Artificial neural network analysis of data from multiple in vitro assays for prediction of skin sensitization potency of chemicals.

In order to develop in vitro risk assessment systems for skin sensitization, it is important to predict a threshold from the murine local lymph node assay (LLNA). We first confirmed that the combination of the human Cell Line Activation Test (h-CLAT) and the SH test improved the accuracy and sensitivity of prediction of LLNA data compared with each individual test. Next, we assessed the mutual correlations among maximum amount of change of cell-surface thiols (MAC value) in the SH test, CV75 value (concentration giving 75% cell viability) in a cytotoxicity assay, EC150 and EC200 values (thresholds concentrations of CD86 and CD54 expression, respectively) in h-CLAT and published LLNA thresholds of 64 chemicals. Based on the results, we selected MAC value and the minimum of CV75, EC150 (CD86) and EC200 (CD54) as descriptors for the input layer of an artificial neural network (ANN) system. The ANN-predicted values were well correlated with reported LLNA thresholds. We also found a correlation between the SH test and the peptide-binding assay used to evaluate hapten-protein complex formation. Thus, this model, which we designate as the "iSENS ver. 1", may be useful for risk assessment of skin sensitization potential of chemicals from in vitro test data.