Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Development of a QSAR model to predict respiratory irritation by individual constituents

: Wehr, Matthias; Karwath, A.; Sarang, S.S.; Rooseboom, M.; Boogaard, P.J.; Escher, Sylvia E.

The Toxicologist 168 (2019), No.1, Abstract PS 1854
ISSN: 0731-9193
Society of Toxicology (Annual Meeting) <58, 2019, Baltimore/Md.>
Fraunhofer ITEM ()

The RespiraTox project, funded by NC3R CrackIT, develops a QSAR model, which predicts the potential of individual compounds to cause irritation in the respiratory tract. We distinguished two mode of actions, i) “sensory irritation”, characterized by a decrease in breathing rates, and ii) “tissue irritation” characterized by primarily histopathological findings. QSAR models rely on high quality datasets. We based the classification “irritating to respiratory tract” on several data types from in vivo studies with inhalation exposure. In a tiered approach, we considered information from i) studies with acute exposure from the ECHA CHEM database (DB), ii) the Hazardous Substance DB, iii) the harmonized classification and labelling inventory from ECHA, and iv) repeated dose studies from the Fraunhofer RepDose DB. For later stage validation, we withhold human data from Fraunhofer Breath DB. The final dataset includes about 2500 irritating and 800 non-irritating compounds. Prior to model development, the CAS numbers and compound structures were quality controlled and corrected. Two kinds of information were generated from the compounds structures: i) structural descriptors (ECFPS), and ii) physico-chemical properties. We explored several machine learning algorithms including Logistic Regression (LR), Random Forests (RF), and Gradient Boosted Decision Trees (BT) to derive a classification model. The internal validation procedure employed stratified k-fold cross-validation (k=5). The overall approach adheres to the five OECD principles. The criteria used to measure performance of a given model is the Area Under ROC-Curve (AUC). The AUC for LR using the combined feature set is 0.71. The optimal performance for both RF and BT is 0.78. The applicability domain is determined by features with highest impact on the final model. The current approach will be further refined and improved (e.g. by differentiating sensory and tissue irritation). The final model will be provided online as user-friendly interface to promote its use by toxicologists, regulators, and overall to reduce the testing of animals.