DOI: 10.1002/minf.201300033 Subchronic Oral and Inhalation Toxicities: a Challenging Attempt for Modeling and Prediction Dimaitar A. Dobchev,* [a, b] Indrek Tulp, [c] Gunnar Karelson, [a, b] Tarmo Tamm, [a, d] Kaido Tämm, [a, c] and Mati Karelson [b, c] 1 Introduction Chronic toxicity is a consequence of the persistent or pro- gressively deteriorating dysfunction of cells, organs or mul- tiple organ systems, resulting from long-term exposure to a chemical. Of relevance for cosmetic products are the oral, dermal and inhalation subacute (28 days) and subchronic (90 days) repeated dose studies in rodents. The 28-day or 90-day oral toxicity tests in rodents are the most commonly used long-term toxicity tests. The highest dose adminis- tered is designed to cause some toxicity, but not lethality. In the notification process of dangerous substances, long-term toxicity studies are required when the substance under consideration is produced or imported in amounts exceeding 1 ton/year. [1] In the case of the development of cosmetic ingredients that have specific biological proper- ties and which will come into contact with human skin for a long period of time, evaluation of the systemic risk is a key element in evaluating the safety of these new ingre- dients, irrespective of the tonnage-linked and possibly re- stricted requirements imposed by the Dangerous Substan- ces Directive. [2] Therefore, in certain cases the use of animal long-term experiments to study one or more potential toxic effects remains to be required by law. However, it is essential that the aim of abolishing animal experiments for testing cosmetic products be pursued and that the prohibi- tion of such experiments becomes effective in the territory of the EU Member State. The 7th Amendment to the Cos- metic Directive 76/768/EEC allows 10 years from the entry into force date on, to come up with validated alternative tests for repeated exposure. [3] Currently, one of the alterna- tive methods which are scientifically adopted/validated by the Organization for Economic Cooperation and Develop- ment [4,5] and applicable to the large part of the chemical sector is Quantitative-Structure Property/Activity Relation- ship (QSPR/QSAR). [6] One of the main advantages of this modeling technique is that it does not require direct labo- ratory experiments on animals but rather already available data on such experiments. A developed QSAR model can be used to predict untested chemicals for their subchronic toxicity without the need of new animal tests. According to OECD guideline 413 [7] the subchronic inha- lation toxicity studies are primarily used to derive regulato- ry concentrations for assessing worker risk in occupation settings. They are also used to assess human residential, transportation, and environmental risk. Such a guideline would enable the characterization of adverse effects follow- [a] D.A. Dobchev, G. Karelson, T. Tamm, K. Tämm MolCode, Ltd. Turu 2, Tartu 51013, Estonia *e-mail: dimitar@molcode.com [b] D. A. Dobchev, G. Karelson, M. Karelson Department of Chemistry, Tallinn University of Technology Akadeemia tee 15, Tallinn 19086, Estonia phone/fax: + 372 6 202 814/ + 372 6 202 819 [c] I. Tulp, K. Tämm, M. Karelson Department of Chemistry, University of Tartu Ravila 14a, Tartu 50411, Estonia [d] T. Tamm Institute of Technology, University of Tartu Nooruse 1, Tartu 50411, Estonia Supporting Information for this article is available on the WWW under http://dx.doi.org/10.1002/minf.201300033. Abstract : The article deals with a challenging attempt to model and predict “difficult” properties as long-term sub- chronic oral and inhalation toxicities (90 days) using nonlin- ear QSAR approach. This investigation is one of the first to tackle such multicomplex properties where we have em- ployed nonlinear models based on artificial neural network for the prediction of NOAEL (no observable adverse effect level). Despite the complex nature of the NOAEL property based on in vivo rat experiments, the successful models can be used as alternative tools to non-animal tests for the initial assessment of these chronic toxicities. The model for oral subchronic toxicity is able to describe 88 %, and the in- halation model 87 % of the statistical variance. For the sake of future predictions, we have also defined in a quantitative way the applicability domain of all neural network models. Keywords: Subchronic oral toxicity · Subchronic inhalation toxicity · Artificial neural network · QSAR · NOAEL Mol. Inf. 2013, 32, 793 – 801  2013 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim 793