Title: Docking vs Screening vs the Gobsmacking Emptiness of Chemical Space

Abstract:
Why are hit rates from virtual and high-throughput screens, particularly for antibiotic and genomic targets, so low? Are low hit rates explained by the well-known problems with docking, with the difficulties of the targets (undruggable targets), challenges of artifacts in screening, or by problems with the library? Here we attempt to disentangle this question with a series of matched computational, experimental, and crystallographic studies. We first compare, prospectively, structure-based docking screens to high-throughput screesn on exactly the same target (AmpC ß-lactamase and Cruzain) with exactly the same library (the MLSMR). Once the artifactual hits are removed—no easy task for some of them—the hit rates for both techniques is depressingly low. This argues for problems in the library, at least for this target. Consistent with this idea, we find much higher hit rates when docking a library of 167,000 fragments against the AmpC structure, where the hit rates rise to close to 50%; representative examples are confirmed by crystallography. Intriguingly we find similarly high hit rates for much a more traditional library of leadlike molecules when docking against more precedented targets, such as the structures of the GPCRs ß2-adrenergic receptor and A2a. Several results emerge: for unprecendented targets, like ß-lactamase and cruzain, hit rates are low because of lack of bias in our libraries. A way to overcome this is with fragment-based screens, for which structure-based docking, at least for this and related targets, seems well-suited. Conversely, our libraries and docking strategies are well-suited to precedented targets, like the small-molecule GPCRs.

Stepping back, however, one might easily wonder not why our hit rates are so low for screening and docking, but why they ever work at all? Chemical space is so vast and so undersampled that one might easily imagine that one would never get any hits at all in a screen. The fact that we do suggests some properties about the “unbiased” libraries that we are screening; several of these are considered from a theoretical standpoint, and implications for library design are proposed.