CON4EI: Evaluation of QSAR models for hazard identification and labelling of eye irritating chemicals

For three computational models (Toxtree, and Case Ultra EYE_DRAIZE and EYE_IRR) performance parameters were calculated. Coverage ranged from 15 to 58%. Coverage was 2 to 3.4 times higher for liquids than for solids. The lowest number of false positives (5%) was reached with EYE_IRR; this model however also gave a high number of false negatives (46%). The lowest number of false negatives (25%) was seen with Toxtree; for liquids Toxtree predicted the lowest number of false negatives (11%), for solids EYE_DRAIZE did (17%). It can be concluded that the training sets should be enlarged with high quality data. The tested models are not yet sufficiently powerful for stand-alone evaluations, but that they can surely become of value in an integrated weight-of-evidence approach in hazard assessment.