May 24, 2018 OpenTox Asia

Transcription

1 May 24, 2018 OpenTox Asia Development of an artificial neural network model for risk assessment of skin sensitization using multiple in vitro sensitization tests and in silico parameter Morihiko Hirota Shiseido Global Innovation Center

2 Agenda 1. About ANN models for predicting LLNA threshold(in vitro/in silico) 2. Application of ANN model in the actual safety assessment.

3 Mechanism of induction phase of allergic contact dermatitis Induction phase Haptens ( ) Dendritic cell(dc) e.g., Langerhans cell) Activation AOP(Adverse Outcome Pathway) Key Events Chemical Structure & Properties Molecular initiating Event Metabolism,Penetration Electrophilic substance(in silico) Covalent interaction with cells protein (protein-binding DPRA) Keratinocyte Migration Cellular Response Dendritic cell activation h-clat Keratinocyte signal activation KeratinoSens Lymph node DC Antigen presentation T cell T cell Proliferation LLNA (in vivo) (OECD TG 429) LLNA:(Local lymph node assay) is the one of in vivo test which detect the lymph node proliferation by the treatment with chemicals.

4 DPRA for hapten-protein binding DPRA (Direct Peptide Reactivity Assay) detects binding among chemicals and model peptides (Cys-peptide/Lys-peptide) using HPLC. 24 h Flow cytometry analysis Ac-RFAACAA-COOH Ac-RFAAKAA-COOH Treatment of chemical Ac-RFAACAA-COOH (K) Each of the % peptide depletion of Cys and Lys was used as a descriptor. h-clat for dendritic cell activation h-clat (human Cell Line Activation Test) detects augmentation of CD86 and CD54 expression on chemical-treated THP-1 cells using flow cytometry. 24 h Flow cytometry analysis Treatment of chemical FITC-anti-CD86 Ab, -anti CD54 Ab Ab; antibody Minimum of threshold concentration of CD86/CD54 expression and CV75 was used as a descriptor.

5 KeratinoSens for ketatinocyte signaling KeratinoSens detects induction of ARE(anti-oxidant response element) dependent luciferase activity in HaCaT. EC1.5 (threshold concentration of ARE-dependent luciferase induction) was used as a descriptor. In silico structure alert for chemical structure & properties Toxtree

6 How to assess skin sensitization risk Hazard Identification: skin sensitizer or non sensitizer Understanding skin sensitization potency 1. skin sensitization category (weak, moderate, strong) 2. Threshold (EC3) Exposure assessment Risk assessment In the risk assessment for skin sensitization, hazard identification, potency and exposure assessment is important factor. In vivo sensitization assay (e.g. LLNA: Local lymph node assay (OECD TG 429) ) can provide the information of hazard(positive/negative) and potency including threshold concentration of chemicals. However, in vivo sensitization test for safety assessment for cosmetics are no longer available in EU.

7 Aim of the study Development of the prediction model for LLNA threshold (called EC3 values) using non-animal assays (in vitro or/and in silico). LLNA: Local lymph node assay (OECD TG 429)

8 Correlation between LLNA, h-clat, DPRA and KeratinoSens LLNA vs DPRA(Cys) LLNA vs h-clat(mit) LLNA vs KetatinoSens (EC1.5) R=0.64(p<0.05) R=0.57(p<0.05) R=0.52(p<0.05) LLNA vs DPRA(Lys) R=0.28(p<0.05) LLNA vs cell toxicity R=0.66(p<0.05) Each descriptor derived from h-clat, DPRA (Cys), DPRA (Lys) and KeratinoSens correlated with LLNA EC3, significantly. However, each indicator is not enough for risk assessment.

9 Artificial Neural Network (ANN) analysis DPRA-Cys Input Layer Hidden Layer Output Layer DPRA-Lys KeratinoSens TM h-clat CD86/CD54 (MIT) Cell tox(cv75) In silico structure alert Toxtree LLNA EC3 LLNA EC3 EC3 values for LLNA positive chemicals For LLNA negative chemicals, 101% was used as EC3 value (converted to µm/cm 2 unit). In vivo test data In vitro/in silico data In this study, the ANN analysis was performed. MIT; Minimum induction threshold of CD86/CD54 in h-clat

10 Log(Predicted LLNA EC3 (um/cm 2 ) ) Prediction of LLNA EC3s using ANN model(in vitro) After 10-fold cross-validation R= 0.90(N=134) RMS error=0.50 Log(Published LLNA EC3 (um/cm 2 )) Log(Predicted LLNA EC3 (um/cm 2 ) ) Model evaluation using test set Log(Published LLNA EC3 (um/cm 2 )) This ANN model showed good correlation between predicted values and tested values. *Root mean square (RMS) error = 2 ((measured value - predicted value) ) / number of data

11 In silico QSAR ANN model Chem3D Input layer Heat of formation Sum H Polarizability (α) Polarizability (γ) Hidden layer Output layer In vivo test data In silico-calculated descriptors from the threedimensional structure based on molecular orbital method were extracted.

12 Prediction of LLNA EC3s using ANN model(in silico(qsar)) In silico data Log (Predicted LLNA EC3) N=206 R=0.726 RMS error= Log (Published LLNA EC3) In vivo data The ANN model consisted of in silico-calculated descriptors of the three-dimensional structures showed good correlation with published LLNA threshold(called EC3).

13 Agenda 1. About ANN models for predicting LLNA threshold(in vitro/in silico) 2. Application of ANN model in the actual safety assessment.

14 Basic strategy of safety assurance on skin sensitization Hazard identification In vitro ANN h-clat/dpra/keratinosens Risk assessment Final confirmation In silico Weight of evidence Human tests RIPT, Use test (as confirmation) Readacross Original QSAR model(tsujita-inoue et al., J. Toxicol. Sci., 40 (2015) ) In our company, we usually assess cosmetic ingredient using in vitro, in silico(qsar), read across and human confirmatory test based on weight of evidence concept.

15 Evaluation flow using ANN models and read across In vitro assays (DPRA, KeratinoSens TM, h-clat) All negative One or more positive *Tsujita-Inoue et al., J. Toxicol. Sci., 40 (2015) In vitro ANN In silico ANN* No potential Lower EC3 is selected for QRA Strong/extreme(Reliability; not low) Margin of safety 10 Read across QRA Non/weak/moderate QRA: quantitative risk assessment QRA Human tests To avoid under-prediction of hazardous compounds, chemicals which were predicted strong or extreme sensitizer in this process were imposed 10 times safety margin to the predicted EC3 value.

16 chemical * : Atobe et al., J. Toxicol. Sci., 40 (2015) Extraction of similar chemicals by read across Protein binding Protein binding Structural similarity Structural similarity (%) Cramer classifi cation Cramer classification Skin absorption (>10% or not) Skin absorption (%) Toxicological alert Toxicological alert Sensitization alert OECD Toolbox DEREK Nexus Reported data In silico ANN *) Sensitization alert Similarity judgment Certain similarity (LLNA, GPMT etc.) sensitization potency (Reported data) Predicted Potential Target chemical No alert 100 Low 25 No alert No alert Moderate Chemical A No alert > 85 Low 30 No alert No alert High Moderate Chemical B No alert > 85 Low 18 Catechol No alert Medium Moderate Chemical C No alert > 85 Low 8 No alert Conjugated diene Medium Non Chemical D No alert > 85 Intermediate 20 No alert No alert Medium Moderate Chemical E No alert > 80 High 5 No alert 1,2-Dihydroxybenzene Low No data Chemical F No alert > 75 Low 0.1 Alkyl alcohol Conjugated diene Low Extream Chemical G Alert > 85 Intermediate 35 No alert No alert Medium No data Chemical H Alert > 85 Low 20 No alert No alert Medium Non Chemical I Alert > 80 Intermediate 5 Alkyl alcohol Conjugated diene Very low Strong Chemical J Alert > 80 High 0.1 Peracid Conjugated diene Very low No data Chemical K Alert > 80 High 0.5 Organic peroxide 1,2-Dihydroxybenzene Very low Non Chemical L Alert > 75 Low 4 No alert No alert Low Strong Skin sensitization potency of similar chemicals When the similar chemicals with strong (extreme) sensitization potential (not low reliability) were detected, the safety margin(x 10) was needed.

17 Example of read across (1,2-cyclohexane dicarboxylic anhydrate ) chemical Protein binding (PB) Structural similarity (%) Skin absorption (%) Cramer classifi cation Toxico logical alert Sensitization alert Similarity judgment Skin sensitization potency of similar chemicals Reliability Read-across Structure Protein binding Structural similarity (%) Skin absorption (%) Cramer classification Toxicological alert Sensitization alert Certain similarity Sensitization potency 1,2-Cyclohexane dicarboxylic anhydrate (Target chemical) Acid anhydride Acetates MA: Direct acylation involving a leaving group Mechanistic Domain: Acylation High Acid anhydride Cyclic acid anhydride Azlactone or analogue Predicted Strong Phthalic anhydride Acid anhydride Acetates MA: Direct acylation involving a leaving group Mechanistic Domain: Acylation > Low Acid anhydride Cyclic acid anhydride Azlactone or analogue High Strong Octahydrocoumarin Active cyclic agents MA: Ring opening acylation Mechanistic Domain: Acylation > High No alert No alert Low Not reported 1,2-cyclohexane dicarboxylic anhydrate: Predicted Strong (Additional safety margin 10)

18 Assessment of 1,2-cyclohexane dicarboxylic anhydrate Chemicals Published LLNA EC3(%) Predicted EC3 (In vitro/in silico) Read Across evaluation 1,2-cyclohexane dicarboxylic anhydrate 0.9% 17% Strong/Extreme Chemicals 1,2-cyclohexane dicarboxylic anhydrate Additional safety margin by read across needed Predicted EC3(%) (by ANN model(toxtree))(adjusted) 1.7%

19 Conclusions & Discussion The ANN models using h-clat, DPRA, KeratinoSens and in silico structure alert detecting models(toxtree) and in silico QSAR model had a capacity to predict LLNA EC3. Each event in the AOP of skin sensitization (protein-binding, dendritic cell activation and keratinocyte signaling) might affect skin sensitization potency level. Our Read-across approach might be useful for extracting similar chemicals. The combination of in vitro/in silico ANN model(s) and read across study (a kind of WoE approach) is useful for avoiding under-predicted results. This integrated model could be useful to evaluate cosmetic ingredients for risk assessment of skin sensitization.

20 Finally, Thank you for your attention.