1、Carcinogenicity prediction for Regulatory Use,Natalja Fjodorova Marjana Novi, Marjan Vrako, Marjan TuarNational institute of Chemistry, Ljubljana, Slovenia,Kemijske Dnevi 25-27 September 2008,UNIVERZA MARIBOR,Overview,1. EU project CAESAR aimed for development of QSAR models for prediction of toxico
2、logical properties of substances, used for regulatory purposes.2. The principles of validations of QSARs which will be used for chemical regulation.3. Carcinogenicity models using Counter Propagation Artificial Network,It is estimated that over 30000 industrial chemicals used in Europe require addit
3、ional safety testing to meet requirements of new chemical regulation REACH. If conducted on animals this testing would require the use of an extra 10-20 million animal experiments. Quantitative Structure Activity Relationships (QSAR) is one major prospect between alternative testing methods to be us
4、ed in a regulatory context.,aimed to develop (Q)SARs as non-animal alternative tools for the assessment of chemical toxicity under the REACH.,FR6- CAESAR European Project Computer Assisted Evaluation of Industrial chemical Substances According to Regulations,Coordinator- Emilio Benfenati- Istituto d
5、i Ricerche Farmacologiche “Mario Negri”,The general aim of CAESAR is,1. To produce QSAR models for toxicity prediction of chemical substances, to be used for regulatory purposes under REACH in a transparent manner by applying new and unique modelling and validation methods.,2. Reduce animal testing
6、and its associated costs, in accordance with Council Directive 86/609/EEC and Cosmetics Directive (Council Directive 2003/15/EC),CAESAR is solving several problems:,Ethical- save animal lifes; Economical- cost reduction on testing; Political- REACH implementation- new chemical legislation,CAESAR aim
7、ed to develop new (Q)SAR models for 5 end-points:Bioaccumulation (BCF), Skin sensitisationMutagenicity Carcinogenicity Teratogenicity,The characterization of the QSAR models follows the general scheme of 5 OECD principles:,A defined endpoint An unambiguous algorithm A defined domain of applicability
8、 Appropriate measures of goodness-of-fit, robustness and predictivity A mechanistic interpretation, if possible.,Principle1- A defined endpoint,Endpoint is the property or biological activity determined in experimental protocol, (OECDTest Guideline).Carcinogenicity is a defined endpoint addressed by
9、 an officially recognized test method (Method B.32 Carcinogenicity test Annex V to Directive 67/548/EEC).,Principle2- An unambiguous algorithm,Algorithm is the form of relationship between chemical structure and property or biological activity being modelled. Examples: 1. Statistically (regression)
10、based QSARs 2. Neural network model, which includes both learning process and prediction process.,Transparency in the (Q)SAR algorithm can be provided by means of the following information: a) Definition of the mathematical form of a QSAR model, or of the decision rule (e.g. in the case of a SAR) b)
11、 Definitions of all descriptors in the algorithm, and a description of their derivation c) Details of the training set used to develop the algorithm.,Principle3- A Defined Domain of Applicability,The definition of the Applicability Domain (AD) is based on the assumption that a model is capable of ma
12、king reliable predictions only within the structural, physicochemical and response space that is known from its training set. List of basic structures (for example, aniline, fluorene) The range of chemical descriptors values.,The assessment of model performance is sometimes called statistical valida
13、tion.,Principle4- Appropriate measures goodness-of-fit, robustness (internal performance) and predictivity (external performance),Principle5- A mechanistic interpretation, if possible,Mechanistic interpretation of (Q)SAR provides a ground for interaction and dialogue between model developer, and tox
14、icologists and regulators, and permits the integration of the (Q)SAR results into wider regulatory framework, where different types of evidence and data concur or compliment each other as a basis for making decisions and taking actions. Example: enhancing/inhibition the metabolic activation of subst
15、ances may be discussed.,National Institute of Chemistry in Ljubljana (NIC-LJU) is responsible for development of models for predicton of carcinogenicity,DATA ON CARCINOGENICITY,1.Studies of carcinogenicity in humans 2.Carcinogenicity studies in animals 3.Other relevant data additional evidence relat
16、ed to the possible carcinogenicity Genetic Toxicology Structure-Activity Comparisons Pharmacokinetics and Metabolism Pathology,Cancer Risk Assessment IARC International Agency for Research of Cancer,Predictive Toxicology Approaches,1. Quantitative models (QSARs) Continuous data prediction on the bas
17、is of experimental evidence of rodent carcinogenic potential (TD50 tumorgenic dose)2. Categorical models based on YES/NO data. (P-positive; NP-not positive),Dataset:,805 chemicals were filtered from 1481compounds taken from Distributed Structure-Searchable Toxicity (DSSTox) Public Database Network h
18、ttp:/www.epa.gov/ncct/dsstox/sdf_cpdbas.html which was derived from the Lois Gold Carcinogenic Database (CPDBAS)The chemicals involved in the study belong to different chemical classes, (noncongeneric substances),Descriptors:,252 MDL descriptors were calculated in program MDL QSAR.2. Descriptors dat
19、aset was reduced to 27 MDL descriptors, using Kohonen map and Principle Component Analisis.,Counter Propagation Artificial Neural Network,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:
20、 prediction of the four-dementional target (toxicity) Ts,Investigation of quantitative models shows us low results RESPONCE- TD50mmol,Correlation coefficient in the external validation is lower then 0.5,Continuouse data models (Quantitative models),Investigation of categorical models shows us satisf
21、actory results,YES/NO principeRESPONCE: P-positive-active NP-not positive-inactive,Characteristics used for validation of categorical model,true positive(TP), true negative (TN) Accuracy(AC), AC=(TN+TP)/(TN+TP+FN+FP) TPrate=Sensitivity(SE)=TP/(TP+FN) TNrate=Specificity(SP)=TN/(TN+FP),Categorical mod
22、el for dataset 805 chemicals (Training=644 and Test=161), using 27 MDL descriptors,Confusion matrix TR(644)/TE(161) classes (Positive- Negative),FP,FN,TP,TN,How we find optimal model, using threshold,Threshold=0.45 Accuracy=0.68 SE=0.73 SP=0.63,Changing of threshold allows us to get models with diff
23、erent statistical performances.,ROC(Receiver operating characteristic) curve,Training set,Test set,The area under the curve is 0.988 and 0.699 in the training and test sets, respectively.,How requrements of REACH reflect development of models,To focus model to high sensitivity in prediction of carci
24、nogenicity From regulatory perspective, the higher sensitivity in predicting carcinogens is more desirable than high specificity Sensitivity- percentage of correct predictions of carcinogens Specificity- percentage of correct predictions of non-carcinogens,Conclusion,1.We have bult the carcinogenici
25、ty models in accordance with 5 OECD principles principle of validation 2. We have got satisfactory results for categorical models with accuracy 68% which is good for carcinogenicity as it meet the level of uncertanty of test data. 3. The goal of our future investigation will be dedicated to research
26、 of relationship between results of carcinogenicity tests and presence of Genotoxic, non Genotoxic alerts using TOX TREE program.,Acknowledgements,The financial support of the European Union through CAESAR project (SSPI-022674) as well as of the Slovenian Ministry of Higher Education, Science and Technology (grant P1-017) is gratefully acknowledged.,THANK YOU,
