical analysis to define which of the early, intermediate, and late genes were shared by these stimuli.



View larger version (61K):
[in this window]
[in a new window]
 
Fig. 1. Representative of hematoxylin and eosin staining bladder cross sections isolated from female mice treated with antigen [ovalbumin (OVA), A], lipopolysaccharide (LPS, B), or substance P (SP, C) indicating signs of inflammation (edema and inflammatory cell infiltrate) used for determination of the inflammatory index (D), as described in MATERIALS AND METHODS.

 
Determining Highly Variable Genes
We used as reference group the gene expression of all control samples (saline-treated mice). The variability of this group was taken as a measure of the instrumental error and used to select highly variable genes. As described in MATERIALS AND METHODS, genes expressed 3 SD above background in at least one time point (by F test) was selected as variable. Of 1,182 genes presented in the array, 437 genes were variable in at least one group. The allocation of variable genes in the different groups is presented in Fig. 2.



View larger version (35K):
[in this window]
[in a new window]
 
Fig. 2. Sharing of the variable genes in experimental groups. Variable genes were selected as deviated from stable position at the extent exceeding instrumental error.

 
Cluster Analysis
F-means cluster analysis produced 14 clusters that represented the dynamics of gene regulation (Fig. 3). The program first selected genes with higher connectivity as described in MATERIALS AND METHODS. These genes, denoted "parents" of clusters, were used as starting points for cluster identification. The clusters were then fully defined by identifying the remaining genes whose expression changed in a statistically significantly manner similar to that of the parent genes. Seven clusters clearly delineated genes that were upregulated early in the time course with a peak around 1 h after bladder stimulation, from those that were intermediately regulated peaking at 4 h, and those late genes peaking at 24 h (Fig. 3, labeled as type A, B, and C clusters, respectively). Similarly, this method also identified clusters of genes that were downregulated in a time-dependent manner (Fig. 3, labeled as -A = downregulated early; -B = downregulated at an intermediate time; and -C = downregulated late, at 24 h).



View larger version (33K):
[in this window]
[in a new window]
 
Fig. 3. F-means clustering of the gene expression dynamics. Gene expressions were normalized as described in MATERIALS AND METHODS. Phases of the response indicated as early (A), intermediate (B), or late (C) events. The numbers of genes within each given cluster are indicated in parentheses.

 
Hierarchical clustering was carried out with SpotFire software and presented in Fig. 4. This analysis presented an interesting dichotomization of the clusters obtained in Fig. 3 into two major patterns: downregulated and upregulated genes. Genes on the upregulated branch of the tree were further divided into two major components: one major component of the hierarchical cluster contained early gene (clusters 1, 5, and 7) and another contained late gene (clusters 2, 3, 6, and -4), further suggesting that the designation of early and late has a statistically significant basis. F-means cluster allocations in experimental groups are summarized in Table 1.



View larger version (21K):
[in this window]
[in a new window]
 
Fig. 4. Hierarchical clustering of the averaged shapes of F-means clusters from Fig. 3. Letters on the right present unified designations for the phases of the responses: early (A), intermediate (B), and late (C). Numbers on the left are the F-means clusters from Fig. 3.

 

View this table:
[in this window]
[in a new window]
 
Table 1. Cluster allocation of variable genes

 
The allocations of genes in different groups are given in Table 2. Table 2A presents genes with completely identical cluster allocation. Similarly, genes presenting the same dynamics could be further grouped into early, intermediate, and late response genes (Table 2B).


View this table:
[in this window]
[in a new window]
 
Table 2. Allocations of genes into different groups

 
Comparing SP, LPS, and OVA-Induced Gene Regulation by Correlation Mosaics and Functional Interconnections by Partial Correlation Coefficients
Correlation analysis was used to identify genes that respond similarly to a given stimulus with respect to time. This method provided a means of identifying statistically significant groups of genes that were correlated in their time response. This was done by determining a threshold below which correlated behavior was likely to occur by chance. This threshold was determined using a simulation experiment in which real data were substituted with random data having the same average and standard deviation over all time points as the real data. Based on this analysis, it was determined that a correlation coefficient of 0.70 could be used to define groups of genes significantly correlated with a particular stimuli.

Results of the correlation analysis were represented as "correlation mosaics." In each mosaic, the pixels were gradually colorized to represent highly negatively correlated (blue) or highly positively correlated (red) genes. These representations provided a visual inspection of similarities and differences in gene expression behavior induced by the three different stimuli. Correlation analysis was done in an iterative manner, with a pair of stimuli compared with the third stimulus in a given analysis. In the initial analysis, SP and LPS-treated bladders were compared with OVA-treated bladders. A striking correlation was found between stimuli-responsive genes in SP- and LPS-treated bladders (Fig. 5). In extreme contrast, the mosaic for the OVA-treated bladders was significantly different than the LPS- and SP-treated groups, suggesting that although OVA may activate some of the same pathways as SP and LPS, the regulation of genes within these pathways is distinct for OVA (Fig. 5).



View larger version (67K):
[in this window]
[in a new window]
 
Fig. 5. Correlation mosaics for genes highly correlated in response to SP and LPS and not correlated with responses to OVA. List of these genes is given in Table 3. Each spot (small square) in the plot presents correlation coefficients of gene expression. Genes highly positively correlated are in red, and highly negatively correlated are in blue. The same order of the genes along axis is used for all three mosaics.

 
The differences in gene regulation could be clearly delineated using partial correlation coefficient analysis. In this analysis, subtle changes in the regulatory relationships among groups of genes were more accurately determined than in cluster or correlation analysis. This was due to the fact that not every pair of highly correlated genes was functionally interconnected, as their coexpression may be due to independent influences. For example, take three highly correlated genes A, B, and C. Gene A may only be correlated with gene C because both A and C are influenced by gene B. By defining these indirect relationships, partial correlation coefficient analysis could be used to identify genes most likely to be directly regulated by a given stimulus. Moreover, both positive and negative correlations in behavior could be identified. This corresponds to genes that were likely to positively influence each other and those that were likely to negatively influence each other. The results of this analysis were represented diagrammatically as a network of interrelated genes with positive and negative influences denoted as solid or broken lines, respectively.

As in the above analysis, partial correlation coefficient analysis was done in an iterative manner, with pairs of stimuli compared with the third stimulus in a given analysis. In the initial partial correlation coefficient analysis, the data from the comparison of SP- and LPS-treated bladders vs. OVA bladders were refined. Significant commonality among all three stimuli was observed (Fig. 6). However, distinct variations were observed in regulation of genes encoding the constituents of the AP-1 transcription factor family including jun-b, jun-c, c-jun, and c-fos, strongly suggesting that there were marked similarities in regulation of AP-1-related genes in SP- and LPS-treated bladders, and marked differences in the regulation of these genes in OVA-treated bladders. The quantitative results of these analyses and gene lists are summarized in Table 3.



View larger version (56K):
[in this window]
[in a new window]
 
Fig. 6. Functional interconnection of genes selected in Fig. 5 as highly correlated in response to SP and LPS and not correlated with responses to OVA. This network was obtained with use of partial correlation determinations as described in MATERIALS AND METHODS. Connections identical in these two networks are presented as dashed lines. Dotted lines indicate interconnections with strong negative partial correlation.

 

View this table:
[in this window]
[in a new window]
 
Table 3. Genes with similar dynamics in SP and LPS and different dynamic in OVA group

 
Correlation and partial correlation analyses were then used to identify genes that respond similarly in OVA- and LPS-stimulated and differently in SP-treated mouse bladders. Fewer genes were identified in these groups compared with common genes induced by SP and LPS, further suggesting that the cellular responses to SP and LPS were more alike then the responses to OVA and LPS (Figs. 7 and 8). The gene that was differentially regulated, IER2, is poorly characterized and known to be regulated by a variety of stimuli (4, 12). The latter suggests that IER2 may play an important but as yet uncharacterized role in the response of tissues to inflammatory and immune stimuli. The quantitative results of these analyses and gene lists are summarized in Table 4.



View larger version (33K):
[in this window]
[in a new window]
 
Fig. 7. Correlation mosaics for genes highly correlated in response to OVA and LPS and not correlated with responses to SP. List of these genes is given in Table 4. All designations are the same as in Fig. 5.

 


View larger version (40K):
[in this window]
[in a new window]
 
Fig. 8. Functional interconnection of genes selected in Fig. 6 as highly correlated in response to OVA and LPS and not correlated with responses to SP. All designations are the same as in Fig. 6.

 

View this table:
[in this window]
[in a new window]
 
Table 4. Genes with similar dynamics in OVA and LPS and different dynamics in SP group

 
Finally, correlation and partial correlation analysis were used to identify genes that respond similarly in OVA- and LPS-stimulated mouse bladders and differently in SP-treated mouse bladders. Once again, the numbers of genes identified were fewer than in the first analysis, further suggesting that SP and LPS have more similar effects on cells than OVA and LPS, or SP and OVA. Strikingly in this analysis, three TNF family members (TNFR1, TNFR2, and CD30) that are key regulators of inflammation were similarly regulated by OVA, LPS, and SP. These are key regulators of inflammation. Previously, CD30 was shown to be structurally related to TNFR1 and TNFR2, and it is also functionally related to a complex immune response in T cells (10). The data from these three comparisons (SP and LPS vs. OVA, OVA and SP vs. LPS, and OVA and LPS vs. SP) demonstrates that the bladder inflammatory responses to SP and OVA are highlighted when the data from these stimuli are paired with LPS for comparison. The quantitative results of these analyses and gene lists are summarized in Table 5.


View this table:
[in this window]
[in a new window]
 
Table 5. Genes with similar dynamics in OVA and SP and different dynamics in LPS group

 
Discriminant Function Analysis
DFA was used to identify genes that maximally discriminate among groups and therefore provided a powerful means of characterizing genes modulating inflammation. In the initial DFA, genes with highest ability to discriminate among the three stimuli (SP, LPS, and OVA) were studied. In this analysis, data from all time points (1 h, 4 h, and 24 h) of a given stimulus were classified as a distinct group. Data from the all arrays performed, including the saline control, were also defined as a distinct group and used to identify the genes that discriminated among groups. The spatial organization of the elements in Fig. 11 provides a measure of the overall variance among groups. Interestingly, the relative positions of SP, LPS, and OVA were roughly equidistant from each other. These data suggest that LPS had the most significant effect on bladder cells. They also suggest that the effects of SP were more similar to those of LPS than OVA. Since LPS and SP had the most significant effects on the bladder cells, the genes modulated by these stimuli and not by OVA dominated the analysis. Not surprisingly, six of the seven genes identified by DFA play a key role in inflammation and repair. These included the Fas I receptor, IL-4 receptor, epidermal growth factor, retinoic acid receptor-{gamma}, c-myc, and BRCA1 (Table 6A). Although the overall picture provided by DFA is more limited than that of the correlation studies described above, the overall conclusions were similar.



View larger version (27K):
[in this window]
[in a new window]
 
Fig. 11. Discriminant function analysis (DFA) of samples from different groups. Genes selected as having higher discriminatory capabilities for the responses to different antigens. List of genes with the parameters of their discriminative potentials is presented in Table 6A.

 

View this table:
[in this window]
[in a new window]
 
Table 6. Highly discriminative genes

 

    DISCUSSION
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 GRANTS
 REFERENCES
 
Inflammation is an inherent response of the organism that permits its survival despite constant environmental challenges. The process normally leads to recovery from injury and to healing. To better understand the mechanisms involved in the inflammatory process, we recently used a profiling methodology, based on self-organizing maps, to define the major components of this pathway (19). The method represents a complementation to the classic quantification of inflammation and provides information regarding the early, intermediate, and late events in gene regulation.

However, gene profiling does not indicate how different stimuli by activating different receptors and signal transduction pathways can elicit a common inflammatory response. Therefore, here we introduced a new statistical technique for inferring functional interconnections between inflammatory pathways underlying classic models of bladder inflammation.

Because of the large amount of data generated with cDNA arrays, some pattern discovery is necessary to provide a high-level overview of a data set. In the present work we used F-means clustering, because each step of analysis has statistical estimations of significance. In addition, we applied this procedure only to highly variable genes whose variations over time were statistically significantly different from common instrumental variations. The decision about cluster allocation was also based on the statistical comparison of instrumental variation and the deviations of each member of the cluster from average. And finally, the number clusters was also a product of statistical analysis. Only shapes statistically significantly different from each other represented a new cluster. As a result, there were as many clusters as it was possible to discriminate by statistical analysis rather than by subjective decision. We also took into consideration that some genes could belong to several clusters simultaneously. The procedure reveals not only highly correlated but also highly anti-correlated genes leading to determination of the significant intercorrelations. Finally, unlikely self-organizing maps that fail to identify negatively correlated genes, this procedure reveals also clusters of genes with negative correlations that may have important in regulatory feedback controls.

The cluster analysis is based on a simple and reasonable idea, specifically that genes with similar expression behavior may be functionally related. However, the correlations identified by clustering do not necessary reflect true interconnection. Functionally distinct genes may be modified by a common regulator and thereby exhibit highly parallel behavior. A more significant assessment of functional interconnection was obtained from partial correlation coefficients. However, only a few publications have used this method to build a molecular network (11, 24, 25), primarily because of the fact that calculating partial correlations is complicated and inefficient. Moreover, the calculations necessary for determining partial correlation coefficients for large data sets are daunting; therefore, a significant portion of the differentially expressed genes from a given study are excluded from analysis to simplify the calculations. Additionally, current methods yield relatively low partial correlation coefficients whose relevance is difficult to determine. Last, significantly robust interconnections between a given pair of functionally related genes under examination and a third gene can unduly diminish the partial correlation coefficients calculated. The method described herein for the calculation of partial correlation is based on a stepwise examination of the separate contribution of all gene interactions. This analysis revealed strong interconnections and, unlike previously described methods, was not susceptible to relatively strong tertiary gene influences. This method is therefore more robust than previously described methods and demonstrated that the principal inflammatory bladder response to SP was similar to LPS. The subtle differences in gene regulation suggest that the SP, LPS, and OVA responses were discriminated primarily due to variations in the regulation of AP-1-encoding genes and TNF receptor family members.

Our initial work regarding the time course of LPS-induced gene regulation indicated that protooncogenes such as c-fos belong to the early response elements (18). This finding may indicate that although SP and LPS initiate inflammation by activation of different receptors, somehow their pathways merge early in the inflammatory cascade. If this is true, we should expect a cross-potentiation of the pro-inflammatory effects between these two stimuli. Interestingly, a cross-sensitization on bladder inflammation and cytokine production was observed between LPS and SP (17). When the results were compared with OVA-induced inflammation, the methodology here employed clearly indicated a strong difference in the pattern of gene regulation (Fig. 5). This was further supported by the connectivity mosaic, which revealed that, when bladders were stimulated with OVA, the connection between AP-1 transcription factor and several proto-oncogenes was broken (Fig. 6). Using the same type of analysis, it was possible to delineate common genes activated by LPS and OVA (Figs. 7 and 8) as well those that were commonly activated by OVA and SP (Figs. 9 and 10). Unlike the pathways shared by LPS and SP that involved primarily early genes, the pathways shared by OVA involved primarily intermediate and late genes more likely to be associated with consequences of inflammation rather than its cause.



View larger version (47K):
[in this window]
[in a new window]
 
Fig. 9. Correlation mosaics for genes highly correlated in response to OVA and SP and not correlated with responses to LPS. List of these genes is given in Table 5. All designations are the same as in Fig. 5.

 


View larger version (35K):
[in this window]
[in a new window]
 
Fig. 10. Functional interconnection of genes selected in Fig. 9 as highly correlated in response to OVA and SP and not correlated with responses to LPS. List of these genes is given in Table 5. All designations are the same as in Fig. 6.

 
In addition to the new statistical analysis introduced here, we also tested the capacity of DFA to identify a set of genes with maximal discriminatory capabilities between control and inflamed tissues. Although the gene lists obtained with this analysis were significantly smaller than those obtained in cluster or correlation-based analyses, this analysis provided a graphical representation (Figs. 11 and 12) of the relative effects of LPS, SP, and OVA on bladder cells that cannot be provided by other means. The major findings of DFA mirrored those obtained using correlation-based analysis by demonstrating that the most significant effects of the three stimuli were inflammatory in nature. Moreover, these results also suggested that the response of the bladder to SP is more like LPS than OVA in regards to its inflammatory potential. DFA highlighted the importance of seven genes worthy of further investigation. The seven genes identified by DFA had also been identified as relevant by either cluster or correlation analyses. In conclusion, we propose that multiple complementary analytical methods can be used to provide a clear picture of the biological significance of microarray data. For example, the validity and relative significance of the genes identified by one method can be assessed and refined by comparing the results obtained in an independent method. Finally, although the catalogs of likely relevant genes provided by these analyses are extremely informative, it is the graphical tools such as cluster analysis and correlation mosaics that allow the broader implications of the biology to be quickly discerned.



View larger version (12K):
[in this window]
[in a new window]
 
Fig. 12. DFA of samples from different groups. Genes selected as having higher discriminatory capabilities for the different phases of the responses to antigens. List of genes with the parameters of their discriminative potentials is presented in Table 6B.

 

    GRANTS
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 GRANTS
 REFERENCES
 
This work was supported by National Institutes of Health (NIH) Grant DK-55828-01 (to R. Saban), Oklahoma Center for the Advancement of Science and Technology Grant HR01-127 (R. Saban), NIH Grant 1-P20-RR-15577 (to M. Centola), and NIH Grant P20-RR-16478 (to M. Centola).


    FOOTNOTES
 
Article published online before print. See web site for date of publication (http://physiolgenomics.physiology.org).

Address for reprint requests and other correspondence: R. Saban, 940 SL Young Blvd., Rm. 666, Oklahoma City, OK 73104 (E-mail: ricardo-saban{at}ouhsc.edu).

10.1152/physiolgenomics.00130.2003.


    REFERENCES
 TOP
 ABSTRACT
 INTRODUCTION
 MATERIALS AND METHODS
 RESULTS
 DISCUSSION
 GRANTS
 REFERENCES
 

  1. Backhed F, Meijer L, Normark S, and Richter-Dahlfors A. TLR4-dependent recognition of lipopolysaccharide by epithelial cells requires sCD14. Cell Microbiol 4: 493–501, 2002.[ISI][Medline]
  2. Bjorling DE, Saban MR, and Saban R. Neurogenic inflammation of the guinea pig bladder. Mediators Inflamm 3: 189–197, 1994.[ISI]
  3. Cho JH, Lee D, Park JH, Kim K, and Lee IB. Optimal approach for classification of acute leukemia subtypes based on gene expression data. Biotechnol Prog 18: 847–854, 2002.[ISI][Medline]
  4. Coleclough C, Kuhn L, and Lefkovits I. Regulation of mRNA abundance in activated T lymphocytes: identification of mRNA species affected by the inhibition of protein synthesis. Proc Natl Acad Sci USA 87: 1753–1757, 1990.[Abstract]
  5. D’Andrea MR, Saban MR, Nguyen NB, Andrade-Gordon P, and Saban R. Expression of protease activated receptor (PAR)s 1, 2, 3, and 4 in control and experimentally inflamed mouse bladder. Am J Pathol 162: 907–923, 2003.[Abstract/Free Full Text]
  6. Dozmorov I and Centola M. An associative analysis of gene expression array data. Bioinformatics 19: 204–211, 2003.[Abstract/Free Full Text]
  7. Dozmorov I, Saban MR, Gerard NP, Lu B, Nguyen NB, Centola M, and Saban R. Neurokinin 1 receptors and neprilysin modulation of mouse bladder gene regulation. Physiol Genomics 12: 239–250, 2003. First published December 19, 2002; 10.1152/physiolgenomics.00141.2002.[Abstract/Free Full Text]
  8. Haak-Frendscho M, Saban R, Shields RL, and Jardieu PM. Anti-immunoglobulin E antibody treatment blocks histamine release and tissue contraction in sensitized mice. Immunology 94: 115–121, 1998.[ISI][Medline]
  9. Haak-Frendscho M, Saban R, Shields RL, and Jardieu PM. Anti-immunoglobulin E antibody treatment blocks histamine release and tissue contraction in sensitized mice. Immunology 94: 115–121, 1998.[ISI][Medline]
  10. Harlin H, Podack E, Boothby M, and Alegre ML. TCR-independent CD30 signaling selectively induces IL-13 production via a TNF receptor-associated factor/p38 mitogen-activated protein kinase-dependent mechanism. J Immunol 169: 2451–2459, 2002.[Abstract/Free Full Text]
  11. Kishino H and Waddell PJ. Correspondence analysis of genes and tissue types and finding genetic links from microarray data. Genome Inform Ser Workshop Genome Inform 11: 83–95, 2000.[Medline]
  12. Latinkic BV and Lau LF. Transcriptional activation of the immediate early gene pip92 by serum growth factors requires both Ets and CArG-like elements. J Biol Chem 269: 23163–23170, 1994.[Abstract/Free Full Text]
  13. Messing E. Interstitial Cystitis and Related Syndromes. Philadelphia: Saunders, 1992.
  14. Moore JH, Parker JS, Olsen NJ, and Aune TM. Symbolic discriminant analysis of microarray data in autoimmune disease. Genet Epidemiol 23: 57–69, 2002.[ISI][Medline]
  15. Saban MR, Hellmich H, Nguyen NB, Winston J, Hammond TG, and Saban R. Time course of LPS-induced gene-expression in a mouse model of genitourinary inflammation. Physiol Genomics 5: 147–116, 2001.[Abstract/Free Full Text]
  16. Saban M, Nguyen NB, Hammond T, and Saban R. Gene expression profiling of mouse bladder inflammatory responses to LPS, substance P, and antigen-stimulation. Am J Pathol 160: 2095–2110, 2002.[Abstract/Free Full Text]
  17. Saban MR, Saban R, Hammond TG, Haak-Frendscho M, Steinberg H, Tengowski MW, and Bjorling DE. LPS-sensory peptide communication in experimental cystitis. Am J Physiol Renal Physiol 282: F202–F210, 2002. First published August 21, 2001; 10.1152/ajprenal.0163.2001.[Abstract/Free Full Text]
  18. Saban MR, Hellmich H, Nguyen NB, Winston J, Hammond TG, and Saban R. Time course of LPS-induced gene expression in a mouse model of genitourinary inflammation. Physiol Genomics 5: 147–160, 2001.[Abstract/Free Full Text]
  19. Saban MR, Nguyen NB, Hammond TG, and Saban R. Gene expression profiling of mouse bladder inflammatory responses to LPS, substance P, and antigen-stimulation. Am J Pathol 160: 2095–2110, 2002.[Abstract/Free Full Text]
  20. Saban MR, Nguyen NB, Hurst RE, and Saban R. Gene expression profiling of inflammatory bladder disorders. Expert Rev Mol Diagn 3: 217–235, 2003.[Medline]
  21. Saban R, Gerard NP, Saban MR, Nguyen NB, DeBoer DJ, and Wershil BK. Mast cells mediate substance P-induced bladder inflammation through a NK1 receptor-independent mechanism. Am J Physiol Renal Physiol 283: F616–F629, 2002. First published May 22, 2002; 10.1152/ajprenal.00096.2002.[Abstract/Free Full Text]
  22. Saban R, Saban MR, Nguyen NB, Hammond TG, and Wershil BK. Mast cell regulation of inflammation and gene expression during antigen-induced bladder inflammation in mice. Physiol Genomics 7: 35–43, 2001.[Abstract/Free Full Text]
  23. Saban R, Saban MR, Nguyen NB, Lu B, Gerard C, Gerard NP, and Hammond TG. Neurokinin-1 (NK-1) receptor is required in antigen-induced cystitis. Am J Pathol 156: 775–780, 2000.[Abstract/Free Full Text]
  24. Toh H and Horimoto K. Inference of a genetic network by a combined approach of cluster analysis and graphical Gaussian modeling. Bioinformatics 18: 287–297, 2002.[Abstract/Free Full Text]
  25. Waddell PJ and Kishino H. Cluster inference methods and graphical models evaluated on NCI60 microarray gene expression data. Genome Inform Ser Workshop Genome Inform 11: 129–140, 2000.[Medline]




This Article
Abstract
Full Text (PDF)
All Versions of this Article:
15/3/209    most recent
00130.2003v1
Citation Map
Services
Email this article to a friend
Similar articles in this journal
Similar articles in ISI Web of Science
Similar articles in PubMed
Alert me to new issues of the journal
Download to citation manager
Search for citing articles in:
ISI Web of Science (6)
Google Scholar
Articles by Dozmorov, I.
Articles by Saban, R.
Articles citing this Article
PubMed
PubMed Citation
Articles by Dozmorov, I.
Articles by Saban, R.


HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
Visit Other APS Journals Online
Copyright © 2003 by the American Physiological Society.