PARENT SESSION

1F - QSAR
Poster Hall
8:30 AM - Tuesday, 29 April 2003
Chair: Schüürmann, G.¹, ¹
Co-chair: Verhaar, H.J.M.², Cronin, M.³, ^{2 3}

(TUP/40) Prediction of Chemical Safety with Neural Network.

Tanabe, Kazutoshi¹, Ohmori, Norihito¹, Ono, Shuichiro¹, Suzuki, Takahiro², Matsumoto, Takatoshi³, Uesaka, Hiroyuki⁴, ¹ Chiba Institute of Technology, Narashino, Chiba, Japan² Toyo University, Bunkyo-ku, Tokyo, Japan³ Tohoku University, Sendai, Miyagi, Japan⁴ Toyama University of International Studies, Ohyama-cho, Toyama, Japan

ABSTRACT- Numerous chemicals have been produced by mankind, but recently their effects on global environment are serious matter. Among chemicals available in market, quite few are ascertained on safety. Animal test to get data on chemical safety takes much cost and period. So it is impossible to get safety data on every unascertained chemicals with animal test. Here a computer prediction with quantitative structure-activity relationship (QSAR) is meaningful for screening. Several systems to predict chemical safety with QSAR have been developed, but the performances of existing systems are all low. Because those systems are based on simple linear analysis methods between chemical structures and safety. So we examined the performance of a neural network as a nonlinear analysis method. A neural network was applied to the prediction of the carcinogenicity of many organic compounds using data of the Challenge contest. Three groups of descriptors, Dragon, tReymers, and Helma were tested. These descriptors were entered into the input layer of a three-layered neural network, and carcinogenicity data were entered into the output layer as teaching data. The network was trained with an error-back-propagation method using 324 training sets, and a test using 168 test sets was carried out to evaluate the performance of this system. The correct classification rate using 839 Dragon descriptors was 70%, and that using 24 Helma descriptors was 71%. On the other hand, it was 85% using only 18 tReymers descriptors. This figure is quite higher any of the contestants. This demonstrates the superiority of a neural network as a nonlinear analysis method. We are going to develop an open system to predict chemical safety based on this result.

Key words: carcinogenicity, QSAR, neural network