Abstract

Structure–activity relationship (SAR) models are powerful tools to investigate the mechanisms of action of chemical carcinogens and to predict the potential carcinogenicity of untested compounds. We describe the use of a traditional fragment-based SAR approach along with a new virtual ligand-protein interaction-based approach for modeling of nonmutagenic carcinogens. The ligand-based SAR models used descriptors derived from computationally calculated ligand-binding affinities for learning set agents to 5495 proteins. Two learning sets were developed. One set was from the Carcinogenic Potency Database, where chemicals tested for rat carcinogenesis along with Salmonella mutagenicity data were provided. The second was from Malacarne et al . who developed a learning set of nonalerting compounds based on rodent cancer bioassay data and Ashby’s structural alerts. When the rat cancer models were categorized based on mutagenicity, the traditional fragment model outperformed the ligand-based model. However, when the learning sets were composed solely of nonmutagenic or nonalerting carcinogens and noncarcinogens, the fragment model demonstrated a concordance of near 50%, whereas the ligand-based models demonstrated a concordance of 71% for nonmutagenic carcinogens and 74% for nonalerting carcinogens. Overall, these findings suggest that expert system analysis of virtual chemical protein interactions may be useful for developing predictive SAR models for nonmutagenic carcinogens. Moreover, a more practical approach for developing SAR models for carcinogenesis may include fragment-based models for chemicals testing positive for mutagenicity and ligand-based models for chemicals devoid of DNA reactivity.

Introduction

Rodent cancer bioassay has proven to be a useful tool in our understanding of the causes and treatment of cancer. There are sufficient physiological, biochemical and metabolic similarities between rodents and humans that allow a high probability that carcinogenicity results obtained in rodents will be predictive of similar effects in humans ( 1–3 ). However, a complete 2-year cancer bioassay as conducted by the National Toxicology Program (NTP) including planning, evaluation and review takes ~5 years to complete, costs between $2 and $4 million and uses 400 animals ( 4 ). There are currently 580 technical reports by the NTP for rodent carcinogenicity using this standardized bioassay ( 5 ). Extending the range of the NTP data, the Carcinogenic Potency Database (CPDB) analyzes and consolidates the world’s diverse literature, including NTP reports, about chronic long-term animal cancer bioassays into a single resource ( 6 ). To date, analyses of 6540 experiments on 1547 chemicals are available on the CPDB’s Web site ( 7 ) and the Distributed Structure-Searchable Toxicity Database Network ( 8 ).

However, there are approximately 75 000 industrial chemicals on the Toxic Substance Control Act’s Chemical Substance Inventory ( 9 ), and the National Institute of Environmental Health Sciences estimates that there are over 84 000 chemicals registered for use in the USA ( 10 ). From a public health perspective, it is evident that not all chemicals in use today will be tested in vivo for carcinogenesis, thus necessitating alternative methods for carcinogen detection. The National Research Council’s 2007 Report on the future of toxicity testing discussed the next generation of toxicity testing, noting that structure–activity relationship (SAR) technology will play an important role ( 11 ). With respect to the work described here, the National Research Council noted that chemical exposures led to adverse outcomes through perturbation of pathways associated with toxicity ( 12 ), with receptor–enzyme interactions being a key interaction ( 13 ). Kavlock et al. (13) noted that ‘few such toxicity pathways are currently understood’.

Whether a nonmutagenic chemical will be carcinogenic is difficult to determine by means other than rodent cancer bioassays. Briefly, the Ames Salmonella mutagenicity assay and other short-term tests for mutagenicity are used to detect carcinogens. Mutagenic carcinogens are typically those that directly damage DNA. Zeiger outlines the current battery of tests for assessing carcinogenic potential such as the Salmonella test with or without Escherichia coli , the in vitro mammalian cell mouse lymphoma assay and/or the in vitro mammalian cell chromosome assay and the in vivo rodent bone marrow chromosome aberration or micronucleus assay ( 14 ). However, a significant number of cancer-causing chemicals are nonmutagenic agents and do not directly interact with DNA but rather are suspected of inducing cancer by a number of alternative mechanisms. Combes included cytotoxic activity along with the inhibition of gap-junctional intercellular communication, inhibition of tubulin polymerization, modulation of apoptosis and induction of cell proliferation and binding to specific molecular targets (receptors) including protein kinase C, the estrogen, peroxisome proliferator and tubulin protein receptors as some of the nonmutagenic mechanisms of carcinogenesis ( 15 ).

An analysis of the ISSCAN v3a (Institute Superiore de Sanità rodent carcinogenicity database) revealed that 311 out of the 387 Salmonella mutagens were carcinogenic (80%) (16) ( Table I ), which is consistent with an earlier analysis of NTP data (see ref. 17 ). As such, the value of the Salmonella assay is significant in identifying potential carcinogens. However, a negative response in Salmonella assay has no predictive value as to rodent carcinogenicity since 209 of 436 nonmutagens are carcinogens (48%) ( Table I ). Zieger concluded that even in established batteries of tests, due to their noncomplementarity, additional tests besides Salmonella do not help in detecting Salmonella -negative carcinogens from Salmonella -negative noncarcinogens ( 14 ). Benigni et al. (16) noted that the Salmonella assay is the most predictive assay for DNA-reactive carcinogens and that no other mutagenicity-based test exists for non-DNA-reactive compounds.

Table I.

ISSCAN v3a comparison between rodent carcinogenicity and Ames Salmonella mutagenicity results (from ref. 16 )

Carcinogenicity Ames test 
Negative Positive Total 
Negative 227 76 303 
Positive 209 311 520 
Total 436 387 823 
Carcinogenicity Ames test 
Negative Positive Total 
Negative 227 76 303 
Positive 209 311 520 
Total 436 387 823 

Based on the electrophilic theory of carcinogenesis put forth by Miller et al. ( 18 ) and subsequent work by Ashby et al. ( 19–22 ), chemical moieties associated with carcinogenesis called ‘structural alerts’ that are mechanistically interpretable and predictive for chemicals that are mutagenic carcinogens were developed. Woo, however, has pointed out that the identification of structural attributes of non-DNA-reactive carcinogens is not as advanced as that for mutagenic carcinogens ( 23 ).

Computational SAR analyses of large diverse sets of chemicals tested in vivo for carcinogenesis have also been successful in understanding mechanisms of carcinogenesis and have provided a means of predicting the carcinogenic potential of chemicals. TOPKAT ( 24 ), DEREK ( 25 ) and MCASE are examples of widely used SAR expert systems for analysis of chemicals for potential carcinogenesis. SAR models are increasingly being used by regulatory agencies worldwide for both human health ( 26 ) and ecological endpoints ( 27 ) including OncoLogic by the U.S. Environmental Protection Agency ( 28 ) and MCASE by the U.S. Food and Drug Administration’s Center for Drug Evaluation and Research ( 29 ).

The Food and Drug Administration is also actively using expert systems to explore the application of SAR models in drug evaluations ( 29–31 ). Analysis of marketed pharmaceuticals, with both mutagenicity and carcinogenicity data, revealed that 116 out of 315 drugs were neither mutagens nor carcinogens, 50 were mutagenic noncarcinogens, 75 were nonmutagenic carcinogens and 74 were mutagenic carcinogens ( 32 ). Also, sensitivity of MCASE, DEREK and TOPKAT in identifying mutagenic drugs was found to range between 43.4 and 51.9% ( 33 ).

Approaches are being developed to assess the carcinogenic potential of nonmutagenic agents. For example, using a set of three nonmutagenic carcinogens and two nonmutagenic noncarcinogens, analysis of gene expression profiles from acutely treated animals was able to discriminate the carcinogens from the noncarcinogens ( 34 ). Malacarne et al. (35) used cancer bioassay data and Ashby’s structural alerts with some additional expertise to develop a learning set of 390 nonalerting compounds (145 carcinogens and 245 noncarcinogens), and this data set was modeled by us (see below). SAR analysis of this data set produced a model that when validated returned a concordance of 65.4% between experimental and predicted values, with a sensitivity of 47% and specificity of 75.6% ( 35 ). However, with only a 47% sensitivity, this approach appears not to be successful at predicting nonmutagenic carcinogens. Rosenkranz et al. (36) using an inductive rule learning program predicted the carcinogenicity of 135 nonmutagens with a concordance, sensitivity and specificity of 70%. In this study, however, results from other short-term genotoxicity assays were used for rule development ( 36 ). It has also recently been demonstrated that incorporating high-throughput screening data into QSAR analysis of rodent carcinogens can improve the overall predictivity of the model ( 37 ).

Recently, we reported a new technique for developing SAR models using computationally derived ligand-binding affinity for over 5500 protein receptors as SAR descriptors. In these studies of rat mammary carcinogens, we observed that a hybrid model based on chemical fragments and ligand-binding potential had superior predictivity over either the traditional fragment-based model or the new ligand-based model ( 38 ). Importantly, considering the application of the ligand-based model to nonmutagenic carcinogenicity, the rat mammary carcinogen–nonmammary carcinogen models had learning sets populated completely with carcinogens (i.e. active compounds were mammary carcinogens and inactive ones were carcinogens to all other sites but the mammary gland); hence, we speculated that DNA reactivity per se was removed from the modeling process since both the active and inactive categories were populated with carcinogens.

Herein we describe the development of SAR models for nonmutagenic compounds using the categorical SAR (cat-SAR) ligand-based approach. We also developed models with the same sets of mutagenic and nonmutagenic compounds with the cat-SAR fragment method for comparison. In all instances, the fragment method was superior to the ligand method for mutagenic carcinogens. However, for analyses that considered only nonmutagenic agents, the new ligand-based approach was superior to the fragment method.

Materials and methods

Learning sets

The CPDB standardizes the experimental results (whether positive or negative for carcinogenicity), including qualitative data on strain, sex, route of compound administration, target organ, histopathology and the author’s opinion and reference to the published paper, as well as quantitative data on carcinogenic potency, statistical significance, tumor incidence, dose–response curve shape, length of experiment, duration of dosing and dose rate ( 39 ). Moreover, a potency value for carcinogens, the TD 50 , is also available. The TD 50 is defined as ‘that dose rate (in mg/kg body weight/day) which, if administered chronically for the standard lifespan of the species, will halve the probability of remaining tumorless throughout that period (39)’.

For this study, we used the rat data from the CPDB where Salmonella mutagenicity data were also provided. This data set consisted of 612 compounds, of which 183 were nonmutagenic noncarcinogens, 134 were nonmutagenic carcinogens, 76 were mutagenic noncarcinogens and 219 were mutagenic carcinogens ( Table II ).

Table II.

Summary of carcinogenic and mutagenic status of compounds use to produce the CPDB rat cancer models

Carcinogenicity Ames test 
Negative Positive Total 
Negative 183 76 259 
Positive 134 219 353 
Total 317 295 612 
Carcinogenicity Ames test 
Negative Positive Total 
Negative 183 76 259 
Positive 134 219 353 
Total 317 295 612 

Additionally, we developed a nonmutagenic carcinogenicity learning set that was originally published by Malacarne et al. ( 35 ). Malacarne et al. (35) used cancer bioassay data and Ashby’s structural alerts with some additional expertise to develop a learning set of 390 nonalerting compounds (145 carcinogens and 245 noncarcinogens). We note that our general approach to rat and mouse cancer models is that while a carcinogen need only test positive in one gender, a negative must be inactive in both; hence, we excluded 41 compounds considered noncarcinogens by Malacarne et al . since they were not tested in both genders. Our final data set consisted of 349 compounds (145 carcinogens and 204 noncarcinogens).

SAR model descriptors: chemical fragments and ligand-binding affinity

Fragment model. The cat-SAR program was developed to generate SAR models based on 2D chemical fragments. For these models, each chemical in the learning set was fragmented into all possible fragments, in this case between three and seven atoms in size. Atom type, bond type and atomic connections were considered. A compound–fragment matrix was then computed, where the rows were intact chemicals and the columns were molecular fragments. Thus, for each chemical, a tabulation of all its fragments was recorded across the table row, and for each fragment, all chemicals that contain it were tabulated down the table column.

Ligand model. For the ligand model, the cat-SAR algorithm was adapted to use compound–protein interaction data as SAR model descriptors. The 612 (CPDB) and 349 (Malacarne) compounds were virtually screened for ligand-like character against a set of virtual screening targets developed by Kellenberger et al. ( 40 ). For this set of ligand–protein crystal structures, Kellenberger et al. extracted binding sites from PDB structures wherein a small molecule and a protein cavity were identified. Solvents, detergents, ions and the ligand were removed, leaving the binding pocket open for virtual screening ( 40 ). For this exercise, we used ‘scPDB 2007’ downloaded on 16 June 2007 ( 41 ), which contained 5655 structures, of which 5494 were amicable to our analyses.

Each of the chemicals was virtually docked into each of the ligand-binding cavities of the sc-PDB using Surflex-Dock 2.3 ( 42 , 43) . The standard scoring function for each compound was calculated to estimate its affinity to the binding site as −log( Kd ) using hydrophobic, polar complementarity, entropic and solvation terms. For each compound, the sc-PDB structures were sorted according to their affinity scores, with high scores suggesting that the compound was a ligand for that particular protein. Range-finding experiments were conducted to determine an appropriate number of sc-PDB structures to be used as cat-SAR descriptors, and the top 250 sc-PDB structures were selected. A compound–ligand matrix was then computed, where the rows were the chemicals, and in this instance, each column represented one of the 5494 proteins. Thus, for each chemical, a tabulation of the top 250 proteins it interacted with was recorded across the table row, and for each protein, all chemicals that interacted with it were tabulated down the table column.

Cat-SAR modeling. To ascertain an association between chemical descriptors (i.e. fragments or ligands) and a chemical’s activity (or inactivity), a set of rules was used to choose ‘important’ from ‘unimportant’ descriptors. The first selection rule (the number rule) was the number of chemicals identified in the learning set that possesses each particular description. For the fragment model, it was the number of chemicals in the learning set that contains the fragment. For the ligand model, it was the number of chemicals determined to be a ligand for each protein. The second selection rule (the proportion rule) was the proportion of active or inactive chemicals that then possesses the particular description. For the fragment model, it was the proportion of active or inactive chemicals that derived the fragment. For the ligand model, it was the proportion of active or inactive chemicals that were classified as ligands for the protein. Since it was not practical to determine the values for the number and proportion rules a priori these values were estimated by the cat-SAR rule optimization routine. The optimization routine in this instance allowed the number rule to range between 1 and 8 and the proportion rule to range between 0.50 and 0.95. Based on leave-one-out (LOO) validations (see below), final values were selected, which yielded both highly accurate (i.e. had a high concordance between experimental and predicted values) and highly predictive models (i.e. made predictions on most of the chemicals in the learning set). Furthermore, another factor considered for selecting the most predictive and inclusive model was adjusting the model’s coverage. The coverage defines the total number of compounds a model predicts. Two sets of models, one in which 80% of the compounds were predicted and another in which 90% of the compounds were predicted, were created. Studying the coverage can verify whether including more chemical space or increasing the concordance by decreasing the coverage will increase the model’s validation.

Model validation and application. The resulting list of important fragments or proteins was used to validate or test the predictivity of the model, for mechanistic analysis, and to predict the activity of an unknown compound. To predict the activity of an untested compound, the cat-SAR program determined which, if any, descriptors from the model’s pool of significant descriptors the untested compound contained. If none were present, no prediction of activity was made for the compound (i.e. there were no default predictions of activity or inactivity). If one or more descriptors were present, the number of active and inactive compounds associated with each descriptor was determined. The probability of activity or inactivity was calculated based on the total number of active and inactive compounds that went into deriving each of the descriptors.

The probability of activity was calculated by cat-SAR by two similar techniques. The summation method added all the active and inactive compounds that went into deriving each fragment and divided the active compounds by the total to determine the probability of activity. For example, if a compound contained two fragments, one being found in 9 of 10 active compounds in the learning set (i.e. 90% active) and the other being found in 3 of 3 inactive compounds (i.e. 0% active), the unknown compound will be predicted to have a probability of activity of 69% (i.e. 9/10 actives + 0/3 actives = 9/13 actives or 69% chance of activity). The average method calculated the average probability of the active and inactive fragments contained in the chemical by averaging the probability of activity associated with each fragment. Using the above example, the two probabilities of activity, 90% and 0%, were averaged, with an activity value of 45%.

A self-fit and two cross-validation analyses were conducted for each model. For the self-fit analysis, after a model was developed, the model was used to predict the activity of the chemicals in its learning set in order to ascertain whether the model was capable of at least fitting its own data. A LOO validation was conducted, wherein each chemical, one at a time, was removed from the model’s learning set and an n − 1 model was derived. Using the same criteria described above, the activity of the removed chemical was predicted using the n − 1 model.

In order to consider concordance, sensitivity and specificity in terms of carcinogenic activity (i.e. whether a carcinogen or not a carcinogen), each compound’s probabilistic activity value was converted back to an active or inactive category value using a cutoff point derived from the LOO validations ( 44 ). Depending on the application of the model, the cutoff point was adjusted wherein a model with the best overall concordance was selected (i.e. a most predictive model), one with near-equal sensitivity and specificity (i.e. a balanced model) or one with high sensitivity (i.e. a risk-averse model). For this study, only results for balanced models are shown. To demonstrate how models with a high number of true-positive and low number of false-negative predictions are selected, relative operating characteristic graphs were derived ( Supplementary Figures 1–6 are available at Carcinogenesis Online).

Results and discussion

For comparison, two previously developed cat-SAR models were considered. The first was a fragment-based model for Salmonella mutagenesis derived from data from NTP ( 45 ). These data were a compilation of test results from different TA strains (with and without S9 activation). The cat-SAR model contained 1678 chemicals, 573 of which were mutagens and 1105 were nonmutagens. The model’s sensitivity, specificity and concordance was 0.79. The second was the rat carcinogenesis model based on the CPDB and consisted of 924 chemicals, 531 of which were categorized as rat carcinogens and 393 as noncarcinogens. The rat carcinogenesis model had a sensitivity of 0.70, specificity of 0.69 and concordance of 0.70.

The mutagen/nonmutagen (Model 1, Table III ) model was based on the complete set of 612 compounds, categorized as mutagens and nonmutagens with no regard to carcinogenicity status. As mentioned, two sets of models were developed, one based on a coverage value of 80% and the other based on a value of 90% (see Table III ). In most instances, the 80% coverage models were slightly more accurate than the 90% coverage models, and we have included results from both models for comparison ( Table III ). At this point, however, we will focus our attention on the 80% models since it is reasonable to accept and analyze fewer predictions that are more concordant with experimental results.

Table III.

Leave-one-out cross-validation results for ligand and fragment cat-SAR models based on 612 with both cancer and mutagenicity data

Model 90%  80% 
Sensitivity Specificity Concordance Sensitivity Specificity Concordance 
MUTAGENS 
Model 1 Fragment 0.78 (210/270) 0.78 (218/280) 0.78 (428/550) 0.79 (207/263) 0.80 (211/264) 0.79 (418/527) 
Mutagen/nonmutagen Ligand 0.59 (175/295) 0.79 (251/317) 0.70 (426/612) 0.61 (142/234) 0.83 (217/263) 0.72 (359/497) 
Model 2 Fragment 0.66 (210/317) 0.66 (155/236) 0.66 (365/553) 0.69 (210/303) 0.65 (147/228) 0.67 (357/531) 
Carcinogen/noncarcinogen Ligand 0.62 (207/332) 0.61 (153/251) 0.62 (360/583) 0.63 (181/286) 0.62 (131/212) 0.63 (312/498) 
Model 3 Fragment 0.80 (161/201) 0.80 (130/162) 0.80 (291/363) 0.80 (161/201) 0.80 (130/162) 0.80 (291/363) 
Carcinogen–mutagen/noncarcinogen–nonmutagen Ligand 0.72 (158/219) 0.69 (126/183) 0.71 (284/402) 0.71 (127/178) 0.73 (107/147) 0.72 (234/325) 
Model 4 Fragment 0.65 (129/199) 0.65 (44/68) 0.65 (173/267) 0.67 (117/175) 0.71 (45/63) 0.68 (162/238) 
Carcinogen–mutagen/noncarcinogen–mutagen Ligand 0.66 (144/219) 0.62 (47/67) 0.65 (191/295) 0.67 (128/190) 0.62 (41/66) 0.66 (169/256) 
NONMUTAGENS 
Model 5 Fragment 0.54 (66/123) 0.55 (97/176) 0.55 (163/299) 0.54 (66/123) 0.55 (97/176) 0.55 (163/299) 
Carcinogen–nonmutagen/noncarcinogen–nonmutagen Ligand 0.63 (80/128) 0.72 (127/176) 0.68 (207/304) 0.64 (68/106) 0.75 (114/151) 0.71 (182/257) 
Model 6 Fragment 0.62 (87/141) 0.60 (125/207) 0.61 (212/348) 0.60 (65/108) 0.65 (113/174) 0.63 (178/282) 
Malacarne Ligand 0.70 (100/142) 0.71 (147/207) 0.71 (247/349) 0.74 (84/113) 0.73 (125/171) 0.74 (209/284) 
Model 90%  80% 
Sensitivity Specificity Concordance Sensitivity Specificity Concordance 
MUTAGENS 
Model 1 Fragment 0.78 (210/270) 0.78 (218/280) 0.78 (428/550) 0.79 (207/263) 0.80 (211/264) 0.79 (418/527) 
Mutagen/nonmutagen Ligand 0.59 (175/295) 0.79 (251/317) 0.70 (426/612) 0.61 (142/234) 0.83 (217/263) 0.72 (359/497) 
Model 2 Fragment 0.66 (210/317) 0.66 (155/236) 0.66 (365/553) 0.69 (210/303) 0.65 (147/228) 0.67 (357/531) 
Carcinogen/noncarcinogen Ligand 0.62 (207/332) 0.61 (153/251) 0.62 (360/583) 0.63 (181/286) 0.62 (131/212) 0.63 (312/498) 
Model 3 Fragment 0.80 (161/201) 0.80 (130/162) 0.80 (291/363) 0.80 (161/201) 0.80 (130/162) 0.80 (291/363) 
Carcinogen–mutagen/noncarcinogen–nonmutagen Ligand 0.72 (158/219) 0.69 (126/183) 0.71 (284/402) 0.71 (127/178) 0.73 (107/147) 0.72 (234/325) 
Model 4 Fragment 0.65 (129/199) 0.65 (44/68) 0.65 (173/267) 0.67 (117/175) 0.71 (45/63) 0.68 (162/238) 
Carcinogen–mutagen/noncarcinogen–mutagen Ligand 0.66 (144/219) 0.62 (47/67) 0.65 (191/295) 0.67 (128/190) 0.62 (41/66) 0.66 (169/256) 
NONMUTAGENS 
Model 5 Fragment 0.54 (66/123) 0.55 (97/176) 0.55 (163/299) 0.54 (66/123) 0.55 (97/176) 0.55 (163/299) 
Carcinogen–nonmutagen/noncarcinogen–nonmutagen Ligand 0.63 (80/128) 0.72 (127/176) 0.68 (207/304) 0.64 (68/106) 0.75 (114/151) 0.71 (182/257) 
Model 6 Fragment 0.62 (87/141) 0.60 (125/207) 0.61 (212/348) 0.60 (65/108) 0.65 (113/174) 0.63 (178/282) 
Malacarne Ligand 0.70 (100/142) 0.71 (147/207) 0.71 (247/349) 0.74 (84/113) 0.73 (125/171) 0.74 (209/284) 

The 80% fragment model had sensitivity of 0.79, specificity of 0.80 and concordance of 0.79. The 80% ligand model for this set had a sensitivity of 0.61, specificity of 0.83 and concordance of 0.72 ( Table III ). This same set of 612 compounds also comprised the carcinogen–noncarcinogen model (Model 2, Table III ) wherein the compounds were classified as carcinogens and noncarcinogens with no regard to their mutagenicity status. Here the 80% fragment model had a sensitivity of 0.69, specificity of 0.65 and concordance of 0.67. The 80% ligand model had a sensitivity of 0.63, specificity of 0.62 and concordance of 0.63.

Considering the electrophilic theory of carcinogenesis, the most rational model is based on mutagenic carcinogens and nonmutagenic noncarcinogens (Model 3, Table III ). In this case, the 80% fragment model had sensitivity, specificity and concordance of 0.80, and the 80% ligand model had a sensitivity of 0.71, specificity of 0.73 and concordance of 0.72.

Considering these three models together, it is apparent that the fragment-based model has an acceptable degree of predictivity when the task is based on mutagenic status, whether it is the entire set of 612 compounds or the subset of mutagenic carcinogens and nonmutagenic noncarcinogens. When carcinogens and noncarcinogens have a mix of mutagens and nonmutagens, the fragment model has a much lower concordance value (i.e. 0.67, Model 2, Table III ). Also, for these three models, the fragment model always outperformed the ligand model.

Though it was possible to derive a working model for mutagens classified as carcinogens and noncarcinogens (Model 4, Table III ), it was, as expected, heavily overweighted with carcinogens (219 mutagenic carcinogens and only 76 mutagenic noncarcinogens). The 80% fragment model had a concordance of 0.68, and the 80% ligand model had a concordance of 0.66.

When just considering the nonmutagenic compounds classified as carcinogens or noncarcinogens, the carcinogen-nonmutagen/noncarcinogen-nonmutagen (Model 5, Table III ) model was developed. Here the 80% fragment model had a sensitivity of 0.54, specificity of 0.55 and concordance of 0.55, and the ligand model had a sensitivity of 0.64, specificity of 0.75 and concordance of 0.71. Interestingly, it appears that the fragment model essentially had no ability to classify nonmutagenic compounds as carcinogens or noncarcinogens and the ligand model had an acceptable concordance of 0.71.

As mentioned, a noncarcinogen-mutagen/noncarcinogen-nonmu tagen model was also developed based primarily on a publication by Malacarne et al. ( 35 ). For this set, 41 compounds considered noncarcinogens that were only tested in one gender were not included. Based on this data set, the 80% fragment model had a sensitivity of 0.60, a specificity of 0.65 and a concordance of 0.63, and the 80% ligand model had a sensitivity of 0.74, a specificity of 0.73 and a concordance of 0.74. The original authors reported a sensitivity of 0.64, specificity of 0.67 and concordance of 0.66 for the 80% fragment model.

It is, therefore, interesting that when considering nonmutagenic compounds, the 80% ligand-based model is superior to the 80% fragment-based model. One the other hand, although considering the electrophilic theory, the carcinogen-mutagen/noncarcinogen nonmutagen model, the fragment model was superior to the ligand model (concordance values of 0.80 and 0.71, respectively). Moreover, as mentioned above, the 80% fragment model had a sensitivity of 0.69, a specificity of 0.65 and a concordance of 0.67, and the 80% ligand model had a sensitivity of 0.63, a specificity of 0.62 and a concordance of 0.63.

Additionally, after LOO predictions were made for the set of 612 compounds, the results were categorized by mutagenicity status ( Table IV ). Looking at subsets of these models based on mutagenicity, the fragment model had a sensitivity of 75.0, a specificity of 74.2 and a concordance of 74.8% for predicting the carcinogenicity of mutagens, and a sensitivity of 50.4, a specificity of 62.0 and a concordance of 57.3% for predicting the carcinogenicity of nonmutagens ( Table IV ). This suggests that the fragment model is capable of learning and predicting the carcinogenicity of mutagenic compounds but essentially unable to accurately assess the nonmutagenic compounds. On the other hand, the ligand model, predicting this mixed pool of mutagens and nonmutagens, had a concordance of 62.9% for the mutagens and 60.7% for the nonmutagens, with minimal differences between the mutagen and nonmutagen subsets ( Table IV ). This indicates that the inclusion of mutagenic compounds in the ligand model’s learning set reduces its overall predictivity, leaving it well suited for analyzing only nonmutagens as described above.

Table IV.

Leave-one-out analysis of Model 2 fragment and ligand model's ability to predict carcinogenicity based on the mutagenicity status of compounds

Fragment model  Correct predictions Total predictions % Correct No predictions 
Carcinogen Mutagen     
Positive Positive 153 204 0.75 15 
Negative Positive 52 70 0.74 
 Total 205 274 0.75  
Positive Negative 57 113 0.50 21 
Negative Negative 103 166 0.62 17 
 Total 160 279 0.57  
 Overall 365 553 0.66  
Ligand model 
Carcinogen Mutagen     
Positive Positive 130 203 0.64 16 
Negative Positive 43 72 0.60 
 Total 173 275 0.63  
Positive Negative 77 129 0.60 
Negative Negative 110 179 0.61 
 Total 187 308 0.61  
 Overall 360 583 0.62  
The bold/italic values labeled “overall” are the grand total for all four combinations of carcinogens and mutagens. 
Fragment model  Correct predictions Total predictions % Correct No predictions 
Carcinogen Mutagen     
Positive Positive 153 204 0.75 15 
Negative Positive 52 70 0.74 
 Total 205 274 0.75  
Positive Negative 57 113 0.50 21 
Negative Negative 103 166 0.62 17 
 Total 160 279 0.57  
 Overall 365 553 0.66  
Ligand model 
Carcinogen Mutagen     
Positive Positive 130 203 0.64 16 
Negative Positive 43 72 0.60 
 Total 173 275 0.63  
Positive Negative 77 129 0.60 
Negative Negative 110 179 0.61 
 Total 187 308 0.61  
 Overall 360 583 0.62  
The bold/italic values labeled “overall” are the grand total for all four combinations of carcinogens and mutagens. 

Conclusions

The NTP’s 11th Report on Carcinogens lists 58 compounds ‘known to be human carcinogens’ and 188 ‘reasonably anticipated to be human carcinogens ( 46 )’. However, only a small number of the upward of 80 000 chemicals manufactured and used in the USA have been analyzed in vivo or in vitro for carcinogenic potential. Computational SAR models are being used by regulatory agencies worldwide for both human health ( 26 ) and ecological endpoints ( 27 ). The present investigation consisted of SAR analysis for expected non-DNA-reactive compounds. One model was based on 612 rat carcinogens reported in the CPDB, which also had the Salmonella mutagenicity data, and the other model was for nonalerting rodent carcinogenicity data published by Malacarne et al. ( 35 ). Cat-SAR analysis of this data set was based on a traditional fragment-based approach along with a new approach using virtual screening to develop SAR descriptors. The rationale for this approach is that these nonalerting carcinogens interact with cellular targets other than DNA.

It appears, as analyzed here, chemical substructures are best suited to modeling mutagens and our novel ligand-based approach is well suited to modeling nonmutagenic compounds. This can be rationalized on the basis that, as described on numerous occasions, mutagenic compounds have clearly definable substructures that are electrophilic or proelectrophilic. And it is these defined chemical structures or proelectrophilic moieties that interact covalently with DNA. Generally, nonmutagenic carcinogens do not have obvious chemical features for their identification. According to Woo, identification of structural attributes of non-DNA-reactive carcinogens is not as advanced as that of genotoxic carcinogens ( 23 ). Moreover, nonmutagenic carcinogens, as a class, do not covalently bind with cellular components like DNA but, as outlined by Combes ( 15 ), may interact with a significant number of cellular targets through noncovalent mechanisms (e.g. receptor–ligand interactions).

The cat-SAR ligand-based SAR model did not follow the paradigm of analyzing chemicals for substructures that could be related to carcinogenic activity but rather analyzed their global structure and determined whether they were potential ligands for a set of 5494 protein targets. Therefore, although this ligand-based approach for analyzing nonmutagenic chemicals had an acceptable predictive concordance of ~70% as determined by LOO, it is possible that this approach has applicability to mechanisms of nonmutagenic carcinogenesis and/or specific molecular targets associated with nonmutagenic carcinogenesis.

We speculate that the identified ligands associated with nonmutagenic carcinogens can be used to explore the previously established or hypothesized mechanisms or, as importantly, to develop new testable hypotheses relating to non-DNA-reactive chemical carcinogens. Considering both the fragment and ligand models, the fragment-based approach is best suited for identifying structural attributes of DNA-reactive chemicals and the ligand-based approach is suited for identifying structural attributes of non-DNA-reactive carcinogens. Hence, separate SAR models for mutagenic and nonmutagenic compounds may be warranted. It is conceivable that defined sets of compounds that induce nonmutagenic carcinogenesis may mostly influence molecular target(s) identifiable by virtual screening. However, although our ligand-based model for nonmutagenic compounds was as predictive as other SAR models for carcinogens, it was superior for nonmutagenic carcinogens, indicating a carcinogen–receptor mode of action.

Keiser et al. ( 47 ), based on virtual screening, confirmed in vitro that individual drugs interact with a significant number of receptors and that the highest affinity targets are not necessarily the drugs’ therapeutic targets. Therefore, as the models presented here used 250 protein targets for each compound, we expect, as is the case in vitro , that a significant number of protein descriptors are probably to be ‘off-target’ interactions.

Finally, the cat-SAR expert system used here is a knowledge-based approach where the knowledge is contained in the learning set rather than a hypothesis-driven approach. As such, the mechanism of action of derived models does not dependent on previous knowledge or assumptions. This is particularly important since as a generality, nonmutagenic carcinogens do not have a specific mechanism of action that has been understood. It is worth mentioning that the selection of chemicals that went into the modeling process was based on a hypothesis that there are structural differences that distinguish nonmutagenic carcinogens from nonmutagenic noncarcinogens.

To determine whether a nonmutagenic chemical will be carcinogenic is an important though difficult task. The ‘importance’ lays in the fact that one or more tests for mutagenicity, being fast and economical, is often used as a first (and perhaps only) line of tests to assess potential carcinogenicity. Although positive results in these assays are considered indicative of carcinogenicity, negative tests are considered indicative of noncarcinogenicity, with the recognition that ~38% of nonmutagenic compounds are carcinogens. Hence, potential exposures to nonmutagenic compounds not further tested for carcinogenicity are a public health concern. Furthermore, not detecting a nonmutagenic carcinogen until late in the developmental phase for new commercial chemicals and drugs has significant economic consequences.

Supplementary material

Supplementary Figures 1–6 can be found at http://carcin.oxfordjournals.org/

Funding

National Institutes of Health (P20 RR018733); James Graham Brown Cancer Center, University of Louisville.

Conflict of Interest Statement: A.R.C. and S.L.C. are co-founders of Gnarus Systems Inc. Gnarus Systems Inc. is a start-up company that will use and license the cat-SAR expert system used in the research described in the manuscript. This research has also been used for a patent application by A.R.C. and J.O.T. (Hybrid Fragment-Ligand Modeling for Classifying Chemical Compounds, Patent Application Serial No. 61/380,048 on 3 September 2011).

References

1.
Rall
D.P.
et al
1987
Alternatives to using human experience in assessing health risks.
Annu. Rev. Public Health
 
8
355
385
2.
Maronpot
R.R.
et al
2004
Relevance of animal carcinogenesis findings to human cancer predictions and prevention.
Toxicol. Pathol.
 
32
40
48
3.
Huff
J.
1994
Chemicals causally associated with cancer in humans and in laboratory animals
. In
Waalkes
M.P.
Ward
J.M.
(eds.)
Carcinogenesis
 
Raven Press, Ltd.
New York, NY
pp.
25
37
4.
NIEHS.
1996
Scientists Try to Match Rodent Results in Testing Chemicals.
http://www.niehs.nih.gov/news/newsroom/releases/news-archive/1996/november19/index.cfm. (last accessed 08/03/12)
5.
NTP.
2012
Long-Term Study Reports & Abstracts.
http://ntp.niehs.nih.gov/?objectid=D16D6C59-F1F6-975E-7D23D1519B8CD7A5. (last accessed (08/03/12).
6.
Gold
L.S.
et al
1997
Overview and update of analyses of the carcinogenic potency database
. In
Gold
L.S.
Zeiger
E.
(eds.)
Handbook of Carcinogenic Potency and Genotoxicity Databases
 
CRC Press
New York, NY
pp.
661
693
7.
Gold
L.S.
2012
Carcinogenic Potency Database.
http://potency.berkeley.edu. (last accessed 08/03/12).
8.
DSSTox.
2012
Distributed Structure-Searchable Toxicity (DSSTox) Database Network.
http://www.epa.gov/NCCT/dsstox/index.html. (last access 08/03/12).
9.
EPA
2012
TSCA Chemical Substance Inventory: Background Information.
http://www.epa.gov/oppt/existingchemicals/pubs/tscainventory/basic.html (last accessed 08/03/12).
10.
11.
Committee on Toxicity Testing and Assessment of Environmental Agents.
2007
Toxicity Testing in the 21st Century: A Vision and a Strategy.
National Research Council
:
Washington D.C.
12.
Andersen
M.E.
et al
2010
The future of toxicity testing.
J. Toxicol. Environ. Health B
 
13
163
196
13.
Kavlock
R.
et al
2010
Computational toxicology as implemented by the U.S. EPA: providing high throughput decision support tools for screening and assessing chemical exposure, hazard and risk.
J. Toxicol. Environ. Health B
 
13
197
217
14.
Zeiger
E.
2010
Historical perspective on the development of the genetic toxicity test battery in the United States.
Environ. Mol. Mutagen.
 
51
781
791
15.
Combes
R.D.
2000
The use of structure-activity relationships and markers of cell toxicity to detect non-genotoxic carcinogens.
Toxicol. In Vitro
 
14
387
399
16.
Benigni
R.
et al
2010
Alternatives to the carcinogenicity bioassay: in silico methods, and the in vitro and in vivo mutagenicity assays.
Expert Opin. Drug Metab.
 
6
1
11
17.
Zeiger
E.
1998
Identification of rodent carcinogens and noncarcinogens using genetic toxicity tests: premises, promises, and performance.
Regul. Toxicol. Pharmacol.
 
28
85
95
18.
Miller
J.A.
et al
1977
Ultimate chemical carcinogens as reactive mutagenic electrophiles
. In
Hiatt
H.H.
Watson
J.D.
Winsten
J.A.
(eds.)
Origins of Human Cancer
 
Cold Spring Harbor Laboratory Press
Cold Spring Harbor, New York
pp.
605
627
19.
Ashby
J.
1985
Fundamental structural alerts to potential carcinogenicity or noncarcinogenicity.
Environ. Mutagen.
 
7
919
921
20.
Ashby
J.
et al
1991
Definitive relationships among chemical structure, carcinogenicity and mutagenicity for 301 chemicals tested by the U.S. NTP.
Mutat. Res.
 
257
229
306
21.
Ashby
J.
et al
1993
The influence of chemical structure on the extent and sites of carcinogenesis for 522 rodent carcinogens and 55 different human carcinogen exposures.
Mutat. Res.
 
286
3
74
22.
Ashby
J.
et al
1994
Prediction of rodent carcinogenicity for 44 chemicals: results.
Mutagenesis
 
9
7
15
23.
Woo
Y.T.
2003
Mechanisms of action of chemical carcinogens, and their role in structure-activity relationships (SAR) analysis and risk assessment
. In
Benigni
R.
(ed.)
Quantitative Structure-Activity Relationship (QSAR) Models of Mutagens and Carcinogens
 
CRC Press
Boca Raton, FL
pp.
41
80
24.
Blake
B.W.
et al
1990
Salmonella mutagenicity and rodent carcinogenicity: quantitative structure-activity relationships.
Mutat. Res.
 
241
261
271
25.
Marchant
C.A.
1996
Prediction of rodent carcinogenicity using the DEREK system for 30 chemicals currently being tested by the National Toxicology Program.
Environ. Health Perspect.
 
105
1065
1073
26.
Cronin
M.T.D.
et al
2003
Use of QSARs in international decision-making frameworks to predict health effects of chemical substances.
Environ. Health Perspect.
 
111
1391
1401
27.
Cronin
M.T.D.
et al
2003
Use of QSARs in international decision-making frameworks to predict ecologic effects and environmental fate of chemical substances.
Environ. Health Perspect.
 
111
1376
1390
28.
EPA (2012) OncoLogic™ - A Computer System to Evaluate the Carcinogenic Potential of Chemicals. http://www.epa.gov/oppt/sf/pubs/oncologic.htm. (last accessed 08/03/12).
29.
Contrera
J.F.
et al
2007
Comparison of MC4PC and MDL-QSAR rodent carcinogenicity predictions and the enhancement of predictive performance by combining QSAR models.
Regul. Toxicol. Pharmacol.
 
49
172
182
30.
Matthews
E.
et al
1998
A new highly specific method for predicting the carcinogenic potential of pharmaceuticals in rodents using enhanced MCASE QSAR-ES software.
Regul. Toxicol. Pharmacol.
 
28
242
264
31.
Matthews
E.
et al
2006
An analysis of genetic toxicity, reproductive and developmental toxicity, and carcinogenicity data: I. Identification of carcinogens using surrogate endpoints.
Regul. Toxicol. Pharmacol.
 
44
83
96
32.
Brambilla
G.
et al
2009
Update on genotoxicity and carcinogenicity testing of 472 marketed pharmaceuticals.
Mutat. Res.
 
681
209
229
33.
Snyder
R.D.
et al
2004
Assessment of the sensitivity of the computational programs DEREK, TOPKAT, and MCASE in the prediction of the genotoxicity of pharmaceutical molecules.
Environ. Mol. Mutagen.
 
43
143
158
34.
Nioi
P.
et al
2008
Erratum to “Prediction of non-genotoxic carcinogenesis in rats using changes in gene expression following acute dosing”.
Chem. Biol. Interact.
 
176
252
260
35.
Malacarne
D.
et al
1995
Molecular fragments associated with non-genotoxic carcinogens, as detected using a software program based on graph theory: their usefulness to predict carcinogenicity.
Chem. Biol. Interact.
 
97
75
100
36.
Lee
et al
1995
Learning rules to predict rodent carcinogenicity of non-genotoxic chemicals.
Mutat. Res.
 
328
127
149
37.
Zhu
H.
et al
2008
Use of cell viability assay data improves the prediction accuracy of conventional quantitative structure-activity relationship models of animal carcinogenicity.
Environ. Health Perspect.
 
116
506
513
38.
Cunningham
A.R.
et al
2010
Mammary carcinogen-protein binding potentials: novel and biologically relevant structure-activity relationship model descriptors.
SAR QSAR Environ. Res.
 
21
463
479
39.
Gold
L.S.
et al
1984
A carcinogenic potency database of the standardized results of animal bioassays.
Environ. Health Perspect.
 
58
9
319
40.
Kellenberger
E.
et al
2006
sc-PDB: an annotated database of druggable binding sites from the Protein Data Bank.
J. Chem. Inf. Model.
 
46
717
727
41.
Rognan
D.
2007
http://bioinfo-pharma.u-strasbg.fr/scPDB/. (last access 08/03/12).
42.
Jain
A.N.
2003
Surflex: fully automatic flexible molecular docking using a molecular similarity-based search engine.
J. Med. Chem.
 
46
499
511
43.
Jain
A.N.
2007
Surflex-Dock 2.1: robust performance from ligand energetic modeling, ring flexibility, and knowledge-based search.
J. Comput. Aided Mol. Des.
 
21
281
306
44.
Cunningham
A.R.
et al
1998
Identification of structural features and associated mechanisms of action for carcinogens in rats.
Mutat. Res.
 
405
9
28
45.
NTP.
2004
Results, Status and Publication Information on All NTP Chemicals
. National Toxicology Program. http://www.predictive-toxicology.org/data/ntp/original_ntp_data.txt (last accessed 08/03/12).
46.
NTP.
2005
Report on Carcinogens, Eleventh Edition. Research Triangle Park, NC: U.S. Department of Health and Human Services, Public Health Service, National Toxicology Program.
47.
Keiser
M.J.
et al
2009
Predicting new molecular targets for known drugs.
Nature
 
462
175
181

Abbreviations:

    Abbreviations:
  • cat-SAR

    categorical SAR

  • CPDB

    Carcinogenic Potency Database

  • LOO

    leave-one-out

  • NTP

    National Toxicology Program

  • SAR

    structure–activity relationship.