Functional Genomics and Gene Expression Profiling in Sepsis: Beyond Class Prediction

Abstract

Functional genomics involving genome-wide expression analyses is rapidly finding applications in clinical medicine. New technologies now permit the simultaneous analysis of mRNA levels for the entire human transcriptome from as few as 1000 cells. This approach is dramatically changing the way we define health and disease, allowing, for the first time, an unbiased view of the global changes in gene expression that are occurring. For the study of trauma biology and sepsis, this technology offers a powerful tool to develop molecular signatures for inflamed tissues and specific cell populations. At present, functional genomics is being used to classify the progress of disease and survival in response to traumatic and burn injury, sepsis and visceral ischemia, and reperfusion injury, as well as to describe patterns of gene expression in response to varying microbial pathogens. As the number of bioinformatics tools increases, functional genomics is beginning to reveal the underlying complexity of the biological response to a variety of inflammatory diseases and is providing new approaches for their exploration. Functional genomics is becoming a standard tool in inflammation research as a means to unravel the basic biological processes.

The successful treatment of sepsis has continued to challenge both physicians and basic scientists alike. Despite improvements in both scientific technology and principles of critical care medicine, the mortality associated with sepsis has only modestly improved [1]. Mortality rates for patients with sepsis generally range from 25%to 45% [2–4]. The incidence of sepsis, however, continues to increase. In the United States alone, there are >750,000 cases of sepsis, resulting in 100,000 deaths [3, 5] The care of these patients has been estimated to cost the American health care system ∼$16.7billion each year [2].

The inflammatory component (systemic inflammatory response syndrome) to sepsis may represent either an exaggeration of a normal physiological response to the infectious process or an imbalance between the pro-and anti-inflammatory processes of the host [6, 7]. To that end, multi pleimmunotherapies have been investigated in an effort to modulate the inflammatory response and alter the clinical course of the patient with sepsis. These approaches have been relatively successful in animal models but have largely failed in clinical trials [6, 7], with the exception of activated protein C [8]. In general, such studies have failed for a number of reasons—for example, the heterogeneity of the treated population, the in ability to recognize the early clinical signs, and an incomplete understanding of the underlying pathophysiology of the septic response [2–4]. In a recent report, in particular, Eichacker et al. [9] have even speculated that the failure of these clinical trials was inevitable given the severity of the illness in the patient populations. In their study, a meta-regression analysis of preclinical studies showed that the beneficial effects of these biological response modifiers were highly dependent on an elevated risk of death, and that preclinical animal studies were conducted at significantly higher control mortality rates than seen in humans with severe sepsis. An analysis of the clinical trials showed that anti-inflammatory agents were also significantly more effective in patients with sepsis with higher risk of death and were potentially harmful in those with a low risk of mortality.

Prediction Of Outcome In Trauma Or Sepsis

Such studies reinforce the conclusion that the host response to traumatic injury and severe sepsis is multifactorial and that its pathogenesis is complex. Thus, it is very unlikely that the ultimate successful treatment of sepsis will encompass as ingle drug or an individual intervention. Rather, the trend in therapeutic interventions has been to focus on multimodal therapies targeting specific pathological components of the host response. Hence, it is probably naive to conclude that the immunologic monitoring of patients with sepsis will rely on the concentration of any one marker or protein or on the expression of any one gene. In fact, prognostic studies conducted over the past 20 years have clearly shown that the measurement of single plasma analytes generally lacks the sensitivity or specificity to predict outcome of severe trauma, infection, and sepsis. Although numerous individual mediators have plasma concentrations as a group that differ between patients with and without sepsis and between those who survive or die from sepsis (e.g.,TNF-α, IL-1β, IL-6, IL-10, and procalcitonin), such measurements have generally not proven effective in predicting which individual patients will survive or respond to therapy (reviewed elsewhere [10]).

To address the complexity of the sepsis response and to predict its outcome, multiple surrogate markers that reflect the nature and severity of the inflammatory response, the status of the coagulation and fibrinolysis systems, and the magnitude of organ injury are likely to be more effective in identifying patients at risk of an adverse outcome and who may benefit from interventional therapies. Such strategies will inherently require the use of multiplex approaches that have sufficiently high throughput to be cost effective. Thus, there arises a compelling need for broad-scale surveying tools of a patient's inflammatory and immune status. The gene-by-gene, or protein-by-protein, approach, therefore, may be rendered obsolete by these requirements.

Gene Expression Profiling As A Class Prediction Tool

The completion of the Human Genome Project has resulted in the preliminary sequencing of the >20,000–30,000 genes in the human genome [11]. One of the analytical tools spawned during this genomic age has been the development of multiplex oligonucleotide or complementary DNA microarrays [12,13]. Strictly speaking, microbar rays are platforms that can be used to simultaneously determine the levels of thousands of specific mRNA species. Current micro arrays, such as the Affymetrix U133+ GeneChip, have >54,000 probesets, representing essentially the entire human genome on a single chip. Hence, from as little as 100 mg of total cellular RNA isolated from whole blood or any tissue, the relative expression of all known genes, expressed sequence tags, and open-reading frames can be evaluated simultaneously. More important, the global interaction among these thousands of genes can be studied without any specific selection bias. By studying global expression, one can devise methods of class prediction based on global gene expression activity, and, more important, this can be done even without a priori knowledge of the function of any of the genes.

The term“genomics” has been used to define these new technologies and approaches that can survey the expression of the entire human genome simultaneously. “Functional genomics” relates these technologies to clinical medicine. Of importance, microarrays and the data they generate are giving scientists and clinicians new perspectives into health and disease. Although most investigations have focused on the changes in one or a relatively few genes in response to a disease or treatment, these newer technologies are exploring the changes in gene expression of the entire genome. This has resulted in functional genomics offering 2 unique perspectives on injury or sepsis biology: the ability to use“patterns” of gene expression to class predict or classify tissue responses (i.e., to develop a“signature” or “fingerprint”for a specific tissue, disease, or outcome) and to explore the underlying biological changes that occur in health and disease, while not being limited to any subset of selected genes.

Conceptually, these technologies are simple [14]. Oligonucleotidesor cDNAs for specific mRNA species in the entire genome are applied to glass slides or membranes, using either photolithography, microjet, or spotting technologies. The “target” mRNA is generally copied and labeled via cRNA or cDNA synthesis with radioactive or immunologic tags and then is allowed to hybridize to the oligonucleotides or cDNA fixed to the solid surface. Hybridization and gene expression are generally quantitated by scanning the slide, membrane, or chip for the degree of probe signal. The real challenge with microarray analyses is not in the data acquisition but rather in the quality control, data analysis, and information extraction [14].

Traditional statistical and bio informational approaches are not appropriate for the simultaneous analyses of 45,000–55,000 analytes. In addition to using probabilities to define confidence intervals among groups, false discovery rates are of ten applied [15, 16]. Equally problematic are the issues of how such vast quantities of data are analyzed or presented in a manner that can be readily assimilated and understood. Instead of bar and line graphs, heat maps, dendrograms, cluster and principal component analyses, and network analyses are generally used to extract meaningful information from the plenitude of data [17, 18].

Cancer Classification

The most common current application of functional genomics has been directed toward class prediction and has yielded some impressive results in cancer research and tumor diagnosis. Several studies indicate that the use of gene expression profiling has promising potential in tumor classification, prognostic implications, and therefore, selection of treatment regimens. Such approaches are currently under consideration by the US Food and Drug Administration for their diagnostic use in determining which patients may be administered specific antineoplastic therapies.

In a now classic report, Golub et al. [19] proposed a system of class discovery and class prediction based on gene expression monitoring in comparison to traditional histopathology for patients with hematologic cancer. At the time, this multiplex approach was revolutionary. The genome analyses were originally applied to 38 blood samples obtained from patients with acute leukemia, a tumor with well-established yet imprecise distinctions and clinically significant differences in treatment. Using a first-generation microarray containing only 6817 probe sets, the study detected ∼1100 genes whose expression was associated with distinction between acute myelogenous leukemia and acute lymphoblastic leukemia. A set of50 closely correlated gene predictors was selected and, in cross-validation studies, correctly assigned 36 of the 38 samples, with assignment for 2 samples remaining uncertain. In addition, the studies also revealed some novel biological insights and potentially new markers of disease, which may be useful in future studies for understanding biological mechanisms. For example, the HOXA9 gene was noted to be consistently over expressed in the 8 patients who were experiencing treatment failure. This correlation may become a useful factor in predicting outcome but still requires further prospective testing and confirmation.

A similar inquiry into diffuse large B cell lymphomas (DLBCLs), which have defied morphologically based classification because of tumor heterogeneity, resulted in the discovery of 2 sub types of tumors based on gene expression profiling [20]. A “Lymphochip” was designed to include 3186 probesets representing genes involved in B cells, lymphomas, and immune activation. Examination of 96 tumor samples revealed that different samples from the same patient were more related to each other than to other patients. Each tumor had its own unique gene expression “signature,” clusters of coordinately expressed genes based on cell type or function. Without identifying information, the algorithm accurately segregated the malignancies into 2 subtypes, a “germinal center B-like DLBCL” andan“activated B-like DLBCL.” The “activated B-like DLBCL” was named for its shared characteristics with the signature of in vivo—activated peripheral blood B cells. Of note, no single gene was useful for this categorization scheme, but several genes of interest were associated with each subtype. These 2 subtypes demonstrated statistically significant differences in 5-year survival after anthracycline-based treatment of 76% for “germinal center B-likeDLBCL,” versus 16% for “activated B-like DLBCL”. In comparison to the conventional International Prognostic Indicator, which is a system based on age, performance status, and extent and location of disease, the proposed subtypes demonstrated independent predictive value. The future use of microarrays to sort DLBCLs may alter treatment recommendations to include early bone marrow transplantation.

Although these approaches were first used in hematologic cancers, their use has not been limited to hematology-oncology but has also been useful with solid tumors. For example, class-prediction efforts have also been applied to breast cancer and have demonstrated a similar class prediction. Hedenfalk et al. [21] examined 6512 cDNA probe sets corresponding to 5361 genes in primary breast tumors from 7 carriers of the BRCA1 mutation, 7 carriers of the BRCA2 mutation, and 7 patients with sporadic cases. Statistical analysis identified 176 genes relevant to tumor classification. Interestingly, the identification of BRCA1 tumors was more accurate than of BRCA2.

The studies discussed in detail represent only a very small fraction of the reports using gene expression profiles in class identification and class prediction. The number of studies of cancer is increasing dramatically. Similar efforts to apply functional genomics have been made in lung cancer, soft tissue sarcomas, and colorectal cancer [22–25]. These studies clearly emphasize the potential utility of this approach to differentiate, on a molecular level, phenotypic differences among histologically similar human tumor sand possible prognostic implications. The sensitivity of this approach is striking, emphasizing the potential usefulness of this approach in clinical practice. What remains unconfirmed, however, is whether this approach can be successfully applied to a better understanding of the host response to tumor growth and metastasis, as well as to the host response to other diseases, such as inflammation and autoimmunity. This is perhaps the greatest challenge in the field of functional genomics: extracting from this vast amount of data a better understanding of the pathways and mechanisms leading to the diseased state.

Understanding Inflammation And Other Complex Physiological Processes

It is interesting that one of the first applications of functional genomics was in inflammation. Studies by Heller et al.[26] almost 10 years ago looked at the patterns of gene expression in synovial tissue during rheumatoid arthritis. Unfortunately, the application of functional genomics and microarray technology to understanding physiological changes in trauma, inflammation, sepsis, or other stressors has been limited. However, it is clear from the much more detailed studies of cancer that such a tool can and will be used in then ear future to better describe the tissue responses to surgical injury and inflammation. At present, most of the studies have used either cell culture systems or rodent models of injury and infection. Only in the past year have these technologies begun to be applied to appropriate patient populations. The majority of studies are still directed at rodent models.

Exploring The Genomics Of Infection

Recognition of microbial invasion is one of the hallmarks of innate immunity. Antigen-presenting cells, as well as cells of epithelial and endothelial origin, express surface receptors that can recognize“pathogen-associated microbialpatterns.” With the recent description of the Toll-like family of receptors as a primary mechanism for the recognition of microbial products, considerable efforts have been directed at examining the genome-wide expression response to different microbial pathogens. Probably most cited is the work of Nau et al. [27], who examined gene expression patterns in differentiated human macrophages in response to different pathogen stimulation. These authors demonstrated that macrophages responded to a broad range of bacteria with a robust, shared pattern of gene expression. This shared response pattern included genes encoding receptors, signal-transduction molecules, and transcription factors. These changes in gene expression are presumed to transform the macrophage into a cell primed to interact with its environment and to mount an innate immune or inflammatory response. The studies also showed that this shared activation program was induced by bacterial components, which included the Toll-like receptor agonists lipopolysaccharide, lipoteichoic acid, and muramyl dipeptide. Similar approaches have been examined in other antigen-presenting cell populations of the innate immune system, including dendritic cells. For example, Huang et al. [28] examined systematically how gene expression in dendritic cells changes in response to different pathogens. They used oligonucleotide microarrays to measure gene expression profiles of dendritic cells in response to Escherichia coli, Candida albicans, and influenza virus, as well as to their molecular components. There results were similar to those of Nau et al in the sense that dendritic cells shared a core response to the different pathogens or pathogen-associated microbial products. Interestingly, however, the authors also observed pathogen-specific programs of gene expression to each of these pathogens. These results demonstrated that dendritic cells can distinguish and respond distinctly to diverse pathogens through pathogen-specific immune responses.

We took a similar approach and used gene expression profiles to explore the human whole blood response to either a gram-negative or gram-positive stimulus [29].We investigated the differential gene regulation of leukocytes stimulated ex vivo withE. coli lipopolysaccharide or heat-killedStaphylococcus aureus. Microarray analysis revealed758 genes significant at the P < .001level that discriminated amonggram-negative—stimulated, gram-positive—stimulated, and unstimulated whole blood leukocytes. Interestingly, the shared response pattern between gram-negative and gram-positive stimulation was relatively modest; only 87 genes were up-regulated in both conditions>2-fold and only 43genes were commonly down-regulated. In contrast, the majority of the identified genes exhibited divergent responses to gram-positive and gram-negative stimulation. Furthermore, as shown in the gene ontology listing in figure 1, the patterns of gene expression were very different. Whereas the gram-positive pathogen markedly increased the expression of genes involved in protein synthesis, ribosomal proteins, and cell cycling, the gram-negative pathogen markedly increased the expression of genes involved in innate immunity and inflammation. The studies emphasize that the host inflammatory response to gram-negative and gram-positive stimuli shares some common response elements, but, more important, it exhibits distinct patterns of cytokine appearance and leukocyte gene expression. Using 4 different prediction models and leave-one-out cross-validation studies, we could accurately predict in every case the source of the stimulus on the basis of the gene expression pattern.

Similar approaches are now being used to describe the in vivo tissue response to a variety of inflammatory responses. For example, we used microarrays to describe temporally the changes in gene expression that occur in response to a second-degree burn injury in rodent skin [30].Those studies demonstrated an integrated genome-wide reprioritization of gene expression in response to a scald burn that continued for at least 2 weeks. As shown in figure 2, they also revealed that the changes in gene expression in response to a scald burn were widespread, with >1100 genes whose expression changed in response to the injury. Recent studies by others have also looked at the hepatic gene expression response to a burn injury in rodents and have begun to survey the dramatic genome-wide changes in gene expression in that organ [31, 32]. Those studies have also emphasized the widespread reprioritization of gene expression for substrate utilization and the manifestation of the hepatic acute phase response.

Cobb et al. [33] and Chung et al.[34] have pioneered the use of expression analysis to distinguish between asepsis response and inflammation. These authors recently reported the splenic and hepatic responses in sepsis due to cecal ligation and puncture versus animals under going sham surgery [33–35]. A7056-element microarray based on588 annotated mouse genes was used. By com paring the signal changes with that of the control group receiving sham laparotomy, many changes induced by sepsis were detected indifferent tissues. Overall, 13liver genes and 32splenic genes produced a5-fold difference in signal intensity. Only one gene was common to both expression profiles, the gene encoding the nicotinic α–cholinergicreceptor. Further data analysis using a consensus of3 analytical strategies demonstrated increased expression after sepsis in 16 genes and decreased expression in 6genes, of which 13were identified in the initial 45. The lack of overlap in gene expression profiles implies that the sepsis response can be distinguished from sham procedures and is organ-, tissue-, and/or cell-specific. Most genes corresponded to recognized inflammatory intermediates, such as CD-14, Iκβ-α, interleukin-1β, TNF-α,and IFN-inducible protein–10. Expression of cell signaling and apoptosis proteins was more common in the spleen, reinforcing the prior knowledge. Genes with increased expression in the liver included integrin b7, low-density lipoprotein receptor, growth hormone receptor, and prolactin receptor. This was the first set of results reported about the use of gene expression profiling to describe the in vivo“sepsistranscriptome.”

The study by Cobbet al. [35] was similar to an earlier report using the cecalligation/puncture model of sepsis in rats. Chinnaiyan etal. [36] used amicroarray with 7398 genes to assess organ-specific responses to sepsis. Samples were collected from rats under going sham surgery and from an untreated control population for comparison. They examined tissue from lung, liver, kidney, spleen, thymus, and brain from multiple time points. By comparing the gene expression profiles of sham-operated rats and controls, changes most likely due to properties of the operation, such as anesthesia, incision, and closure, could be filtered out in search of changes induced by sepsis. Several patterns emerged in the subsequent analysis, including a predominant cluster of up-regulated genes, including many familiar inflammatory mediators, such as IL-1β,phospholipase A₂, and complement components. One of the powers of this approach is the ability to identify without any selection bias previously unreported genes whose expression changed in response to the injury. For example, the study identified N-chimaerin and phorbolester receptor as genes actively up-regulated in the sepsis response, requiring further inquiry. Other clusters included down-regulated genes mostly involving the lung and spleen. These genes, including aquaporin 5, maintain the extra cellular matrix.

Similar studies were also performed by Yu et al. [37], who looked at the differential gene expression response in livers of healthy mice and after the administration of live E. coli or S. aureus.They observed both a common response pattern in the livers of the mice to the 2 different pathogens and a differential expression of 17genes. The authors conclude that, for the liver, both infectious agents share a final common pathway involved in the pathogenesis of sepsis, but certain genes are differentially expressed.

Exploring The Genomics Of Inflammation

One important goal of functional genomics has been to extend the class prediction models used with tumors to outcome in patients with trauma and sepsis. In particular, the question is whether patterns of gene expression can predict a successful or adverse out come in response to a severe traumatic injury. If so, the second question is whether those genes can provide insights in to the underlying mechanisms that contribute to this differential outcome. We addressed these questions in a preliminary exploration of gene expression profiles in patients undergoing thoracic aortic surgery and its consequential visceral ischemia and reperfusion injury. This injury has been shown to induce systemic inflammatory response syndrome and multi organ dysfunction in both human and animal models. We examined patterns of gene expression in peripheral blood leukocytes in response to visceral ischemia produced by a surgical repair of a thoracic aortic aneurysm [38]. Gene expression patterns were determined both preoperatively and at 2and 24 h after the surgery. Using a supervised analysis of gene expression from >12,000 genes simultaneously, we identified the expression of 148 genes that changed in response to the surgical injury. Not surprisingly, many of these genes were early activation genes involved in the innate immune and inflammatory response. However, when a supervised analysis was used to identify those genes whose expression predicted which patients would develop multi system organ failure after the surgical injury, it identified an additional 135genes whose expression correlated with outcome. Surprisingly, the expression of only 1gene overlapped between the lists of genes that changed in response to the injury and could predict outcome. These findings suggest that the changes in gene expression that occur after this type of surgical injury are different from the changes seen in patients who are likely to develop multisystem organ failure.

Interestingly, the genes whose expression changed in patients who developed multisystem organ failure were not limited to genes associated with an exaggerated inflammation. As shown in figure 3, the expression of a number of inflammation-related genes changed at least2-fold in patients who developed multisystem organ failure. These included chemokines and proteases, whose expression was increased. However, a number of other genes were equally discriminatory between the 2 groups. For example, decreases in the expression of a number of HLA class I genes and increases in the expression of heat shock proteins and serine proteases were all seen inpatients who went onto develop multisystem organ failure.

Perhaps most surprisingly, several of these changes in gene expression were seen before the surgical insult, suggesting that these patients had either a genetic predisposition or some prior immune or inflammatory response. Of importance, these findings imply that patients at risk from this procedure may be identified before surgery. Thus, it may be possible to include some version of expression profiling as a means to predict which patients are at risk of developing post surgical complications, such as multisystem organ failure.

Conclusions

For the study of trauma and sepsis, functional genomics offers a powerful tool to develop molecular signatures or fingerprints for inflamed tissues and specific cell populations. The majority of functional genomics studies have been performed for class prediction in cancer research, but the tools are now being used more frequently in inflammation research. Current applications in sepsis and trauma include the classification of disease progress and survival in response to traumatic and burn injury, sepsis and visceral ischemia, reperfusion injury, and various microbial pathogens. Functional genomics is beginning to reveal the underlying complexity of the biological response to a variety of inflammatory diseases, including the description of shared activation programs in response to differing microbial pathogens. The future of functional genomics is to provide new approaches to unravel the basic biological processes in inflammation, trauma, and sepsis.

Acknowledgements

Financial support. National Institutes of Health (grants R37 GM-40586,R01 GM-63213, and T32GM-08721).

Potential conflicts of interest. All authors: no conflicts.

References