The difficulties of developing predictive computational models of toxicity are discussed in relation to their internal and external validation, the selection of relevant physicochemical data and the need to characterise the structure-activity relationship landscapes obtained with training sets of chemicals by using recently published methods. It is concluded that the developers of in silico systems for toxicity prediction should apply such methods to ensure adequate and continuous sampling of chemical space, especially when external validation cannot be undertaken due to lack of sufficient test chemicals not used in the training set. This, combined with discriminate selection of molecular descriptors, and the use of reliable toxicity data, should improve model predictivity.