Approaches to Skin
Sensitisation Prediction52nd ICGM
Senior Scientist
Dr Rachael Tennant
Overview
• Skin sensitisation• Drive towards non-animal methods
• Adverse Outcome Pathway (AOP)
• OECD-validated assays
• Derek Nexus
• Lhasa’s defined approach• Results
• Conclusions
What is skin sensitisation?
• Common occupational disease
• Estimated to cost the EU €600 million and 3 million lost working days
• High concentrations and/or repeated exposure increases the likelihood of
becoming sensitised to a specific chemical
• Not life-threatening but is life-long
• In vivo assays no longer permitted for predicting skin sensitisation of
cosmetics in the EU
• In chemico / in vitro assays now used where possible
• Assays cannot be used alone - must be used in combination with other
information sources - known as a defined approach (DA)
Skin sensitisation AOP
Figure adapted from OECD 2012, The Adverse Outcome Pathway for Skin Sensitisation Initiated by
Covalent Binding to Proteins Part 1: Scientific Evidence, Series on Testing and Assessment, No. 168.
MIE = Molecular Initiating Event
KE = Key Event
DC = Dendritic Cells
AO = Adverse Outcome
Organism responseOrgan responseCellular responseMolecular Initiating
Event (MIE)
Haptenation T-cell activation
Activation of DC
Stress response
Skin sensitisation
Skin sensitisation Adverse Outcome Pathway (AOP)
KE3
KE2
AOMIE/KE1 KE4
Chemical structure &
properties
Metabolism & penetration
Electrophilic substance
a DA consists of a fixed data interpretation
procedure (e.g. decision tree) using a defined set
of information sources to derive a prediction (e.g.
in silico predictions, in chemico, in vitro data)
Skin sensitisation AOP
• In silico tools can be used as an information source
• Derek Nexus (Derek) covers multiple aspects of the AOP and would be a
valuable complement to a defined approach
• Lhasa’s defined approach combines in chemico/in vitro assay data with
Derek predictions to provide a predicted skin sensitisation outcome
Figure adapted from OECD 2012, The Adverse Outcome Pathway for Skin Sensitisation Initiated by
Covalent Binding to Proteins Part 1: Scientific Evidence, Series on Testing and Assessment, No. 168.
MIE = Molecular Initiating Event
KE = Key Event
DC = Dendritic Cells
AO = Adverse Outcome
Organism responseOrgan responseCellular responseMolecular Initiating
Event (MIE)
Haptenation T-cell activation
Activation of DC
Stress response
Skin sensitisation
Skin sensitisation Adverse Outcome Pathway (AOP)
KE3
KE2
AOMIE/KE1 KE4
Chemical structure &
properties
Metabolism & penetration
Electrophilic substance
Derek Nexus
• In silico toxicity prediction tool which covers multiple KEs in the skin sensitisation AOP
• Consists of a knowledge base (KB) containing structural alerts (patterns) written by
Lhasa expert scientists
• Provides hazard as well as potency predictions (EC3%)
• Fulfils the OECD (Q)SAR validation principles
• Consists of three main components:
• Alerts
• Potency predictions
• Negative predictions
OECD. 2007. Guidance Document on the Validation of (Quantitative)Structure-Activity Relationships [(Q)SAR] Models.
Limitations Not applicable for:
Alerts based mainly on public data (and some
proprietary data donated by Lhasa Limited members)
Substances of unknown or variable composition, complex
reaction products or biological materials (UVCBs)
Potency predictions not provided for all alerts
Derek Nexus: Typical alert
Derek Nexus: Potency predictions
• EC3 data from the local lymph node assay (LLNA) is used for model
• EC3 - effective concentration required to cause a 3-fold increase in T-cell
proliferation
• Measure of potency (lower EC3 = more potent sensitiser)
All EC3 data in model
Canipa et al., 2017. J. Appl. Toxicol. 37, 985–995.
Derek Nexus: Potency predictions
• Derek’s alerts are used to form a mechanistic domain for query chemicals
• Sensitisers undergoing the same mechanism expected to have similar potencies
• EC3 is predicted using the weighted mean of the nearest neighbours’ (NN)
potency values
Canipa et al., 2017. J. Appl. Toxicol. 37, 985–995.
EC3 data within same
alert/mechanistic domain
Derek Nexus: Negative predictions
Query fragment present in
Lhasa dataset?
Query fragment contained
in known false negatives?Outcome Example
Yes No
Non-sensitiser with
no misclassified or unclassified features
Yes Yes
Non-sensitiser with
misclassified features
No N/A
Non-sensitiser with
unclassified features
No alert
fired
Ionic form not present
in fragment libraryFragment similar to
known false negative
Fragment
library
Defined Approach: Hypothesis
• Use a Key Event (KE) approach
• Assays measuring the same KE will have similar limitations and applicability
domains
• Apply exclusion criteria to chemicals:
• Based on known assay limitations and confidence in Derek predictions
• Ensure the most relevant information source(s) are used for a given chemical
(class)
• By de-prioritising results from less applicable assays and/or predictions
• Use prioritised results until a concordant result is obtained
• 2 out of 3 majority call
Defined Approach: Exclusion criteriaExclusion criteria Derek MIE KE2 KE3 Comment
Metabolism Prohapten ✓ ✗ ✓ ✓Assays lacking metabolic competency are
deprioritised as they are less likely to predict prohaptens well
logP
> 3.5 ✓ ✓ ✓ ✗Cell-based assays are deprioritised for
chemicals with a logP > 3.5 (KE3) and logP > 5 (KE2) as more lipophilic chemicals may
lack high solubility in these cell-based
assays> 5 ✓ ✓ ✗ ✗
Lysine
reactiveExclusive ✓ ✓ ✗ ✓
The Nrf2-ARE pathway is associated with
cysteine binding - lysine-reactive chemicals may not be reliably predicted
Reasoning
levelEquivocal ✗ N/A
Alerts with a likelihood of equivocal have
less evidence of skin sensitisation potential than other likelihoods (e.g. certain) and are
thus deprioritised
Negative
prediction
Misclassified
features✗ N/A Negative predictions with ‘misclassified
features’ or ‘unclassified features’ are deprioritised as these are associated with
higher uncertainty.Unclassified
features✗ N/A
Defined Approach: Hazard prediction
Potency
prediction
model
1st assay
Equivocal
Non-sensitiser with
misclassified or
unclassified
features
2nd assay
3rd assay
2nd assay
Certain
Probable
Plausible
Non-sensitiser
Doubted
Improbable
Impossible
Derek
alert
outcome
Potency category 5/6
(GHS not classif ied)
QueryUse Derek outcome to
determine decision tree branch
Prioritise in chemico/in vitro
assays using exclusion criteria
Potency category 1 (GHS 1A)
Potency category 2 (GHS 1A)
Potency category 4 (GHS 1B)
Run in chemico/in vitro assays in order of AOP (MIE → KE2 → KE3) unless de-
prioritised by exclusion criteria
Potency prediction using k- nearest neighbours model
Hazard prediction using ‘2 out of 3’ approach
Exclusion
criteria
sensitiser
sensitiser
non-
sensitiser
non-
sensitiser
sensitiser
non-
sensitiser
Blue italics = Derek outcome
Red arrow = positive result
Green arrow = negative result
+
-
+
+
+
1st assay
+
+
+
+2nd assay
1st assay
2nd assay
-
-
-
-
-
-
-
+
Potency category 3 (GHS 1B)
Defined Approach: Potency prediction
Human data1-2Mouse data
(EC3)Combine
and curate
Combined
dataset
ℎ𝑢𝑚𝑎𝑛 > 𝑚𝑜𝑢𝑠𝑒
n = 199 n = 672
n = 762
1. Basketter et al., 2014. Dermatitis, 11-21.
2. Api et al., 2017. Dermatitis, 299-307.
3. Basketter, 2016. Altern. Lab. Anim., 431–436.
0 1.0
Non-sensitiser (6)
Pote
ncy C
ate
gory
Extreme (1)
Strong (2)
0.5
Query compound with predictedpotency category
Tanimoto Similarity
Top 10 Nearest Neighbours
Moderate (3)
Weak (4)
Very weak (5)
Compounds that fire alert
Human potency category
name
GHS
Classification
Human potency
category
Equivalent EC3
value (%)3
extreme 1A 1 < 0.2
strong 1A 2 0.2 – 2
moderate 1B 3 2 – 20
weak 1B 4 20 – 80
very weak/non-sensitiser Not classified 5 > 80
non-sensitiser Not classified 6 negative
Results - Hazard
BA = Balanced Accuracy
Se = SensitivitySp = SpecificityBased on a dataset of 210 chemicals from Urbisch et al., 2015, Regul. Toxicol. Pharmacol., 71, 337–351.
Full analysis published in Macmillan & Chilton, Reg. Toxicol. Pharmacol. 2019, 101, 35-47.
Results - Potency (GHS Classification)
Human
n = 79
Acc = 76%
Defined approach prediction vs in vivo outcome
LLNA
n = 174
Acc = 73%
no cat no cat
no
cat
no
cat
Under-prediction
Over-prediction
Under-prediction
Over-prediction
Based on a dataset of 210 chemicals from Urbisch et al., 2015, Regul. Toxicol. Pharmacol., 71, 337–351.
Full analysis published in Macmillan & Chilton, Reg. Toxicol. Pharmacol. 2019, 101, 35-47.
Defined Approach: Prototype
• Lhasa have developed a prototype application that can be accessed through the web• Data can be input manually or by uploading a Derek report
• Supply in chemico/in vitro data and the prediction is given
• Examples are available in the application with the Derek and in chemico/in vitro data already populated
• Web app available at: https://skinsensda.lhasacloud.org
Conclusions
• Lhasa Limited have developed a defined approach (DA) which incorporates in
silico predictions with results from OECD-validated in chemico / in vitro assays• Correctly predicts LLNA hazard for 82%, and LLNA GHS classification for 73% of
the dataset analysed
• Correctly predicts human hazard for 85%, and human GHS classification for 76%
of the dataset analysed
• The DA is currently being considered by the OECD for inclusion in the
upcoming guidance on DAs
• Future work will focus on:• Using any new applicability domain knowledge to amend the exclusion criteria
• Incorporating results from new OECD-validated assays
• Extending predictions from potency (GHS 1A/1B) to quantitative
predictions which may be more useful for risk assessment
Acknowledgements
• Donna Macmillan
• Martyn Chilton
• Sam Webb