• School of Materials Science and Engineering, Shanghai University, People's Republic of China
• Department of Chemistry, College of Sciences, Shanghai University, People's Republic of China

Early studies of structure-activity relationships (SAR) among cytotoxic DNA intercalating agents that act as topoisomerase (topo) inhibitors suggested a positive correlation between cytotoxic potency and the strength of reversible DNA binding. As bis-intercalation can theoretically greatly increase DNA binding, substituted bis[(acridine-4-carboxamides)propyl]methylamines have been reported as a new class of anticancer agents, which was in Phase II clinical trials on the basis of its dual inhibition of both topoI and topoII, and activity in resist cell lines and in experimental solid tumor models. How to design and synthesis new analogues of the acridinecarboxamide derivative N-[(2-dimethylami-no)ethyl]acridine-4-carboxamide compound with higher activity interested many scientists.
People have already measured its biologic activity, but there are little quantitative structure-activity relationship (QSAR) studies have been reported. Hence, in this work, are studied with the application of SVM in the field QSAR.
After feature selection, Electronic Energy, Nuclear Energy, Volume, LogP were selected as descriptors in building QSAR model. By using the leave-one-out cross-validation (LOOCV) test, one QSAR model of substituted bis[(acridine-4-carboxamides)propyl]methylamines is attained. As a result, the MRE (mean relative error) of this QSAR model achieved 6.819%, which is better than the result of multiple linear regression. And the promising result show that SVM may hold a high potential for improving the quality in predicting the activity of substituted bis[(acridine-4-carboxamides)propyl]methylamines as well. Hence, as a robust algorithm, SVM has a bright future as a powerful tool for drug design and related fields. Or at the very least, it can be regard as a supplement to many of the existing algorithms in this regard.

• Istituto Superiore di Sanità, Environment and Health Department, Rome, Italy

The European Chemicals Bureau (ECB) has developed (through IdeaConsult Ltd.) a software tool called ToxTree that is able to estimate different types of toxic hazards by applying structural rules. This is a freely available application from the ECB website (http://ecb.jrc.it/QSAR). Recently, a new module with rules for predicting the carcinogenicity and mutagenicity of chemicals has been implemented. The core of the module is a list of Structural Alerts (SA) for carcinogenicity. The SAs for carcinogenicity are molecular functional groups or substructures known to be linked to the carcinogenic activity of chemicals. As one or more SAs embedded in a molecular structure are recognized, the system flags the potential carcinogenicity of the chemical. The SAs in Toxtree refer mainly to the knowledge on the action mechanisms of genotoxic carcinogenicity (thus they apply also to the mutagenic activity in bacteria), but include also a number of SAs flagging potential nongenotoxic carcinogens. The list of SAs derives from the critical evaluation of different existing proposed sets of SAs. Because of their nature, the SAs have the role of pointing to chemicals potentially toxic, whereas no conclusions or indications about nontoxic chemicals are possible (except by exclusion). Thus the SAs are not a discriminant model on the same ground of the Quantitative Structure-Activity Relationships (QSAR) models, that produce estimates for both positive and negative chemicals. In addition to the SAs, this software includes QSAR models for: 1) the mutagenic activity of aromatic amines in the Salmonella typhimurium TA100 strain (Ames test); 2) the carcinogenic activity of the aromatic amines in rodents (summary activity from rats and mice); 3) the mutagenic activity of αβ-unsaturated aldehydes in the Salmonella typhimurium TA100 strain (Ames test).

• Kiwa Water Research, Chemical Water Quality and Health, Nieuwegein, Netherlands

Dutch drinking water and especially its sources are continuously analyzed chemically for the presence of contaminants to safeguard its good quality and safety. In evaluating the health risk for drinking water consumers from these compounds, we find that for around one third, no or insufficient toxicity data can be found. Therefore, we have recently begun to apply Quantitative or Qualitative Structure Activity Relationships (QSARs) to screen these compounds on toxicity potential and subsequently prioritize them for further study. We have first performed a limited comparative study with several different commercial QSAR software packages (DEREK, TOPKAT, CASE, HazardExpert and OncoLogic), in order to choose which one to purchase and use. Ultimately, we had a preference for TOPKAT, but by budget limitations we have chosen to purchase HazardExpert. We have now analyzed a number of detected and expected water contaminants with HazardExpert, of which most gave no alerts for the covered toxic effects. However, some substances, such as some benzotriazoles and benzothiazoles, did give serious alerts.

• National Institute of Chemistry Slovenia, Laboratory of Chemometrics, Ljubljana, Slovenia

In the context of EU legislation, such as REACH and the Cosmetics Directive (Council Directive 2003/15/EC), it is anticipated that (Q)SARs will be used more extensively, in the interests of time- and cost-effectiveness and animal welfare. Thousands of predictive models have been published in recent years, but typically they are not suitable for regulatory purposes, because they have not taken into account essential factors for validation or quality assurance, and specific requirements for Regulation. CAESAR is one of European Projects aimed at development the models for five properties relevant for chemical management and to provide the users all details necessary to fulfill the requirements of Five principles of validation. In this presentation we will show the example of QSAR modelling of chemical carcinogens in accordance with principles of validation adopted by OECD in scope of European project CAESAR using Counter Propagation Neural Networks (CP ANN). Statistical performance of different models will be discussed.

• Freiburg University, FDM (Center for data analysis and modelling), Freiburg, Germany
• In Silico toxicology, Freiburg, Germany

In this study we compared two different models for the prediction of biological activities of chemicals (e.g. IRIS and CPDB carcinogenicity). They are designed for diverse training sets and use a robust instance-based approach with linear fragments as chemical descriptors and activity specific chemical similarities. A confidence value is associated with each prediction to indicate the applicability domain. A support vector regression model and a similarity weighted multilinear model with prior objective feature selection and principal components analysis are trained for each prediction on a subset of the training instances similar to the current query structure (neighbors). The applicability domain for each prediction is assessed using gaussian smoothed chemical similarity of the training set. Leave-one-out crossvalidation has been performed for several recent databases obtained from the DSSTOX project including quantitative carcinogenicity estimates. The results indicate a comparable performance for both models with precision of over 80% (within 1 log unit error) and squared correlation values between 64 and 85% depending on the cutoff value for the applicability domain.

• Lhasa Ltd, Leeds, United Kingdom

Derek for Windows is a knowledge-based expert system designed to predict the toxicity of a chemical from its structure. The knowledge base is composed of alerts, example compounds and rules each of which contributes to predictions made by the system. Recently, a new feature has been introduced into Derek for Windows which allows validation data for an alert to be displayed when it is activated by a query compound. Such data assist the user in understanding the reliability of an alert and contribute to compliance with the OECD Principles for (Q)SAR validation. The first step in the development of this feature has been the addition of validation data for mutagenicity. Currently, there are 88 alerts describing Ames test mutagenicity in the knowledge base and the predictive performance of each has been assessed using four data sets, including a collection of proprietary Ames test data arising from a data sharing initiative. Further work is in progress to extend the feature to alerts for chromosome damage and carcinogenicity, alongside additional analyses of mutagenicity alerts using proprietary data held in-house by Derek for Windows member organisations.

• Istituto di Ricerche Famacologiche Mario Negri, Milan, Italy
• BioChemics Consulting SAS, Olivet, France

It has often been proposed that models derived for assessing mutagenicity can be used to model carcinogenicity as well. Usually this implies a lower accuracy for carcinogenicity due to the fact that only genotoxic carcinogens - directly binding to DNA to exhibit their activity - can be detected. Moreover this assumption does not allow dealing with epigenetic carcinogens. Here we propose a possible scheme to overcome this drawback consisting in a hierarchical application of two classification trees: the first one for mutagenicity the latter working to better detect non mutagenic carcinogens. The dataset investigated consists of 504 compounds extracted from the CPDBAS database and provided with binary activity classes for both mutagenicity and carcinogeniticy. We firstly built a classification tree trained to classify between mutagens and not mutagens, using as training set 449 molecules with defined mutagenicity activity class. A second tree for carcinogenicity has been then developed using as training set the non mutagens compounds. The variables used to develop the trees are mainly fragment counts extracted by literature or computed with DRAGON software. The underlying idea was to couple these two trees in order to make an initial selection using the mutagenicity tree and then to use the carcinogenicity tree to classify the compounds marked as non mutagenic by the first tree. The performances of the entire architecture have been evaluated and compare with that of the two separated trees. The models were also used to predict an external test set of compounds.

• Istituto di Ricerche Famacologiche Mario Negri, Milan, Italy
• Institute of Geology and Geophysics, Tashkent, Uzbekistan

Elements of SMILES are one-symbol and two-symbols fragments. Examples of the two-symbols fragments are 'Cl', 'Br', 'N⁺', 'O^-', etc.; in other words they are some undivided semantic units of the SMILES strings. By means of calculation (by Monte Carlo method) of the correlation weights for the elements and for their combinations, models for the carcinogenicity and mutagenic potentials have been obtained. Statistical characteristics of these models are robust. Taking into account relative prevalence of the SMILES elements and their combinations in the training and test set one can formulate transparent criterions for applicability domain of these models. The main principle of the definition is the balance of the prevalence (probability) of each the SMILES elements and a combination of the SMILES elements in the training and test sets. If the prevalence of the SMILES attribute is considerable higher for the training set, it can leads to overtraining. If the prevalence of the SMILES attribute is considerable higher for the test set, it can leads to absence of the enough information for adequate model.

• Liverpool John Moores University, School of Pharmacy and Chemistry, Liverpool, United Kingdom

Aliphatic α,β-unsaturated compounds and chloronitrobenzenes are able to interact with proteins, enzymes and DNA through various mechanisms, and are able to stimulate a range of environmental toxicities and adverse health effects. In this study, groups of aliphatic α,β-unsaturated compounds and chloronitrobenzenes were investigated for their reactivity and toxicity for different endpoints. The whole set of α,β-unsaturated compounds, which are thought to act by the same mechanism of action (Michael-type addition), can be classified into two sub-groups according to their reactivity and mutagenicity, such that within each sub-group all compounds are similarly reactive. A group of chloronitrobenzenes, known to react via nucleophilic aromatic substitution (SNAr), with defined structures and data for different endpoints were analysed. Their activity across endpoints can be associated with different modes and molecular mechanisms of action. The results indicate that a category can be formed that allows structural information and compounds¿ reactivity and toxicity to be elucidated. The funding of the EU FP6 InSilicoTox Marie Curie Project (MTKD-CT-2006-42328) is gratefully acknowledged.

• Liverpool John Moores University, School of Pharmacy and Chemistry, Liverpool, United Kingdom

Aliphatic α,β-unsaturated carbonyl compounds are an important group of industrial chemicals. They are able to interact with proteins, enzymes and DNA through various mechanisms, and are able to stimulate a range of environmental toxicities and adverse health effects. In this study, a "category" of α,β-unsaturated carbonyl compounds with mutagenicity data for Salmonella typhimurium (strain TA100) was modelled. These compounds (including aldehydes and ketones) are thought to act by the same mechanism of action, namely Michael-type addition. The results of the modelling of the toxicity of the α,β-unsaturated carbonyl domain, currently limited to models for narrow sub-domains, were compared to acute aquatic toxicity to Tetrahymena pyriformis. In these two toxicities (mutagenicity and acute aquatic toxicity) the chemicals of the α,β-unsaturated carbonyl domain were shown to exhibit different reactivity. The whole set of compounds for mutagenicity examined in this study can be classified into sub-groups, such that within each sub-group all compounds are similarly reactive. The funding of the EU FP6 InSilicoTox Marie Curie Project (MTKD-CT-2006-42328) is gratefully acknowledged.

• Liverpool John Moores University, School of Pharmacy and Chemistry, Liverpool, United Kingdom

A chemical¿s electrophilic potential to react with DNA and proteins may be responsible for many mutagenic and skin sensitising responses via a number of differing mechanisms of action. An understanding of these mechanisms has resulted in the development of two sets of chemical rules aimed at identifying potentially mutagenic and skin sensitising compounds. This study aimed to assess the overlap between the two sets of chemical rules with the intention being to identify similarities and differences between the mechanisms for mutagenicity and skin sensitisation. Identification of similarities and differences has the potential to allow for mechanistic chemical category formation which may apply across endpoints, or be specific to a particular endpoint. This in turn could lead to read across predictions being made which may allow for a reduction in the need for animal testing. The funding of the EU FP6 OSIRIS Integrated Project (GOCE-037017-OSIRIS), the EU FP6 CAESAR Specific Targeted Project (SSPI-022674-CAESAR) and EU FP6 InSilicoTox Marie Curie Project (MTKD-CT-2006-42328) is gratefully acknowledged.

• DEI, Politecnico di Milano

Mutagenicity prediction has a long tradition in the QSAR world. Besides various indications about the role of specific substructures, for regulatory purposes it is important to obtain a good classification also on chemical families not well studied and developed. In this area it is important to develop methods that make use of the statistical analysis on great numbers, and that can be further refined using cooperative models (for instance local models of given classes) to improve the results, confirm the value of the prediction, give more insights into the domain.
Continuing on the road of some recent papers that have introduced good predictive models using novel classification algorithms as Support Vector Machines, we defined the first step of our modelling task. We considered a large data set of more than 4000 chemicals with about 1000 molecular descriptors calculated by commercial softwares, carefully checked in the CAESAR project of EU, and the corresponding Ames test results.
For this data set a stratified subdivision in train(80%) and test(20%) set was also suggested.
In our approach we set a target performance-rate, in terms of percentage accuracy, of about 85%, which is the reliability of the Ames test itself.
We devised an approach based on various data mining tools, most of them provided in WEKA 3.5.7: starting from the entire set of descriptors, we tried mathematical and statistical methods to transform and reduce the descriptors space while maintaining a good predictive performance.
Our results have reached the expectations: we produced a few models, of decreasing complexity, all beyond the threshold of 83% of correctly classified instances in stratified 10-fold cross-validation.
Whereas, with the given training/test set split for the hold out procedure, the best of our models reached an accuracy on the test set near to 85%.
These rates give evidence of the robustness and the prediction ability of our models.