Anna Lombardo

Assessment and validation of the CAESAR predictive model for bioconcentration factor (BCF) in fish

BackgroundBioconcentration factor (BCF) describes the behaviour of a chemical in terms of its likelihood of concentrating in organisms in the environment. It is a fundamental property in recent regulations, such as the European Community Regulation on chemicals and their safe use or the Globally Harmonized System for classification, labelling and packaging. These new regulations consider the possibility of reducing or waiving animal tests using alternative methods, such as in silico methods. This study assessed and validated the CAESAR predictive model for BCF in fish.ResultsTo validate the model, new experimental data were collected and used to create an external set, as a second validation set (a first validation exercise had been done just after model development). The performance of the model was compared with BCFBAF v3.00. For continuous values and for classification purposes the CAESAR BCF model gave better results than BCFBAF v3.00 for the chemicals in the applicability domain of the model. R2 and Q2 were good and accuracy in classification higher than 90%. Applying an offset of 0.5 to the compounds predicted with BCF close to the thresholds, the number of false negatives (the most dangerous errors) dropped considerably (less than 0.6% of chemicals).ConclusionsThe CAESAR model for BCF is useful for regulatory purposes because it is robust, reliable and predictive. It is also fully transparent and documented and has a well-defined applicability domain, as required by REACH. The model is freely available on the CAESAR web site and easy to use. The reliability of the model reporting the six most similar compounds found in the CAESAR dataset, and their experimental and predicted values, can be evaluated.

CORAL: Building up the model for bioconcentration factor and defining it's applicability domain

CORAL (CORrelation And Logic) software can be used to build up the quantitative structure--property/activity relationships (QSPR/QSAR) with optimal descriptors calculated with the simplified molecular input line entry system (SMILES). We used CORAL to evaluate the applicability domain of the QSAR models, taking a model of bioconcentration factor (logBCF) as example. This models based on a large training set of more than 1000 chemicals. To improve the model is predictivity and reliability on new compounds, we introduced a new function, which uses the Delta(obs) = logBCF(expr)--logBCF(calc) of the predictions on the chemicals in the training set. With this approach, outliers are eliminated from the phase of training. This proved useful and increased the models predictivity.

A new bioconcentration factor model based on SMILES and indices of presence of atoms.

Indices of the presence of atoms (IPA) encode the presence or absence of atoms, such as nitrogen, oxygen, sulphur, phosphorus, fluorine, chlorine, and bromine in a molecule. They are calculated with the simplified molecular input line entry system (SMILES). Using the Monte Carlo method for correlation weights of these indices, one can improve the predictive ability of optimal SMILES-based descriptors in quantitative structure-activity relationships (QSAR) for bioconcentration factor. The model without IPA gave the following results: n=503, r(2)=0.6803, q(2)=0.6781, s=0.759, F=1066 (subtraining set); n=322, r(2)=0.8181, r(pred)(2)=0.8159, s=0.565, F=1439 (calibration set); n=105, r(2)=0.6703, r(pred)(2)=0.6577, R(m)(2)=0.6628, s=0.728, F=209 (test set); n=106, r(2)=0.6624, r(pred)(2)=0.6502, R(m)(2)=0.6212, s=0.757, F=204 (validation set) The model with IPA gave: n=503, r(2)=0.7082, q(2)=0.7062, s=0.725, F=1216 (subtraining set); n=322, r(2)=0.8401, r(pred)(2)=0.8383, s=0.528, F=1682 (calibration set); n=105, r(2)=0.7489, r(pred)(2)=0.7402, R(m)(2)=0.7252, s=0.637, F=307 (test set); n=106, r(2)=0.7306, r(pred)(2)=0.7217, R(m)(2)=0.7010, s=0.680, F=282 (validation set).

In silico models for predicting ready biodegradability under REACH: A comparative study

REACH (Registration Evaluation Authorization and restriction of Chemicals) legislation is a new European law which aims to raise the human protection level and environmental health. Under REACH all chemicals manufactured or imported for more than one ton per year must be evaluated for their ready biodegradability. Ready biodegradability is also used as a screening test for persistent, bioaccumulative and toxic (PBT) substances. REACH encourages the use of non-testing methods such as QSAR (quantitative structure-activity relationship) models in order to save money and time and to reduce the number of animals used for scientific purposes. Some QSAR models are available for predicting ready biodegradability. We used a dataset of 722 compounds to test four models: VEGA, TOPKAT, BIOWIN 5 and 6 and START and compared their performance on the basis of the following parameters: accuracy, sensitivity, specificity and Matthews correlation coefficient (MCC). Performance was analyzed from different points of view. The first calculation was done on the whole dataset and VEGA and TOPKAT gave the best accuracy (88% and 87% respectively). Then we considered the compounds inside and outside the training set: BIOWIN 6 and 5 gave the best results for accuracy (81%) outside training set. Another analysis examined the applicability domain (AD). VEGA had the highest value for compounds inside the AD for all the parameters taken into account. Finally, compounds outside the training set and in the AD of the models were considered to assess predictive ability. VEGA gave the best accuracy results (99%) for this group of chemicals. Generally, START model gave poor results. Since BIOWIN, TOPKAT and VEGA models performed well, they may be used to predict ready biodegradability.

CORAL: QSAR models for acute toxicity in fathead minnow (Pimephales promelas).

CORrelation And Logic (CORAL) is a software that generates quantitative structure activity relationships (QSAR) for different endpoints. This study is dedicated to the QSAR analysis of acute toxicity in Fathead minnow (Pimephales promelas). Statistical quality for the external test set is a complex function of the split (into training and test subsets), the number of epochs of the Monte Carlo optimization, and the threshold that is a criterion for dividing the correlation weights into two classes rare (blocked) and not rare (active). Computational experiments with three random splits (data on 568 compounds) indicated that this approach can satisfactorily predict the desired endpoint (the negative decimal logarithm of the 50% lethal concentration, in mmol/L, pLC50). The average correlation coefficients (r2) are 0.675 ± 0.0053, 0.824 ± 0.0242, 0.787 ± 0.0101 for subtraining, calibration, and test set, respectively. The average standard errors of estimation (s) are 0.837 ± 0.021, 0.555 ± 0.047, 0.606 ± 0.049 for subtraining, calibration, and test set, respectively. The CORAL software together with three random splits into subtraining, calibration, and test sets can be downloaded on the Internet (http://www.insilico.eu/coral/).

A new in silico classification model for ready biodegradability, based on molecular fragments

Regulations such as the European REACH (Registration, Evaluation, Authorization and restriction of Chemicals) often require chemicals to be evaluated for ready biodegradability, to assess the potential risk for environmental and human health. Because not all chemicals can be tested, there is an increasing demand for tools for quick and inexpensive biodegradability screening, such as computer-based (in silico) theoretical models. We developed an in silico model starting from a dataset of 728 chemicals with ready biodegradability data (MITI-test Ministry of International Trade and Industry). We used the novel software SARpy to automatically extract, through a structural fragmentation process, a set of substructures statistically related to ready biodegradability. Then, we analysed these substructures in order to build some general rules. The model consists of a rule-set made up of the combination of the statistically relevant fragments and of the expert-based rules. The model gives good statistical performance with 92%, 82% and 76% accuracy on the training, test and external set respectively. These results are comparable with other in silico models like BIOWIN developed by the United States Environmental Protection Agency (EPA); moreover this new model includes an easily understandable explanation.

Using Toxicological Evidence from QSAR Models in Practice

Leading QSAR models provide supporting documentation in addition to a predicted toxicological value. Such information enables the toxicologist to explore the properties of chemical substances as well as to review and to increase the reliability of toxicity predictions. This article focuses on the use of this information in practice. We explore the supporting documentation provided by the EPISuite, T.E.S.T. and VEGA platforms when evaluating the bioconcentration factor (BCF) of three example compounds. Each compound presents a different challenge: to recognize high reliability, analyze complex evidence of reliability, and recognize uncertainty. In each case, we first describe and discuss the supporting documentation provided by the QSAR platforms. We then discuss the judgments on reliability across sectors from 28 toxicologists who used this supporting information and commented on the process. The article demonstrates both the use of QSAR models as tools to reduce or replace in vivo testing, and the need for scientific expertise and rigor in their use.

Predicting persistence in the sediment compartment with a new automatic software based on the k-Nearest Neighbor (k-NN) algorithm.

The ability of a substance to resist degradation and persist in the environment needs to be readily identified in order to protect the environment and human health. Many regulations require the assessment of persistence for substances commonly manufactured and marketed. Besides laboratory-based testing methods, in silico tools may be used to obtain a computational prediction of persistence. We present a new program to develop k-Nearest Neighbor (k-NN) models. The k-NN algorithm is a similarity-based approach that predicts the property of a substance in relation to the experimental data for its most similar compounds. We employed this software to identify persistence in the sediment compartment. Data on half-life (HL) in sediment were obtained from different sources and, after careful data pruning the final dataset, containing 297 organic compounds, was divided into four experimental classes. We developed several models giving satisfactory performances, considering that both the training and test set accuracy ranged between 0.90 and 0.96. We finally selected one model which will be made available in the near future in the freely available software platform VEGA. This model offers a valuable in silico tool that may be really useful for fast and inexpensive screening.

A k-NN algorithm for predicting oral sub-chronic toxicity in the rat

Repeated dose toxicity is of the utmost importance to characterize the toxicological profile of a chemical after repeated administration. Its evaluation refers to the Lowest-Observed-(Adverse)-Effect-Level (LO(A)EL) explicitly requested in several regulatory contexts, such as REACH and EC Regulation 1223/2009 on cosmetic products. So far in vivo tests have been the sole viable option to assess repeated dose toxicity. We report a customized k-Nearest Neighbors approach for predicting sub-chronic oral toxicity in rats. A training set of 254 chemicals was used to derive models whose robustness was challenged through leave-one-out cross-validation. Their predictive power was evaluated on an external dataset comprising 179 chemicals. Despite the intrinsically heterogeneous nature of the data, our models give promising results, with q²≥0.632 and external r²≥0.543. The confidence in prediction was ensured by implementing restrictive user-adjustable rules excluding suspicious chemicals irrespective of the goodness in their prediction. Comparison with the very few LO(A)EL predictive models in the literature indicates that the results of the present analysis can be valuable in prioritizing the safety assessment of chemicals and thus making safe decisions and justifying waiving animal tests according to current regulations concerning chemical safety.

Identification of structural alerts for liver and kidney toxicity using repeated dose toxicity data

BackgroundThe potential for a compound to cause hepatotoxicity and nephrotoxicity is a matter of extreme interest for human health risk assessment. To assess liver and kidney toxicity, repeated-dose toxicity (RDT) studies are conducted mainly on rodents. However, these tests are expensive, time-consuming and require large numbers of animals. For early toxicity screening, in silico models can be applied, reducing the costs, time and animals used. Among in silico approaches, structure–activity relationship (SAR) methods, based on the identification of chemical substructures (structural alerts, SAs) related to a particular activity (toxicity), are widely employed.ResultsWe identified and evaluated some SAs related to liver and kidney toxicity, using RDT data on rats taken from the hazard evaluation support system (HESS) database. We considered only SAs that gave the best percentages of true positives (TP).ConclusionsIt was not possible to assign an unambiguous mode of action for all the SAs, but a mechanistic explanation is provided for some of them. Such achievements may help in the early identification of liver and renal toxicity of substances.