You are here:
QSAR Model Development for Skin Sensitization
Citation:
Vegosen, L. AND T. Martin. QSAR Model Development for Skin Sensitization. QSAR 2021 International Workshop on QSAR in Environmental and Health Sciences, Virtual, NC, June 07 - 10, 2021. https://doi.org/10.23645/epacomptox.15070311
Impact/Purpose:
Presentation to the QSAR 2021 International Workshop on QSAR in Environmental and Health Sciences June 2021. This presentation describes development of a variety of QSAR models for skin sensitization using publicly available toxicity data. Estimated toxicity values are useful in applications such as chemical prioritization under TSCA. The best QSAR approaches will be added to EPA tools such as T.E.S.T. (Toxicity Estimation Software Tool).
Description:
This study develops QSAR models for predicting skin sensitization as a binary toxicity endpoint, assesses the performance of different modeling methods and descriptor sets, develops consensus models, and assesses methods for defining applicability domain (AD). A local lymph node assay (LLNA) dataset of 1355 chemicals was compiled from the NICEATM LLNA Database, OECD QSAR Toolbox, and eChemPortal. Records were mapped to unique substances in the U.S. Environmental Protection Agency’s (EPA) Distributed Structure-Searchable Toxicity (DSSTox) Database. Using 10 different modeling methods, models were developed in the Online Chemical Database with Modeling Environment (OCHEM). Models were built using 25 descriptor sets available in OCHEM and two additional descriptor sets: PaDEL descriptors and descriptors developed for EPA’s Toxicity Estimation Software Tool (T.E.S.T.). Consensus models were developed using the best performing models in OCHEM. The best-performing modeling methods, including Associative Neural Networks, Support Vector Machines, and WEKA-random forest, produced validation set balanced accuracies of 77%. T.E.S.T. and PaDEL descriptors performed comparably to the best-performing descriptor sets in OCHEM. Consensus models, achieving balanced accuracies of 79%, performed better than individual models. For the binary skin sensitization endpoint, the applicability domains in OCHEM generally did not improve the results (considering the tradeoff between balanced accuracy and prediction coverage) so other methods for defining AD should be explored. Models will be made publicly available in OCHEM and T.E.S.T., which will contribute to improving the performance and acces
URLs/Downloads:
DOI: QSAR Model Development for Skin Sensitization
QSAR2021_VEGOSEN.PDF (PDF, NA pp, 321.118 KB, about PDF)