Hormesis Predicts Low-Dose Responses Better Than Threshold Models

Abstract

This study evaluated characteristics of the concentration-response relationships of chemicals from the U.S. National Cancer Institute (NCI) Yeast Anticancer Drug Screen database with respect to the threshold and the hormetic dose-response models. The database reported concentration-response studies of 2189 chemicals from a broad range of chemical classes. The biological end point was growth in 13 strains of yeast (Saccharomyces cerevisiae), most of which contain genetic alterations affecting DNA repair or cell cycle control. The analysis was limited to studies that satisfied a priori entry criteria for evaluation, including having two or more concentrations in the nontoxic zone (below a Benchmark Dose). The mean growth response compared to untreated controls of these doses was significantly greater than 100% in all 13 yeast strains, ranging from ~105% to ~111%. Under a threshold model, one would expect values more closely approximating 100%. Moreover, the distribution of responses below the BMD₅ for chemicals was shifted upwardly from the expectations of a threshold model for all strains. These results indicate that for the chemicals and yeast strains studied, the responses are more consistent with a hormetic model than a threshold model, and they strengthen previous results presented by Calabrese et al. (2006, Toxicol. Sci. 94:368–378). Taken together, the analyses provide strong evidence for hormesis, a phenomenon with a broad range of biomedical and toxicological implications.

Critical analysis and validation of low-dose-zone predictions by dose-response models, especially the threshold model, are important because this model has long been used as the default model for toxicological risk assessment in the United States for all noncarcinogenic chemicals entering commerce and food products (Beck et al. 2001). In some countries the threshold model may also be considered in assessing risks for carcinogenic agents (Seeley et al. 2001). The use of different default models has important implications in many areas, including the establishment of limits for chemical exposures. A better understanding of hormesis can also be important for the prevention of degenerative diseases and the development of safe, effective regimens for the treatment of cancer and other disorders. The implications of hormesis for a broad range of biomedical sciences and for risk assessment in toxicology and radiation biology have recently been reviewed (Calabrese 2008; Arumugam et al. 2006).

In a recent study Calabrese et al. (2006) used the U.S. National Cancer Institute (NCI) Yeast Anticancer Drug Screen database to compare the ability of a hormetic model and a threshold model to predict responses to prospective antitumor drugs at concentrations below estimated toxic thresholds. A Benchmark Dose (BMD) methodology was used to estimate toxic thresholds. The NCI database contains 2189 chemicals, with each agent studied over five concentrations with approximately half-log concentration spacing (1.2 to 100 μM). Thirteen yeast strains, including strains with unique genetic alterations that affect cell cycle regulation or DNA repair functions and their wild-type counterparts, were used to test the 2189 chemicals under similar conditions. Analysis of patterns of response below a broad range of BMDs (2.5% to 12.5%) showed that the patterns were more consistent with a hormetic model than with a threshold model. These findings were consistent with earlier analyses of below-threshold responses in published literature, showing the hormetic dose-response model to be more broadly applicable than the threshold model in assessments made across a range of environmental and pharmaceutical agents, biological models, and end points (Calabrese and Baldwin 2001, 2003). A limitation of the pattern analysis used in Calabrese et al. (2006) is that it did not evaluate the magnitude of the hormetic effect below the toxic threshold or the consistency of individual chemical responses with hormesis. The current investigation extends the analyses applied to the U.S. NCI Yeast Anticancer Screen database in Calabrese et al. (2006) by estimating the average magnitude of the hormetic effect for each strain and the distribution of the hormetic effect over chemicals.

METHODS

A detailed description of the NCI database, experimental design, and evaluation methods is given in Calabrese et al. (2006). Briefly, data from stage 2 of the NCI testing procedure were evaluated on 2189 compounds considered prospective antitumor agents based on preliminary testing. Each agent was tested at 1.2, 3.7, 11, 33, and 100 μM.

The chemicals were tested in 13 strains, 11 of which contain mutations in genes that can affect susceptibility to toxicants and radiation by altering the capacity for DNA repair or cell cycle controls (Simon 2001; Holbeck and Simon 2007). The genetic alterations of interest are bub3, mec2, mgt1, mlh1, rad14, rad18, rad50, rad52, sgs1, and overexpression of CLN2. None of the strains are wild type in a strict genetic sense, but the strains designated “wild type,” also called SPY50644 (MATa erg6Δ::LEU2 pdr1Δ::LEU2 pdr3Δ::hisG::URA3::hisG ade2 ade3 leu2 ura3), and SPY50780 (MAT α erg6Δ::TRP1 pdr1Δ::LEU2 pdr3Δ:: hisG ade2 ade3 leu2 trp1 ura3) carry the wild-type alleles of the genes of interest in the other strains (Holbeck and Simon 2007; http://dtp.nci.nih.gov/yacds/index.html; http://dtp.nci.nih.gov/yacds/spy50644.html; http://dtp.nci.nih.gov/yacds/spy50780.html). With the exception of strain rad50EPP⁺, the strains have enhanced susceptibility to chemicals owing to mutations in erg6, pdr1, and pdr3 (Dunstan et al. 2002). The erg6 mutation confers enhanced permeability by depleting membrane ergosterol, whereas the pdr1 and pdr3 mutations eliminate transcription factors required for the expression of chemical-efflux transporters and thereby confer pleiotropic drug resistance.

The responses in the NCI database were obtained from the growth of the yeast strain exposed to the compound relative to the growth of the same yeast strain in a solvent (i.e., DMSO) control. Yeast cells in the exponential phase of growth were inoculated into synthetic complete medium containing 2% glucose and the test chemical. The initial cell density was 10⁴ cells per well containing 200 μl of medium. Each agent was assessed four times at the same five concentrations in each yeast strain. Chemicals were tested in 96-well plates, with 80 chemicals at the same concentration on one plate. The remaining 16 peripheral wells were used as controls, of which 4 were unexposed controls, 8 solvent controls, and 4 controls containing cycloheximide. The assay was deemed invalid if growth occurred in the presence of cycloheximide. All concentrations of a drug were incubated over the same 12-h period on different plates such that there were five plates run on the same chemical at the same time. The chemical location in the 96-well plate was systematic rather than randomly allocated. Employing a different source of chemical on each day and different daily yeast cultures maximized variability in response.

The response data consisted of a ratio of the optical density (OD) of the response well for the treatment divided by the mean of the OD readings of eight solvent-control wells for each concentration. OD readings were at 600 nm. This process was repeated on a second day, and the ratios from the 2 days were averaged. We refer to the average response as the replication response. Two replication responses were produced for each concentration in each strain. Although the data on the NCI website give the average of the two response values and the difference between the two values, the original OD values are not available due to a computer malfunction at the Fred Hutchinson Cancer Research Center (FHCRC) facility. In light of the loss of the original data, special efforts were made to ensure that a detailed understanding of the entire research methodology and all quantitative methods was obtained prior to undertaking our analyses and that our written understandings were confirmed by the Principal Investigator (Julian Simon) of the original yeast study (Calabrese et al. 2007).

Identifying Chemicals with High Concentration Inhibition

The concentration-response patterns for chemicals with concentrations below an estimated BMD were evaluated in order to estimate the average magnitude of response below the toxic threshold. To achieve this goal within a five-concentration study design, a priori entry criteria were created. Evidence of toxicity (i.e., a response of ≤80% of the control, called the inhibition response) at the highest concentration (100 μM) was required, in addition to having either two or three concentrations (1.2 and 3.7 μM or 1.2, 3.7, and 11.0 μM) below the BMD, excluding concentrations used to determine the BMD. The BMD₅ was estimated by a linear interpolation on the log scale between the concentration immediately above and below the 95% response, similar to Calabrese et al. (2006). Concentrations used to derive the BMD were not used in the assessment of the predictions of the threshold and hormetic dose-response models.

Setting the Benchmark Dose (BMD)

The BMD₅ is the concentration at which the response is estimated to have decreased 5% below the control value (Crump 1984). The BMD₅ was chosen for the primary analysis because 5% is near the midpoint for the standard deviation of the yeast strains (which ranged from 3.0% to 7.5% for the 13 strains).

A BMD₅ was calculated for each of the 28,457 concentration-response experiments (2189 chemicals and 13 strains) using the average of two replications as the response. The BMD₅ was estimated through the following procedure:

1. The largest concentration with an average response below 95% was identified. Let this concentration be C _below, and let the associated response be R _below.

2. If the average response at the next smallest concentration was at least 95%, then let this concentration be C _above, and let the associated response be R _above. The BMD₅ is estimated by linear interpolation on the log concentration scale: BMD₅ = exp[log(C _above)+ (0.95 − R _above)(log(C _below)−log(C _above))/(R _below − R _above)].

3. If the average response at the next lowest concentration below C _below was less than 95%, then let this concentration be C _below with response R _below, and return to step 2.

Selection of Studies and Concentrations below the BMD₅

We identified all studies where there was one or more concentration below the concentration corresponding to C _above, which we refer to as the low-concentration range. We then constructed a set of data that included only the studies with one or more concentration in the low concentration range, and we analyzed responses from these studies for different chemicals for each strain of yeast. We report analyses for sets of chemicals with two concentrations below C _above separately from analyses for sets of chemicals with three concentrations below C _above. Our analysis does not include an assessment of chemicals with only one concentration below the estimated BMD₅ because such data are less robust than those for agents with two or three concentrations below the BMD₅ and because the only available concentration (i.e., 1.2 μM) is so close to the estimated toxic threshold that responses for some of these agents are apt to reflect residual (i.e., carryover) toxicity (Calabrese and Baldwin 2003). Data supporting the basis of this exclusion were published in Calabrese et al. (2006).

Evaluating Responses for Concentrations in the Low-Concentration Range

The response at each concentration was reported as the mean response constructed from two replications. The response for individual replications was not recorded, but the absolute value of the difference in response between the replications was reported. We added and subtracted one half of this difference to the mean response to form the basic replicate responses at each concentration.

For each yeast strain we use a linear mixed model to predict the average response in the low concentrations range for chemicals. In our models, response is reported as percent of control, as indicated in the methods. The model is specified in the context of a large population of possible chemicals, indexed by s = 1, . . . , N, where for each chemical, measures of response indexed by k = 1, . . . , r are made at each low concentration indexed by t = 1, . . . , d. In the yeast study, we fit separate models for chemicals with d = 2 low concentrations and for chemicals with d = 3 low concentrations. At each concentration, there are two replicates, such that r = 2. We represent a model for the kth measure of response at concentration t for chemical s as

$$Y_{stk}={\mu }_{st}+E_{stk}$$

where E (Y _stk ) = μ _st corresponds to the expected response at concentration t for chemical s. We represent var (E _stk ) = σ _st ². We allow the response variance to depend on the chemical and the concentration.

Our primary interest is in the average response for low concentrations, which we define for chemical s as ${\mu }_s=\frac{1}{M}{\sum }_{t=1}^M{\mu }_{st}$ . Using this definition, we define a concentration effect for chemical s as d _st = μ _st − μ _s . We use the mean response over all chemicals, which is given by ${\mu }_s=\frac{1}{M}{\sum }_{t=1}^M{\mu }_{st}$ to define the chemical effect as a _s = μ _s − μ. With these definitions, we can represent response for the kth replication of chemical s at concentration t by

$$Y_{stk}=\mu +{\alpha }_s+{\delta }_{st}+E_{stk}$$[1]

The only random term in the model is the replication error term, E _stk .

The mixed model that we fit is parallel to model (1). The difference is that we make the additional assumption that the chemicals included in the analysis correspond to a simple sample from the larger set of chemicals, and the concentrations below the BMD₅ are a simple random sample of concentrations from that range. We distinguish the chemicals in the population from a selected chemical by the subscript i, and index the selected concentration by the subscript j. Representing the chemical effect and the corresponding concentration effects as random effects, the resulting model is the mixed model given by

$$Y_{ijk}=\mu +C_i+D_{ij}+E_{ijk}$$[2]

where E(C _i ) = 0, E(D _ij ) = 0, var(C _i ) = σ _C ² and var(D _ij ) = σ _D ², with var(E _ijk ) = σ _e ², corresponding to the average replication variance over concentrations and chemicals (as indicated inStanek and Singer [2004]). This model is a nested mixed model and is fit using SAS Proc Mixed (Littell et al. 2006).

There are three variance components in model (2), and each has meaning. The first variance component σ _C ² is the variance of the distribution of mean response for chemicals. The size of this component provides evidence of the variability of response across chemicals. Relative to the average response for a chemical, σ _D ² provides a measure of how variable concentration-specific responses are from the chemical mean. This variance is formed by averaging the chemical-specific concentration variances over chemicals in the population. The third variance component, σ _e ², is the replication variance, averaged over chemicals and concentrations.

We fit this model to the yeast data for chemicals that have concentrations below the BMD₅ and estimated the mean response in addition to variance components for each strain of yeast. In addition, we used the results to predict the average response for each of the sample chemicals using the best linear unbiased predictor (BLUP), and we generated confidence intervals for these predictors.

RESULTS

Using the criterion of r _inhib = 80, concentration-response relationships were identified that met the criterion of a high-concentration inhibition (≤80% of control value at the highest concentration tested). The results indicate that 77% (21,977 of 28,457 studies) satisfied the criterion for inhibition at the highest concentration.

The concentration-response studies that were eligible showed some differences among strains. The strain designated “wild type” had the lowest proportion of chemicals satisfying the inhibition entry criterion at 67%, and the rad50 strain had the highest value (83%). The result for the wild-type strain is consistent with the fact that the genetic defects in most of the other strains would be expected to confer increased sensitivity to the toxic effects of some chemicals. In the case of rad50, which is deficient for recombinational repair by homologous recombination and nonhomologous end joining (Holbeck and Simon 2007), the enhanced sensitivity would encompass many agents that cause DNA damage.

Before exploring the low-concentration zone, we tabulated numbers of chemicals that gave a toxic response at the highest concentration. Nearly half the chemicals (1093 of 2189) had an inhibitory response equal or below r _inhib = 80 at their highest concentration (100 μM) for all the strains. In contrast, 164 chemicals had no concentration-response studies with an inhibitory response equal or below r _inhib = 80 at their highest concentration in any strain. There were 338 chemicals for which an inhibitory response was not apparent at the highest concentration for only one strain. Of these 338 chemicals, 229 (67.8%) occurred with the rad50EPP⁺ strain, which has wild-type membrane permeability and an efficient efflux transporter system, thereby reducing susceptability relative to the other strains tested.

Of the 21,977 studies that satisfied the high concentration toxicity requirement, 12,602 (57.3%) had at least one concentration below the BMD₅. There were 2451 studies that met the inhibition criterion at the highest concentration and had three concentrations below the BMD₅, whereas 5235 studies had two concentrations below the BMD₅, and 4916 studies had one concentration below the BMD₅.

Among the studies that showed evidence of high-concentration inhibition, the strain with the largest number of studies with concentrations below the BMD₅ is SPY50780, whereas rad50EPP⁺ had the lowest number (Table 1). The low number for rad50EPP⁺ is consistent with this strain’s chemical efflux pump and diminished permeability reducing susceptibility to toxicants. The high number for SPY50780 is consistent with its wild-type repair and cell cycle control having led to fewer concentration responses being eliminated for reasons of excessive toxicity. The strain-specific responses reflect the entry criteria whereby responses are eliminated for nontoxicity at high concentration or excessive toxicity at low concentration.

We were also interested in differences among chemicals for numbers of studies where there was one (or more) concentration below the BMD₅ (Table 2). We only used concentration-response curves that met the high-concentration inhibition criterion. Of the 2189 chemicals in the database, 2025 chemicals had one (or more) concentration-response with an inhibition at the highest concentration. Of these, 118 chemicals had an inhibition for all strains, along with concentrations below the BMD₅ range.

Table 3 summarizes the mean responses of each yeast strain for all chemicals with a BMD₅ value between 33 and 100 μM. For these chemicals, there are three concentrations below the BMD. The estimate, standard error, and variance components were evaluated by fitting model (2) to each strain. There was considerable variation among strains for numbers of chemicals with three concentrations below the BMD₅, with CLN2oe, SPY50780, and wild type having the most and rad50EPP⁺ having the fewest. The mean values for the 13 yeast strains range from 106.5% to 111.6%, compared to 100% for the control group’s cell proliferation. Table 3 provides information on variation for each yeast strain indicated by the SD for chemical, concentration, and replication separately and their integration in the standard error value. Notice that the estimated variance of response for concentrations about the mean response, σ̂ _D , is zero for most strains. The estimate of zero is most likely due to the small number of observations per chemical and the similar response for concentrations relative to the variability in replicated response at a concentration. For all strains except rad50 and rad50EPP⁺, the standard deviation between chemicals is larger than the response variance. A similar assessment was made for chemicals with BMD₅ values within the concentration range beginning at 11 μM but less than 33 μM. These chemicals have two concentrations (i.e., 1.2 and 3.7 μM) below the BMD and gave results consistent with those in Table 3 with respect to the mean increase across strains (i.e., ~5% to 9.5%) and the standard error (Table 4).

Figure 1 shows a graph of the predicted responses of 253 chemicals with three concentrations below the BMD for the wild-type strain. The figure was constructed by first ordering from smallest to largest the predictor of each chemical (considered to be a realized random effect in model (2)), and then plotting the predictor along with the 95% prediction interval (PI) in increasing order. The abscissa represents the response (in percent) under or over 100. The ordinate has been scaled to 100%, representing 100% of the chemicals in the analysis. Each of the 253 horizontal lines in the plot represents a 95% prediction interval for a chemical. The dark line at the center of the prediction intervals connects the predictors for adjacent chemicals.

The continuous line on the left (black line with a value of 0 at 50%) is a plot of the mean response for the chemicals under the assumption that a threshold model holds for each chemical in the analysis. The departure of that line from the vertical line at 0 (on the abscissa) is due to replication and concentration variability. It is calculated by multiplying the percentile from a standard normal distribution by $\sqrt{\frac{1}{d}({\widehat{\sigma }}_D^2+\frac{{\widehat{\sigma }}_e^2}{r})}$ , the estimated standard deviation of the chemical mean. Note that the threshold model predicts that some chemicals will display responses greater than 100% (upper end of the distribution) and that others will display responses less than 100% (lower end of the distribution). Such deviations from 100% are ascribable to chance.

A comparison of the actual data for the wild-type strain with the predicted responses of the threshold model indicates a consistent shift to the right across the entire distribution of 253 chemicals with three concentrations below the BMD₅ (Figure 1). These findings indicate that the threshold model does not match the responses in the low-concentration (i.e., subtoxic) zone for this set of data. The upshift in response occurs across the entire distribution of chemicals. All chemicals gave responses compatible with a hormetic dose-response relationship. Below the 65th percentile, the overall upshift in response is about 4% to 5%. However, there is a marked and progressive increase in the upshift starting around the 65th to 74th percentile of the chemicals.

A similar comparison is presented in Figure 2 for the wild-type strain for 394 chemicals with two concentrations below the BMD₅. The overall shape and quantitative features of the plotted data are very similar to Figure 1. Figure 3, which superimposes Figures 1 and 2 on each other, demonstrates that the findings are quantitatively very similar. Thus, the hormetic dose response predominated regardless of whether there were two or three concentrations below the BMD₅, situations that reflect the grouping of agents that differ in toxic potency by about threefold. As in the previous case, a response compatible with hormesis was observed for all chemicals assessed. Similar results for each of the remaining 12 yeast strains are given in Figure 4/Table 5 and Figure 5/Table 6. All responses are upshifted from the estimates based on a threshold dose-response model regardless of whether there were two or three concentrations below the BMD₅.

A similar response trend is seen for each of the 13 strains of yeast regardless of the BMD criterion (i.e., 2.5, 5.0, 7.5, or 10.0) and number of concentrations examined. Although the findings using different BMD criteria all support the occurrence of hormesis, there is an increase in the strength of the evidence for hormesis as the BMD criterion decreases from 10 to 2.5 (Figure 6). The most likely explanation is that the BMD₁₀ reflects a low level of toxicity. This residual toxicity diminishes as the BMD criterion decreases toward 2.5. All 13 yeast strains gave responses supportive of the hormesis model, regardless of their diverse genetic alterations.

DISCUSSION

The present analysis indicates that the distribution of predicted responses in the low-concentration range is upshifted in a manner that is inconsistent with the threshold dose-response model. Although this was the case across the entire distribution of chemicals, it was most striking for the upper 20% to 40% of chemicals. The upshift was consistent across the 13 strains and did not appear to depend on the inherent toxicity of the chemicals, as indicated by comparisons of chemicals with two or three concentrations below the BMD₅. Figure 7 demonstrates the consistency in the wild-type strain of this upshifted distribution of responses relative to the threshold model predictions for chemicals with two or three concentrations below the BMD₅. These findings are consistent across all strains. The general upshift in the predicted responses is consistent with a hormetic dose-response model. The findings confirm and extend an earlier report of Calabrese et al. (2006) in which a different mode of analysis of the same database showed that the hormetic dose-response model better predicted below-threshold responses than did the threshold dose-response model. Taken together, the two analyses provide a more substantial and integrative perspective of hormesis in this large yeast database.

The data reveal that about 20% to 40% of the chemicals, depending on the yeast strain, show strong evidence of hormesis on an individual basis. The remaining 60% to 80% give responses that also support the hormetic model, in that they consistently differ from the prediction of the threshold model in the direction expected for hormesis. The strong evidence of hormesis for some chemicals suggests that beyond a general stress response there may be effects dependent on chemical structure. However, the NCI study protocol used only one time point for the measurement of cell proliferation; therefore one cannot discern whether differences in response magnitude are more likely to be related to chemical structure or to temporal factors that can affect adaptive responses.

The methodology that we used adjusts for the likelihood that variability apparent with small sample sizes will regress toward the average variation with repeat testing/sampling. This approach, commonly referred to as best linear unbiased prediction (BLUP) or empirical Bayes, provides more accurate predictors of the true chemical mean response than the simple mean. The regression towards the mean affects only chemicals whose predictor differs from the mean, not the mean itself.

The present methodology compared the distribution of the predicted response in the low-concentration region to results anticipated on the basis of the threshold dose-response model with the same concentration and response variances. Although the actual data were consistently upshifted compared to expectations of the threshold model, some treatments yielded results not only <110% of control responses but also less than 100%. Although such low values would not normally be considered to display a hormetic effect, we regard chemicals with such responses as consistent with hormesis because they displayed a clear upshift in response compared to the threshold-model distribution. Thus, the below-BMD treatments were shown to be upshifted across the entire population of chemical agents.

We had hoped that the use of genetically altered yeast strains with well-characterized genetic alterations affecting DNA repair processes and cell cycle control might provide mechanistic insight into the hormetic stimulatory process, but the 13 strains were similar to one another in showing responses consistent with hormesis. However, further analysis of specific chemicals and yeast strains and their interactions may provide an opportunity to better define hormetic response pathways. The fact that the distribution of predicted responses consistently supported a hormetic model suggests that the hormetic response is general and not strongly affected by the diverse genetic changes built into the yeast strains of the NCI screening program.

These findings clearly indicate that elevated responses (over 100%) are to be expected for chemicals at concentrations in the low-concentration range in yeast, a conclusion consistent with the hormetic dose-response model. The threshold dose-response model was adopted in the 1930s and became accepted by national regulatory agencies for safety assessments without thorough critical evaluation with respect to the low-dose zone (Calabrese 2005a). In addition to the present findings, a large volume of peer-reviewed literature and experimental data are more consistent with a hormetic model than the threshold model for predicting responses to low-dose exposures (Calabrese and Blain 2005; Calabrese 2005b). These results further support calls for a reexamination of the threshold dose-response model and its use by regulatory agencies and suggests that the hormetic dose-response model may be useful for interpreting responses in the low-dose zone. Finally, these data suggest that antitumor drugs have the potential to enhance tumor cell proliferation at low concentrations. A recent comprehensive review of the published literature concerning human tumor cell lines is consistent with this perspective (Calabrese 2005c).