nuovi modelli in silico - sitox.org · pdf filenuovi modelli in silico ... genotoxic sas ashby...
TRANSCRIPT
NUOVI MODELLI IN SILICO UTILI PER LA VALUTAZIONE DELLA
TOSSICITÀ SISTEMICA
Nuovi approcci in silico e strategie integrate nella valutazione del rischio
Milano, 27 maggio 2015
Alessandra Roncaglioni [email protected]
IRCCS - Istituto di Ricerche Farmacologiche “Mario Negri”
• Mutagenicity
• Carcinogenicity
• Developmental toxicity
• Repeated dose toxicity
Addressed endpoint
MUTAGENICITY
Mutagenicity (Ames test) Detection of mutations that restore the functional capability of the bacteria to synthesize an essential amino acid. Positive results indicate that a substance induces point mutations by base substitutions or frameshifts in Salmonella typhimurium and/or Escherichia coli.
In vitro bacterial mutagenesis
Ames mutagenicity is also one of the most modeled endpoints by predictive computational methods.
A critical component for the assessment of carcinogenesis
Approaches to model mutagenicity
QSAR models of noncongeneric compounds to predict
mutagenicity can use TWO APPROACHES:
1 : STRUCTURAL ALERTS
2 : STATISTICS
Classification models
Expert knowledge codified in SA • All available scientific
information for a specific toxicological endpoint collected, evaluated, and weighted
• A generalized relationship between structural moieties (alerts) and biological activity identified
• Structural information translated into a knowledge base by human experts
• Prediction of new chemical entitiesis based on the existence of alerts
Ashby, 1985
Overlapping: lists of SA for Ames
Benigni /Bossa 28 (46)
genotoxic SAs
Ashby 19 SAs
Bailey 33 SAs
LMC 17 SAs
Kazius 29 SAs
ToxTree – genotoxicity
Genotoxic SA (predictivity for mutagenesis)
ToxTree – cancerogenesis
CAESAR model for mutagenicity
Dataset
• Extracted by Kazius-Bursi Mutagenicity Dataset (Kazius et al. J Med Chem, 2005), containing 4225 compounds, 2358 classified as mutagens and 1867 classified as non-mutagens by Ames test
• For validation, the dataset has been divided into training (80%) and test (20%) sets
Descriptors
• 2D descriptors: MDL software
Models
• Classification: SVM (Support Vector Machines)
• 10 fold cross-validation
CAESAR integrated model
Support Vector Machines (basic model)
ToxTree 1st checkpoint
ToxTree 2nd checkpoint
if negative…
if negative…
if negative…
positives
positives
positives
MUTAGENIC
MUTAGENIC
NON-MUTAGENIC
3 STEPS IN CASCADE:
statistical model (based on chemical descriptors)
knowledge-based
filter (based on structural alerts)
The tested models Software Method Free Applicability
domain
ADMET predictor (Simulation Plus v
6.06.0007) Artificial Neural Network Ensembles No Outsider of the
sub-models
T.E.S.T v 4.0.1 Consensus method: average of 3 models: NN, FDA and Hierarchical clustering Yes
The model processes only the compounds inside
AD
TOPKAT (Accelerys discovery studio v
3.1)
QSAR statistical method on 2D descriptors (e-state, topological) No Optimum
Predictive Space
Tox Suite (ACD Labs v 2.95)
Statistical model employing binominal PLS, predefined set of fragmental descriptors, local
correction to baseline, using experimental data for similar compounds
No Reliability index
VEGA Caesar v 2.1.10 QSAR statistical model based on Artificial Neural Network + rule based model Yes Quantitative AD
measurement
VEGA SARpy v 1.0.5-Beta Structural Alerts data mining Yes Quantitative AD
measurement
Derek Nexus v 2.0 Collection of knowledge - based rules No Not applicable
Toxtree v 2.5.0 Collection of knowledge - based (SA) Yes Not applicable
Benchmark Data set for in silico prediction of Ames mutagenicity, Hansen et al., 6512 compounds
Evaluation of models
Statistical Models
Knowledge Based
Models
Statistical Rule Based
Models
FRAGMENTATION
EVALUATION
RULESET EXTRACTION +
VALIDATION
PREDICTOR
Data mining: SARpy steps
SMILES ACTIVITY
O=CC 1
O/N=C/C 0
O=C(N)C 1
… …
NEW MOLECULE
PREDICTION
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
ACD CAESAR Toxtree T.E.S.T. SARpy Topkat Derek ADMET
Sensitivity
Specificity
Accuracy
Performance on the whole Dataset
6065 chemicals
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
CAESAR SARpy ACD T.E.S.T. ADMET
Accuracy
out AD
in AD
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
CAESAR SARpy ACD T.E.S.T. ADMET
Sensitivity
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
CAESAR SARpy ACD T.E.S.T. ADMET
Specificity
Performance out of training in/out AD
Evaluation on REACH compounds
27,144 studies corresponding to 2,975 unique CAS RN
START
Keep only data respecting OECD 471 and eliminate UVCB/inorganics, false mono constituents and ambiguous (neither positive or negative)
results
765 CAS RN 170 P and 595 N
Keep Klimisch codes = 1 & 2
Results for chemicals in AD
MCC
0.65 0.65 0.71 0.34 0.78
75%
94%
97%
75%
57%
83%
Without AD!
free models
Reminder : “Consensus VEGA + MN rules” (765 chemicals) = 0.82, 0.74, 0.84, MCC = 0.53
commercial models
Results for chemicals in AD
• Among systemic toxicity endpoints mutagenicity is one of the endpoints with greater success in the modelling
• Qualitative assessment, primarily focused on in vitro bacterial reverse test
• Used for the assessment of potential genotoxic impurities in pharmaceuticals
Mutagenicity conclusions
CARCINOGENICITY
Data to assess POTENTIAL CARCINOGEN:
Animal carcinogenicity studies (rat, mouse)
Epidemiological studies
Studies of genotoxicity in short-term in-vitro and in-
vivo
QSAR, read-across, general toxicity (subchronic) and specific (organ toxicity), toxicokinetics, pathology
Carcinogenicity
Berkeley Carcinogenic Potency Database (CPDB)
> 1300 chemical structures
Data fromTechnical Reports of NCI/NTP (rodent assays and data from peer reviewed sources)
Selection of 805 compounds (only male and female rats, no other species have been taken into account)
TD50
CAESAR model
Counter Propagation ANN
Step1: mapping of molecule Xs
(vector representing structure)
into the Kohonen layer
Step2: correction of weights in both, the Kohonen and the Output layer
Step3: prediction of the four-dementional target (toxicity) Ts=carcinogenicity
Model Training set Test set
Total compounds (number) 644 161
Accuracy,% 91 73
Cross-validation (leave 20% out), % 66
False Positive (FP) (number) 44 22
False Positive Rate, % 14 31
False Negative (FN) (number) 13 22
False Negative Rate, % 4 25
Positive Predictive Value (PPV) (precision), %
88 75
Negative Predictive Value (NPV), % 95 69
Sensitivity (Carcinogen), % 96 75
Specificity (Non-Carcinogen), % 86 69
CAESAR statistics
Caesar v 1.0.0.6
Free
QSAR statistical model based on CP-ANN
Rat (both sexes) yes
Lazar Free
K-nearest neighbours, escluding
identical compound Rat (both sexes)
yes
MultiCASE AF1 Module
Commercial Statistical method, fragment based +
modulating factors Male rat yes
Topkat Commercial Qsar statistical method on 2D
descriptors
Rat (both sexes)
yes
Toxtree V 1.60
Free
Structural Alerts interpretation
Studies in different
species
Not based on training set
Derek for Windows DfW_11.0.0
Commercial Collection of knowledge - based rules Studies in different
species
Not based on training
set
Hazard Expert Module in Pallas 3.0
Commercial Collection of Structural Alerts +
modulating factors for bioavailability
Studies in different
species
Not based on training set
software availability method Input basis Info training set
Software/tested models
1544 (805+739) chemicals with carcinogenicity values towards rat (males or females)
Berkeley Carcinogenic Potency Database (CPDB)
FDA 2009 SAR Carcinogenicity – SAR Structures
(from the Leadscope database)
Public; More than 1300 chemical structures
Commercial database containing 2090 compounds
Data extracted from the Technical Reports of the NCI / NTP concerning rodent essays and data extracted from multiple sources subject to review.
Information about chemical structures and carcinogen activity based on different species of mammalians
Assays data from other species as well as rat and mouse and includes a wide variety of experimental protocols; selected 805 compounds related to rats (male and female)
Compounds in common with CPDB (655) evaluated to determine the suitability of the experimental class of carcinogenicity assigned from the two sources (concordance 95%)
Quantitative measure of the carcinogenic potential (TD50)
Compounds already present in the CPDB excluded; selected 739 compounds with experimental data on rats reliable
Dataset
0.81 0.87
0.57
0.89
0.72
0.61
0.52
0.43
Topkat CAESAR LAZAR MultiCase
Sensitivity
In DB
Out DB
0.68
0.82
0.73
0.89
0.41
0.58
0.8 0.78
Topkat CAESAR LAZAR MultiCase
Specificity
In DB
Out DB
0.74
0.85
0.64
0.89
0.57 0.6
0.67 0.6
Topkat CAESAR LAZAR MultiCase
Accuracy
In DB
Out DB
Performance of the models for the training (In DB) and test sets (Out DB)
Results
Percentage of matched predictions stratified by mechanism of carcinogenicity
Mechanisms of carcinogenicity
TOXTREE HAZARDEXPERT DEREK LAZAR CAESAR TOPKAT
Acylating direct acting
77,8 55,6 66,7 44,4 77,8 44,4
Alkylating direct acting
58,6 53,7 58,6 61,5 70,9 57,8
Alkylating indirect acting
79,2 70,7 78,8 75,7 83,0 67,6
Intercalating and DNA adduct
forming Indirect acting
68,0 68,9 76,7 45,6 70,9 54,4
Aminoaryl DNA adducts forming Indirect acting
64,8 60,6 65,4 63,5 74,3 58,7
Non genotoxic 41,6 56,2 65,2 64,0 71,9 59,6
No Alerts 65,2 63,4 63,1 64,9 70,5 49,5
MoA information
• Complex endpoint, data uncertainty
• Advantage of the use of mutagenicity information
• Combination with read-across in a weight of evidence approach would increase reliability
Carcinogenicity conclusions
DEVELOPMENTAL TOXICITY
Targeted endpoints
Sexual maturation
Release of gametes
Fertilisation
Transport of the zygote
Implantation
Embryogenesis
Foetogenesis
Birth
Postnatal development
Growth and development
ReproTox (fertility)
DevTox
- Different effects: • death of the developing organism(fetal death, pre/post-
implantation loss) • structural abnormality • altered growth (growth retardation, weight variation) • functional deficiency
- Confounding factors (maternal toxicity)
Developmental toxicity is a complex endpoint because:
DevTox – issues (1)
- Different tests • Prenatal developmental toxicity test [OECD TG 414] • Reproductive/Developmental toxicity screening assay
[OECD TG 421] • Two-generation reproduction toxicity study [OECD TG 416] • Extended one-generation reproductive toxicity study
[OECD TG 443]
- Existing QSAR models are mainly classifiers (POS/NEG)
• Translation of NOAEL/LOAEL values in classification scheme (POS/NEG)
• Classification based on different lines of evidence (risk-based, epidemiological findings)
DevTox – issues (2)
Model Type of model
CAESAR Descriptor-based
SARpy Fragment-based, extracted by the
software
DART (from P&G) Categories and
substituents identified by experts
DevTox – freely available models
Model Type of model
LeadScope Model Applier Fragment & descriptor based
statistical software
MultiCASE Rules extracted by the
software
Derek Nexus Rules identified by experts
Kindly provided by S. Kulkarni and T. Barton-Maclaren Health Canada
DevTox – commercial models
FDA pregnancy categories
A No risk in controlled human studies
B No risk in other studies: Negative animal studies, no human studies OR Positive animal studies but negative human studies
C Risk not ruled out: Positive animal studies, no human studies, but risk-benefit analysis may warrant use of the drug in pregnant women
D Positive evidence of risk: Positive evidence of effects in human studies, but risk-benefit analysis may warrant use of the drug in pregnant women
X Contraindicated in Pregnancy: Animal OR human studies show abnormalities AND/OR evidence of fetal risk based on human experience, risks for pregnant women clearly outweigh potential benefits
Negative n = 91
Positive n = 201
Random forest (RF) model based on 13 2D-descriptors
CAESAR model
P&G dataset
total compounds 672
positive compounds 615
negative compounds 57
Library of about 185000 positive
chemicals
P&G model
main symptoms sub symptom
fetal growth
fetal growth
retardation
fetal weight decrease
fetal survival
fetal death
post-implantation lost
pre-implantation lost
structural
dysmorphogenesis
visceral organ toxicity
LeadScope dataset
total compounds 1320
positive compounds 692
negative compounds 628
LeadScope dataset classification
Two steps model
First step: rules with high accuracy
Positive fragment 53
Negative fragment 51
Second step: rules with low accuracy
Positive fragment 77
Negative fragment 77
Compounds not predicted
SARpy model
Rules comparison - DART
ReproTox DevTox
• Devtox is a difficult endpoint to be estimated for the high number of sub-endpoints, relatively low number of chemicals with experimental data, classification of data and heterogeneity of source information
• Few models exists but so far without satisfactory performance
DevTox conclusions
REPEATED DOSE TOXICITY
1. Repeated dose toxicity approaches towards LO(A)EL
and NO(A)EL
2. k-nearest neighbors (k-NN) approach
3. Chemical categories for hepatotoxicity and
nephrotoxicity
4. Structural alerts for hepatotoxicity and
nephrotoxicity
Quantitative
Qualitative
General overview
This study provides information on:
• adverse effects on organs and tissues
• lowest observed (adverse) effect level (LO(A)EL) and no observed (adverse)
effect level (NO(A)EL)
Repeated Dose Toxicity
Lowest Observed (Adverse) Effect Level (LO(A)EL)
the lowest dosage where (adverse) effects can be observed
No Observed (Adverse) Effect Level (NO(A)EL)
It is the highest dosage for which no (adverse) effects can be observed
20 mg/kg/die 40 mg/kg/die 60 mg/kg/die 80 mg/kg/die
NO(A)EL LO(A)EL
Control
The NO(A)EL and LO(A)EL should be expressed as mg/kg body weight/day.
Many different effects (steatosis, seizure, hematuria etc…) on organs and
tissues (liver, kidney, blood, CNS etc…)
Experimental design: different species used, strain used, spacing between
doses, exposure time, exposure route, variability of the responses between
animals within the dose groups, the definition of “adversity” of an effect,
specific life stages differences between animals…
NO(A)EL and LO(A)EL values are not derived from a dose-response curve
Uncertainty and variability of Repeated Dose Toxicity data
*In the calculation of the MoS, the lowest obtained LO(A)EL value is used when a NO(A)EL is
not available (SCCS, 2012).
Munro database
613 compounds
HESS database
502 compounds
IRIS’EPA database
557 compounds
Final dataset
255 compounds
(138 HESS, 99
Munro, 18 IRIS)
84-98 days
Oral exposure (gavage, feed,
drinking water)
Rats studies (interspecies
difference)
Organic compounds
When more data were
available for the same
compound: used the lowest
K-NN:Training set
LO(A)EL data
Final dataset 179 compounds
84-98 days
Oral exposure (gavage, feed,
drinking water)
Rats studies
Organic compounds
Rejected data already present
in the training set
Fraunhofer RepDose
database
Over 650 compounds
K-NN:Validation set
LO(A)EL data
Refined algorithm K-NN: quantitative approach
Target compd
k most similar
compds
SI ≥ 0.85
Δ exp values < 1
SI ≥ 0.90
Error
in prediction of
similar compds
< 1
Similar
compds have another
similar compd
Cross-validation
of most similar
compds
Not
predicted
Prediction as
average of exp
values
Prediction as
average of exp
values
Not
predicted
Not
predicted
Not
predicted
Not
predicted
n = 1 n = 0
n > 2
Non predicted Predicted
Yes
Yes
Yes
Yes
No
No
No
No
k 2 3 4 5 6
q2 0.699 0.778 0.797 0.793 0.792
RMSE 0.451 0.401 0.388 0.392 0.398
n 68 / 255 52 / 255 51 / 255 51 / 255 49 / 255
% 27% 20% 20% 20% 19%
Training set (255 compounds)
k = 2 k = 3 k = 4
k = 5 k = 6
K-NN: quantitative approach
k 2 3 4 5 6
r2 0.534 0.549 0.551 0.672 0.679
RMSE 0.623 0.663 0.658 0.632 0.647
n 27 / 179 24 / 179 23 /179 21 / 179 20 /179
% 15% 13% 13% 12% 11%
Validation set (179 compounds)
k = 2 k = 3 k = 4
k = 5 k = 6
K-NN: quantitative approach
The algorithm was developed by using only freely available tool
(IstSimilarity) and can be easily implemented in VEGA
platform
The models could be easily improved by simply updating the
model dataset, without any need of new training procedures
The real model’s predictivity power was evaluated using an
external dataset
The models provide reliable predictions only for few chemicals.
K-NN: quantitative approach
LO(A)EL
Chemical categories and structural alerts: qualitative approach
Organ toxicity
liver
kidney
Chemical categories and structural alerts: qualitative approach
Hepatotoxicity
Nephrotoxicity
Hepatotoxicity dataset (218 compounds)
Nephrotoxicity dataset (200 compounds)
a) effects on LIVER at the LOAEL ACTIVE 121 compds
INACTIVE 97 compds b) effects on other organs than
liver at the LOAEL
a) effects on KIDNEY at the
LOAEL ACTIVE 87 compds
INACTIVE 113 compds b) effects on other organs than
kidney at the LOAEL
If a compound reported effects both on kidney and liver: included in both dataset
HESS database (OECD Toolbox)
JAVA application based on the VEGA core libraries
(http://vega-qsar.eu) that performs the analysis of a given dataset of
molecules on the basis of a set of relevant chemical features (functional
groups and atom centered fragments).
INPUT: dataset where each molecule is assigned to a class of toxicity
(ACTIVE-INACTIVE)
OUTPUT: statistics for each relevant feature, with the n. of matching and
the % for each user defined class
IstChemFeat-Chemical categories
The application is released under the GPLv3 open source license. The application has
been developed under the EC funded project CALEIDOS (http://www.caleidos-
life.eu/) by Kode s.r.l. (http://kode-solutions.net).
IstChemFeat-Chemical categories
Chemical category Occurences
n. TP n. % TP
Aromatic Nitriles 4 4 100
Sulfonamides 8 7 87
Aromatic 2° amines 7 6 86
Aromatic ketones 6 5 83
Esters (aromatic and aliphatic) 9 6 67
Chemical category Occurences
n. TP n. % TP
Aromatic ethers 14 11 79
Aromatic hydroxyl 22 17 77
3° amines (aromatic and aliphatic) 8 6 75
Aliphatic esters 7 5 71
Aromatic ketones 10 7 70
Aromatic Carboxylic acids 3 2 67
Furanes 6 4 67
Aromatic nitro group 18 12 67
Halobenzenes 16 10 62
Hepatotoxicity
Nephrotoxicity
IstChemFeat-Chemical categories
Halobenzenes Cyp 450 Cyp 450 Oxidation
Convalently
bound to
macromolecules
Liver damage
Chan et al., 2007, Sakuratani et al., 2013
Sulfonamides Metabolites, N-acetyl
derivatives
Low solubility
excreted in large amounts
in the urine
Intratubular crystal
deposition
Renal damage
Parazzella, 1999
A: any atoms except H
AH: any atoms including H
X: halogens
02
All compounds
(active-inactive)
SARpy rules
extraction
Human expert rules
analysis
STEP 1
65% TP
LR ≥ 2
Ruleset procedure
STEP 2
Human expert rules
extraction
Literature
information and
chemical categories
(IstChemFeat)
F
i
n
a
l
R
u
l
e
s
e
t
SARpy-Hepatotoxicity alerts
8 7 86
19 16 84
8 6 75
8 6 75
Structures Tot TP %TP
16 11 69
53 36 68
11 7 64
Structures Tot TP %TP
AH: any atoms including H
SARpy fragments Human-based fragments
Halobenzenes,
at least 2 X
SARpy-Nephrotoxicity alerts
4 4 100
14 12 86
7 5 71
17 12 70
9 6 67
Structures Tot TP %TP
4 4 100
Structures Tot TP %TP
SARpy fragments
Human-based fragments
10 7 70
AH: any atoms including H
SARpy alerts
IstChemFeat
sulphanilamide
4-hydroxylaminobenzenesulfonamide
(4-HABSA)
Severe renal damage
(Interstitial nephritis,
polyarteritis, intratubular
crystal deposition)
Eyanagi et al., 1985; Parazzella, 1999, Mustafa et al., 2014
Ahmed et al., 2007
naphtalene
decrease in GSH levels, hepatic
lipid peroxidation,
increase of ALT, AST and AP
Liver injury
oxidative damage,
cholestasis,
DNA fragmentation
AH: any atoms including H
Sulfonamides
This approach focuses not on whole body toxicity but on single
organ and tissue toxicity (liver and kidney)
Structural alerts and chemical categories
The alerts and the chemical categories are not already tested with
external data
For some alerts and chemical categories it is possible to find a
mechanistic explanation in the literature
These SA and chemical categories will be implemented in VEGA for
predicting renal and liver toxicity and they will be implemented
also in the new VEGA read across-tool
• Models highly affected by the study design
• Quantitative models capable to deal with a low number of compounds
• Category approach/SA can support read-across and weight of evidence
Repeated dose conclusions