In the early stage of drug discovery, especially for computer-aided drug design, a large number of molecules will be proposed as potential leads and the bioactivity risk of these molecules is expected to be evaluated prior to synthesis. The rule-based expert systems have been used for the aim, while mining of a large amount of toxicological data can provide us with another promise. In the present study, we first make an initial analysis to a toxicological database of chemicals. Then a stepwise strategy is provided to data mine of toxic chemicals, which combines QSAR study with structure pattern clustering. Finally, on the basis of the analysis and aiming to be applied in the early stage of drug design, we propose one system as an alternative to toxicity prediction of chemicals, which is based on the database and dynamically combined with the database.