Consensus prediction of skin permeability and sensitization using cheminformatics approaches

Abstract

We have compiled, curated, and integrated the largest publicly-accessible datasets of human skin sensitization (381 compounds) and skin permeability (186 and 96 compounds for human and rodents, respectively). For both end points, we have built rigorously validated QSAR models using Random Forest with SiRMS and Dragon molecular descriptors. These models showed reasonable (53-62%) coverage and external predictive power as high as CCRext = 80% for skin sensitization and Q2ext= 0.87-0.89 for human and rodent skin permeability; models are available on the Chembench webportal (http://chembench.mml.unc.edu/). Virtual screening of the ToxCast chemical library using selected QSAR models identified 56 compounds as potential skin sensitizers. Furthermore, we have explored a hypothesis that skin permeability and sensitization could be related. The concordance analysis showed no significant global correlation; however, clustering chemicals by structural similarity revealed that the correlation does exist for certain chemical classes suggesting possible mechanistic link between skin permeability and sensitization.