In the framework of pesticide risk assessment, a fragment-based QSAR approach is presented to correlate LC50-96 h acute toxicity to the rainbow trout (Oncorhynchus mykiss). While there are other fragment-based modeling routes, our approach exploits the possibility of prioritizing fragments' contributions to toxicity. On the assumption that one fragment might be mainly responsible for the molecular toxicity, we developed a three-stage modeling strategy to select the most important moieties and to establish their priorities at a molecular level. This strategy was tested on a heterogeneous dataset containing 282 pesticides, collected under the EU-funded project Demetra. Quantitative toxicity prediction yielded good results for the training set (R2TR = 0.85) and the test set (R2TS = 0.75). The advantages and limitations of the current priority strategy are examined.