Validation and subsequent development of the DEREK skin sensitization rulebase by analysis of the BgVV list of contact allergens

The DEREK knowledge-based computer system contains a subset of approximately 50 rules describing chemical substructures (toxophores) responsible for skin sensitization. This rulebase, based originally on Unilever historical in-house guinea pig maximization test data, has been subject to extensive validation and is undergoing refinement as the next stage of its development. As part of an ongoing program of validation and testing, the predictive ability of the sensitization rule set has been assessed by processing the structures of the 84 chemical substances in the list of contact allergens issued by the BgVV (German Federal Institute for Health Protection of Consumers). This list of chemicals is important because the biological data for each of the chemicals have been carefully scrutinized and peer reviewed, a key consideration in an area of toxicology in which much unreliable and potentially misleading data have been published. The existing DEREK rulebase for skin sensitization identified toxophores for skin sensitization in the structures of 71 out of the 84 chemicals (85%). The exercise highlighted areas of chemistry where further development of the rulebase was required, either by extension of the scope of existing rules or by generation of new rules where a sound mechanistic rationale for the biological activity could be established. Chemicals likely to be acting as photoallergens were identified, and new rules for photoallergenicity have subsequently been written. At the end of the exercise, the refined rulebase was able to identify toxophores for skin sensitization for 82 of the 84 chemicals in the BgVV list.     

Theoretical studies to estimate the skin sensitization potential of chemicals of the Schiff base domain

Skin sensitization occurs when an exogenous chemical substance forms a covalent adduct with a dermal protein electrophile or nucleophile. This instigates an immune response which leads to inflammation. The local lymph node assay is an in vivo model used in the assessment of relative skin sensitizing potency of chemicals. The method is time consuming and expensive, as well as poses ethical questions given that a number of mice must be sacrificed for each compound assessed. In this work, we investigate the use of an inexpensive, rapid, and ethical method to predict the skin sensitization potential of Schiff base chemicals. We employ quantum chemical methods to rationalize the sensitization potential of 22 compounds with a diverse range of activities. To this end, we have evaluated the mechanistic profile associated with this type of reaction using gas-phase models. We subsequently use the predicted rate determining barriers and key physico-chemical parameters (such as logP) to establish stucture activity relationship (SAR) guidelines to predict the skin sensitization potential for new chemicals. We find that the predicted rate determining barriers for aldehydes, ketone, and 1,2 and 1,3 diones generally decrease in the given order, which concurs with the overall trends in sensitization. We find that lipophilicity also plays a role, with those chemicals displaying both low barriers to reaction, and lower lipophilicity (ie, diones), being more likely to display undesirable skin sensitization effects. These findings are in line with experiment-based observations in the literature and point to the value 3D quantum chemical calculations could have if combined with other orthogonal approaches to estimate skin sensitization potential of chemicals.     

Evaluation of in silico tools to predict the skin sensitization potential of chemicals

Public domain and commercial in silico tools were compared for their performance in predicting the skin sensitization potential of chemicals. The packages were either statistical based (Vega, CASE Ultra) or rule based (OECD Toolbox, Toxtree, Derek Nexus). In practice, several of these in silico tools are used in gap filling and read-across, but here their use was limited to make predictions based on presence/absence of structural features associated to sensitization. The top 400 ranking substances of the ATSDR 2011 Priority List of Hazardous Substances were selected as a starting point. Experimental information was identified for 160 chemically diverse substances (82 positive and 78 negative). The prediction for skin sensitization potential was compared with the experimental data. Rule-based tools perform slightly better, with accuracies ranging from 0.6 (OECD Toolbox) to 0.78 (Derek Nexus), compared with statistical tools that had accuracies ranging from 0.48 (Vega) to 0.73 (CASE Ultra – LLNA weak model). Combining models increased the performance, with positive and negative predictive values up to 80% and 84%, respectively. However, the number of substances that were predicted positive or negative for skin sensitization in both models was low. Adding more substances to the dataset will increase the confidence in the conclusions reached. The insights obtained in this evaluation are incorporated in a web database www.asopus.weebly.com that provides a potential end user context for the scope and performance of different in silico tools with respect to a common dataset of curated skin sensitization data.     

Development of an expert system rulebase for the prospective identification of photoallergens

Relationships between the structure and properties of chemicals can be programmed into knowledge-based systems such as DEREK (an acronym for 'Deductive Estimation of Risk from Existing Knowledge'). The DEREK knowledge-based computer system contains a sub-set of over 50 rules describing chemical substructures (toxophores) responsible for skin sensitization. This rulebase, based originally on Unilever historical in-house guinea pig maximisation test data, is largely complete and is undergoing refinement as the next stage of its development. As part of an ongoing program of validation and testing, the predictive ability of the sensitization rule set was assessed by processing the structures of over 100 chemical substances in the list of contact allergens identified by the BgVV (German Federal Institute for Health Protection of Consumers). The exercise highlighted areas of chemistry where further development of the rulebase was required, either by extension of the scope of existing rules or by generation of new rules where a sound mechanistic rationale for the biological activity could be established. Several chemicals likely to be acting as photoallergens were identified and rules for photoallergenicity were written covering three classes of chemicals. This paper describes work to extend the DEREK rules for photoallergenicity as part of the European Phototox Project.     

A fast clustering algorithm for analyzing highly similar compounds of very large libraries

As a result of the recent developments of high-throughput screening in drug discovery, the number of available screening compounds has been growing rapidly. Chemical vendors provide millions of compounds; however, these compounds are highly redundant. Clustering analysis, a technique that groups similar compounds into families, can be used to analyze such redundancy. Many available clustering methods focus on accurate classification of compounds; they are slow and are not suitable for very large compound libraries. Here is described a fast clustering method based on an incremental clustering algorithm and the 2D fingerprints of compounds. This method can cluster a very large data set with millions of compounds in hours on a single computer. A program implemented with this method, called cd-hit-fp, is available from http://chemspace.org.     

Random forests for functional covariates

We propose a form of random forests that is especially suited for functional covariates. The method is based on partitioning the functions' domain in intervals and using the functions' mean values across those intervals as predictors in regression or classification trees. This approach appears to be more intuitive to applied researchers than usual methods for functional data, while also performing very well in terms of prediction accuracy. The intervals are obtained from randomly drawn, exponentially distributed waiting times. We apply our method to data from Raman spectra on boar meat as well as near‐infrared absorption spectra. The predictive performance of the proposed functional random forests is compared with commonly used parametric and nonparametric functional methods and with a nonfunctional random forest using the single measurements of the curve as covariates. Further, we present a functional variable importance measure, yielding information about the relevance of the different parts of the predictor curves. Our variable importance curve is much smoother and hence easier to interpret than the one obtained from nonfunctional random forests.     

QSAR classification models for the screening of the endocrine-disrupting activity of perfluorinated compounds

Perfluorinated compounds (PFCs) are a class of emerging pollutants still widely used in different materials as non-adhesives, waterproof fabrics, fire-fighting foams, etc. Their toxic effects include potential for endocrine-disrupting activity, but the amount of experimental data available for these pollutants is limited. The use of predictive strategies such as quantitative structure-activity relationships (QSARs) is recommended under the REACH regulation, to fill data gaps and to screen and prioritize chemicals for further experimentation, with a consequent reduction of costs and number of tested animals. In this study, local classification models for PFCs were developed to predict their T4-TTR (thyroxin-transthyretin) competing potency. The best models were selected by maximizing the sensitivity and external predictive ability. These models, characterized by robustness, good predictive power and a defined applicability domain, were applied to predict the activity of 33 other PFCs of environmental concern. Finally, classification models recently published by our research group for T4-TTR binding of brominated flame retardants and for estrogenic and anti-androgenic activity were applied to the studied perfluorinated chemicals to compare results and to further evaluate the potential for these PFCs to cause endocrine disruption.     

A comparison of some hierarchal agglomerative clustering algorithms for structure—property correlation

Six different hierarchal clustering algorithms were used to cluster eleven sets of compounds for which associated property data were available. The effectiveness of the clustering in each case was assessed by inspection of the resulting tree diagram representing the classification and by the utility of the classification for molecular property prediction. The best predictions were generally obtained by using the algorithm proposed by Ward while the widely-used nearest neighbour algorithm performed very badly. Unstandardised data gave better results than standardised data although the difference was less when structurally related, rather than structurally disparate, groups of compounds were tested.     

Evaluation of SARs for the prediction of skin irritation/corrosion potential: structural inclusion rules in the BfR decision support system

The German Federal Institute for Risk Assessment (BfR) has developed a Decision Support System (DSS) to assess certain hazardous properties of pure chemicals, including skin and eye irritation/corrosion. The BfR-DSS is a rule-based system that could be used for the regulatory classification of chemicals in the European Union. The system is based on the combined use of two predictive approaches: exclusion rules based on physicochemical cut-off values to identify chemicals that do not exhibit a certain hazard (e.g., skin irritation/corrosion), and inclusion rules based on structural alerts to identify chemicals that do show a particular toxic potential. The aim of the present study was to evaluate the structural inclusion rules implemented in the BfR-DSS for the prediction of skin irritation and corrosion. The following assessments were performed: (a) a confirmation of the structural rules by rederiving them from the original training set (1358 substances), and (b) an external validation by using a test set of 200 chemicals not used in the derivation of the rules. It was found as a result that the test data set did not match the training set relative to the inclusion of structural alerts associated with skin irritation/corrosion, albeit some skin irritants were in the test set.     

Evaluation of SARs for the prediction of skin irritation/corrosion potential–structural inclusion rules in the BfR decision support system†

The German Federal Institute for Risk Assessment (BfR) has developed a Decision Support System (DSS) to assess certain hazardous properties of pure chemicals, including skin and eye irritation/corrosion. The BfR–DSS is a rule-based system that could be used for the regulatory classification of chemicals in the European Union. The system is based on the combined use of two predictive approaches: exclusion rules based on physicochemical cut-off values to identify chemicals that do not exhibit a certain hazard (e.g., skin irritation/corrosion), and inclusion rules based on structural alerts to identify chemicals that do show a particular toxic potential. The aim of the present study was to evaluate the structural inclusion rules implemented in the BfR–DSS for the prediction of skin irritation and corrosion. The following assessments were performed: (a) a confirmation of the structural rules by rederiving them from the original training set (1358 substances), and (b) an external validation by using a test set of 200 chemicals not used in the derivation of the rules. It was found as a result that the test data set did not match the training set relative to the inclusion of structural alerts associated with skin irritation/corrosion, albeit some skin irritants were in the test set.     

New Models to Predict the Acute and Chronic Toxicities of Representative Species of the Main Trophic Levels of Aquatic Environments

To assess the impact of chemicals on an aquatic environment, toxicological data for three trophic levels are needed to address the chronic and acute toxicities. The use of non-testing methods, such as predictive computational models, was proposed to avoid or reduce the need for animal models and speed up the process when there are many substances to be tested. We developed predictive models for Raphidocelis subcapitata, Daphnia magna, and fish for acute and chronic toxicities. The random forest machine learning approach gave the best results. The models gave good statistical quality for all endpoints. These models are freely available for use as individual models in the VEGA platform and for prioritization in JANUS software.     

Local and global quantitative structure-activity relationship modeling and prediction for the baseline toxicity

The predictive accuracy of the model is of the most concern for computational chemists in quantitative structure-activity relationship (QSAR) investigations. It is hypothesized that the model based on analogical chemicals will exhibit better predictive performance than that derived from diverse compounds. This paper develops a novel scheme called "clustering first, and then modeling" to build local QSAR models for the subsets resulted from clustering of the training set according to structural similarity. For validation and prediction, the validation set and test set were first classified into the corresponding subsets just as those of the training set, and then the prediction was performed by the relevant local model for each subset. This approach was validated on two independent data sets by local modeling and prediction of the baseline toxicity for the fathead minnow. In this process, hierarchical clustering was employed for cluster analysis, k-nearest neighbor for classification, and partial least squares for the model generation. The statistical results indicated that the predictive performances of the local models based on the subsets were much superior to those of the global model based on the whole training set, which was consistent with the hypothesis. This approach proposed here is promising for extension to QSAR modeling for various physicochemical properties, biological activities, and toxicities.     

A random forest of combined features in the classification of cut tobacco based on gas chromatography fingerprinting

We applied the random forest method to discriminate among different kinds of cut tobacco. To overcome the influence of the descending resolution caused by column pollution and the subsequent deterioration of column efficacy at different testing times, we constructed combined peaks by summing the peaks over a specific elution time interval Δt. On constructing tree classifiers, both the original peaks and the combined peaks were considered. A data set of 75 samples from three grades of the same tobacco brand was used to evaluate our method. Two parameters of the random forest were optimized using out-of-bag error, and the relationship between Δt and classification rate was investigated. Experiments show that partial least squares discriminant analysis was not suitable because of the overfitting, and the random forest with the combined features performed more accurately than Naïve Bayes, support vector machines, bootstrap aggregating and the random forest using only its original features.     

Computer‐aided prediction of toxicity with substructure pattern and random forest

Toxicity of chemicals induced by different factors is an important consideration, especially during the drug research and development process. Thus, there is urgent need to develop computationally effective models that can predict the toxicity or adverse effects of chemicals for a specific class of chemicals. In this study, random forest (RF) was used to classify five toxicity data sets from Distributed Structure‐Searchable Toxicity database network, using substructure fingerprints calculated directly from simple molecular structure. Three model validation approaches, out‐of‐bag validation incorporated in RF, fivefold cross‐validation, and an independent validation set, were used for assessing the prediction capability of our models. The chemical space analysis of data sets was explored by multidimensional scaling plots, and outlying molecules were also detected by the proximity measure in RF. At the same time, the important substructure fingerprints, recognized by the RF technique, gave some insights into the structure features related to toxicity of chemicals. The results obtained showed that these in silico classification models with substructure patterns and RF are applicable for potential toxicity prediction of chemical compounds. Copyright © 2012 John Wiley & Sons, Ltd.