An Online Integrated Testing Strategy Development for Skin Sensation Assessment
Allergic contact dermatitis (aka., skin sensitization) accounts for 20% of all contact dermatitis cases and hence has an estimated annual cost up to $200 million. It is also a public health problem, responsible for more than seven million outpatient visits annually. Currently, there are more than 3700 substances that are identified as contact allergens. Traditionally, skin sensitization hazards are assessed using in vivo animal experiments. However, the legislation is increasingly put in place to encourage the replacement of such experiments with non-animal methods. Cosmetic products in the European Union are now banned from using animals for hazard testing. A prohibition is also in place for the sales of products that have been previously tested on animals or that contain ingredients which have undergone such testing following the ban. Multiple tests have been suggested to reflect the current mechanistic understanding of the processes leading to skin sensitization and many framework proposals were reported to guide classification decisions with alternative non-animal data.
We present an implementation for an integrated testing strategy approach previously described by Jaworska et al. The Integrated Testing Strategy (ITS-3) utilizes a Bayesian network to assesses skin sensitization potency by combining information from three validated alternative assays, DPRA, KeratinoSens and h-CLAT as well as in silico predictions for bioavailability to provide a quantitative estimation for skin sensitization potency (a stimulation index = 3; EC3) across a four-category system (Non-, weak, moderate or strong sensitizers). The one-step look-ahead hypothesis was used to calculate the VoI from all network variables which were therefore ranked according to their importance in the decision process as suggested by Jaworska et al. and is used as the methodology to guide testing (i.e., by suggesting the variable (in vitro or in silico) for which its knowledge would maximize information gain about skin sensitization).
The calculated Bayesian factors (BF) serve to correct the imbalance in chemical representation in the training set as well as provide a quantitative assessment of the confidence in prediction. The threshold for strength of evidence suggested by Goodman were used. The posterior probability of chemicals is corrected to reflect the anti-inflammatory effect of Michael acceptors. The decision of whether a compound is considered a direct Michael acceptor for which correction is needed is based on SMARTS patterns.
The deployed web application allows risk assessors to use a web-interface to perform safety assessment for skin sensitization in an interactive manner. As structures are being sketched or entered as SMILES strings, the application calculates in silico parameters for solubility, lipophilicity, partition coefficient and protein binding. It provides instant feedback if chemicals are thought to be out of experiments’ applicability domain and suggests next experiment to be performed in order to achieve the highest value of information. The robust architecture allows the integration of the tool into workflow systems such as KNIME and Pipeline Pilot through he use of application programming interfaces (APIs) or in custom applications.