Discovery of medicinal molecules based on similarity networks

Abstract

At a molecular scale, ligands work by binding to a therapeutic target of interest, usually a protein, to solicit a response or block its function. Many modern drug discovery projects start with a computational search for these "active" molecules, called Virtual Screening (VS). Big data techniques are adequate for such problems that involve huge molecular datasets. In this dissertation we present a Ligand-Based Virtual Screening (LBVS) approach leveraging the parallelism and scalability of Apache Spark. We created multiple molecular similarity networks to discover new putative ligand binders for a given protein target, and to find other proteins which are likely to interact with a given ligand. For network abstraction weusedGraphFrames, which to our best knowledge, this is the first work that uses this approach for tackling VS problems. Weevaluatedwithsuccessourapproachby comparing its enrichment and recall with similar tools, using A Database of Useful Decoys: Enhanced(DUD-E)dataset. A web application for scientists was created to run in silico experiments using ChEMBL data and visualise the results. Finally, we used our web application to run a VS experiment to find putative targets for maltanedienol, which was not in our dataset. We correctly found oestrogen receptor and rejected farnesyl pyrophosphate synthase as protein targets, which were also confirmed via in vitro testing.