Applied In Vitro Toxicology

Isothiazolinones (ITs) are widely used as antimicrobial preservatives in cosmetics and as additives for preservation of consumer and industrial products to control bacteria, fungi, and algae. Although they are effective biocides, they have the potential to produce skin irritation and sensitization, which poses a human health hazard. In this project, we evaluated nonanimal defined approaches (DAs) for skin sensitization that can provide point-of-departure estimates for use in quantitative risk assessment for ITs.

The skin sensitization potential of six ITs was evaluated using three internationally harmonized nonanimal test methods: the direct peptide reactivity assay, KeratinoSens™, and the human cell line activation test. Results from these test methods were then applied to two versions of the Shiseido Artificial Neural Network DA.

Sensitization hazard or potency predictions were compared with those of the in vivo murine local lymph node assay (LLNA). The nonanimal methods produced skin sensitization hazard and potency classifications concordant with those of the LLNA. EC3 values (the estimated concentration needed to produce a stimulation index of three, the threshold positive response) generated by the DAs had less variability than LLNA EC3 values, and confidence limits from the DAs overlapped those of the LLNA EC3 for most substances.

The application of in silico models to in chemico and in vitro skin sensitization data is a promising data integration procedure for DAs to support hazard and potency classification and quantitative risk assessment.

We are developing computational models for basic nervous system pathways for use in toxicology, pharmacology, and medicine. γ-Amino butyric acid (GABA) is the major inhibitory neurotransmitter in the central nervous system, and acts through GABA_A (cys-loop) or GABA_B (G-protein–coupled) receptors.

We mined publicly available data for compounds reported to interact (actives) or not interact (negative controls) with either of these receptors, and used these data to train binary random-forest machine-learning models to predict such interactions for novel compounds. Using the Konstanz Information Miner (KNIME), we developed two types of models for interaction with these receptors; both types derive and analyze structural fingerprints, and conserved scaffolds with the second type also incorporating information on binding energies and topologies from protein–ligand docking.

The scaffold/fingerprint-based models were highly sensitive (85.5%–93.5%) and accurate (89.8%–94.0%). Incorporation of pharmacophore data (derived from docking) showed sensitivities of 89.7%–96.5% and improved the GABA_B model, which had fewer active compounds (330 vs. >4000 for the GABA_A receptor). The models exhibited high positive and negative predictivity (82.6%–97.5%). Although robust sensitivity was achieved for the GABA_A receptor with 17% of available data in a fingerprint-based model, sensitivity for the less well-studied GABA_B receptor was lower. We anticipate that model performance for GABA_B should improve with the generation of additional data.

The models can be used in an integrated testing approach to complement in vitro screening and as a covariate in Quantitative Structure Activity Relationship ((Q)SAR) models for other neurological endpoints.