Regression analysis such as linear regression and logistic regression has often been employed to construct toxicogenomic predictive models, which forecast toxicological effects of chemical compounds in human or animals based on gene expression data. While in general these techniques can generate an accurate and sparse model when a regularization term is added to a loss function, they ignore structural relationships behind genes which form vast regulatory networks and interact with each other. Recently, several reports proposed structured sparsity-inducing norms to incorporate prior structural information and make a model reflecting relationships between variables. In this study, assuming that genes regulated by the same transcription factor should be selected together, we applied the latent group Lasso technique on toxicogenomic data with transcription factor networks as prior knowledge. We compared generated classifiers for liver weight gain in rats between the latent group Lasso and Lasso. The latent group Lasso was comparable or superior to the Lasso in terms of predictive performances (balanced accuracy: 74% vs. 72%, sensitivity: 62% vs. 62%, specificity: 86% vs. 83%). Besides, groups selected by the latent group Lasso suggested involvement of Wnt/β-catenin signaling pathway. Such mechanism-related analysis could not have been possible with the Lasso and is one of the advantages of the latent group Lasso.