Using Bayesian modeling on molecular fragments features for virtual screening

Virtual screening enables to search large small-molecule compound libraries for active molecules with respect to given macromolecular target. In ligand-based virtual screening, this goal is achieved by utilizing information about fragments or patterns present in existing known active compounds. Typically, the patterns are encoded as fingerprints which are used to screen a database of candidate compounds. In this work, we introduce an approach which uses Bayesian inference to encode activity-related information. Unlike previous approaches, our method does not utilize simple fragments, but rather uses features of these fragments. For each molecule, we generate a set of molecular fragments and extract molecular features for each of them. Next, we remove correlated features and use the remaining ones to build a Bayes model of activity. To score a previously unseen molecule, the molecule's fragment feature vectors are passed to the model and a score is obtained as the aggregation of their probability scores. When screening a database, this score is used to rank the compounds database. We show on datasets with various levels of difficulty that using fragments features rather then fragments themselves results in improvement of retrieval rates with respect to the best state-of-the art molecular fingerprints.