Virtual screening is widely used to discover new chemical tools and leads for drug discovery. Unfortunately, the technique remains difficult to use, and has thus been restricted to a few expert laboratories. Here, we create databases and tools to bring virtual screening to a wide biological audience, much expanding its impact and usefulness, and develop a chemoinformatics method to identify the “on” and “off” targets for drugs and reagents.
A long-term goal is to bring chemical reagents to biology, extending the development of tools on which the community increasingly depends: ZINC (http://zinc.docking.org
), DUDE (http://dude.docking.org
), DOCK Blaster (http://blaster.docking.org
) and SEA (http://sea.docking.org
) among them. A second goal explores the fundamental bases and implications of a ligand-based organization of pharmacology, leveraging it to predict biologically-relevant polypharmacology, we argue for the first time in the field. 1. New public tools to bring chemistry to biology. A. We develop ZINC tools that enable one to input a target and find all available reagents known for it, or to input a molecule to find its known and predicted targets, testing several of these. B. Targets operate in pathways and networks. We introduce tools that allow one to island hop from target-to-target in “clickable” networks organized by both chemo- and bio-informatic similarity. C. By bringing chemoinformatics directly into DOCK Blaster, investigators can search their hit lists for analogs, scaffolds, functional groups, and their docked interactions. D. We develop a library that addresses the key problem of phenotypic screening, target ID, by pre-annotating all library compounds to targets, with each target having two or more orthogonal molecules annotated (with Novartis & Sigma-Aldrich). When such orthogonal molecules share a phenotype in a screen, it suggests the underlying target. 2. Comparing and combining ligand-based, structure-based & bioinformatic similarity. Linking targets by ligand similarity reveals associations very different from what bioinformatics would suggest. This is gratifying but puzzling. Here we A. Comprehensively compare bioinformatic target networks to those predicted by ligand similarity. Preliminary results suggest that these networks are mostly orthogonal but have intriguing areas of overlap (explored in C, below). B. To understand the structural basis for the binding of identical ligands by “unrelated” proteins, we compare the x-ray structures of hundreds of complexes where two dissimilar proteins bind exactly the same ligands, using widely-used, structure-based site comparison programs. Can these programs recognize the binding sites in the unrelated proteins as, in fact, similar? How many ways can proteins recognize the same ligand functional groups? C. In the areas where the bio- and chemoinformatics neighborhoods overlap, the bioinformatics implicates pairs of targets in a disease, while chemoinformatics suggest that that the pair can be co-modulated by a reagent. For these pairs, shared polypharmacology is functionally meaningful. We predict and test 50. Whereas these goals are ambitious, their plausibility is supported by extensive preliminary results.