Quantitative structure-activity relationships (QSARs) for skin corrosivity of organic acids, bases and phenols: Principal components and neural network analysis of extended datasets

Quantitative structure-activity relationships (QSARs) relating skin corrosivity data of organic acids, bases and phenols to their log(octanol/water partition coefficient), molecular volume, melting point and pK(a) have been extended to substantially larger datasets. In addition to principal components analysis, as used in earlier work, the datasets have also been analysed using neural networks. Plots of the first two principal components of the four independent variables, which broadly model skin permeability and cytotoxicity, for each of the extended datasets confirmed that the analysis was able to discriminate well between corrosive and non-corrosive chemicals. Neural networks using the same parameters as inputs, were trained to an output in the range 0.0 to 1.0, with non-corrosive chemicals being assigned the value 0 and corrosive chemicals the value 1. As well as yielding classification predictions in agreement with those in the training sets, predicted outputs in the 0 to 1 range gave a useful indication of the confidence of the predicted classification. These QSARs are useful (a) for the prediction of the skin corrosivity potentials of new or untested chemicals and (b) for determining the confidence of predictions in regions of 'biological uncertainty' which exist at the classification threshold between corrosive and non-corrosive chemicals.