Generally, many of QSAR models with a training set that consisted of structurally diverse compounds don't give us good prediction for the prediction set. In this paper, in order to resolve such a problem and explore higher performance in the prediction, we have suggested two alternative approaches for building the QSAR models. One is Cluster-based Modeling (CM) and the other is Similarity-based Modeling (SM). To compare the predictive performances of the two methods and that of the ordinal approach, structure-toxicity modeling was carried out with an aquatic toxicity database. Here the authors used TFS (Topological Fragment Spectra) for the numerical description of chemical structure information, which represents the topological structure profile of a chemical compound. Partial Least-Squares (PLS) method was employed for building the QSAR model. It is concluded that both the cluster-based modeling and the similarity-based modeling could be more successful than the ordinal approach, although the former has some difficulties to identify the potential major clusters and for the latter the result depends on how to find proper neighbors of the particular sample to be predicted, and especially the similarity-based modeling would be more useful when dealing with a set of structurally diverse compounds and when increasing the available data all the time.