two million rodent carcinogens? the role of sar and qsar in their detection

Mutation Research, 305 (1994) 3-12 3 © 1994 Elsevier Science B.V. All rights reserved 0027-5107/94/$26.00

MUT 05347

INTERNATIONAL COMMISSION FOR PROTECTION AGAINST

ENVIRONMENTAL MUTAGENS AND CARCINOGENS

Two million rodent carcinogens? The role of SAR and QSAR in their detection

J o h n A s h b y *

Zeneca Central Toxicology Laboratory, Alderley Park, Macclesfield, Ches. SKI O 4TJ, UK

(Received 16 August 1993) (Accepted 27 August 1993)

Keywords: SAR, QSAR; Rodent carcinogens; Prediction of carcinogenicity

Summary

The accurate prediction of chemical carcinogenicity can only be achieved by a balanced consideration of the following factors: the chemistry and metabolism of the test agent, the interaction between toxicity and genetic toxicity, the possibility of non-genotoxic events that trigger subsequent non-targeted mutagenesis, the difference between activities observed in vitro and in vivo, and the possible inadequacy and/or partiality of all datasets and observations. Extrapolation of activities within a series of congeners is usually possible, but predictions across different chemical classes/mechanisms of carcinogenicity are difficult. Artificial intelligence systems can be used to predict one or more of the above parameters given adequate learning sets, but the hope for a single, coherent and self-contained method of predicting all instances of carcinogenicity is unreal. The future of carcinogen/mutagen prediction lies with data-rich artificial intelligence systems based on known mechanistic principles used selectively within the context of chemical and biological human insight. The major current obstacle to progress is the assumption that mutagenicity and carcinogenicity are unitary phenomena that can be learned and predicted by artificial intelligence systems operating in isolation.

* Corresponding author.

This manuscript is listed as ICPEMC Library No. 0097.

ICPEMC is affiliated with the International Association of Environmental Mutagen Societies (IAEMS) and the Institute de la Vie. Secretary ICPEMC: Ms. T. Berndt, c /o Hazleton Laboratory, 9200 Leesburg Turnpike, Vienna, VA 22182, USA.

Vainio et al. (1991) have described 55 discrete situations in which humans have succumbed to chemically induced carcinogenesis. Gold et al. (1991) have described the activities of 533 organic chemicals found to be carcinogenic after evaluation in rats and mice. Chemical Abstracts (1992) makes reference to over 4.5 million chemical entities (vide infra, Fig. 2). Gold and Ames (1990) have observed that one half of all natural and

SSDI 0027-5107(93)E0154- I

synthetic chemicals tested to date are carcinogenic to rodents. One can therefore estimate that about 2 million rodent carcinogens await detection. St ructure-act iv i ty relationships (SAR) should have a key role to play in this task. Some of the issues to be faced if SAR are to fulfill this task are discussed in the present paper, itself written mindful of the 5 papers that follow it - that ground is not covered twice. Quantitative SAR (QSAR) are a refinement of SAR; this chapter is concerned mainly with SAR.

Structural alerts to genotoxicity

James and Elizabeth Miller (review: Miller and Miller, 1977) provided the insight that most chemical carcinogens contain a critical electrophilic atom around which the rest of the molecule is arrayed either as an aid or as an encumbrance to its electrophilic reactivity a n d / o r stability. Some carcinogens are naturally electrophilic while others become electrophilic as the result of metabolic transformation. Metabolic transformations are potentially of infinite variety

due to the different classes and balances of activating and deactivating enzymes in different cell types of a tissue, in different tissues of different sexes of different strains of different species of different genera. Carcinogenicity and mutagenicity, the most frequently monitored phenotypic changes induced by electrophiles, are therefore also expected to be variable phenomena. Two premises underly the prediction of potential human mutagens and carcinogens. First, that effects produced in cultured cells using Aroclor-induced rat-liver $9 mix are predictive of effects likely to occur in any or all tissues of rodents. Second, that effects observed in rodents are likely also to occur in exposed humans.

The major chemical groupings associated with electrophilic attack on DNA (genotoxicity) have been gathered into a model structure (Fig. 1; Ashby 1985; Tennant and Ashby 1991; Ashby and Tennant, 1991; Ashby and Paton, 1993 and refs. cited therein). Use of this model structure to predict possible genotoxic carcinogenesis and mutagenesis requires informed caution - an alert to activity is provided, not a confirmation of activity.

(t) Halogenated methanes co0 4 X=H, F, CI, Br, I in any combination

(a)

?CH3 0 ~-S ~0

i (b) O\(d ) .> ........... (NO 2

(r) ;' - - ',, (c) /N ----=~, ,,CH 3 O ~ ,S ! /:,/ (h) NlN(g) ,,:~......:~/>--N ..CH 3

', ........... ,, c.=oH CliO o . . /H (q) //, ,,, I I ,,ell 3

"~.---..-C. H,IN-----< ~ ...... N---CH,~---CH-NH--N (t') H2N" \CH '= "~'~ / ' " "CH 3

i (s) / ..... '" ",~. (i) O~=N - -N /e l l2 ---Ci ...... C H2-~,~,,, , /~'-CH " CH -CI O)

I CH 2 ~ N(CH 2 CH 2 CI)2 (P) CH3 I CH2CI CH ,,; ......... ,,

0 (n) I "~CH---CH2--~'.' N.--Cl (k)

oJ',..= ""% ') (o) (u) NO2 xl,l ........... CH2(I):

0, 7 ........ O

Fig. 1. Composite model structure indicating substituents or moieties associated with genotoxicity (DNA reactivity). For discussion of this model see Tennant and Ashby (1991).

The learning set for the model structure (Fig. 1) is the learning set of chemistry itself - essentially completed for the present purposes by the start of the present century (Fig. 2). This last observa- tion leads to the most important conclusion of this paper, namely, that as chemical carcinogenicity is so evidently a chemical as well as a biological phenomenon it seems essential that its prediction should involve chemical as well as biological input. The critical learning set that led to the 'rules' of chemical reactivity and metabolic transformation are well known to chemists and they should form an active input to model systems based on artificial intelligence (AI). A1 systems must remain as valuable aids rather than as an end in themselves. An AI model system may reveal correlations of value that have not been identified before, but such correlations should not automatically be accepted at face value, attempts should be made to rationalize such associ- ations mechanistically and chemically. For example, the recent finding using the CASE system that the -CH 2 - 0 - substructure is associated with clastogenicity (Rosenkranz and Klopman, 1992) makes no immediate sense, and until it does it should be used with care.

Standard alerts to non-carcinogenicity

Certain substitucnts in a molecule predispose to non-carcinogenicity (Ashby, 1985). Examples

No of chemicals (xlO s) 5o 45- 4o 35 3o 25 2o

5,000 12,000

1o 5 ! o ~

1900 1925 Year

Structural Ned Learning Set essentially in place

1950 1975 lgg2

Fig. 2. Rate of discovery of organic chemicals leading to the 4.5 million now known. The basics of organic chemistry were

in place by 1900 A.D.

Steric inhibition of amino group

NH 2

C H 3 ~ 3

SOaH

J solubility/conjugation labile site

Fig. 3. Model chemical showing 3 possible features that may predispose to non-carcinogenicity.

are sulphonic acids which lead to rapid excretion due to enhanced water solubility, steric crowding of substituents that might otherwise be metabolized to an electrophile, and labile metabolic sites in a molecule that would lead to rapid bio- deactivation in vivo (Fig. 3). Despite the potential value of such factors as modifiers of carcinogenic potential (Fig. 3) their use is dangerous and their relative weightings in predictive methods should be low. This is vividly illustrated by an erroneous prediction of possible carcinogenicity made by the author in 1978 (unpublished) for glycidol. Glycidol

/ O ~ (CH 2-- CH 2-- CH 2OH)

has a naturally electrophilic epoxide grouping and a -CH2OH group that might be expected to be metabolized to an acid group (-CO2H) leading to rapid elimination of the molecule from rodents. Seemingly consistent with this, early re- ports established glycidol as a direct acting mutagen to Salmonella that was rendered non-mutagenic by the addition of $9 mix. Surely, a case of a likely non-carcinogen to rodents. In 1990 the US NTP published the oral gavage carcinogenicity bioassay results for glycidol (Anon., 1990). It was carcinogenic to both sexes of rat (75 mg/kg) and mouse (50 mg/kg), inducing tumours in 15 different tissues, arguably the most effect carcinogen yet defined. Caution is indicated when predicting carcinogenicity or its absence for a chemical.

==.--- Nitrobenzenes

CI CH 3 ~ Alkylbenzenes

Chlorobenzenes

Fig. 4. Illustration of the 3 classes info which the chemical shown could be placed. From the viewpoint of carcinogenicity

the nitro group will dominate.

Grouping of chemicals by functional class reactivity

Sometimes in order to make predictions of carcinogenicity it is necessary to summon up courage and impose subjective classifications on a database using chemical knowledge. For example, the model chemical shown in Fig. 4 could be entered into each of 3 possible chemical classes, but in terms of its possible carcinogenicity/

mutagenicity it is only a derivative of nitroben- zene. To describe this material as a chloroben- zene or an alkylbenzene is in itself correct but practically confusing. Likewise, the two groupings of chemicals shown in Fig. 5 are chemically dis- tinct. Thus, while it is justifiable to study the horizontal series of haloalkanes as a series of congeners it is wrong to link the methyl groups in the vertical column. The simple practical expedi- ent of introducing each of these methyl com- pounds to a thiol will reveal two alkylating agents, 1 possible instance of trans-esterification [(-)] and 4 cases of non-reactivity. An instance of such a possible false association of chemically unre- lated structures occurred in the CASE learning sets for omeprazole (Ashby, 1992).

Chemical classification were recently imposed (Ashby and Paton, 1993) on the Gold database of 522 rodent carcinogens (Gold et al., 1991) and the Vainio database of 54 human carcinogens (Vainio et al., 1991). The strengths and weak- nesses of this approach to chemical grouping is discussed in detail therein.

R - " S H

R - - S - - C H s

( - - )

R ~ S ~ C H s

CHs /

CHs -

CH3 -

CHs -

CHs -

CH ~ -

CHs -1

H N - - P h

CI Et - - CI Pr -- CI

OH

O ~ Ph

/o

O--C \

E1

Bu ~ CI Bu t _ CI

Can be modelled: Relmlve • v o ~ l i t y

=reactivity • ,poml,city • mdckJct r e l ~ r • Idori¢

/ / O

o - - p

I \ OR OR

0 - - CH~ Cannot be modelled: ~ i ~ ~ ~ .

Fig. 5. Congeneric (horizontal) and non-congeneric (vertical) series of chemicals. Only 2 of the non-congenerics will react with R-SH.

Newly discovered structural alerts to carcinogenicity

With the aquisition of new data, new structural alerts to carcinogenicity will become apparent. For example, several additions to the model multi-carcinogen (Fig. 1) have been made since it was first described (see Tennant and Ashby, 1991). Likewise, it is probable that computer assisted SAR systems will discern structural alerts not hitherto realized. It is suggested, however, that such new activating fragments or biophores should be rationalized mechinstically before they are used predictively. In Fig. 6, two sets of superficially common fragments are shown that are in fact mutually inconsistent when viewed chemically - thus, were an AI model system to associ- ate the MeS-group of the phosphate and the MeS group of the thiophenol, known chemical differences between these two MeS functionalities should act as an alert to their probable false association by the AI model system.

In a similar way, the rodent carcinogenicity of the putatively non-genotoxic rodent carcinogen 2,3,7,8-tetrachlorodibenzdioxin (TCDD) and its analogs appears to be a function of the ability of each analogue to interact with the Ah receptor. In such cases, substructure fragments are not of significance per se, but only in so much as they enable metabolic stability and receptor interac- tions to remain high (Fig. 7). Thus, the azo ( - N = N - ) link in the bottom chemical of Fig. 7 maintains the overall shape and size of TCDD and this analogue remains active at least as a

CH 3 - -$ - - ~ NH=

CI CI

CI CI

CI A '%. -J~ S . - / " ~ - C I

Site of detoxiflcation in vivo

Fig. 7. Two active chloracnegens (shaded boxes) and three inactive analogues. The shaded box represents size require-

ments for Ah receptor interaction.

chloroacnegen. There is, however, no further significance to the azo group - it does not confer activity on other molecules.

The examples shown in Figs. 6 and 7 indicate that care is required when seeking new substructures associated with potential carcinogenicity - each potential new one should be challenged for chemical a n d / o r biological significance before being generally adopted. A clear distinction between genotoxic and non-genotoxic carcinogenesis should also be maintained.

Carcinogenicity as a continuous (non-binary) phenomenon

o C H , - - s - P . / / C H ~ ~ S

\ (OR) =

Fig. 6. Illustrations of how the same substructure (S-C = C or CH3-S) can have different chemical properties dependent

upon their chemical environment.

By whatever method carcinogenicity is predicted it is important to be clear about how cancer bioassay results are best described mathe- matically. The simplest assumption is that carcinogenicity is a binary phenomenon - an agent is either carcinogenic or non-carcinogenic. Such a

Agent A

Agent B

RAT MOUSE

o" ~ o" ?

+

Jr - -

T ' I ' 2 CARCINOGENS 2 NON-CARCINOGENS?

p: l p=l

Chemical NTP ~ ] + 1 Bioessay ,w...-- ] p=0.2

J - p=0.5

p=0

Fig. 9. Levels of carcinogenicity used by the US NTP. Three possible mathematical descriptions of these levels of evidence

are shown and discussed in the text.

Agent C K - -

Agent D - - L

i i

I 2 EQUAL CARCINOGENS?

Fig. 8. Two carcinogens (A and B) that of fer the chance of

being regarded as non-carcinogens if only partial databases exist for them. The selective carcinogens (C and D) are not

optimally grouped as carcinogens.

simple model allows the use of superficially meaningful estimates of the strength of a prediction - from p = 0 to p = 1, representing certainty of non-carcinogenicity and carcinogenicity, re- spectively. Reference to Fig. 8 suggests, however, that carcinogenicity cannot be so simply defined. According to how much data have been gathered, agents A and B (Fig. 8) could be regarded as either two carcinogens or as two non-carcinogens. Given that databases are often incomplete this is a serious concern as such instances could confound a learning set. Similarly, it is hard to accept

that the two highly specific carcinogens D and E are best described simply as two rodent carcinogens. The best prediction for agents C and D could well be argued to be for the average response of rodents to these agents, in which case each are predominantly non-carcinogens.

Another dimension is added by the terms used by the US NTP to describe the level of evidence for carcinogenicity (clear evidence, some evidence, equivocal evidence and no-evidence; Fig. 9). It is the usual practice of the NTP scientists to group EE with NE yielding the binary classification of + / - shown in Fig. 9. It is argued here, however, that the p value for a prediction of carcinogenicity should range across the C E / S E classes as shown in Fig. 9 (0.2 taken arbitrarily as a low certainty of carcinogenicity). In contrast, there is a viewpoint current that interprets EE as p = 0.5 for a prediction (shown on the right of Fig. 9). The exact mathematical meaning of the term EE will be important to gain agreement on before consideration of the predictions of car-

I p=l

p=O.2

p=l

Chemical

p=0.2

p=o.2

I I p=o

GENOTOXIC: / POTENT, MULTIPLE /

SITE/SPECIES /

(~ M L only

(3" R K only I Non-genotoxic

O I R TG only

EE (ie not even some evidence)

Genotoxic non-carcinogen

Non-carcinogens

Fig. 10. Representation of carcinogenicity as a continuum with putative non-genotoxic carcinogens mid-spectrum and requiring separate prediction.

cinogenicity made for the 44 agents currently under bioassay by the US NTP (Tennant et al., 1991).

An alternative and preferred method of pro- ceeding to that implied in Fig. 9 is to predict separately genotoxic and non-genotoxic carcinogenicity, as shown in Fig. 10. In Fig. 10 carcinogenicity is shown as 3 discrete parts of a full spectrum - genotoxic carcinogenicity, non-genotoxic carcinogenicity and non-carcinogenicity (in- cluding equivocal evidence). Rather than at- tempting to predict carcinogenicity or its absence for a chemical (the binary boxes) it is suggested that separate predictions of genotoxic and non- genotoxic carcinogenicity should be made (each with an associated probability value). When these predictions jointly become weak (p = 0.2 suggested) the agent should be regarded as a predicted non-carcinogen (p = 0.2-0).

The problem of changing classifications of activity

The permanent problem besetting attempts to discern SAR is that mutagenicity and carcinogenicity data are not absolute entities - they can vary legitimately. By a small change in an assay parameter a confirmed positive response can become a confirmed negative response, and vice versa. Data for SAR learning sets, either based on artificial intelligence or human deduction, are usually gathered without reference to the precise conditions of test, and as a consequence they are intrinsically variable. A few of these sources of variability are shown schematically in Fig. 11.

Activities for a carcinogenic and mutagenic alkylating agent are shown in Fig. 11 to be capable of genuine variations that can confound derived correlations.

R-Cl NBP Salm. CA in vitro GT in viva GT in viva Carcin. Carcin. Test Site of appl. Systemic Bioassay Bioassay

Topical Systemic

~ , . . - - - ~ ~ ~ . . ~ : : . . ~ - - ~ ~!ii~iii~ ~ -

CH CH CH CH I 21 2 I 21 2 CI CI GS CI

• 4 - Ve

- - V8

ip dosing

50 wk '

:~.,..:~:$i$~:~:~$~$~ • ..

102 wks

rats

inhalation Fig. 11. Levels at which mutagenicity and carcinogenicity data can be measured. Shaded boxes represent positive test responses,

open boxes represent negative test responses (see text).

10

In Fig. 11 a simple alkylating agent is followed from the conduct of a confirmatory chemical alky- lation test [4-nitrobenzylpyridine (NBP)], through to its bioassay for rodent carcinogenicity - the simple assumption being that it will be positive in all assays conducted. The instance of 1,2-dichlo- roethane (EDC), a weakly reactive alkylating agent, demonstrates that a negative NBP test can be followed by a positive Salmonella assay response (Salm). This conflict of observations is caused by glutathione activating EDC to a Salmonella mutagen. Some agents, such as ben- zene and acrylamide, are genotoxic when as- sessed in a chromosomal aberration assay (CA in vitro), yet are non-mutagenic to Salmonella. This divergence of observations is due to either genetic or metabolic differences between mam- malian and bacterial cells. When genotoxicity is measured in rodent (GT in vivo) an effect may be seen (perhaps uniquely) at the site of application - the nasal tissue, the skin or the stomach. Diver- gences in activities may arise even between these 3 tissues due to the protective barr ier afforded to the skin by the stratum corneum. For agents active systemically as genotoxins (GT in vivo systemic) sensitive organs may vary with the route of exposure. In the examples represented in Fig. 11 a negative genotoxic response in the liver following oral dosing may become positive when the liver is bathed with the agent as would happen following its intraperitoneal injection. Finally, as with genotoxicity, carcinogenicity may only be observed for a reactive alkylating agent following its topical application - such agents may or may not be systemically carcinogenic. In the final column of Fig. 11 three ways in which a non- carcinogen may later be found to be carcinogenic (and vice versa) are listed - following its more extensive bioassay, or a change in the test species, or a change in the route of its application.

There are clearly many possible divergent datasets that could be derived selectively from Fig. 11, both real and theoretical - yet beneath them all is the potential to understand the mechanistic basis for the divergence. The danger alerted to here is that AI learning sets are usually culled from mixed datasets and this leads to an enhanced chance of false data entries if mutagenicity or carcinogenicity are reduced to the

binary status of positive or negative. A recent example is provided by Dichlorvos. In 1978 the NTP reported this agent to be a non-carcinogen following dietary administration (Anon., 1978). Following its repeat evaluation for carcinogenicity by oral gavage it was reported by the NTP to have clear evidence of carcinogenicity (Anon., 1989). Clearly, caution is required when selecting data for SAR studies or AI learning sets - wher- ever possible a single source of coherent data should be employed (e.g. same route, test species, assay duration, etc.).

During the 1992 Church of England Synod a speaker emphasized that words should be inter- preted within the context of their use - the word fire, for example, being initially an equal invita- tion to fall flat on the ground or to seek a pail of water. So it is with the word carcinogen - it is not a word of single meaning. The appellation 'male mouse non-genotoxic liver carcinogen' should provoke a different response to the appellation ' t rans-species/multiple-si te genotoxic carcinogen'. Different mechanisms of carcinogenesis, in particular relating to the major division between genotoxic and non-genotoxic carcinogenesis, require the derivation of different SAR learning sets, two examples of which will be used to end this discussion.

Alkylating agents (AA). An alkylating agent can alkylate, by definition. Vogel (elsewhere herein) has demonstrated the wide range of chemical factors that can influence electrophilicity and the dependent degree of which different AA's produce critical adducts on DNA. A separate range of biological factors then determine the mutational consequences, and a different range of biological factors then influence the extent to which certain of these mutagenic changes lead on to carcinogenesis. Given suffi- ciently detailed and appropriate information an AI model system could be primed to predict with accuracy the mutagenic/carcinogenic potential of a new alkylating agent. The AI system would derive firm conclusions of intrinsic reactivity based on iterative recognition of the key structural requirements for electrophilicity. The AI model system would then have to be separately (and repetitively) primed to predict the range of

different mutagenic /carc inogenic consequences of this predicted adduction of DNA. Specifically, Drosphilia will respond differently to a given adduction of D N A than will the cells of the Fischer 344 rat Zymbal 's gland. Given sufficient data input an AI model system should therefore make differential predictions - e.g. that an A A will be non-mutagenic to mei mutants of Drosophila, strongly mutagenic to Salmonella, non-carcinogenic to the mouse liver and strongly carcinogenic to the rat kidney. The critical point here is recognition of the fact that an appropriately pr imed AI model system should be capable of making firm prediction of both carcinogenicity and non-carcinogenicity for a given alkylating agent: a single prediction of 'carcinogen' or 'non-carcinogen' merely displays the relative naivety of the prediction system - be it human or artificial. An appropriately primed AI model system should be greatly superior to human deduction when predicting electrophilicity, mutagenesis and genotoxic carcinogenesis.

Peroxisome proliferators (PP). It is now al- most certain that there is no critical centre (akin to electrophilicity) within the structure of a PP that confers on it the property of carcinogenicity to the rodent liver. One is therefore not expecting to recognize critical substructures (biophores) and if such are discovered by an AI model system relying on a limited learning set of PP's one should respond with great caution. Underlying the carcinogenicity of PP's is the ability of the PP, an hydrolysis product or a metabolite, to affect lipid homeostasis in the rodent liver. A receptor has been recognized as associated with PP activity in the liver, but it is unlikely that the PP per se is the ligand for this receptor (Issemann and Green, 1990). It is true but of little predictive value that most PP's either already are, or are capable of being transformed into, an organic acid. Clearly, empirical SAR's local to a particular congeneric chemical series of PP will be capable of discernment, but the derivation of a coherent SAR capable of encompassing both diethyl- hexyl phthalate (DEHP) and trichloroacetic acid (TCA), for example, must await recognition of the critical biological stimulus to the train of biological events that lead to cancer within this

11

broad collection of chemicals (see also Ashby and Leigibel, 1992).

SAR's in mutagenesis and carcinogenesis can be no bet ter than our understanding of these phenomena themselves. As such SAR's will remain imperfect for the present. Nonetheless, if sought and derived objectively, SAR's can con- tribute significantly to the prediction of mutagenesis and carcinogenesis.

Acknowledgement

I am grateful to Paul Lohman, David Clayson and Henry Pitot for comments on this paper. Stuart Kettle provided the data for Fig. 2 and the title of Gold and Ames (1990) provided the idea for the title of this paper.

References

Anon. (1978) Bioassay of dichlorvos for rodent carcinogenicity, US NTP Tech. Rep. 10.

Anon. (1989) Bioassay of dichlorvos for rodent carcinogenicity, US NTP Tech. Rep. 342.

Anon. (1990) Evaluation of glycidol for carcinogenicity in rats and mice, NTP Tech. Rep. 374, National Toxicology Pro- gram, NIEHS, NC.

Ashby, J. (1985) Fundamental structural alerts to potential carcinogenicity and non-carcinogenicity, Environ. Muta- gen., 7, 919-921.

Ashby, J. (1992) Consideration of CASE predictions of genotoxic carcinogenesis for omeprazole, methapyrilene and azathioprine, Mutation Res., 272, 1-7.

Ashby, J., and U. Leigibel (1992) Transgenic mouse mutation assays: Potential for confusion of genotoxic and non- genotoxic carcinogenesis, A proposed solution, Environ. Mol. Mutagen., 20, 145-147.

Ashby, J., and D. Paton (1993) The influence of chemical structure on the extent and sites of carcinogenesis for 522 rodent carcinogens and 54 different chemical carcinogen exposures, Mutation Res., in press.

Ashby, J., and R.W. Tennant (1991) Definitive relationships among chemical structure, carcinogenicity and mutagenicity for 301 chemicals tested by the US NTP, Mutation Res., 257, 229-306.

Ashby, J., and U. Leigibel (1992) Transgenic mouse mutation assays: Potential for confusion of genotoxic and non- genotoxic carcinogenesis, A proposed solution, Environ. Mol. Mutagen., 20, 145-147.

Gold, L.A., and B.N. Ames (1990) Too many rodent carcinogens? Science, 249, 970-971.

Gold, L.S., T.H. Slone, N.B. Manley and L. Bernstein (1991) Target organs in chronic bioassays of 533 chemical carcinogens, Environ. Hlth. Perspect., 93, 233-246.

12

Issemann, I. and S. Green (1990) A member of the steroid hormone receptor superfamily activated by peroxisome proliferators, Nature (London), 347, 645-650.

Miller, J.A., and E.C. Miller (1977) Ultimate carcinogens as reactive mutagenic electrophiles, in: H.H. Hiatt, J.D. Wat- son and J.A. Winstein (Eds.), Origins of Human Cancer, Cold Spring Harbor Laboratory, Cold Spring Harbor, NY, pp. 605-628.

Rosenkranz, J.S., and G. Klopman (1992) 1,4-Dioxane: prediction of in vivo clastogenicity, Mutation Res., 280, 245- 251.

Tennant, R.W. and J. Ashby (1991) Classification according to chemical structure, mutagenicity to Salmonella and level of carcinogenicity of a further 39 chemicals tested by the US NTP, Mutation Res., 257, 209-227.

Tennant, R.W., J. Spalding, S. Stasiewicz and J. Ashby (1989) Prediction of rodent carcinogenicity for 44 chemicals currently being evaluated by the US NTP, Mutagenesis, 5, 3-14.

Vainio, H., M. Coleman and J. Wilbourn (1991) Carcinogenic- ity evaluation and ongoing studies: The IARC databases, Environ. Hlth. Perspect., 96, 5-9.

two million rodent carcinogens? the role of sar and qsar in their detection

Documents