Support Vector Machines In The QSAR Study Of Some Dioxin-like Organic Pollutants

Support vector machine (SVM), a novel type of learning machine has been used for the quantitative structure & activity relationship (QSAR) study of dioxin-like organic pollutants: PXDDs (polychlorinated dibenzo-p-dioxins (PCDDs), polybrominated dibenzo-p-dioxins (PBDDs), polybrominated- chlorinated dibenzo-p-dioxins (PBCDDs) ), Polybrominated diphenyl ethers (PBDEs) and Polychlorinated naphthalenes (PCNs).With molecular weight and 8 net atomic charge parameters, which have been selected from more than 80 quantum chemical parameters, as molecular descriptors, SVM was used for the first time to develop a QSAR study of organic pollutants for the Ah receptor binding affinities of 14 polychlorinated PCDDs, 8 PBDDs and 3 PBCDDs. Partial Least Squares (PLS) has been utilized to compare with the results obtained by SVM. For predicting the binding affinities, the SVM model with squared cross-validated correlation coefficients (q2) of 0.842 outperforms the PLS model with q2 of 0.712. Furthermore, the SVM model with selected descriptors performs far better than the one with original 82 descriptors.By the use of partial least square with leave-one-out cross-validation, five net atomic charge descriptors and the first-order hyper-polarizability have been extracted from more than 80 quantum descriptors for predicting the aryl hydrocarbon receptor relative binding affinities (RBA) of PBDEs. Using the SVM and the radial basis function neural network (RBFN), the RBAs of 18 PBDE congeners have been correlated with the extracted 6 quantum chemical descriptors. The SVM model performs well in avoiding the over-training. The q2 for the SVM and RBFN models are 0.841 and 0.927, respectively.Then, SVM has been also used to establish QSAR models with three sets of toxicity data of the dioxin-like toxicities (relative potencies) of 20 PCNs as dependent variable, while the descriptors are selected from more than 90 quantum descriptors including molecular polarizabilities, molecular weight (Mw), net charge on each atom (Q) and electrostatic potential on each atom (ESP). The q2 of the three models are 0.805, 0.890 and 0.936, respectively. Furthermore, PLS has been utilized to compare with the results obtained by SVM.