Mutagenesis, Vol. 17, No. 4, 321-329,
July 2002
© 2002 UK Environmental Mutagen Society/Oxford University Press
Comparison of the computer programs DEREK and TOPKAT to predict bacterial mutagenicity
2 Safety Assessment, GlaxoSmithKline Inc., 5 Moore Drive, Research Triangle Park, NC 27709, USA, 3 Safety Assessment, GlaxoSmithKline Inc., 2FA14, Park Road, Ware, Hertfordshire, UK SG12 0DP and 4 Mechanism and Extrapolation Technologies, GlaxoSmithKline Inc., 3030 Cornwallis Road, Research Triangle Park, NC 27709, USA
 |
Abstract |
|---|
 Top Abstract IntroductionMaterials and methodsResultsDiscussionReferences |
|---|
The performance of two computer programs, DEREK and TOPKAT, was examined with regard to predicting the outcome of the Ames bacterial mutagenicity assay. The results of over 400 Ames tests conducted at Glaxo Wellcome (now GlaxoSmithKline) during the last 15 years on a wide variety of chemical classes were compared with the mutagenicity predictions of both computer programs. DEREK was considered concordant with the Ames assay if (i) the Ames assay was negative (not mutagenic) and no structural alerts for mutagenicity were identified or (ii) the Ames assay was positive (mutagenic) and at least one structural alert was identified. Conversely, the DEREK output was considered discordant if (i) the Ames assay was negative and any structural alert was identified or (ii) the Ames assay was positive and no structural alert was identified. The overall concordance of the DEREK program with the Ames results was 65% and the overall discordance was 35%, based on over 400 compounds. About 23% of the test molecules were outside the permissible limits of the optimum prediction space of TOPKAT. Another 4% of the compounds were either not processable or had indeterminate mutagenicity predictions; these molecules were excluded from the TOPKAT analysis. If the TOPKAT probability was (i)
0.7 the molecule was predicted to be mutagenic, (ii) 
0.3 the compound was predicted to be non-mutagenic and (iii) between 0.3 and 0.7 the prediction was considered indeterminate. From over 300 acceptable predictions, the overall TOPKAT concordance was 73% and the overall discordance was 27%. While the overall concordance of the TOPKAT program was higher than DEREK, TOPKAT fared more poorly than DEREK in the critical Ames-positive category, where 60% of the compounds were incorrectly predicted by TOPKAT as negative but were mutagenic in the Ames test. For DEREK, 54% of the Ames-positive molecules had no structural alerts and were predicted to be non-mutagenic. Alternative methods of analyzing the output of the programs to increase the accuracy with Ames-positive compounds are discussed. |
Introduction |
|---|
|
TopAbstractIntroductionMaterials and methodsResultsDiscussionReferences |
|---|
Several computer packages have been developed to study structure–activity relationships (SAR) in toxicology (for reviews see Lewis, 1992
; Wang and Milne, 1993; Richard, 1994, 1998). These programs are designed to predict toxicological outcomes such as skin sensitization, LD50 and aquatic toxicity; in particular, the predictions of end points of mutagenicity and carcinogenicity have received considerable attention (Ashby and Tennant, 1994; Lewis, 1994; Benigni, 1995, 1997; Zeiger et al., 1996). In general, the approaches followed by these toxicity prediction packages fall into two categories: (i) rule-based expert systems that rely on a set of chemical structure alerts and (ii) correlative structure–activity relationship methods based on statistical analysis. We present our experience with DEREK, a rule-based system, and TOPKAT, a correlative SAR system, for the prediction of bacterial mutagenicity.
DEREK (deductive estimate of risk from existing knowledge) uses a set of rules derived from the collective expertise of toxicologists from academia, industry and government. The premise of the system was clearly stated by its originators (Sanderson and Earnshaw, 1991):
if chemical structure feature is present then specific toxic action is a possibility
DEREK is a system which indicates whether a specific toxic response may occur; it does not provide a quantitative estimate of the prediction. DEREK has several rule bases, consisting of descriptions of molecular substructures (structural alerts) that have been associated with toxic end points (e.g. mutagenicity, carcinogenicity or skin irritation). Since substructures can exist in a variety of molecular contexts, the rules are not chemical-specific, but rather serve as broad generalizations with regard to the chemical structure (e.g. alkylating agent, acid or halogen-containing molecule). The development of the rules is a continuous process that is monitored by the DEREK Users Group.
TOPKAT (toxicity prediction by komputer assisted technology), on the other hand, quantifies electronic, bulk and shape attributes of a structure in terms of electrotopological state (E-states) values (Hall et al., 1991) of all possible two-atom fragments (Gombar, 1999), atomic size adjusted E-states computed from rescaled count of valence electrons, molecular weight, topological shape indices (Kier, 1986; Gombar and Jain, 1987) and symmetry indices. The methodology is an extension of classic quantitative structure–activity relationships (QSAR).
We report here a comparison of the DEREK and TOPKAT output for over 400 diverse compounds that have been tested by Glaxo Wellcome (now GlaxoSmithKline) for bacterial mutagenicity.
|
Materials and methods |
|---|
|
TopAbstractIntroductionMaterials and methodsResultsDiscussionReferences |
|---|
Bacterial mutagenicity assays
Compounds were tested in a standard Ames reversion assay (Maron and Ames, 1983
), the preincubation modification of the Ames assay (Yahagi et al., 1975) or a `mini-Ames' assay (Brooks, 1995; Burke et al., 1996) over the past 15 years. The most commonly used strains were Salmonella typhimurium TA98, TA100, TA1535 and TA1537 and Escherichia coli WP2uvrA(pKM101). Not all compounds were tested in all five strains. The compounds were in various stages of pharmaceutical development and ranged from chemical intermediates used in the manufacture of pharmaceuticals to marketed drugs. A compound was designated as mutagenic if (i) a dose–response relationship existed and (ii) one dose level produced at least a doubling over background in E.coli or S.typhimurium TA98 or TA100 or (iii) a tripling in TA1535 or TA1537.
DEREK and TOPKAT analysis
DEREK v.17.1 (Java client) and TOPKAT 5.01 for Windows were used. DEREK was run in batch and the rules that fired for each compound were captured in a text file output which was subsequently parsed into a Microsoft Access database for further analysis.
Individual structures were imported into the KATLOG of TOPKAT and processed one-by-one using the Ames Mutagenicity Prediction module v.3.1. The following parameters were captured for processing in the Microsoft Access database: (i) probability of mutagenicity; (ii) whether the molecule was outside the optimum prediction space (OPS); (iii) if a molecule was outside the OPS, whether the molecule was still within permissible limits.
When a query structure is determined to be inside all dimensions of a model's OPS, the computed value of toxicity can be considered acceptable. However, if a structure is found outside one or more dimensions, the computed toxicity may or may not be acceptable depending on the distance of the query from the OPS.
The distance of a query structure from the OPS is a complex function of the query location in each dimension. Every TOPKAT model has a suggested permissible limit of distance from the OPS, precalculated and stored with each model. If the query structure distance from the OPS is greater than the permissible limit, the toxicity value is considered unacceptable.
|
Results |
|---|
|
TopAbstractIntroductionMaterials and methodsResultsDiscussionReferences |
|---|
A total of 416 compounds were tested in the Ames assay and the structures were submitted to both DEREK and TOPKAT. Of these, 79.8% (332) were not mutagenic, 19.7% (82) were mutagenic and 0.5% (2) were considered equivocal. Equivocal compounds were excluded from the present analysis; thus 414 compounds (332 non-mutagens and 82 mutagens) were submitted to DEREK and TOPKAT for evaluation.
DEREK
DEREK does not provide a quantitative assessment of the probability of mutagenicity of a particular compound, but rather provides a series of structural alerts. To begin to assess the performance of DEREK, the DEREK output was considered concordant with the Ames assay if (i) the Ames assay was negative (not mutagenic) and no structural alerts for mutagenicity were identified or (ii) the Ames assay was positive (mutagenic) and at least one structural alert was identified. Conversely, the DEREK output was considered discordant if (i) the Ames assay was negative and any structural alert was identified or (ii) the Ames assay was positive and no structural alert was identified. Note that only mutagenic structural alerts are being considered.
DEREK could not process five of the 414 compounds since their molecular weights were too high. The overall performance of the program based on 409 compounds is given in Table I. Of the 409 compounds, 82 (20%) were mutagenic in the Ames assay and 327 (80%) were not mutagenic.
|
Using the ranking system described above, the overall concordance of the DEREK program with the Ames results was 65% (264 compounds) and the overall discordance was 35% (145 compounds).
For the discordant results, 44 of 82 (54%) of the Ames-positive compounds had no structural alerts and were therefore predicted by DEREK to be negative. Of the Ames-negative compounds, 101 of 327 (31%) triggered one or more structural alerts and were predicted to be positive.
The discordant results for compounds which were positive in the Ames assay but which had no alerts represented a wide variety of structures. However, there were some features of these compounds that were present in three or more molecules, and these features are shown in Figure 1. It may be useful to evaluate these features as possible DEREK alerts that could be incorporated into the software.
|
Certain structural alerts were prevalent in the discordant results for chemicals which were negative in the Ames assay but produced an alert. The alerts and the number of times they were triggered by the Ames-negative compounds are given in Table II
. In particular, the rule for aromatic amine mutagenicity (rule 827) was triggered 43 times in the Ames-negative data set. The top three rules that were triggered inappropriately were aromatic amine (rule 827, 43 instances), 
,ß-unsaturated amide (rule 821, 10 instances) and aromatic amine with diaryl fusion (rule 848, 10 instances).
|
For the concordant results, one or more rules were triggered for 38 compounds that were positive in the Ames assay. These results are given in Table III
.
|
Interestingly, the rule for aromatic amines was the most frequently triggered rule in both the discordant and concordant data sets. This illustrates one of the difficulties of simplistically applying a rule-based approach.
An extensive literature exists regarding the mutagenicity and carcinogenicity of aromatic amines. On a qualitative level, important features which can modulate mutagenicity and carcinogenicity include: (i) the number and nature of the aromatic rings; (ii) the position, nature and size of other ring substituents; (iii) the position and nature of the amine groups; (iv) the polarity, size and shape of the molecule (Lai et al., 1996).
On a quantitative level, numerous QSARs have been developed for the mutagenicity and carcinogenicity of aromatic amines (for a review see Benigini et al., 2000). QSAR equations have been developed for Salmonella mutagenicity of aromatic amines which make use of logP, semi-empirical AM1 molecular orbital calculations, HUMO, LUMO and a wide variety of geometrical, topological, electrostatic and quantum chemical descriptors.
DEREK essentially informs the user that an aromatic amine exists; further interpretation and analysis of the likelihood of mutagenicity is up to the user. DEREK provides some textual explanation of the logic used by the rule, however, compound-by-compound evaluation can become impractical when processing hundreds or thousands of compounds.
TOPKAT
TOPKAT could not process 1% (five) of the 414 input structures for one of the following reasons: (i) a charge on a molecule other than anionic oxygen or cationic nitrogen; (ii) compounds with manganese; (iii) mixtures.
About 23% of the remaining compounds were outside the OPS TOPKAT considers acceptable for model application. TOPKAT determines whether a structure is within the OPS of a quantitative structure–toxicity relationship model based on values of descriptors in the model. Of the 409 compounds, 96 exceeded the permissible limit of the OPS and TOPKAT models are not applicable to these structures. These molecules were excluded from the analysis.
TOPKAT computes a probability of mutagenesis ranging from 0 to 1 based on a linear discriminant function corresponding to the submodel to which the input structure belongs. It is generally recommended that the compounds with predicted probabilities (i) 0.7 be labeled mutagenic, (ii) 0.3 be considered non-mutagenic and (iii) between 0.31 and 0.69 be considered inconclusive. Ten compounds produced an inconclusive probability and these were not considered in the analysis.
Thus, TOPKAT excluded a high fraction of the molecules, nearly 26% that could be processed but were either outside the permissible limits of the OPS or had an inconclusive probability. A total of 303 compounds remained in the analysis, of which 250 (82.5%) were not mutagenic in the Ames assay and 53 (17.5%) were mutagenic.
The performance of TOPKAT on these 303 molecules is given in Table IV. The overall concordance of the TOPKAT program with the Ames results was 73% and the overall discordance was 27%.
|
For the discordant results, 32 of 53 (60%) Ames-positive compounds had a TOPKAT probability
0.3 and these compounds were therefore predicted to be negative. Of the Ames-negative compounds 49 of 250 (20%) had a TOPKAT probability 0.7 and are predicted to be positive. |
Discussion |
|---|
|
TopAbstractIntroductionMaterials and methodsResultsDiscussionReferences |
|---|
The ability of two software programs to predict the outcome of the Ames bacterial mutation assay was examined by comparing the results of the Ames test with the output of the computer programs. The overall concordance of the DEREK program with the Ames results was 65% and the overall discordance was 35%, based on ~400 compounds. The overall concordance of TOPKAT with the Ames results was 73% and the overall discordance was 27%, based on ~300 compounds.
While the overall concordance of TOPKAT was higher than that of DEREK, TOPKAT placed a high fraction of the molecules into the `uncertain' category. About 26% (106 of 409) of the compounds were either outside the permissible limits of the OPS or produced probability estimates in the indeterminate region (0.31–0.69); the prediction of mutagenicity for these compounds is considered unreliable. DEREK, an expert system, does not have a defined chemical space and consequently a much higher fraction of compounds could be processed.
It should be noted that ~400 compounds were processed by DEREK to estimate the Ames concordance and discordance, while only ~300 compounds were considered for TOPKAT. While TOPKAT offers a diagnostic mechanism to signal extrapolation of a model, the fact that it showed a higher concordance with the Ames data could be due to the restricted compound set analyzed by TOPKAT. We therefore analyzed the DEREK predictions for the 303 compounds which were accepted by TOPKAT. The overall concordance of the DEREK program improved slightly, from 64.5% with the large 409 compound set to 66.0% with the smaller 303 compound set. The 303 compound set produced the following: (i) an overall discordance of 34%, with 9.7% Ames-positive with no rules firing and 24.3% Ames-negative and any rule firing; (ii) an overall concordance of 66.0%, with 8.0% Ames-positive and any rule firing and 58.0% Ames-negative and no rule firing.
TOPKAT v.5.01 is unable to process compounds in batch mode, which makes evaluations of very large data sets unfeasible. Compound names and structures are stored in a TOPKAT KATLOG which is neither searchable nor ordered alphabetically; locating a given molecule in a KATLOG of several hundred compounds can be difficult.
Neither DEREK nor TOPKAT performed particularly well in the critical Ames-positive category. While the overall concordance of the TOPKAT program was higher than DEREK, TOPKAT fared more poorly than DEREK in the Ames-positive category. Among the mutagens, 60% of the compounds were incorrectly predicted to be non-mutagenic by TOPKAT. In comparison, 55% of the Ames-positive compounds did not produce a structural alert and were predicted to be negative by DEREK.
We wished to determine the performance of both programs when used in conjunction.
Eighty-two compounds were positive in the Ames assay. All Ames-positive compounds could be processed using DEREK, while a valid probability was produced for 53 of the Ames-positive compounds using TOPKAT.
The fact that TOPKAT could not produce a valid probability for 35% of the mutagenic compounds indicates that the training set of TOPKAT and the test set of mutagenic pharmaceuticals in this report do not have a great deal of overlap. Future releases of TOPKAT may benefit from expanding the training set with `mutagenic pharmaceuticals'.
DEREK carries out no analysis of its valid `chemical space' and is consequently able to process all of the mutagenic molecules.
Only 13 compounds were predicted to be mutagenic by both programs. Thus, it appears that DEREK and TOPKAT are recognizing different portions of the Ames-positive molecules during processing. Twenty-one Ames-positive molecules were predicted to be negative by both programs. Therefore, even if one accepts the output of either program as a criterion for mutagenicity prediction, about one-quarter (26%) of the Ames-positive molecules will be missed; clearly, the false positive rate will be increased if the programs are used in this fashion.
Of the non-mutagenic compounds, 327 could be processed by DEREK, while TOPKAT could process 250 molecules. Of the Ames-negative compounds, 226 of 327 were correctly identified by DEREK as not mutagenic, while TOPKAT identified 201 of 250 Ames-negative compounds correctly; 146 Ames-negative compounds were predicted by both programs to be non-mutagenic and 21 compounds were predicted to be mutagenic by both.
We wished to determine whether the concordance of the programs could be increased in the Ames-positive category using DEREK. Only the DEREK structural alerts for mutagenicity were considered for the preceding analysis and discussion. We wished to determine the effect of evaluating the entire data set using both the DEREK mutagenicity and carcinogenicity rule bases.
Thus, a concordant result for the re-analysis is a positive Ames test and firing of any DEREK mutagenicity or carcinogenicity rule. Fifty-four of 82 (66%) of the Ames-positive compounds would be properly predicted using the expanded rules, compared to 44 (54%) using only the mutagenic rule base. However, as expected, this comes at the expense of false positives in which a rule fired but the compound was not mutagenic; 176 compounds were in this category with the expanded rules, while 101 were found using only the mutagenic rule base. Using the expanded rule base the overall concordance of DEREK drops to 50%.
To determine whether the performance of TOPKAT can be improved in the Ames-positive category, the data set was re-evaluated using a different TOPKAT threshold for indeterminate predictions. TOPKAT produces a probability between 0 and 1 for a given compound and the data set was initially evaluated using the recommended values of: (i) 0.7 as the threshold for a mutagen; (ii) 0.3 for a non-mutagen; (iii) between 0.3 and 0.7 as indeterminate.
We widened the indeterminate category to see if the concordance of the program would be improved. This has the effect of placing more compounds into the indeterminate category, where they will be excluded from the analysis; this process can be thought of as increasing the stringency for accepting a negative or positive prediction. The results of this analysis are given in Table V.
|
Naturally, as the indeterminate boundaries grow larger, a greater number of compounds fall into this category, so the ability of TOPKAT to process compounds is reduced. A rather modest increase in the number of indeterminate compounds occurs up to and including the range 0.10–0.90 for indeterminates; past this level the number of indeterminate compounds increases sharply.
As expected, the overall discordance is reduced and the concordance is increased as the indeterminate category grows larger. The reduction in discordant predictions occurs for both the Ames-positive and Ames-negative compounds as the indeterminate category is widened. This is not the case, however, for the concordant predictions. While the overall concordance is increased as the indeterminate category is widened, this comes exclusively by more accurately predicting the Ames-negative compounds; the program actually fares more poorly in predicting the Ames-positive compounds as the indeterminate bounds are widened.
In certain situations it may be acceptable to decrease the overall concordance of the program to increase the accuracy in the Ames-positive class. In the pharmaceutical industry there is a trend towards computer-based assessment of compounds before they have been physically synthesized. Parameters such as solubility, absorption and metabolic stability can be predicted. In such a scenario, before the molecules have been chemically synthesized, it may be acceptable to alter the DEREK and TOPKAT parameters to increase the likelihood that a compound when actually synthesized is not mutagenic. Of course, this means increasing the fraction of molecules that will be discarded, even though they will not actually be mutagenic.
As an aside, we also examined the effect of ignoring the permissible limits that are computed by TOPKAT for each molecule. About 23% of the 409 molecules analyzed by TOPKAT were outside the permissible limits and were excluded from the analysis. Surprisingly, including all molecules had very little effect on the performance of TOPKAT, as shown in Table VI.
|
These data suggest that the OPS function in TOPKAT does not add significant value in the Ames module for predicting the Glaxo Wellcome dataset, but rather unnecessarily restricts application of the model. Therefore, in our case it appears appropriate to include all molecules in the analysis. Our pharmaceutical dataset presents a challenge to the OPS function in the mutagenicity module. It should be noted that the OPS may add value to the mutagenicity module when used with other datasets.
One of the difficulties in writing this manuscript was that many of the chemical structures are proprietary and cannot be revealed. Confidentiality must be balanced against the need of the scientific community to fairly evaluate the methods and conclusions put forth in this paper, which can be best done by examination of structures.
To better allow the reader to evaluate the methodology used in this paper, 52 molecules which triggered a DEREK alert for an aromatic amine were subject to intellectual property review; nine of these compounds were mutagenic in the Ames assay and 43 were not. Forty-two compounds are in the public domain, and these structures are given in Figure 2; the remaining 10 compounds cannot be revealed. Seven compounds shown in Figure 2 were mutagenic in the Ames assay and 35 were not.
|
All 42 structures in Figure 2
triggered a DEREK alert for an aromatic amine, however, only seven of these compounds were mutagenic. It has been recognized for some time among DEREK users that DEREK `overpredicts' the mutagenicity of aromatic amines. Additional research to refine the DEREK rules for the mutagenicity of aromatic amines would be very beneficial.
TOPKAT did not fare well in predicting the mutagenic aromatic amines in Figure 2. The TOPKAT values for probability of mutagenicity, several OPS parameters, similarity distances, use of similar molecules to predict mutagenicity and the predicted and actual mutagenicity of the similar molecules are given in Table VII.
|
Of the seven mutagenic aromatic amines, two were outside the OPS and the permissible limits and are therefore not used. However, for the remaining five mutagenic aromatic amines the predicted probability of mutagenicity by TOPKAT was 0, i.e. not mutagenic. The similarity distances of the nearest molecule for four of the five mutagenic compounds was <0.21, indicating that these four molecules have reasonably `similar' molecules in the OPS TOPKAT submodel for aromatic amines. All of the nearest molecules in the training set were non-mutagenic and TOPKAT correctly predicted these molecules in the training set to be non-mutagenic; this increases confidence in the prediction of the unknowns. Taken together, the predicted values of zero for at least four of the five mutagenic aromatic amines appear reasonably sound, however, all the predictions are wrong.
Nine of the 35 non-mutagenic aromatic amines in Figure 2 were not within TOPKAT permissible limits and are not considered further. Seven of the 26 non-mutagenic aromatic amines within the permissible limits were predicted to be mutagenic, as these molecules yielded a probability of mutagenicity of >0.7; one of the seven predicted mutagens had molecules that showed a high similarity distance and the TOPKAT probability could be taken as somewhat unreliable. Therefore, the percentage of aromatic amines incorrectly predicted to be mutagenic is between 23 (6/26) and 27% (7/26).
In summary, both DEREK and TOPKAT fared more poorly than we had hoped. It appears that laboratory testing for Ames mutagenicity will not be supplanted by computer methods in the imminent future. The programs can be adapted for specific uses by reducing the false positive rate or the false negative rate, however, it is not possible to lower both simultaneously. A new version of TOPKAT has been released, but the Ames module is unchanged. However, a new version of DEREK is now available, v.5.0.1, in which the mutagenicity rule base has been updated. In a subsequent publication we will examine the performance of the new rule base.
|
Notes |
|---|
1 To whom correspondence should be addressed. Tel: +1 919 483 6782; Fax: +1 919 483 6858; Email: nfc20355{at}gsk.com
|
References |
|---|
|
TopAbstractIntroductionMaterials and methodsResultsDiscussionReferences |
|---|
-
Ashby,J. and Tennant,R.W. (1994) Prediction of rodent carcinogenicity for 44 chemicals: results. Mutagenesis, 9, 7–15.
Benigni,R. (1995) Predicting chemical carcinogenesis in rodents: the state of the art in light of a comparative exercise. Mutat. Res., 334, 103–113.[Web of Science][Medline]
Benigni,R. (1997) The first US National Toxicology Program exercise on the prediction of rodent carcinogenicity: definitive results. Mutat. Res., 387, 35–45.[Web of Science][Medline]
Benigini,R., Giuliani,A., Franke,R. and Gruska,A. (2000) Quantitative structure–activity relationships of mutagenic and carcinogenic aromatic amines. Chem. Rev., 100, 3697–3714.[Web of Science][Medline]
Brooks,T.M. (1995) The use of a streamlined bacterial mutagenicity assay, the MINISCREEN. Mutagenesis, 10, 447–448.
Burke,D.A., Wedd,D.J. and Burlinson,B. (1996) Use of the Miniscreen assay to screen novel compounds for bacterial mutagenicity in the pharmaceutical industry. Mutagenesis, 11, 201–205.
Gombar,V.K. (1999) Reliable assessment of logP of compounds of pharmaceutical relevance. SAR QSAR Environ. Res., 10, 371–380.[Web of Science][Medline]
Gombar,V.K. and Jain,D.V.S. (1987) Quantification of molecular shape and its correlation with physicochemical properties. Indian J. Chem., 26A, 554–555.
Hall,L.H., Mohney,B. and Kier,L.B. (1991) The electrotopological state: structure information at the atomic level. J. Chem. Inf. Comput. Sci., 31, 76–82.
Kier,L.B. (1986) Shape indices of orders one and three from molecular graphs. Quant. Struct. Activ. Relat., 5, 1–7.
Lai,D.Y., Woo,Y.T., Argus,M.F. and Arcos,J.C. (1996) Cancer risk reduction through mechanims-based molecular designe of chemicals. In Vito,S.C. and Garrett,R.L. (eds) Designing Safer Chemicals. Green Chemistry for Pollution Prevention. American Chemical Society, Washington, DC, pp. 1–13.
Lewis,D.F.V. (1992) Computer-assisted methods in the evaluation of chemical toxicity. In Lipowitz,K.B. and Boyd,D.B. (eds) Reviews in Computational Chemistry. VCH Publishers, New York, NY, pp. 173–222.
Lewis,D.F. (1994) Comparison between rodent carcinogenicity test results of 44 chemicals and a number of predictive systems. Regul. Toxicol. Pharmacol., 20, 215–222.[Web of Science][Medline]
Maron,D.M. and Ames,B.N. (1983) Revised methods for the Salmonella mutagenicity test. Mutat. Res., 113, 173–215.[Web of Science][Medline]
Richard,A.M. (1994) International Commission for Protection Against Environmental Mutagens and Carcinogens. Application of SAR methods to non-congeneric data bases associated with carcinogenicity and mutagenicity: issues and approaches. Mutat. Res., 305, 73–97.[Web of Science][Medline]
Richard,A.M. (1998) Structure-based methods for predicting mutagenicity and carcinogenicity: are we there yet? Mutat. Res., 400, 493–507.[Web of Science][Medline]
Sanderson,D.M. and Earnshaw,C.G. (1991) Computer prediction of possible toxic action from chemical structure; the DEREK system. Hum. Exp. Toxicol., 10, 261–273.[Web of Science][Medline]
Wang,S. and Milne,G.W. (1993) Applications of computers to toxicological research. Chem. Res. Toxicol., 6, 748–753.[Web of Science][Medline]
Yahagi,T., Degawa,M., Seino,Y., Matsushima,T. and Nagao,M. (1975) Mutagenicity of carcinogenic azo dyes and their derivatives. Cancer Lett., 1, 91–96.[Web of Science][Medline]
Zeiger,E., Ashby,J., Bakale,G., Enslein,K., Klopman,G. and Rosenkranz,H.S. (1996) Prediction of Salmonella mutagenicity. Mutagenesis, 11, 471–484.
Received on May 9, 2001; revised on February 2, 2002; accepted on February 26, 2002.
CiteULike 
Connotea 
Del.icio.us What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||