Chronic Rat Toxicity Prediction of Chemical Compounds Using Kernel Machines

Abstract

A recently published study showed the feasibility of chronic rat toxicity prediction, an important task to reduce the number of animal experiments using the knowledge of previous experiments. We benchmarked various kernel learning approaches for the prediction of chronic toxicity on a set of 565 chemical compounds, labeled with the Lowest Observed Adverse Effect Level, and achieved a prediction error close to the interlaboratory reproducibility. ε-Support Vector Regression was used in combination with numerical molecular descriptors and the Radial Basis Function Kernel, as well as with graph kernels for molecular graphs, to train the models. The results show that a kernel approach improves the Mean Squared Error and the Squared Correlation Coefficient using leave-one-out cross-validation and a seeded 10-fold-cross-validation averaged over 10 runs. Compared to the state-of-the-art, the Mean Squared Error was improved up to MSE_loo of 0.45 and MSE_cv of 0.46±0.09 which is close to the theoretical limit of the estimated interlaboratory reproducibility of 0.41. The Squared Empirical Correlation Coefficient was improved to $\text{Q}^2_{\text{loo}}$ of 0.58 and $\text{Q}^2_{\text{\text{cv}}}$ of 0.57±0.10. The results show that numerical kernels and graph kernels are both suited for predicting chronic rat toxicity for unlabeled compounds.