Prediction of chemical toxicity with local support vector regression and activity-specific kernels

We propose a new kernel, based on 2-D structural chemical similarity, that integrates activity-specific information from the training data, and a new approach to applicability domain estimation that takes feature significances and activity distributions into consideration. The new kernel provides superior results than the well-established Tanimoto kernel, and activity-sensitive feature selection enhances prediction quality. Validation of local support vector regression models based on this kernel has been preformed with three publicly available datasets from the DSSTox project. One of them (Fathead Minnow Acute Toxicity) has been already modelled by other groups, and serves as a benchmark dataset, the other two (Maximum Recommended Therapeutic Dose, IRIS Lifetime Cancer Risk) have been modelled for the first time according to the knowledge of the authors. For all three models predictive accuracies increase with the prediction confidences that indicate the applicability domain. Depending on the confidence cutoff for acceptable predictions we were able to achieve > 90% predictions within 1 log unit of the experimental data for all datasets.