Testing by artificial intelligence: computational alternatives to the determination of mutagenicity

Applications of intelligent technology in the evaluation of mutagenicity

Bioassays are widely used in assessment of mutagenicity. Alternative methods have also been developed, including "intelligent evaluation", which depends on the quality of data, strategies, and techniques. CISOC-PSMT is an Ames test prediction system. The strategies and techniques for intelligent evaluation and four applications of CISOC-PSMT are presented; roles in pesticide management, environmental protection, drug discovery, and safety management of chemicals are introduced.

In vitro and in silico genetic toxicity screening of flavor compounds and other ingredients in tobacco products with emphasis on ENDS

Electronic nicotine delivery systems (ENDS) are regulated tobacco products and often contain flavor compounds. Given the concern of increased use and the appeal of ENDS by young people, evaluating the potential of flavors to induce DNA damage is important for health hazard identification. In this study, alternative methods were used as prioritization tools to study the genotoxic mode of action (MoA) of 150 flavor compounds. In particular, clastogen-sensitive (γH2AX and p53) and aneugen-sensitive (p-H3 and polyploidy) biomarkers of DNA damage in human TK6 cells were aggregated through a supervised three-pronged ensemble machine learning prediction model to prioritize chemicals based on genotoxicity. In addition, in silico quantitative structure-activity relationship (QSAR) models were used to predict genotoxicity and carcinogenic potential. The in vitro assay identified 25 flavors as positive for genotoxicity: 15 clastogenic, eight aneugenic and two with a mixed MoA (clastogenic and aneugenic). Twenty-three of these 25 flavors predicted to induce DNA damage in vitro are documented in public literature to be in e-liquid or in the aerosols produced by ENDS products with youth-appealing flavors and names. QSAR models predicted 46 (31%) of 150 compounds having at least one positive call for mutagenicity, clastogenicity or rodent carcinogenicity, 49 (33%) compounds were predicted negative for all three endpoints, and remaining compounds had no prediction call. The parallel use of these predictive technologies to elucidate MoAs for potential genetic damage, hold utility as a screening strategy. This study is the first high-content and high-throughput genotoxicity screening study with an emphasis on flavors in ENDS products.

Computational techniques for the prediction of toxicity

Computational methods for the prediction of biological activity have been in use since the early 1960s. The application of these approaches in attempts to predict toxicological endpoints was reported within a few years of the establishment of the techniques, although early work was criticized on a number of grounds. This report reviews three approaches to the prediction of toxicity with examples of each approach. A new commercially available system (APEX) for the identification of pharmacophores (toxicophores) based on the three-dimensional structures of database compounds has been applied to a literature dataset of mutagenicity results. This program achieved a success rate of approximately 75% when trained on a set of 105 compounds; comparable results from other prediction systems, although trained on a much larger dataset, are approximately 85-90%.

Prediction of mutagenic toxicity by combination of Recursive Partitioning and Support Vector Machines

The study of prediction of toxicity is very important and necessary because measurement of toxicity is typically time-consuming and expensive. In this paper, Recursive Partitioning (RP) method was used to select descriptors. RP and Support Vector Machines (SVM) were used to construct structure-toxicity relationship models, RP model and SVM model, respectively. The performances of the two models are different. The prediction accuracies of the RP model are 80.2% for mutagenic compounds in MDL's toxicity database, 83.4% for compounds in CMC and 84.9% for agrochemicals in in-house database respectively. Those of SVM model are 81.4%, 87.0% and 87.3% respectively.

A multiple in silico program approach for the prediction of mutagenicity from chemical structure

We have conducted an evaluation of three of the most widely used commercial toxicity prediction programs, Toxicity Prediction by Komputer Assisted Technology (TOPKAT), Deductive Estimation of Risk from Existing Knowledge (DEREK) for Windows (DfW) and CASETOX. The three programs were evaluated for their ability to predict Ames test mutagenicity using 520 proprietary drug candidate (Test set 1) and 94 commercial (Test set 2) compounds. The study demonstrates that these three commercially available programs are useful, with limitations in their ability to predict mutagenicity over a wide range of chemical space, i.e. global predictivity. Individually, each of the programs performed at an acceptable level for overall accuracy, i.e. the ability to predict the correct outcome. However, analysis of the predictions indicates that the overall accuracy figure is heavily weighted by the ability of the programs to correctly predict non-mutagens, whereas none of the programs individually performed well in the prediction of novel mutagenic structures, i.e. Ames positive compounds. The performance of these programs' in predicting Ames positive mutagens appeared to be independent of the chemical utility of the compound, i.e. industrial, agricultural or pharmaceutical. The combination of program predictions provided some improvement in overall accuracy, sensitivity and specificity.

Quantitative structure-based modeling applied to characterization and prediction of chemical toxicity

Quantitative modeling methods, relating aspects of chemical structure to biological activity, have long been applied to the prediction and characterization of chemical toxicity. The early linear free-energy approaches of Hansch and Free Wilson provided a fundamental scientific framework for the quantitative correlation of chemical structure with biological activity and spurred many developments in the field of quantitative structure-activity relationships (QSARs). In addition to modeling of chemical toxicity, these methods have been extensively applied to modeling of medicinal properties of chemicals. However, there are important differences in the nature and objectives of these two applications, which have led to the evolution of different modeling approaches (namely, the need for treating sets of noncongeneric toxic compounds). In this paper are discussed those approaches to chemical toxicity that have taken a more "personalized" configuration and have undergone implementation into software programs able to perform the various steps of the assessment of the hazard posed by the chemicals. These models focus both on a variety of toxicological endpoints and on key elements of toxicity mechanisms, such as metabolism.

Toxic effects of griseofulvin: disease models, mechanisms, and risk assessment

Griseofulvin (GF) has been in use for more than 30 years as a pharmaceutical drug in humans for the treatment of dermatomycoses. Animal studies give clear evidence that it causes a variety of acute and chronic toxic effects, including liver and thyroid cancer in rodents, abnormal germ cell maturation, teratogenicity, and embroyotoxicity in various species. No sufficient data from human studies are available at present to exclude a risk in humans: therefore, attempts were made to elucidate the mechanisms responsible for the toxic effects of GF and to address the question whether such effects might occur in humans undergoing GF therapy. It is well documented that GF acts as a spindle poison and its reproductive toxicity as well as the induction of numerical chromosome aberrations and of micronuclei in somatic cells possibly may result from disturbance of microtubuli formation. Likewise, a causal relationship between aneuploidy and cancer has been repeatedly postulated. However, a critical survey of the data available on aneuploidogenic chemicals revealed insufficient evidence for such an association. Conceivably, other mechanisms may be responsible for the carcinogenic effects of the drug. The induction of thyroid tumors in rats by GF is apparently a consequence of the decrease of thyroxin levels and it is unlikely that such effects occur in GF-exposed humans. The appearance of hepatocellular carcinomas (HCC) in mice on GF-supplemented diet is preceded by various biochemical and morphological changes in the liver. Among these, hepatic porphyria is prominent, it may result from inhibition of ferrochelatase and (compensatory) induction of ALA synthetase. GF-induced accumulation of porphyrins in mouse liver is followed by cell damage and necrotic and inflammatory processes. Similar changes are known from certain human porphyrias which are also associated with an increased risk for HCC. However, the porphyrogenic effect of GF therapy in humans is moderate compared with that in the mouse model, although more detailed studies should be performed in order to clarify this relationship on a quantitative basis. A further important effect of GF-feeding in mice is the formation of Mallory bodies (MBs) in hepatocytes. These cytoskeletal abnormalities occur also in humans, although under different conditions; their appearance is associated with the induction of liver disease and HCC. Chronic liver damage associated with porphyria and MB formation, enhanced cell proliferation, liver enlargement, and enzyme induction all may contribute to the hepatocarcinogenic effect of GF in mice. In conclusion, further investigation is required for adequate assessment of health risks to humans under GF therapy.

Methods for predicting carcinogenic hazards: new opportunities coming from recent developments in molecular oncology and SAR studies

Without epidemiological evidence, and prior to either short-term tests of genotoxicity or long-term tests of carcinogenicity in rodents, an initial level of information about the carcinogenic hazard of a chemical that perhaps has been designed on paper, but never synthesized, can be provided by structure-activity relationship (SAR) studies. Herein, we have reviewed the interesting strategies developed by human experts and/or computerized approaches for the identification of structural alerts that can denote the possible presence of a carcinogenic hazard in a novel molecule. At a higher level of information, immediately below epidemiological evidence, we have discussed carcinogenicity experiments performed in new types of genetically engineered small rodents. If a dominant oncogene is already mutated, or if an allele of a recessive oncogene is inactivated, we have a model animal with (n-1) stages in the process of carcinogenesis. Both genotoxic and receptor-mediated carcinogens can induce cancers in 20-40% of the time required for classical murine strains. We have described the first interesting results obtained using these new artificial animal models for carcinogenicity studies. We have also briefly discussed other types of engineered mice (lac operon transgenic mice) that are especially suitable for detecting mutagenic effects in a broad spectrum of organs and tissues and that can help to establish mechanistic correlations between mutations and cancer frequencies in specific target organs. Finally, we have reviewed two complementary methods that, while obviously also feasible in rodents, are especially suitable for biomonitoring studies. We have illustrated some of the advantages and drawbacks related to the detection of DNA adducts in target and surrogate tissues using the 32P-DNA postlabeling technique, and we have discussed the possibility of biomonitoring mutations in different human target organs using a molecular technique that combines the activity of restriction enzymes with polymerase chain reaction (RFLP/PCR). Prediction of carcinogenic hazard and biomonitoring are very wide-ranging areas of investigation. We have therefore selected five different subfields for which we felt that interesting innovations have been introduced in the last few years. We have made no attempt to systematically cover the entire area: such an endeavor would have produced a book instead of a review article.

Computational predictive programs (expert systems) in toxicology

The increasing number of pollutants in the environment raises the problem of the toxicological risk evaluation of these chemicals. Several so called expert systems (ES) have been claimed to be able to predict toxicity of certain chemical structures. Different approaches are currently used for these ES, based on explicit rules derived from the knowledge of human experts that compiled lists of toxic moieties for instance in the case of programs called HazardExpert and DEREK or relying on statistical approaches, as in the CASE and TOPKAT programs. Here we describe and compare these and other intelligent computer programs because of their utility in obtaining at least a first rough indication of the potential toxic activity of chemicals.

Estimation of the optimal data base size for structure-activity analyses: the Salmonella mutagenicity data base

In the present study, the effects of data base size on predictivity, informational content and structural overlap of derived Structure-Activity Relationship (SAR) models were investigated. It was found that indices of predictivity (i.e., sensitivity, specificity, and concordance between experimental and predicted results (OCP) increased with increasing size of the data base until the range is 300-400 chemicals, at which point they plateau. The greater the size of the data base, the greater the informational content of the model; however, the rate of this increase is no longer optimal when the size of the data base exceeds 400 chemicals.

A computerized connectivity approach for analyzing the structural basis of mutagenicity in Salmonella and its relationship with rodent carcinogenicity

We have applied a new software program, based on graph theory and developed by our group, to predict mutagenicity in Salmonella. The software analyzes, as information in input, the structural formula and the biological activities of a relatively large database of chemicals to generate any possible molecular fragment with size ranging from two to ten nonhydrogen atoms, and detects (as predictors of biological activity) those fragments statistically associated with the biological property investigated. Our previous work used the program to predict carcinogenicity in small rodents. In the current work we applied a modified version of the program, which bases its predictions solely on the most important fragment present in a given molecule, considering as practically negligible the effects of additional less important fragments. For Salmonella mutagenicity we used a database of 551 compounds, and the program achieved a level of predictivity (73.9%) comparable to that obtained by other authors using the Computer Automated Structure Evaluation (CASE) program. We evaluated the relative contributions of biophores and biophobes to overall predictivity: biophores tended to be more important than biophobes, and chemicals containing both biophores and biophobes were more difficult to predict. Many of the molecular fragments identified by the program as being strongly associated with mutagenic activity were similar to the structural alerts identified by the human experts Ashby and Tennant. Our results tend to confirm that structural alerts useful to predict Salmonella mutagenicity are generally not very strong predictors of rodent carcinogenicity. Although the predictivity level achieved for oncogenic activity improved when the program was directly trained with carcinogenicity data, carcinogenicity as a biological endpoint was still more difficult to predict than Salmonella mutagenicity.

The structure-toxicity relationship challenge at hazardous waste sites

It is shown that the use of a combination of two programs, MULTICASE and META can help assess the carcinogenic risk factor posed by the disposal of industrial organic materials in the ecosystem. MULTICASE is a knowledge-based computer system that had been trained to identify molecular substructures believed to be conducive to carcinogenic potential and META is an expert system trained to predict the aerobic biodegradation products of organic molecules. The programs can be used to assess the health hazard of the discarded chemicals by evaluating their chemical structure, their biodegradability and the structures of the predicted biodegradation products. Several examples of the application of the methodology are described in this paper.

Molecular fragments associated with non-genotoxic carcinogens, as detected using a software program based on graph theory: their usefulness to predict carcinogenicity

We assembled 390 chemicals with a structure non-alerting to DNA-reactivity (145 carcinogens and 245 non-carcinogens) for which rodent carcinogenicity data were available. These non-alerting chemicals were defined by the absence in their molecules of DNA-reactive (directly or after metabolic activation) alerting structures, as described by Ashby and coworkers (Mutat. Res., 204 (1988) 17-115; Mutat. Res., 223 (1989) 73-103; Mutat. Res., 257 (1991) 209-227; Mutat. Res., 286 (1993) 3-74). Using our software program based on graph theory we analyzed the compounds in order to estimate the program's ability to predict nonalerting carcinogens. Our software fragmented the structural formula of the chemicals into all possible fragments of contiguous atoms with size between 2 and 8 (non-hydrogen) atoms and learned about statistically significant fragments from a training set of chemicals. These fragments were used to predict carcinogenicity or lack thereof in a verification set of compounds. For 390 runs of the software program we used (n - 1) of the chemicals as a training set, to predict the excluded chemical at each run (as a test set). Using two different probability thresholds to select significant fragments (P = 0.05 and P = 0.125 1-tailed according to binomial distribution), we performed two analyses: in the better one (P = 0.05) 19% of the molecules tested lacked significant fragments, for the remaining 81% the observed level of accuracy of the prediction was 66.0% against an expected level of accuracy of 51.7%. The difference was highly significant (P < 0.0001). We also examined the more significant activating fragments (biophores) and discussed at length both their biological plausibility and the working hypothesis that additional alerting structures for carcinogenicity (not only those related to genotoxicity) can be detected using this type of SAR approach. This new class of alerting structures could identify subfamilies of congeneric analogs active through mechanisms of receptor mediated carcinogenesis.

Structural basis for the induction of preneoplastic glutathione S-transferase positive foci by hepatocarcinogens

A data base consisting of 100 chemicals tested for the ability to enhance the formation of glutathione-S-transferase (GST) positive preneoplastic lesions were analyzed by the CASE structure-activity relational system. A number of structural determinants associated with the induction of GST-positive foci were recognized. The majority of these describe non-electrophilic moieties. It is concluded that there is a structural basis for the induction of these neoplastic lesions; interestingly, it was found that this activity is associated with structures that are non-electrophilic. Reconstruction experiments have indicated that the identified structures are meaningful and that their significance could be better understood with the availability of test results on additional chemicals.