S2: Development of an artificial neural network model for risk assessment

Development of an artificial neural network model for risk assessment in skin sensitization using multiple in vitro sensitization tests and in silico parameters

Development of an artificial neural network model for risk assessment, Morihiko Hirota , OpenTox Asia 2018
PRESENTING AUTHOR: 

Morihiko Hirota

INSTITUTION / COMPANY : 

Shiseido Global Innovation Center, Kanagawa, Japan

POSITION: 

Toxicologist

REFERENCES: 
Hirota et al., Development of an artificial neural network model for risk assessment of skin sensitization using human cell line activation test,
direct peptide reactivity assay, KeratinoSens™ and in silico structure alert parameter.
J Appl Toxicol. 2017 (doi: 10.1002/jat.3558) 
ABSTRACT CONTENT / DETAILS: 

In the safety assessment of cosmetic ingredients, the importance of both in vitro and in silico tests has increased since 2013. In the field of skin sensitization, three in vitro methods, the direct peptide reactivity assay (DPRA), KeratinoSens™, and the human Cell Line Activation Test (h-CLAT), were adopted under OECD guidelines. Recently, several prediction model(s) for skin sensitization based on the integrated testing strategy (ITS) concept have been reported. However, most of these models were not made as a quantitative risk assessment. So, we focused on threshold values (EC3) of the murine local lymph node assay (LLNA) and tried to make a model for predicting this EC3 value. In this study, an artificial neural network (ANN) analysis was used as a tool to combine  in vitro and in silico parameters.

 We investigated the relationship between LLNA EC3 and the indicators derived from in vitro tests (h-CLAT, cell toxicity, DPRA and KeratinoSens™ (n=134)). First, predictions based on ANN analysis using combinations of parameters from all three in vitro tests (not in silico parameter) showed a good correlation with LLNA EC3 values (r=0.87, RMS error=0.55). However, when the ANN model was applied to a testing set of 28 chemicals that had not been included in the training set, predicted EC3s were overestimated for some chemicals. Incorporation of an additional in silico or structure alert descriptor (obtained with TIMES-M or Toxtree software) in the ANN model improved prediction performance (r=0.89-0.91, RMS error=0.47-0.51) and the overestimated results of certain chemicals.

These refined results suggested that the incorporation of in silico parameters is useful for improving prediction performance of the ANN model. The modified ANN model could contribute to the evaluation of cosmetic ingredients in their potential effects regarding skin sensitization.