Oversampling to overcome overfitting: exploring the relationship between data set composition, molecular descriptors, and predictive modeling methods.

The traditional biological assay is very time-consuming, and thus the ability to quickly screen large numbers of compounds against a specific biological target is appealing. To speed up the biological evaluation of compounds, high-throughput screening is widely used in the fields of biomedical, biological information, and drug discovery. The research presented in this study focuses on the use of support vector machines, a machine learning method, various classes of molecular descriptors, and different sampling techniques to overcome overfitting to classify compounds for cytotoxicity with respect to the Jurkat cell line. The cell cytotoxicity data set is imbalanced (a few active compounds and very many inactive compounds), and the ability of the predictive modeling methods is adversely affected in these situations. Commonly imbalanced data sets are overfit with respect to the dominant classified end point; in this study the models routinely overfit toward inactive (noncytotoxic) compounds when the imbalance was substantial. Support vector machine (SVM) models were used to probe the proficiency of different classes of molecular descriptors and oversampling ratios. The SVM models were constructed from 4D-FPs, MOE (1D, 2D, and 21/2D), noNP+MOE, and CATS2D trial descriptors pools and compared to the predictive abilities of CATS2D-based random forest models. Compared to previous results in the literature, the SVM models built from oversampled data sets exhibited better predictive abilities for the training and external test sets.