Abstract
We have previously reported an Artificial Neural Network (ANN) model for predicting LLNA EC3 to assess the risk of skin sensitization using in vitro tests. In this study, we report on the ANN model developed using Amino acid Derivative Reactivity Assay (ADRA), which can be performed at lower concentrations and is Known for low frequency of co-elution, as opposed to Direct Peptide Reactivity Assay (DPRA), which is a known method for evaluating the covalent binding of test substance and skin proteins.
We defined LLNA EC3 as the objective variable, and used characteristic values obtained from ADRA, KeratinoSens™, h-CLAT, and toxicity alerts from Toxtree as descriptors. We developed two models based on data of ADRA at molar and gravimetric concentrations, and used Python and scikit-learn, a machine learning library, to develop the ANN models. In addition, we evaluated the robustness of the models by cross-validation and compared their prediction accuracy with our previously reported ANN model1) using characteristic values of DPRA.
We compared the two models with the DPRA-based model using R² and RMSE, the results were comparable in both models. In addition, no substances of external data were underestimated for the LLNA Category. Furthermore, the predictive value between the two models developed using ADRA showed a high correlation (R²=0.865). Finally, we consider that the results suggest the usefulness of ANN models using ADRA for risk assessment of skin sensitization.
Reference
1) Hirota et al., J Appl Toxicol. 2018, 38, 514-526
