Negative Predictions

Negative Predictions
ICGM Phoenix, March 2014
Dr Chris Barber
Director of Science
[email protected]
Negative Predictions
OUTLINE
• Impact of changes driven by M7
• Negative predictions in Derek for mutagenicity
 The science
 Performance
 Using it in practice…
• What further information / development would you like to see?
 Focus of the first workshop
in silico predictions for M7
• Use of models that predict Ames outcomes
• 2 complementary methods should be applied
 One expert rule-based + one statistical-based
 Models should follow OECD Principles for QSAR
 The absence of alerts from both is sufficient to conclude that the
impurity is of no concern

Seems unlikely that ‘out of domain’ will be considered a prediction!
• Expert review is needed to provide additional evidence for
any prediction
…and to explain conflicting results
Lhasa is hosting a webinar on April 16…
• 2014 vICGM: Members of the FDA present on ICH M7
 Naomi Kruhlak and Mark Powley of the Food and Drug
Administration will be presenting on the proposed ICH M7
guidelines

Naomi - “FDA/CDER Current Practices for (Q)SAR Analysis under ICH
M7”
o overview of (Q)SAR models and methods used at FDA/CDER for the
prediction of bacterial mutagenicity, including specific expert analysis steps
applied.

Mark - “Reconciling Conflicting (Q)SAR Predictions in Impurity
Evaluations”
o will address the potential regulatory implications of discordant (Q)SAR
predictions and describe strategies used to reconcile such results
Negative Predictions
OUTLINE
• Impact of changes driven by M7
• Negative predictions in Derek for mutagenicity
 The science
 Performance
 Using it in practice…
• What further information / development would you like to see?
 Focus of the first workshop
Derek Nexus – an expert knowledge base
• Built using public & confidential data
• Comprises of >110 alerts for mutagenicity
 Can be further customised by members with private knowledge
• Derek is the preferred system for mutagenicity predictions
 In silico methods combined with expert knowledge rule out mutagenic
potential of pharmaceutical impurities: An industry survey
 Regulatory Toxicology and Pharmacology, 2012, 62, 449–455
– Pfizer, Novartis, GSK, AZ, Lilly, Hoffmann-La Roche, Covance, Merck,
J&J
 Use of in silico systems and expert knowledge for structure-based
assessment of potentially mutagenic impurities
 Regulatory Toxicology and Pharmacology, 2013, 67, 39
– Bayer, Sanofi, AZ, Hoffmann-La Roche, Computational Toxicology
Services LLC, BMS, Pfizer, Servier, Novartis, J&J, Abbott, Merck,
Boehringer, NCSP
Data sharing remains critical for Derek’s performance
New / modified
alerts since 2012 KB
Sensitivity
86%
30
Public
Proprietary
82%
20
78%
74%
10
70%
66%
0
D2012KB
Feb-2013
Apr-2013
Jul-2013
Aug-2013
D2014KB
• SOT Poster - Can public data improve mutagenicity predictions for
proprietary compounds?
 Richard V Williams and Chris Barber
Enhancing Derek Nexus for mutagenicity
• Designed to support expert analysis for M7
 Provides additional supporting information
 Recommends where expert should focus analysis
• If no alerts for mutagenicity were found, Derek Nexus would
return ‘Nothing to report’
 With your support, we have developed a robust way to extend
this and provide further information
 Next release of Derek Nexus will make an explicit prediction of
inactivity for mutagenicity
Supporting expert analysis
• In the absence of a positive alert, experts ask
“Is there any reason to be concerned with this prediction?”
 “Are there any unusual features in my molecule?”
 “Are there features associated with false negative predictions?”
 “Do I have additional confidential information?”
a feature = a property derived from structure
Negative predictions for mutagenicity
• Lhasa experts have developed two lists
 Features known to the model



present in the Lhasa Ames Test Ref Set
encoded within structural alerts
present in Derek examples
features not in this list are Unclassified
 Features found in non-alerting mutagens

features present in mutagens that Derek predicts non-mutagenic
o These may be coincidental or contributory
features in this list are Misclassified
Lhasa Ames Test Reference Set contains Vitic, Hansen, FDA, ISSSTY, CGX, Marketed
Pharmaceuticals…
Current Derek Nexus
• Absence of alert returns ‘nothing to report’
Q
Match
alert or
example
N
Nothing to report
Y
Prediction with
supporting
details
•
•
•
•
•
•
•
•
Prediction
Likelihood
Substructure highlighted
Markush
Expert comments
Validation metrics
References
Examples
Next Release of Derek Nexus
• Assesses the query using the two expert-derived lists…
Match
alert or
example
Q
Y
N
• Query contains unclassified features?
• Query contains misclassified features?
Inactive
Inactive with
misclassified
features
Prediction with
supporting
details
•
•
•
•
•
•
•
•
Inactive with
unclassified
features
Prediction
Likelihood
Substructure highlighted
Markush
Expert comments
Validation metrics
References
Examples
Inactive with
unclassified &
misclassified
features
How are compounds classified?
• Positive predictions are unchanged
Nothing to report
Inactive
Inactive with
misclassified
features
Inactive with
unclassified
features
Inactive with
unclassified &
misclassified
features
Partition of public
data
1524
89%
176
10%
12
1%
1
-
Vitic
intermediates
464
94%
20
4%
9
2%
0
-
Private member
data 1
280
86%
29
9%
15
5%
1
-
Private member
data 2
372
89%
31
7%
13
3%
0
-
How accurate are these classifications?
Nothing to report
Inactive
Inactive with
misclassified
features
Inactive with
unclassified
features
Inactive with
unclassified &
misclassified
features
FN 132 / TN 1392
FN 165 / TN 11
FN 1 / TN 11
FN 1 / TN 0
91%
6%
92%
-
Vitic
intermediates
FN 65 / TN 399
FN 4 / TN 16
FN 3 / TN 6
FN 0 / TN 0
86%
80%
67%
-
Private member
data 1
FN 16 / TN 264
FN 4 / TN 25
FN 0 / TN 15
FN 0 / TN 1
94%
86%
100%
-
Private member
data 2
FN 47 / TN 325
FN 2 / TN 11
FN 2 / TN 29
FN 0 / TN 1
87%
86%
94%
-
Partition of public
data
FN = false negatives : TN = true negatives : Accuracy %
Example with an unclassified feature
 No alerts contain this system
 No examples in the Lhasa Ames Test Reference Set
• Highlights where to focus to increase confidence

Database search / proprietary data could alleviate concerns
inactive
Example 1 with a misclassified feature
Example 1 with a misclassified feature
• Review examples where the feature was seen in false negative
predictions
 No reason for an expert to over-rule the prediction of Derek
• Proceed with confidence in a negative prediction
Example 2 with a misclassified feature
Example 2 with a misclassified feature
• Review examples where the feature was seen in false negative
predictions
 No reason for an expert to over-rule the prediction of Derek
• Proceed with confidence in a negative prediction
Derek Nexus – negative predictions - recap
• Only for mutagenicity (Ames) alerts
• Providing additional direction to focus attention on
 which features are associated with uncertainty
• It is unusual for features to be highlighted.
 Less than 10% of the time against a number of datasets
• Accuracy remains high
 We are confident in making a negative prediction
Future development
• Will be driven by your feedback
 Features
 Supporting data
 Performance
• Negative prediction lists are closely coupled to a
particular version of the Knowledge Base and Vitic
 Unclassified and Misclassified features will change with each
release…
Negative Predictions
OUTLINE
• Impact of changes driven by M7
• Negative predictions in Derek for mutagenicity
 The science
 Performance
 Using it in practice…
• What further information / development would you like to see?
 Focus of the first workshop
Structure
2. Unambiguous
algorithm
3. Applicability
domain
5. Mechanism
1. Defined endpoint
Prediction
4. Performance
References
Likelihood
Key examples