[Genetic algorithms in 3D-QSAR: The use of multiple ligand orientations for improved predictions of toxicity]

In a recent article in this journal, we discussed the application of a 3D-QSAR technique to the prediction of the toxicity of dibenzodioxins, dibenzofurans, and biphenyls (Vedani et al., 1999a). The use of such methods is legitimate because there is strong evidence that the toxicity is mediated by the Aryl hydrocarbon (Ah) receptor, a regulatory element involved in the mammalian metabolism of xenobiotics. In an extention to a concept developed at our laboratory (Vedani et al., 1998), we now show that safer predictions are possible if instead of a single - and, therefore, necessarily biased - assumption about the mutual orientation of the toxins, an ensemble of possible orientations is used for model construction. The contribution of a single entity within this ensemble to the toxin-receptor interaction energy is determined by a Boltzmann criterion. While in the single-orientation model two of the 26 toxins of an external test set were predicted false positive or false negative, all test substances are correctly predicted in the multiple-orientation model - including up to four different orientations per molecule - within a factor 10 of the experimental binding affinity. These results demonstrate that 3D-QSAR techniques based on a genetic algorithm can be used to predict the toxicity of chemical and pharmacological substances "in computo" if a receptor-mediated mechanism can be assumed. Consequently, the method can be used in toxicological screening assays, thereby replacing stressful tests on animals.