Mode of action-based local QSAR modeling for the prediction of acute toxicity in the fathead minnow.

The ultimate intention of quantitative structure-activity relationship (QSAR) study in toxicology is to predict the toxic potential of untested compounds with great accuracy. As QSAR has been based on the assumption that compounds from the same chemical domain will behave in similar manner, the QSAR model built upon the analogical chemicals is hypothesized to exhibit better performance than that derived from the miscellaneous data set. In this paper, the acute toxicity, 96 h LC(50) (median lethal concentration) for the fathead minnow from database EPAFHM_v2a_617_1Mar05 served as the interested toxicity endpoint, and the mode of action (MOA) in toxic response was employed as a criterion to compartmentalize the chemical domains. MOA-based local QSAR models were built by partial least squares (PLS) regression for each subset with single mode of action such as Narcosis I, Narcosis II or Reactive, and global model was also developed for the combined data set containing several subsets above. By comparing the performances of these two types of models, the local models were superior to the global model in that the relative standard error (R.S.E.) of the former was much lower for both the training set and the test set of any subset. In addition, the influence of the reliability of MOA determination on the performance of local model was also investigated and the statistical results for subsets with MOAs at A and B confidence level were better than those at C and D confidence level. Therefore, the MOA-based local QSAR models are promising to improve the accuracy of toxicity prediction as long as the assessment of MOA is of high reliability.