Background: Quantitative structure-activity relationship (QSAR) models have become popular tools to help identify promising lead compounds in anticancer drug development. Few QSAR studies have investigated multitask learning, however. Multitask learning is an approach that allows distinct but related data sets to be used in training. In this paper, a suite of three QSAR models is developed to identify compounds that are likely to (a) exhibit cytotoxic behavior against cancer cells, (b) exhibit high rat LD50 values (low systemic toxicity), and (c) exhibit low to modest human oral clearance (favorable pharmacokinetic characteristics). Models were constructed using Kernel Multitask Latent Analysis (KMLA), an approach that can effectively handle a large number of correlated data features, nonlinear relationships between features and responses, and multitask learning. Multitask learning is particularly useful when the number of available training records is small relative to the number of features, as was the case with the oral clearance data.
Results: Multitask learning modestly but significantly improved the classification precision for the oral clearance model. For the cytotoxicity model, which was constructed using a large number of records, multitask learning did not affect precision but did reduce computation time. The models developed here were used to predict activities for 115,000 natural compounds. Hundreds of natural compounds, particularly in the anthraquinone and flavonoids groups, were predicted to be cytotoxic, have high LD50 values, and have low to moderate oral clearance.
Conclusion: Multitask learning can be useful in some QSAR models. A suite of QSAR models was constructed and used to screen a large drug library for compounds likely to be cytotoxic to multiple cancer cell lines in vitro, have low systemic toxicity in rats, and have favorable pharmacokinetic properties in humans.