Animal Models for the Study of Childhood Leukemia: Considerations for Model Identification and Optimization to Identify Potential Risk Factors

Abstract

Leukemias are the most common pediatric malignancies diagnosed in western industrialized societies. In spite of the substantial incidence of childhood leukemia in the United States and other countries, neither epidemiology studies conducted in human populations nor hazard identification studies conducted using traditional animal models have identified environmental or other factors that are directly linked to increased risk of disease. Molecular biology data and mathematical modeling of incidence patterns suggest that pediatric leukemogenesis may occur through a multistage or “multihit” mechanism that involves both in utero and postnatal events. The authors propose that pediatric leukemias can be modeled experimentally using a “multihit” paradigm analogous to the “initiation-promotion” and “complete carcinogenesis” models developed for tumor induction in mouse skin and rat liver. In this model for childhood leukemia, an initial genetic alteration occurs during in utero or early postnatal development, but clinical disease develops only upon additional genetic or nongenetic events that occur during the postnatal period. Application of this multistage or “multihit” model to hazard assessment studies conducted in transgenic or knockout mice carrying relevant molecular lesions may provide a sensitive approach to the identification of environmental agents that are important risk factors for childhood leukemia.

Leukemias are the most common neoplasms diagnosed in children in the United States and other western industrialized societies, and are the primary cause of pediatric cancer mortality in this country (Smith et al. 1999). More than 90% of childhood leukemias demonstrate a clinical pattern of rapid disease onset and progression, and are therefore classified as acute leukemias. Approximately 75% of pediatric leukemias are acute lymphoblastic leukemias (ALL). ALL arise in lymphoid precursor cells in the bone marrow, and most are of B-cell lineage (Look, 1997). Approximately 20% of childhood leukemias are acute myelogenous leukemias (AML); these neoplasms arise in myeloid progenitor cells in the bone marrow.

The American Cancer Society (ACS) estiates that approximately 2700 new cases of leukemia will be diagnosed in children in the United States in 2004 (ACS 2004); approximately 2000 of these cases will be ALL, and 550 will be AML. The majority of pediatric leukemias are diagnosed in children less than 5 years of age: National Cancer Institute (NCI) statistics demonstrate that the combined incidence for ALL + AML peaks at approximately 90 cases per million population in children between ages 2 and 3, and declines substantially in later years (Smith et al. 1999).

In spite of considerable advances in our understanding of the molecular biology of pediatric leukemia, the risk factors underlying most cases remain largely unknown. Inherited genetic diseases such as Down Syndrome and ataxia telangiectasia are strong predictors of pediatric leukemia risk (Miller 1967; Ross et al. 1994); disease incidence has also been linked to exposure to ionizing radiation and to cancer chemotherapeutic drugs such as cyclophosphamide and etoposide (Pui et al. 1989; Ross et al. 1994; Tucker et al. 1987). Comparisons of international disease incidence patterns suggest a general relationship between degree of industrialization and rates of pediatric leukemia (Parkin et al. 1998). Associations have also been proposed between disease risk and viral infection (Kinlen 1995) and increasing socioeconomic status (McWhirter 1982; Ross et al. 1994). Beyond these associations, however, no specific environmental factor or factors have been conclusively identified as directly increasing the risk of pediatric ALL or AML. As a result, the search continues to identify chemical and physical agents to which exposure may underlie the induction of pediatric leukemia.

EARLY EVENTS IN THE PATHOGENESIS OF CHILDHOOD LEUKEMIA

A substantial number of genetic alterations, including mutations in or amplification of core-binding factors such as AML1 (RUNX1), or the presence of translocation-associated chimeric fusion genes such as TEL-AML1, AML1-ETO, BCR-ABL, MLL-AF4, or E2A-PBX1 have been identified in pediatric leukemias (Speck and Gilliland 2002; Staudt 2002). However, recent evidence suggests that a single mutation, translocation associated fusion, or other genetic alteration may not be sufficient for leukemia development. Mathematical modeling of leukemia incidence patterns suggests that pediatric leukemia develops through a “multihit” mechanism, in which an initial genetic event occurs in utero, and other events required for disease promotion or progression occur during early childhood (Smith, Chen, and Simon 1997).

Support for this hypothesis comes from a growing number of studies in which alterations in leukemia-associated genes have been identified in newborns. In a series of retrospective studies, alterations in several leukemia-associated genes have been identified in cord blood samples collected at birth from children who later developed acute leukemia (Gale et al. 1997; Wiemels et al. 1999, 2002a; Mori et al. 2002; Yagi et al. 2000). The presence of alterations in leukemia-associated genes at birth in children who were (at least temporarily) disease-free clearly demonstrates (a) that alterations in leukemia-associated genes occur in utero, and (b) that additional postnatal events are required for development of clinical disease. This hypothesis is further supported by the observation that the number of neonates carrying chimeric fusion genes or other leukemia-associated genetic alterations is far greater than is the later incidence of clinical disease (Mori et al. 2002). Leukemia development in neonates carrying these “covert leukemia clones” (Mori et al. 2002) would therefore appear to depend on additional events (or “hits”) that occur during the postnatal period.

Additional support for a multihit mechanism of leukemia induction comes from the relatively modest concordance of ALL in monozygotic twins (Buckley et al. 1996), and from differences in the time course of leukemia development in monozygotic twins carrying identical genetic alterations (Ford et al. 1993, 1998). In these cases, differences in disease outcome in individuals with identical genetic alterations suggest that both in utero and postnatal events are important determinants of disease progression. Further evidence that key leukemogenic events can occur during early childhood comes from studies demonstrating the postnatal appearance of gene translocations and fusions that underlie the development of specific ALL subtypes (Wiemels et al. 2002b).

When considered together, these molecular data suggest that the pathogenesis of pediatric leukemia involves an initial in utero genetic alteration, followed by additional, postnatal event(s) that support clonal expansion and progression of preleukemic cells into clinically significant disease (Figure 1). Should the “multihit” mechanism of leukemogenesis parallel that defined experimentally for many solid tumors, leukemia could result from either (a) the induction of multiple genetic lesions in the hematopoietic progenitor cell, thereby conferring both a malignant phenotype and a proliferative advantage over normal cells; or (b) a single genetic alteration in the hematopoietic progenitor cell, followed by nonmutational alteration of differentiation, proliferation, and/or apoptotic pathways. Functional alterations in regulatory pathways occurring during the postnatal period could confer a proliferative and/or survival advantage on genetically altered hematopoietic precursor cells through mechanisms that are independent of additional mutations. As such, both prenatal and postnatal events appear to play critical roles in the etiology of childhood leukemia, and a “multihit” mechanism may provide a useful framework in which to model the disease process.

MODELING EARLY EVENTS IN CHILDHOOD LEUKEMIA

The hypothesized interaction of multiple factors in the etiology of childhood leukemia suggests that the process of leukemia development may follow a mechanistic path that is analogous to the multistage paradigm of experimental tumorigenesis in models such as mouse skin and rat liver. In mouse skin tumorigenesis, neoplasms can be induced by local application of either a single very high dose of a genotoxic chemical, or by repeated administration of lower doses of genotoxic agents (Boutwell 1964). These genotoxic chemical exposures induce an array of genetic lesions that is responsible for neoplastic development in the target tissue; this mechanism for cancer induction has been termed “complete carcinogenesis.” Alternatively, skin tumors may be induced by a multistage (initiation-promotion) sequence (Boutwell 1964) involving a single exposure to a low dose of a genotoxic agent (tumor initiator) followed by chronic exposure to a nongenotoxic compound (tumor promoter). The tumor promoter acts by one or more nonmutational mechanisms to stimulate cell proliferation and promote clonal expansion of genetically altered (initiated) cells (reviewed in Pitot and Dragan 2001).

For nearly 50 years, multistage tumorigenesis studies in mouse skin were performed using a chemical carcinogen or ionizing radiation to induce the initiating genetic lesion. More recent studies in transgenic mice have demonstrated that neoplasia can also be initiated by a heritable genetic lesion (such as an activated ras gene) whose expression is targeted to the tissue at risk (Leder et al. 1990). Skin tumors in both carcinogen-initiated (Boutwell 1946) and transgenic (Leder et al. 1990; Spalding et al. 1993; Tamaoki 2001) mice can be promoted by repeated topical application of nongenotoxic agents that stimulate tumorigenesis by mechanisms that appear to include inhibition of apoptosis, stimulation of cell proliferation, and suppression of normal checkpoint controls and other cellular regulatory processes.

Considering the molecular evidence and mathematical support for a multistage etiology of pediatric leukemogenesis, it appears likely that an interaction between genetic events and nonmutagenic environmental or other factors may be also causally associated with the induction of this family of diseases. A multistage mechanism of neoplastic development analogous to that described for mouse skin appears to provide a reasonable approach to the study of the etiology of childhood leukemia. As in mouse skin tumorigenesis, the initial genetic event in pediatric leukemia is not phenotypically detectable, yet begins a multistage or multievent process of neoplastic development. This genetic event predisposes the child to leukemia, but is not, in itself, sufficient to support progression to clinically significant malignancy. Disease development occurs only upon additional exposures to genotoxic or nongenotoxic agents that stimulate malignant progression of the initiated clone of cells.

USE OF ANIMAL MODELS TO IDENTIFY RISK FACTORS FOR CHILDHOOD LEUKEMIA

Substantial recent progress has been made in identifying molecular lesions that are associated with childhood leukemia, and a number of specific genetic lesions have been linked to increased disease risk (Speck and Gilliland 2002; Mori et al. 2002; Wiemels et al. 2002b). However, because the prevalence of these predisposing genetic lesions in the population is considerably greater than is the ultimate incidence of pediatric leukemia (Mori et al. 2002), factors in addition to these genetic lesions appear to be important determinants of the disease process. Epidemiologic investigations and basic molecular studies have yet to identify these other, potentially critical determinants of leukemia induction or progression. Until such agents are identified, our understanding of the etiology of pediatric leukemia, and our knowledge of possible gene-environment interactions that may underlie disease risk, must be considered to be incomplete.

In cases where human studies fail to identify risk factors for disease, animal models can play an important role in identifying exogenous agents to which exposure may underlie disease induction. Animal models may be particularly important in identifying risk factors for apparently multifactorial diseases such as pediatric leukemia. Well-designed and conducted studies in animal model systems permit the investigator to control for exogenous exposures and eliminate possible confounders, thereby establishing a causal link between observed effects and exposure to the agent being studied. It must be recognized, however, that each animal model has specific strengths and weaknesses, and that the rational application of any model to studies of disease processes requires an understanding of these strengths and weaknesses. The attributes and limitations of individual experimental model systems are defined in the context of the specific biological processes being modeled, the scientific questions (causality, mechanism, etc.) to be addressed, and a broad range of logistic factors that could have a substantive impact on study performance and data generation.

CRITERIA FOR ANIMAL MODEL SELECTION

General Considerations

A number of variables will determine the utility of an animal model to study disease etiology and/or identify environmental agents to which exposure may be causally linked to disease induction. The biology of the experimental process is clearly of primary importance: In the ideal case, the experimental model will demonstrate both phenotypic and genotypic concordance with the human disease being modeled. For utility in hazard identification, the process of disease development must also be sensitive to modulation: Even if a model demonstrates both a genetic lesion and a phenotype that are identical to those seen in humans, the model will be of limited value for agent identification if disease incidence cannot be reproducibly increased or decreased by exposure to exogenous chemical or physical agents.

The utility of an animal model for hazard identification is generally increased if that model is also used to develop human risk assessments and/or develop regulatory standards. Animal models used for these purposes are often supported by historical databases that both (a) provide evidence of the predictive nature of the model for human responses, and (b) define a context in which observed effects may be interpreted. If the model is to be used for broadly based screening of agents for their ability to induce, prevent, or cure the disease process under investigation, logistic considerations associated with study conduct (technical complexity, model reproducibility, disease latent period, background incidence of disease, statistical power, and cost) become increasingly important.

An important complication of modeling pediatric leukemia lies in the fact that the disease entity is, in reality, a family of related diseases that arise in different cell types and with different kinetics and disease courses. Because approximately ¾ of all pediatric leukemias are ALLs, it can be argued that a rodent model of ALL is the most appropriate surrogate for human pediatric leukemia. Accepting this argument, however, carries with it the recognition that a rodent ALL model may not be an appropriate system to study the remaining 25% of all pediatric leukemias. This limitation may be most important when attempting to identify risk factors and mechanisms for the 20% of acute pediatric leukemias that arise is in the myeloid component of the bone marrow. Although our knowledge of risk factors for pediatric leukemia is clearly incomplete, available data do suggest that both disease demographics and the identified risk factors for ALL and AML are at least somewhat different (Smith et al. 1999). For example, ALL is seen in higher incidence in whites than in blacks, in males than in females, and demonstrates a peak incidence in children between ages 2 and 5. The risk of ALL is increased by both in utero and postnatal exposure to ionizing radiation. By contrast, the highest incidence of AML is seen in Hispanics, AML does not demonstrate the clear agerelated peak in incidence that is seen with ALL, and an increased risk of AML is seen in children receiving post-natal exposure to ionizing radiation or to alkylating agents (Smith et al. 1999).

In consideration of the heterogeneity of this family of diseases, it appears very unlikely that a single “perfect” animal model will be developed for pediatric leukemia. For this reason, model selection becomes an iterative process of endpoint prioritization and optimization, whose result will be both (a) selection of an animal model that appears most appropriate to address the specific questions under investigation, and (b) development of an understanding of the strengths and limitations of the model selected. Restated, the goals of such a process are to identify the best available model system, and to determine the types of information that the model can, and perhaps more importantly, cannot provide.

Primary (Autochthonous) versus Transplantable Tumor Models

Over the past 40 years, a wide range of animal model systems has been developed for use in experimental studies of the biology of leukemia, lymphoma, and related hematopoietic malignancies. These models can be broadly classified into two groups: primary (autochthonous) leukemia models, in which the entire process of leukemogenesis occurs in the host animal, and transplantable tumor models, in which fully malignant or premalignant cells are transplanted into a suitable host.

The processes of leukemia development in these two types of animal models, and hence their utility for studies of different stages of leukemogenesis, are quite different. Models of primary leukemia induction require weeks to months for disease development, and may involve essentially all stages of leukemogenesis from the earliest genetic event through promotion and progression. On this basis, primary leukemia models are useful both for studies of the mechanisms underlying essentially all stages of neoplastic development in the bone marrow, and for hazard assessments and studies designed to identify potential leukemogens. The primary limitations of primary tumor models are logistic: technical complexity, the long latent period required for tumor development, and their resulting high cost and limited throughput.

By contrast, transplantable tumor models involve very rapid onset of clinical disease (days to weeks), and offer relatively rapid, low-cost experimental systems in which to study disease processes. Transplantable leukemia models such as the murine P388 and L1210 leukemias have been used for decades to screen new drug entities for antineoplastic activity (Burchenal 1975). It must be noted, however, the use of the P388 model for drug efficacy screening is predicated on its extremely rapid development of clinical disease following cell transplantation, resulting in mortality within 2 to 4 weeks. Although the course of clinical disease in other transplantable leukemia models is less rapid, these experimental systems can be used to study the modulation of only the latest stages of the disease. On this basis, transplantable tumor models appear to provide a relatively poor platform for studies of leukemia etiology, and for use in studies to identify potential leukemogens. For this reason, primary leukemia models provide the more suitable experimental systems for leukemia hazard identification.

Species Selection

Strong scientific and logistic arguments support the use of rodent species to model pediatric leukemia. Although primate, canine, and other large animal species are used widely in nonclinical safety assessment and to model certain human diseases, several key issues suggest that these nonrodent models are not optimal for studies of pediatric leukemia. Primary among these arguments are issues associated with study logistics (time, cost, and animal availability), animal welfare and use considerations, and statistical power issues related to group size. Another key consideration is our comparatively limited understanding of the genetics of nonrodent species that are commonly used in biomedical research, and a resulting lack of molecular insight into disease induction in these species. As a result, primate, canine, and other nonrodent species have had very limited application as models for hematopoietic neoplasia; the vast majority of model development for leukemia and lymphoma has been performed using rodent systems.

The use of rodent model systems for studies of pediatric leukemia addresses many of the issues of logistics, statistical power, and animal welfare that limit the use of large animal models for this purpose. The primary limitation to the use of rodent models is associated with possible species differences in disease mechanisms. In this regard, however, rapid advances in genetic engineering technology, and in characterization of the rodent (primarily murine) genome suggest that studies in mice, and especially in transgenic or gene deletion (knockout or conditional knockout) mice, may provide an approach in which both disease phenotype and genotype can be developed to simulate the human situation. An integrated approach to leukemia modeling, in which rodent disease phenotype may be linked to underlying molecular events, provides the opportunity to identify both agents that may be responsible for the induction of pediatric leukemia, and the mechanisms through which these agents act to induce or promote the disease process.

Strain Selection

Animal strains currently used for nonclinical safety assessments (toxicity testing of drugs or environmental chemicals) and/or studies of disease etiology include both traditional and genetically engineered (transgenic and knockout) models. The advantage of many traditional (i.e., not genetically engineered) animal models lies primarily in their long history of use to study disease processes, often resulting in a known ability to predict human responses. The rationale supporting the use of animal species and strains that are commonly used in nonclinical safety assessment is also strengthened by the availability of historical databases for disease incidence, thereby providing data (in addition to concurrent study controls) against which disease induction data can be evaluated.

The application of genetically engineered animal model systems to hazard identification and safety assessment is relatively recent. Even considering their relatively short period of use, however, the value of a number of these models in the study of disease etiology is well established. Most transgenic and knockout mouse models were initially developed to address specific mechanistic questions in the pathogenesis of disease, viz., to determine the role of a specific gene or pathway in disease induction. Following the establishment of a causal linkage between the gene and the disease, transgenic or knockout models in which the gene of interest has been inserted or deleted from the germ line may provide a valuable approach to hazard assessment and the identification of agents to which exposure may stimulate or inhibit disease induction. Several such model systems have been used to identify agents that may induce leukemia and lymphoma (Repacholi et al. 1997; Harris et al. 1998; McCormick et al. 1998), and additional models are under development.

A key attribute of transgenic and knockout animals is their ability to provide a model system in which the initial genotoxic event (or “hit”) has been accomplished by manipulation of the germ line. Using mouse skin as a paradigm, these transgenic or knockout models of disease may be considered to be constitutively initiated (Leder et al. 1990). As such, these models appear to provide highly sensitive systems in which to identify agents that act by mechanisms involving later events in disease development. As discussed previously, these late events could involve either additional genetic damage, or nongenetic alterations that support clonal expansion of genetically altered cells.

Although transgenic and knockout models appear to provide significant advantages for use in hazard assessment, these models do have several potentially important disadvantages. Ashortterm disadvantage of these new models is the absence of validation data. Because genetically engineered mouse models have been used in hazard assessments for only a short time, they are not supported by large historical databases and demonstrated evidence of predictiveness for human responses. With the exception of the p53^+/− knockout and ras transgenic strains such as the TG.AC and ras H2 (Gulezian et al. 2000), most genetically engineered mouse models have been used in only a small number of laboratories, and as such have not been broadly validated for use in hazard assessment. Secondly, because most transgenic models were initially bred and are currently maintained in small research colonies, their availability in numbers required to support hazard assessments is often quite limited. Available supporting data for these strains often do not include demonstrations of genetic stability over multiple generations or under the intensive type of breeding required to generate animals in sufficient quantities for broadly based hazard assessment studies. As the use of individual genetically engineered models in hazard assessment grows, these issues will be resolved. At the present time, however, the lack of background validation data and unknown predictiveness for human responses appear to be the primary limitations associated with their use.

Biology (Phenotype/Histogenesis/Histopathology)

Throughout most of the 20th century, gross and microscopic pathology were the predominant factors determining the relevance of an animal model to human neoplasia. Although advances in molecular biology have expanded the metrics on which comparisons can be based, demonstration of comparable patterns of lesion histopathology and histogenesis in an animal cancer model and in the corresponding human neoplasm remains an essential component of model suitability. Evidence of histopathologic concordance of human and experimental lesions is further strengthened by the presence of similar patterns of preneoplastic lesions.

Demonstration that an in vivo model responds to an exogenous agent or manipulation in a manner comparable to that of the corresponding human disease is also a key element of animal model validation. Although the fundamental regulatory mechanisms underlying most types of human neoplasia remain unknown, factors that can modulate proliferative and other responses in human malignancies have often been identified. Where such data exist, similarities and differences between the responses of human cells, tissues, or in vivo neoplasms and those of the animal model to hormones, ionizing radiation, drugs, ablative surgery, nutritional manipulations, or other interventions can provide important evidence on which the biological relevance of the animal model can be evaluated.

Molecular Biology

The broad application of genetic engineering technology to animal model development permits the study of the etiologic role of specific genes that have been either inserted into or deleted from the murine genome. Transgenic and knockout models are commonly used to identify genes whose altered expression or function is linked to disease, and their use in hazard assessments is becoming increasingly common. The application of genetically engineered models to environmental toxicology promises to provide improved experimental systems for the study of gene-environment interactions, and also provides an approach to study disease process in susceptible populations.

Advances in molecular medicine now permit parallel interspecies comparisons of disease phenotype and the molecular events linked to disease induction. Integrated analysis of morphologic and molecular similarities and differences between experimental models and human neoplasms provides a considerable advantage over earlier comparisons made on the basis of histopathology alone. When considered in combination with similarities in disease histogenesis and histopathology, evidence that a disease process in an experimental model is associated with molecular events that mimic those in the human provides a considerable degree of support for the scientific relevance of the model system.

EVALUATION OF THE UTILITY OF SPECIFIC ANIMAL MODELS TO IDENTIFY RISK FACTORS FOR CHILDHOOD LEUKEMIA

Standard (Outbred, Inbred, and Hybrid) Rodent Models

Although a number of important limitations can be identified, rat and mouse strains that are commonly used as models for nonclinical safety assessments have substantial value as systems to identify environmental leukemogens. The primary strength of these models lies in the existence of large historical databases for the incidence of leukemia and malignant lymphoma in animals receiving either no treatment or treatment with vehicle only, and existing data demonstrating the sensitivity of these models to the induction of hematopoietic malignancy. The most widely used rodent strains are the Sprague-Dawley (and Sprague-Dawley–derived strains, such as the CD) rat, the Fischer (F344) rat, the Wistar rat, the CD-1 mouse, and the B6C3F1 mouse. Among these models, the incidence, latency, and histopathology of spontaneous hematopoietic malignancies varies considerably by species, strain, and sex; this permits the investigator considerable flexibility in model system selection in order to optimize the sensitivity of the bioassay. All five models are used widely by both government agencies and by commercial entities to evaluate the potential oncogenicity of a broad range of agents, and are considered to be generally predictive of human responses. Finally, all are sensitive to the induction of hematopoietic neoplasia by chemical and physical agents, and large data bases supporting their sensitivity to leukemogenesis or lymphomagenesis are available from the United States National Toxicology Program (2004) and other organizations.

The strengths of these model systems must be considered in parallel with some important limitations (Table 1). First, hematopoietic malignancies are routinely observed in these rodent strains only in chronic oncogenicity bioassays that use group sizes of 50 to 100 animals per sex per group and exposure durations of 18 to 24 months. Chronic studies of this duration, which include extensive histopathologic evaluation of tissues, are very costly, technically complex, and labor intensive. Furthermore, because neoplastic development in these models is not specifically targeted or limited to the hematopoietic system, concomitant effects on tumorigenesis in other tissues may alter the influence of test agents on leukemia and lymphoma. Finally, and perhaps most importantly, these standard animal strains may be of limited sensitivity to detect agents whose activity is limited to later stages of hematopoietic neoplasia. Should leukemia development proceed through a “multihit” mechanism that is effected by a single genotoxic event followed by stimulation of the proliferation of altered clones of cells, the absence of the initiating event in these models could limit their ability to identify leukemogens that act through nonmutational or promotional mechanisms. Benzene provides an important example of the limitations of these model systems for the identification of leukemogens. Although exposure to benzene is a known risk factor for human leukemia (Rinsky 1989), studies in F344 rats, CD-1 mice, B6C3F1 mice, and several other strains exposed to benzene developed tumors in multiple sites, but not in the bone marrow (Snyder et al. 1988; Huff et al. 1989).

High-Incidence Rodent Models

Rodent strains that demonstrate a high incidence of background or “spontaneous” hematopoietic neoplasms in otherwise untreated animals may offer greater sensitivity to detect etiologic agents for childhood leukemia than do the standard-bred rodent models used routinely in safety assessments. The primary hypothesis supporting the use of high-incidence models is that the sensitivity of leukemogen detection will increase in animals in which the disease process is active. Such increases in sensitivity could be of key importance in identifying agents with weak to moderate activity as leukemogens, and could also be of mechanistic relevance when used to identify risk factors for an apparently multifactorial disease such as pediatric leukemia.

High-incidence rodent models have been identified for a large number of neoplastic diseases. These models commonly demonstrate genomic integration of viral sequences that predispose animals to malignancy in a manner analogous to being exposed to a genotoxic agent. Although these “hits” are untargeted and their molecular biology may be of limited direct relevance to human leukemia, within the animal model they perform the same function as a carcinogen-induced somatic mutaton or the insertion of a disease-associated transgene into the germ line.

The most widely studied high incidence model of hematopoietic neoplasia is the AKR mouse. Originated approximately 70 years ago by Furth and colleagues (Furth, Seibold, and Rathbone 1933), the AKR mouse develops a high incidence of leukemias and thymic lymphomas as a result of the genomic integration and spontaneous activation of oncogenic viral sequences (Steffen et al. 1979; Hiai 1996). Mathematical analyses have demonstrated that the development of hematopoietic and lymphoid neoplasia in the AKR mouse requires fewer genotoxic “hits” than does disease development in standard-bred or “spontaneous” tumor models (Frei 1980). These analyses suggest that the AKR mouse may provide a sensitive test system in which to identify potential leukemogens.

Because tumorigenesis follows an accelerated time course, hazard assessment studies in high-incidence model systems require shorter in-life periods than do studies performed in standard animal models. In addition to reduced technical complexity and cost, decreases in study duration may also reduce the incidence of tumors in nontarget tissues; this reduction can improve study data by reducing or eliminating mortality or morbidity associated with neoplastic development in organs and tissues other than the hematopoietic system. Group size is driven by statistical considerations associated with the response in sham (or vehicle) controls, but is ordinarily considerably smaller than the 50 to 100 animals per sex per group used in studies in spontaneous tumor models.

The strengths and limitations of high-incidence models such as the AKR mouse for use in the identification of leukemogens are considerably different than (and generally oppose) those identified for standard animal strains. When compared to rodent strains used in safety assessments, the reduced number of “hits” required for disease induction suggests that AKR mouse may provide improved sensitivity in agent detection. However, data to support this statement are sparse, and unlike the rodent strains commonly used in safety assessment, the AKR mouse lacks a substantial historical database to support data interpretation. The acceleration of leukemogenesis in AKR mice exposed to whole body ionizing radiation (Jullien and Rudali 1967) does provide evidence of concordance with one known risk factor for human leukemia. However, the ability of this mouse strain to predict human responses to either genotoxic or non-genotoxic chemicals is less clear. Perhaps most importantly, the viral genes that are responsible for leukemogenesis in the AKR mouse (Hartley et al. 1977) are apparently unrelated to genetic alterations that underlie human leukemogenesis (Speck and Gilliland 2002;Wiemels et al. 2002b). This lack of genetic concordance suggests that substantial improvements in model relevance may be achieved by the use of transgenic mice carrying genes of known relevance to the etiology of pediatric leukemia.

Multistage or Multiagent Rodent Models

The ability to identify risk factors for neoplastic disease, and particularly those agents whose activity is limited to the later stages of neoplastic development, may be increased substantially by the use of multistage or multiagent models. Several multiagent models have been developed to study the etiology and pathogenesis of leukemia and lymphoma; optimally, neoplastic development in these models is targeted to the hematopoietic system in a site-specific manner.

In these models, exposure to a virus, chemical carcinogen, or ionizing radiation causes hematologic malignancies to develop more rapidly than in spontaneous model systems, presumably through the induction of initial “hits” that result from exposure to these agents. Evidence of leukemogenic activity for a test agent is demonstrated by increases in disease incidence and/or a further acceleration of disease induction in groups treated with the inducing agent (virus/carcinogen/ radiation) plus the test agent versus groups treated with the inducing agent (virus/carcinogen/radiation) alone.

Most multistage or multiagent models for leukemia were developed decades ago, using observational, rather than mechanistic, criteria for model development. Representative models in this class include those in which hematopoietic neoplasia is induced by exposure to whole body x-rays (Kaplan and Brown 1952), viruses (Yefenof and Kotler 1995), or chemicals such as procarbazine (Rogers, Akhtar, and Zeisel 1990).

The strengths and limitations of multistage and multiagent models to identify leukemogens are generally similar to those of the high-incidence models. Disease induction is targeted to bone marrow and/or lymphoid tissues, thereby reducing competing risks associated with neoplastic development in nontarget tissues. Study size and duration are both reduced in comparison to studies performed in standard-bred animal models. Because tumor latency in multistage models systems is generally shorter than is observed in oncogenicity studies performed using standard animal model systems, costs are reduced and study throughput increases. As was the case in the high-incidence models, the ability to detect leukemogens with weak or moderate activity may be improved versus that in model systems used in traditional safety assessments.

However, in spite of their many years of availability, multistage models for hematopoietic neoplasia have been used much less frequently for hazard assessment than have comparable models for neoplastic development in sites such as the skin (Yuspa 1994), mammary gland (McCormick and Moon 1985), and liver (Dragan and Pitot 1992). As such, their biology and predictiveness for human responses are not well characterized, and the background database supporting their use is limited. Finally, the genetic lesions underlying disease induction in these models have not been characterized, so the relevance of individual models to the molecular biology of human leukemia is not established.

Genetically Engineered Rodent Models

The application of genetically engineered murine models to hazard assessment offers the potential for major improvements in the identification of potential risk factors for pediatric leukemia. In a fully optimized genetically engineered test system, risk factors for leukemia induction can be identified using a model that demonstrates both phenotypic and genotypic concordance with human disease. As such, an optimized transgenic, knockout, or knockin animal model will demonstrate both a genetic lesion that corresponds directly to those identified in human leukemias, and a phenotype that closely simulates at least one subtype of the human disease. After model validation, transgenic or other genetically engineered in vivo leukemia models may address all of the key limitations discussed previously for the standard-bred, high-incidence, and multistage animal model systems.

From a mechanistic perspective, integration of the initial genetic lesion into the germ line of a transgenic mouse provides a reasonable surrogate for the proposed first “hit” in leukemia induction. A genetically initiated animal provides the requisite sensitivity to identify leukemogens that act in later stages of disease induction, and parallels the prenatal/postnatal sequence of events that has been proposed as the etiologic basis for human leukemia (Smith, Chen, and Simon 1997). Finally, the importance of the concordance between genetic alterations in the model and in human leukemia cannot be overstated for the purposes of hazard identification or studies of disease mechanisms.

That said, however, a fully optimized genetically engineered mouse model that is suitable for use in leukemia hazard assessment does not yet exist. Although a number of transgenic mouse strains with both relevant genetic lesions and disease phenotypes have been developed, issues of model validation and human predictiveness remain to be addressed. More importantly, however, current models appear to require substantial modification in order to yield disease incidence and time-response parameters that are suitable for use in hazard assessments.

Genetically Engineered Models Used in Safety Assessment

The heterozygous p53 knockout mouse (p53^+/− mouse) and two transgenic mouse strains that carry activated ras genes (the TG.AC mouse and the ras H2 mouse) have received intensive study for possible use as intermediate-term bioassays to identify chemical carcinogens. The p53 knockout mouse is of particular interest for possible application in leukemia hazard assessments on the basis of its high-incidence of “spontaneous” lymphomas that develop between approximately 6 and 12 months of age (Donehower et al. 1992). For this reason, the p53^+/− mouse has been used for the purposes of leukemia hazard assessment (McCormick et al. 1998).

Six-month oncogenicity studies using these genetically engineered mouse models are being considered by regulatory agencies as possible replacements for 2-year oncogenicity bioassays in standard-bred mice. For this reason, these models have undergone extensive validation efforts to determine their reproducibility and predictive nature (reviewed in Gulezian et al. 2000). It appears likely that the p53^+/− mouse model and at least one ras transgenic mouse model may be accepted as intermediate term assays for oncogenicity, and as a result, will support rapidly growing historical databases. The ongoing validation of these animal models for application in carcinogen detection and hazard assessment is an important strength. However, the value of these models for the study of leukemia induction may be reduced by their sensitivity to oncogenesis in a wide spectrum of tissues (Gulezian et al. 2000); neoplastic development in tissues other than the bone marrow could provide competing risks in the host animal, and thereby reduce the sensitivity of these models for use in the identification of leukemogenic agents.

It is also important to note that the genetic lesions in the p53 knockout mouse and the two ras transgenic mouse mod demonstrate little concordance with genetic alterations identified in human leukemias. Molecular analyses of neoplastic cells obtained from pediatric leukemia patients clearly demonstrate that p53 alterations and ras activation are low frequency events (Lübbert et al. 1990; Kawamura et al. 1995, 1999; Yokota et al. 1998). Because neither p53 inactivation or deletion nor ras activation appear to be mechanistically linked to human leukemogenesis, animal models carrying these genetic alterations lack the requisite molecular linkage to the human disease.

Genetically Engineered Leukemia/Lymphoma Models

In addition to p53 knockout mice, several other genetically engineered mouse strains develop hematopoietic neoplasms either spontaneously or in response to chemical carcinogens. These models may offer considerable value in studies of mechanisms of hematopoietic neoplasia. However, as is the case with p53 knockout and ras transgenic mice, the utility of these mouse strains as platforms for leukemia hazard assessment is reduced by their lack of genetic concordance with human leukemia.

The Eμ-ret transgenic mouse carries an RFP/RET fusion gene driven by the immunoglobulin heavy chain enhancer, and develops B-cell leukemias lymphomas within several months after birth (Wasserman, Zeng, and Hardy 1998). This model is of potential mechanistic relevance to the multistage or “multihit” mechanism of leukemogenesis: Evidence from the laboratory in which the Eμ-ret mouse was developed suggests that the RFP/RET fusion gene initiates neoplastic development in utero, but requires additional, postnatal events for the development of clinical disease (Wasserman, Zeng, and Hardy 1998; Zeng et al. 1998). It has been hypothesized that these secondary factors induce leukemia via altered regulation of cell proliferation and apoptosis (Wasserman, Zeng, and Hardy 1998), providing support for the possible role of nongenotoxic agents in the multistage or “multihit” etiology of hematopoietic neoplasia.

The Eμ-myc transgenic mouse, which carries a c-myc gene driven by the immunoglobulin heavy chain enhancer, develops a >90% incidence of pre-B and B-lymphomas during its first 5 months of its life (Harris et al. 1988). Although its genetic lesion does not demonstrate concordance with human leukemia, the Eμ-myc model is of considerable mechanistic interest as a result of demonstrated cooperativity between c-myc and other oncogenes (ras, raf, N-myc, and bcl-2, but not abl) in the induction of hematopoetic neoplasia (Landgon, Harris, and Cory 1989; Rosenbaum et al. 1989; Strasser et al. 1990). The stimulation of lymphoma development by the interaction of multiple oncogenes provides further evidence of the “multihit” nature of hematopoietic oncogenesis.

The Eμ-pim transgenic mouse (or PIM mouse) develops lymphoma as a result of genomic integration of the pim-1 gene (Breuer et al. 1989), an oncogene for which there is no known human homologue. Although it demonstrates no genetic concordance to human hematopoietic neoplasia, the PIM mouse model is of interest for several reasons. First, the PIM mouse develops B-cell lymphomas with a long latent period (>12 months; Repacholi et al. 1997), but develops T-cell lymphomas rapidly (<6 months) in response to a single exposure to a chemical carcinogen such as N-ethyl-N-nitrosourea (van Lohuizen et al. 1989). The incidence and latency of lymphoma induction can be modulated by altering the dose of carcinogen, and disease induction and associated mortality can be suppressed by the administration of cancer chemopreventive agents (McCormick et al. 1996). These data clearly demonstrate the sensitivity of oncogenesis in the PIM transgenic model system to modulation by exogenous agents. On the basis of its disease phenotype and sensitivity to modulation, the PIM mouse has been used in hazard assessment studies to evaluate the possible role of exposure to power frequency magnetic fields in the etiology of hematopoietic neoplasia (Repacholi et al. 1997; McCormick et al. 1998; Harris et al. 1998).

Genetically Engineered Models Carrying Genes Associated with Human Leukemia

Over the past several years, a number of investigators have developed transgenic mouse models that carry genes (either BCRABL or AML1 fusion genes) that have been identified in human leukemias. The genetic basis of disease induction in these transgenic model systems provides a major advance in in vivo model development for the study of leukemia mechanisms, and for possible use as experimental platforms for hazard assessment studies. In different models, hematopoietic neoplasms originate in both the lymphoid and myeloid compartments of the bone marrow, and demonstrate considerable differences in both leukemia incidence and latency. Disease phenotypes for several of these model systems are summarized in Table 2.

When compared to other transgenic and nontransgenic model systems previously discussed, the genotypic and phenotypic relevance of transgenic mouse models that carry genetic alterations that are present in human leukemias provide major advantages for use in hazard assessment. In addition to mechanistic relevance, however, several other criteria must be evaluated in the process of model optimization. These include disease incidence and latency, reproducibility of disease patterns, predictiveness for human responses, and ability to up- or down-regulate neoplastic development. Because most transgenic models carrying human leukemia genes have been developed only very recently, there are often few relevant data available on which an assessment of model suitability can be based.

Disease incidence and latency are critical determinants of model suitability for hazard assessment: Should leukemia develop very rapidly and in high-incidence in otherwise untreated transgenic animals, it may be impossible to demonstrate a statistically significant increase in risk resulting from exposure to an exogenous agent. This limitation appears to apply to virtually all currently available BCR-ABL transgenic mouse models (see Table 2); in these models, the process of leukemogenesis is very robust, resulting in a high-incidence of clinical disease within weeks to a few months. The high-incidence of disease in these models may preclude demonstration of statistically significant increases in disease risk; similarly, the short latency of disease induction complicates any demonstration of accelerated disease induction in response to xenobiotic exposure. In their current stage of development, it is concluded that leukemia development in BCR-ABL transgenic mouse models may be too robust for these models to serve as sensitive hazard assessment platforms for ALL. However, differences in disease latency among several regulatable (conditional) BCR-ABL transgenic mouse models (Huettner et al. 2000) suggest that further development of transgenic lines in which leukemia develops more slowly may yield models systems whose disease incidence patterns are suitable for use in hazard assessment.

The converse situation is seen with transgenic mouse models carrying AML-1 fusion genes (Table 2). These mice fail to develop hematopoietic malignancies over their lifetimes, even though the fusion gene is expressed (Rhoades et al. 2000; Andreasson et al. 2001). The lack of leukemia response in these transgenic models may be interpreted two ways: (a) transgenic models carrying AML-1 fusion genes are generally unresponsive to leukemogenesis, and will therefore be insensitive as hazard assessment platforms; or (b) incorporation of the AML-1 fusion gene provides only the first “hit” in a multihit model, and additional events are required for the development of malignancy. If the latter interpretion is borne out, this type of model may prove to be ideal for leukemia hazard assessment. Such a model would provide a potentially highly sensitive approach to leukemogen detection, with minimal influence on tumors in transgenic control groups. However, successful induction of leukemia by a number of exogenous agents must be demonstrated in order to support this interpretation of model applicability. As seen with the BCR-ABL transgenics, this class of animal model holds considerable promise for leukemia hazard assessment, but requires substantial further development.

Reproducibility of disease incidence and latency patterns, model predictiveness for human responses, and model sensitivity to up- or down-regulation are other issues that must be considered in the evaluation of transgenic mice bearing BCRABL, AML fusion genes, or other human leukemia genes as models for leukemia hazard assessment. At the present time, the development stage of these models is not sufficiently mature to permit adequate evaluation of such parameters.

SUMMARY

Patterns of genetic alterations in pediatric leukemias and mathematical modeling based on epidemiology data both suggest that pediatric leukemias may be effectively modeled using a “multihit” paradigm. In the “multihit” model for childhood leukemia, an initial genetic alteration occurs during in utero or early postnatal development, but clinical disease develops only upon additional genetic or nongenetic events that occur during the postnatal period. Optimization of animal models for leukemia hazard assessment should incorporate this “multihit paradigm” and have the capacity to detect agents that influence leukemia development either early or late in the disease process. In the ideal case, experimental models used for leukemia hazard assessment will demonstrate (a) a genetic lesion that is concordant with genetic alterations identified in human pediatric leukemias, and (b) disease incidence and latency patterns that can be reproducibly up- or down-regulated by exogenous agents.

Lack of genetic concordance between experimental hematopoietic neoplasms and human leukemias is a key limitation to the use of many animal models to leukemia hazard assessment. This limitation applies to all standard-bred animal strains used in nonclinical safety assessment (e.g., Sprague-Dawley and F344 rats, CD-1 and B6C3F1 mice); to “high-incidence” animal leukemia models (e.g., AKR mouse); to multistage models using traditional animal strains, and to many genetically engineered animal models (e.g., p53 knockout and ras transgenic mice). By contrast, application of the multistage or “multihit” model to hazard assessment studies conducted in genetically engineered mice carrying human leukemia genes may provide a genetically relevant, highly sensitive approach to identify environmental agents that are important risk factors for childhood leukemia. Currently available transgenic mouse models bearing human leukemia genes demonstrate patterns of disease induction that are not optimal for use in hazard assessment, and the reproducibility and predictive nature of these model systems have not been validated. However, the development of new regulatable (conditional) mouse models carrying human leukemia genes, and/or manipulation of existing leukemia models to reduce disease incidence and increase disease latency appear to provide the most direct path to establish and optimize animal models to identify risk factors for childhood leukemia.