SaferSkin™ web interfaceTutorial

The SaferSkin™ Application provides an open implementation of integrated testing strategies (ITS) that assesses skin sensitisation potency based on three previously described approaches:

  1. The Bayesian Network: Initially developed at Procter & Gamble (Jaworska et al. 2015). The predictor allows the user to combine information from in silico, in chemico and in vitro assays to reach a comprehensive judgement on the skin sensitisation potential of a chemical. This implementation combines information from three validated alternative assays (DPRA, KeratinoSens and h-CLAT) with in silico predictions for bioavailability using open cheminformatics tools from the community.
  2. Multiple Regression: This approach described in Natsch et al. (2015) uses a multiple regression model based on the the results of in vitro assays (KeratinoSens™ and peptide reactivity assay) and physico-chemical properties of the compound to predict LLNA category and pEC3.
  3. ‘2 out of 3’ Voting: Uses a weight of evidence approach (as presented in Bausch et al. (2012) by taking a majority of two congruent results of the first 3 key events in the skin sensitisation Adverse Outcome pathway (AOP) (represented here by the DPRA, KeratinoSens™, and h-CLAT in vitro assays), and returns a simple prediction of the compound being a sensitiser or non-sensitiser.

Step 1: Submit your molecule

Enter the molecule of interest either by entering the SMILES representation (Figure 1) or by sketching the molecule (Figure 2). You can also select one of the predefined examples from the drop-down menu as seen in Figure 1.

Figure 1. Submit molecule as SMILES or select a predefined example.
Figure 2. Submit molecule by sketching.

Step 2: Review/enter molecular descriptors and experimental values

Once a molecule has been entered in step 1, the in silico descriptors are calculated for each method. The descriptors for each method are listed below. While most are supplied by the application, a few need to be entered by the user.

Bayesian network input

Molecular descriptors: The Water solubility @ pH7, LogKow (Octanol/Water partition coefficient), Protein binding, and logD @ pH7 are calculated based on our underlying QSAR models and supplied by the application (Figure 3a). If descriptor values were not cached before, they will be calculated on the spot as shown in Figure 3b. Please allow for a few seconds for all calculations to complete before proceeding to the next step. The TIMES classification and Michael acceptor class (yes/no) need to be entered by the user.

Figure 3a. Review of molecular descriptors for Bayesian Network approach
Figure 3b. Calculation of molecular descriptors.

You may also choose to enter all of the physico-chemical and in silico descriptors on your own.

Experimental values: If available, enter the experimental values from the validated alternative assays, DPRA, KeratinoSens™ and h-CLAT as shown in Figure 4. This is recommended to improve the prediction estimate. Clicking on the ? buttons of each assay entry provides more information about the assay and the expected input values.

Figure 4. Inputs for experimental values.

Note on Assay applicability domain:
You will receive an alert if a compound is out of an assay’s applicability domain. If the compound is outside the applicability domain (judgement based on the compound’s solubility and its fraction ionized at pH = 7, see Jaworska et al. [1] for more details), the assay values are ignored and the corresponding input textbox turns orange (as shown in Figure 5) as an indication that the chemical is out of applicability domain.

Figure 5. Alert shown when compound is out of an assay’s applicability domain.

Multiple regression input

Molecular descriptors: The only molecular descriptor required is the vapor pressure of the compound (measured in Pa). This is a mandatory entry in order to make a prediction.

Figure 6. Molecular descriptors for multiple regression method.

Experimental values: It is mandatory to enter the experimental values (KEC 1.5 and IC50 in µM) from the KeratinoSens™ assay and the Kmax from the Covalent binding to skin proteins or Cor1-C420 assays as shown in Figure 7. Clicking on the ? buttons of each assay entry provides more information about the expected input values.

Figure 7. Experimental values needed for multiple regression method.

‘2 out of 3’ voting input

This method requires no molecular descriptors only the experimental values from the three validated assays (DPRA, KeratinoSens™ and h-CLAT) as shown in Figure 8. For the h-CLAT assay positive/negative entries are required for the CD54 and CD86 measurements.

Figure 8. Experimental values needed for ‘2 out of 3’ voting method.

Step 3: Make the prediction

Finally, click on “Make prediction” button on the right side (Figure 9).

Figure 9. To make a prediction click on the “Make prediction” button.

This will return a prediction of the compound’s skin sensitization which is the LLNA pEC3 class of non, weak, moderate or strong sensitiser for the Bayesian Network and Multiple Regression methods, and a sensitiser/not a sensitizer classification for the ‘2 out of 3’ voting method (Figure 10).

Figure 10. Summary result.

In the event that no/incomplete in silico or experimental values are provided in step 2 for the Bayesian Network method, after making the prediction (step 3) the app also makes recommendations on which experimental or in silico values will provide the highest value of information and improve prediction (Figure 11).

Figure 11. Recommendation for the experiment providing the highest value of information (Vol).
References
  • (1) Jaworska, J. S.; Natsch, A., Ryan, C., Strickland, J., Ashikaga, T., and Miyazawa, M. Bayesian Integrated Testing Strategy (ITS) for Skin Sensitization Potency Assessment: A Decision Support System for Quantitative Weight of Evidence and Adaptive Testing Strategy. Archives Toxicol. 2015.
  • (2) A. Natsch, R. Emter, H. Gfeller, T. Haupt and G. Ellis. Toxicol. Sci. 2015.
  • (3) C. Bausch, S. N. Kolle, T. Ramirez, T. Eltze, E. Fabian, A. Mehling, W. Teubner, B. van Ravenzwaay and R. Landsiedel. Regul. Toxicol. Pharmacol. 2012.