We have developed classification QSAR models to predict chemical skin sensitization potential (non-sensitizers, weak, moderate, and strong/extreme sensitizers). We have (i) compiled, curated, and integrated the largest publicly-available dataset comprising 515 chemicals tested in the LLNA assay; (ii) used this data to generate and validate multi-class QSAR models; (iii) established SAR rules for skin sensitizers; and (iv) employed QSAR models for virtual screening of the COSMOS database of cosmetics inventory. Model developed with SiRMS and PASS/QNA descriptors using random forest and deep learning techniques showed total average accuracy of 63% and coverage of 59% as applied to several external validation sets. Virtual screening of the COSMOS inventory yielded 250 putative strong skin sensitizers. Structural fragments promoting or interfering with skin sensitization potential as well as structural modifications required to decrease chemical toxicity were identified. Models developed herein could be used to guide the rational design of safe cosmetic products.

Development and application of multiclass QSAR models for predicting human skin sensitization

251st National Meeting (2016)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.