Post on 22-Jun-2020
transcript
Skin Sensitisation
Negative predictions, Validations and
Integrated Testing Strategies
Chris Barber
Acknowledgements – Donna MacMillan & Martyn Chilton
chris.barber@lhasalimited.org
Outline
• Skin sensitisation is an important endpoint within Derek
• Cosmetic/personal care members can not undertake animal tests
• Important endpoint to protect workers in chemical manufacturing
• In vitro and in vivo assays are expensive – desire to minimise
• Improvements implemented and planned within Derek…
• Improved alerts
• Prediction of potency
• Integrated Testing Strategy (ITS)
• Negative predictions (not yet implemented)
Improving alerts for skin sensitisation - performance
• Particularly grateful for member donations
• Significantly improved predictivity within their chemical space
• Currently 88 skin sensitisation alerts.
• Performance against public data is good:
Acc Sens Spec PP NP No. of alerts
Derek 2015/2016 74 78 70 73 76 88
Derek 2014/2015 72 72 71 72 72 80
*Analysis based on a data set of 1316 sensitisers and 1283 non-sensitisers based on conservative
combination of results from the LLNA and/or guinea pig assays.
Improving alerts for skin sensitisation - performance
• Poster presented at SOT – available from our website
Improving alerts for skin sensitisation - text
Skin Potency (LLNA) EC3 model – released Jan2016
• Lhasa collated and curated all available public LLNA data
• Over 1000 EC3 values were assessed as good quality
• 600 unique compounds were added to our Master EC3 data base
• Investigated different methods for predicting EC3 from structure
• …wanted a transparent, interpretable, mechanistically-based,
scientifically robust, approach that captured what we know – and
what we don’t know…
• Expert derived structural alerts for different potency categories
• Linear regression models using physicochemical descriptors
• Nearest neighbour approach using structural fingerprints
Skin Potency (LLNA) EC3 - Model Methodology
• k-Nearest Neighbors (kNN) model
• using weighted average scaled by the Tanimoto distance between the
query compound and dataset compounds
• Model approach:
• A valid model compound must fire the same Derek alert (mechanistic domain)
• Similarity is based upon structural fingerprint (interpretable)
• Minimum of 3 model compounds required (robust)
• Supporting examples - data and references are provided (supported)
• User can inspect, add or remove examples (enable expert analysis)
• EC3 prediction based upon provided training compounds (transparent)
Skin Potency (LLNA) EC3 model
• Interactive display showing nearest neighbours and
supporting data including source reference
Skin Potency (LLNA) EC3 model - Performance
20%13% 16%
11%
23%14%
18%11%
44% 60%51%
64%
29% 50% 36%
64%
36%27%
33%24%
48%
37%46%
25%
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
3-fold 5-fold ECETOCClassification
GHSClassification
3-fold 5-fold ECETOCClassification
GHSClassification
More potent EC3 prediction
Within EC3 prediction range
Less potent EC3 prediction
Internal validationn = 45
External validationn = 103
• Model is stable and relatively conservative with <20% under-
predicted before expert review
Derek as part of an Integrated Testing Strategy (ITS)
Integrated Approach to Testing and Assessment (IATA)
• No single in vitro or in silico assay can replace LLNA
• Many studies have looked at combinations…
• Some have integrated in silico into the approach
• Most apply a simple weight-of-evidence approach
• ‘Test in everything and take the majority call’
• We looked for an approach that
• Integrated Derek as a transparent in silico model
• Supported expert assessment
• Reduced the number of assays needed
• Look at available data and then decide which assay is appropriate
Derek as part of an Integrated Testing Strategy (ITS)
• Grounding the approach in the AOP…
Predicting skin sensitisation using a decision tree integrated testing strategy with an in silico model and in chemico/in vitro assays.
Donna Macmillan…. Regulatory Toxicology and Pharmacology 76 (2016) 30
Derek as part of an Integrated Testing Strategy (ITS)
Predicting skin sensitisation using a decision tree integrated testing strategy with an in silico model and in chemico/in vitro assays.
Donna Macmillan…. Regulatory Toxicology and Pharmacology 76 (2016) 30
ITS performance: Lhasa decision tree
Urbisch et al., Reg. Tox. Pharmacol., 2015, 71, 337-351
ITS Acc PP NP Coverage
2/3 in vitro WoE 0.79 0.89 0.59 0.85
2/3 DX + in vitro WoE 0.84 0.90 0.70 0.80
Lhasa DT 0.85 0.86 0.81 0.75
Derek as part of an Integrated Testing Strategy (ITS)
• We are working with members…
• …to evaluate the performance of this approach
• Have you any data to support the joint publication?
• …and with regulators…
• …who are yet to develop a clear position on the use of ITS
Skin sensitisation – Negative Predictions
• We extended mutagenicity from ‘nothing to report’ to
• Negative
• Negative with Misclassified features
• Negative with Unclassified features• It’s difficult, but important, to make negative predictions.
Williams, R..., Regul. Toxicol. Pharmacol. 2016, 76, 79
• This approach also works well with skin sensitisation
• Reactivity-based mechanism
• Fragments capture well reactivity
• We have a sufficiently large dataset to have good coverage
• Research shows this method is robust with good reliability
• It will be integrated into a future Derek release
Skin sensitisation – Negative Predictions
• Misclassified and unclassified features occur infrequently. However the data indicate that making negative
predictions for these outcomes is valid: these features can be considered weak arguments against the
negative prediction which may necessitate further expert analysis depending on the usage of the prediction
No potency prediction
In vitro dataITS
QUERYCOMPOUND
Derek Roadmap for Skin Sensitisation
Likelihood
Hazard prediction
SS alert?
Nothing to report
NO
YES
New functionality
Future functionality
Existing functionality
Negative predictions
Non-sensitiser
Application within ITS
EC3
Potency prediction
No potency prediction
In vitro dataITS
QUERYCOMPOUND
Derek Roadmap for Skin Sensitisation
Likelihood
Hazard prediction
SS alert?
Nothing to report
NO
YES
New functionality
Future functionality
Existing functionality
Negative predictions
Non-sensitiser
Application within ITS
EC3
Potency prediction
Summary – Skin Sensitisation in Derek
• Improved alerts for skin sensitisation
• Performance improved through shared data
• Potency predictions now made
• Robust interactive approach with good performance
• Promoting the use of in silico methods within ITS
• Safe, cost effective, designed for expert assessment
• Tested methodology for negative predictions
• ...will be implemented within Derek
Questions?