Ligand-based virtual screening (LBVS) uses machine readable representations of chemicals to learn a mapping function that can predict binding interactions with protein labels. Because it is highly scalable it is increasingly used in drug development in academic and pharmaceutical contexts. We have identified assumptions commonly used in LBVS that are false, which collectively can be described as the missing label problem. Firstly, many of the binding interactions in the bioactivity databases typically used to train LBVS models have never been tested before, but the absence of a label is interpreted by most models as a true negative. Secondly, many proteins have multiple binding sites with unrelated shapes but the associated ligands are grouped together under the one protein label. These assumptions frustrate the ability of the model to learn a correct mapping function. Here we use statistical techniques to predict values for the missing labels and binding sites and show how this improves the ability of LBVS models to rank ligands correctly. In the process we introduce a new technique for removing bias during model evaluation based on data blocking from experimental design theory. All data and code for analysis and generating figures is publicly available on github (https://github.com/ljmartin/Missing_label_problem).
InstitutionThe University of Sydney
ORCID For Submitting Author0000-0002-8621-4258
Declaration of Conflict of InterestNo conflict of interest.