Tobacco Documents Online

---

---

Environmental and Molecular Mutagenesis 29:312- 322 (1997) Predicting Rodent Carcinogenicity From Mutagenic Potency Measured in the Ames Salmonella Assay Bethel A. Fetterman,! Byung Soo Kim,2 Barry H. Margolin,° Jonathan S. Schildcrout,1 Melissa G. Smith,1 S. Michelle Wagner,~ and Errol Zeigera ~ Department of Biostatistics, University ~f North Carolina, Chapel Hill, North Carolina 2Department of Applied Statistics, Yor~sei University, Seoul, South Korea 3Environmental Toxicology Program, Nation&l Ins#tote of Environmental Health Sciences, Research Triangle Park, North Carolina Many in vitro tests have been developed to identify chemicals that can damage cellular DNA or cause mutations, and secondarily to identify potential car- cinogens. The test receiving by far the most USe and attention has been the Salmonella (SAL) mutagene- sis test developed by Ames and colleagues [(1973): Proc Natl Acad Sci USA 70:2281-2285; (1975): Murat Res 31:347-364], because of its initial promise of high qualitative (YES/NO) predictivily for cancer in rodents and, by extension, in humans. In addition to the initial reports of high qualitative predictivity, there was also an early report by Mes- elson and Russell [in Hiatt HH et al. (1977): "Ori- gins of Human Cancer, Book C: Human Risk Assess- ment," pp 1473-1481] of a quantitative relation- ship between mutagenic potency measured in SAL and carcinogenic potency measured in rodents, for a small number of chemicals. However, other re-" ports using larger numbers of chemicals have found only very weak correlations. The primary purpose of this study was to determine whether mutagenic potency, as measured in a number of different ways, could be used to improve predictivity of carci- nogenicity, either qualitatively or quantitatively. To this end, eight measures of SAL mutagenic potency were used. This study firmly establishes that the pre- dictive relationship between mutagenic potency in SAL and rodent carcinogenicily is, at best, weak. When predicting qualitative carcinogenicity, only qualitative mutagenicity is useful; none of the quanti- tative measures of potency considered improves the carcinogenicity prediction. In fact, when qualitative mutagenicity is forced out of the model, the quantita- tive'measores are still not predictive of carcinogenic- ify. When predicting quantitative carcinogenicity, several, possible methods were considered for sum- marizing potency over all experiments; however, in 611 cases, the relationship between mutagenic po- tency predictors and quantitative carcinogenicity is very weak. Environ. Mol. Mutagen. 29:312-322, 1997. © 1997 Wiley4.iss, Inc. Key words: statistics; National Toxicology Program; carcinogenic potency; mutagenic potency; Salmonella INTRODUCTION Many in vitro short-term tests (STTs) have been devel- oped to identify chemicals that can damage cellular DNA or cause mutations and, secondarily, to identify potential carcinogens. The somatic mutation theory of carcinogene- sis holds that mutations are a common first step in the development of a cancer cell; consequently, interest in these STI's has been fueled by their ability to identify potential carcinogens. The test receiving by far the most use and attention has been the Salmonella (SAL) muta- genesis test developed by Ames and colleagues [1973, 1975]. The SAL test gets its most extensive use as a prelimi- nary screen during chemical and drug development. The results of this test are often the only toxicology-related information used by industry to decide whether to proceed © 1997 Wiley-Liss, Inc. with development of the chemical and t~o more definitive toxicological tests, or to label it a potential carcinogen, and put it ~ide, with no further testing. Only qualitative responses are used in this process. It would be extremely . valuable for these users of SAL to know if there is a relationship between SAL test potency and rodent cancer potency because it could permit them to proceed with development of chemicals estimated to be weak carcino- gens. Often there are no animal toxicity data at this point. Contract grant sponsor:. National Institute of Environmental Health Sci- ences: Contract grant number:. 273-90-1-0005; Contract grant sponsor: Korea Research Foundation (1995 Overseas Research Program for Uni- versity Professors. *Correspondence to: Dr. Barry Margolin, Department of Biostatistics. CB# 7400. University of North Carolina, Chapel Hill, NC 27599. Received 29 June 1995; revised and accepted 18 September 1996.

TABLE h Correlation of Mutagenic and Carcinogenic Log Potencies Among Chemicals That are Both Mutagenic in SAL and Carcinogenic in Rodents Study # of Chemicals Correlation Predicting Carcinogenicity From Mutagenic Potency 313 TABLE Ih Comparison of Operating Characteristics of the 73- and 42-Chemical Databases, and their Combination" 73 42 115 chemicals chemicals chemicals Meselson and Russell [ 1977] I 0 0.94~" Meselson and Russell [19771 14 0.36a Meselson and Russell [ 19771 9 0.92~ Parodi et al. [19901 88 0.39 McCann et al. [1988] 80 0.40 McCann et al. [19881 77 0.24d Piegorsch and Hoel [ 19881 28 0.44 aAs cited by Parodi et al. [1990]. bNitroso chemicals not considered. ~As.calculated by McCann et al. [19881 using 9 non-nitroso chdmicals. '~After elimination of "3 extreme values." SAL positives (%) 33 Carcinogen[city positives (%) 60 Positive predictivity ¢%) 83 Negative predictivity (%) 51 Sensitivity (%) 45 Specificity (%) 86 Concordance with rodent cancer (%) 62 Significance of association (P value) 29 32 56 " 59 1130 89 62 55 52 48 100 91 73 66 0.004 <0.001 <0.001 ~Adapt~d froha Haseman et al. [1990] and Zeiger et al. [1990]. The current study is the first in a series of studies that examines the predictivity of carcinogen[city from mutage- nicity, assessed by using four in vitro STTs (SAL, mouse lymphoma, sister chromatid exchange, and chromosome aberrations), alone or in combination. A number of differ- ent potency measures for the Ames SAL test are evaluated and applied to estimate rodent carcinogen[city. The primary force behind the development and exten- sive use of the SAL mutagenicity test has been its ease of use, its low cost, and its initial promise of high qualitative (YES/NO) predictivity for cancer in rodents and, by ex- tension, in humans, Early reports showed approximately 89-95% predictivity of cancer from SAL results, and 87- 93%. concordance between the results of SAL tests and rodent cancer tests [McCann et al., 1975; Purchase et al., 1976; Sugimura etal., 1976]. Later reports using chemi- cals tested by the U.S. National Toxicology Program (NTP) demonstrated a wider range of positive predictivi- ties (69-100%) and lower concordances (61-66%) [Ten- nant et al., 1987; Zeiger, 1987; Zeiger and Tennant, [986; Zeiger et al., 1990]. In addition to the initial reports of high qualitative pre- dictivity, there was also an early report of a significant and substantial quantitative relationship between mutagenic potency measured in SAL and carcinogenic potency mea- sured in rodents. Meselson and Russell [1977] examined the quantitative relationship between mutagenic and car- cinogenic potency for 14 carcinogens and reported a nearly perfect linear relationship (correlation = 0.94) be- tween carcinogenic and mutagenic log potency for 10 of the 14 chemicals. This was particularly exciting because it appeared to extend the SAL assay from the prediction solely of carcinogenic hazard to a test that could also predict carcinogenic risk. Subsequent attempts to replicate this high correlation by using larger numbers of chemi- cals, however, have not been successful (Table I). Because of its interest in the relatiqnship between muta- genicity and carcinogen[city, the NTP developed a data- base of mutagenicity test results in parallel with its carci- nogenesis testing program. The study reported here evalu- ates two groups of chemicals taken from the NTP database: 73 tested for carcinogen[city in rats and mice by the NTP in studies ending December 1976 or later, with peer-review approval dates of January 1985 or earlier [Tennant et al., 1987] and an additional 42 chemicals~ peer-reviewed and accepted between January 1985 and May 1988 [Haseman et al., 1990; Zeiger et al., 1990]. Both groups of chemicals had similar operating character- istics with respect to predicting rodent carcinogen[city [Haseman et al., 1990; Zeiger et al., 1990]. Table II illus- trates the results of the carcinogen[city and SAL mutage- nicity tests for these two groups of chemicals. These studies showed that for the NTP database, which reflects chemicals that are of interest to the federal regula- tory agencies, other government organizations, and the public, the SAL test provides good positive predictivity for carcinogens (89%), but the test "misses" 52% of the carcinogens (48% sensitivity). The studies by Tennant and colleagues [1987], Haseman and colleagues [1990], and Zeiger and colleagues [ 1990] were designed to exam- ine the qualitative relationships between SAL mutagenic- ity and rodent carcinogen[city. There still remains a need to examine thoroughly the quantitative relationships be- tween SAL mutagenicity and rodent carcinogen[city re- sults to determine whether the use of quantitative mea- sures (potency) of the former improves predictivity of the latter. ~ These reports discuss 41 chemicals, rather than 42. This is because tetrakis(hydroxymethyl)phosphonium chloride and tetrakis(hydroxy- methyl)phosphonium sulfate were tested as two separate chemicals but were evaluated as one chemical because they differed only in the salt and produced similar responses in the four STFs and in the carcinogenic- ity assays. However, the measures of interest for our purposes--muta- genic potency and TD50--are different for these two chemicals, so they remain distinct in our evaluation.

314 Fetterman et al. TABLE III. Frequencies of Salmonella Strain and Activation Experiments for Each Dataset 73 chemicals 42 chemicals Activation TA98 TA 100 TA 1535 TA98 TA 100 TA 1535 None 215 213 207 131 130 108 Rat liver 208 227 214 141 136 103 Hamster liver 213 223 216 142 137 103 Just as qualitative results can vary according to the decision" mechanism used to define a positive response, so there can be more than one working definition of the potency of a response. A number of measures of potency have been used for SAL mutagenicity data; they fall into two general categories: those based on the slope of the dose response, and those based on the concept of an effec- tive dose (e.g., the lowest effective dose). Similarly, po- tency measures of carcinogenicity can be based on the slope of the carcinogenic response, or on an effective dose (e.g., the TD50 [Gold et al., 1984]) as used here. In the current study, eight measures of potency were defined for the SAL test and were compared for their degree of agreement with each other and for their predictivity for carcinogenicity. The potency measures were used to ana- lyze the 73-chemical dataset, and the results obtained were validated by using the 42-chemical dataset. A key consideration in this study was whether the inferences drawn would change depending upon the SAL potency measure used. DESCRIPTION OF THE DATA Because of the specific SAL test protocol used for the NTP database, test data are available for all chemicals in strains TA98 and TAI00. However, as a rule, only chemi- cals that were negative or equivocal in these two strains bta were tested in strains TA1535 or TA1537. If TA1537 were included here, it would contribute 734 experiments (19% of the total database) and be heavily biased toward the nonmutagens. An earlier study [Zeiger et al., 1985] showed that approximately 89% of the NTP-identified mutagens could be identified through the use of strains TA100 and TA98, and 88% through strains TAI00 and TA1535. TA1537 uniquely detected only 4% of the muta- gens. Previous analyses of the NTP database included SAL strain TA1537 [Tennant et al., 1987; Haseman et al., 1990; Zeiger et al., 1990], but this strain was excluded from all of our analyses because it is not currently in use by the NTP as a primary screen, and it has not been used consistently in the past. The dataset used here for the Ames SAL assay consists of three tester strains--TA98, TAI00. and TA1535-- or and three metabolic activation systems--no activation, rat liver S-9. and hamster liver S-9. Table III shows the frequency distribution for strain and activation in each chemical dataset. There are 1936 and 1131 SAL experiments available for estimating potency in the 73-chemical and 42-chemi- cal datasets, respectively. Each of these experiments con- tains a solvent control dose and at least one treatment dose. Table IV shows a frequency distribution for the number of analyzable treatment doses for each dataset; 93% of the experiments in the 73-chemical dataset and 85% in the 42-chemical dataset consisted of a solvent control and five treatment doses, reflecting the design of the NTP SAL protocol. DESCRIPTION OF MUTAGENIC POTENCY MEASURES There is no universally accepted way to measure the potency of a mutagenic response in any existing assay. In order to be thorough, eight measures of mutagenic potency were considered. The first three measures are "specific to the SAL assay because they attempt to esti- mate low-dose mutagenic potency after adjustment for toxicity" [Margolin et al.. 1994]. The other five measures are more general and directly applicable to other STTs. The bM potency measure, described by Margolin and coworkers [1981 ], is based on a family of nonlinear dose- response models of the SAL assay. These models describe "po, the probability that a single treated microbe yields a mutant colony when exposed to dose D of a test chemical, by po = {I - exp[-(a + bD)]}*T(D) where (a + bD) is ~0, and T(D) is one of two functions describing the toxicity of dose D of the test chemical, either T(D) = exp(-gD), T(D) = [2 - exp(gD)]÷,

Predicting Carcinogenicity From Mutagenic Potency TABLE IV. Frequency Distribution for Number of Treatment Doses 315 Number of analyzable treatment doses Dataset I 2 3 4 5 6 7 8 9 73 chemical 3 II 21 88 1798 I1 0 4 0 42 chemical 6 3 24 120 957 14 0 3 4 where [x]÷ = max(0, x) and (gD) is >0. In these models, b reflects the mutagenic effect per unit dose, adjusted for concomitant toxicity. Doses that exhibit substantial toxicity are excluded from the analysis [Margolin et al., 1988]. The estimate of b from the model, denoted by bra, is used as an estimate of mutagenic potency and reflects the low-dose slope of the dose response. Bernstein and colleagues [1982] describe an estimate of potency based on the assumption that the underlying dose-response curve is initially linear. The slope of the regression line is computed under the assumption of iin- earity; if the observed mean response at the highest dose significantly departs from linearity of the remaining re- sponses, the observations at the highest dose are set aside, and the slope is recomputed from regression on the re- maining data. The process is repeated until no significant departure from linearity occurs. If there are at least three doses (including the control dose) with data remaining, then the estimate of the slope of the regression line for the remaining observations is used as the estimate of po- tency and is denoted by bB~; otherwise the potency esti- mate is undefined. A slight modification of this procedure yields bB2, in which the estimate of the slope of the regression line is used even if there are only two doses with observations remaining after the discard proces~ is concluded. bx An estimate of potency commonly appearing in the mutagenicity literature is the maximum observed pairwise slope for each dose group relative to controls, that is, the maximum observed average colony count per plate per unit dose applied. Let Yk be the mean number of revertants per plate observed at dose Dk, where 0 = Do < Dt < • • • < Dr and r = number of treatment doses. Then bx = max[(Yk - Yo)l(Dk - Do)], k where k is between I and r inclusive. bR br~ is the straightforward estimate of the slope of the linear regression of observed counts per plate regressed on dose, which equals ba~ when no dose groups are dis- c~rded. bo!, bD2 Margolin and Risko [1988] describe the dose per unit increase (DUI) potency measure as the dose needed to induce a unit increase in response over control, based on a model that allows a quadratic dose-response curve. Two estimates of DUI are considered: bot is the inverse of the DUI measure obtained from a linear regression that includes the quadratic term only if it is statistically sig- nificant; b02 is the inverse of the DUI measure obtained from a linear regression that includes the quadratic term whether or not it is statistically significant. Because the inverses of the respective DUI measures are used, bot and b02. are directly comparable to other potency measures. bLED Th~ lowest effective dose (LED) is the lowest dose that has a mean colony count per plate significantly greater than the mean count per plate at the control dose. Significance is assessed using a variance estimate based on the nonlinear dose-response model described for the measure bM and a Bonferroni correction for multiple com- parisons. Note that for all potency measures described except LED, a smaller value corresponds to lower muta- genic potency; for LED a smaller value implies higher mutagenic potency. Therefore, to make the comparison of potency measures simpler, the reciprocal of LED-- denoted bLED--is used as the estimate of mutagenic po- tency. DESCRIPTIVE STATISTICS FOR POTENCY MEASURES The measures bx and bR will always be defined as long as there are data for the control and one treatment dose; however, there are cases in which the other potency mea- sures may not be defined. The Bernstein measures and bB2 require at least two and one remaining treatment doses, respectively, after excluding high doses that depart

316 Fetterman et al. TABLE V. Percent of Experiments for Which Potency Measure Is Computable Dataset b~ bnt bn2 bx b~ bDi bo2 73 chemical 94 77 100 100 100 99.7 99.7 94 42 chemical 93 77 100 100 100 99.7 99.7 93 TABLE Vl. Percentiles for bM Before and After Logging Percentile bM In( 1 + bM) 0% 0.00 0.00 25% 0.00 0.00 50% 0.00 0.00 75% 0.05 0.05 90t~ 1.93 1.08 95% 55.47 4.03 99% 2035.94 7.62 100% 15864.36 9.67 from linearity of the response. The bM measure requires at least two. treatment doses remaining after excluding high doses that exhibit substantial toxicity. The bL~ mea- sure has the same restriction as that for br~ because it relies on results from the nonlinear dose-response model used for that measure. Table V shows the percent comput- ability of the eight measures for the NTP data. Because the potency measures bM, bin, and bB_, are so similar in principle, these measures were examined more thoroughly to see if they produce similar potency results. The two Bernstein measures are very similar; results for the comparison of bM with bat are almost identical to results for the comparison of br~ with There are 203 experiments for which bB~ or bB2 are defined, but b~a is undefined. Examination of the observed dose-response curves for these experiments shows that, in most cases, there is a peak response at the zero dose or at the first treatment dose, thereby explaining why b~ is undefined. The distribution of estimated potencies obtained for each measure is highly skewed. Logging the measures creates distributions that are better behaved, although still not normally distributed, and have a reasonable range for analysis purposes. The exact transformation used is In (I + potency); adding 1 to the potency before logging allows ~s to deal with potency measures that yield value 0 for some experiments. As an example of the effect of logging on the distribution of estimated potencies, Table VI shows the percentiles for the unlogged and logged b~ measure. Table VII shows the observed Pearson correlations be- tween the various logged potency measures.The correla- tion between bDI and bD2 is very high (r = 0.98) because the two measures are so similar in definition. In fact, the maximum absolute difference between logged bo~ and bm is 5.4, and 99% of the experiments have an absolute difference of 0.91 or less between the two measures. The correlations of bm and ba: with bu are high as well (r = 0.94). This high correlation is expected because each of these measures attempts to estimate the slope of the dose- response curve at low doses. The correlations of ba2 and bu with bx are also high (r > 0.92). Recall that bx is the maximum observed slope between the control and all test doses. For dose-response curves that level off or turn down at high doses, the value for bx would be obtained from the low doses; this should then be similar to ba~. and bM, which also estimate the slope at low doses. SUMMARIZING SALMONELLA EXPERIMENTS Condensing the data for each element in the database is key to the analyses described here. The first issue is how to handle replicate experiments. For a given chemi- cal, within each of the nine strain/activation combinations considered, several replicate experiments were generally available. Although McCann and colleagues [ 1988] used the median to summarize replicate experiments, the choice here was to use the mean as a summary measure. The mean and median summaries are highly correlated (r = 0.96) in this database because most of the experi- ments have only two replicates (in which case the mean and median are identical). Using the maximum as a sum- mary measure was judged to be misleading; if one repli- cate experiment has abnormally high results, its use as a summary measure would not be representative. Thus for each of the eight methods of measuring mutagenic po- tency, there are up to nine possible stra#dactivation co~n- bination potency measures estimated per chemical, ob- tained by averaging replicates. The next step in the summarization of the data before the prediction analysis is to attempt to summarize all the experiments for a given chemical across the different strain/activation combinations. Rather than considering nine different potency measures for a chemical (one for each strain/activation combination), it may be easier to consider only one potency measure per chemical. Pie- gorsch and Hoel [1988] used the maximum over all exper- iments (MAX), ignoring strain and activation distinctions, to summarize mutagenic potency for a given chemical. However, this is not the only possibility. Six possible methods of summarizing the many experiments per chem- ical are considered here (Table VIII): these will be re- ferred to as potency summar3" statistics.

Predicting Carcinogenicity From Mutagenic Potency TABLE VII. Observed Correlation Matrix for Logged Measures 317 b.u 1.00 bs~ .94 1.00 bs:.94 .91 1.00 bx .92 .94 .94 b~ .79 .77 .79 bo~ .89 .87 .89 bo, .91 .89 .91 bt~o .85 .81 .83 1.00 .77 1.00 .86 .90 1.00 .88 .89 .98 1.00 .79 .82 .87 .89 TABLE VIII. Potency Summary Statistics--Methods of Summarizing All Experiments for a Given Chemical 1. MAX of all experiments [Piegorsch and Hoel, 1988] 2. MEAN of all experiments 3. MEDIAN of all experiments 4. MAX of all 9 strain/activation combinations (averaged over replicates) 5. MEAN of all 9 strain/activation combinations (averaged over replicates) 6. MEDIAN of all 9 strain/activation combinations (averaged over replicates) PREDICTING QUALITATIVE CARCINOGENICITY USING SALMONELLA MUTAGENICITY The main purposes of this investigation were to evalu- ate a number of different measures of SAL mutagenic potency, and to examine the predictivity of the different potency measures for rodent carcinogenicity. Because there is no standard or generally accepted measurement of SAL mutagenic potency, we evaluated a number of different measures that had been used or proposed. Eight measures of potency were selected for this evaluation; most of the potency measures were a function of the mutagenic dose-response, but one was simply a function of dose, without considering the magnitude of the re- sponse. For each of the eight SAL potency measures con- sidered, we computed summary mutagenic potency pre- dictors for each strain/activation combination, and in- cluded the overall qualitative NTP mutagenicity decision. We then compared the predictive abilities of these eight potency measures and determined separately the best pre- dictors for qualitative and quantitative carcinogeni~ity. In order to validate findings obtained, a "split-sample" approach was used for this exploratory analysis. Previous comparisons of the 73- and 42-chemical datasets [Hase- man et al., 1990; Zeiger et al., 1990] have shown that they are not significantly different. Various models of predictivity were examined using the 73-chemical dataset and validated using the 42-chemical dataset. Because vali- dation results from the 42ochemical dataset were almost identical to results from the 73-chemical dataset, the sam- TABLE IX. Models Chosen for Predicting Qualitative Carcinogenicity, for Each Potency Measure Considered Predictors used in logistic regression* Final model selected 1. MUTA not included 2. MUTA included, but not forced into the model 3. MUTA forced into the model first intercept-only model intercept + MUTA intercept + MUTA *In addition to the 9 strain/activation combinations. ples were combined after validation and results are re- ported for the entire set of 115 chemicals. The analyses reported by Tennant and colleagues [1987], Haseman and colleagues [1990], and Zeiger and colleagues [1990] examined the prediction of qualitative (YES/NO) rodent carcinogenicity as a function of qualita- tive (YES/NO) mutagenicity. In our study, the initial anal- ysis was for the prediction of qualitative rodent carcinoge- nicity as a function of qualitative and quantitative mutage- nicity as measured by SAL. Logistic regression is the approach adopted here to explore the prediction of quali- tative (YES/NO) carcinogenicity. All of the mutagenic chemicals in the 42-cbemical dataset have positive carci- nogenicity. Logistic models cannot be fit to data that have only one response level for a particular value of the pre- dictor. Therefore, for the purpose of this analysis, one arbitrarily chosen chemical from the dataset was changed to be mutagenic and noncarcinogenic, so that verification analyses could be performed. Separately for each of the eight individual potency measures, the nine logged strain/activation combinations were considered as predictors in stepwise logistic analy- sis. The effect of qualitative SAL mutagenicity (the NTP decision, denoted by MUTA) on prediction was examined by comparing the results obtained by forcing MUTA to be included first in the model, including MUTA as a predictor but not forcing inclusion, and excluding MUTA from the model. Within each of these three procedures, a stepwise logistic regression analysis was performed for each of the eight potency measures. The same model was chosen regardless of which potency measure was

318 Fetterman et al. 4 2 1 o Maximum of Logged Rat & Mouse TDSOs Fig. 1. Frequency distribution for maximum of logged rat and mouse TD50s. considered. Table IX lists the predictors used in the step- wise logistic regression procedures and the final model chosen by each procedure, with a significance level of 0.10. When MUTA was excluded from the predictor list, an intercept-only model was statistically significant; none of the nine strain/activation combination predictors were selected for inclusion. When MUTA was included, whether by force or not, MUTA and the intercept were the only predictors with any significance. Another interesting question is, if MUTA is already included in the model, which strain/activation combina- tion best fits next in the model? For each potency measure, we determined which strain/activation combination would be included after MUTA. The "best choice" among all models and all potency measures was strain TA1535 with rat liver activation, which has a P value of 0.16 from the score test for inclusion after MUTA. All other P values are much higher, indicating that no other strain/activation combination even comes close to improving the predic- tivity of carcinogenicity once MUTA is included in the logistic model. Next, for each individual potency measure, an assess- ment was made as to whether any of the overall potency summary statistics listed in Table VIII might be a better predictor than the individual strain/activation combina- tions. Following the same three procedures described above for examining the effect of MUTA, and including the six potency summary statistics in the list of predictors, the same results as those shown in Table IX were ob- tained. The models discussed above compare different meth- ods of summarizing all of the potency information for a chemical, for a given method of estimating mutagenic potency. Finally, the different methods of calculating mu- tagenic potency were compared by including the eight different potency measures in the logistic regression pre- dictor list. The effect of MUTA on the prediction was also considered, as described previously. For both the 73- chemical and the 42-chemical datasets, the results were again the same as those shown in Table IX, regardless of which summary statistic described in Table VIII was used to obtain the potency measures. When the models were fitted to the combined sample of 115 chemicals, results were identical to those shown in Table IX, with two ex- ceptions. When MUTA was not considered in the predic- tion, bR and bB_, were chosen as predictors of carcinogenic- ity for the summary method of MAX over all nine strain/ activation combinations, and bD~ and bx were chosen as predictors of carcinogenicity for the summary method of MEDIAN over all nine strain/activation combinations. Conclusion I No matter how mutagenic potency is measured or sum- marized, only qualitative mutagenicity is useful for the prediction of qualitative rodent carcinogeni¢ity; none of the quantitative measures of SAL mutagenicity improves the prediction of qualitative carcinogenicity. PREDICTING QUANTITATIVE CARCINOGENICITY USING SALMONELLA MUTAGENICITY One measure of the carcinogenic potency of a chemical is the TD50. defined as the "dose rate (in mg/kg body weight/day) which, if administered chronically for the stan- dard life span of the species, will halve the probability of remaining tumorless throughout that period" [Gold et al., 1984]. Carcin~enic potency is inversely related to TD50. Estimates of the TD50 for both rats and mice are available for 67 of the chemicals in the exploratory. 73-chemical

TABLE X. Models Chosen for Predicting Quantitative Carcinogenicity From Strain/Activation Combinations, for All Chemicals Potency Predictors in F-test measure final model R" P-value Predicting Carcinogenicily From Mutagenic Potency 319 TABLE XI. Models Chosen for Predicting Quantitative Carcinogenicity From Strain/Activation Combinations and Overall MAX, for Combined Sample of 108 Chemicals Pot.~ncy F-test measure Predictors in final model R: P-value bM hi535" 0.05 0.0164 bM bBi hi00. rl00 0.13 0.0097 bn~ bB., hi00. n98, n1535 0.10 0.0108 bx hi00. r1535 0.11 0.0023 bx ba r1535 0.09 0.0019 bR bD~ rl535 0.09 0.0017 bo~ bD., r1535 0.07 0.0057 bD2 bLeD h 1535 0.08 0.0042 *Strain/activation combination. The first letter represents the activation (none, hamster liver, rat liver) and the numbers represent the Salmonella strain. h 1535* 0.05 0.01 64 hi00, rl00 0.13 0.0097 MAX. hi(X), n98, n1535 0.15 0.0023 MAX, rl535 0.13 0.0010 r 1535 0.09 0.13019 rl535 0.09 0.0017 r1535 0.07 0.0057 h ! 535 0.08 0.0042 *Strain/activation combination. The first letter represents the activation (n_one, hamster liver, _rat liver) and the numbers represent the Salmonella strain. dataset, and for 41 of the chemicals in the confirmatory 42-chemical dataset. Figure 1 provides the frequency distri- bution of logged TD50s for the 108 chemicals for which TD50s are available. Note that the noncarcinogenic chemi- cals have higher TD50s than do the carcinogenic chemi- cals, but that there is substantial overla'p. The single carcinogenic potency measure used in the analyses reported here is CARCIN = max[ln(l + rat TD50), ln(l + mouse TD50)]. The TD50s are logged because the predictors are logged; thus the outcome variable will be on the same scale as the predictors. For any one carcinogen, the maximum of rat and mouse TD50s among all the sexes and tissues examined is used as a conservative summary. The '*split-sample" approach was used for these analy- ses as well. Exploratory analysis was performed on the 67-chemical dataset, and validation was performed on the 41-chemical dataset. The samples were then combined, and results are reported for the entire set of 108 chemicals. In order to predict quantitative carcinogenicity (CAR- CIN), stepwise linear regression models were used. First, for each of the eight SAL potency measures, the nine strain/activation combinations were considered as pre- dictors of CARCIN. Table X provides the final model and the model's R-" value for each of the potency mea- sures. The strain/activation combination(s) chosen by the stepwise regression procedure for predicting TD50 vary with the potency measure considered. However, in all cases, R2 is very small, indicating that the relationships between the mutagenic potency measures and TD50 are weak. Piegorsch and Hoel [1988] used the maximum over all experiments (MAX) as a summary mutagenic potency for a given chemical. In order to examine the effect of this TABLE XII. Models Chosen for Predicting Quantitative Carcinogenicity From 6 Summary Statistics, for Combined Sample of 108 Chemicals Summary F-test statistic Predictors in final model Rz P-value bM bx bR bD2 b~ intercept only intercept only MAX over reps, MEAN over reps 0.12 0.0015 MAX over reps, MEAN over reps 0.15 0.0002 MAX over reps, MEAN Overall 0.10 0.0031 MAX over reps, MEDIAN over reps 0.08 0.0150 MAX over reps, MEDIAN over reps 0.08 0.0117 intercept only measure on prediction, MAX was included in the list of predictors and the above stepwise linear regressions were repeated. Results of this analysis are shown in Table XI. In the analysis of the 73-chemical dataset, MAX was the only predictor in the final model for all potency measures except bB~. In the analysis of the 42-chemical dataset, MAX did not enter into any of the final models. In the analysis of the combined sample, MAX enters the final model for two of the potency measures, but other strain/ activation combinations are included as well. In the split- sample and combined analyses, all R: values are very small, indicating very weak relationships between the pre- dictors and TD50. Differences in models chosen among the analyses of different samples can be attributed primar- ily to sample size. Conclusion II When only the nine strain/activation combinations are used to predict quantitative rodent carcinogenicity, the best model for predicting quantitative carcinogenicity depends on the potency measure being considered. Inclusion of the overall MAX in the model does little to improve prediction;