The paradigm of chemical safety assessment is shifting from 'chemical management focusing on single chemicals' to 'product management extending to mixtures and articles'. However, because of the enormous combinatorial complexity, testing the toxicity of all conceivable mixture products is currently not feasible. There exist only few models that allow predicting the synergistic toxicity potentially caused by toxicological interactions among components. In this study, we present a novel approach to qualitatively predict the synergistic toxicity of binary mixtures to Vibrio fischeri. On the basis of information derived from protein-chemical and protein-protein interaction networks, we trained machine learning models for classifying chemical mixtures to have synergistic or nonsynergistic toxicity with accuracies and an area under the receiver operating characteristic (ROC) curve (AUC) up to 0.73. The numbers of shared targets and their neighborhood were found to be the most important features for classifying chemicals into synergistic and nonsynergistic groups.