A review of mammalian carcinogenicity study design and potential effects of alternate test procedures on the safety evaluation of food ingredients.

Extensive experience in conducting long term cancer bioassays has been gained over the past 50 years of animal testing on drugs, pesticides, industrial chemicals, food additives and consumer products. Testing protocols for the conduct of carcinogenicity studies in rodents have been developed in Guidelines promulgated by regulatory agencies, including the US EPA (Environmental Protection Agency), the US FDA (Food and Drug Administration), the OECD (Organization for Economic Co-operation and Development) for the EU member states and the MAFF (Ministries of Agriculture, Forestries and Fisheries) and MHW (Ministry of Health and Welfare) in Japan. The basis of critical elements of the study design that lead to an accepted identification of the carcinogenic hazard of substances in food and beverages is the focus of this review. The approaches used by entities well-known for carcinogenicity testing and/or guideline development are discussed. Particular focus is placed on comparison of testing programs used by the US National Toxicology Program (NTP) and advocated in OECD guidelines to the testing programs of the European Ramazzini Foundation (ERF), an organization with numerous published carcinogenicity studies. This focus allows for a good comparison of differences in approaches to carcinogenicity testing and allows for a critical consideration of elements important to appropriate carcinogenicity study designs and practices. OECD protocols serve as good standard models for carcinogenicity testing protocol design. Additionally, the detailed design of any protocol should include attention to the rationale for inclusion of particular elements, including the impact of those elements on study interpretations. Appropriate interpretation of study results is dependent on rigorous evaluation of the study design and conduct, including differences from standard practices. Important considerations are differences in the strain of animal used, diet and housing practices, rigorousness of test procedures, dose selection, histopathology procedures, application of historical control data, statistical evaluations and whether statistical extrapolations are supported by, or are beyond the limits of, the data generated. Without due consideration, there can be result conflicting data interpretations and uncertainty about the relevance of a study's results to human risk. This paper discusses the critical elements of rodent (rat) carcinogenicity studies, particularly with respect to the study of food ingredients. It also highlights study practices and procedures that can detract from the appropriate evaluation of human relevance of results, indicating the importance of adherence to international consensus protocols, such as those detailed by OECD.

Hierarchical models for probabilistic dose-response assessment.

Probabilistic risk assessment is gaining acceptance as the most appropriate way to characterize and communicate uncertainties in estimates of human health risk and/or reference levels of exposure such as benchmark doses. Although probabilistic techniques are well established in the exposure-assessment component of the National Research Council's risk-assessment paradigm, they are less well developed in the dose-response-assessment component. This paper proposes the use of hierarchical statistical models as tools for implementing probabilistic dose-response assessments, in that such models provide a natural connection between the pharmacokinetic (PK) and pharmacodynamic (PD) components of dose-response models. The results show that incorporating internal dose information into dose-response assessments via the coupling of PK and PD models in a hierarchical structure can reduce the uncertainty in the dose-response assessment of risk. However, information on the mean of the internal dose distribution is sufficient; having information on the variance of internal dose does not affect the uncertainty in the resulting estimates of excess risks or benchmark doses. In addition, the complexity of a PK model of internal dose does not affect how the variability in risk is measured via the ultimate endpoint.

Classification ensembles for unbalanced class sizes in predictive toxicology.

This paper investigates the effects of the ratio of positive-to-negative samples on the sensitivity, specificity, and concordance. When the class sizes in the training samples are not equal, the classification rule derived will favor the majority class and result in a low sensitivity on the minority class prediction. We propose an ensemble classification approach to adjust for differential class sizes in a binary classifier system. An ensemble classifier consists of a set of base classifiers; its prediction rule is based on a summary measure of individual classifications by the base classifiers. Two re-sampling methods, augmentation and abatement, are proposed to generate different bootstrap samples of equal class size to build the base classifiers. The augmentation method balances the two class sizes by bootstrapping additional samples from the minority class, whereas the abatement method balances the two class sizes by sampling only a subset of samples from the majority class. The proposed procedure is applied to a data set to predict estrogen receptor binding activity and to a data set to predict animal liver carcinogenicity using SAR (structure-activity relationship) models as base classifiers. The abatement method appears to perform well in balancing sensitivity and specificity.

Using dose addition to estimate cumulative risks from exposures to multiple chemicals.

The Food Quality Protection Act (FQPA) of 1996 requires the EPA to consider the cumulative risk from exposure to multiple chemicals that have a common mechanism of toxicity. Three methods, hazard index (HI), point-of-departure index (PODI), and toxicity equivalence factor (TEF), have commonly been considered to estimate the cumulative risk. These methods are based on estimates of ED(10) (point of departure) and reference doses from the dose-response functions of individual chemicals. They do not incorporate the actual dose-response function of the mixture from multiple chemical exposures. Dose addition is considered to be an appropriate approach to cumulative risk assessment because it assumes that the chemicals of interest act in accordance with a common mode of action (a similar action). This paper proposes a formal statistical procedure to estimate the cumulative risk by fitting the dose-response model of the mixture under dose addition. The relative potency between two chemicals is estimated directly from the joint dose response model of the mixture. An example data set of four drugs representing four chemicals is used to illustrate the proposed procedure and compare it to the HI, PODI, and TEF methods.

Statistical issues in the analysis of low-dose endocrine disruptor data.

UNLABELLED: The National Institute of Environmental Health Sciences (NIEHS) and the U.S. Environmental Protection Agency (U.S. EPA) recently cosponsored the Endocrine Disruptors Low-Dose Peer REVIEW: The purpose of this meeting was to examine data supporting the presence or absence of low-dose effects of endocrine disruptors in specific studies and then to evaluate the likelihood and significance of these and/or other potential low-dose effects for humans. All invited speakers agreed to provide their raw data in advance of the meeting to a Statistics Subpanel, which was asked to reevaluate the authors' experimental design, data analysis, and interpretation of experimental results. The purpose of this statistical reevaluation was to provide an independent assessment of the experimental design and data analysis used in each of the studies and to identify key statistical issues relevant to the evaluation and interpretation of the data. This paper presents a summary of the Statistics Subpanel's evaluation. Specific examples are presented to illustrate problems that arose in the experimental design and data analysis of certain studies. The statistical principles and issues that are discussed in this paper are not unique to endocrine disruptor studies and should provide important guidelines regarding appropriate experimental design and statistical analysis for other types of laboratory investigations.

A mechanistic approach to modelling the risk of liver tumours in mice exposed to fumonisin B1 in the diet.

Data from the National Toxicology Program's carcinogenesis study of fumonisin B1 in B6C3F1 mice, conducted at the National Center for Toxicological Research, were used to fit the Moolgavkar-Venzon-Knudson (MVK) two-stage, clonal-expansion model of carcinogenesis. In addition to tumour data from the conventional 2-year bioassay, the study included data on tissue weights, cell proliferation, cell death, and sphingolipid metabolism in primary target organs. The model was used to predict 2-year liver tumour rates in female and male mice based on differences among dose groups in the effect of fumonisin B1 on the growth of normal tissue and on the proliferation of preneoplastic cells as a compensatory response to sphinganine-induced cell death. Fumonisin B1 was assumed to be non-genotoxic, i.e. the model did not include any effect of fumonisin B1 on either of the two mutation rates of the MVK model. The model was able to reproduce reasonably well the observed tumour rates in both female and male mice, predicting substantially increased rates above background only at the highest doses of fumonisin B1 in females.

Dose-response trend tests for tumorigenesis adjusted for differences in survival and body weight across doses.

A relationship between rodent body weight and tumor incidence for some tissue/organ sites has been demonstrated in many studies. It is not uncommon for a chemical tested for carcinogenicity to also affect body weight due to toxicity and/or food consumption. In such cases, comparisons of tumor incidence may be biased by body weight differences across dose groups. A simple procedure was investigated for reducing this bias. This procedure divides the animals into a few groups on the basis of body weight. Body weight at 12 months was used, before the appearance of a tumor was likely to affect body weight. Statistics for dose-response trend tests are calculated within body weight strata and pooled to obtain an overall dose-response trend test. This procedure is analogous to stratifying animals on the basis of age at the time of removal from a study to account for differences in ages of animals across dose groups that can affect comparisons of tumor incidence. In this paper, differences in survival times of animals were adjusted by the Poly-3 technique used by the National Toxicology Program. This technique does not require the assignment of cause of death. Several examples from rodent chronic bioassays were investigated, where the high dose group had reduced body weights and associated reductions in tumor incidence. When we analyzed the data by body weight strata, some positive dose-response trends for tumor incidence were demonstrated. In one case, the body weight adjusted analysis indicated that a negative dose-response trend in tumor incidence was a real effect in addition to a body weight reduction. These examples indicate that it is important to consider the effects of body weight changes as low as 10%, and perhaps less, as possibly being caused by chemicals in 2-year bioassays for carcinogenesis. The simple procedure of analyzing tumor incidence within body weight strata can reduce the bias introduced by body weight differences across dose groups.

Incorporating model uncertainties along with data uncertainties in microbial risk assessment.

Much research on food safety has been conducted since the National Food Safety Initiative of 1997. Risk assessment plays an important role in food safety practices and programs, and various dose-response models for estimating microbial risks have been investigated. Several dose-response models can provide reasonably good fits to the data in the experimental dose range, but yield risk estimates that differ by orders of magnitude in the low-dose range. Hence, model uncertainty can be just important as data uncertainty (experimental variation) in risk assessment. Although it is common in risk assessment to account for data uncertainty, it is uncommon to account for model uncertainties. In this paper we incorporate data uncertainties with confidence limits and model uncertainties with a weighted average of an estimate from each of various models. A numerical tool to compute the maximum likelihood estimates and confidence limits is addressed. The proposed method for incorporating model uncertainties is illustrated with real data sets.

Risk assessment for quantitative responses using a mixture model.

A problem that frequently occurs in biological experiments with laboratory animals is that some subjects are less susceptible to the treatment than others. A mixture model has traditionally been proposed to describe the distribution of responses in treatment groups for such experiments. Using a mixture dose-response model, we derive an upper confidence limit on additional risk, defined as the excess risk over the background risk due to an added dose. Our focus will be on experiments with continuous responses for which risk is the probability of an adverse effect defined as an event that is extremely rare in controls. The asymptotic distribution of the likelihood ratio statistic is used to obtain the upper confidence limit on additional risk. The method can also be used to derive a benchmark dose corresponding to a specified level of increased risk. The EM algorithm is utilized to find the maximum likelihood estimates of model parameters and an extension of the algorithm is proposed to derive the estimates when the model is subject to a specified level of added risk. An example is used to demonstrate the results, and it is shown that by using the mixture model a more accurate measure of added risk is obtained.

Bioassays of shortened duration for drugs: statistical implications.

Declining survival rates in rodent carcinogenesis bioassays have raised a concern that continuing the practice of terminating such studies at 24 months could result in too few live animals at termination for adequate pathological evaluation. One option for ensuring sufficient numbers of animals at the terminal sacrifice is to shorten the duration of the bioassay, but this approach is accompanied by a reduction in statistical power for detecting carcinogenic potential. The present study was conducted to evaluate the loss of power associated with early termination. Data from drug studies in rats were used to formulate biologically based dose-response models of carcinogenesis using the 2-stage clonal expansion model as a context. These dose-response models, which were chosen to represent 6 variations of the initiation-promotion-completion cancer model, were employed to generate a large number of representative bioassay data sets using Monte Carlo simulation techniques. For a variety of tumor dose-response trends, tumor lethality, and competing risk-survival rates, the power of age-adjusted statistical tests to assess the significance of carcinogenic potential was evaluated at 18 and 21 months, and compared to the power at the normal 24-month stopping time. The results showed that stopping at 18 months would reduce power to an unacceptable level for all 6 submodels of the 2-stage clonal expansion model, with the pure-promoter and pure-completer models being most adversely affected. For the 21-month stopping time, the results showed that, unless pure promotion can be ruled out a priori as a potential carcinogenic mode of action, the loss of power is too great to warrant early stopping.

Percentiles of the product of uncertainty factors for establishing probabilistic reference doses.

Exposure guidelines for potentially toxic substances are often based on a reference dose (RfD) that is determined by dividing a no-observed-adverse-effect-level (NOAEL), lowest-observed-adverse-effect-level (LOAEL), or benchmark dose (BD) corresponding to a low level of risk, by a product of uncertainty factors. The uncertainty factors for animal to human extrapolation, variable sensitivities among humans, extrapolation from measured subchronic effects to unknown results for chronic exposures, and extrapolation from a LOAEL to a NOAEL can be thought of as random variables that vary from chemical to chemical. Selected databases are examined that provide distributions across chemicals of inter- and intraspecies effects, ratios of LOAELs to NOAELs, and differences in acute and chronic effects, to illustrate the determination of percentiles for uncertainty factors. The distributions of uncertainty factors tend to be approximately lognormally distributed. The logarithm of the product of independent uncertainty factors is approximately distributed as the sum of normally distributed variables, making it possible to estimate percentiles for the product. Hence, the size of the products of uncertainty factors can be selected to provide adequate safety for a large percentage (e.g., approximately 95%) of RfDs. For the databases used to describe the distributions of uncertainty factors, using values of 10 appear to be reasonable and conservative. For the databases examined the following simple "Rule of 3s" is suggested that exceeds the estimated 95th percentile of the product of uncertainty factors: If only a single uncertainty factor is required use 33, for any two uncertainty factors use 3 x 33 approximately 100, for any three uncertainty factors use a combined factor of 3 x 100 = 300, and if all four uncertainty factors are needed use a total factor of 3 x 300 = 900. If near the 99th percentile is desired use another factor of 3. An additional factor may be needed for inadequate data or a modifying factor for other uncertainties (e.g., different routes of exposure) not covered above.

Dose-response trend tests for tumorigenesis, adjusted for body weight.

Several studies have demonstrated a relationship between rodent body weight and tumor incidence for some tissue/organ sites. It is not uncommon for a chemical tested for carcinogenicity to also affect body weight. In such cases, comparisons of tumor incidence may be biased by body-weight differences across dose groups. A simple procedure was investigated for reducing this bias. This procedure divides the animals into a few groups based on body weight. Body weight at 12 months was used, before the appearance of a tumor was likely to affect body weight. Statistics for dose-response trend tests are calculated within body weight strata and pooled to obtain an overall dose-response trend test. This procedure is analogous to that currently used, of stratifying animals, based on their age at the time of removal from a study. Age stratification is used to account for differences in animal age across dose groups, which can affect comparisons of tumor incidence. Several examples were investigated where the high-dose group had reduced body weights and associated reductions in tumor incidence. When the data were analyzed by body-weight strata, some positive dose-response trends for tumor incidence were demonstrated. In one case, the weight-adjusted analysis indicated that a negative dose-response trend in tumor incidence was a real effect, in addition to a body weight reduction. These examples indicate that it is important to consider the effects of body weight changes as low as 10%, and perhaps below, that were caused by chemicals in 2-year bioassays for carcinogenesis. The simple procedure of analyzing tumor incidence within body-weight strata can reduce the bias introduced by weight differences across dose groups.

Caloric restriction abolishes enhanced metabolic efficiency induced by ectopic agouti protein in yellow mice.

Caloric restriction (CR), from approximately 3 months of age, at 70% of the ad libitum (AL) caloric intake prevented development of overt obesity in female "viable yellow" Avy/A (BALB/cStCrlfC3Hf/Nctr x VY/WffC3Hf/Nctr-Avy) F1 hybrid mice. In adult yellow Avy/A mice, caloric restriction eliminated the increased metabolic efficiency associated with the presence of agouti protein in ectopic sites. At 4 weeks of age, the yellow Avy/A mice were approximately 14% heavier and by 12 weeks of age, when caloric restriction began, they were approximately 24% heavier than the congenic agouti A/a mice. Between 4 and 12 weeks, the yellow mice gained approximately 63% in body weight, whereas the agouti mice gained only approximately 44%. While the comparable AL Avy/A mice gained approximately 128% and the AL A/a mice gained approximately 41% between 12 and 51 weeks of age, the CR Avy/A and A/a mice gained only 16% and 15%, respectively. Mean brain weights of CR mice of both genotypes were lower than those of the comparable ad libitum-fed (AL) groups; however, CR Avy/A mice had slightly, but significantly (P < 0.0001), higher brain weights than CR A/a mice. The larger mean brain weight and retention, during caloric restriction, of the somewhat greater prerestriction Avy/A mean body weight compared with prerestriction A/a mice were consonant with the hypothesis that ectopic agouti protein affects somatic growth directly or indirectly. This may be related to altered developmental/metabolic programming in yellow mice, indicated by greater metabolic efficiency and by an early transient increase in circulating IGF-1 levels. The specific cellular processes modulated by the agouti protein in ectopic sites remain to be identified.

A unified approach to risk assessment for cancer and noncancer endpoints based on benchmark doses and uncertainty/safety factors.

A fundamental goal of toxicology is to determine safe levels of human exposure to toxic substances. In the absence of information to establish dose-response relationships at low exposure levels generally experienced by humans, high-dose to low-dose linear extrapolation is generally used for estimating carcinogenic risks and the no-observed-adverse-effect-level divided by uncertainty (safety) factors is widely used for establishing human exposure guidelines for noncancer effects. The basis and impact of this dichotomy is examined and questioned. It is proposed that a unified approach be adopted for establishing human exposure guidelines for both cancer and noncancer endpoints. It is suggested that a lower confidence limit on the dose estimated to produce an excess incidence of adverse health effects in 10% of the individuals in a human study or 10% of the animals in laboratory experiments be used as a point-of-departure. This dose would be divided by appropriate uncertainty factors to establish human exposure guidelines. For severe irreversible adverse health effects we suggest a total default uncertainty factor (divisor) for animal data on the order of 10,000, which is comparable to current guidelines. For reversible biological effects a smaller default uncertainty factor on the order of 1000 may be employed. This is comparable to the divisor often used currently when the point-of-departure is the lowest-observed-adverse-effect-level. It is asserted that the toxicological information generally available does not warrant numerical estimates of risk at low levels of human exposure. Rather, we support a unified approach for all adverse health effects of dividing a benchmark dose by appropriate uncertainty factors to establish guidelines for human exposures to toxic substances.

Combining uncertainty factors in deriving human exposure levels of noncarcinogenic toxicants.

Acceptable levels of human exposure to noncarcinogenic toxicants in environmental and occupational settings generally are derived by reducing experimental no-observed-adverse-effect levels (NOAELs) or benchmark doses (BDs) by a product of uncertainty factors (Barnes and Dourson, Ref. 1). These factors are presumed to ensure safety by accounting for uncertainty in dose extrapolation, uncertainty in duration extrapolation, differential sensitivity between humans and animals, and differential sensitivity among humans. The common default value for each uncertainty factor is 10. This paper shows how estimates of means and standard deviations of the approximately log-normal distributions of individual uncertainty factors can be used to estimate percentiles of the distribution of the product of uncertainty factors. An appropriately selected upper percentile, for example, 95th or 99th, of the distribution of the product can be used as a combined uncertainty factor to replace the conventional product of default factors.

A comparison of methods of benchmark-dose estimation for continuous response data.

Methods of quantitative risk assessment for toxic responses that are measured on a continuous scale are not well established. Although risk-assessment procedures that attempt to utilize the quantitative information in such data have been proposed, there is no general agreement that these procedures are appreciably more efficient than common quantal dose-response procedures that operate on dichotomized continuous data. This paper points out an equivalence between the dose-response models of the nonquantal approach of Kodell and West and a quantal probit procedure, and provides results from a Monte Carlo simulation study to compare coverage probabilities of statistical lower confidence limits on dose corresponding to specified additional risk based on applying the two procedures to continuous data from a dose-response experiment. The nonquantal approach is shown to be superior, in terms of both statistical validity and statistical efficiency.

Maternal epigenetics and methyl supplements affect agouti gene expression in Avy/a mice.

'Viable yellow' (Avy/a) mice are larger, obese, hyperinsulinemic, more susceptible to cancer, and, on average, shorter lived than their non-yellow siblings. They are epigenetic mosaics ranging from a yellow phenotype with maximum ectopic agouti overexpression, through a continuum of mottled agouti/yellow phenotypes with partial agouti overexpression, to a pseudoagouti phenotype with minimal ectopic expression. Pseudoagouti Avy/a mice are lean, healthy, and longer lived than their yellow siblings. Here we report that feeding pregnant black a/a dams methyl-supplemented diets alters epigenetic regulation of agouti expression in their offspring, as indicated by increased agouti/black mottling in the direction of the pseudoagouti phenotype. We also present confirmatory evidence that epigenetic phenotypes are maternally heritable. Thus Avy expression, already known to be modulated by imprinting, strain-specific modification, and maternal epigenetic inheritance, is also modulated by maternal diet. These observations suggest, at least in this special case, that maternal dietary supplementation may positively affect health and longevity of the offspring. Therefore, this experimental system should be useful for identifying maternal factors that modulate epigenetic mechanisms, especially DNA methylation, in developing embryos.

Oral squamous cell carcinoma in ad libitum-fed and food-restricted Brown-Norway rats.

A high incidence of oral squamous cell carcinoma was present in male and female Brown-Norway rats fed ad libitum or food-restricted dietary formulations. One hundred eight-nine rats were examined from 4 dietary treatment groups: male ad libitum, male food-restricted, female ad libitum, and female food-restricted. The ad libitum treatment groups for both males and females had significantly more cases of oral squamous cell carcinoma than cohort food-restricted groups. In ad libitum rats, 10 of 47 (21%) males and 15 of 47 (32%) females had oral squamous cell carcinoma, whereas only 4 of 47 (9%) males and 5 of 48 (10%) females in the food-restricted groups were similarly affected. The food-restricted rats lived significantly longer than ad libitum cohorts, so the higher incidence of squamous cell carcinoma was not dependent on extended lifespans. In addition to the dietary influence, a significant difference in oral squamous cell carcinoma incidence occurred between various familial lines. Family lines having representatives in both ad libitum and food-restricted groups had lower oral squamous cell carcinoma incidences in the food-restricted group whether comparing affected litters or individuals. Results suggest that the incidence of oral squamous cell carcinoma in our colony of Brown-Norway rats can be influenced by both the dietary treatment group and genetic predilection within certain pedigrees.