Undersampling: case studies of flaviviral inhibitory activities

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Infectious diseases and virology

Background:

• Imbalanced datasets are a significant challenge in drug discovery, particularly for neglected tropical diseases.
• Effective modeling requires robust strategies to handle datasets with many more inactive than active compounds.

Purpose of the Study:

• To evaluate undersampling strategies for building predictive models of Dengue Virus 2 (DENV2), West Nile Virus (WNV), and Zika Virus (ZIKV) inhibitory activities.
• To compare the performance of data pruning versus data selection algorithms in this context.

Main Methods:

• Development of ensemble classifiers using GT-STAF information indices and molecular fragmentation approaches (connected subgraphs, substructure, alogp atom types).
• Application and comparison of various undersampling algorithms on curated DENV2, WNV, and ZIKV datasets.
• Evaluation of model performance on external test sets using balanced accuracy (BACC).

Main Results:

• Data pruning algorithms outperformed data selection algorithms in predictive modeling.
• The one-sided selection algorithm achieved the best overall performance, yielding BACC values of 0.84 (DENV2), 0.74 (WNV), and 0.77 (ZIKV).
• Combining molecular fragmentation strategies enhanced ensemble predictivity.

Conclusions:

• Undersampling, particularly the one-sided selection method, is crucial for developing accurate predictive models from imbalanced datasets in drug discovery.
• The developed models show potential for screening compounds against DENV, WNV, and ZIKV.
• ADMET modelers should integrate undersampling techniques into their workflows for imbalanced data challenges.