chemical - Société Française de Toxicologie

Le Read Across et QSAR
EMILIO BENFENATI1, Rodolfo Gonella Diaza1,Giuseppa Raitano1,
Enrico Mombelli2, Antonio Cassano2
1Istituto
di Ricerche Farmacologiche MARIO NEGRI
Laboratory of Environmental Chemistry and Toxicology
2INERIS - Institut National de l'Environnement Industriel et des riSques
Unité Modèles pour l'Ecotoxicologie et la Toxicologie (METO)
Congrès Annuel de la Société Française de Toxicologie
« Les Nouvelles Approches en Evaluation du Risque :
TTC, MOA, MOS, MOE, WOE, QSAR, et Read Across »
Paris, les 14 & 15 novembre 2013
The Nobel Prize in Chemistry 2013
The Nobel Prize in Chemistry 2013 has gone to Michael Levitt, Martin Karplus
and Arieh Warshel, who “took the chemical experiments into cyberspace”
1
COMPARISON
• A case-to-case basis
Read-across is a method of filling in data
gaps for a substance by using surrogate
data from another substance.
• Reliability supported by specific
explanation
• Supporting data needed (generic and/or
substance-specific)
• Subjective expert assessment
• Pre-built model
(Q)SARs are computer based models
designed to predict properties from
knowledge of the chemical structure.
• Reliability supported by the applicability
domain
• Supporting examples from training sets
• Objective output (though it requires an
evaluation by the expert)
2
ANTARES
Evaluating the existence and suitability of Non-Testing Methods for
REACH
Alternative Non-Testing methods
Assessed for REACH Substances
Contract LIFE08 ENV/IT/000435
www.antares-life.eu
3
ANTARES Contribution to ANNEX XI
According to REACH Regulation (Annex XI)
a QSAR Model is VALID IF
1.
the model is recognized scientifically valid;
• ANTARES contributed to assess
model’s validity
2.
the substance is included in the applicability
domain of the model;
• ANTARES provided results per
chemical classes and MoA
• VEGA improved ADI
3.
results are adequate for classification and
labelling and for risk assessment;
• VEGA introduced safety margin
• Evaluation done in regression
and classification
4.
adequate documentation of the methods
is provided.
• VEGA provided material (figures,
framments, guidance to expert)
4
ANTARES
HOME
www.antares-life.eu
5
ANTARES
MODELS PAGE – Specific Properties
www.antares-life.eu/software.php
6
FOCUS ON 8 ENDPOINTS
Mutagenicity (Ames)
Carcinogenicity
LD50
HUMAN TOXICITY
Fish Acute Toxicity
Daphnia Acute Toxicity
ECOTOXICOLOGY
BCF
Ready Biodegradability
ENVIRONMENTAL
Water Solubility
PHYSICO-CHEMICAL
www.antares-life.eu/software.php
7
MUTAGENICITY: Performance
Total dataset (6065 compounds)
1,00
0,90
0,80
0,70
0,60
Accuracy
0,50
Sensitivity
0,40
Specificity
0,30
0,20
0,10
0,00
ACD
T.E.S.T.
Topkat
CAESAR
Sarpy
Derek
Toxtree
ADMET
The first 4 models showed the best accuracy values very
close to the in vitro reproducibility of Ames test (0.85)
8
MUTAGENICITY: Performance
In & out train chemicals and in & out Applicability Domain
Accuracy
1
0,9
0,8
0,7
0,6
0,5
0,4
0,3
0,2
0,1
0
Excluding compounds in the training
set: T.E.S.T. and CAESAR gave the
highest accuracy. There is a decrease
in the predictive performance
considering molecules out of training
set of the models.
in train
out train
CAESAR
ACD
T.E.S.T.
SARpy
ADMET
An increase in the performance
was seen after selecting the
compounds inside the
Applicability Domain for each
model.
Accuracy
1
0,9
0,8
0,7
0,6
0,5
0,4
0,3
0,2
0,1
0
out AD
in AD
CAESAR
SARpy
ACD
T.E.S.T.
Topkat
ADMET
9
MUTAGENICITY: Performance
Chemicals out of train distributed by AD
Accuracy
1
0,9
0,8
0,7
0,6
0,5
out AD
0,4
in AD
0,3
0,2
0,1
0
CAESAR
SARpy
ACD
T.E.S.T.
ADMET
Applying the information on the
applicability domain improves results.
For compounds out of training and within AD,
CAESAR and SARpy gave the highest sensitivity.
10
VEGA
www.vega-qsar.eu
13
VEGA and the APPLICABILITY DOMAIN
The different checks done by VEGA for the definition of the Applicability Domain
Index
•
Visualisation of similar substances
•
Similarity index (chemical; sub-indices)
•
Chemiometric check (descriptor space)
•
Atom centred-fragment (chemical)
•
Check of the descriptor sensitivity (algorithm)
•
Uncertainty (algorithm)
•
Fragments for outliers (output space)
•
Prediction Accuracy (output space)
•
Prediction Concordance (tox exploration)
www.vega-qsar.eu
14
APPLICABILITY DOMAIN INDEX
How the ADI information is visualized
The Applicability Domain
Index is summarized in one
value, in top of the table of
the Prediction Summary
All the measured
components contributing to
the AD global index are
shown for an easy
visualization of some
potentially critical aspects.
www.vega-qsar.eu
15
SOME EXAMPLES
27
SOME EXAMPLES
CAESAR
SARpy
ToxTree
SOME EXAMPLES
MUTAGENS
29
ADEQUACY OF A MODEL
This chart shows (for BCF case) the predicted value
together with its conservative
confidence interval for safe classification
Compound safely classified
as not bioaccumulative (<3.3)
1
SAFETY MARGIN
2.2 + 0.6 l.u.
2
B
vB
3.3 l.u.
3.7 l.u.
threshold
threshold
3
Predicted
logBCF
logBCF
2.2 l.u.
No. Comp. = 492
Pred.
logBCF
Exp. logBCF
nB
B/vB
nB
359
0
B/vB
60
73
VEGA shows not only the predicted value (2.2 l.u.) but also its
uncertainty, and how far it is from the threshold (3.3 l.u. for logBCF).
The safety margin (2.2 l.u. plus a conservative interval of 0.6 l.u.) is
calculated specifically for each chemical, considering the ADI of the
specific compound. In addition, it is determined in a way to provide
no false negative prediction.
The confusion matrix verified on a set of 492 compounds is shown.
www.vega-qsar.eu
16
SUPPORTING DOCUMENTATION
VEGA provides additional material to support the prediction: DETAIL
ON MOLECULAR DESCRIPTORS
logBCF
In this example the experimental logBCF
values versus the predicted logP values for
the chemicals in the training set of the
model are shown, as blue dots.
In red we can see the predicted logBCF and
logP values of the target compound.
The user can evaluate if the predicted
values are within the typical trend of the
compounds, or if an unusual behaviour
appears.
MlogP
www.vega-qsar.eu
17
CALEIDOS STARTED WHERE ANTARES ENDED
http://www.antares-life.eu/
IT ADDRESSED THE OVERALL PERFORMANCE OF QSAR METHODS AND IDENTIFIED
RELIABLE QSAR MODELS USING GOOD QUALITY DATASETS.
http://www.caleidos-life.eu/
CALEIDOS ADDRESSES THE REGISTERED DATA.
11
From ANTARES and CALEIDOS to VEGA
Identification of the BEST MODELS
Characterization of the AD
Integration of DIFFERENT MODELS
Implementation into a UNIQUE PLATFORM
Integration with READ ACROSS
12
CALEIDOS ENDPOINTS
PHYSICO-CHEMICAL PROPERTY
FATE and BEHAVIOUR
LogP
BCF
ECOTOXICOLOGICAL PROPERTY
Fish acute toxicity
HUMAN TOXICOLOGICAL
PROPERTY
Mutagenicity (Ames test)
Carcinogenicity
Reprotoxicity
18
DATA PRUNING OF REACH DOSSIERS : MUTAGENICITY
Starting point: 27144 studies on 2975 unique CAS-RN
After filtering: 388 unique CAS-RN
Main features of the data selected for the mutagenicity endpoint:
Reliability: “1” (Klimisch’s code); Composition: “mono constituent
substance” & Origin: “organic”
Metabolic activation: “with and without”; Degree of purity
91 % of chemicals non mutagen
60 % of chemicals without alert (OECD Toolbox)
19
Global STATISTICS of the Models
CAESAR
SARPY
TTVEGA
TTIDEA
T.E.S.T.
MUTA
OASIS DB
OECD DB
EVALUATION SET A (766)
Predicted
766
766
766
766
755
765
766
ACC
0.75
0.75
0.77
0.78
0.78
0.78
0.67
SE
0.69
0.53
0.60
0.63
0.43
0.59
0.65
SP
0.77
0.82
0.82
0.82
0.88
0.83
0.68
MCC
0.41
0.33
0.40
0.42
0.34
0.41
0.28
EVALUATION SET B (388)
Predicted
388
388
388
388
384
388
388
ACC
0.76
0.79
0.79
0.80
0.86
0.82
0.66
SE
0.65
0.56
0.61
0.65
0.42
0.59
0.65
SP
0.77
0.81
0.81
0.81
0.90
0.84
0.67
MCC
0.27
0.25
0.31
0.31
0.27
0.31
0.18
20
The VEGA consensus model for mutagenicity
CAESAR
±AD 2
INPUT
SARPY
TOXTREE
±AD 1
WOE
CONSENSUS
MODEL
OUTPUT
±AD 3
1. Equation:
2. The prediction with greater AD is chosen – different thresholds AD were considered
21
The VEGA consensus model for mutagenicity
VEGA MODELS
CAESAR
SARPY
TTVEGA
CONSENSUS 1
equation
CONSENSUS 2
AD>
EVALUATION SET A (766)
Predicted
766
766
766
766
766
ACC
0.75
0.75
0.77
0.81
0.81
SE
0.69
0.53
0.60
0.68
0.67
SP
0.77
0.82
0.82
0.85
0.85
MCC
0.41
0.33
0.40
0.50
0.49
EVALUATION SET B (388)
Predicted
388
388
388
388
388
ACC
0.76
0.79
0.79
0.83
0.82
SE
0.65
0.56
0.61
0.63
0.61
SP
0.77
0.81
0.81
0.85
0.85
MCC
0.27
0.25
0.31
0.37
0.35
22
Chemical classes
• 6065 chemicals
• 56% mutagen
• 44% nonmutagen
Antares
Chemical
classes
• Descriptors
• Functional
groups
• Statistical
methods
• Mode of action
• Expert’s work
New rules
Greater
precision in
the old rules
23
SOME EXAMPLES from set of 6065 chemicals
Chemical
classes/SA
Aromatic
Amines
SA_28: primary
aromatic amine,
hydroxyl amine
and its derived
esters
SA_28bis:
Aromatic monoand dialkylamine
SA_28ter:
aromatic N-acyl
amine
N. Of
chemicals
fired
N. Of
mutagens
True
Positives
Rate
EXCEPTION RULES
1043
718
68,84%
Many exception rules
79.52%
• Chemicals with ortho-disubstitution, or with an ortho
carboxylic acid substituent are excluded.
• Chemicals with a sulfonic acid group (-SO3H) on the
same ring of the amino group are excluded .
54,55%
• Chemicals with ortho-disubstitution, or with an ortho
carboxylic acid substituent are excluded.
• Chemicals with a sulfonic acid group (-SO3H) on the
same ring of the amino group are excluded .
83
11
14
66
6
12
85.71%
• Chemicals with ortho-disubstitution, or with an ortho
carboxylic acid substituent are excluded.
• Chemicals with a sulfonic acid group (-SO3H) on the
same ring of the amino group are excluded .
Exclusion rule for Aromatic Amines
NH2
If -SO3H on the same ring, in meta or in para, of the amino
group the chemical is non-mutagen.
If -SO3H on the same ring, in orto, the chemical is mutagen
N. Of chemicals fired: 21
N. Of mutagens: 7
True Positives Rate: 33,3%
S
OH
O
O
O
Some examples
Mutagen
CH 3
Cl
O
HO
H2N
HO
OH
H 2N
Non-Mutagen
HO
S
O
O
O
S
O
O
S
O
OH
S
OH
O
NH2
S
O
OH
NH2
NH 2
25
Conclusions
Advantages in the use of multiple
QSAR models (consensus)
Advantages in the combined use of
QSAR and Read Across for the
assessment of the individual chemical
Advantages in the combined use of
QSAR and Read Across for automatic
assessment of set of chemicals
26
JUNE 16 - 20, 2014 – Milan, ITALY
SELECTED TOPICS
PERSPECTIVES ON:
QSAR &:
• Chemoinformatics and software
solutions for modelling
• Toxicology data, curation and mining
• QSAR & Adverse Outcome Pathways
• ADME, PBPK & predictive toxicology
• Chronic toxicity
• Environmental toxicity and fate
•
•
•
•
•
•
REACH
Cosmetics
Food safety
Pesticides
Nanomaterials
Drugs
VOICE FROM:
• Industry
• Regulators
• Scientistis