We consider a wide range of ligand-based machine learning approaches for virtual screening on two targets, and present a strategy for choosing which algorithm is best for prospective compound prioritization. Recent studies report superior predictive performance from both multi-task models and deep learning models. In contrast, our workflow identifies a simple single-task random forest model as the best algorithm for these targets over the more sophisticated neural network-based models. This reinforces that the ideal virtual screening method can depend on the training data available, target attributes, and other factors. Therefore, careful target-specific cross-validation is important to optimize prospective performance. One cannot assume that deep learning models will be dominant for all targets and all virtual screening scenarios.

Practical approach for selecting a virtual screening model

2019 Spring National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.