Ž . Chemometrics and Intelligent Laboratory Systems 43 1998 135–145 Hybrid toxicology expert system: architecture and implementation of a multi-domain hybrid expert system for toxicology Giuseppina Gini a, ) , Vito Testaguzza a , Emilio Benfenati b , Roberto Todeschini c a Dipartimento di Elettronica e Informazione, Politecnico di Milano, Piazza Leonardo da Vinci 32, 20133 Milan, Italy b Laboratory of EnÕironmental Chemistry and Toxicology, Department of EnÕironmental Health Sciences, Istituto di Ricerche Farmacologiche ‘Mario Negri’, Milan, Italy c Dipartimento di Scienze dell’Ambiente e del Territorio, UniÕersita Statale di Milano, Milan, Italy ` Received 9 January 1998; accepted 21 June 1998 Abstract Ž . A hybrid expert system prototype using artificial neural networks ANN and classical rules has been developed for pre- dicting toxicology of compounds. Modularity was a must for the architecture of the system. The study of chemicals was approached by establishing classes. When appropriate descriptors are calculated for the molecule, the ANN classifier assigns the chemical class to the compound. Then the toxic activity is quantitatively predicted of by one of the trained ANN in the Ž . system. After that, a qualitative prediction activernon-active is made by a rule-based system, calling only the correct Ž . knowledge base KB for the assigned class. This last step enabled us to give an explanation of the results. All the rules in the KBs have been obtained with automated learning techniques. q 1998 Elsevier Science B.V. All rights reserved. Keywords: Toxicology; Expert systems; Artificial neural networks; Feature selection; QSAR models; WHIM descriptors; Automated rule extraction 1. Introduction The increasing number of pollutants in the envi- ronment raises the problem of toxicological charac- terization of these chemicals. Toxicology is the sci- ence that defines the limits of safety of chemical wx agents 1 . The traditional way of assessing the toxic risk of a compound is to test them in animals. The results are then extended to humans using safety factors and dose relationships. ) Corresponding author. Tel.: q39-2-2399-3626; Fax: q39-2- 2399-3411; E-mail: gini@elet.polimi.it This approach, however, suffers many drawbacks wx 2: Ž Ø cost of the experiments ) 1 million US$ per . compound ; Ž . Ø the duration of the tests 3–5 years ; Ø public pressure to reduce or eliminate the use of animals in scientific experiments. To overcome these problems, several computer- aided tools have been developed to help toxicologists assess the risk for new compounds. Many algorithms have been proposed to explain toxic effects in different situations where homoge- neous classes of chemicals showed various activities. However, in most cases these algorithms are suitable 0169-7439r98r$ - see front matter q 1998 Elsevier Science B.V. All rights reserved. Ž . PII: S0169-7439 98 00125-7