An ensemble model of QSAR tools for regulatory risk assessment

Prachi Pradeep, Richard J. Povinelli, Shannon White, Stephen J. Merrill

Quantitative structure activity relationships (QSARs) are theoretical models that relate a quantitative measure of chemical structure to a physical property or a biological effect. QSAR predictions can be used for chemical risk assessment for protection of human and environmental health, which makes them interesting to regulators, especially in the absence of experimental data. For compatibility with regulatory use, QSAR models should be transparent, reproducible and optimized to minimize the number of false negatives. In silico QSAR tools are gaining wide acceptance as a faster alternative to otherwise time-consuming clinical and animal testing methods. However, different QSAR tools often make conflicting predictions for a given chemical and may also vary in their predictive performance across different chemical datasets. In a regulatory context, conflicting predictions raise interpretation, validation and adequacy concerns. To address these concerns, ensemble learning techniques in the machine learning paradigm can be used to integrate predictions from multiple tools. By leveraging various underlying QSAR algorithms and training datasets, the resulting consensus prediction should yield better overall predictive ability. We present a novel ensemble QSAR model using Bayesian classification. The model allows for varying a cut-off parameter that allows for a selection in the desirable trade-off between model sensitivity and specificity. The predictive performance of the ensemble model is compared with four in silico tools (Toxtree, Lazar, OECD Toolbox, and Danish QSAR) to predict carcinogenicity for a dataset of air toxins (332 chemicals) and a subset of the gold carcinogenic potency database (480 chemicals). Leave-one-out cross validation results show that the ensemble model achieves the best trade-off between sensitivity and specificity (accuracy: 83.8 % and 80.4 %, and balanced accuracy: 80.6 % and 80.8 %) and highest inter-rater agreement [kappa (κ): 0.63 and 0.62] for both the datasets. The ROC curves demonstrate the utility of the cut-off feature in the predictive ability of the ensemble model. This feature provides an additional control to the regulators in grading a chemical based on the severity of the toxic endpoint under study.