Prediction of toxicity of aliphatic carboxylic acids using adaptive neuro-fuzzy inference system

Toxicity of 38 aliphatic carboxylic acids was studied using non-linear quantitative structure-toxicity relationship (QSTR) models. The adaptive neuro-fuzzy inference system (ANFIS) was used to construct the nonlinear QSTR models in all stages of study. Two ANFIS models were developed based upon different subsets of descriptors. The first one used log ow K and LUMO E as inputs and had good prediction ability; for the training set of 28 compounds 2 Training R was 0.86 and for the test set of 10 compounds, the corresponding statistic was 2 Test R =0.97. Two outliers were detected for this ANFIS model and removing them improved the quality of the model. Another ANFIS model was constructed based on PEOE_VSA_FPNEG and G3u descriptors chosen by exhaustive search of all two combinations of calculated descriptors by Dragon and MOE softwares. The later ANFIS model showed better performance than the former ( 2 Training R =0.92 and 2 Test R =0.90) and no outlier was detected.